Subset (2)

default · 33.4k rows

Split (3)

train · 28.4k rows

text stringlengths 2.85k 2.55M	label class label 11 classes
arXiv:1604.04973v1 [math.GR] 18 Apr 2016 Factorization numbers of finite rank 3 abelian p-groups Marius Tărnăuceanu March 22, 2016 Abstract In this short note we give a formula for the factorization number F2 (G) of a finite rank 3 abelian p-group G. This extends a result in our previous work [9]. MSC (2010): Primary 20D40; Secondary 20D60. Key words: factorization number, subgroup commutativity degree, Möbius function, finite abelian p-group. 1 Introduction Let G be a finite group, L(G) be the subgroup lattice of G and H, K be two subgroups of G. If G = HK, then G is said to be factorized by H and K and the expression G = HK is said to be a factorization of G. Denote by F2 (G) the factorization number of G, that is the number of all factorizations of G. The starting point for our discussion is given by the papers [6, 1], where F2 (G) has been computed for certain classes of finite groups. Then, in [9], we have obtained explicit formulas for the factorization numbers of an elementary abelian p-group and of a rank 2 abelian p-group. These are based on the connection between F2 (G) and the subgroup commutativity degree sd(G) of G (see [7, 8]), namely sd(G) = X 1 F2 (H). \|L(G)\|2 H≤G 1 Obviously, by applying the well-known Möbius inversion formula to the above equality, one obtains X (1) F2 (G) = sd(H)\|L(H)\|2µ(H, G). H≤G In particular, if G is abelian, then we have sd(H) = 1 for all H ∈ L(G), and consequently X X (2) F2 (G) = \|L(H)\|2µ(H, G) = \|L(G/H)\|2 µ(H). H≤G H≤G This formula will be used in the following to calculate the factorization number of a rank 3 abelian p-group. First of all, we recall a theorem due to P. Hall [2] (see also [4]), that permits us to compute explicitly the Möbius function of a finite p-group. Theorem 1. Let G be a finite p-group of order pn . Then µ(G) = 0 unless n G is elementary abelian, in which case we have µ(G) = (−1)n p ( 2 ) . We also need to know the total number of subgroups of a finite rank 3 abelian p-group. It has been determined by using different methods in [3, 5, 10]. Theorem 2. The total number of subgroups of Zpλ1 ×Zpλ2 ×Zpλ3 , where λ1 ≥ λ2 ≥ λ3 ≥ 1, is (3) f (λ1 , λ2 , λ3 ) = (p2 A , − 1)2 (p − 1) where A =(λ3 +1)(λ1 −λ2 +1)pλ2 +λ3 +5 +2(λ3 +1)pλ2 +λ3 +4 −2(λ3 +1)(λ1 −λ2 )pλ2 +λ3 +3 −2(λ3 +1)pλ2 +λ3 +2 +(λ3 +1)(λ1 −λ2 −1)pλ2 +λ3 +1 −(λ1 +λ2 −λ3 +3)p2λ3 +4 −2p2λ3 +3 +(λ1 +λ2 −λ3 −1)p2λ3 +2 +(λ1 +λ2 +λ3 +5)p2 +2p−(λ1 +λ2 +λ3 +1). We are now able to give the main result of our note. 2 Theorem 3. The following equality holds F2 (Zpλ1 ×Zpλ2 ×Zpλ3 )=−p3 f 2 (λ1 −1, λ2 −1, λ3 −1)+p(f 2 (λ1 −1, λ2 −1, λ3 ) +pf 2 (λ1 −1, λ2 , λ3 −1)+p2 f 2 (λ1 , λ2 −1, λ3 −1)) −(f 2 (λ1 −1, λ2, λ3 )+pf 2 (λ1 , λ2 −1, λ3 ) +p2 f 2 (λ1 , λ2 , λ3 −1))+f 2 (λ1 , λ2 , λ3 ), where the quantities f (λ1 , λ2 , λ3 ) are given by (3). We remark that for λ3 = 0 the above equality leads to the formula in Theorem 3 of [9]. It simplifies in the particular case λ1 = λ2 = λ3 = λ. Corollary 4. We have F2 (Zpλ ×Zpλ ×Zpλ )=−p3 f 2 (λ−1, λ−1, λ−1)+p(1+p+p2 )f 2 (λ, λ−1, λ−1) −(1+p+p2 )f 2 (λ, λ, λ−1)+f 2 (λ, λ, λ). However, even in this case, an explicit formula for the factorization number is too difficult to be written, but we can do it for small values of λ1 , λ2 , λ3 . Examples. a) F2 (Zp3 ×Zp2 ×Zp ) = 9p6 + 15p5 + 21p4 + 16p3 + 20p2 + 11p + 13. b) F2 (Zp2 ×Zp2 ×Zp2 ) = 5p8 +7p7 +16p6 +15p5 +21p4 +16p3 +20p2 +11p+13. 2 Proof of Theorem 3 It is well-known that G = Zpλ1 ×Zpλ2 ×Zpλ3 has a unique elementary abelian subgroup of order p3 , say M, and that G/M ∼ = Zpλ1 −1 ×Zpλ2 −1 ×Zpλ3 −1 , = Φ(G) ∼ where Φ(G) is the Frattini subgroup of G. Moreover, all elementary abelian subgroups of G are contained in M. Denote by Mi , i = 1, 2, ..., p2 + p + 1, the subgroups of order p and by Mi′ , i = 1, 2, ..., p2 + p + 1, the subgroups of order p2 in M. Then every quotient G/Mi is isomorphic to a maximal subgroup of G, which are: p2 of type (λ1 , λ2 , λ3 − 1), p of type (λ1 , λ2 − 1, λ3), and 3 1 of type (λ1 − 1, λ2 , λ3 ). Similarly, every quotient G/Mi′ is isomorphic to a subgroup of index p2 of G that contains Φ(G), and these are: p2 of type (λ1 , λ2 −1, λ3 −1), p of type (λ1 −1, λ2 , λ3 −1), and 1 of type (λ1 −1, λ2 −1, λ3 ). In this way the equality (2) becomes X p2 +p+1 F2 (G) =\|L(G/M)\|2 µ(M) + \|L(G/Mi )\|2 µ(Mi ) i=1 X p2 +p+1 + \|L(G/Mi′ )\|2 µ(Mi′ ) + \|L(G)\|2 µ(1), i=1 in view of Theorem 1. Since we have µ(M) = µ(Z3p ) = −p3 , µ(Mi ) = µ(Zp ) = −1, µ(Mi′ ) = µ(Z2p ) = p, ∀ i = 1, 2, ..., p2 + p + 1, and µ(1) = 1, one obtains F2 (G)=−p3 \|L(Zpλ1 −1 ×Zpλ2 −1 ×Zpλ3 −1 )\|2 +p(\|L(Zpλ1 −1 ×Zpλ2 −1 ×Zpλ3 )\|2 +p\|L(Zpλ1 −1 ×Zpλ2 ×Zpλ3 −1 )\|2 +p2 \|L(Zpλ1 ×Zpλ2 −1 ×Zpλ3 −1 )\|2 ) −(\|L(Zpλ1 −1 ×Zpλ2 ×Zpλ3 )\|2 +p\|L(Zpλ1 ×Zpλ2 −1 ×Zpλ3 )\|2 +p2 \|L(Zpλ1 ×Zpλ2 ×Zpλ3 −1 )\|2 )+\|L(Zpλ1 ×Zpλ2 ×Zpλ3 )\|2 . Under the notation of Theorem 2 this leads to the desired formula. References [1] M. Farrokhi D.G., Factorization numbers of finite abelian groups, Int. J. Group Theory 2 (2013), 1-8. [2] P. Hall, A contribution to the theory of groups of prime-power order, Proc. London Math. Soc. 36 (1933), 29-95. [3] M. Hampejs and L. Tóth, On the subgroups of finite abelian groups of rank three, Annales Univ. Sci. Budapest., Sect. Comp., 39 (2013), 111124. [4] T. Hawkes, I.M. Isaacs and M. Özaydin, On the Möbius function of a finite group, Rocky Mountain J. Math. 19 (1989), 1003-1033. [5] J.M. Oh, An explicit formula for the number of subgroups of a finite abelian p-group up to rank 3, Commun. Korean Math. Soc. 28 (2013), 649-667. 4 [6] F. Saeedi and M. Farrokhi D.G., Factorization numbers of some finite groups, Glasgow Math. J. 54 (2012), 345-354. [7] M. Tărnăuceanu, Subgroup commutativity degrees of finite groups, J. Algebra 321 (2009), 2508-2520. [8] M. Tărnăuceanu, Addendum to ”Subgroup commutativity degrees of finite groups”, J. Algebra 337 (2011), 363-368. [9] M. Tărnăuceanu, On the factorization numbers of some finite p-groups, accepted for publication in Ars Combin., arXiv:1502.04778. [10] M. Tărnăuceanu and L. Tóth, On the number of subgroups of a given exponent in a finite abelian group, accepted for publication in Publ. Inst. Math. Beograd, arXiv:1507.00532. Marius Tărnăuceanu Faculty of Mathematics “Al.I. Cuza” University Iaşi, Romania e-mail: tarnauc@uaic.ro 5	4math.GR
On deep-holes of Gabidulin codes arXiv:1711.11202v2 [cs.IT] 28 Feb 2018 Weijun Fanga Li-Ping Wangb† Daqing Wanc a Chern Institute of Mathematics and LPMC, Nankai University, Tianjin 300071, China Email: nankaifwj@163.com b† Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100093, China Data Assurance and Communications Security Research Center, Chinese Academy of Sciences, Beijing 100093, China, University of Chinese Academy of Sciences, Beijing, China, Email: wangliping@iie.ac.cn c Department of Mathematics, University of California, Irvine CA 92697, USA, Email: dwan@math.uci.edu Abstract In this paper, we determine a class of deep holes for Gabidulin codes with both rank metric and Hamming metric. Moreover, we give a necessary and sufficient condition for deciding whether a word is not a deep hole for Gabidulin codes, by which we study the error distance of two special classes of words to certain Gabidulin codes. Keywords: Gabidulin codes, rank metric, deep holes, covering radius, error rank distance 1 Introduction Let q be a power of a prime p, m ≥ 1 an integer and let Fnqm be an n-dimensional vector space over the finite field Fqm . In this paper we only consider the case when n ≤ m. Let β = (β1 , . . . , βm ) be a basis of Fqm over Fq . Let Fi be the map from Fqm to Fq where Fi (u) is the i-th coordinate of u in the basis β. To any u = (u1 , . . . , un ) in Fnqm , we associate the matrix ū ∈ Mm,n (Fq ) in which ūi,j = Fi (uj ). The rank weight of a vector u can be defined as the rank of the associated matrix ū, denoted by wR (u). Thus we can define the rank distance between two vectors u and v as dR (u, v) = wR (u − v). We refer to [18] for more details on codes for the rank distance. For integers 1 ≤ k ≤ n, a rank code C of length n and dimension k over Fqm is a subspace of dimension k of Fnqm embedded with rank metric. The minimum rank distance of the code C, denoted by dR (C), is the minimum rank weight of the non-zero vectors in C. A k × n matrix is called the generator matrix of C if its rows span the code C. A rank code of length n and dimension k over Fqm is called a maximum rank distance separable (MRDS) code if dR (C) = n − k + 1. The error rank distance of any word u ∈ Fnqm to C is defined as dR (u, C) = min{dR (u, c)\|c ∈ C}. The error rank distance plays an important role in the decoding of rank codes. The maximum error rank distance ρR (C) = max{dR (u, C)\|u ∈ Fnqm } is called the covering radius of C. If the rank distance from a word to the code C achieves the covering radius of the code, the word is called a deep hole of the code C. Covering radius and deep holes of linear code embedded with Hamming metric were studied extensively [1, 2, 3, 4, 5, 10, 12, 14, 16, 21, 22, 23, 24, 25, 26], in which MDS codes such as generalized Reed-Solomon codes, standard Reed-Solomon codes and projective Reed-Solomon codes were explored deeply. Gabiduli codes were first introduced by Gabidulin in [7] and independently by Delsarte in [6]. Gabidulin codes can be seen as the q-analog of Reed-Solomon codes. Furthermore, Gabidulin codes are MRDS codes. Over the last decade there has been increased interest in Gabidulin codes, mainly because of their relevance to network coding [15, 19]. Covering radius for Gabidulin codes also were studied in [8, 9, 20]. However, little is known about the deep holes for such codes. In this paper we determine a class of deep holes for Gabidulin codes with both rank metric and Hamming metric. Meanwhile we also obtain the covering radius by our method. Moreover, we give a necessary and sufficient condition for deciding whether a word is not a deep hole for Gabidulin codes, by which we study the error rank distance of a special class of words to certain Gabidulin codes. Note that we refer to rank metric if Hamming metric is not explicitly pointed out in this paper. 2 Linearized polynomials Gabidulin codes exploit linearized polynomials instead of arbitrary polynomials and so we recall some results about linearized polynomials. A q-linearized polynomial over Fqm is defined to be a polynomial of the form L(x) = d X i ai xq , ai ∈ Fqm , ad 6= 0 i=0 where d is called the q-degree of f (x), denoted by degq (f (x)). Note that L(x) has no constant term. One can easily check that L(x1 + x2 ) = L(x1 ) + L(x2 ) and L(λx1 ) = λL(x1 ) for any x1 , x2 ∈ Fqm and λ ∈ Fq , from which the name stems. In particular, L(x) induces an Fq -linear endomorphism of the Fq -vector space Fqm . The set of all q-linearized polynomials over Fqm is denoted by Lq (x, Fqm ). The ordinary product of linearized polynomials need not be a linearized polynomial. However, the composition L1 (x) ◦ L2 (x) = L1 (L2 (x)) is also a linearized polynomial. The set Lq (x, Fqm ) forms a non-commutative ring under the operations of composition ◦ and ordinary addition. It is also an Fq -algebra. Lemma 1. [17] Let f (x) ∈ Lq (x, Fqm ) and Fqs be the smallest extension field of Fqm that contains all roots of f (x). Then the set of all roots of f (x) forms an Fq -linear vector space in Fqs . 2 Let U be an Fq -linear subspace of Fqm . Then polynomial of U . Q g∈U (x − g) is called the q-annihilator Lemma 2. [17] Let U be an Fq -linear subspace of Fqm . Then polynomial over Fqm . Let β1 , . . . , βn ∈ Fqm and denote the k × n Moore  β1 β2  β1q β2q  Mk (β1 , . . . , βn ) :=  .. ..  . . q k−1 q k−1 β1 β2 Q g∈U (x−g) is a q-linearized matrix by ... ... .. . βn βnq .. . k−1 . . . βnq    .  Furthermore, if g1 , . . . , gn is a basis of U , one can write Y (x − g) = λ det(Mn+1 (g1 , . . . , gn , x)) g∈U for some non-zero constant λ ∈ Fqm . Clearly, its q-degree is n. In addition, we have the notion of q-Lagrange polynomials. Let g = {g1 , . . . , gn } ⊂ Fqm and r = {r1 , . . . , rn } ⊂ Fqm , where g1 , . . . , gn are Fq linearly independent. For 1 ≤ i ≤ n, we define the matrix Di (g, x) as Mn (g1 , . . . , gn , x) without the ith column. The q-Lagrange polynomial with respect to g and r is defined to be n X det(Di (g, x)) (−1)n−i ri Λg,r (x) = ∈ Fqm [x]. det(Mn (g)) i=1 Proposition 1. [13] The q-Lagrange polynomial Λg,r (x) is a q-linearized polynomial in Fqm [x] and Λg,r (gi ) = ri for i = 1, . . . , n. Proposition 2. [13] Let L(x) ∈ Lq (x, Fqm ) be such that Q L(gi ) = 0 for all i. Then there exists an H(x) ∈ Lq (x, Fqm ) such that L(x) = H(x) ◦ g∈<g> (x − g), where < g > is the Fq -vector space spanned by g. 3 Deep holes of Gabidulin codes Let g1 , . . . , gn ∈ Fqm be linearly independent over Fq , which also implies that n ≤ m. Let g = {g1 , . . . , gn } and < g > is the Fq -vector space spanned by g. A Gabidulin code G ⊆ Fnqm is defined as the linear block code with generator matrix Mk (g1 , . . . , gn ), where 1 ≤ k ≤ n. Using the isomorphic matrix representation, we can interpret G as a matrix . The rank distance is defined as above. code in Fm×n q The Gabidulin code G with length n has dimension k over Fqm and minimum rank distance n − k + 1, and so G is an MRDS code [7]. The Gabidulin code G can also be defined as follow: G = {(m(g1 ), . . . , m(gn )) ∈ Fnqm \| m(x) ∈ Lq (x, Fqm ) and degq (m(x)) < k}. 3 (1) Note that this interpretation of the code G will be used throughout the rest of the paper. It is the q-analogue of the generalized Reed-Solomon code. Let Y Y ( (x − g)) = Lq (x, Fqm ) ◦ (x − g) g∈<g> g∈<g> Q be the left ideal generated by the element g∈<g> (x − g) in the non-commutative ring Q Lq (x, Fqm ) with respect to the composition product. In particular, ( g∈<g> (x − g)) is an Q m m Fq -linear additive subgroup of Lq (x, Fq ). It follows that Lq (x, Fq )/( g∈<g> (x − g)) is an Fq -vector space. Define an Fq -linear evaluation map Y σ : Lq (x, Fqm )/( (x − g)) −→ Fnqm g∈<g> given by σ(f (x)) = (f (g1 ), . . . , f (gn )). We have the following property. Proposition 3. The above defined map σ is an Fq -vector space isomorphism. Q Proof. First, σ is well-defined since the polynomial g∈<g> (x − g) vanishes at every gi . Second, if f (gQ i ) = 0 for all i = 1, . . . , n, then there exists H(x) ∈ Lq (x, Fq m ) such that f (x) = H(x)◦ g∈<g> (x−g) by Proposition 2 and so σ is one-to-one. Third, we show that σ is surjective. For a given r = (r1 , . . . , rn ) ∈ Fnqm , we have the q-Lagrange polynomial Λg,r (x) satisfying Λg,r (gi ) = ri for i = 1, . . . , n by Proposition 1. The result is proved. Q The q-linearized polynomial g∈<g> (x−g) has q-degree n. It follows that any element f (x) ∈ Lq (x, Fqm ) can be written uniquely in the form Y f (x) = h(x) ◦ (x − g) + r(x), g∈<g> where h(x), r(x) ∈ Lq (x, Fqm ) and r(x) has q-degree smaller than n. This is the q-division algorithm inQthe non-commutative ring Lq (x, Fqm ). As Fq -vector spaces, the quotient Lq (x, Fm q )/( g∈<g> (x − g)) is thus represented by all q-linearized polynomials of q-degree less than n. That is, Y Lq (x, Fqm )/( (x − g)) = {f ∈ Lq (x, Fqm )\| degq (f ) < n}. g∈<g> Using the isomorphism σ, we can identify any word u ∈ Fnqm with σ(f ) for a unique polynomial f (x) ∈ Lq (x, Fqm ) with degq (f ) < n. When degq (f ) ≤ k − 1, it is easy to see that the error distance dR (σf , G) = 0 by the definition. Thus, when k ≤ degq (f ) < n, we provide the following upper and lower bounds for dR (σf , G). 4 Theorem 1. Let f (x) ∈ Lq (x, Fqm ) with degq (f ) < n and let σf = σ(f ) ∈ Fnqm be the corresponding word. If k ≤ degq (f ) < n, then n − degq (f ) ≤ dR (σf , G) ≤ n − k. Furthermore, we suppose f is monic, then dR (σf , G) = n − degq (f ) if and only if there exists a degq (f )-dimensional subspace H of < g > such that f (x) − v(x) = Y (x − h), h∈H for some v(x) ∈ Lq (x, Fqm ) with degq (v) ≤ k − 1. Proof. Applying the q-division algorithm, we can write Y f (x) − r(x) = h(x) ◦ (x − g), g∈<g1 ,··· ,gk > for some unique r(x) with degq (r) < k. This implies that dR (σf , G) ≤ n − k. The right hand of the inequality holds. For the lower bound, let u(x) be any q-polynomial over Fqm . We consider the Fq -linear map defined by πu : < g1 , · · · , gn >→< u(g1 ), · · · , u(gn ) > n n n X X X ξi gi ). ξi u(gi ) = u( ξi gi 7→ i=1 i=1 i=1 It is clear that the map πu is surjective and ker(πu ) ⊆ Root(u) (the set of roots of u(x)). So dimFq ker(πu ) ≤dimFq Root(u) ≤ degq (u). Then dimFq < u(g1 ), · · · , u(gn ) > = dimFq < g1 , · · · , gn > −dimFq ker(πu ) ≥ n − degq (u). Thus dR (σf , G) = = ≥ min rank((f − v)(g1 ), · · · , (f − v)(gn )) min dimFq < (f − v)(g1 ), · · · , (f − v)(gn ) > degq (v)<k degq (v)<k min (n − degq (f − v)) = n − degq (f ). degq (v)<k The last equality holds since degq (f − v) = degq (f ) for any q-polynomial v(x) with degq (v) < k. 5 Furthermore, from the above proof, we know that dR (σf , G) = n − degq (f ) if and only if dimFq Root(f − v) = dimFq ker(πf −v ) = degq (f − v) = degq (f ) for some q-polynomial v(x) with degq (v) < k, which is equivalent to f (x) − v(x) = Y (x − h), h∈H for some degq (f )-dimensional subspace H of < g1 , · · · , gn >. The theorem is proved. By Theorem 1, we immediately deduce the following two corollaries. Corollary 1. For the Gabidulin code G, the covering radius of G is n − k. This corollary had also been obtained in [9]. Corollary 2. The set {σf : degq (f (x)) = k, f (x) ∈ Lq (x, Fqm )} are deep holes of the Gabidulin code G and so the number of deep holes of G is at least (q m − 1)q mk . According to the definition in (1), we may also study the Gabidulin codes under the Hamming metric. It was showed that the Gabidulin codes are MDS codes in [7]. We use dH (u, v) and dH (u, G) to denote Hamming distance between vectors u and v and Hamming distance of a word u to G, respectively. Similarly, we have the following theorem. Theorem 2. Let f (x) ∈ Lq (x, Fqm ) with degq (f ) < n and let σf = σ(f ) ∈ Fnqm be the corresponding word. If k ≤ degq (f ) < n, then n − degq (f ) ≤ dH (σf , G) ≤ n − k. Furthermore, suppose f is monic, then dH (σf , G) = n − degq (f ) if and only if there exists a subset E = {gi1 , . . . , gidegq (f ) } of {g1 , · · · , gn } such that f (x) − v(x) = Y (x − g), g∈<E> for some v(x) ∈ Lq (x, Fqm ) with degq (v) ≤ k − 1. Proof. The proof of dH (σf , G) ≤ n − k is same as Theorem 1. Let now t = n − dH (σf , G). By definition of the Hamming distance, there exists some v(x) and non-zero H(x) ∈ Lq (x, Fqm ) with degq (v) < k such that f (x) − v(x) = H(x) ◦ Y g∈<gi1 ,...,git > 6 (x − g) for some indices 1 ≤ i1 < · · · < it ≤ n. Comparing the q-degrees of both sides, we deduce that t ≤ degq (f ). This proves that n − degq (f ) ≤ dH (σf , G). Furthermore, if f is monic, the equality t = degq (f ) holds if and only if H(x) = x, in which case, we obtain Y f (x) − v(x) = (x − g) g∈<gi1 ,...,git > The theorem is proved. By Theorem 2, we deduce the results in Corollaries 1 and 2 still hold under Hamming metric. 4 The error distance of a special class of words to certain Gabidulin codes We hope to obtain more deep holes of Gabidulin codes and so consider monic f (x) of k k+d−2 k+d−1 k+d +· · ·+(−1)d ad xq +· · · . +a2 xq degq (f ) = k +d, d ≥ 1, where f (x) = xq −a1 xq Q k k+d−1 k+d + · · · + (−1)d hd xq + · · · and − h1 xq In Theorem 1, if we write h∈H (x − h) = xq let β1 , β2 , . . . , βk+d ∈ Fqm be a basis of H, then dR (σf , G) = n − degq (f ) is equivalent to ai = hi = det(Rk+d−i (β1 , · · · , βk+d )) , for all 1 ≤ i ≤ d, det(Mk+d (β1 , · · · , βk+d )) where Rk+d−i (β1 , · · · , βk+d ) denotes the matrix Mk+d+1 (β1 , · · · , βk+d ) deleting the row k+d−i q k+d−i ). As a result, we have dR (σf , G) = n − (k + d) if and only if there (β1q , · · · , βk+d exist k + d linearly independent elements β1 , β2 , . . . , βk+d of < g1 , · · · , gn > such that ai = det(Rk+d−i (β1 , · · · , βk+d )) , for all 1 ≤ i ≤ d, det(Mk+d (β1 , · · · , βk+d )) where Rk+d−i (β1 , · · · , βk+d ) denotes as the above. When d = 1 i.e., degq (f ) = k + 1, then by Theorem 1, σf is not a deep hole of G if and only if dR (σf , G) = n − (k + 1). Thus, we have k+1 k − a1 xq + · · · . Then σf is not a deep hole of G if and Proposition 4. Let f (x) = xq only if there exist k + 1 linearly independent elements β1 , β2 , . . . , βk+1 of < g1 , · · · , gn > such that det(Rk (β1 , · · · , βk+1 )) , a1 = det(Mk+1 (β1 , · · · , βk+1 )) k where Rk (β1 , · · · , βk+1 ) denotes the matrix Mk+2 (β1 , · · · , βk+1 ) without the row (β1q , · · · , qk ). βk+1 7 Similar to the above discussion, we get the result for Hamming metric case by Theorem k k+d−2 k+d−1 k+d + · · · + (−1)d ad xq + · · · . Then dH (σf , G) = + a2 x q − a1 x q 2. Let f (x) = xq n − (k + d) if and only if there exist k + d distinct elements gi1 , gi2 , . . . , gik+d of {g1 , · · · , gn } such that det(Rk+d−i (gi1 , · · · , gik+d )) ai = , for all 1 ≤ i ≤ d, det(Mk+d (gi1 , · · · , gik+d )) where Rk+d−i (gi1 , · · · , gik+d ) denotes as the above. k+1 k − a1 xq + · · · . Then σf is not a deep hole of G with Proposition 5. Let f (x) = xq Hamming metric if and only if there exist k + 1 distinct elements gi1 , gi2 , . . . , gik+1 of {g1 , · · · , gn } such that det(Rk (gi1 , · · · , gik+1 )) a1 = , det(Mk+1 (gi1 , · · · , gik+d )) k where Rk (gi1 , · · · , gik+d ) denotes the matrix Mk+2 (gi1 , · · · , gik+d ) without the row (giq1 , · · · , k giqk+1 ). In the following we discuss the error distance of a special class of words to certain Gabidulin codes by using the above results. Before that, we give two lemmas. Lemma 3. [17] Let a be in a finite field Fq and p be the characteristic of Fq . Then the trinomial xp − x − a is irreducible in Fq [x] if and only if TrFq /Fp (a) 6= 0. For a finite field Fq , the integer-valued function v on Fq is defined by v(b) = −1 for b ∈ F∗q and v(0) = q − 1. Lemma 4. [17] For even q, let a ∈ Fq with trFq (a) = 1 and b ∈ Fq , then the number of solutions of the equation x21 + x1 x2 + ax22 = b is q − v(b). We first consider the finite field F2m . Let h(x1 , x2 ) = x21 + x1 x2 + x22 . For any b ∈ F2m , let the set S(h(x1 , x2 ) = b) = {(c1 , c2 ) ∈ F2m × F2m \|h(c1 , c2 ) = b, c1 6= c2 , ci 6= 0, i = 1, 2} and N (h(x1 , x2 ) = b) = \|S(h(x1 , x2 ) = b)\|. We consider two cases: Case 1: m is odd, which implies that Tr2m (1) = 1. If b = 0, the number of solutions of the equation h(x1 , x2 ) = b is 2m − v(b) = 1 by Lemma 4. Since (0, 0) is a solution, S(h(x1 , x2 ) = b) = ∅ and N (h(x1 , x2 ) = b) = 0. If b 6= 0, the number of solutions of the equation h(x1 , x2 ) = b is 2m + 1 by Lemma 4. Thus, N (h(x1 , x2 ) = b) = 2m + 1 − 2 since any element in F2m is a square. We also obtain the corresponding S(h(x1 , x2 ) = b). 8 Case 2: m is even, which implies that Tr2m (1) = 0. By Lemma 3, x2 + x + 1 is reducible over F2m and so it can be written as x2 + x + 1 = (x + α)(x + β) where α, β ∈ F2m , α 6= 1, β 6= 1 and α 6= β. Thus, x21 + x1 x2 + x22 = (x1 + αx2 )(x1 + βx2 ) = b and so the number of solutions of h(x1 , x2 ) = b is 2m + 2m − 1 if b = 0 or 2m − 1 if b 6= 0. If b = 0, then N (h(x1 , x2 ) = b) = 2m+1 − 2 and also we get S(h(x1 , x2 ) = b). If b 6= 0, N (h(x1 , x2 ) = b) = 2m − 1 − 2 since any element in F2m is a square. We also get S(h(x1 , x2 ) = b). From the above discussion, we get the following result. Proposition 6. Let G be the Gabidulin code over F2m with g = {g1 , . . . , gn }, dimension k = 1 and 3 ≤ n ≤ m. Let f (x) = x4 + bx2 + cx, where b, c ∈ F2m . Then σf is not a deep hole of G if and only if there are two elements β1 and β2 in < g1 , · · · , gn > such that (β1 , β2 ) ∈ S(h(x1 , x2 ) = b). In particular, if n = m, then σf is a deep hole of G if and only if b = 0 and m is odd. Proof. Note that two nonzero elements β1 and β2 are linearly independent over F2 if and only if β1 6= β2 . Thus, by Proposition 4, σf is not a deep hole of G if and only if there are two distinct nonzero elements β1 and β2 in < g1 , · · · , gn > such that b = β12 + β1 β2 + β22 , i.e., (β1 , β2 ) ∈ S(h(x1 , x2 ) = b). In particular, if n = m, then < g1 , · · · , gn >= F2m . By the above discussion, σf is a deep hole only when b = 0 and m is odd. For the other cases, N (h(x1 , x2 ) = b) is at least 1. Therefore, the desired result is obtained. Remark 1. Proposition 6 may not hold for the case of Hamming metric by Proposition 5 since it is possible that h(x1 , x2 ) = b has no solutions in {g1 , · · · , gn } when b 6= 0 although h(x1 , x2 ) = b always has solutions in F2m . Proposition 7. Let G be the Gabidulin code over Fqm where m is odd, 3 ≤ n ≤ m, any 2 linearly independent set g = {g1 , . . . , gn } and dimension k = 1. Let f (x) = xq + cx with c ∈ Fqm . Then σf is a deep hole of G. Proof. Suppose that σf is not a deep hole of G. By Proposition 4, there are two linearly 2 2 independent elements β1 and β2 in < g1 , · · · , gn > such that b = 0 = β1 β2 (β2q −1 − β1q −1 ). 2 Thus, (β2 β1−1 )q −1 = 1. Since m is odd, gcd(q 2 − 1, q m − 1) = q − 1. So we have (β2 β1−1 )q−1 = 1, which implies that ββ21 ∈ Fq i.e., β1 and β2 are linearly dependent over Fq . This contradicts with the assumption of β1 and β2 . Finally, we consider the deep holes of the Gabidulin codes with large dimension k = n − 2 and m = n. 9 Proposition 8. Let G be the Gabidulin code over Fqn with linearly independent set g = n−1 + f≤n−3 , where f≤n−3 is a q{g1 , . . . , gn } and dimension k = n − 2. Let f (x) = xq linearized polynomial over Fqn of degree less than or equals to n − 3. Then σf is a deep hole of G. Proof. Suppose that σf is not a deep hole of G. By Proposition 4, there are n − 1 linearly independent elements β1 , . . . , βn−1 in < g1 , · · · , gn > such that det(Rn−2 (β1 , · · · , βn−1 )) = 0. det(Mn−1 (β1 , · · · , βn−1 )) (2) For any matrix A = (aij ), denote by A(q) the matrix (aqij ). Then  β1q 2 β1q .. .    (q) Rn−2 (β1 , · · · , βn−1 ) =    qn−2  β1 n β1q β2q 2 β2q .. . n−2 β2q n β2q ... ... .. . q βn−1 q2 βn−1 .. . n−2 q . . . βn−1 qn . . . βn−1     .    n Note that βiq = βi . Thus  β1q 2 β1q .. .    (q) Rn−2 (β1 , · · · , βn−1 ) =    qn−2  β1 β1 β2q 2 β2q .. . n−2 β2q β2 ... ... .. . q βn−1 q2 βn−1 .. . n−2 q . . . βn−1 . . . βn−1     ,    (q) and det(Rn−2 (β1 , · · · , βn−1 )) = (−1)n−1 det(Mn−1 (β1 , · · · , βn−1 )). It is easy to see that det(A(q) ) = (det(A))q , for any matrix A over finite field with characteristic p. Thus (q) (det(Rn−2 (β1 , · · · , βn−1 )))q = det(Rn−2 (β1 , · · · , βn−1 )) = (−1)n−1 (det(Mn−1 (β1 , · · · , βn−1 )))q 6= 0, i.e., det(Rn−2 (β1 , · · · , βn−1 )) 6= 0, which contradicts to Eq. (2). Thus σf is a deep hole of G. The proposition holds. Remark 2. Indeed, according to the process of the above proof, the conclusion of Proposin−2 n−1 + f≤n−3 over Fqn , − axq tion 8 still holds for any q-linearized polynomial f (x) = xq n−1 1−q ∗ where a 6= (−1) b for all b ∈ Fqn . Propositions 7 and 8 still hold for Hamming metric after similar analysis. 10 5 Conclusions In this paper, we study the deep-holes of Gabidulin codes with both Hamming metric and rank metric. The general results for Hamming metric case (see Theorem 2 and Proposition 5) depend on the choice of the set {g1 , . . . , gn }, while the results for rank metric case (see Theorem 1 and Proposition 4) only depend on the subspace of Fqm spanned by g1 , . . . , gn . In particular, when n = m, the latter case does not depend on the choice of g1 , . . . , gn since < g1 , . . . , gn > equals to the whole space Fqm . Hence the problem about deep-holes of Gabidulin codes with Hamming metric seems more complicated than the rank metric case. On the other hand, for generalized Reed-Solomon codes, it has been proved that the problem of determining if a received word is a deep hole is NP-hard [11]. For Gabidulin codes, the problem seems more complicated although we give a necessary and sufficient condition for this problem. So we state it as a conjecture. Conjecture 1. Deciding deep holes of the Gabidulin code is NP-hard. References [1] D. Bartoli, M. Giulietti and I. Platoni, On the covering radius of MDS codes, IEEE Transactions on Information Theory, vol. 6, No. 2, pp. 801-811, 2015. [2] Q. Cheng and E. Murray, On deciding deep holes of Reed-Solomon codes, Lecture notes in Computer Science, vol. 4484, pp. 296-305, 2007. [3] Q. Cheng and D. Wan, On the list and bounded distance decodability of Reed-Solomon codes, SIAM Journal on Computing, vol. 37, no. 1, pp. 195-209, 2007. [4] G. Cohen, M. Karpovsky, H. Mattson and J. Schatz, Covering radius–survey and recent results, IEEE Transactions on Information Theory, vol. 31, No. 3, pp: 328-343, 1985. [5] G. Cohen, A. C. Lobstein and N. Sloane, Further results on the covering radius of codes, IEEE Transactions on Information Theory, vol. 32, no. 5, pp: 680-694, 1986. [6] P. Delsarte, Bilinear forms over a finite field with applications to coding theory, Journal of Combinatorial Theory, Series A, vol. 25, No.3, pp: 226-241, 1978. [7] E. M. Gabidulin, Theory of codes with maximum rank distance, Problemy Peredachi Informatsii, vol. 21, No. 1, pp: 3-16, 1985. [8] M. Gadouleau and Z. Yan, Packing and covering properties of rank metric codes, IEEE Transactions on Information Theory, vol. 54, no. 9, pp. 3873-3883, 2008. [9] M. Gadouleau and Z. Yan, Properties of codes with the rank metric, in Proc. IEEE Globecom 2006, San Francisco, CA, 2006. 11 [10] R. Graham and N. Sloane, On the covering radius of codes, IEEE Transactions on Information Theory, vol. 31, no. 3, pp: 385-401, 1985. [11] V. Guruswami and A. Vardy, Maximum-likelihood decoding of Reed-Solomon codes is NP-hard, In Proceeding of SODA, pp. 2249-2256, 2005. [12] T. Helleseth, T. Klove and J. Mykkeltveit, On the covering radius of binary codes, IEEE Transactions on Information Theory, vol. 24, no. 5, pp: 627-628, 1978. [13] A. Horlemann-Trautmann and M. Kuijper, Gabidulin decoding via minimal bases of linearized polynomial modules, Mathmatics, 2014. [14] M. Keti and D. Wan, Deep holes in Reed-Solomon codes based on Dickson polynomials, Finite Fields & Applications, Vol. 40, pp: 110-125, 2016. [15] R. Kötter and R. R. Kschischang, Coding for errors and erasures in random networking coding, IEEE Transactions on Information Theory, vol. 54, No. 8, pp. 3579-3591, 2008 . [16] Q. Liao, On Reed-Solomon codes, Chinese Annals of Mathematics, no. 1, pp. 89-98, 2011. [17] R. Lidl and H. Niederreiter, Finite fields, Cambridge University Press, Cambridge, London. [18] P. Loidreau, Properties of codes in rank metric, http://arxiv.org/abs/cs/ 0610057. [19] D. Silva, F. R. Kschischang and R. Kötter, A rank-metric approach to error control in random network coding, IEEE Transactions on Information Theory, vol. 54, No.9, pp: 3951-3967, 2008. [20] W. B. Vasantha, N. Suresh Babu, On the covering radius of rank-distance codes. Gaita Sandesh 13 (1999), no. 1, 4348. [21] D. Wan and Y. Li, On error distance of Reed-Solomon codes, Science in China, vol. 51, no. 11, pp: 1982-1988, 2008. [22] R. Wu and S. Hong, On deep holes of standard Reed-Solomon codes, Science China Mathematics, vol. 55, no. 12, pp. 2447-2455, 2012. [23] J. Zhang, F.-W. Fu and Q. Liao, New deep holes of generalized Reed-Solomon codes, Scientia Sinica, vol. 43, no. 7, pp. 727-740, 2013. [24] J. Zhang and D. Wan, On deep holes of projective Reed-Solomon codes, International Symposium on Information Theory, pp: 925-929, 2016. [25] J. Zhang and D. Wan, Explicit Deep Holes of Reed-Solomon Codes, arXiv:1711.02292. [26] J. Zhuang, Q. Cheng and J. Li, On determining deep holes of generalized ReedSolomon codes, IEEE Transactions on Information Theory, vol. 62, no. 1, pp. 199-207, 2016. 12	7cs.IT
ASYMPTOTIC BOUNDS ON THE GLOBALLY OPTIMAL POSITIONS OF ORTHOGONAL STIFFENERS FOR RECTANGULAR PLATES IN ELASTOSTATIC BENDING Nathan Perchikov∗ School of Mechanical Engineering, Faculty of Engineering, Tel Aviv University, Ramat Aviv, Tel Aviv 69978, Israel Abstract The present paper treats the problem of finding the asymptotic bounds for the globally optimal locations of orthogonal stiffeners minimizing the compliance of a rectangular plate in elastostatic bending. The essence of the paper is the utilization of a method of analysis of orthogonally stiffened rectangular plates first presented by Mazurkiewicz in 1962, and obtained herein in a closed form for several special cases in the approximation of stiffeners having zero torsional rigidity. Asymptotic expansions of the expressions for the deflection field of a stiffened plate are used to derive limit-case globally optimal stiffening layouts for highly flexible and highly rigid stiffeners. A central result obtained in this work is an analytical proof of the fact that an array of flexible enough orthogonal stiffeners of any number, stiffening a simply-supported rectangular plate subjected to any lateral loading, is best to be put in the form of exactly two orthogonal stiffeners, one in each direction. Keywords Elastic plate bending; Orthogonal stiffeners; Fredholm's 2nd kind integral equation; Asymptotic analysis; Globally optimal positions 1 INTRODUCTION This paper addresses the well-known problem of optimal positioning of straight, uniform, orthogonal beam-stiffeners of a given number or, integrally attached to a thin, rectangular, isotropic and homogeneous plate subjected to lateral, elastic, quazistatic loading, in a layout minimizing the total work done by the loads, in a formulation resembling an earlier version by Dems et al. (1989). Herein, as in a majority of previous works, it is assumed that the neutral lines of the stiffeners coincide with the neutral surface of the plate. Both the analysis and the optimal design problems have gained substantial attention in the literature. The early papers of Nowacki (1954a,1954b) and Mazurkiewicz (1962a,1962b) solved the analysis problem by assuming the stiffeners positioned symmetrically about the neutral surface of the plate and enforcing geometric compatibility between the plate and the stiffeners as well as the static equilibrium of both the stiffeners (the Euler-Bernoulli equation) and the plate (the Love-Kirchhoff equation). In the work of Mazurkiewicz (1962), the case of a simply-supported plate was treated and the influence of the stiffeners was viewed as effective distributed line loads expressed as double sine series. The procedure employed by Mazurkiewicz for bidirectional stiffening with beams having non-negligible torsional rigidity produced infinite sets of linear algebraic equations that had to be cut off and solved simultaneously. A proof of the convergence of such algebraic systems is given by Fletcher and Thorne (1955), within the context of the analysis of unstiffened plates having arbitrary boundary conditions. In the work of Dems et al. (1989), the equilibrium equations for the plate and the stiffeners are replaced by the use of the Green functions of the bending of simply-supported plate and beams, and the compatibility equation is formulated as Fredholm's integral equation of the 1st kind, which is solved numerically by assuming that the stiffeners are attached to the plate in a finite number of points and the effective point loads corresponding to geometric compatibility at those points are computed, again by the simultaneous solution of cut infinite linear algebraic equations. Various aspects of the analysis of quazistatic bending, free vibrations and buckling of orthogonally stiffened rectangular simply-supported plates using double sine series representations were examined (Konchkovskii 1984; Savin et al. 1964; Grigolyuk et al. 1980; Andrianov et al. 1985; Kalamkarov 1992). The (sometimes called inverse) optimization problem, of optimal positioning of stiffeners has gained not less an extensive coverage. In the work of Dems et al. (1989) the locations of the stiffeners are varied to decrease the work done by a given quazistatic lateral loading, where the optimization procedure utilizes an optimality criterion derived variationally. In their work, Grayhack and Mahar (1990) consider highly flexible stiffeners positioned in a layout increasing the minimum buckling load and obtain an analytic form of the equations determining the optimal locations. The objective of the present paper is the derivation of a closed-form expression for the elastic strain ∗ E-mail: perchico@gmail.com 1 energy in a simply supported thin rectangular plate stiffened with a single set of parallel stiffeners and subjected to a harmonic loading, an asymptotic expansion of the elastic strain energy with respect to a nondimensional parameter representing the ratio of the bending rigidity of the stiffeners and the plate, and analytical globally optimal minimization of the compliance with respect to the locations of the stiffeners. In addition, a closed form, globally optimal bound for the optimal layout of asymptotically flexible stiffeners is derived for the case of bidirectional stiffening. The essence of the results presented herein is the explicitness, simplicity and global optimality of the expressions for the optimal layouts of highly flexible and highly rigid stiffeners. In the present work, in contrast to the procedure followed by Dems et al. (1989), the geometric compatibility equation (between the plate and the stiffeners) is formulated as Fredholm's 2nd kind integral equation, which, unlike the 1st kind equation can be solved analytically for the case of a double-sine expressible Green's kernel function. This method is equivalent to the method used by Nowacki (1954) and Mazurkiewicz (1962) with the exception that here a closed form expression is obtained, mainly (but not solely) due to the restriction to unidirectional stiffening. A thorough review of solutions of contact problems in plate and shell bending by reformulation of the governing equations into 2nd kind Fredholm's equations can be found in the work of Grigolyuk et al. (1980). A different approach is described in the books of Andrianov et al. (1980) and Kalamkarov (1992), where the method of homogenization is applied. Finding an anisotropic plate structurally equivalent to a ribstiffened isotropic plate is widely used in development of finite elements, topological optimization and analysis of uniformly stiffened panels. Although obviously advantageous in the aforementioned cases, the method of homogenization is not necessarily more suitable for the purpose of finding the globally optimal locations of asymptotically rigid stiffeners. A discussion on certain mathematical aspects of the method is given by Andrianov et al. (1980). A review of applications of the method is carried out by Kalamkarov (1992). The outline of the present paper is as follows. In section 2, the deflection field of a stiffened plate is derived self-consistently, assuming continuous interfaces between the plate and the stiffeners, resulting in a closed form expression for a unidirectionally stiffened simply supported plate subjected to bisinusoidal loading, and in double-sine series for general loading cases. In section 3, the solution is generalized to account for bidirectional stiffening. In section 4, torsional rigidity is taken into account for the simpler case of unidirectional stiffening. In section 5, on the basis of the results of sections 2, 3 and 4, asymptotic, explicit and globally optimal bounds for the optimal locations of highly flexible and highly rigid uni- or bidirectional stiffeners with or without torsional rigidity are obtained analytically. 2 ANALYSIS OF UNIDIRECTIONALLY STIFFENED PLATES This section outlines the procedure of derivation of the linear elastic deflection field of a thin, rectangular, homogenous, isotropic, simply supported plate of a uniform thickness, subjected to a prescribed lateral loading and stiffened with straight uniform parallel open-section beams, all perpendicular to one of the sides of the plate, and having locations specified parametrically. Graphical representation of the problem follows in Fig. 1, where ξ and η denote the planar Cartesian coordinates, a, b and t the dimensions of the plate, p(ξ,η) the distribution of the lateral loading and x1, x2 the locations of the stiffeners. The bending rigidities of the plate and the stiffeners are denoted by D, E1I1 and E2I2, respectively. Fig. 1 Representation of the problem (the lower half of a plate stiffened symmetrically with respect to its neutral surface) 2 The Love-Kirchhoff model is used to describe the bending of the plate. The stiffeners are assumed to behave as Euler-Bernoulli beams, positioned symmetrically with respect to the neutral surface of the plate and having the same deflection as the neutral surface, along their longitudinal axes. The influence of the ith stiffener on the plate is represented by a distributed force per unit length, qi(xi,η), defined as positive when acting upwards. The stiffeners and the plate are assumed to have the same boundary conditions. In the case considered in the present section the beams are all simply supported. Each beam is assumed to be subjected to a distributed force per unit length qi(xi,η), defined as positive acting when downwards, as shown in Fig. 2. (a) (b) Fig. 2 (a) The distributed force that the plate applies on the ith stiffener and (b) The distributed force that the ith stiffener applies on the plate The solution procedure presented here is based on the infinitesimal deformation assumption, according to which the deflection of the plate is small relatively to its thickness, and the thickness, in turn, is small relatively to the (say, geometric) average of the lengths of the edges of the plate. According to this assumption, the bending of the plate is uncoupled with the in-plane stretch and can therefore be described by a single force balance equation for the static equilibrium deflection of the neutral surface (usually the middle surface) of the plate, and the appropriate boundary conditions. This paper treats the problem of small-strain elastic bending of a thin, simply supported plate. This means that the plate is treated as if it had a line of hinges on each of two of its adjacent edges and a line of moving supports, say, moving bearings, on each of the other two edges. Accordingly, the deflection field of the plate would follow from the Love-Kirchhoff assumptions. The derivation of the elastic infinitesimal deflection field will follow from linear superposition: N w(ξ ,η , x) = w p (ξ ,η ) + ∑ wqi (ξ ,η , x) = (2.1) i =1 =∫ b ∫ a 0 0 N a i =1 0 p (α , β ) w1 (ξ ,η ,α , β )dβdα +∑ ∫ qi ( xi , β , x) w1 (ξ ,η , xi , β )dβ where w(ξ,η,x) is the total deflection of the stiffened plate at a point which coordinates are denoted by ξ and η, namely, at (ξ,η), wp(ξ,η) is the deflection of the plate alone due to the external load alone at (ξ,η), w1(ξ,η,α,β) is the deflection of the unstiffened plate at (ξ,η) due to a unit lateral force acting at (α,β), p(ξ,η) is the function describing the distribution of the given external lateral load, x is the vector of the ξ coordinates of the stiffening beams and q(ξ,η,x) is the vector of the distributed forces per unit length representing the forces acting between the plate and the stiffeners (due to the external load) when the system is in equilibrium, each of which components corresponds to each of the N stiffeners. Equation (2.1) assumes that p(ξ,η) is everywhere small enough for the deflection of the plate to be small relatively to its thickness. According to Newton's third law, qi ( xi ,η , x ) = − qiB ( xi ,η , x ) (2.2) where the superscript B represents the fact that the distributed force is applied at the beams representing the stiffeners. It is now due to geometric compatibility, representing the fact that the stiffeners are considered integrally attached to the plate, that the deflection fields of the stiffeners are equal to the deflection field of the stiffened plate, along their interfaces, namely: w( xi ,η , x) = wiB ( xi ,η , x ) (2.3) Next, assuming that the stiffeners can be treated as Euler-Bernoulli beams, one can write: qiB ( xi ,η , x) = Ei I i ∂ 4 wiB ( xi ,η , x) ∂η 4 Substituting equations (2.2) and (2.3) into the above relation, results in: 3 (2.4) q i ( xi , η , x ) = − E i I i ∂ 4 w( xi ,η , x) ≡ − Ei I i wη( IV ) ( xi ,η , x) ∂η 4 (2.5) Next, substituting equation (2.5) into equation (2.1) produces the integro-differential equation for the deflection field of the stiffened plate (which can be turned into a standard Fredholm integral equation of the 2nd kind – and solved using the standard procedure – either by differentiating or by performing integration by parts four times consecutively): w(ξ ,η , x) = ∫ b a ∫ 0 0 N a i =1 0 p(α , β ) w1 (ξ ,η ,α , β )dβdα − ∑ Ei I i ∫ wβ( IV ) ( xi , β , x)w1 (ξ ,η , xi , β ) dβ (2.6) Here, use is made of Green’s function for the bending of a Love-Kirchhoff simply-supported plate as obtained by Navier (Szilard (1974)): ∞ 4 sin(nπα / b)sin( sπβ / a )sin(nπξ / b)sin( sπη / a ) ∑ 4 π abD n,s =1 ( n 2 / b2 + s 2 / a 2 ) 2 w1 (ξ ,η , α , β ) = (2.7) and uniform beams are assumed: E j I j (η ) = E j I j ∀ 1≤ j ≤ N (2.8) Next, taking a double-sine transform of the loading, 4 b a p(α , β ) sin(nπα / b) sin( rπβ / a )d β dα ab ∫0 ∫0 Pnr = (2.9) and introducing the auxiliary function: λ jn ≡ E j I j sin(nπ x j / b) (2.10) the deflection field of a unidirectionally stiffened simply supported plate for a general loading becomes: w(ξ ,η , x) = T 1 ∞ sin(mπξ / b) sin(rπη / a)  2r 4  ∞ λ m sin(nπx / b)  2r 4  P + ×  ∑ ∑ mr  a 4 bD  n =1 (n 2 / b 2 + r 2 / a 2 ) 2  a 4 bD π 4 D m, r =1 (m 2 / b 2 + r 2 / a 2 ) 2  ∞  λ µ sin( µπx / b) 2r sin(nπx T / b) 1 + 4 ∑  a bD µ =1 ( µ 2 / b 2 + r 2 / a 2 ) 2  T 4     −1 ∞ ∑ (ν ν =1 (2.11)    − Pnr    /b +r /a )    λ ν Pνr 2 2 2 2 2 The elastic strain energy stored in the deformed structure, consistent of the energy stored in the stiffened plate and the energy stored in the stiffeners, is equal to the work done by the external loads, and can be obtained accordingly as: U= 1 a b p (ξ ,η ) w(ξ ,η , x)dξdη 2 ∫0 ∫0 (2.12) Next, if one assumes a sinusoidal loading distribution, p(ξ ,η ) = Pgh sin( gπξ / b) sin(hπη / a) ⇒ Pkr = Pghδ gk δ hr (2.13) The deflection field of the stiffened plate becomes: w(ξ ,η , x) = + Pgh sin( gπξ / b) sin(hπη / a) + π 4 D (g 2 / b 2 + h 2 / a 2 ) 2 4h 8 Pgh π a b D 4 8 2 ∞ 3 ∑ m =1 λ Tm sin(mπξ / b) sin(hπη / a) (m / b + h / a ) 2 2 2 2 2 ∞ ∑ (n n =1 2 −1 sin(nπx / b) × / b2 + h2 / a2 )2 T  λg 2h 4 ∞ λ µ sin( µπx / b)  × sin(nπx T / b) 1 + 4 ∑ 2 2 2 2 2 2 2  a bD µ =1 ( µ / b + h / a )  ( g / b + h 2 / a 2 ) 2   4 T 2h Pgh ∞ λ sin(mπξ / b) sin(hπη / a) sin( gπx / b) − 4 4 2∑ m 2 2 2 2 2 2 π a bD m =1 (m / b + h / a ) (g / b 2 + h 2 / a 2 ) 2 4 (2.14) Hence, more compact a form of the deflection field of a stiffened plate can be obtained. First, using a well-known trigonometric identity: ∞ ∑ n =1 2 sin( nπx j / b) sin( nπxi / b) (n / b + r / a ) 2 2 2 2 2 ∞ =∑ n =1 cos[ nπ ( x j − xi ) / b] (n / b + r / a ) 2 2 2 2 2 ∞ −∑ cos[ nπ ( x j + xi ) / b] (n 2 / b 2 + r 2 / a 2 ) 2 n =1 (2.15) the double sine sums in (2.14) can be written as a difference between two single cosine sums. Then, summing the cosine series analytically by relating them to a function that produces a known sine series, the following result is obtained (this matter is further addressed in section 5.2): ∞ d ∞ cos(nπx / b) ( nπ / b) sin( nπx / b) = − = ∑ ∑ 2 2 2 2 2 dx n =1 ( n 2 / b 2 + r 2 / a 2 ) 2 n =1 ( n / b + r / a ) = (2.16) ab 2 sinh(πrx / a ) − abx sinh(πrb / a ) cosh(πr[b − x ] / a ) 4π −3 r sinh 2 (πrb / a ) By integrating (2.16) and substituting the result into (2.14), an expression with no infinite sums is obtained. Defining, component-wise, the following quantities (i and j being the indices related to the numbers of the stiffeners): B jir = cosh(πr[b − x j − x i ] / a) r 2 E j I j  cosh(πr[ x j − x i ] / a) − cosh(πr[ x j + x i ] / a ) +a b 3 2 2 πr sinh(πrb / a) 2π a D  sinh (πrb / a) (2.17) cosh(πr[b − xi − x j ] / a) x j − x i sinh(πr[b − x j − x i ] / a) − ( x j + x i ) sinh(πr[b − x i − x j ] / a)  −a +  πr sinh(πrb / a) sinh(πrb / a)  cosh(πh[b − x j − ξ ] / a) Pgh h 2 sin( hπη / a) E j I j  cosh(πh[ x j − ξ ] / a ) − cosh(πh[ x j + ξ ] / a ) +a b 3 2 2 πh sinh(πhb / a) 2π a D sinh (πhb / a )  (2.18) cosh(πh[b − ξ − x j ] / a) x j − ξ sinh(πh[b − x j − ξ ] / a) − ( x j + ξ 0 ) sinh(πh[b − ξ − x j ] / a)  −a +  πh sinh(πhb / a) sinh(πhb / a)  s jh = S jih =  cosh(πh[ x j − x i ] / a ) − cosh(πh[ x j + x i ] / a ) cosh(πh[b − x j − x i ] / a ) h2 +a b 3 2 2 πh sinh(πrb / a ) 2π a D  sinh (πhb / a )  (2.19) cosh(πh[b − x i − x j ] / a ) x j − x i sinh(πh[b − x j − x i ] / a ) − ( x j + x i ) sinh(πh[b − x i − x j ] / a )  −a +  πh sinh(πrb / a ) sinh(πhb / a )  the deflection field for a bisinusoidal loading in a closed form – although not totally explicitly, as a solution of a finite linear system of equations is required – becomes: Pgh sin( gπξ / b) sin( hπη / a ) + s Th [S h (1 + B h ) λ g − sin( gπx / b)] −1 wgh (ξ ,η , x) = π 4 D( g 2 / b 2 + h 2 / a 2 ) 2 (2.20) Having the deflection field corresponding to a bisinusoidal loading, the deflection field for a general loading (expressible as a double sine series) can be derived by superposition: w(ξ ,η , x) = ∞ ∑ g , h =1 Pgh sin( gπξ / b) sin(hπη / a) + s Th [S h (1 + B h ) λ g − sin( gπx / b)] −1 π 4 D( g 2 / b 2 + h 2 / a 2 ) 2 (2.21) Numerical verification of the presented method is omitted here, as an equivalent method has been presented before by Nowacki (1954) and Mazurkiewicz (1962). The analytical summations performed here can only improve convergence and help avoid the Gibbs phenomena. However, in the present paper, the above operations are needed in order to obtain closed-form asymptotic expansions rather than in order to produce a better numerical analysis module. Therefore, no comparison of deflections and stresses of stiffened plates for various loading cases (calculated once with the employed method and once with commercial software tools), is given here. 5 3 GENERALIZATION OF THE ANALYSIS FOR BIDIRECTIONAL STIFFENING The treatment of bidirectional stiffening is not much different from that of the unidirectional one. The deflection field can be obtained using linear superposition: N b i =1 0 w(ξ ,η , x) = w p (ξ ,η ) + ∑ wqi (ξ ,η , x) = ∫ Nη a + ∑ ∫ qi ( xi , β , x) w1 (ξ ,η , xi , β )dβ + 0 i =1 ∫ a p(α , β ) w1 (ξ ,η , α , β )dβdα + 0 Nη + N ξ (3.1) ∑∫ b qi (α , xi , x) w1 (ξ ,η , α , xi )dα 0 i = Nη +1 where qi are the distributed forces per unit length exerted on the plate by the stiffeners, N is the total number of the stiffeners, Nη is the number of η-axis aligned stiffeners, and Nξ is the number of ξ -axis aligned stiffeners. Denoting the interaction forces applied on the ξ-axis aligned stiffeners by the η-axis aligned stiffeners by fij, one can write the following relations: Ei I i Nη ∂ 4 wi 's Beam (ξ ) ∂ 4 w(ξ , xi ) = E I = q ( ξ ) = − q ( ξ , x ) + f ij δ (ξ − x j ) ∑ i i on i ' s i i ∂ξ 4 ∂ξ 4 Beam j =1 (3.2) where the first equality sign is due to the assumption that the neutral axes of the stiffeners coincide with the neutral surface of the plate, the second is due to the Euler-Bernoulli beam theory for uniform beams and the third is obtained by summing up all the forces exerted on the i’s, ξ-axis aligned stiffener. The first term after the third equality sign is the distributed force per unit length exerted by the plate on the i’s, ξ-axis aligned stiffener, as arises from Newton’s third law. The function δ(x) denotes Dirac’s Delta function. For the η-axis aligned stiffeners one can write: Ei I i Nη + Nξ ∂ 4 w( xi ,η ) = − q ( x , η ) + ∑ (− f ij )δ (η − x j ) i i ∂η 4 j = Nη +1 (3.3) using the same arguments as before and having (– fij) instead of fij this time, due to Newton’s third law. Next, deriving qi from (3.2) and (3.3) and substituting them into (3.1), one has: w(ξ ,η , x) = ∫ b 0 Nη − ∑∫ i =1 + a 0 Nη + N ξ Nξ 0 Nξ p (α , β ) w1 (ξ ,η , α , β )dβdα − ∑ ∫ Ei I i wβ( IV ) ( xi , β , x)w1 (ξ ,η , xi , β )dβ ∑ f δ (β − x ij j )w1 (ξ ,η , xi , β )dβ − i = Nη +1 Nη b 0 ij i =1 Nη + N ξ i =1 + δ (α − x j )w1 (ξ ,η , α , xi )dα = ∫ 0 ∑ ∑f Nη Nη + N ξ ij b 0 a i = Nη +1 i =1 0 ∑∫ i = Nη +1 − ∑ ∫ Ei I i wβ( IV ) ( xi , β , x)w1 (ξ ,η , xi , β ) dβ − Nη + N ξ Nη a i =1 Nη + N ξ ∑ ∫ ∑f i = Nη +1 ∫ a w1 (ξ ,η , x j , xi ) − ∑ ∑f ij ∫ b 0 a 0 Ei I i wα( IV ) (α , xi , x)w1 (ξ ,η , α , xi ) dα + p (α , β ) w1 (ξ ,η , α , β )dβdα Nη + N ξ ∑∫ i = Nη +1 b 0 (3.4) Ei I i wα( IV ) (α , xi , x)w1 (ξ ,η , α , xi )dα w1 (ξ ,η , xi , x j ) i =1 i = Nη +1 Owing to the fact that the influence function for the bending of a simply-supported plate, w1, is symmetric (insensitive to an interchange of the arguments), as one can learn from (2.7), the last two terms in (3.4) cancel each other. Therefore, in deriving the compatibility equation for the bending of an orthogonally stiffened simply-supported plate, one can disregard the interaction forces between the stiffeners, as they cancel out in the final governing equation. Consequently, the governing equation becomes: w(ξ ,η , x) = ∫ b 0 ∫ a 0 Nξ a p (α , β ) w1 (ξ ,η , α , β )dβdα − ∑ ∫ E i I i wβ( IV ) ( xi , β , x)w1 (ξ ,η , xi , β )dβ i =1 − Nη + N ξ ∑∫ i = Nη +1 b 0 0 Ei I i wα( IV) (α , xi , x)w1 (ξ ,η , α , xi ) dα 6 (3.5) Hence, using expressions (2.7)-(2.10) and the orthogonality of the sine functions, and following the strategy for the solution of integro-differential equations mentioned in section 2, one obtains two sets of compatibility equations for the distributed forces between the stiffeners and the plate, one for each stiffening direction. Making use of (2.17) and introducing the following quantities: b a 0 0 z in ≡ Ei I i ∫ wα( IV) (α , xi , x) sin(nπα / b) dα ; vis ≡ Ei I i ∫ wβ( IV ) ( xi , β , x) sin( sπβ / a)dβ u jr ≡ λ jn Pnr µ js Pms r4 ∞ bm 4 ∞ ; y jm ≡ 4 ∑ 2 2 ; µ js ≡ E j I j sin( sπx j / a) ∑ 3 2 2 2 2 2 2a D n =1 (n / b + r / a ) 2a D s =1 (m / b + s 2 / a 2 ) 2 C jirn ≡ 2r 4 λ jn sin(rπxi / a) a 4 bD(n 2 / b 2 + r 2 / a 2 ) ; D jims ≡ 2 2m 4 µ js sin(mπxi / b) a 5 D(m 2 / b 2 + s 2 / a 2 ) ∞ ; F jim ≡ ∑ 2 s =1 (3.6) 2m 4 µ js sin( sπxi / a) a 5 D( m 2 / b 2 + s 2 / a 2 ) 2 and (analytically) decoupling the two sets of compatibility equations mentioned above, one gets a single system defined by the quantities: ∞   g n ≡ (1 + Fn ) −1 y n − ∑ D ns (1 + B s ) −1 u s  s =1   (3.7) ∞ Μ nm ≡ ∑ (1 + Fn ) −1 D ns (1 + B s ) −1C sm s =1 which are, respectively, the vector and matrix infinite sets representing the final form of the compatibility equation for the bidirectionally stiffened simply supported thin rectangular plate: ∞ z n = g n + ∑ M nm z m (3.8) m =1 Equation (3.8) is an infinite inhomogeneous linear system of equations in which every equation is by itself a finite system of equations where the number of equations is equal to the number of the stiffeners in the direction parallel to the ξ axis. Obviously the relation M nm = M nδ mn does not hold for any Mn and therefore the infinite system cannot be solved exactly and one has to truncate it in order to solve it. Thus, the system that would have to be solved would be: mf z n = g n + ∑ M nm z m ; 1 ≤ n ≤ m f << ∞ (3.9) m =1 In order to solve the system, it is necessary to write the finite systems of equations one beneath the other, resulting in a total number of mTot = N y m f unknowns and equations. That is if one has 17 stiffeners in the direction parallel to the ξ axis and they taking 100 harmonics in the Fourier series representation of the Green influence function of the bending of the plate, then one would write the first 17 equations for m = 1, and then the 17 equations for m = 2 and so forth up to m = 100, so that eventually they would have 1700 equations or rows in the representing matrix. Similarly, in the vectors one would have 17 components for m = 1 and then 17 components for m = 2 and so forth, with resultant vectors of 1700 components, corresponding to 1700 columns in the representative matrix. Naturally, one could have written the 100 equations corresponding to the 1st stiffener first and have 17 bunches of 100 and not 100 bunches of 17. However, it is better to keep the equations corresponding to the different stiffeners in bunches so that the truncation error will affect them as similarly as possible, so that no stiffener is represented worse than the other. Of course, there still remains the order within each bunch, but this is a smaller effect, there has to be some order anyway. Now, if the solution of the system of equations is achieved by Gauss elimination, then the last equations will have the smallest round-off error as the first to be solved, so that it would be beneficiary to put the first, more important harmonics in the end, that is, to reverse the order described above. Anyhow, defining: 7  M m f ,m f  gmf   z mf      M  z m f −1   g m f −1   m f −1,m f g  z  M m f − 2 , m f z ≡  m f −2  ; g ≡  m f −2  ; M ≡  .  .   .    .   .   .       z   g   M 1,m f  1   1  M m f ,m f −1 M m f −1,m f −1 M m f ,m f − 2 M m f −1,m f − 2 M m f − 2 ,m f −1 . . . . . M 1,m f −1 . . . M m f ,1  . . M m f −1,1  . . .   . . .  . . .   . . M 1,1  (3.10) every component in the vectors being a column vector itself and every component in the matrix being a matrix itself, results in the equation: (3.11) z = g + M z ⇒ (1 − M ) z = g m ×m where the unit matrix is 1 ∈ R f f . The solution of equation (3.8) would produce a series of the approximate values of a finite number of the necessary integrals for the evaluation of the deflection field. If the matrix (1 − M ) is regular and the system is convergent, then the solution can be written as: z = (1 − M ) −1 g (3.12) Obviously, this is only formal writing. The solution of a system of equations is not achieved by taking the inverse of a large matrix. The system can be solved with the Matlab solver based on LU factorization and Gauss elimination. Next, the question of convergence can be addressed in a two-fold way. First, a proof by Fletcher and Thorne [2] guarantees that a system equivalent to (3.8)-(3.11) is convergent and may be solved to any desired degree of accuracy (unlike, say, an asymptotic system, which can only be solved to a certain, limited degree of accuracy). Second, the system in (3.11) was solved by the author for realistic problem parameters, showing monotonic convergence upon increase in the number of considered harmonics and correlating well with a FEM solution. Now, whereas the solution of a unidirectionally stiffened plate presented in section 2 can be considered analytical and exact, the solution of a bidirectionally stiffened plate may be considered exact and semianalytical. ‘Exact’ due to the convergence of the above series, since exactness of a solution is one’s ability to have it up to any desired degree, which is possible in the present case and ‘semi-analytic’ due to the truncation of the infinite series. The procedure one has to follow in order to obtain a solution after (3.12) is not discussed thoroughly here. Shortly, one has to solve (3.11)-(3.12) and substitute the solution into (3.5) using definitions (3.10) and (3.6) and (2.7), thus obtaining the third row in (3.5). For the second row in (3.5) one has to substitute the solution of (3.11) along with the definitions in (3.6) into the following relation, obtained from the analytical decoupling of (3.5): mf   v r = (1 + B r ) −1  u r − ∑ C rn z n  n =1   4 (3.13) UNIDIRECTIONAL STIFFENERS WITH NON-NEGLIGIBLE TORSIONAL RIGIDITY One can account for the torsional rigidities of the stiffeners (which are relevant especially for the case of stiffeners having closed cross-sections). The torques that the stiffeners apply on the plate can be related to the twisting angles of the plate along the plate-stiffeners interfaces, as follows: for each stiffener, one can employ the constitutional relation of a linear elastic shaft, denoting the torque applied on the plate by T, the shear modulus and the geometric torsion rigidity parameter by G and C respectively, the coordinate changing along the stiffener being η and the perpendicular being ξ, assuming small twisting angles. The superscript b implies that a quantity is related to the stiffener. Ti ( xi ,η , x) = − Ti b ( xi ,η , x) = − C i Gi Newton 3rd Law Shaft Law ∂θ ib ∂θ ′′ ( xi ,η , x) = − C i Gi i = − C i Gi wξη ∂η wb = w ∂η θ i <<1 Superimposing, the deflection field of the (for simplicity, unidirectionally) stiffened plate becomes: 8 (4.1) N N b i =1 i =1 0 0 w(ξ ,η , x) = w p (ξ ,η ) + ∑ wqi (ξ ,η , x) + ∑ wTi (ξ ,η , x) = ∫ N a N a i =1 0 i =1 0 ∫ a p(α , β ) w1 (ξ ,η ,α , β ) dβdα + (4.2) + ∑ ∫ qi ( xi , β , x) w1 (ξ ,η , xi , β )dβ + ∑ ∫ Ti ( xi , β , x) wT =1 (ξ ,η , xi , β ) dβ Substituting (4.1) and (2.5) into (4.2) one gets the following integro-differential equation for the deflection field, where only the deflection due to a unit couple, wT =1 , is yet to be determined: w(ξ ,η , x) = ∫ b 0 ∫ a 0 N a i =1 0 p(α , β ) w1 (ξ ,η , α , β )dβdα − ∑ Ei I i ∫ wβ( IV) ( xi , β , x)w1 (ξ ,η , xi , β )dβ N a i =1 0 (4.3) ′′ ( xi , β , x)wT =1 (ξ ,η , xi , β )dβ − ∑ Gi Ci ∫ wξβ By Betti’s theorem, the deflection of an elastic structure due to a unit couple at a point is equal to the rotation angle (in the direction of the couple – perpendicular to the stiffeners, in the present case) – or the slope, for small rotation angles – due to a unit force at that point, namely: ∂w (ξ ,η , α , β ) (4.4) wT =1 (ξ ,η , xi , β ) = 1 ∂α α =x i Substituting (4.4) into (2.7) this yields: wT =1 (ξ ,η , x i , β ) = ∞ n cos( nπxi / b ) sin( sπβ / a ) sin( nπξ / b ) sin( sπη / a ) 4 ∑ 2 (n 2 / b 2 + s 2 / a 2 ) 2 π ab D n , s =1 3 (4.5) Substituting (2.7) and (4.5) into (4.3) produces a fully defined integro-differential equation for the deflection field. Naturally, similar steps can be taken if a second, perpendicular, stiffening array is present. Treating the obtained integro-differential equation in a manner similar to the one described in the previous sections, and introducing the following quantities (referring to definition (2.10) again): τ jn ≡ G j C j cos( nπx j / b) , Tr ≡ 2πr 4 ∞ nλ n cos( nπx T / b) ∑ a 4 b 2 D n =1 ( n 2 / b 2 + r 2 / a 2 ) 2 (4.6) one obtains a first set of compatibility equations: v r = u r − B r v r − Tr t r (4.7) where use is made of the definitions of Br from (2.17) and ur and vr from (3.6) and: a ′′ ( xi , β , x) sin(sπβ / a)dβ t is ≡ Gi Ci ∫ wξβ (4.8) 0 The equation in (4.7) is obtained by calculating the distributed forces-induced deflections at the platestiffeners interfaces from the left-hand side of (4.3). Calculating the distributed torques-induced deflections at the plate-stiffeners interfaces from (4.3) results in a second set of compatibility equations: ∞ ∞ s =1 s =1 t r = υ r − ∑ Λ rs v s − ∑ Ω rs t s (4.9) where the following definitions have been made: υr ≡ 1 ∞ rs − (−1) r + s s 2 ∑ r2 − s2 π 3 bD s =1 Λ rs ≡ Ω rs ≡ ∞ ∑ (n n =1 r+s 2 r+s 2 4 rs − (−1) s 3 2 π ab D r2 − s2 4 rs − ( −1) s 2 3 π ab D r2 − s2 ∞ nτ n Pns / b2 + s 2 / a2 )2 nτ n sin(nπx T / b) 2 / b2 + s 2 / a2 )2 ∑ (n n =1 n 2 τ n cos(nπx T / b) 2 / b2 + s 2 / a2 )2 n =1 ∞ ∑ (n One can decouple the two systems (4.7) and (4.9) analytically. 9 2 (4.10) Defining the following auxiliary quantities: ∞ Ψ rs ≡ Ω rs − Λ rs (1 + B s ) −1 Ts + δ rs 1 , φ r ≡ υ r − ∑ Λ rs (1 + B s ) −1 u s (4.11) s =1 where δrs is Kronecker’s Delta, the following infinite set of equations is obtained for tr: ∞ ∑Ψ t = φr rs s (4.12) s =1 Now, naturally, in order to be solved, this system has to be truncated. If the system is convergent, and we assume that it is, as we had assumed for the other systems in this paper (relying on a proof of convergence by Fletcher and Thorne (1955)), then an approximate solution with any desired degree of accuracy, can be obtained from: sf ∑Ψ t = φ r ; 1 ≤ r ≤ s f << ∞ rs s (4.13) s =1 Next, applying the argumentation from section 3 and denoting by t the truncated vector (t1T , t T2 ,..., t Tsf ) T and doing the same for the other quantities in (4.13), one can write the last form of the compatibility equation for a rectangular, simply supported plate stiffened with a single set of uniform stiffeners, parallel to one of the sides of the plate, taking the torsional rigidity of the stiffeners into account, as follows: Ψt = φ (4.14) This system would have to be solved numerically, after defining the auxiliary quantities introduced above. Hence, (4.15) θ r ≡ (1 + B r ) −1 u r ; Θ r ≡ (1 + B r ) −1 Tr and denoting by a corresponding symbol but with an upper bar the truncated vectors ( v1T , v T2 ,..., v Ts f ) T and (θ1T , θ T2 ,..., θ Ts f ) T and constructing the truncated matrix Θ in a manner similar to the one employed in equation (3.10) in section 3, we obtain the second set of unknowns required for the solution: v = θ − Θt (4.16) And the solution becomes: f 4 sin(nπξ / b) sin( sπη / a)  ab nπ  Pns − sin(nπx T / b) v s − cos(nπx T / b)t s  ∑ 4 b π abD n ,s =1 ( n 2 / b 2 + s 2 / a 2 ) 2  4  s w(ξ ,η , x) = (4.17) Last, in order to be substituted into (4.17), the solutions of (4.14) and (4.16) should be separated into the appropriate sub-vectors again, in a way similar to the way they were constructed. 5 ASYMPTOTIC ANALYSIS: GLOBALLY OPTIMAL LAYOUTS OF HIGHLY FLEXIBLE AND HIGHLY RIGID STIFFENERS This section is concerned with the determination of the optimal layout of stiffeners in the limit case where the stiffeners are much less or much more rigid than the ground structure. The examination of this limit case is important for the estimation of the correct asymptotic behavior of the optimal locations of stiffeners with respect to their rigidities and the determination of the bounds for the optimal positions of realistic stiffeners. The derivation assumes a thin, isotropic, homogenous, rectangular plate in elastostatic bending, subjected to a lateral pressure and stiffened with two sets of straight, parallel uniform stiffeners. 10 5.1 Highly flexible bidirectional stiffeners with no torsional rigidity – 1st order asymptotics The linear elastostatic strain energy for the case can be derived, using the results of sections 2 and 3, as follows. Introducing the small parameter ε ≡ EI /(aD) and deriving the auxiliary quantities required for the calculation of the strain energy for the asymptotic case, one gets: ∞ ∞   g n = (1 + Fn ) −1 y n − ∑ D ns (1 + B s ) −1 u s  → y n − ∑ D ns u s → y n  ε <<1 ( 3.7 )  ε <<1O s =1 s =1 O (ε ) O (ε ) (ε ) O (ε ) O (ε )  O (ε ) (5.1.1) O (ε 2 ) Μ nm = ∞ ( 3.7 ) ∑ (1 + F ) −1 n s =1 O (ε ) ∞ D ns (1 + B s ) −1C sm → ∑ D ns C sm = O(ε 2 ) ε <<1 s =1 O (ε ) O (ε ) O (ε ) −1 ⇒ z = (1 − M g → y ⇒ zn → yn ) g ε→ ( 3.9 ) <<1 ε <<1 ε <<1 (5.1.2) O (ε 2 ) mf mf mf v r = (1 + B r ) −1 (u r − ∑ C rn z n ) → u r − ∑ C rn z n → u r − ∑ C rn y n → u r ⇒ v r → u r ε <<1 ( 3.13) ε <<1 ε <<1 ε <<1 n =1 n =1 n =1 O (ε ) O (ε ) (5.1.3) O (ε ) O (ε ) Recalling the expression for the deflection (which is not given explicitly in section 3 but can easily be derived): w(ξ ,η , x) = mf Nη Nη + N ξ  4 sin( nπξ / b )in ( sπη / a )  abPns − sin( n π x / b ) v − sin( sπxi / a ) z in   ∑ ∑ ∑ i is 4 2 2 2 2 2 π abD n, s =1 (n / b + s / a )  4 i =1 i = Nη +1  (5.1.4) integrating it multiplied by half the loading, thus obtaining an expression for the work or the strain energy, U ( x, N ξ , N η ) = 1 2π 4 mf ∑ (n D n , s =1 Pns 2 2 / b + s 2 / a 2 )2 Nη + N ξ  abPns Nη  − ∑ sin( nπxi / b)vis − ∑ sin( sπxi / a ) z in   i =1 i = Nη +1  4  (5.1.5) and substituting (5.1.3) into (5.1.5), one gets an explicit expression for the strain energy of the stiffened plate: m Nη + N ξ  abPns Nη  Pns 1 U (x, N ξ , Nη ) → − n x b u − s x a y sin( / ) sin( / ) π π   ∑ ∑ ∑ i is i in 2 2 2 2 2 min{ E I }<< aD 2π 4 D n , s =1 ( n / b + s / a )  i =1 i = Nη +1   4 f j j j mf m f abPns2 Pns 1 = − ∑ ∑ 4 2 2 2 2 2 4 2 2 2π D n ,s =1 4( n / b + s / a ) 2π D n ,s =1(n / b + s 2 / a 2 ) 2 1 − 1 2π 4 1 mf ∑ (n D n , s =1 1 − 4 2π D 2a 2 2 Pns / b + s 2 / a 2 )2 2 Nη + N ξ ∑ sin( sπx i / a ) y in = i = Nη +1  m f P sin( nπxi / b) ε i ∑ s  ∑ 2ns 2 ∑ 2 2 2 i =1 s =1  n =1 (n / b + s / a ) Nη mf 4 ( 3..3), ( 2.1.10 ) 2 1 2π 4 Nη ∑ sin(nπx mf ∑ 4( n D  1 b  − 4  2π D 2a 3  i / b)u is (5.1.6) i =1 n , s =1 2 ns 2 abP / b2 + s 2 / a 2 )2  m f Pns sin( sπx j / a) ε j ∑ n  ∑ 2 2 ∑ 2 2 2 j = Nη +1 n =1  s =1 (n / b + s / a ) Nη + N ξ mf 4     2 This asymptotic expansion obtained for the energy, U, consists of a constant term, U0, from which two other terms are subtracted. Naturally then the energy, U, attains a minimum if and only if the terms that are subtracted from the constant, U0, attain maxima with respect to x. As the domain is unconstrained, obviously, the subtracted sums of squares attain their maxima when each of the squares in the sums attains its maximum. As the first sum in the rightmost part of (5.1.6) is symmetric with respect to i and the second sum is symmetric with respect to j, clearly: (5.1.7) xi = xη ∀ 1 ≤ i ≤ Nη ; x j = xξ ∀ Nη + 1 ≤ j ≤ Nη + N ξ is a globally optimal solution. In other words, the energy of a bidirectionally stiffened plate attains a global minimum with respect to the locations of two orthogonal sets of straight, parallel, uniform, infinitesimally rigid stiffeners, for the case of a general lateral loading distribution, if all of the stiffeners in each of the two sets are located at the same place. Knowing the optimal layout one can calculate the energy at the global optimum, substituting (5.1.7) into (5.1.6): 11 U min ( N ξ , Nη ) → min ε i << 1 i mf 1 2π 4 ∑ 4(n D n , s =1 2 abPns2 / b2 + s 2 / a2 )2  Pns sin(nπxξ / b) 1 1 − s 4  ∑ 2 2 4 2 ∑ 2π D 2a s =1  n =1 (n / b + s 2 / a 2 ) 2 mf mf     (5.1.8) 2 N η  m f Pns sin( sπxη / a ) 1 b 4 ε − n ∑ ∑ ∑ i 2π 4 D 2a 3 n =1  s =1 (n 2 / b 2 + s 2 / a 2 ) 2 i =1 mf     2 N +N η ξ ∑ε j j = Nη +1 The same energy corresponds to a plate stiffened with one "horizontal" stiffener and one "vertical" stiffener having, respectively, the following bending rigidities: Nη Nη + N ξ i =1 j = Nη +1 Eη Iη ≡ aD ∑ ε i ; Eξ I ξ ≡ aD ∑ε (5.1.9) j Therefore if one has a certain amount of material in the form of two orthogonal sets of parallel highly flexible stiffeners that can be integrally attached to a given plate to minimize the work done by external loads, a globally optimal way to use this material is to split it between exactly two orthogonal beams. For the special case of a symmetric loading, Pns=Psn, and a square plate, a = b, the result given in (5.1.8) shows that for flexible enough stiffeners, stiffening with two orthogonal beams is energetically equivalent to stiffening with a single beam, having twice the bending rigidity. This analytically obtained result coincides with numerical observation concerning the same matter, given in the work of Dems et al. (1989). The power of this result is that a globally optimal stiffening layout is determined analytically for bidirectional stiffening for an arbitrary number of stiffeners having arbitrary cross-sectional parameters and for a general loading case. The limitation of this result is that it is only relevant for highly flexible stiffeners and that it does not say for what actual bending rigidities it is true, for what relative bending rigidities does it cease to be true, and what happens to the optimal positions of the stiffeners for stiffeners having higher bending rigidities. The reason for this shortcoming is the fact that first order asymptotics with respect to the stiffeners do not capture the ‘split’ of the stiffeners as an optimal stiffening topology. The relative stiffness for which the first order approximation is valid is not high enough for it to be optimal for the stiffeners to be positioned separately. The following section shows that, at least for the special case of a bisinusoidal loading and unidirectional stiffening, with two beams positioned symmetrically with respect to the mid-span of the plate, even for a second order asymptotic expansion of the energy of the stiffened plate with respect to the relative stiffness of the beams, the degenerated layout given in (5.1.7) is still globally optimal. A comparison between the first and the second order expansions of the energy enables one to determine for what actual rigidities it is still optimal to place the stiffeners together. 5.2 Highly flexible unidirectional stiffeners with no torsional rigidity – 2nd order asymptotics Repeating the procedure shown in the preceding subsection and retaining terms up to the order of O(ε2), ∞ ∞   g n = (1 + Fn ) −1 y n − ∑ D ns (1 + B s ) −1 u s  → y n − Fn y n − ∑ D ns u s  s =1 ( 3.7 )  ε <<1O s =1 O (ε ) O (ε ) (ε ) O (ε ) O (ε ) O (ε ) O ( ε )  Μ nm = ( 3 .7 ) ∞ ∑ (1 + F ) n s =1 O (ε ) −1 D ns (1 + B s ) −1C sm → ∑ D ns C sm ε <<1 s =1 O (ε ) O (ε ) O (ε ) ∞ −1 ⇒ z = (1 − M (1 + M g +M y n − Fn y n − ∑ D ns u s ) g ε→ ) g ε→ g → g ⇒ z n ε→ (3.9 ) <<1 <<1 <<1 2 2 s =1 O (ε ) (5.2.1) ∞ O (ε ) O (ε ) O (ε ) O (ε 3 ) (5.2.2) O (ε ) mf mf ∞   v r = (1 + B r ) −1  u r − ∑ C rn z n  → u r − B r u r − ∑ C rn z n → u r − B r u r − ∑ C rn y n ( 3.13) ε <<1 ε <<1 n =1 n =1 n =1   O (ε ) (5.2.3) and substituting them into the expression for the work done by the external loads given in (5.1.5), one gets: 12 1 ⇒ U ( x, N η , N ξ ) → ε <<1 2π 4 mf ∑ (n D n , s =1 Nη + Nξ  abPns Nη  Pns − − sin( n x / b ) u sin( sπxi / a) yin  + π  ∑ ∑ i is 2 2 2 2 2 / b + s / a )  4 i =1 i = Nη +1  + Nη + Nξ  Nη  Pns sin( n x / b ) B u + sin( sπxi / a) Fijn y jn  + π  ∑ ∑ i ijs js 4 2 2 2 2 2 ∑ 2π D n ,s =1( n / b + s / a ) i , j =1 i , j = Nη +1  + Pns ∑ 4 2 2 2π D n ,s =1( n / b + s 2 / a 2 ) 2 1 1 mf mf (5.2.4) m f Nη + N ξ Nη  m f Nη Nη + Nξ  sin( n π x / b ) C y + sin( sπxi / a ) Dijnk u jk  ∑∑ ∑ ∑ ∑ ∑ i ijsr jr k =1 i = Nη +1 j =1  r =1 i =1 j = Nη +1  Next, recalling (3.6) and rearranging, results in the following 2nd order asymptotic expansion of U: ab ∞ Pns2 ∑ 2 2 2 22 ε <<1 8π 4 D n ,s =1 ( n / b + s / a ) ⇒ U (x, N ξ , Nη ) → 1 − 4 3 2 4π a D + + 1 2π 4 a 7bD 3 b 4 9 3 2π a D 1 + 4 8 3 2π a D + 1 2π 4 a 8 D 3 2  ∞ s 2 sin(nπx / b) P  b Ei I i ∑ ∑ 2 2 i 2 2 ns2  − 4 4 2 ∑ 4π a D i =1 s =1  n =1 ( n / b + s / a )  Nη Nη ∞ ∞ ∞ ∑∑ ∑ sin( µπxi / b) sin( µπx j / b) (µ 2 / b2 + s 2 / a 2 )2 i , j =1 s =1 µ =1 Nη + Nξ ∞ ∞ ∑ ∑∑ i , j = Nη +1 n =1 w=1 Nη Nη + Nξ ∑∑ ∑ i =1 j = Nη +1 r ,s =1 Nη Nη + Nξ ∞ ∑∑ ∑ j =1 i = Nη +1 n ,k =1 ∞ sin( wπxi / a) sin( wπx j / a) (n 2 / b 2 + w2 / a 2 ) 2 (r 2 / b 2 + s 2 / a 2 ) 2 sin(nπx j / b) sin(kπxi / a) (n 2 / b 2 + k 2 / a 2 ) 2 2  ∞ n 2 sin( sπx / a) P  Ei I i ∑ ∑ 2 2 i 2 2 ns2  + ∑ i = Nη +1 n =1  s =1 ( n / b + s / a )  ∞ 4 Ei I i s 4 sin(nπxi / b) Pns ∞ E j I j s sin(mπx j / b) Pms + ∑ ∑ ( n 2 / b 2 + s 2 / a 2 ) 2 m=1 ( m 2 / b 2 + s 2 / a 2 ) 2 n =1 sin(rπxi / b) sin( sπx j / a) ∞ Nη + Nξ ∞ ∑ s =1 (5.2.5) 4 Ei I i n 4 sin( sπxi / a) Pns ∞ E j I j n sin(qπx j / a) Pnq + ∑ (n 2 / b 2 + s 2 / a 2 ) 2 q=1 (n 2 / b 2 + q 2 / a 2 ) 2 4 Ei I i s 4 sin(nπxi / b) Pns ∞ E j I j r sin(qπx j / a) Prq + ∑ ∑ (n 2 / b 2 + s 2 / a 2 ) 2 q =1 (r 2 / b 2 + q 2 / a 2 ) 2 n =1 ∞ ∞ ∑ E j I j k 4 sin(mπx j / b) Pmk (m 2 / b 2 + k 2 / a 2 ) 2 m =1 Ei I i n 4 sin( sπxi / a) Pns ∑ (n 2 / b 2 + s 2 / a 2 ) 2 s =1 ∞ At this point, as the result is anyway an asymptotic one, several assumptions are made in order to get an analytic form of the solution of the optimal stiffening problem. First, the loading is assumed to be bisinusoidal, with half-periods equal to the lengths of the sides of the plate. Second, for simplicity and clarity of illustration, unidirectional stiffening is assumed, with only two stiffeners, positioned symmetrically with respect to the center of the plate (the a priori symmetry of the stiffening layout seems justified in view of the symmetry of the loading). With these assumptions, the work done by the loading becomes: abp 2 p02ε {sin 2 (π [1 / 2 + δ ]) + sin 2 (π [1 / 2 − δ ])} 1 U (δ , ε , p0 , a, b, D) → 4 0 − + ε <<1 8π D (1 / b 2 + 1 / a 2 ) 2 4π 4 a 2 D (1 / b 2 + 1 / a 2 ) 4 + p02ε 2 sin 2 (π [1 / 2 + δ ]) ∞ sin 2 (nπ [1 / 2 + δ ]) + ∑ 2π 4 a 5bD (1 / b 2 + 1 / a 2 ) 4 n=1 (n 2 / b 2 + 1 / a 2 ) 2 + p02ε 2 sin(π [1 / 2 + δ ]) sin(π [1 / 2 − δ ]) ∞ sin(nπ [1 / 2 + δ ]) sin(nπ [1 / 2 − δ ]) + ∑ π 4 a 5bD (1 / b 2 + 1 / a 2 ) 4 (n 2 / b 2 + 1 / a 2 ) 2 n =1 + p02ε 2 sin 2 (π [1 / 2 − δ ]) ∞ sin 2 (nπ [1 / 2 − δ ]) ∑ 2π 4 a 5bD (1 / b 2 + 1 / a 2 ) 4 n=1 (n 2 / b 2 + 1 / a 2 ) 2 (5.2.6) where δ is the distance of each of the stiffeners from the center of the plate divided by b, ε ≡ EI / aD and a is the length of each of the stiffeners. Employing trigonometric identities and recalling the definition: β ≡ b / a one can obtain a simpler expression for the (non-dimensionalized) elastic strain energy of the loaded structure, consistent of the plate and the stiffeners: ∞ ∞ cos 2 (πδ )  1 cos(2nπδ )   (5.2.7) β 5  Uˆ (δ , ε , β ) → 2 1 − (4β 3ε ) 2 1 + 2β 3ε ∑ + 2β 3ε ∑   2 2 2 2 2 2 2 2  ε <<1 ( β + 1)  ( β + 1)  n =1,3, 5 ( β + n ) n =1, 3, 5 ( β + n )     where the actual energy can be obtained from the non-dimensional form by: U (δ , ε , β , a, p0 , D ) = 13 a 6 p02 ˆ U (δ , ε , β ) 8π 4 D (5.2.8) Next, the series in (5.2.7) can be expressed in terms of the Fourier cosine series of coshz and zsinhz, z being a linear function of \|δ\|. Thence, substituting these functions instead of their Fourier series representations, one obtains: Uˆ (δ , β , ε ) → ε <<1 β5 ( β + 1) 2 2 − 4 β 8ε cos 2 (πδ ) ( β 2 + 1) 4 − 8β 11ε 2 cos 2 (πδ )  π coth(πβ ) π 2 coth 2 (πβ ) π coth(πβ / 2) π 2 coth 2 (πβ / 2) 3π 2  + − − −   3 ( β 2 + 1) 4 4β 2 8β 3 16 β 2 16 β 2   4β − 8β 11ε 2 cos 2 (πδ ) π cosh[πβ (1 − 2 δ )] π cosh(πβ ) cosh[πβ (1 − 2 δ )] π (1 − 2 δ ) sinh[πβ (1 − 2 δ )] + −  3 ( β 2 + 1) 4 4β 2 sinh 2 (πβ ) 4 β 2 sinh(πβ )  4 β sinh(πβ ) 2 (5.2.9) 2 π cosh[πβ (1 / 2 − 2 δ )] π 2 cosh(πβ / 2) cosh[πβ (1 / 2 − 2 δ )] π 2 (1 − 4 δ ) sinh[πβ (1 / 2 − 2 δ )]  − − +  8β 3 sinh(πβ / 2) 16 β 2 sinh 2 (πβ / 2) 16 β 2 sinh(πβ / 2)  Differentiating with respect to δ one gets: 4πβ 8 ε sin( 2πδ ) ∂ ˆ U (δ , β , ε ) → + ε <<1 ∂δ ( β 2 + 1) 4 + [ 4 coth(πβ ) + 4πβ coth 2 (πβ ) − 2 coth(πβ / 2) − πβ coth 2 (πβ / 2) − 3πβ ] + − − + π 2 β 8 ε 2 sin( 2πδ ) + 2( β 2 + 1) 4 π 2 β 8 ε 2 sin(2πδ )  4[sinh(πβ ) + πβ cosh(πβ )] cosh[πβ (1 − 2 δ )] 4πβ (1 − 2 δ ) sinh[πβ (1 − 2 δ )] −  (5.2.10) sinh(πβ ) 2( β 2 + 1) 4 sinh 2 (πβ )  [2 sinh(πβ / 2) + πβ cosh(πβ / 2)] cosh[πβ (1 / 2 − 2 δ )] sinh (πβ / 2) 2 + πβ (1 − 4 δ ) sinh[πβ (1 / 2 − 2 δ )]   sinh(πβ / 2)  3 3 β 10 ε 2 cos 2 (πδ )  4π (1 − 2 δ ) cosh[πβ (1 − 2 δ )] 4π cosh(πβ ) sinh[πβ (1 − 2 δ )] − +  sinh(πβ ) ( β 2 + 1) 4 sinh 2 (πβ )  π 3 cosh(πβ / 2) sinh[πβ (1 / 2 − 2 δ )] sinh 2 (πβ / 2) − π 3 (1 − 4 δ ) cosh[πβ (1 / 2 − 2 δ )]   sgn(δ ) sinh(πβ / 2)  Now, in order to determine whether placing the stiffeners together is optimal, one could have examined whether the second derivative of the energy with respect to δ at δ =0 changes its sign as ε increases. Such a change of sign would have indicated that unified stiffeners cease to correspond to a global minimum of the energy and start to correspond to its local maximum as their rigidity is increased. This would imply that a second order asymptotic expansion of the energy enables one to find a critical relative rigidity, ε, above which the stiffeners are best to be split and that this critical relative rigidity is small enough for the second order approximation to be a good one. However, the second derivative with respect to δ of the second order expansion of the energy with respect to ε is undefined at δ =0 due to the non-continuousness of the sign function. Therefore, instead, one can check whether the first derivative vanishes anywhere besides at δ =0 as ε increases. In fact, it is enough to check the sign of lim δ −1∂Uˆ / ∂δ . If the sign is positive for every positive ε, δ →0 + then placing the stiffeners together never corresponds to a local maximum of the second order expansion of the energy. Section 5.1 shows that placing the stiffeners together is globally optimal in the limit of a small positive ε. Therefore if δ =0 happens to be a local minimum of the energy for every ε in the range of the validity of the second order asymptotics and there is another, global, minimum elsewhere then, due to the continuity of the energy with respect to both δ and ε there has to be a local maximum between these two local minima. Numerical examination shows that when ε is high enough such that δ =0 ceases to be the global optimum, the globally optimal solution corresponds to δ continuously increasing with ε. Therefore in the limit case where ε is only infinitesimally higher than its critical value, the energy would have three extrema (two local minima and a local maximum in between) located at a line segment with a length tending to zero. This would imply infinite derivatives and therefore contradict with the fact that the energy is Hölder continuous in the vicinity of zero (as one can show by expanding δ −1∂Uˆ / ∂δ as a Taylor series around δ = 0. By expanding 14 δ −1∂Uˆ / ∂δ rather than ∂Û / ∂δ one eliminates the non-continuous functions). This contradiction can only be avoided by the assumption that wherever δ =0 is a local minimum of the energy, it is also a global minimum and that in order for it not to be a global minimum it has to be a local maximum. Also, a function Hölder continuous in the vicinity of zero behaves as a power function near zero and thus cannot have three extrema within a line segment starting at zero and having an arbitrarily small length. Hence, by proving that lim δ −1∂Uˆ / ∂δ is positive for every ε for which the second order expansion of the δ →0 + energy is valid, one proves that placing the stiffeners together is globally optimal (within the abovementioned range of rigidities). Next, expanding the hyperbolic functions of δ in (5.2.10) using the identities of hyperbolic trigonometry, dividing by δ and computing the limits, recalling that: sinh( 2πβ δ ) sgn(δ ) /[ 2πβδ ] = sinh( 2πβδ ) /[ 2πβδ ] → 1 δ →0 (5.2.11) the following result is obtained: 8π 2 β 8ε (1 − ε / ε cr ) ε <<1 ( β 2 + 1) 4 lim+ δ −1∂Uˆ / ∂δ → δ →0 (5.2.12) where ε cr ≡ 4 π {2(1 + β )[coth(πβ / 2) − 2 coth(πβ )] + πβ ( β 2 − 1)[4 coth 2 (πβ ) − coth 2 (πβ / 2) − 3]} 2 (5.2.13) Obviously, the limit in (5.2.13) is positive if εcr is negative for every positive β. Asymptotic expansions of εcr(β) show that this is indeed the case for the limit cases of β <<1, β >>1 and \|1–β\| <<1: ε cr → − β >>1 2 π (1 + β 2 ) < 0 , ε cr → − β <<1   24 π 2 β 3[12 + 5π 2 β (1 − β )] ε cr → − π −1[2 coth(π ) − coth(π / 2)]−1 1 + 1 + π 1− β <<1   ≅ −0.8831(1 − 0.6069β ) < 0 (∀ β < 1.6477) <0  3 + coth 2 (π / 2) − 4 coth 2 (π )   (1 − β )  ≅ 2 coth(π ) − coth(π / 2)   (5.2.14) (5.2.15) Numerical verification shows that εcr(β) is negative for all positive finite values of β. Therefore the limit in (5.2.12) is positive and thus for small enough positive finite values of ε it is globally optimal to stiffen a simply supported thin rectangular plate in elastostatic bending under a lateral bisinusoidal loading with two stiffeners positioned at the same place, in the geometric center of the plate, in parallel to one of the edges of the plate. The fact that the global optimality of this stiffening layout holds for finite values of ε, at least for the specified loading, is due to the fact that this result is true for both the first and the second order asymptotic expansions of the elastic energy of the stiffened plate, and it is the essence of the present subsection. The (non-dimensionalized) structural strain energy corresponding to the globally optimal layout discussed above can be obtained in a second order approximation with respect to ε by setting δ =0 in (5.2.9): Uˆ opt ( β , ε ) → ε <<1 β5 ( β + 1) 2 2 − 4β 8ε [1 + f ( β )ε ] ( β 2 + 1)4 (5.2.16) where: f ( β ) ≡ [4 coth(πβ ) + 4πβ coth 2 (πβ ) − 2 coth(πβ / 2) − πβ coth 2 (πβ / 2) − 3πβ ]π / 4 (5.2.17) Both the first and the second order approximations of the energy hold for small enough values of ε, formally, when f(β)ε = γ = o(1). Then the higher orders contribute negligibly to the first and second order approximations of the energy for which the degenerated stiffening layout is proven herein to be globally optimal for general and bisinusoidal loadings, respectively. Recalling the definition of the relative bending rigidity and assuming blade stiffeners with cross-sectional width d and static height h and a plate with thickness t and Poisson's ratio ν, and that the plate and the stiffener have the same Young modulus, one can obtain an approximation for the maximal static height of the stiffener for which it is still globally optimal not to split it. Hence, assuming that the cross-sectional width of the stiffener is much smaller than its length, as one might expect from a beam, namely that d/a = ζ = o(1), and making the approximation γ~ ζ based on the fact that both these quantities should be smaller than unity and are compliant to no additional restrictions (the values of γ and ζ can each be, say, 0.01), and, last, taking ν to be 1/3, which is typical for structural materials, one obtains the following approximation: 15 1/ 3 h 1 9  < t 2  f ( β )  (5.2.18) Taking the limit of (5.2.17) for β>>1 one obtains the result that for a short stiffener the critical static height is hmax ~ [9 /(4π )]1/ 3 t ≅ 0.9t . For a square plate, β = 1, hmax ~ t. That is, if the stiffener is not relatively long, it can have about the same static height as the thickness of the plate and it would still be flexible enough for it to be globally optimal not to split it. This result is impractical, since the stiffener is assumed to be placed symmetrically with respect to the mid-surface of the plate and thus if its static height is equal to the thickness of the plate and Young’s modulus of the plate and the stiffener are the same, then it is as having no stiffener at all, and it does not matter whether it is split or not. However, if the stiffener is longer, its length makes it more flexible such that its static height can be larger for the optimality of stiffening with a single stiffener yet to hold. For example, it follows from (5.2.17) and (5.2.18) that for the case β = 1/2 the critical static height is hmax ~ 1.5t, for β = 1/3, hmax ~ 2t and for β = 1/6, hmax ~ 4t. This simple calculation shows that, at least for long stiffeners, the degenerated stiffening layout is globally optimal for rather realistic parameters and not only, as one could presume, for stiffeners of absolutely negligible bending rigidity. In the results obtained above, the globally optimal layout corresponds to a thin plate with thin stiffeners. The fact that the bending rigidity of the stiffeners normalized by the bending rigidity of the plate is small does not imply that for the stiffeners to be realistic the plate should be thick, since a stiffener of width and height of the order of the thickness of the plate has a much smaller bending rigidity than that of the plate, as it appears from the definition of the relative rigidity, ε ≡ EI / aD , which is proportional to the ratio of the width of the stiffener and the breadth of the plate, which is usually a small number. Thus the above derivations do not contradict with the assumptions of the first order Love-Kirchhoff plate bending theory. The following subsection examines the opposite case, where the stiffeners are highly rigid compared to the plate. If flexible stiffeners have an asymptotically optimal layout in which they are combined to form a single stiffener, then, as is it turns out, highly rigid stiffeners have their own asymptotically optimal layout, where they are optimally spread, and to which they tend as their rigidities increase. 5.3 Highly rigid unidirectional stiffeners with no torsional rigidity – 1st order asymptotics The present subsection examines the case where the stiffeners are highly rigid with respect to the plate. By means of a first order asymptotic analysis with respect to the flexibilities of the stiffeners, it is shown that rigid enough stiffeners are best to be positioned separately, in locations determined by the functionality of the loading. It is shown that the optimal positions of the stiffeners approach their asymptotic limits as the rigidities of the stiffeners approach infinity. The functionality of this approach is given analytically for bisinusoidal loading with half-periods equal to the lengths of the edges of the plate, unidirectional stiffening with two stiffeners positioned symmetrically with respect to the mid-span of the plate and the condition that the stiffeners are relatively long. Substituting the expression for the deflection field for the case, as given in (2.20) into (2.12) and making use of (2.9), employing the definitions of β and ε used in the previous subsections, and defining xˆ ≡ x / b one obtains the strain energy of the stiffened plate, as follows: ( ) −1 P 2 a 6 sin(π xˆ T ) ε −11 + Bˆ sin(π xˆ ) β P2a6 ˆ U ( x) = 4 − 8π D ( β −2 + 1) 2 4π 4 D ( β −2 + 1) 4 (5.3.1) where xˆ = ( xˆ ,1 − xˆ )T . Evaluating expression (2.17) for the abovementioned assumptions and dividing it by ε, the coefficients defining the matrix in (5.3.1) can be readily obtained: 1  1 − cosh(2πβ xˆ ) cosh(πβ ) − cosh(πβ [1 − 2 xˆ ]) xˆ sinh(πβ [1 − 2 xˆ ])  Bˆ11 = 3  β + − 2β  2 π sinh(πβ ) 2π  sinh (πβ ) sinh(πβ )  1  1 − cosh(2πβ [1 − xˆ ]) cosh(πβ ) − cosh(πβ [1 − 2 xˆ ]) [1 − xˆ ]sinh(πβ [1 − 2 xˆ ])  Bˆ 22 = 3  β + + 2β  2 2π  sinh (πβ ) sinh(πβ ) π sinh(πβ )  1  cosh(πβ [1 − 2 xˆ ]) − cosh(πβ ) cosh(2πβ xˆ ) − 1 (1 − 2 xˆ ) sinh(2πβ xˆ )  Bˆ12 = Bˆ 21 = 3  β + +β  2π  sinh 2 (πβ ) sinh(πβ ) π sinh(πβ )  16 (5.3.2) Considering ε large enough for the spectral norm of ε −1Bˆ −1 to be much smaller than unity and expanding the expression for the energy in (5.3.1) into a first order Taylor series of the relative flexibility of the stiffeners, ε–1, one obtains the following asymptotic relation: β P2a6 P 2 a 6 [sin(π xˆ T )Bˆ −1 sin(π xˆ ) − ε −1 sin(π xˆ T )Bˆ −2 sin(π xˆ )] − ε >>1 8π 4 D ( β −2 + 1) 2 4π 4 D ( β −2 + 1) 4 U ( xˆ ) → (5.3.3) Noting the symmetry of the sine function, inverting the 2x2 matrices in (5.3.3) analytically, and introducing the auxiliary variables, (5.3.4) Bij ≡ 2π 4 sinh 2 (πβ ) Bˆ ij one gets the following representation of the energy: 2 2 βP 2 a 6 P 2 a 6 sinh 2 (πβ )  B22 − 2 B12 + B11 −1 ( B22 − B12 ) + ( B11 − B12 )  2 − − ε   sin (πxˆ ) (5.3.5) ε >>1 8π 4 D ( β − 2 + 1) 2 2 D( β − 2 + 1) 4  B11 B22 − B122 ( B11 B22 − B122 ) 2  U ( xˆ ) → Taking a Taylor series expansion with respect to β of the Bij components for a small β and retaining only the first non-vanishing powers of β results in: B12 = B21 = → 2(πβ )5 xˆ 2 (1 − 2 xˆ 2 ) / 3 , B22 → B11 → 4(πβ )5 xˆ 2 (1 − xˆ )2 / 3 β <<1 β <<1 β <<1 (5.3.6) Substituting (5.3.6) into (5.3.5), one gets an asymptotic representation of the energy for β<<1 (which, due to the high order of the expansion, is a fair approximation for any β<1): β P 2a6 P 2 a 6 sinh 2 (πβ ) − ε >>1 8π 4 D ( β −2 + 1) 2 2 D ( β −2 + 1) 4 (πβ )5 β <<1 U ( xˆ ) →  3sin 2 (π xˆ ) 9sin 2 (π xˆ )  −5 −1 − ( ) πβ ε  ˆ2  3 2(3 xˆ 2 − 4 xˆ 3 ) 2   3 x − 4 xˆ (5.3.7) The asymptotic limit for the optimal location of long rigid stiffeners can be obtained by minimization of the limit of (5.3.7) for ε → ∞,0 < β << 1, β 5ε → ∞ (that is finitely long and infinitely rigid stiffeners):  12π sin 2 (πxˆ )  as xˆopt = arg min lim U ( xˆ ) = arg min 1 −  ≅ 0.3445 3xˆ 2 − 4 xˆ 3  xˆ∈[ 0 ,1 / 2 ] ε →∞ xˆ∈[ 0 ,1 / 2 ]  β <<1 (5.3.8) β 5ε →∞ as Next, substituting the definition δ ≡ xˆ − xˆopt into (5.3.7), performing a 2nd order asymptotic expansion with respect to δ around zero, and differentiating with respect to δ, noting that this derivative should vanish for δ , ε −1 → 0 , one obtains an asymptotic expansion of ∂U (ε −1 , β , δ << 1) / ∂δ (which in this case becomes a linear function of δ). Finding the single root of the abovementioned derivative produces an expression for the globally optimal location of a stiffener (the one closer to the origin of the coordinate system) in a nondimensional form, x̂ , as a function of ε and β for β , ε −1 , β −5ε −1 << 1 , as follows: 1 / 2 − xˆopt as 1 / 2 − xˆopt → 1 − qβ −5ε −1 ε −1 ,β ,β −5ε −1 <<1 (5.3.9) where q is: as  cot(πxˆ opt  ) 12 − as  5 as 2 as as  2π ( xˆ opt )  (1 / 2 − xˆopt ) xˆ opt (3 − 4 xˆ opt )  q= ≅ 0.268 as as as 12 xˆopt − 3 1 / 2 − xˆopt 1 / 2 − xˆ opt as 2 as as − 12 + (3 − 4 xˆopt )[cot (πxˆopt ) − 1] − 24 cot(πxˆopt ) as 2 as as ( xˆopt ) xˆopt xˆopt 3 1 (5.3.10) As an illustration of the result obtained above, one might consider the following stiffening parameters: β = 0.9 (which seems acceptable due to the high order of the expansion with respect to β), ν =1/3, d/a = 0.01 and h/t = 10, which imply ε = 11.25. According to (5.3.9) and (5.3.10), the corresponding optimal position is: xˆopt ≅ 0.35 . It is therefore established that the globally optimal stiffening layout for even moderately long stiffeners consists of a single stiffener for a finite range of rigidities, and that for a certain value of the relative rigidity, ε, the stiffener should be split into two. Then, for two rigid stiffeners in an optimal layout, the distance between the stiffeners should increase (hyperbolically) with their rigidities until it approaches its asymptotic limit. The rate of this increase would in case of an optimal layout be approximately proportional 17 to the 5th power of the aspect ratio of the plate, defined as the length of the stiffeners divided by the length of the edge of the plate perpendicular to the stiffeners. 5.4 Highly flexible unidirectional stiffeners with non-negligible torsional rigidity – 1st order asymptotics The linear elastostatic strain energy for the case can be derived, using the results of sections 3 and 4, as follows. Introducing the small parameters ε ≡ EI /(aD) , γ ≡ GC /(bD) and deriving the auxiliary quantities required for the calculation of the strain energy for the asymptotic case, one gets: ∞ ∞ φ r ≡ υr − ∑ Λ rs (1 + B s ) −1 u s → υr − ∑ Λ rs (1 − B s ) u s → υr ε <<1 ε <<1 ( 4.11) s =1 s =1 O (γ ) O (γ ) O (ε ) O (γ ) O (ε ) O (γ ) O (γ ) O ( ε ) O (ε ) (5.4.1) −1 Ψ rs ≡ Ω rs − Λ rs (1 + B s ) Ts + δ rs 1 → Ω rs − Λ rs (1 − B s ) Ts + δ rs 1 → δ rs 1 ( 4.11) ε << 1 ε << 1 O (γ ) O (γ ) O (ε ) O (ε ) O (γ ) O (γ ) O (ε ) O (ε ) −1 θ r ≡ (1 + B r ) u r → (1 − B r ) u r → u r ; Θ r ≡ (1 + B r ) −1 Tr → (1 − B r ) Tr → Tr ε <<1 ε <<1 ε <<1 ε <<1 ( 4.15 ) ( 4.15) O (ε ) O (ε ) O (ε ) O (ε ) ∞ O (ε ) O (ε ) O (ε ) O (ε ) O (ε ) O (ε ) ∞ ⇒ υ r ← φ r = ∑ Ψ rs t s → ∑ (δ rs 1)t s = t r ⇒ t s → υ s γ <<1 ( 4.12) s =1 γ <<1 γ <<1 s =1 O (γ ) (5.4.2) O (γ ) −1 ⇒ v r = (1 + B r ) ( u r − Tr t r ) → (1 − B r )( u r − Tr t r ) → u r − Tr υ r → u r ε <<1 ( 4.7 ) ε <<1 ε <<1 O (ε ) O (ε ) O (ε ) O (ε ) O (ε ) O (ε ) O (ε ) O ( ε ) O (γ ) O (ε ) Then, substituting (5.4.1) and (5.4.2) into the expression for the deflection field given by (4.17), gives: ∞ 4 sin(nπξ / b) sin( sπη / a)  ab nπ  Pns − sin( nπx T / b)u s − cos(nπx T / b) υ s  ∑ 4 2 2 2 2 2  ε ,γ <<1 π abD 4 b ( n / b + s / a )   n , s =1 w(ξ ,η , x) → (5.4.3) Integrating it multiplied by half the loading (thus obtaining the expression for the external work or the total structural strain energy), ∞ Pns 1 nπ  ab  U ( x) → Pns − sin( nπx T / b)u s − cos( nπx T / b)υ s  (5.4.4) ∑ 4 2 2 2 2 2  ε ,γ <<1 2π D 4 b + ( n / b s / a )   n , s =1 and recalling definitions (2.10, (3.6), (4.6) and (4.10), produces the following asymptotic expression:  abN −1 U ( x) → ∑  4 ε i ,γ <<1 i =1  8π D  N − ∞  ∞ s 2 Pns sin( nπxi / b)  εi Pns2 − ∑ 2 2 2 2 2 4π 4 a 2 D ∑ ∑ 2 2 2 2 2  n , s =1 ( n / b + s / a ) s =1  n =1 ( n / b + s / a )  ∞ 2 (5.4.5) ∞ γj s − q (−1) s + q ∞ nPns cos(nπxi / b) ∞ mPmq cos(mπxi / b)  q  ∑ ∑ ∑ 6 2π bD s ,q =1 ( s − q )( s + q ) n =1 ( n 2 / b 2 + s 2 / a 2 ) 2 m =1 (m 2 / b 2 + q 2 / a 2 ) 2  As before, if the expression in the curved brackets attains its global minimum for a certain i at xi = x0, there is no reason why for different values of i, xi should not be equal x0 if one wishes to find a global minimum point of U. This is true whenever, N k N i =1 j =1 i =1 N k U (x) = ∑U i ( xi ) = ∑∑U i j ( xi ) = ∑∑U i j ( xi ) (5.4.6) i =1 j =1 and there is no additional equality constraint on the variables. Then the principle of separation of variables holds and the minimum of the sum is the sum of the minima (due to linearity, which always holds for a 1st order asymptotics). Then, all the minima are attained at the same point due to the fact that U i j ( xi ) is the same function of its argument for every i. Hence, for flexible enough stiffeners – and small enough bending rigidity necessarily implies small enough torsional rigidity, since E and G are usually of the same order of magnitude and I and C for closed, thin-walled sections are both approximately proportional to the smaller dimension and to the third power of the larger dimension of the cross section – the optimal layout, in the since of minimal structural compliance, is a single stiffener. That is, there should be no splitting of stiffeners before a critical bending rigidity is supplied. This phenomenon was shown for bidirectional stiffening when neglecting torsional rigidity and for unidirectional stiffening when accounting for it. It is reasonable to believe that this should be at least 18 approximately true for bidirectional stiffening with beams of non-negligible torsional rigidity (when the torsional rigidity is non-negligible with respect to the bending rigidity). Moreover, in order to emphasize that the obtained result is not absolutely trivial, it should be noted that the 2nd order expansion in section 5.2 showed coupling between the variables implying that the objective function cannot be written as in (5.4.6). In the case examined there, it is not obvious that the stiffeners should not be split. However, subsection 5.2 proves that this is still true even if not by the argument shown in (4.6), but by a more delicate reasoning. The combination of the results of sub-sections 5.1, 5.2, and 5.4 thus makes it not unreasonable to believe that for realistic bidirectional stiffening with closed cross-sections of non-negligible torsional rigidity, it may still be optimal to use only two orthogonal major stiffeners, for various loading distributions. Beyond the abovementioned, numerical analysis of stiffened plates performed with the use of commercial software, showed that the effect of torsional rigidity, at least for the blade stiffeners examined in section 5.2, corresponds to no significant change in the total work done by the external loading. Numerical optimization also showed that when the optimal layout consists of one or two perpendicular stiffeners, it is true for various cross-sections, both when accounting for and neglecting the effect of torsional rigidity. Last, in order to illustrate the result in (5.4.5), it should be pointed out that for a symmetric loading, xi = 1/2 makes the 2nd term in (5.4.5) maximal in absolute value and the third term zero. This coincides with the understanding that for symmetric loading, the mid-span would not twist and therefore a stiffener placed there would apply no torque and torsional rigidity would not contribute to the strain energy of the structure. 6 CONCLUSIONS This paper presented several results of an investigation in the subject of optimal stiffening of rectangular plates in elastostatic bending. A previously published method for the analysis of stiffened plate was obtained in more compact forms in sections 2, 3 and 4 herein and utilized to derive globally optimal stiffening layouts for limit-case parameters. It was shown that the optimal layout of each of two orthogonal sets of parallel, highly flexible stiffeners with negligible torsional rigidity consists of two orthogonal beams having increased rigidities positioned perpendicularly. In addition, the asymptotically globally optimal layouts for long stiffeners were obtained for both highly flexible and highly rigid stiffeners, to the 1st and 2nd orders of approximation for the former and the latter, respectively. The effect of the interaction of perpendicular stiffeners was shown to cancel out for stiffeners of negligible torsional rigidity. The effect of torsional rigidity was examined from the perspective of its ability to influence the optimal layout of asymptotically flexible stiffeners and was found to be insufficient, at least for a 1st order asymptotic expansion of the structural compliance with respect to the rigidities of the stiffeners, and for the special case of unidirectional stiffening. In addition, it should be mentioned that, although not shown here, the structural compliance minimization problem was solved for beams of different cross-sectional parameters, producing the result that the higher the bending rigidity of the stiffeners with respect to that of the plate, the more unidirectional the stiffening becomes in the optimum. Further noteworthy results correlating with the present work can be found in the works of Schade (1940), Clarckson (1965), Mróz and Rozvany (1975), Fuchs (1976), Samsonov (1978), Fuchs and Brull (1979), Cheng and Olhoff (1981), O’Leary and Harari (1985), Lam and Santhikumar (2003), Lagaros et al. (2004), Perchikov and Fuchs (2006) and Szczepanik (2006). The essence of the contribution of this work to the existing body of knowledge is the utilization of a method of solution enforcing continuous interfaces between a plate and a set of stiffeners, including a derivation of a compact form of the strain energy of the structure, based on the method, for the purpose of the derivation of globally optimal positions of uniform stiffeners, at least for the asymptotic cases of highly flexible and highly rigid stiffeners and under the simplifying assumptions that the loading is harmonic and the stiffeners are at least relatively long. The obtained results seem to give insight into the structural behavior of such widely used structural elements as stiffened panels. As globally optimal layouts are hard to obtain either in geometric or in topological optimization, and even more so when it comes to the determination of the symbolic dependence of the optimal layouts on the parameters, asymptotic analysis, together with a closed form of the solution of the elastostatic bending problem, are what it seems that might shed additional light on the area. Although asymptotic approach to the optimal stiffening problem has been employed before, the specific combination of assumptions and the results presented herein seem not too marginal to contribute to the understanding of the subject. On another subject, the deflection field of a thin rectangular plate subjected to arbitrary boundary conditions and a general double-sine-series-expressible loading is given by Fletcher and Thorne (1955) and discussed by Goriupp (1947) and Timoshenko (1959). As the solution procedure given by Fletcher and Thorne produces the deflection field of a plate having arbitrary boundary conditions given the deflection 19 field of a simply-supported plate subjected to a prescribed loading, clearly all one has to do is to supply the solution procedure given by these authors with the deflection field of a simply-supported stiffened plate instead of that of an unstiffened one. In this paper, the effect of the torsional rigidity of the stiffeners was accounted for in a way similar to the way it is done by Mazurkiewicz (1962). More compact expressions for the entities representing the solution in the case where torsional rigidity is non-negligible can be obtained using analytic summation of the series in section 4 in the same manner in which it is done in sections 2 and 3 and in the work of Fletcher and Thorne (1955). Acknowledgments The present work was partly supported by the Israeli Science Foundation under grant No. 13/05. REFERENCES 1. Andrianov IV, Lesnichaya VA, Manevich LI (1985) Metod usrednienia v statikie i dinamike rebristikh obolochek. Moskva , Nauka 2. Cheng KT, Olhoff N (1981) An investigation concerning optimal design of solid elastic plates. Int J Solids Struct 17:304-323 3. Clarkson J (1965) The Elastic Analysis of Flat Grillage. Cambridge University Press, Cambridge 4. Dems K, Mróz Z, Szelag D (1989) Optimal design of rib-stiffeners in disks and plates. Int J solids Struct 25:973-998 5. Fletcher HJ, Thorne CJ (1955) Bending of thin rectangular plates. In: Proceedings of the 2nd U.S. National Congress on Applied Mechanics, Ann Arbor, Michigan, 1954 6. Fuchs MB (1976) Substitute function methods in structural optimization and their application to continuous beams. Sc.D. dissertation, Technion 7. Fuchs MB, Brull MA (1979) A new strain energy theorem and its use in the optimum design of continuous beams. Comput Struct 10:647-657 8. Goriupp K (1947) Die dreiseitig gelagerte Rechteckplatte. Arch Appl Mech 16:153–163 9. Grayhack WT, Mahar TJ (1990) Buckling of rib-stiffened plates: an asymptotic approach. SIAM J. appl. Math. 50:1126-1133 10. Grigolyuk EI, Tolkachev VM (1980) Kontaktnyie zadachi teorii plastin i obolochek. Moskva, Mashinostroenie 11. Kalamkarov AL (1992) Composite and reinforcement elements of construction. Wiley, New York 12. Konchkovskii Z (1984) Plity. Staticheskie raschety. Moskva, Stroiizdat 13. Lagaros ND, Fragiadakis M, Papadrakakis M (2004) Optimum design of shell structures with stiffening beams. AIAA J 42:175-184 14. Lam YC, Santhikumar S (2003) Automated rib location and optimization for plate structures. Struct Multidisc Opt 25:35-45 15. Mazurkiewicz Z (1962) Bending and buckling of rectangular plate reinforced transversely by ribs with variable rigidities. Bull Acad Polon Sci Ser Sci Tech 10: 231-239 16. Mazurkiewicz Z (1962) Buckling of rectangular plate obliquely reinforced by ribs with variable flexural rigidity. Bull Acad Polon Sci Ser Sci Tech. 10:329-339 17. Mróz Z, Rozvany GIN (1975) Optimal design of structures with variable support conditions. J Optim Theory Applic 15:85-101 18. Nowacki W (1954) Statecznosc plyt prostokatnych wzmocnionych zebrami. Arch Mech Stos. 6:317342 19. Nowacki W (1954) Zagadnienia statyki i dynamiki plyt wzmocnionych zebrami. Arch Mech Stos 6:601-638 20. O’Leary JR, Harari I (1985) Finite element analysis of stiffened plates. Comput Struct 21:973-985 21. Perchikov N, Fuchs MB (2006) Optimal layouts of stiffeners for plates in bending – Topology optimization approach. Paper presented at the 3rd European conference on computational mechanics, solids, structures and coupled problems in engineering, LNEC, Lisbon, 5-8 June 2006 22. Samsonov AM (1978) The Optimal Location of a Thin Rib for an Elastic Plate. Izv Acad Nauk USSR, MTT (Mechanics of Solids) No.1, 132-138 23. Savin GN, Fleishman NP (1964) Plastinki i obolochki s rebrami zhestkosti. Kiev, Naukova Dumka 24. Schade HA (1940) The orthogonally stiffened plate under uniform lateral load. J Appl Mech 7 25. Szczepanik M (2006) Optimization of topology and stiffener locations in 2D structures using evolutionary methods. Paper presented at the 3rd European conference on computational mechanics, solids, structures and coupled problems in engineering, LNEC, Lisbon, 5-8 June 2006 26. Szilard R (1974) Theory and Analysis of Plates: Classical and numerical methods, Prentice Hall 27. Timoshenko SP, Woinowsky-Krieger S (1959) Theory of Plates and Shells, 2nd edn. McGraw Hill, New York 20	5cs.CE
arXiv:1712.05181v2 [cs.CL] 15 Dec 2017 Rasa: Open Source Language Understanding and Dialogue Management Tom Bocklisch Rasa tom@rasa.ai Joey Faulkner Rasa joey@rasa.ai Nick Pawlowski Rasa nick@rasa.ai Alan Nichol Rasa alan@rasa.ai Abstract We introduce a pair of tools, Rasa NLU and Rasa Core, which are open source python libraries for building conversational software. Their purpose is to make machine-learning based dialogue management and language understanding accessible to non-specialist software developers. In terms of design philosophy, we aim for ease of use, and bootstrapping from minimal (or no) initial training data. Both packages are extensively documented and ship with a comprehensive suite of tests. The code is available at https://github.com/RasaHQ/ 1 Introduction Conversational systems are becoming pervasive as a basis for human computer interaction as we seek more natural ways to integrate automation into everyday life. Well-known examples of conversational AI include Apple’s Siri, Amazon’s Alexa and Microsoft’s Cortana but conversational systems are becoming widespread with platforms like Facebook Messenger opening up to chatbot developers. Common tasks for conversational systems include scheduling meetings1 , booking flights2 and customer support tasks3 . Modern open source libraries are held to a high standard of professionalism, and this extends to implementations of machine learning algorithms. There is a large amount of non-research work involved in maintaining a widely used project, and code produced by research groups often falls short of expectations4 . Rasa NLU and Core aim to bridge the gap between research and application, bringing recent advances in machine learning to non-experts who want to implement conversational AI systems. We introduce Rasa NLU and Core as easy to use tools for building conversational systems, since until now there was no widely-used statistical dialogue system intended for non-specialists. Rasa is already used by thousands of developers worldwide. As with many other conversational systems, our tools are split into natural language understanding (Rasa NLU) and dialogue management (Rasa Core). Section 3 describes the code architecture, in 4 we outline the developer experience, and in 5 demonstrate an example application. 1 e.g. Meekan, see https://meekan.com e.g. KLMs BlueBot, see https://bb.klm.com/en 3 see http://rasa.ai/enterprise/case-studies/ 4 see https://www.reddit.com/r/MachineLearning/comments/6l2esd/d_why_cant_you_guys_ comment_your_fucking_code/ 2 NIPS 2017 Conversational AI workshop, Long Beach, CA, USA. 2 Related Work Rasa takes inspiration from a number of sources. Rasa’s API uses ideas from scikit-learn [8] (focus on consistent APIs over strict inheritance) and Keras [4] (consistent APIs with different backend implementations), and indeed both of these libraries are (optional) components of a Rasa application. Text classification is loosely based on the fastText approach [5]. Sentences are represented by pooling word vectors for each constituent token. Using pre-trained word embeddings such as GloVe [9], the trained intent classifiers are remarkably robust to variations in phrasing when trained with just a few examples for each intent. Braun et al [3] show that Rasa NLU’s performance compares favourably to various closed-source solutions. Custom entities are recognised using a conditional random field [6]. Rasa Core’s approach to dialogue management is most similar to [13], but takes a different direction from many recent research systems. There is currently no support for end-to-end learning, as in [12] or [2], where natural language understanding, state tracking, dialogue management, and response generation are all jointly learned from dialogue transcripts. Rasa’s language understanding and dialogue management are fully decoupled. This allows Rasa NLU and Core to be used independently of one another, and allows trained dialogue models to be reused across languages. For language generation, we encourage developers to generate a variety of responses by authoring multiple templates for each response. This is currently easier and more reliable than using (for example) a neural network to generate grammatically coherent and semantically correct responses as in [11]. Rasa Core also does not account for uncertainty in voice transcription and NLU, as would typically be achieved with Partially Observable Markov Decision Processes (POMDP) [15]. While there is support for training via reinforcement learning (currently in alpha), we do not emphasise this for new users. Rather than implementing a reward function and a simulated user, or immediately placing humans in the loop, we encourage developers to train the dialogue policy interactively (see section 4.2). Recent work by Williams follows a similar machine teaching approach [14], one difference being that in that work the dialogue policy is directly exposed to the surface form of the user utterances, whereas in Rasa Core the dialogue policy only receives the recognised intent and entities. 2.1 Related Software Libraries As mentioned in the introduction, the large body of high-quality research into statistical dialogue systems from the last decades has not translated into widely used software libraries. A notable contribution is to this end is PyDial [10], a recently released toolkit for dialogue research. Compared to PyDial, Rasa Core emphasises the needs of non-specialist software developers over those of researchers in the field. Other open toolkits include OpenDial [7] and RavenClaw [1]. There are a number of general-purpose natural language processing (NLP) libraries in widespread use, and it is not necessary to mention them all here. There are also a number of online services for natural language understanding (NLU), as the term is understood in the dialogue research community: converting short user messages into dialogue acts comprising an intent and a set of entities. Like the online services, Rasa NLU hides implementation details from new users, but with the advantage that slightly more experienced users can fully customise their NLU system. Language understanding is performed by a number of components implementing a common API, and is therefore easily configurable to suit the needs of a particular project. 3 Description of the Code Rasa’s architecture is modular by design. This allows easy integration with other systems. For example, Rasa Core can be used as a dialogue manager in conjunction with NLU services other than Rasa NLU. While the code is implemented in Python, both services can expose HTTP APIs so they can be used easily by projects using other programming languages. 3.1 Architecture Dialogue state is saved in a tracker object. There is one tracker object per conversation session, and this is the only stateful component in the system. A tracker stores slots, as well as a log of all the 2 events that led to that state and have occurred within a conversation. The state of a conversation can be reconstructed by replaying all of the events. When a user message is received Rasa takes a set of steps as described in figure 1. Step 1 is performed by Rasa NLU, all subsequent steps are handled by Rasa Core. Figure 1: 1. A message is received and passed to an Interpreter (e.g. Rasa NLU) to extract the intent, entities, and any other structured information. 2. The Tracker maintains conversation state. It receives a notification that a new message has been received. 3. The policy receives the current state of the tracker. 4. The policy chooses which action to take next. 5. The chosen action is logged by the tracker. 6. The action is executed (this may include sending a message to the user). 7. If the predicted action is not ‘listen’, go back to step 3. 3.2 Actions We frame the problem of dialogue management as a classification problem. At each iteration, Rasa Core predicts which action to take from a predefined list. An action can be a simple utterance, i.e. sending a message to the user, or it can be an arbitrary function to execute. When an action is executed, it is passed a tracker instance, and so can make use of any relevant information collected over the history of the dialogue: slots, previous utterances, and the results of previous actions. Actions cannot directly mutate the tracker, but when executed may return a list of events. The tracker consumes these events to update its state. There are a number of different event types, such as SlotSet, AllSlotsReset, Restarted, etc. The full list is in the documentation at https: //core.rasa.ai. 3.3 Natural Language Understanding Rasa NLU is the natural language understanding module. It comprises loosely coupled modules combining a number of natural language processing and machine learning libraries in a consistent API. We aim for a balance between customisability and ease of use. To this end, there are pre-defined pipelines with sensible defaults which work well for most use cases. For example, the recommended pipeline, spacy_sklearn, processes text with the following components. First, the text is tokenised and parts of speech (POS) annotated using the spaCy NLP library. Then the spaCy featuriser looks up a GloVe vector for each token and pools these to create a representation of the whole sentence. Then the scikit-learn classifier trains an estimator for the dataset, by default a mutliclass support vector classifier trained with five-fold cross-validation. The ner_crf component then trains a conditional random field to recognise the entities in the training data, using the tokens and POS tags as base features. Since each of these components implements the same API, it is easy to swap (say) the GloVe vectors for custom, domain-specific word embeddings, or to use a different machine learning library to train the classifier. There are further components for handling out-of-vocabulary words and 3 many customisation options for more advanced users. All of these are detailed in the documentation at https://nlu.rasa.ai. 3.4 Policies The job of a policy is to select the next action to execute given the tracker object. A policy is instantiated along with a featurizer, which creates a vector representation of the current dialogue state given the tracker. The standard featurizer concatentates features describing: • what the last action was • the intent and entities in the most recent user message • which slots are currently defined The featurization of a slot may vary. In the simplest case, a slot is represented by a single binary vector element indicating whether it is filled. Slots which are categorical variables are encoded as a one-of-k binary vector, those which take on continuous values can specify thresholds which affect their featurisation, or simply be passed to the featurizer as a float. There is a hyperparameter max_history which specifies the number of previous states to include in the featurisation. By default, the states are stacked to form a two-dimensional array, which can be processed by a recurrent neural network or similar sequence model. In practice we find that for most problems a max_history value between 3 and 6 works well. 4 Usage 4.1 Training Data Formats Both Rasa NLU and Core work with human-readable training data formats. Rasa NLU requires a list of utterances annotated with intents and entities. These can be specified either as a json structure or in markdown format. The markdown syntax is especially compact and easy to read and can be rendered by many text editors and web applications like GitHub. ## intent: restaurant_search - show me [chinese](cuisine) restaurants The json format is slightly more cumbersome to read, but is not whitespace sensitive and more suitable for transmission of training data between applications and servers. { "text": "show me chinese restaurants", "intent": "restaurant_search", "entities": [ { "start": 8, "end": 15, "value": "chinese", "entity": "cuisine" } ] } Rasa Core employs markdown to specify training dialogues (aka ‘stories’). ## story_07715946 * greet - utter_ask_howcanhelp * inform{"location":"rome","price":"cheap"} - utter_on_it 4 - utter_ask_cuisine * inform{"cuisine":"spanish"} - utter_ask_numpeople * inform{"people":"six"} - action_ack_dosearch A story starts with a name preceeded by two hashes ## story_03248462 The choice of name is arbitrary but can be helpful for debugging. The body of a story is a sequence of events, separated by newlines. An event such as * inform{"location":"bombay", "price":"expensive"} is a user utterance, annotated as a dialogue act. The general format is * intent{"entity_name": "entity_value"} where entities are key-value pairs separated by commas. System actions are also events and are specified by lines starting with a dash. The end of a story is denoted by a newline. 4.2 Machine Teaching In addition to supervised learning, Rasa Core supports a machine teaching approach where developers correct actions made by the system. We find that this is a practical approach for generating training data, and exploring the space of plausible conversations efficiently. 5 Here is an example of a user engaged in machine teaching of a restaurant recommender system (described in Section 5) . The user is presented with the following prompt5 : -----Chat history: bot did: [] bot did: action_listen user said: /greet whose intent is: greet we currently have slots: cuisine: None, people: None, price: None, location: None -----The bot wants to [utter_ask_howcanhelp] due to the intent. Is this correct? 1. 2. 3. 0. Yes No, intent is right but the action is wrong The intent is wrong Export current conversations as stories and quit >> If the user inputs 2, indicating that the action is wrong, they will be provided with a prompt as below which lists the possible actions and the probability assigned to them by the dialogue policy: -----what is the next action for the bot? 0 action_listen 0.12 utter_ask_cuisine utter_ask_helpmore utter_ask_howcanhelp utter_ask_location utter_ask_moreupdates utter_ask_numpeople action_search_restaurants 0.03 0.03 0.19 0.04 0.03 0.05 0.03 ... 8 9 10 11 12 13 14 ... >> Choosing the correct action creates a new training data point. Rasa Core then partially trains the dialogue policy and moves the conversation forward to the next step. Once completed, the trained model is persisted and the newly generated training data is saved to a file. 4.3 Visualisation of Dialogue Graphs Rasa Core also has the capability to visualise a graph of training dialogues. A story graph is a directed graph with actions as nodes. Edges are labeled with the user utterances that occur in between the execution of two actions. If there is no user interaction between two consecutive actions, the edge label is omitted. Each graph has an initial node called START and a terminal node called END. Note that the graph does not capture the full dialogue state and not all possible walks along the edges necessarily occur in the training set. To simplify the visualization a heuristic is used to merge similar nodes. A generated graph before and after running the simplification is shown in Figure 2. During the simplification two nodes are merged by replacing them with a single node that inherits all the 5 In Rasa a conversational system is referred to as a ‘bot’ 6 incoming and outgoing edges, removing duplicates in the process. This makes the resulting graphs easier to interpret. Nodes are merged if these two conditions are met: • they represent the same action • the previous max_history turns are identical for all dialogues leading to the nodes. START greet utter_ask_howcanhelp START greet greet utter_ask_howcanhelp utter_ask_howcanhelp inform{"location": "rome", "price": "cheap"} inform{"cuisine": "spanish"} inform{"location": "rome", "price": "cheap"} inform{"cuisine": "spanish"} utter_ask_cuisine utter_ask_location utter_ask_cuisine utter_ask_location inform{"cuisine": "spanish"} inform{"location": "rome", "price": "cheap"} inform{"cuisine": "spanish"} action_ack_dosearch action_ack_dosearch inform{"location": "rome", "price": "cheap"} action_ack_dosearch END END (a) Example story graph without simplification (b) Example story graph with simplification Figure 2: Minimal example illustrating how story graphs are simplified. The training data contains two stories with the same first interaction. These nodes are therefore considered equivalent and merged. 4.4 Deployment in a Production Environment The repositories for Rasa NLU and Core both contain Dockerfiles for producing static virtual machine (VM) images. This aids in reproducibility and ease of deployment to a variety of server environments. The web servers running the HTTP API support both thread-based and process-based parallelism, allowing them to handle large request volumes in a production environment. 5 Demonstration To demonstrate the usage of Rasa Core, we use the BAbl dialogue dataset [2]. This is a simple slot-filling exercise where the system is asked to search for a restaurant and has to fill several slots before being able to perform a successful search. The system may ask the user for their preference for any slot. The available slots are location, the number of people, the cuisine and the price range. This is an interesting dataset due to the inherent non-linearity of the problem - there are multiple ways to get the same information so there isn’t a single ‘correct’ action in every case. Accuracy and precision are therefore not the most appropriate metrics for evaluating a dialogue policy. Instead we consider how the system chooses actions depending on what information is already available. It should attempt to fill slots which are empty. In Figure 3 we can see the probabilities which Rasa Core attaches to each action given the slots it already knows. We see that Core follows a rough pattern of asking about cuisine, followed by location, followed by number of people - as it does in the training data. However it recognises that it could also ask about the other unfilled slots by attributing a non-zero probability to each one. The filled slots (the lower diagonal of the ‘grid’) are given vanishingly small probability. This illustrates how Rasa Core can use contextual clues to learn non-linear conversations. 6 Outlook Both Rasa NLU and Core are under active development. They serve as a platform for making applied research in conversational AI usable by non-specialist developers, and as such will never be ‘finished’. A number of topics are under active development, including improved support for reinforcement learning, making NLU robust to typos and slang, and supporting more languages. We also plan to 7 Figure 3: Here we plot the probabilities of choosing actions for the bAbl example (Section 5) . We sequentially inform the system of the correct (top-bottom): cuisine, location, number of people and price, and the system chooses the next action based on the information it already has (which is listed in the box on the top left). The system favours asking about slots it has not been informed of, until all of the slots are filled and it searches for the restaurant. release real-world datasets for comparing the performance of different models. The authors welcome external contributions to the project, the specifics of which can be found in the repositories on GitHub. Acknowledgements The authors are indebted to the users of both libraries for providing invaluable feedback and creating a supportive community around these tools. Special acknowledgement is owed to the external contributors to both libraries. Up-to-date lists of contributors may be viewed at https:// github.com/RasaHQ/rasa_nlu/graphs/contributors and https://github.com/RasaHQ/ rasa_core/graphs/contributors. References [1] D. Bohus and A. I. Rudnicky. The ravenclaw dialog management framework: Architecture and systems. Comput. Speech Lang., 23(3):332–361, July 2009. [2] A. Bordes and J. Weston. Learning end-to-end goal-oriented dialog. CoRR, abs/1605.07683, 2016. [3] D. Braun, A. Hernandez-Mendez, F. Matthes, and M. Langen. Evaluating natural language understanding services for conversational question answering systems. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue, pages 174–185, 2017. [4] F. Chollet et al. Keras, 2015. 8 [5] A. Joulin, E. Grave, P. Bojanowski, and T. Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. [6] J. Lafferty, A. McCallum, and F. C. Pereira. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. 2001. [7] P. Lison and C. Kennington. Opendial: A toolkit for developing spoken dialogue systems with probabilistic rules. ACL 2016, page 67, 2016. [8] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al. Scikit-learn: Machine learning in python. Journal of Machine Learning Research, 12(Oct):2825–2830, 2011. [9] J. Pennington, R. Socher, and C. Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532–1543, 2014. [10] S. Ultes, L. M. R. Barahona, P.-H. Su, D. Vandyke, D. Kim, I. Casanueva, P. Budzianowski, N. Mrkšić, T.-H. Wen, M. Gasic, et al. Pydial: A multi-domain statistical dialogue system toolkit. Proceedings of ACL 2017, System Demonstrations, pages 73–78, 2017. [11] T.-H. Wen, M. Gasic, N. Mrksic, P.-H. Su, D. Vandyke, and S. Young. Semantically conditioned lstm-based natural language generation for spoken dialogue systems. arXiv preprint arXiv:1508.01745, 2015. [12] T.-H. Wen, D. Vandyke, N. Mrksic, M. Gasic, L. M. Rojas-Barahona, P.-H. Su, S. Ultes, and S. Young. A network-based end-to-end trainable task-oriented dialogue system. arXiv preprint arXiv:1604.04562, 2016. [13] J. D. Williams, K. Asadi, and G. Zweig. Hybrid code networks: practical and efficient end-to-end dialog control with supervised and reinforcement learning. arXiv preprint arXiv:1702.03274, 2017. [14] J. D. Williams and L. Liden. Demonstration of interactive teaching for end-to-end dialog control with hybrid code networks. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue, pages 82–85, 2017. [15] S. Young, M. Gašić, B. Thomson, and J. D. Williams. Pomdp-based statistical spoken dialog systems: A review. Proceedings of the IEEE, 101(5):1160–1179, 2013. 9	2cs.AI
SENATE: A Permissionless Byzantine Consensus Protocol in Wireless Networks Zhiyuan Jiang1 , Bhaskar Krishnamachari2 , Sheng Zhou1 , Zhisheng Niu1 , arXiv:1803.08694v1 [cs.IT] 23 Mar 2018 1 {zhiyuan, sheng.zhou, niuzhs}@tsinghua.edu.cn, Tsinghua University, Beijing, China 2 bkrishna@usc.edu, University of Southern California, Los Angeles, USA Abstract—The blockchain technology has achieved tremendous success in open (permissionless) decentralized consensus by employing proof-of-work (PoW) or its variants, whereby unauthorized nodes cannot gain disproportionate impact on consensus beyond their computational power. However, PoWbased systems incur a high delay and low throughput, making them ineffective in dealing with real-time applications. On the other hand, byzantine fault-tolerant (BFT) consensus algorithms with better delay and throughput performance have been employed in closed (permissioned) settings to avoid Sybil attacks. In this paper, we present Sybil-proof wirelEss Network coordinAte based byzanTine consEnsus (SENATE), which is based on the conventional BFT consensus framework yet works in open systems of wireless devices where faulty nodes may launch Sybil attacks. As in a Senate in the legislature where the quota of senators per state (district) is a constant irrespective with the population of the state, “senators” in SENATE are selected from participating distributed nodes based on their wireless network coordinates (WNC) with a fixed number of nodes per district in the WNC space. Elected senators then participate in the subsequent consensus reaching process and broadcast the result. Thereby, SENATE is proof against Sybil attacks since pseudonyms of a faulty node are likely to be adjacent in the WNC space and hence fail to be elected. Index Terms—Byzantine fault-tolerant consensus, Sybil attack, wireless network, permissionless blockchain, distance geometry 1. Introduction Recent years have witnessed an explosive development in digital cryptocurrency, both in academic fields and financial markets. Behind its tremendous success, the key enabling technology of digital cryptocurrency is the blockchain [1], [2] which combines several judiciously designed techniques to facilitate trusted distributed ledgers such that intermediary can be eliminated during transactions. In particular, the Bitcoin blockchain ingeniously adopts the proof-of-work (PoW) for mining to, among other purposes, deal with identity attacks (Sybil attacks) in open (permissionless) systems wherein the identities of participating nodes are not assumed to be known a priori. Specifically, each node, whether it is authentic or a pseudonym, must solve a cryptographic puzzle to participate in the block generation process and obtain rewards (mining). Therefore, the impact of a mining Figure 1. Time-critical application scenarios, e.g., autonomous driving systems wherein vehicles and pedestrians go through intersections based on distributed consensus (left); Internet-of-Things where terminals (drones, sensors and actuators) act based on coordinated and synchronized behavior (right). node is directly tied to its computational power, irrespective of the number of identities it has. According to the necessity of a prior identity authorization procedure, blockchain technologies can be categorized by permissionless and permissioned blockchains. Permissionless blockchains, such as Bitcoin [1] and Ethereum [3], are applied in open systems wherein faulty nodes may apply Sybil attacks, the counteraction of which usually involves enforcing a strict coupling between the consensus impact of a node and the computational power (PoW for Bitcoin) or the resources (proof-of-stake for Ethereum Casper) of the node. Despite the robustness against Sybil attacks, a price is payed that permissionless blockchains usually suffers from high processing delay (6 blocks are recommended in Bitcoin which amounts to one hour) and low throughput (at most 7 transactions per second [4] for Bitcoin). Many existing works try to remedy this issue [4], [5], however, the inherent mining based probabilistic consensus reaching technique is the key limiting factor. On the other hand, permissioned blockchains, such as Hyperledger Fabric [6], need not be wary of Sybil attacks since all participating nodes have gone through an explicit authentication process such that they can be trusted. Adopting a long line of existing research on byzantine fault-tolerant (BFT) protocols [7]–[9], the processing delay and throughput of permissioned blockchains can be dramatically improved; a nice comparison between traditional BFT protocols and permissionless blockchains is presented by Vukolić [10]. In this paper, a large-scale dense wireless network scenario is considered, which is likely to be encountered in the future in Internet-of-Things deployments as well as in TABLE 1. C OMPARISONS WITH BLOCKCHAIN TECHNOLOGIES Open system Delay Permissionless blockchain Yes Large Permissioned blockchain No Small SENATE Yes Small the context of vehicle networks and intelligent transportation (Figure 1). We come up with a solution for Sybil-proof BFT consensus which offers the benefits of both permissioned and permissionless blockchains (cf. Table 1 for details), i.e., low-delay and high-throughput BFT consensus in permissionless systems. The proposed protocol, namely Sybilproof wireless network coordinate (WNC) based byzantine consensus (SENATE), consists of three major phases: sortition, senator selection and byzantine agreement. SENATE thwarts the Sybil attack by exploiting the fact that even a faulty node cannot forge its wireless channel to other nodes such that a unique wireless fingerprint can be leveraged to identify nodes; a fully decentralized approach is proposed to achieve this. Notations: Throughout the paper, we use boldface uppercase letters, boldface lowercase letters and lowercase letters to designate matrices, column vectors and scalars, respectively. The transpose of a matrix is denoted by (·)T . Xi,j and xi denotes the (i, j)-th entry and i-th element of matrix X and vector x, respectively. The `p and nuclear norm of a matrix are denoted by k·kp and k·k∗ , respectively. The vector consisting of the diagonal entries of a square matrix is denoted by diag[·]. The trace of a matrix is denoted by tr[·]. The matrix with all entries being one is denoted by 1, and likewise zero matrix is denoted by 0. 1.1. Related Work The idea of only allowing selected nodes to participate in BFT consensus reaching is shared by, e.g., NEO [11] and Algorand [4]. NEO is a delegated BFT consensus based blockchain in which a small number of servers are statically configured to run consensus on behalf of a larger open network. Similar with SENATE, Algorand counteracts Sybil attacks by adopting a sortition phase; the difference is that a random verifiable function based solution combining with proof-of-stake (PoS) is leveraged by Algorand, whereas SENATE is based on the underlying wireless channels. Because SENATE is not employing PoS, it is not tied to a digital currency and thus can be applied more broadly to achieving consensus in wireless systems; moreover, it avoids the unfairness introduced by PoS which intentionally favors participants with more resources. Proof-of-location (PoL) in wireless networks is adopted by Dasu et al. [12] to replace PoW for faster transactions. The PoLs are generated by authorities such as wireless network operators and hence some notion of centralization is introduced. SENATE also uses the concept of PoL whereas, on a high level, nodes generate PoLs in a fully-decentralized manner without any trusted authority. There have been some work where wireless channel fingerprints are utilized to protect against Sybil attacks (cf. [13] for a survey and [14], [15] for a signalprint-based approach). However, most existing work relies on a trusted authority to verify the wireless channels of nodes [14], or pre-distribute encrypted keys [13]. In [15], a wireless ad hoc network of commodity 802.11 devices is considered; a view selection policy based on signalprint observations is proposed. Our work considers a fully decentralized wireless network and a novel WNC based protocol (SENATE) is proposed; compared with existing work, SENATE has much lower running time and better understandability and hence more favourable for real-world implementations. 2. System Model We consider a wireless network consisting of N geographically distributed nodes with full connectivity, namely any pair of nodes in the network are within each other’s radio communication range. The system is open, or permissionless, in the sense that any node can join the system without prior identity authentication. The objective is for the good nodes to reach a valid (the definition for validity is addressed later) consensus on a set of values over a certain time period such that deterministic concerted actions can be carried out, in the meantime, subject to malicious behaviors by faulty nodes. Note that the considered scenario distinguishes from the state machine replication wherein a log is proposed by a client and different nodes agree on the same log record; here different nodes may have different set of initial values, e.g., sensory data from environment, and hence a reasonably good (valid) consensus needs to be reached. Unlike existing work on byzantine consensus which mainly adopts the Internet as the overlay network, a wireless overlay is considered. In this regard, the behavior of a faulty node should be clarified. Specifically, the following assumptions are made in this paper. • • • The objective of a faulty node is to rig the consensus reaching process to benefit itself, rather than halting the process. To achieve its purpose, possible malicious behaviors include: (1) Byzantine node [7], namely it does not comply with the protocol and can report arbitrary messages; (2) Sybil attack [16]: it can generate pseudonyms to gain inappropriate power in the process of reaching consensus. In the overlay wireless network, a faulty node does not block or interfere with other nodes’ transmissions and messages. The first assumption describes the motive of a faulty node and therefore has implications on the other two assumptions. The second assumption simply states that, on a message level, there is no restriction on the behavior of a faulty node, both from the perspectives of the message content and the identity of the message sender. In most existing byzantine agreement protocols with the Internet as overlay [7], the third assumption is also implied which limits the malicious behavior of a node to itself; whereas in wireless networks with the broadcast nature of electromagnetic waves, this assumption has more implications, meaning that a faulty node is assumed to comply with the communication protocol. For instance, a faulty node would not transmit when another node is scheduled (by the consensus protocol). This assumption stems in large part from the first assumption, since messing with the communication protocol, e.g., transmitting with a high power and thus blocking other nodes, leads to retransmissions and hence halting the consensus reaching process. Besides, the following two reasons also justify the assumption: (1) an attack becomes trivially devastating without this assumption, namely a faulty node with sufficient transmit power can block other transmissions all the time to prevent reaching consensus; (2) a node not complying with the communication protocol is obviously malicious and easy to spot. In this work, we assume nodes can obtain ranging estimations based on others’ pilot signals. However, we do not focus on specific methods to obtain the distance estimations; they can be based on receive signal strength (RSS), time of arrival (ToA) or other approaches which have been studied extensively [17]. The net effect of ranging estimations is considered by a statistical model, i.e., dˆij = σij dij + nij , where dij denotes the geographic distance between node-i and node-j and hence dji = dij , the distance estimation at node-j from node-i is denoted by dˆij , the estimation error is introduced by multiplicative and additive random coefficients σij and nij , respectively. In the wireless localization literature, it is usually assumed that • • For ToA-based ranging estimations, σij = 0 and nij is modeled as a Gaussian distributed variable. For RSS-based ranging estimations, the shadow fading is modeled as σij which is assumed to be Gaussian distributed; it is often termed as log-normal shadow fading by taking logarithm on both sides. Considering the ranging estimation error, it is observed thereby that the RSS-based approach is effective with short distances since there is a multiplicative error component; the ToA-based approach applies to a more wide range of distances although it may require a central node to calibrate the clocks of terminals to ensure synchronization. In the ALOHA game, a critical requirement is that every good player (node) knows the total number of users (including faulty nodes) in the system to determine its action; in the considered scenario, this requirement poses a challenge in the presence of faulty nodes, namely faulty nodes can let good nodes believe there are fewer nodes in the system such that good nodes may behave more conservatively and hence benefit faulty nodes. To prevent this, a chorus procedure is proposed which leverages the unique pattern of channel power delay profile (PDP) to estimate the total number of participating nodes. The key observation is that a faulty node cannot forge multi-path components (MPCs) and hence this feature can be utilized, especially with line-of-sight (LoS) transmission environment, to estimate the node quantity. Chorus: In this procedure consisting of T time slots, each node is supposed to randomly select one time slot to receive and, on the other hand, transmit pilot symbols in the remaining T − 1 time slots. An analysis in Section 4 shows that even in the presence of faulty nodes, this procedure is robust. In the receive time slot of, e.g., nodei, it estimates the PDP of receive signal and calculates the number of MPCs. In an LoS environment, this measurement gives an accurate estimation of the number of transmitters in the system; moreover, even a faulty node cannot generate multiple MPCs given its location. This is analogous to let all nodes perform chorus first such that every node can have an estimation of the population based on nodes’ unique timbre. The procedure lasts T time slots since we assume that nodes cannot operate in the full-duplex mode; otherwise the procedure can be shortened to one time slot wherein every node transmit and receive simultaneously. The detailed procedure, as well as the analysis for appropriate T , is given in Section 4. ALOHA game based sortition: Given each node’s estimations about the total number of nodes in the system, we let every node be selfish in the ALOHA game to prevent faulty nodes from gaining advantages; this all-be-selfish methodology is essentially identical with the blockchain technology which allows all miners to compete for the opportunity to register a block. In particular, an ALOHA random access game with selfish users is implemented. It is roughly described as follows. • 3. Overview SENATE consists of three major phases: sortition, senator selection and byzantine agreement. 3.1. Sortition In the sortition process, the objective is to prevent faulty nodes to generate arbitrarily many pseudonyms; note that this does not mean we eliminate the Sybil attack by sortition completely. The key to achieve this is by developing an ALOHA game with selfish users [18] such that no one can cheat based on the Nash equilibrium arguments. • Every node is selfish, in the sense that they all want to transmit as soon as possible in a collision-free time slot. However, after a successful transmission, a good node would stop competing whereas a faulty node might keep on transmitting to launch Sybil attacks. Once a node successfully transmits in a time slot (collision-free), it is selected as the s-th candidate where s ∈ {1, ..., S} denotes the s-th successful transmission in the ALOHA game; then it transmits a pilot signal for ranging estimations immediately afterwards. By this definition, it can be shown (Section 4) that a Nash equilibrium exists based on which every node adopts the same mixed-strategy [18]: transmit in each time slot with a probability p with 0 < p < 1, and no one can benefit by changing the strategy unilaterally. After a quorum of S candidates is reached, the sortition phase terminates with S candidates going into the senator selection in the next phase. Note that a Sybil attack is still possible; a faulty node may occupy several seats among the candidates. 3.2. Senator Selection This phase is dedicated to further removing the pseudonyms generated by faulty nodes, by cross-checking the ranging estimations among nodes in a fully distributed manner. After S candidates are selected, they no longer follow the ALOHA-based random access protocol. Instead, each candidate is assigned a unique time slot to transmit in a frame of S time slots in this phase. Since every candidate has transmitted a pilot signal, every candidate has obtained the distance estimations from other candidates. For candidate-i, its distance estimations are denoted by a vector h iT d̂i = dˆ1i , ..., dˆ(i−1)i , 0, dˆ(i+1)i , ..., dˆSi , i = 1, ..., S. (1) The estimated Euclidean distance matrix (EDM) [19] with squared norm is hence h 2 i 2 D̂ , d̂1 , ..., d̂S . (2) Distance feedback and symmetry verification: Each candidate feeds back its d̂i in its dedicated time slot. Afterwards, every candidate obtains the distance estimations between any pair of candidates (double-directional) in the network. Note that dij = dij , ∀i, j , and thereby every node can remove suspicious distance feedback based on checking the estimated EDM (here we assume the feedback is perfect) dˆ2ij − dˆ2ji < (3) where is a constant related to σij and nij . In this case, as long as (3) does not hold, both dˆij and dˆij are removed since we cannot tell if node-i or node-j is lying. Robust WNC generation: Despite the fact that the symmetry verification can, to some extent depending on distance estimation error, eliminate untruthful distance feedback, a faulty node can still launch what we refer to as a “shout attack”1 . Definition 1 (Shout Attack). A shout attack is that a faulty node pretends to be further away to other nodes, by synchronously adding to the distance estimations to other nodes. In particular, for ToA-based ranging estimations, a faulty node can purposely transmit pilot signals later than supposed, and, accordingly, feed back tampered 1. Likewise, a “whisper attack” can be defined by which the faulty node pretends to be nearer to other nodes. For ease of exposition, we use the shout attack for illustration henceforth. (larger) distance estimations; for RSS-based ranging estimations, a faulty node can purposely amplify its pilot signal power and, accordingly, feed back tampered (larger) distance estimations. By definition, a shout attack cannot be detected by symmetry verification and gives a faulty node arbitrarily many fake geographical locations that are arbitrarily far from its real one. The purpose of a shout attack is hence to create pseudonyms and facilitate the Sybil attack, which causes a severe challenge to SENATE since SENATE uses the location information for Sybil protection. To thwart the shout attack, we introduce the seesaw test based on the following intuition. In the real world with (at most) 3-dimensional space, it is increasingly unlikely that a faulty node, which launches the shout attack, is further away to other nodes proportionally, as the number of nodes grows. This is analogous to placing elastic sticks between each pair of nodes in the system, with the lengths of sticks given by the distance estimations. The circumstance for a faulty node launching shout attack in the 2-dimensional space is illustrated in Figure 2; its related sticks are bent dramatically and hence the elastic force levers it out (screened out by the seesaw test), like being on the lighter side of seesaws. This argument is mathematically formalized in Theorem 2 which states the local error is proportional to the number of good nodes. Good Node Node 1 Faulty Node Node 4 Node 2 Node 3 Figure 2. A seesaw test. A faulty node launching the shout attack would be “levered out”. An iterative WNC calculation with seesaw tests is proposed, the essence of which can be illustrated as follows. In each round of the iteration, every node, e.g., node-i, uses the distance estimation to another node-j to push (resp. pull) node-j if the distance estimation is larger (resp. smaller) than the predicted distance between the two nodes by the current WNC; then node-j moves accordingly. At the end of each round, the node with the largest local error, i.e., its related seesaws are bent the most, is identified as a faulty node and therefore removed; note that a termination criterion (specified later) is added such that the WNC calculation terminates when the system has small prediction error. K -means clustering. After obtaining the coordinates of all candidates, a K -means clustering algorithm [20] is applied to the coordinates such that all candidates are divided into K clusters. Then K representatives, each from one cluster, are selected as senators; this prevents Sybil attacks because pseudonyms of one faulty candidate are likely to fall into the same cluster. 3.3. Byzantine Agreement The K senators run a byzantine agreement protocol to reach consensus. We primarily consider the median validity for consensus, which is defined as follows [21].2 Assume a single consensus value is to be reached upon. Denote by G the sorted array of the initial values of good nodes. Among n nodes, f nodes are faulty and it is assumed that f ≤ t and hence G = [G[0], ..., G[n − f − 1]]. Definition 2 (Median Validity). We call a value x medianvalid, if it holds that n−f n−f −1−t ≤x≤G − 1 + t . G 2 2 (4) Thereby, we adopt the Jack algorithm [21] which ensures the following properties, as long as the number of faulty senators, i.e., f , satisfies K ≥ 3f + 1. • • • Agreement: For every selection of input values and every selection of faulty senators, all good senators can decide on the same value. Termination: Every good senator can decide on a value in finite time. Median Validity: The decision is median-valid. Upon agreement, every senator broadcasts its consensus value, and every good node in the system adopts the majority value. Since consensus is reached in the senate, the majority value can reflect the consensus value and ensures safety and agreement among all nodes. Algorithm 1: Sortition Input: Node-1, ..., N ; Output: Candidate-1, ..., S ; 1 Chorus Procedure 2 Every node (node-i) uniform-randomly selects 1 ≤ ti ≤ T . 3 for t=1:T do 4 if t = ti then 5 Node-i transmits a pilot symbol in the t-th time slot. 6 else 7 Node-i receives signals and estimates the number of transmitting nodes based on the receive PDP; the estimation is denoted by σi . 8 The total number of nodes estimated by node-i is N̂i = 1 + T T−1 σi . 9 10 11 12 13 ALOHA Game based Sortition Every node can transmit in every time slot. for s=1:S do The s-th candidate is selected when a new successful transmission happens and the corresponding node is candidate-s; afterwards, it transmits a pilot signal. return Candidate-1, ..., S . all faulty nodes can conjure is by always transmitting, and thereby, 4. Sortition N̂i ≤ 1 + The sortition phase is described in Algorithm 1. We elaborate on several details as follows. 4.1. Analysis on Chorus Duration When a node, e.g., node-i, receives signals, assuming its estimation on the number of transmitting nodes is correct, the probability of a good node transmitting in this time slot is 1 pj = 1 − , node-j is a good node and j 6= i. (5) T Therefore, the unbiased estimate should be T qi , (6) T −1 where qi denotes the estimation of transmitting node in node-i’s receive time slot. The optimal attack a faulty node can launch is to let good nodes believe there are more nodes in the system such that the transmission strategy of the latter would be more conservative. Therefore, the worst effect N̂i = 1 + 2. Nevertheless, basically any BFT protocol can be plugged in to SENATE in this phase, and works well in an open system since we have achieved Sybil-proof in the previous phases. T T F+ T −1 T −1 N −F X−1 mn , (7) n=1 where mn is a Bernoulli random variable with parameter 1 − T1 . By letting T be sufficiently large compared with N , then T N N̂i ≤ 1 + F + (N − F − 1) + O T −1 T F N =N+ +O ,F ≤N (8) T −1 T where O (·) denotes infinitesimal. In an LTE system with time slot of 0.5 ms and 100 nodes, a chorus procedure lasting 1 s yields N T = 0.05. 4.2. Nash Equilibrium of ALOHA Game An ALOHA game is described as follows. Every node participates in the game. • In a time slot a node successfully transmits (collision-free), the node receives a payoff of 1 − c where c ∈ [0, 1] denotes the one-time transmission cost and leaves the game.3 3. A faulty node may otherwise stay in the game and keep playing. • In a time slot if a collision happens, every transmitting node receives a payoff of −c. A detailed payoff function is described in Table A. Every node’s goal is to maximize its payoff in a single time slot and the game is repeated. Based on the game setting, we can prove the existence of Nash equilibrium. Theorem 1. There exists a Nash equilibrium that every node adopts the same mixed-strategy: transmit with probability p(c, N ) in each time slot, where √ (9) p(c, N ) = 1 − N −1 c. Algorithm 2: Senator Selection Input: Candidate-1, ..., S ; Output: Senator-1,...,K ; validSenate. 1 Distance estimation and feedback 2 for i=1:S do 3 Based on the pilot signals received in the sortition phase, candidate-i estimtes its distance to other candidates and feeds back its distance vector d̂i . 4 EDM symmetry verification T 5 6 Proof: The proof is based on [18]. See Appendix A for details. Remark 1. Theorem 1 indicates that by allowing every node to be selfish, a symmetric equilibria exists, namely even a faulty node cannot improve its payoff by changing its transmission probability unilaterally. Specifically, if it increases its transmission probability, there would be more collisions and the payoff decreases due to cost c; if it decreases its chance then its success chance also decreases. The transmission cost per time slot c clearly plays a critical role here, which denotes the relative cost per transmission as compared with one successful transmission. In practice, we propose that, aside from the power and resource cost due to wireless transmissions, an economic approach can be applied whereby a small fee is charged for every sortition transmission to enhance the robustness of the process. 7 8 9 10 11 12 13 14 15 Σ , \|D̂ − D̂ \|. for Every element Dij in D̂ do if Σij > ∆d then D̂ij ← invalidValue. Robust WNC generation At every terminal, generate the WNC simultaneously based on the same following procedure: terminate ← false; X , [x1 , ..., xS ]T ← 0S×2 ; e ← 0S . while terminate = false do for {i, j} ∈ {1, ..., S} × {1, ..., S} do if Dij 6= invalidValue or 0 then i w ← ei e+e ; j ei ← f← xi ← xi + γf . 16 maxError ← maxi {ei }; errIndex ← arg P maxi {ei }. if maxError > β Sei then Remove candidate-errIndex, and its corresponding entries in X , e and D̂ ; S ← S − 1. else terminate ← true 17 5. Senator Selection 18 In this phase, the S candidates transmit in a round-robin fashion. The detailed algorithm description of the phase is presented in Algorithm 2, and some explanations follow. 19 20 5.1. Robust WNC Generation 21 The rationale for robust WNC generation is as follows. In the face of EDM estimation error introduced by faulty nodes, i.e., denote 22 D̂ = D + E, 25 (10) where the entries of E can be arbitrarily large considering malicious behavior, our goal is to recover D . Towards this end, two structures can be exploited: (a) although the faulty nodes can cause arbitrarily large error, the error is sparse in terms of entries of E , i.e., majority is still good; (b) the EDM stems from space of limited dimensionality4 and hence there are mature tools in distance geometry [19] that can be utilized to verify its authenticity. Thereby, considering 2dimensional space, the EDM can be written as T D = −2XX T + 1diag XX T + diag XX T 1T , (11) 4. We use 2-dimensional space for ease of exposition in this paper. However, the generalization to 3-dimensional is considered straightforward. 23 24 \|kxi −xj k2 −dˆij \| δw + (1 − δw)ei ; dˆij \|kxi −xj k2 −dˆij \| w (xi − xj ); ˆ dij K -means clustering if S < K then validSenate ← false. 26 else S ← Kmeans(X, K); validSenate ← true 27 return S ; validSenate. where X ∈ RS×2 is the geographical location coordinates T of candidates, i.e., X , [x1 , ..., xS ] . We then formulate the WNC generation problem as follows, exploiting the sparse error property. P1: minimize kEk0 , subj. (10), (11) and rank(X) = 2. X,E (12) 0 The ` norm based formulation in P1 is notoriously nonconvex and in fact NP-complete based on compressive sensing theory. Therefore, the `0 norm in P1 is relaxed to `1 norm which often exhibits near optimal performance [22], i.e., P2: minimize kEk1 , subj. (10), (11) and rank(X) = 2. X,E (13) We adopt a data-driven gradient-descend-based method to solve P2. Based on an estimation dˆij , we can update xi (or xj ) based on the gradient of the objective function in P2: xi ← xi + µ ∂ kxi − xj k22 − dˆ2ij ∂xi , (14) which corresponds to the 15-step in Algorithm 2. Also note that in the algorithm, we keep track of the local error array e whose element ei represents the squared distance error related to candidate-i; that is how much candidate-i is levered in the seesaw test (Figure 2). Therefore in the 15-step, we take into account the fact that a candidate with small error should not be updated based on the location of a candidate with large error; the latter is likely to be a faulty node. Based on this argument, we remove the candidate with the largest error at the end of each round, until the error is evenly distributed among candidates which means the error is introduced by ranging estimation instead of faulty nodes. In the case that the selected senators do not reach a quorum of K , Algorithm 2 returns validSenate = false. Intriguingly, this method is similar with the spring network based method where any pair of nodes are connected by a spring in, e.g., [17], [23]; the objective in those works is to minimize the elastic potential energy of the system (equivalent with the total square error (TSE) of distance prediction) given the current lengths of springs (distance estimations) by placing the nodes (the distances among nodes are the rest lengths of the springs) on a plane. Although the objective in P2 is not minimizing the TSE, the presented data-driven gradient-descend-based method turns out to be similar with the Vivaldi algorithm [23], except for the faulty detection. 5.2. Analysis on Seesaw Test The seesaw test is based on the rationale that a faulty node implementing the shout attack can be detected because its resultant location would be out of the 2D space. A question arises accordingly: how out-of-space the faulty node is given a certain strength of its shout attack, and moreover the effect of the number of good nodes. This question is important because its answer can quantitatively characterize the effectiveness of the seesaw test against forged locations. In seeking for a concise and illustrative answer, we consider a simplified scenario where there is one faulty node, without loss of generality located at x0 = (0, 0), who is trying to launch a shout attack to M good nodes located at xm = (xm , ym ), ∀m ∈ 1, ..., M . Concretely, we consider that the faulty node adds an arbitrary (independent with real node-locations) error vector to the entries in the EDM that are related to it; note that this is more general than the shout attack whereby the error is added synchronously. The arbitrary error is written based on (10) as 0 eT E= . (15) e 0M ×M Note that no ranging estimation error is considered in this subsection to focus on the synthetic error by the faulty node. The level of out-of-space of the faulty node is measured by 2 h(ς 2 ) , EX min min 2 X̂ − Z , Z, rank(Z)=2 e, kek1 =M ς 2 (16) where X̂ denotes the reconstructed coordinates of nodes given the tempered EDM in (10) and (15). In other words, the level is quantified by the minimum squared Euclidean distance from the reconstructed coordinate space to its projection into any 2D space, given that the faulty node implements the attack that minimizes this distance. It is essential to note the sequence of minimization, meaning that the faulty node first chooses the error then the closest 2D space is selected. Since this quantity is affected by the locations of good nodes by noting that closer good nodes produce stronger lever force in the seesaw test given the same strength of shout attack, the expectation in (16) is taken over a given location distribution. In the following theorem, we adopt the 2D Gaussian distribution for ease of exposition. Theorem 2. Assume that the faulty node is at (0, 0) with the attack strength of kek1 = M ς 2 , and the good nodes’ coordinates are i.i.d. generated based on a Gaussian distribution with zero mean and variance of σ 2 , i.e., (xm , ym ) ∼ (0, σ 2 I), ∀m ∈ 1, ..., M . When the error e is independent with (xm , ym ), ∀m ∈ 1, ..., M , then h(ς 2 ) = min (M − 1)σ 2 , (M − 2)ς 2 . (17) Proof: See Appendix B. Remark 2. It is shown that a faulty node cannot conceal its lie by noting that h(ς 2 ) scales with the attack strength ς 2 , until ς 2 is comparable with the squared distance measurement (σ 2 ) whereby the attack becomes quite obvious. In addition, the effect is amplified by approximately M times; this is intuitive since it becomes increasingly more difficult to lie to more good nodes when they form a concrete 2D space. Another note is that h(ς 2 ) > 0 as long as M ≥ 3, because at least 3 nodes can determine a 2D space. Remark 3. The theorem assumes that the error matrix E is independent with the coordinates of good nodes, which requires that a faulty node is not aware of the coordinates of other nodes (assumed mostly good nodes) in the sortition phase; this is reasonable because the coordinate information is not accessible in the sortition phase before any distance feedback occurs. Corollary 1. In L-dimensional space, min (M − L + 1)σ 2 , (M − L)ς 2 . h(ς 2 ) = 1 Byzantine consensus w/o Sybil attack SENATE 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 20 30 40 50 60 70 80 90 100 Figure 3. The probability a valid consensus is reached with 100 nodes. Remark 4. This corollary generates the effectiveness of the seesaw test to higher dimensions, e.g., 3D scenarios with applications for, e.g., drone swarms. 6. Byzantine Agreement In this phase, since we have removed Sybil nodes from the selected senators to a great extent, basically any byzantine agreement protocol can be implemented among the K senators. In particular, we adopt the Jack scheme proposed in [21] which consists of the following two major stages: • • Setup stage: Each senator broadcasts its initial value and receives other senators’ initial values. Thereby, each senator broadcasts its acceptable values and a proposed value, jointly considering its and other senators’ initial values. Search stage: Rotating among (t + 1) predetermined leaders, in each round a leader receives proposals from other senators and accordingly proposes a value based on its acceptable values; if an agreement is reached based on proposals, the leader would propose the agreed value. It is proved that as long as one leader among the (t + 1) leaders is a good node, a valid agreement would be reached. Therefore, at most t faulty nodes are allowed in this phase. The validity is assured by the median validity specified in Section 2. The termination and agreement properties are also proved in [21]. In fact, this protocol achieves the optimum safety and 2-approximate of the optimum median validity. 7. Simulations We run a computer simulation to test the performance of SENATE. The ranging estimations at nodes are assumed to be perfect, and the information exchange is simulated by direct modifications of data arrays for nodes. In Figure 3, where there are N = 100 nodes and SENATE selects S = 50 candidates and K = 7 senators, the number of faulty nodes is shown by x-axis and the performance of consensus probability is shown which is obtained by running the algorithms for 1000 episodes. The node locations are randomly generated in a square with a side length of 200 meter in each episode. The faulty nodes are assumed to always launch Sybil attacks and propose randomly generated values which are deviating from the true values; specifically, the good and faulty nodes’ initial values are uniformly randomly generated from interval [−1, 1] and [99, 101], respectively. In the figure, we also simulate the Jack algorithm [21] without faulty nodes launching Sybil attacks for comparisons; the Jack algorithm under this scenario and attack assumptions can ensure consensus when the number of faulty nodes does not reach majority (50%), and cannot otherwise; this is direct consequences of the design of the Jack algorithm and the definition of median validity. It is observed that SENATE perform close-to the conventional BFT protocol as if there is no Sybil attack; this verifies that SENATE is Sybil-proof. We also observe that SENATE perform better when faulty nodes reach majority, which is because the SENATE randomly selects the senators such that there is a probability that the selected senators are dominated by good nodes; the Jack algorithm can also perform, at least, as well as SENATE if a sortition phase is added, but the effect is not shown in the figure to be in line with the original Jack algorithm. 8. Conclusion SENATE is a real-time distributed BFT protocol which is applicable to fully-connected wireless-networked systems without prior identity authentication. In order to prevent malicious nodes to generate an arbitrary number of pseudonyms, SENATE leverages the wireless signals transmitted by nodes to cross-verify their identities in a fully-decentralized (no-trust) manner, based on the fact that pseudonyms are likely to be adjacent geographically. Thereby, only selected nodes, i.e., senators, participate in the final consensus reaching process. Computer simulations show that SENATE is Sybil-proof, by comparing the consensus probabilities between systems wherein faulty nodes launch and do not launch Sybil attacks, respectively. Acknowledgement This work is sponsored in part by the Nature Science Foundation of China (No. 61701275, No. 91638204, No. 61571265, No. 61621091), the China Postdoctoral Science Foundation, and Hitachi Ltd. Appendix A. Proof of Theorem 1 First, we invoke the following lemma which ensures the existence of a Nash equilibrium. Lemma 1 ( [24]). A finite symmetric game has a symmetric mixed-strategy equilibrium. The finite symmetric game in Lemma 1 denotes a game wherein every player has the same finite action set, and the payoff received by each player given the same action (and other players’ actions) is identical, irrespective with the specific player. A mixed strategy is in contrast with a pure strategy; the latter employs a fixed action each time whereas the former can be viewed as a mix (randomized strategy) of the latter. The proof of Lemma 1 is based on the Brouwer fixed point theorem and is omitted for brevity. Provided the existence of symmetric equilibria, we are ready to derive the transmission probability p for every node. Due to symmetry, it suffices to consider a single node, whose game choice is shown in Table A. TABLE 2. PAYOFF FUNCTIONS All other nodes are silent (p1 ) 1−c 0 Transmit Silent Some other node transmits (p2 ) −c 0 T T rank(X̂ X̂ ) ≤ rank(X̄ X̄ ) + rank(1eT ) + rank(e1T ) = 4. (21) In addition, the eigenspace is spanned by T eigenspace(X̂ X̂ ) = span{x, y, 1, e}, (22) where x , [x1 , ..., xK ]T and y , [y1 , ..., yK ]T ; this can be seen from the following equation. 1 1eT + e1T 2 I 0 = [x, y, 1, e] [x, y, e, 1]T . 0 12 I T X̄ X̄ + (23) Expected payoff (1 − c)p1 − cp2 0 N −1 p1 = Pr{success} = (1 − p) ; p2 = Pr{collision} = 1 − (1 − p)N −1 . Note that we assume E is independent with X̄ , and that the power of the error is ς 2 , i.e., 1eT + e1T 2 The probability of each event is (18) Based on the principle of indifference [25], both expected payoffs should be zero, which yields (9) and concludes the proof. Appendix B. Proof of Theorem 2 Based on (11), the EDM without error can be written as T D = −2XX T + 1diag XX T + diag XX T 1T h i h iT T = −2XX T + 1 0, β T + 0, β 1T , (19) 0 0 2 T = where β 0 , d , ..., d2M 0 , and X T 10 0 x1 · · · xM T T , [0, X̄ ] . The attack is 0 y1 · · · yM implemented on EDM and we can derive the resultant coordinate covariance, which is tampered by the attack, based on [19]: T T First, we notice that the eigenspace of X̂ X̂ in (20) is at most 4-dimensional (the other eigenvalues are zeros) by the rank inequality X̂ X̂ 1 T T T T T =− D + E − 1[0, β T + e ] − [0, β + e ] 1 0 0 2 " # " #! T 0 0T 0 0 1 T + = XX + 2 0 1eT 0 e1T " # 0 0T = (20) . T 0 X̄ X̄ + 12 1eT + e1T = ς 2 M. (24) ∗ Since the error is assumed to be independent with the coordinates and that i.i.d. Gaussian coordinate vectors are uniformly directed in space, a constant share of the error power is leaked out of the coordinate eigenspace spanned by [x, y]. Considering the objective of the faulty node is to minimize the leakage power beyond any 2D space, which in this case is equivalent to maximizing the power in the 2D space given that the total power is fixed, the best attack the faulty node can implement is to concentrate its error power to the same linear space spanned by 1, i.e., eopt = ς 2 1. (25) Note that the solution e = −eopt also satisfies the conditions, T however, it results in a negative eigenvalue of X̂ X̂ . This violates with [26, Theorem 2] which proves an important T property of EDM that if X̂ X̂ is not positive-semi-definite, then there exists at least three distance measurements violating the triangle inequality. In other words, a shout attack may put the faulty node in a higher-dimensional space whereas a whisper attack would lead to violation of the triangle inequality which is much easier to spot. Let us consider the minimization inside the expectation in (16), this is an optimal low-rank approximation problem whose solution is well known to be the dominant 2dimensional singular space of X̂ , i.e., the eigenspace of T X̂ X̂ . Denoting the singular value decomposition (SVD) T of X̂ as X̂ = Û Σ̂V̂ (singular values are always arranged in non-increasing order), then T Z opt = Û 2 Σ̂2 V̂ 2 , (26) wherein Û 2 and V̂ 2 denote the first two columns of Û and Û , respectively, and Σ̂2 contains the two dominant singular values. It follows that the minimum projection error is 2 min Z∈RM ×M , rank(Z)=2 X̂ − Z 2 2 = X̂ − Z opt = 2 M X Σ2i,i , i=3 (27) which is the coordinate power leakage beyond 2D space due to tampered EDM. To solve for this quantity, we adopt the Gram–Schmidt orthogonalization process on the set {x, y, \|ς\|1} (given the optimal attack derived in (25)). Based on symmetry that x and y are both i.i.d. Gaussian distributed, it is clear that the direction of the third vector is√not relevant and hence it is replaced by w where w , ς 2 M [1, ..., 0]T with the same power. For brevity, the detailed process is omitted as the orthogonal basis vectors and the expected leakage power h(ς 2 ) are given as below: u1 = x; uT y u2 = y − 1 2 u1 ; ku1 k2 uT w uT w u3 = w − 1 2 u1 − 2 2 u2 ; ku1 k2 ku2 k2 h(ς 2 ) = min {E[uT u ]} i i i={1,2,3} = min {M σ 2 , (M − 1)σ 2 , (M − 2)ς 2 }, (28) i={1,2,3} [9] R. Guerraoui, N. Knežević, V. Quéma, and M. Vukolić, “The next 700 BFT protocols,” in Proceedings of the 5th European Conference on Computer Systems, ser. EuroSys ’10. New York, NY, USA: ACM, 2010, pp. 363–376. [10] M. Vukolić, “The quest for scalable blockchain fabric: Proof-ofwork vs. BFT replication,” in Open Problems in Network Security, J. Camenisch and D. Kesdoğan, Eds. Cham: Springer International Publishing, 2016, pp. 112–125. [11] Online; accessed 20-Mar-2018; https://neo.org/. [12] T. Dasu, Y. Kanza, and D. Srivastava, “Unchain your blockchain,” in Proc. Symposium on Foundations and Applications of Blockchain, vol. 1, 2018, pp. 16–23. [13] J. Newsome, E. Shi, D. Song, and A. Perrig, “The Sybil attack in sensor networks: Analysis & defenses,” in Proceedings of the 3rd International Symposium on Information Processing in Sensor Networks, ser. IPSN ’04. New York, NY, USA: ACM, 2004, pp. 259–268. [14] D. B. Faria and D. R. Cheriton, “Detecting identity-based attacks in wireless networks using signalprints,” in Proceedings of the 5th ACM Workshop on Wireless Security. New York, NY, USA: ACM, 2006, pp. 43–52. [15] M. Demirbas and Y. Song, “An RSSI-based scheme for sybil attack detection in wireless sensor networks,” in Proceedings of the International Symposium on on World of Wireless, Mobile and Multimedia Networks, Washington, DC, USA, 2006, pp. 564–570. [16] J. R. Douceur, “The Sybil attack,” in International Workshop on Peerto-Peer Systems. Springer, 2002, pp. 251–260. [17] N. Patwari, J. N. Ash, S. Kyperountas, A. O. Hero, R. L. Moses, and N. S. Correal, “Locating the nodes: cooperative localization in wireless sensor networks,” IEEE Signal Process. Mag., vol. 22, no. 4, pp. 54–69, Jul 2005. and the conclusion follows immediately. [18] A. B. MacKenzie and S. B. Wicker, “Selfish users in Aloha: a gametheoretic approach,” in IEEE Veh. Tech. Conf., vol. 3, 2001, pp. 1354– 1357. References [19] I. Dokmanic, R. Parhizkar, J. Ranieri, and M. Vetterli, “Euclidean distance matrices: Essential theory, algorithms, and applications,” IEEE Signal Process. Mag., vol. 32, no. 6, pp. 12–30, Nov 2015. [1] S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system,” 2008. [2] F. Tschorsch and B. Scheuermann, “Bitcoin and beyond: A technical survey on decentralized digital currencies,” IEEE Communications Surveys Tutorials, vol. 18, no. 3, pp. 2084–2123, thirdquarter 2016. [20] J. MacQueen et al., “Some methods for classification and analysis of multivariate observations,” in Proc. the fifth Berkeley symposium on mathematical statistics and probability, vol. 1, no. 14. Oakland, CA, USA., 1967, pp. 281–297. [3] I. Bentov, C. Lee, A. Mizrahi, and M. Rosenfeld, “Proof of activity: Extending Bitcoin’s proof of work via proof of stake,” SIGMETRICS Performance Evaluation Review, vol. 42, no. 3, pp. 34–37, 2014. [4] Y. Gilad, R. Hemo, S. Micali, G. Vlachos, and N. Zeldovich, “Algorand: Scaling byzantine agreements for cryptocurrencies,” in Proceedings of the 26th Symposium on Operating Systems Principles. New York, NY, USA: ACM, 2017, pp. 51–68. [5] I. Eyal, A. E. Gencer, E. G. Sirer, and R. V. Renesse, “BitcoinNG: A scalable blockchain protocol,” in 13th USENIX Symposium on Networked Systems Design and Implementation (NSDI 16). Santa Clara, CA: USENIX Association, 2016, pp. 45–59. [21] D. Stolz and R. Wattenhofer, “Byzantine agreement with median validity,” in International Conference on Principles of Distributed Systems (OPODIS), 2016, pp. 1–14. [22] E. J. Candès et al., “Compressive sampling,” in Proc. international congress of mathematicians, vol. 3. Madrid, Spain, 2006, pp. 1433– 1452. [23] F. Dabek, R. Cox, F. Kaashoek, and R. Morris, “Vivaldi: A decentralized network coordinate system,” SIGCOMM Comput. Commun. Rev., vol. 34, no. 4, pp. 15–26, Aug. 2004. [24] J. F. Nash, “Equilibrium points in n-person games,” Proceedings of the national academy of sciences, vol. 36, no. 1, pp. 48–49, 1950. [6] C. Cachin, “Architecture of the Hyperledger blockchain fabric,” in Workshop on Distributed Cryptocurrencies and Consensus Ledgers, 2016. [25] S. fen Cheng, D. M. Reeves, Y. Vorobeychik, and M. P. Wellman, “Notes on equilibria in symmetric games,” in International Workshop On Game Theoretic And Decision Theoretic Agents (GTDT), 2004, pp. 71–78. [7] M. Castro and B. Liskov, “Practical byzantine fault tolerance,” in OSDI. Berkeley, CA, USA: USENIX Association, 1999, pp. 173– 186. [26] J. C. Gower, “Euclidean distance geometry,” Math. Sci, vol. 7, no. 1, pp. 1–14, 1982. [8] D. Ongaro and J. Ousterhout, “In search of an understandable consensus algorithm,” in 2014 USENIX Annual Technical Conference (USENIX ATC 14). Philadelphia, PA: USENIX Association, 2014, pp. 305–319.	7cs.IT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 1 Estimation and Mitigation of Channel Non-Reciprocity in Massive MIMO arXiv:1706.04626v1 [cs.IT] 14 Jun 2017 Orod Raeesi, Student Member, IEEE, Ahmet Gokceoglu, Member, IEEE, and Mikko Valkama, Senior Member, IEEE Abstract—Time-division duplex (TDD) based massive MIMO systems rely on the reciprocity of the wireless propagation channels when calculating the downlink precoders based on uplink pilots. However, the effective uplink and downlink channels incorporating the analog radio front-ends of the base station (BS) and user equipments (UEs) exhibit non-reciprocity due to non-identical behavior of the individual transmit and receive chains. When downlink precoder is not aware of such channel non-reciprocity (NRC), system performance can be significantly degraded due to NRC induced interference terms. In this work, we consider a general TDD-based massive MIMO system where frequency-response mismatches at both the BS and UEs, as well as the mutual coupling mismatch at the BS large-array system all coexist and induce channel NRC. Based on the NRC-impaired signal models, we first propose a novel iterative estimation method for acquiring both the BS and UE side NRC matrices and then also propose a novel NRC-aware downlink precoder design which utilizes the obtained estimates. Furthermore, an efficient pilot signaling scheme between the BS and UEs is introduced in order to facilitate executing the proposed estimation method and the NRC-aware precoding technique in practical systems. Comprehensive numerical results indicate substantially improved spectral efficiency performance when the proposed NRC estimation and NRC-aware precoding methods are adopted, compared to the existing state-of-the-art methods. Index Terms—Beamforming, channel non-reciprocity, channel state information, frequency-response mismatch, linear precoding, massive MIMO, mutual coupling, time division duplexing (TDD). I. I NTRODUCTION M ASSIVE MIMO is one of the key potential technologies for upcoming 5G systems [1] where base stations (BSs) deploy very large antenna arrays, e.g., several tens or hundreds of antenna units per array, to facilitate high beamforming and spatial multiplexing gains. In such systems, it is not feasible to transmit downlink pilots from each BS antenna in order to estimate the corresponding spatial channels at user equipments (UEs) and feedback the channel state information (CSI) to BS, as the amount of overhead in such approach is proportional to the number of antennas in the BS side [2]. Massive MIMO O. Raeesi, A. Gokceoglu, and M. Valkama are with the Department of Electronics and Communications Engineering, Tampere University of Technology, Tampere 33720, Finland (e-mail: orod.raeesi@tut.fi; ahmet.gokceoglu@tut.fi; mikko.e.valkama@tut.fi). This work was supported by the Finnish Funding Agency for Technology and Innovation (Tekes) under the project “5th Evolution Take of Wireless Communication Networks (TAKE-5)”, by the Academy of Finland under the projects 284694 and 288670 and TUT Graduate School. This work has been submitted to IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. systems are thus envisioned to primarily deploy time-division duplex (TDD) based radio access and rely on the reciprocity of the physical uplink and downlink channels when obtaining CSI at BS. This, in turn, requires substantially smaller pilot or reference signal overhead being only proportional to the number of UEs [3]. While it is a common assumption in TDD systems that the physical propagation channels are reciprocal within a coherence interval [2], [3], the impacts of the BS and UE side transceiver analog front-ends on the effective downlink and uplink channels are not reciprocal. This hardware induced phenomenon is often referred to as the channel non-reciprocity (NRC) problem [4], [5]. Typically, the mismatches in the frequency-responses (FRs) of both the BS and UE side radio front-ends at transmit and receive modes are seen as the main cause of NRC. Another source of NRC considered in literature is the differences in mutual coupling (MC) of BS antenna units and associated RF transceivers under transmit and receive modes [6], [7]. The impacts of the NRC on the achievable system performance have been studied in various works in the recent literature. To this end, [5] provides downlink sum-rate analysis for a general multi-user MIMO system with zero-forcing (ZF) and eigen-beamforming types of precoding under NRC due to FR mismatch. Then, specifically focusing on massive MIMO systems, [8], [9] study achievable downlink sum-rates for maximum-ratio transmission (MRT) and ZF precoding schemes, demonstrating significant performance degradation under practical values of the NRC parameters. There is also a large amount of work reported in the literature addressing the estimation and mitigation of NRC in TDD based MIMO systems [2], [4], [6], [10]–[16]. These studies can be divided into three main categories as follows: i) BS carries out “self-calibration” using a reference antenna with the help of additional circuitry [4], [6]. This method is capable of estimating BS side NRC only. ii) BS carries out “self-calibration” without additional circuitry. Mutual coupling between antennas is utilized when exchanging pilot signals with the reference antenna [2], [10]–[13]. Similar to i), also this method estimates only the BS side NRC, and also commonly neglects the mutual coupling mismatch. iii) BS transmits specific pilot signals to UEs and UEs send back the received signals in certain properly precoded forms to facilitate BS side NRC parameter estimation, which is often referred to as over-the-air (OTA) approach [6], [14]–[16]. 2 In this work, we focus on OTA-based estimation and mitigation of NRC in a multi-user massive MIMO system context deploying MRT or ZF precoding. The novelty and contributions of this paper can be summarized as follows: IEEE TRANSACTIONS ON SIGNAL PROCESSING element-wise and return real and imaginary parts of complexvalued arguments, respectively. The element in the i’th row and j’th column of matrix V is represented by vij , whereas the i’th element on the main diagonal of a diagonal matrix C is shown by ci . The complex-valued zero-mean circularly symmetric Gaussian distribution with variance σ 2 is denoted 2 as CN 0, σ . Finally, In and 0n denote the n × n identity and all-zero matrices, respectively. 1) We consider generalized NRC induced by coexisting FR mismatches of all associated radio transceivers at UE and BS sides as well as the mutual coupling mismatches in the BS side large-array antenna system, unlike many of the earlier works that consider only FR mismatch such as II. S YSTEM M ODEL AND P ROBLEM F ORMULATION [2], [4], [10]–[16]. In this respect, only [6] reports similar We consider a TDD based single-cell multi-user downlink modeling, however, the proposed mitigation scheme in transmission scenario where the BS with a large number of [6] is suitable mainly for small scale MIMO systems, e.g., antenna units, denoted by N , transmits to K single-antenna UEs 2-4 BS antennas. simultaneously in the same time-frequency resource, where 2) We address the estimation and mitigation of the NRC N K. All signal and system models are written for an sources of both the UE and BS sides, unlike many other arbitrary subcarrier of the underlying orthogonal frequency works that address only BS side NRC, e.g., [2], [4], [10]– division multiplexing/multiple access (OFDM/OFDMA) wave[13], [15]. As shown in [17], with popular assumption of form, that is, before IFFT and after FFT on the TX and RX not having downlink demodulation pilots, UE side NRC sides, respectively. can be a major cause of performance degradation in multiIn an ideal TDD massive MIMO system, the effective uplink user massive MIMO systems, thus strongly motivating and downlink channels consist of only the reciprocal physical to incorporate such effects in the NRC estimation and channels. Building on that, the downlink transmission is done mitigation processes. by beamforming the multi-user downlink data based on the 3) Unlike other massive MIMO NRC mitigation works [2], estimated channels from uplink pilot sequences of length τu [10]–[13] which all assume the availability of downlink symbols [2], [3]. In this work, we assume the same procedure pilots in the UE side, we consider the appealing massive for the downlink transmission, however, we consider more MIMO scenario in which there are no downlink pilots generalized uplink and downlink effective channel models and thus UEs rely on the statistical properties of the which are non-reciprocal due to radio front-end mismatches beamformed channels to decode the received downlink and non-idealities. In this respect, the uplink model for channel signals [8], [18]–[21]. estimation phase [18] and the corresponding downlink received 4) We demonstrate the performance of our proposed scheme signal model in beamformed data transmission phase under under imperfect uplink CSI, unlike other works which the non-reciprocal effective channels can be expressed as commonly rely on the perfect uplink CSI assumption [4], √ [11]–[13], [15]. Uplink Training : Y = ρu τu G + Zu (1) √ The organization of the paper is as follows. Fundamental Downlink Transmission : r = ρd Hx + zd , signal models of the considered massive MIMO system with where x denotes the precoded user data, whereas G and H MRT and ZF-based precoding schemes under NRC are first are the effective non-reciprocal uplink and downlink multipresented in Section II. Then, the NRC-aware downlink user MIMO channels, respectively, which will be elaborated in precoding approach is formulated for given NRC estimates. In detail later in Section II-A. Zu is the processed noise matrix at Section III, novel pilot signaling method between the BS and the BS, while zd denotes the UE side multi-user thermal noise UEs is introduced which is followed by the proposed novel vector, both assumed to consist of i.i.d. CN (0, 1) elements. iterative estimation of BS and UE side NRC matrices. The The average signal to noise ratios (SNRs) in the uplink and results of empirical performance evaluations in terms of the downlink are denoted as ρu and ρd , respectively. This basic achievable system spectral efficiency are presented in Section system framework is largely based on and following the seminal IV, incorporating the proposed estimation-mitigation scheme towork by Marzetta in [18], [22] where reciprocal channels were gether with existing state-of-the-art NRC estimation/mitigation assumed. methods for reference. Finally, conclusions are drawn in Section V. A. Effective and Relative Uplink and Downlink Channels Notations: Throughout the paper, vectors and matrices are As illustrated in Fig. 1, the complete description of the denoted with lower and upper case bold letters, respectively, ∗ T H uplink and downlink effective channels appearing in (1) can e.g., vector x, matrix Y. The superscripts (.) , (.) , (.) , and † be expressed as (.) indicate complex-conjugation, transposition, Hermitian) transpose, and Moore-Penrose pseudo inverse operations, G = Er PFt ⇒ H 6= GT , (2) respectively. Expectation operator is shown by E[.], while H = Fr PT Et Tr (.) represents the trace operator. diag (.) operator transforms a vector v to a diagonal matrix with the elements of v at with Er = Lr Mr and Et = Mt Lt . In above, F is the joint its diagonal, and vice versa, reads the diagonal elements of frequency-response matrix of the UEs, L is the frequencythe input matrix into a column vector. <{.} and ={.} work response matrix of the BS, M is the mutual coupling matrix MITIGATION OF CHANNEL NON-RECIPROCITY IN MASSIVE MIMO SYSTEMS UE #1 f1TX f 2TX l1RX RX m21 mNRX1 UE #2 Multi-user MIMO Propagation Channels P UE #K f KTX B = IN + B0 , where B0 accounts for the deviation of diagonal and off-diagonal entries from the ideal reciprocal response. The detailed modeling of the entries of the above matrices is based on the practical NRC modeling introduced in [6], in which σF2 is denoting the variance of diagonal elements in Ft and Fr , while the corresponding variance of diagonal elements in Lt and Lr is denoted by σL2 . The power of elements in Mt and Mr is controlled by input reflection coefficients which 2 have the variance σM . The characterization as given in (2) and/or (3) is generally referred to in literature as channel non-reciprocity [5], [6]. The ideal reciprocal channel model is a special case where A = B = I, i.e., A0 = B0 = 0. Base Station m11RX Mutual Coupling l2RX m12RX mNRX2 RX m22 m1RNX m2RNX 3 lNRX RX mNN (a) Base Station l1TX l2TX lNTX B. Channel Estimation and Beamforming under NRC Mutual Coupling m11TX UE #1 f1RX TX m21 mTNX1 m12TX mTNX2 UE #2 Multi-user MIMO Propagation Channels TX m22 f 2RX PT m1TNX m2TXN First, we shortly address the influence of NRC when the downlink transmission is carried out without any processing against the NRC, i.e., NRC-blind precoding is adopted. In this respect, the required downlink channel estimate in BS is obtained from the orthogonal uplink training signals, with the observation model given already on the first line of (1), complemented, e.g., with LMMSE channel estimator as described in [18], [22]. This yields formally UE #K Ĥ = ĜT , f KRX TX NN m (b) Fig. 1. Basic system models for (a) uplink and (b) downlink transmission and reception including physical propagation channels, transceiver frequency responses and antenna mutual coupling in the devices. (4) where Ĥ and Ĝ are the estimated downlink and uplink effective channels, respectively. Using the estimated downlink effective channel in (4), T the user data vector s = [s1 , ..., sK ] ∈ CK which is assumedh to have element-wise power normalization of the i 2 form E \|sk \| = 1, is precoded as of the BS, P is the reciprocal physical channel, while the x = Us, (5) subscripts t and r denote the transmitting and receiving modes, respectively. Note that the frequency-response matrices, F and where the NRC-blind linear precoding matrix U reads [22]  L, are diagonal, while the mutual coupling matrix M in general β ĤH , for MRT has both non-zero diagonal and off-diagonal entries. −1 U= (6) In general, the effective channels with above assumptions β ĤH ĤĤH , for ZF. and modeling are clearly non-reciprocal, i.e., H 6= GT , due to differences in the TX and RX modes of the radio front-end In above, without loss of generality, the scalar β can be chosen and array responses, i.e., Ft 6= Fr , Lt 6= Lr and Mt 6= MT r . to satisfy unit average transmit power constraint as [18] Hence, the effective uplink and downlink channels can be q −1 described relative to each other as β= E[Tr (UH U)] . (7) H = AGT B, (3) C. Received Signal at UE under NRC −1 T −1 where, A = Fr F−1 and B = L M M L . t t t r r The multi-user received downlink signal vector is given by In general, A is a diagonal matrix and the k’th diagonal the second line of (1). Plugging the precoded symbol vector entry, denoted as ak , corresponds to the frequency-response expression in (5) into (1), the received signal for k’th user ratio of k’th UE at TX and RX modes. In the following, corresponding to the k’th element of r can be written as similar to [5], [6], [8], we will use the decomposition of the K X form A = IK + A0 , where the diagonal matrix A0 measures √ √ rk = ρd βhT u s + ρ β hT (8) d k k k k ui si + zd,k , the deviation from unity frequency-response ratio. The k’th i=1,i6 = k diagonal entry of A0 is denoted as a0k , such that ak = 1 + a0k . In (3), B is a full matrix that incorporates both the frequency- where uk and hT k denote the k’th column and row vectors of the responses and mutual coupling at the BS side. In the following, precoder and effective downlink channel matrices, respectively. for notational convenience, we will use the decomposition Notice that by denoting the k’th column of the uplink effective 4 IEEE TRANSACTIONS ON SIGNAL PROCESSING channel matrix as gk , the effective downlink channel towards the k’th user can be expressed as T hT k = ak gk B. (9) In general, conventional MIMO systems employ downlink pilots to acquire downlink CSI for detection purposes. However, in massive MIMO systems, as shown in [8], [18]–[21], it is generally assumed that UEs employ only the statistical properties of the beamformed channel, namely E βhT k uk , as the downlink CSI to decode the received signal. This assumption is justified by the law of large numbers which T implies that hT u → E h u k k k k , commonly known as the channel hardening concept [22], [23]. Utilizing such approach in acquiring downlink CSI in UEs eliminates the need for sending downlink pilots which directly reduces downlink overhead. Building on this and plugging (9) into (8), the received signal under NRC can be re-written in a general form as √ (10) rk = ρd E βhT k uk sk + zSI,k + zIUI,k + zd,k , where the self-interference (SI), zSI,k , and inter-userinterference (IUI), zIUI,k , are given by √ zSI,k = ρd β ak gkT Buk − E hT sk k uk zIUI,k = √ ρd β K X ak gkT Bui si . (11) i=1,i6=k Based on (11), it can be clearly observed that the NRCblind precoder u which is constructed based on the estimated uplink effective channel Ĝ, through Ĥ = ĜT , cannot take into account the NRC effects from ak and B, which results into increased interference levels and thus reduced downlink spectral efficiency. This is illustrated through an elementary system spectral efficiency evaluation in Fig. 2, with the detailed evaluation assumptions being described in Section IV. It can be noticed that in particular in the ZF precoder case, NRC-blind precoding results into substantial performance degradation, hence strongly motivating to develop efficient NRC estimation and mitigation techniques. D. NRC-Aware Downlink Precoding Principle As shown in Section II-C above, if MRT and/or ZF precoders are applied naively without accounting for NRC, there are additional SI and IUI terms that can substantially degrade the quality of the received signal at the UE side. Here, we introduce a novel NRC mitigation approach, called NRC-aware precoding, which seeks to cancel out the effects of NRC by properly modifying the precoder. Assuming that the BS has already estimates of the NRC matrices A and B, denoted by Â and B̂, the NRC-aware precoding approach transforms the basic linear precoders given in (6) as Unrc = B̂−1 UÂ−1 . (12) Note that, in the special case where the NRC estimation method is capable of estimating the BS side NRC only, (12) reduces to Unrc = B̂−1 U. Fig. 2. Spectral efficiency vs. downlink SNR (ρd ) for N = 100, K = 20, τu = K, ρu = 0 dB, T = 250. The system spectral efficiency performance with the NRCaware precoder, assuming ideal NRC estimates, is shown in Fig. 2. As can be observed, the NRC-aware precoder achieves the ideal system performance, i.e., the performance with fully reciprocal channels. The evaluation setup and details of spectral efficiency calculations will be described in Section IV. III. P ROPOSED E STIMATION OF NRC M ATRICES The NRC mitigation method, i.e., the NRC-aware precoder described in Section II-D requires the knowledge of the matrices A and B at the BS. The information about these matrices is not readily available, hence calling for efficient estimation approaches. Thus, in this section, we will propose a novel iterative OTA estimation framework for acquiring accurate estimates of A and B, based on a novel pilot signaling concept between the BS and UEs. 2 In general, the NRC variances σF2 , σL2 , σM , and the corresponding realizations of the elements of A and B depend on hardware characteristics and operating conditions, e.g., temperature, which vary slowly in time. Thus, the NRC characteristics and the corresponding realizations of A and B can be assumed to stay constant over many propagation channel coherence intervals [14]. Therefore, it is sufficient to perform the NRC estimation very infrequently, e.g, once in every 10 minutes or once a day [2], [10], which makes the estimation overhead negligible, when compared to signaling and pilot overhead that commonly rises from channel estimation procedures. A. Proposed Pilot Signaling In order to estimate the matrices A and B, we propose the following round-trip pilot signaling approach: 1) BS transmits an N × N orthonormal pilot matrix Xnrc . 2) Upon reception, without decoding, UEs send back the conjugated versions of the received signals. Based on the above scheme, the received multi-user signal matrix at UE side can be written as p Rnrc = ρ̃d HXnrc + Z̃d , (13) MITIGATION OF CHANNEL NON-RECIPROCITY IN MASSIVE MIMO SYSTEMS 5 CSI and NRC estimation ( test ) CSI acquisition Uplink pilots ttra test NRC acquisition Downlink pilots Uplink pilots Downlink Uplink Downlink Uplink Fig. 3. Assumed radio frame or sub-frame structure incorporating CSI and NRC estimation as well as actual data transmission phases. where ρ̃d is the downlink SNR and Z̃d is the K × N multiuser receiver noise matrix with i.i.d. CN (0, 1) entries. The tilde sign is used in above and what follows to distinguish these variables between the actual data transmission and pilot signaling phases. Then, the corresponding received signal at BS with the UEs sending back the conjugated version of (13) reads p Ynrc = ρ̃u GR∗nrc + Z̃u p p (14) = ρ̃u ρ̃d GH∗ X∗nrc + Z̃tot , processed received signal in (15) and the corresponding UL channel estimate are available at multiple parallel sub-carriers in an OFDM/OFDMA based radio system. Hence, the above iterative estimation scheme can be carried out in a per subcarrier manner as well. Furthermore, as mentioned in [6], transceivers’ baseband-to-baseband behavior can be modeled by allpass-like transfer functions, therefore it is reasonable to assume that the NRC matrices A and B are largely similar over a set of adjacent subcarriers Csc where typically Csc ≤ 10, whereas G is subject to variations depending on the frequency selectivity of the propagation channels. Based on these assumptions, the estimates Â and B̂ can be obtained by averaging the per subcarrier estimates over Csc neighboring subcarriers, i.e., where ρ̃u is the uplink SNR and Z̃u is the N × N BS receiver noise matrix with i.i.d. CN (0, 1) entries. The√total effective noise matrix seen at BS is denoted as Z̃tot = ρ̃u GZ̃∗d + Z̃u . In above, the duration of the described overall NRC-related Csc 1 X pilot signaling is 2N symbols where the uplink and downlink Â = Âl channels are assumed to be fixed. The coherence time of the Csc l=1 (16) physical channels is typically in the order of several hundred Csc X 1 symbol intervals, determined mostly by the mobility of the B̂ = B̂l , Csc UEs and the system center-frequency. Hence, we assume l=1 a scenario where the coherence time is at least 2N + K where l denotes the subcarrier index. Next we will present the symbols, taking into account both NRC-related pilot signaling actual proposed methods to obtain the estimates for A and B. and uplink channel estimation. As mentioned in the previous To simplify the notation, we will drop the subcarrier index l. section, matrices A and B are expected to change very slowly compared to channel coherence time and hence it is assumed that their values are fixed during the above pilot signaling. C. Proposed Estimation of B Fig. 3 illustrates the overall assumed radio frame or subAs described earlier, B̂ is iteratively refined using the current frame structure of the considered massive MIMO TDD system estimate of A. The proposed estimator builds on solving the including the proposed NRC estimation phase. matrix equation in (15) based on minimizing the Frobenius norm criterion. With this setting, the refined estimate of B can be formulated as B. Overall Estimation Framework 2 p p ∗ T As the initial step in estimating A and B, the BS processes B̂(m + 1) = argmin Q − ρ̃ ρ̃ Ĝ Â(m) Ĝ B , u d ∗ F the received signal Ynrc in (14) as Q = Ynrc XH B nrc . Since (17) the pilot matrix Xnrc has the property XH nrc Xnrc = IN , the where the subscript F denotes√the √ Frobenius norm. processed signal can be expressed as Next, by denoting T(m) = ρ̃u ρ̃d Ĝ∗ Â(m)ĜT , we have p p Q = ρ̃u ρ̃d G∗ AGT B + V, (15) the following identity K where the processed noise matrix is given by V = Z̃∗tot XH X nrc . 2 2 \|\|Q − T(m)B\|\|F = \|\|qj − T(m)bj \|\| , (18) Now the target is to estimate A and B from (15) assuming j=1 that the BS has the uplink channel estimate Ĝ. In this respect, denoting the estimates at m’th iteration as Â(m) and B̂(m), where qj and bj denote the j’th column of Q and B, we propose the following iterative estimation framework: respectively. Since the j’th term in the sum depends only 1) Initialize, Â(0) = IK , obtain the estimate B̂(1). on bj , minimizing the total sum is equivalent to separately 2) Substitute B̂(1) for B in (15) and obtain estimate Â(1). minimizing each term \|\|qj − T(m)bj \|\|2 . Thus, the estimation 3) Successively refine the estimates Â and B̂ by fixing the of matrix B is eventually simplified to estimation of each current value of one and solving for the other from (15). column of B, independently. In above, IK is used for initialization since the deviation As mentioned earlier, the BS NRC matrix incorporates both matrix A0 in A = IK + A0 is in practice small. Notice that the the frequency-responses and mutual coupling at the BS side. 6 IEEE TRANSACTIONS ON SIGNAL PROCESSING  /2 The solution to (20) is then given by  /2 † red b̂red j (m + 1) = Tj (m) qj .  D  0, C ( D )  0  D  1, C ( D )  4  D  2, C (D)  8 corresponding antenna unit Fig. 4. Illustration of sparsity threshold D on 10 × 10 rectangular BS antenna grid with λ/2 antenna spacing. The power of mutual coupling between two different antenna units is related to their physical distance, thus the off-diagonal elements in B become smaller as the distance between the two corresponding antenna units increases. Here, in estimating the BS NRC matrix B, we treat those off-diagonal entries which are corresponding to two antennas with a distance larger than a pre-defined threshold, called sparsity threshold D, as zeros, yielding a sparse matrix structure for B̂. We also define the maximum number of coupled neighboring antenna elements as C(D). In Fig. 4 an example 10 × 10 rectangular antenna layout with λ/2 antenna spacing between the neighboring elements√is shown with 3 different values of D, namely D = 0, 1 and 2, measured as multiples of λ/2. When D = 0, it is assumed that there is no√mutual coupling and C(D) = 0, whereas for D = 1 and D = 2, the central antenna elements are coupled with at most C(D) = 4 and C(D) = 8 closest neighboring antenna elements. Note that, the antenna elements close to the edges of the grid are coupled with less number of antenna units. This is illustrated in Fig. 4 where for D = 1, the bottom left antenna element is assumed to be coupled with only 2 ≤ C(1) = 4 antennas. The following column-wise estimator will build on the assumption that B has a sparse structure and the number of non-zero row entries in any column j, denoted as Rj , satisfies Rj ≤ R (D) , (19) where R(D) = C(D)+1. It is further assumed that the index of non-zero entries of bj are known, which is directly determined by the antenna system architecture and geometry, and the assumed pre-defined coupling threshold discussed above. With these assumptions, we define a reduced vector of dimension Rj × 1, bred j , that contains the non-zero entries of bj . If the j’th row is kept when constructing bred j , then similarly, the j’th column is kept to construct Tred (m). Based on these, we j can formulate the estimation of columns of B̂(m + 1) through a reduced system of equations as red red b̂red j (m + 1) = argmin qj − Tj (m)bj bred j 2 . (20) (21) Once the b̂red j (m + 1) is solved from (21), then b̂j (m + 1) can be obtained straightforwardly by appending zeros to the appropriate rows. Note that, when A ≈ IK , we also have G∗ AGT ≈ G∗ GT , ∗ where the matrix G∗ GT = GGH is positive semi-definite matrix and of rank K if G is of rank K. The obtained b̂red j and the corresponding minimum expression from (20) depend on the corresponding values of K and D. The column space of Tred j (m) has higher dimensionality for larger K. Thus, when D is fixed, for larger K one can solve for b̂red from (21) j red 2 which yields smaller values of qj − Tred (m)b . j j D. Proposed Estimation of A Next, given B̂(m) from (17), the (refined) estimate of A can be formulated based on minimizing the Frobenius norm criterion as 2 p p Â(m) = argmin Q − ρ̃u ρ̃d Ĝ∗ AĜT B̂(m) . (22) F A For diagonal A, the solution to (22) can be obtained as Â(m) = diag ξ̂ . (23) where ξ̂ = [IK , iIK ] ψ̂ and the vector ψ̂ is given as  −1 N N X X ψ̂ =  W̄jT W̄j  W̄jT q̄j . j=1 (24) j=1 T T In above, q̄j = <{qT and defining the N × j }, ={qj } , h i √ √ N matrix Wj = ρ̃u ρ̃d Ĝ∗ diag ĜT B̂(m) , with j h i ĜT B̂(m) being the j’th column of ĜT B̂(m), W̄j is given j as " # <{Wj }, −={Wj } W̄j = . (25) ={Wj }, <{Wj } Proof: See Appendix. IV. N UMERICAL E VALUATIONS AND A NALYSIS A. Basic Evaluation Settings and Performance Measures In this section, by using extensive computer simulations, we evaluate the performance of the proposed NRC estimation and mitigation scheme. We also compare its performance to the performance of two other existing schemes in literature, namely the direct-path based least squares (LS) known as “Argos” [2] and the generalized neighbor LS [11]. The latter is the optimized version of the generalized LS method presented in [10] and is shown in [11] to have the best performance amongst several existing NRC estimation methods. Both LS based methods estimate the BS NRC by the means of mutual coupling between BS antennas, while they depend on the downlink pilots to compensate the NRC in the UE side. where the expectation is over different NRC realizations and channel coherence intervals. The length of downlink pilots in symbols is denoted by τd and T is the total number of symbols in a channel coherence interval. SINR is the instantaneous signal to interference and noise ratio (SINR), which can be written, based on (8), as √ 2 ρd α̂k sk SINR = (27) √ 2, rk − ρd α̂k sk where α̂k is the scaling of the useful signal term available at the receiver of the k’th UE. In the context of the proposed NRC estimation and mitigation method, no DL CSI pilots are used. Hence, for the proposed estimation method, α̂k = E βhT k uk . On the contrary, the other two estimation methods utilizes τd ≥ K downlink pilots for DL CSI acquisition as described in [22]. The other relevant performance metric is the normalized mean squared error (MSE) for NRC estimation which is defined as  2   B − B̂   F   , for BS side  2 \|\|B\|\|F 2 (28) δe = 2   diag (A) − diag Â   F   , for UE side.  2 \|\|diag (A)\|\|F 10!1 Proposed method - D = 0 Proposed method - D = 1p Proposed method - D = 2 BS 10!2 UE 10!3 -40 -35 -30 -25 -20 -15 -20 -15 2 <M (dB) (a) 120 Spectral E/ciency (bits/s/Hz) We consider the DL spectral efficiency as the key performance metric, which is defined as τu + τd ηs = K 1 − E [log2 (1 + SINR)] , (26) T 7 NRC Estimation Normalized MSE (/e2 ) MITIGATION OF CHANNEL NON-RECIPROCITY IN MASSIVE MIMO SYSTEMS 100 Reciprocal Channel ZF p Proposed method - D = 2 Proposed method - D = 1 Proposed method - D = 0 No NRC mitigation 80 60 40 -40 MRT -35 -30 -25 2 <M (dB) (b) Fig. 5. (a) NRC estimation normalized MSE and (b) system spectral efficiency 2 ) for different values of the sparsity vs. input reflection coefficients variance (σM threshold D with N = 100, K = 20, τu = K, ρu = 0 dB, T = 250. As a baseline simulation scenario, we consider a BS which is equipped with N = 100 infinitely thin λ/2 dipole antennas in a 10 × 10 square layout with λ/2 spacing as illustrated in Fig. 4. The input and the mutual impedances are computed based subcarriers, Csc = 10, over which the NRC realizations are on [24] for the assumed carrier-frequency of fc = 3.5 GHz. assumed to be constant. Finally, the variances of transceivers The impedances are assumed to be frequency-independent, as frequency-responses in both BS and UE side are assumed to the modulated signal RF bandwidth is much smaller than the be −20 dB, i.e., σL2 = σF2 = −20 dB. These are the baseline carrier frequency. The uplink channel matrix G is assumed simulation settings, while some of the parameter values are to have i.i.d. CN (0, 1) elements. The BS serves K = 20 also varied in the evaluations. single-antenna UEs simultaneously on the same time-frequency resource through either ZF or MRT precoding. We assume a scenario where each coherence interval contains T = 250 B. Numerical Results OFDM symbols. The number of uplink pilots sent by each UE 1) Effect of Sparsity Distance Threshold D: Here, we in each coherence interval is equal to the number of scheduled will study the effect of D on the normalized MSE and the UEs, τu = K, and the uplink SNR in this phase is assumed to system spectral efficiency. In this respect, Fig. 5a illustrates be ρu = 0 dB. In the scenarios where UEs rely on downlink the normalized MSE of UE and BS NRC estimation under the 2 pilots for decoding purposes, i.e., direct-path based LS and baseline system settings, with the value of σM being varied. It generalized neighbor LS methods, the number of downlink can be seen that the choice of D = 0, i.e., estimating only the pilots in each coherence interval is set to be τd = K [22], and diagonal elements of B, yields the lowest MSE for UE NRC their SNR is equal to the downlink SNR in data transmission estimation. Note that, in the proposed NRC estimation method, phase which is assumed to be ρd = 20 dB. The SNR of the the choice of D influences the UE side estimation as well since coupling channel between two neighboring antennas is set to A and B are estimated iteratively as described in Section III-B. be 80 dB for the two mentioned NRC mitigation methods [11]. On the other hand, the highest BS NRC estimation accuracy is 2 The uplink and downlink SNRs for the pilot signaling in the achieved for D = 0 only when σM ≤ −21 dB, whereas higher 2 proposed NRC estimation framework are set to be ρ̃u = 0 estimation accuracy is obtained for D = 1 when σM > −21 dB and ρ̃d = 10 dB, respectively. In the proposed method, dB. Following that, the spectral efficiencies plotted in Fig. 5b the estimated NRC matrices are averaged over 10 neighboring illustrate the combined effect of UE and BS NRC estimation. 100 Proposed method - D = 0 Proposed method - D = 1p Proposed method - D = 2 10!1 BS 10!2 10!3 UE !4 10 10 20 30 40 50 60 Number of UEs (K) 70 80 NRC Estimation Normalized MSE (/e2 ) IEEE TRANSACTIONS ON SIGNAL PROCESSING NRC Estimation Normalized MSE (/e2 ) 8 100 2 <F2 = <L2 = <M = !15 dB 2 <F2 = <L2 = <M = !20 dB 10!1 BS 10!2 UE !3 10 0 1 2 3 4 5 Iteration Number (a) Fig. 7. NRC estimation normalized MSE vs. NRC estimation iteration number for N = 100, K = 20, τu = K, ρu = 0 dB, T = 250. Spectral E/ciency (bits/s/Hz) 250 sufficient amount of iterations such that convergence is obtained. This is commonly in the order of 4 iterations, as illustrated more specifically next. 150 ZF 2) Effect of the Number of Iterations: Fig. 7 illustrates 100 the reduction in NRC estimation normalized MSE over NRC p estimation iteration steps. It is shown in Fig. 7 that, even with Proposed method - D = 2 2 high NRC levels of σL2 = σF2 = σM = −15 dB, having 4 50 Proposed method - D = 1 Proposed method - D = 0 iteration rounds is sufficiently good for the proposed NRC MRT No NRC mitigation estimator to converge. Therefore, in the continuation, we set 0 10 20 30 40 50 60 70 80 the number of iteration rounds to 4. Number of UEs (K) 3) Effect of Number of Scheduled Users K: In Fig. 8, the (b) NRC estimation normalized MSE and the system spectral Fig. 6. (a) NRC estimation normalized MSE and (b) system spectral efficiency efficiency are plotted against the number of scheduled UEs K vs. number of UEs (K) for N = 100, τu = K, ρu = 0 dB, T = 250. 2 for σM = −20 dB. Fig. 8a shows that while direct-path based LS has the worst performance, the proposed method is the ≥ 20 with a high As can be seen, the highest spectral efficiency is achieved for best option for estimating BS NRC for K−3 accuracy where MSE is in the order of 10 . For direct-path 2 2 D = 0 when σM ≤ −22 dB and for D = 1 when σM > −22 based LS [2] and generalized neighbor LS [11], the normalized dB. MSE for UE side NRC is not shown. It is mentioned in [2] 2 2 2 For fixed NRC characteristics of σL = σF = σM = −20 and [11] that additional downlink pilot signaling per coherence dB, Fig. 6 evaluates the normalized estimation MSE and the system spectral efficiency for different values of D and against interval can be used together with UE side estimation for the number of scheduled UEs K. Fig. 6a shows that higher UE side NRC acquisition. However, no detailed description UE and BS NRC estimation accuracy is achieved for D = 0 is provided on the actual pilot signal structure or the actual when K < 20, whereas when the number of scheduled users estimation method. The corresponding system spectral efficiency performances exceeds K ≥ 20 the choice of D = 1 yields the highest BS are evaluated and shown in Fig. 8b. The proposed NRC NRC estimation accuracy. For K ≥ 20, UE NRC estimation estimation and mitigation scheme clearly outperforms the directperformances are largely similar for all three choices of D. path based LS and generalized neighbor LS methods. The Following these, Fig. 6b illustrates that from spectral efficiency difference between the performance of the proposed method perspective, the optimum sparsity distance threshold is D = 0 and the other two methods increases as K grows. Remarkably, for K < 20 and D = 1 for K ≥ 20. Thus, in the continuation for K = 70, the difference between the proposed method and D = 0 and D = 1 will be used under the settings of K < 20 the other two methods is already in the order of 100 bits/s/Hz. and K ≥ 20, respectively. As discussed in the previous section, ∗ T Another advantage in utilizing the proposed NRC estimation when A ≈ IK , G AG which is used in the estimation scheme is that the optimum number of UEs Kopt , which is process is of rank K. Therefore, having higher number of defined as the number of scheduled UEs which maximizes the K increases the accuracy of the BS NRC estimation in the spectral efficiency, is higher compared to the other two NRC proposed method which facilitates the estimation of more nonestimation methods. For instance under ZF precoding Kopt is diagonal elements in B, i.e., higher values for D. between 60 and 70 for the proposed method whereas for LS It should be noted that for all the cases in Fig. 5 and based methods K is only around 50. opt Fig. 6, the proposed iterative NRC estimator is executed over a 200 9 10!1 120 Spectral E/ciency (bits/s/Hz) NRC Estimation Normalized MSE (/e2 ) MITIGATION OF CHANNEL NON-RECIPROCITY IN MASSIVE MIMO SYSTEMS 10!2 10!3 10!4 10 Direct-path based LS [2] - BS Generalized neighbor LS [11] - BS Proposed method - BS Proposed method - UE 20 30 40 50 60 70 100 ZF 80 Proposed method Generalized neighbor LS [11] Direct-path based LS [2] 60 40 -40 80 MRT -35 -30 -25 -20 -15 2 <M (dB) Number of UEs (K) (a) 2 ) for Fig. 10. Spectral efficiency vs. input reflection coefficients variance (σM N = 100, K = 20, τu = K, ρu = 0 dB, T = 250. Spectral E/ciency (bits/s/Hz) 250 Proposed method Generalized neighbor LS [11] Direct-path based LS [2] 200 ZF 150 100 50 MRT 0 10 20 30 40 50 60 70 80 Number of UEs (K) (b) Fig. 8. (a) NRC estimation normalized MSE and (b) system spectral efficiency vs. number of UEs (K) for N = 100, τu = K, ρu = 0 dB, T = 250. Spectral E/ciency (bits/s/Hz) 120 100 MRT 60 40 20 • ZF 80 under ZF precoding. 5) Effect of Input Reflection Coefficient: Fig. 10 shows the impact of the variance of the input reflection coefficients on the achievable spectral efficiency. The proposed estimation and mitigation method again outperforms the other two LS based methods. The difference between the proposed method and the 2 other two methods increases as σM grows, which is due to the ability of the proposed method to estimate the non-diagonal elements in BS NRC matrix. It should be noted that D = 1 is used for obtaining the results in Fig. 10, and there is still room for improving the performance of the proposed method by adaptively selecting the optimum D according to the level 2 of σM as shown already in Fig. 5b. 6) Summary of the Obtained Results: Overall, as observed through extensive numerical evaluations in various scenarios, the proposed NRC estimation method outperforms the other two state-of-the-art methods. Selected technical aspects can be summarized as follows: Proposed method Generalized neighbor LS [11] Direct-path based LS [2] 0 5 10 15 20 25 30 • SNR (dB) Fig. 9. Spectral efficiency vs. downlink SNR (ρd ) for N = 100, K = 20, τu = K, ρu = 0 dB, T = 250. 4) Effect of Downlink SNR: In Fig. 9, the system spectral 2 efficiency is plotted against the downlink SNR when σM = −20 dB. The results show clear advantage in employing the proposed method in estimating NRC for all SNR values. The proposed estimation and mitigation method outperforms the LS based methods for both low and high SNR regions. Especially, the performance difference is most visible for high SNR region • Employing the proposed NRC estimation method eliminates the need to send downlink demodulation pilots since the proposed OTA framework facilitates estimating both the BS side and UE side NRC characteristics in the base station. Therefore, more time-frequency resources can be allocated in each coherence interval for actual downlink data transmission purposes which improves the spectral efficiency. The proposed NRC estimation method is more and more superior over the two reference methods when the number of scheduled UEs K grows. The reason is that increasing K is forcing the other two NRC estimation methods to spend more time for downlink pilot transmission in each coherence interval, while a larger number of scheduled users improves the accuracy of the proposed NRC estimation method. Due to the ability to estimate also non-diagonal elements of the BS NRC matrix, the difference between the performance of the proposed NRC estimation method and the other two methods increases as the power of BS antenna mutual coupling mismatch grows. 10 IEEE TRANSACTIONS ON SIGNAL PROCESSING V. C ONCLUSION In this work, we proposed an efficient NRC estimation and mitigation framework for multi-user massive MIMO TDD networks to compensate the jointly coexisting BS and UE side NRC. In general, even relatively modest NRC levels can cause significant performance loss in the achievable spectral efficiency when only standard NRC-blind MRT or ZF downlink precoding is employed. A novel OTA-based approach incorporating a dedicated round-trip pilot signaling with small pilot overhead together with sparsity-aided efficient iterative estimation techniques were proposed for the acquisition of NRC matrices at BS. Unlike the existing state-of-the-art methods, the proposed NRC estimation method acquires both the UE transceiver NRC as well as the BS transceiver NRC, and does not rely on downlink pilot transmission during the actual data transmission phase to compensate the NRC in the UE side. Therefore, it can be efficiently employed in massive MIMO systems that rely only on the statistical knowledge of the beamformed downlink channels at terminals for data decoding with very low system pilot overhead. The extensive computer simulations showed that for practical values of the NRC levels, SNRs and the number of spatially multiplexed users, the proposed estimation and mitigation method clearly outperforms the existing state-of-the-art methods in terms of the system spectral efficiency. A PPENDIX P ROOF FOR ESTIMATION OF A Let ∆ L= Q− p p ρ̃u ρ̃d Ĝ∗ AĜT B̂(m) 2 . (29) F Then, L= N X qj − p ρ̃u p h i ρ̃d Ĝ∗ A ĜT B̂(m) j=1 2 j h N i X p p T ∗ qj − ρ̃u ρ̃d Ĝ diag Ĝ B̂(m) = ξ 2 , j j=1 (30) ∆ T where ξ = [a1 , a2 , · · · , aK ] . h i √ √ ∗ T By using Wj = ρ̃u ρ̃d Ĝ diag Ĝ B̂(m) , (30) j can be re-written as L= N X 2 \|\|qj − Wj ξ\|\| j=1 = N X (31) q̄j − W̄j ψ 2 ∆ = L̄ (ψ) , j=1 T ∆ where ψ = <{ξ T }, ={ξ T } . Therefore, the solution Â(m) can be obtained by Â(m) = diag [IK , iIK ] ψ̂ , (32) where ψ̂ = argmin L̄ (ψ) . ψ∈RK (33) Since L̄ (ψ) is convex, (33) can be solved from the partial derivative equation ∂ L̄(ψ) ∂ψ = 0, which finally yields the solution given by (24). ACKNOWLEDGMENT The estimation of A is building on the work presented in [25]. The proof in Appendix A is a result of discussion with the authors of [25], H. Q. Ngo and E. G. Larsson, whose technical guidance is greatly acknowledged. R EFERENCES [1] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski, “Five disruptive technology directions for 5G,” IEEE Communications Magazine, vol. 52, no. 2, pp. 74–80, February 2014. [2] C. Shepard, H. Yu, N. Anand, E. Li, T. Marzetta, R. Yang, and L. Zhong, “Argos: Practical many-antenna base stations,” in Proceedings of the 18th Annual International Conference on Mobile Computing and Networking, ser. Mobicom ’12. New York, NY, USA: ACM, 2012, pp. 53–64. [3] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive MIMO for next generation wireless systems,” IEEE Communications Magazine, vol. 52, no. 2, pp. 186–195, February 2014. [4] A. Bourdoux, B. Come, and N. Khaled, “Non-reciprocal transceivers in OFDM/SDMA systems: Impact and mitigation,” in Radio and Wireless Conference, 2003. RAWCON ’03. Proceedings, Aug 2003, pp. 183–186. [5] Y. Zou, O. Raeesi, R. Wichman, A. Tolli, and M. Valkama, “Analysis of channel non-reciprocity due to transceiver and antenna coupling mismatches in TDD precoded multi-user MIMO-OFDM downlink,” in 2014 IEEE 80th Vehicular Technology Conference (VTC2014-Fall), Sept 2014, pp. 1–7. [6] M. Petermann, M. Stefer, F. Ludwig, D. Wubben, M. Schneider, S. Paul, and K. D. Kammeyer, “Multi-user pre-processing in multi-antenna OFDM TDD systems with non-reciprocal transceivers,” IEEE Transactions on Communications, vol. 61, no. 9, pp. 3781–3793, September 2013. [7] H. Wei, D. Wang, and X. You, “Reciprocity of mutual coupling for TDD massive MIMO systems,” in Wireless Communications Signal Processing (WCSP), 2015 International Conference on, Oct 2015, pp. 1–5. [8] W. Zhang, H. Ren, C. Pan, M. Chen, R. C. de Lamare, B. Du, and J. Dai, “Large-scale antenna systems with UL/DL hardware mismatch: Achievable rates analysis and calibration,” IEEE Transactions on Communications, vol. 63, no. 4, pp. 1216–1229, April 2015. [9] F. Athley, G. Durisi, and U. Gustavsson, “Analysis of massive MIMO with hardware impairments and different channel models,” in 2015 9th European Conference on Antennas and Propagation (EuCAP), May 2015, pp. 1–5. [10] R. Rogalin, O. Y. Bursalioglu, H. Papadopoulos, G. Caire, A. F. Molisch, A. Michaloliakos, V. Balan, and K. Psounis, “Scalable synchronization and reciprocity calibration for distributed multiuser MIMO,” IEEE Transactions on Wireless Communications, vol. 13, no. 4, pp. 1815– 1831, April 2014. [11] J. Vieira, F. Rusek, and F. Tufvesson, “Reciprocity calibration methods for massive MIMO based on antenna coupling,” in 2014 IEEE Global Communications Conference, Dec 2014, pp. 3708–3712. [12] H. Wei, D. Wang, H. Zhu, J. Wang, S. Sun, and X. You, “Mutual coupling calibration for multiuser massive MIMO systems,” IEEE Transactions on Wireless Communications, vol. 15, no. 1, pp. 606–619, Jan 2016. [13] H. Wei, D. Wang, J. Wang, and X. You, “TDD reciprocity calibration for multi-user massive MIMO systems with iterative coordinate descent,” Science China Information Sciences, vol. 59, no. 10, p. 102306, 2015. [14] M. Guillaud, D. T. M. Slock, and R. Knopp, “A practical method for wireless channel reciprocity exploitation through relative calibration,” in Proceedings of the Eighth International Symposium on Signal Processing and Its Applications, 2005., vol. 1, August 2005, pp. 403–406. [15] Y. Zou, O. Raeesi, and M. Valkama, “Efficient estimation and compensation of transceiver non-reciprocity in precoded TDD multi-user MIMOOFDM systems,” in 2014 IEEE 80th Vehicular Technology Conference (VTC2014-Fall), Sept 2014, pp. 1–7. [16] M. Guillaud and F. Kaltenberger, “Towards practical channel reciprocity exploitation: Relative calibration in the presence of frequency offset,” in 2013 IEEE Wireless Communications and Networking Conference (WCNC), April 2013, pp. 2525–2530. MITIGATION OF CHANNEL NON-RECIPROCITY IN MASSIVE MIMO SYSTEMS [17] O. Raeesi, A. Gokceoglu, Y. Zou, and M. Valkama, “Performance analysis of multi-user massive MIMO downlink under channel non-reciprocity and imperfect CSI,” submitted to IEEE Transactions on Communications, 2016. [Online]. Available: http://arxiv.org/abs/1610.06769 [18] H. Yang and T. L. Marzetta, “Performance of conjugate and zero-forcing beamforming in large-scale antenna systems,” IEEE Journal on Selected Areas in Communications, vol. 31, no. 2, pp. 172–179, February 2013. [19] J. Jose, A. Ashikhmin, P. Whiting, and S. Vishwanath, “Channel estimation and linear precoding in multiuser multiple-antenna TDD systems,” IEEE Transactions on Vehicular Technology, vol. 60, no. 5, pp. 2102–2116, Jun 2011. [20] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot contamination and precoding in multi-cell TDD systems,” IEEE Transactions on Wireless Communications, vol. 10, no. 8, pp. 2640–2651, August 2011. [21] E. G. Larsson and H. V. Poor, “Joint beamforming and broadcasting in massive MIMO,” IEEE Transactions on Wireless Communications, vol. 15, no. 4, pp. 3058–3070, April 2016. [22] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Massive MU-MIMO downlink TDD systems with linear precoding and downlink pilots,” in Communication, Control, and Computing (Allerton), 2013 51st Annual Allerton Conference on, Oct 2013, pp. 293–298. [23] B. M. Hochwald, T. L. Marzetta, and V. Tarokh, “Multiple-antenna channel hardening and its implications for rate feedback and scheduling,” IEEE Transactions on Information Theory, vol. 50, no. 9, pp. 1893–1909, Sept 2004. [24] S. A. Schelkunoff and H. T. Friis, Antennas Theory and Practice. New York: John Wiley & Sons, 1952. [25] H. Q. Ngo and E. G. Larsson, “EVD-based channel estimation in multicell multiuser MIMO systems with very large antenna arrays,” in 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), March 2012, pp. 3249–3252. 11	7cs.IT
Concept Formation and Dynamics of Repeated Inference in Deep Generative Models arXiv:1712.04195v1 [stat.ML] 12 Dec 2017 Yoshihiro Naganoa,d , Ryo Karakidab,d , Masato Okadaa,b,c,∗ a Graduate School of Frontier Sciences, The University of Tokyo, 5-1-5 Kashiwanoha, Kashiwa, Chiba 277-8561, Japan b Artificial Intelligence Research Center, National Institute of Advanced Industrial Science and Technology, 2-3-26 Aomi, Koto-ku, Tokyo 135-0064, Japan c Brain Science Institute, RIKEN, 2-1 Hirosawa, Wako, Saitama 351-0198, Japan d Research Fellow of the Japan Society for the Promotion of Science, Chiyoda, Tokyo 102-0083, Japan Abstract Deep generative models are reported to be useful in broad applications including image generation. Repeated inference between data space and latent space in these models can denoise cluttered images and improve the quality of inferred results. However, previous studies only qualitatively evaluated image outputs in data space, and the mechanism behind the inference has not been investigated. The purpose of the current study is to numerically analyze changes in activity patterns of neurons in the latent space of a deep generative model called a “variational auto-encoder” (VAE). What kinds of inference dynamics the VAE demonstrates when noise is added to the input data are identified. The VAE embeds a dataset with clear cluster structures in the latent space and the center of each cluster of multiple correlated data points (memories) is referred as the concept. Our study demonstrated that transient dynamics of inference first approaches a concept, and then moves close to a memory. Moreover, the VAE revealed that the inference dynamics approaches a more abstract concept to the extent that the uncertainty of input data increases due to noise. It was demonstrated that by increasing the number of the latent variables, the trend of the inference dynamics to approach a concept can be enhanced, and the generalization ability of the VAE can be improved. ∗ Corresponding author Email addresses: nagano@mns.k.u-tokyo.ac.jp (Yoshihiro Nagano), karakida.ryo@aist.go.jp (Ryo Karakida), okada@k.u-tokyo.ac.jp (Masato Okada) Preprint submitted to Neural Networks December 13, 2017 Keywords: Deep Generative Models, Variational Auto-Encoder, Inference, Concept Formation 1. Introduction Research on deep generative models, which extract essential features from an unlabeled dataset, is currently an active area. Deep generative models have been reported to be useful in a broad range of applications such as generating images [13, 24, 7, 23], movies [32, 33, 26], and text [34, 16, 28]. In particular, the conventional bidirectional network structure for recognition and generation of images has made it possible to eliminate noise from cluttered images and smoothly interpolate between different images. In detail, recognition is the process of mapping from a data point to a latent variable, and generation is the inverse of that process. Several studies have qualitatively highlighted the importance of repeated inferences between data space and latent space. In the present study, repeated inferences are defined as a process by which a deep generative model repeats the recognition and generation of images. It was shown that using noise-containing images as initial values, deep generative models can eliminate noise by repeating recognition and generation several times [24]. Moreover, compared to generating an output image from latent space to smoothly morph one image into another, repeating inferences several times improves the quality of the output image [3]. However, most of these studies only qualitatively evaluate output data. To fill that gap in the literature, we quantified the dynamics of repeated inferences in latent space to investigate why repeating inferences is effective for a wide range of applications. In this study, we focused on the dynamics of repeated inferences in a “variational auto-encoder” (VAE) [13, 24], which is a typical type of deep generative model. First, noise-containing images are presented as initial inputs to the VAE, which denoises the images and generates clean outputs by using repeated inferences. Various factors, such as noise in real environments, cause the data to deviate from the original distribution. According to the manifold hypothesis, most data (e.g., natural images) widely used in typical applications are “likely to concentrate in the vicinity of non-linear sub-manifolds of much lower dimensionality” rather than the high dimensional space the data actually presented [25]. Therefore, it is suggested that the inference begins outside the sub-manifolds of the training data when noise was added to the initial inputs. However, little works have paid attention to how the activity patterns of neurons are drawn into the original sub-manifolds from the initial point in latent space. To understand the patterns 2 of repeated inferences, we used a dataset with clear cluster structures which could be intuitively grasped by our eyes (e.g., numeric characters). To examine how a VAE expresses the cluster structures in latent space, we utilized a mathematical notion known as the concept. First introduced in the study of associative memory models [1], the concept was referred as the centroid of multiple correlated data points (memories), and it was analytically demonstrated to spontaneously evolve to an equilibrium state [1]. This phenomenon is called concept formation. Furthermore, the dynamics of neural activity patterns has also been studied in terms of the associative memory model with multiple correlated memories [18]. It was revealed that the dynamics of neural activity patterns first approach the concept and then move toward each memory pattern [18]. In summary, there are four major findings in our study. First, consistent with these reports, we demonstrated that the dynamics of repeated inferences is drawn to a unique memory via a corresponding concept, which is the center of each cluster in the latent space. Second, by averaging all clusters in the latent space, we defined the abstract concept; now by definition, memories, concepts, and the abstract concept are hierarchically related in ascending order. We found that the inference dynamics approaches the abstract concept, to the extent that the uncertainty of input data increases due to noise. The result suggested that the model selects appropriate inference strategies in accordance with the fraction of the noise added to the input data. Third, we identified the approximate necessary number of latent variables to map all memories in the latent space. As the number of latent variable increases, the internal representations of the clusters tend to become orthogonal and it makes the dynamics of repeated inferences approaches each corresponding concept. Finally, we checked the generalization error of the VAE. The result demonstrated that the generalization performance of the model has been improved, to the extent the concept is observed to attract the dynamics of repeated inferences. 2. Method A variational auto-encoder (VAE) is a generative model consisting of two neural networks, namely, an encoder and a decoder [13, 24]. An encoder sends a mapping from data such as natural images to a latent variable space, and the decoder gives an inverse mapping. The objective function of the VAE is obtained by P finding the variational lower bound of log-likelihood i log p(x(i) ) for N training n oN data X = x(i) . In the following, parameter θ that maximizes log-likelihood i=1 3 log pθ (x(i) ) at each data point is considered. Using latent variable z and its conditional probability distribution qφ (z \| x(i) ) and taking the variational lower bound of the log-likelihood gives the following objective function: h i log pθ (x(i) ) ≥ −DKL qφ (z \| x(i) ) p(z) + Eqφ (z\|x) log pθ (x(i) \| z) = L(θ, φ; x(i) ). (1) (2) In the above equation, p(z) is the prior distribution of latent variable z, and DKL (qkp) is the Kullback-Leibler divergence [14] of probability distributions q and p. The first term of the objective function corresponds to the regularization, and the second term corresponds to the reconstruction error. The VAE models conditional distributions pθ (x(i) \| z) and qφ (z \| x(i) ) by using respective neural networks. To optimize parameters θ and φ by backpropagation, samples are generated by a method called reparameterization trick with encoder qφ (z \| x(i) ). The latent variable is modeled as z = gφ (, x) = µ + σ , (3) (4) to decompose z into random variable and deterministic variables µ and σ. Giving as a sample from the standard Gaussian distribution eliminates the need for a complicated integral during training. If are assumed, and the h the above conditions i expected reconstruction error Eqφ (z\|x) log pθ (x(i) \| z) is approximated by sample average, eq. (2) can be rewritten as J L 1X 1 X (i) 2 (i) 2 2 ) + 1 + log (σ(i) ) − (µ ) − (σ log pθ (x(i) \| z(i,l) ). L(θ, φ; x ) ' j j j 2 j=1 L l=1 (5) The outputs of the encoder are µ and σ. Parameter φ for determining them and parameter θ of the decoder were trained by the gradient ascent method to maximize eq. (5). The output of the decoder was set as the probability of the Bernoulli distribution, and the expectation of the conditional probability, namely, the second term of the objective function, was approximated by average of L samples. A separate three-layer fully-connected neural network was used for encoder qφ (z \| x(i) ) and decoder pθ (x(i) \| z) mentioned above. The number of units in the (i) 4 middle layer was set to 1,024, the number of samples for calculating the reconstruction error was set to L = 2, and the activation function was set as tanh. Adam [12] was used as a parameter optimization algorithm, and the learning rate was reduced in descending order: 0.0005, 0.0001, and 0.00005. The number of units of latent variable Nz was set to 100 unless otherwise noted. The Modified National Institute of Standards and Technology (MNIST) database (which consists of 28 × 28-pixels ‘0’-‘9’ handwritten images with 60,000 training data and 10,000 test data) was used as the dataset for training. These data are considered to have cluster structures consisting of 10 types of labels, namely, ‘0’-‘9’. In this study, noisy MNIST data were inferred by using the trained network according to the following procedure, and the time evolution of latent variable z(t) was obtained. First, noise was added to an image of the training dataset. Pixels with probability p were selected from 784 pixels, the image intensities of the selected pixels were swapped, and that image was set as x0 . Next, the data variable in step t = 0 was taken as x(0) = x0 . Finally, generation and recognition were repeated T times according to the following two equations, x(t + 1) = E pθ (x\|z(t)) [x], z(t) = Eqφ (z\|x(t)) [z], (6) (7) to obtain the time evolution of data variable x(t) and latent variable z(t). The dynamics of x(t) and z(t) were numerically analyzed as shown below. The deep learning framework Keras [6] version 2.0.2 on Theano [30] backend version 0.9.0, running on CUDA 8.0 with CuDNN v5.1 on NVIDIA Tesla K80, was used for all numerical simulations. 3. Results 3.1. Dynamics of inference trajectory: approach to a concept In our first analysis, we demonstrate that the transient dynamics of inference in latent space of the VAE is first attracted to the concept, which is the center of the memorized patterns, and then moves to each memory. The representation in the latent space of the VAE captures the cluster structures hidden behind the MINIST data, and the VAE displays corresponding clusters in its latent space (fig. A.1). The details of the expression in the latent space of the VAE are given in Appendix A. Since the cluster structure exists in the data and the latent variable of the VAE 5 reflects it, the time evolution of inference also seems to reflect the cluster structure. The center of the cluster of each numerical label, namely, ξ̄num = 1 N num X Nnum i ξ(i) num , (8) was therefore defined as a concept, where ξ(i) num means the activity pattern of the latent variable for i-th training data having label num ξ(i) [z]. num = Eqφ (z\|x(i) num ) (9) The definition of the concept is based on two studies on an associative memory model [1, 18]. The relationship between the time development of inference and the concept of each label (‘0’-‘9’) represented in the MNIST data is numerically analyzed in the following sections. Figure 1: Consecutive samples in data space (from left to right, one row after the other). The image of ‘6’ with p = 0.2 noise applied was used as an initial value. The image generated by concept E pθ (x\|ξ̄6 ) [x] is shown on the right. Here, it is numerically demonstrated that the transient dynamics of the activity pattern are first attracted to the concept, which is the center of the memorized patterns, and then moves to each memory. Figure 1 expresses the consecutive samples in data space. The time development of the activity pattern in data space x(t) is aligned from left to right, one row after the other. The upper left image corresponds to the initial value, x0 . The image of ‘6’ with p = 0.2 noise applied was used as an initial value. The image generated by the concept, E pθ (x\|ξ̄6 ) [x], is shown on the right. From the figure, the VAE removes the noise contained in the image in the first few steps, and then it gradually shifts to the specific image of ‘6’. It is qualitatively suggested that the result of the VAE inference approaches the concept of ‘6’ once. The gradual changes of output images in the data space were quantitatively evaluated in the latent space. The time evolution of the Euclidean distance, namely, 6 15.1 18.6 16.1 1 15.8 2 17.5 3 4 Distance \|\|z(t) num\|\|2 0 6.7 0 17.5 8.2 80 0 16.7 5 0 5.4 80 7.1 80 0 80 17.1 6 0 10.2 80 6.9 0 15.7 7 0 7.2 80 7.4 80 0 16.6 8 0 8.4 80 6.3 80 9 0 80 Step Figure 2: Time development of the distance from concepts ξ̄num for all labels of the MNIST data. Shades represent ±1 standard error of the mean (300 trials). All figures were generated with noise fraction p = 0.2. z(t) − ξ̄num 2 , (10) between neural activity patterns and concepts for every label of MNIST data (fig. 2) in the latent space was evaluated. The distance between the concept and 300 different initial images was calculated. Each figure corresponds to each label, which is used as the initial input for the VAE. The x-axis expresses the time step t of repeated inference, and the y-axis expresses the Euclidean distance (eq. (10)). It was clarified that the trajectory of the VAE’s inference first rapidly approaches the concept, and then moves away from it. This result is qualitatively consistent with all other labels of MINIST data. The VAE makes the latent space in which the activity patterns are orthogonal in relation to another cluster and correlated with the same cluster in a lowdimensional space. The relationships between concepts of each label are shown in fig. 3. The activity patterns in the latent variable space of each numerical concept are shown in fig. 3a. The heat map expresses the activity pattern of each neuron, which corresponds to the latent variable, where the x-axis represents the neuron’s index, and the y-axis represents the label. Only a few neurons out of 100 contribute to information representation, and many neurons are pruned and not active. According to our observation, 14 out of 100 neurons were active. The dimensions of the latent space were examined by using the cumulative contribution ratio determined by the principal component analysis. The cumulative contribution ratio of each principal component when the training images are given to the VAE is shown in fig. 3b. The variance of the latent space was explained entirely 7 (c) 10 20 30 40 50 z index 60 70 80 0 0.8 0.8 2 0.4 0.6 4 0.0 0.4 6 Input label 0 1.0 0.2 90 0.4 8 9 8 7 6 5 4 3 2 1 0 1.0 0.5 0.0 0.5 1.0 Cumulative ratio (b) (a) 0 10 20 Components 30 0.8 0 2 4 6 8 Figure 3: (a) Activity pattern in the latent variable space of each concept. The x-axis represents the neuron index of the latent variable, the y-axis represents the label, and the heat map shows the activity pattern of each neuron. (b) Cumulative contribution ratio of principal components. (c) Cosine similarity between activity patterns of each label. by 14 dimensions, and 70% of that was explained by nine dimensions. To examine the relationships between concepts, the cosine similarity between the activity patterns of each concept i and j (fig. 3c) was obtained as Ci j = ξ̄i · ξ̄ j ξ̄i 2 ξ̄ j . (11) 2 By definition, the cosine similarity between concepts of the same label (shown in diagonal terms in the figure) is 1. On the other hand, the cosine similarity between concepts of different labels in non-diagonal terms is minimal, namely, near zero. The figure shows that the concepts of each label are orthogonal in the latent space. It is suggested that the activity patterns corresponding to different training data of the same label are correlated and those of the different label are orthogonal to each other in the latent space of the VAE. 3.2. Relationship between data hierarchy and inference Previously, we arbitrarily determined the amount of noise added to the initial input images. In our second analysis, to examine the effect of noise on the dynamics of repeated interferences, we modulated the amount of noise. Since noise in input images causes the data to deviate from the original distribution, we created another class, the abstract concept, which averaged all the labels’ concepts, in addition to concept and memory. By respectively measuring the distance between the trajectory of each neural activity pattern and its corresponding classes in the 8 latent space, which one of these classes attracts respective neural activity patterns the most is identified. The abstract concept is defined as ξ̄all = 9 1 X ξ̄ . 10 num=0 num (12) The three classes (memories, concepts, and the abstract concept) are in a hierarchical relationship (from detailed to coarse information) in the order ξ(i) num , ξ̄ num , and ξ̄all . We calculated the minimum distances between respective neural activity patterns z(t) and corresponding classes, min kz(t) − ξk2 . (13) t Figure 4 shows those minimum distances according to the noise fraction. In the figure, the x-axis represents noise fraction p, which is the probability that the image intensities of the pixels are swapped. For every noise fraction, the minimum distances between firing pattern z(t) and hierarchical concepts were calculated by changing the initial image 500 times. The dots in the figure express the mean of the minimum distance, and the bars are ±1 standard error of the mean (500 trials). Parameter regions are divided into three stages, I, II, and III, which correspond to the minimum distance between firing pattern z(t) and hierarchical concepts, ξ6(i) , II III mint z(t) − ξ 2 I ξ̄all ξ̄6 ξ6(i) Noise fraction p Figure 4: Minimum distances from concepts according to noise fraction p. 9 ξ̄6 , and ξ̄all , respectively. In stage I, the firing activity is closest to original pattern ξ(i) 6 with a small amount of noise. Interestingly, the closest concept was ξ̄ 6 with moderate noise in stage II. And the activity came close to concept ξ̄all in stage III. In stages I and II, the memory was successfully retrieved because the inference path is close to the cluster to which the input data belongs; however, in stage III, the model could not determine the original cluster, so the recall failed. We discovered that when a noisy environment makes recognizing objects difficult, the neural activity pattern wanders around the center of all memories. Accordingly, the model achieves the inference dynamics depending on the input uncertainty. Non-numeric space ξ̄all Num eric spa Concept ξ̄ 6 ce ξ6(i) Figure 5: Schematic diagram of firing patterns in latent state space As shown in section 3.1 and above, the VAE extracts the cluster structures inherent in MNIST data and infers images through the center of each cluster. The dynamics of this inference is shown as a schematic diagram in fig. 5. It is considered that adding noise to an image corresponds to moving the initial value in a direction orthogonal to the original data space. Results from our first analysis suggest that when the inference starts from the position orthogonal to the space expressing the MNIST data, the activity patterns first approach the corresponding concepts at high speed and then transitions to the corresponding memories. Approaching a concept in repeated inferences indicates that the concept formation 10 occurs in the latent space of the VAE. In neuroscience, the activities in the visual cortex of a macaque monkey [29, 5] and the human brain measured by MEG [17] have been reported to process global information before detailed information. The previous studies on associative memory models explain these behaviors by spontaneous stabilization of the concept and its effect on memory retrieval [1, 18]. The VAE also recalled the concept before each memory pattern in the inference phase. Our results in the VAE are consistent with these findings in studies on the associative memory models, and the cerebral cortex. 3.3. Effect of Model Architecture on Internal Representations In our third analysis, the mechanism by which the trajectory of VAE’s inference approaches the concept is clarified as follows. First, the following question must be answered: How does the tree-like memory structure mentioned above affect the time evolution of inference? In other words, how such a memory structure is changed by controlling a hyperparameter of the VAE was verified, and the trajectory of inference in the five circumstances was examined respectively (fig. 6). The time evolution of distance from the concept of ‘6’ with different model hyperparameters are compared in figs. 6a to 6e. The number of neurons in latent variable Nz was controlled in the order 2, 5, 10, 20, and 100. Under the condition 2 4 6 8 2 4 6 8 6 \|2 6 \|2 0.8 2 4 6 8 \|z 80 0 0.8 0 0.8 0.4 2 0.4 0.0 4 (j) 0.0 0.4 6 0.4 Step 0 2 0.0 2.3 4 0 0.4 0 80 0.8 8 0.8 Step 0 6 0.4 0.8 2 4 0.0 1.9 (i) 4 0 2 0.4 0 80 6 0.8 0.8 6 0.4 Step 0 (h) 8 0 2 4 0.0 6 0.4 8 0.8 1.9 8 80 4.0 \|z 6 \|2 Step 0 (g) (f) 0 1 Nz = 100 (e) 3.3 \|z 6 \|2 80 Nz = 20 (d) 3.3 \|z 6 \|2 Step 0 Nz = 10 (c) 3 \|z 0 Nz = 5 (b) 1 8 Nz = 2 (a) 0 2 4 6 8 0.4 0.8 0 2 4 6 8 Figure 6: (a-e): Time development of distance with the concept of ‘6’. The number of elements of the latent variable is written as Nz . (f-j): Cosine similarity between memory patterns of each concept in the corresponding model of (a-e). 11 Nz = 10, 20, 100, the trajectory approached the concept of ‘6’ once, whereas under condition of Nz = 2, 5, it did not approach the concept. Also, between Nz = 5 and Nz = 10, the trajectory gradually approached its corresponding concept from Nz = 6 (the figure is omitted). The cosine similarity (eq. (11)) between the concepts of each label in the above parameters is shown in figs. 6f to 6j. For Nz = 100, the similarity of the off-diagonal term is approximately zero. On the other hand, as the number of the latent variables decreases, the orthogonality of each concept decays. These results in our third analysis suggest that orthogonality between concepts is necessary for the trajectory of inference to be drawn into the concept. Since the number of latent variables decreases, it is necessary to express data in fewer dimensions, and the orthogonality between concepts is lost. The reduction in the number of latent variables is considered to cause instability of the memory patterns corresponding to each training data, and stabilize only the concepts. As a result, the trajectory of inference goes straight to a stable point. We also numerically assessed whether other labels confused the repeated inferences in the VAE (e.g., although an inference starts from the label ‘1’, it is incorrectly attracted to the concept associated with the label ‘7’). The result of this assessment is shown in Appendix B. 3.4. Engineering significance of concept in the latent representation In our fourth analysis, the engineering significance of the attraction to the concept is explained as follows. Figure 7 is the generalization performance of the model according to Nz . The performance of the model was evaluated using the variational lower bound (eq. (2)) of the log-likelihood for the test MNIST data. In each Nz , parameters that minimize the generalization error at epoch 100 with a total of nine conditions were selected from learning rates of 0.01, 0.001, and 0.0001 and mini-batch size of 50, 100, and 200. The generalization error was the minimum value in the vicinity of Nz = 14, and it did not change significantly after that. Fourteen of 100 latent variable neurons express training data under condition Nz = 100 (fig. 3a), and the number of neurons that minimize the generalization error was consistent with this result. These results suggest that about 14 latent neurons are required to express MNIST data in the network structure used in this study. Moreover, in the vicinity of Nz = 14, the cluster structure appears in the representation of the latent variable space, and the trajectory of an inference is drawn into the concept. These results suggest that it is possible to judge the generalization performance of the model without computing generalization error or 12 Validation loss 150 140 130 120 110 0 20 40 Nz 60 80 100 Figure 7: Generalization error for the number of elements of latent variables Nz . The y-axis represents the variational lower bound of the log-likelihood for test data. orthogonality of the internal representations by simply observing the dynamics of repeated inference. 4. Conclusion & Future Work It was found that a VAE extracts the cluster structures inherent in MNIST and infers images via the center of each cluster. Our results in first analysis suggest that when the inference starts from a point far away from the original data distribution, the repeated inferences first approach the concept at high speed and then slowly move toward each memory pattern. The learning and inference of multiple memory patterns has been widely studied by using associative memory models [9, 2, 22]. In an associative memory model with embedded multiple correlated patterns, the centroid of those correlated patterns spontaneously evolves to a fixed point [1], and the time evolution of activity patterns approaches the concept [18]. The results of our first and second analyses are qualitatively consistent with these findings, suggesting that the mechanism underlying dynamics of repeated inferences in the VAE is related to the traditional associative memory model. Originally, Matsumoto et al. proposed a model to explain the dynamics of the neural activities in macaque monkey’s visual cortex studied by Sugase et al. [29]. Although the VAE investigated in this study, the model proposed by Matsumoto et al., and the experiment conducted by Sugase et al. have different architectures by nature, our results infer that they share the universal working principle 13 at the abstract level. The behavior of the repeated inferences in the VAE was qualitatively consistent with the time evolution of firing patterns observed in neuroscience literature [29, 17, 5]. In our current study, we used a VAE to study a simple hierarchical deep generative model. Taking account that the VAE exhibits patterns similar to biological activities, our future work will further examine the dynamics of repeated inferences introducing biologically plausible mechanisms such as short-term plasticity [11, 19], common noise [10], and spontaneous firing based on log-normal weight [20]. This future investigation will be useful both for neuroscience and engineering applications. Previously, several studies demonstrated that repeated inferences successfully denoise [24] and improve the quality of inferred images [3]. Our study suggests that the dynamics of repeated inferences approaching the center of the cluster inherent in the data leads to denoising and improving the quality of output images, which were quantitatively observed in data space. It is critical to take a sufficient number of the latent variables to precisely represent the concept inherent in the data; if the number of the latent variables is not sufficient, the cluster structures will not be realized in the latent space, so the concept will be hardly identified. Our results suggest that stage II in fig. 4 appears only when the number of latent variables is sufficiently large, and the number of latent variables qualitatively changes the dynamics of repeated inferences. In this study, we introduced hierarchical concepts ξ(i) num , ξ̄ num , and ξ̄ all , which reflect a hierarchical structure of the MNIST dataset. Previous works have discussed the relationship between the hierarchy of data and deep neural networks. For example, deep neural networks are claimed to express abstract information in deeper layers [15, 4]. In particular, Bengio et al. stated that deep layers speed up the mixing of Markov chains by using their ability to manifest abstract information. On the other hand, Saxe et al. analytically showed that deep neural networks learn the data in order from large to small modes, and the internal representations branch accordingly [27]. These hypotheses and previous studies pointed out that the representations and the learning dynamics of deep neural networks reflect the data’s hierarchy. Our study suggests that the inference process of the deep generative model is also related to the hierarchy of the data. Recently, researchers are actively working on models that can capture features inherent in data as forms of internal representations [8, 31, 21]. The VAE used in this study embeds the data points in a simple isomorphic Gaussian distribution. As a next step to expand on these works, using other deep generative models, we aim to further investigate what factors influence the behavior of repeated inferences approaching the concept. Likewise, we will analyze the dynamics of the repeated 14 inferences in another model, in which we used training datasets with more and various hierarchies. Acknowledgements This work was supported by a Grant-in-Aid for JSPS Fellows (grant number: 17J08559) and a Grant-in-Aid for Research Activity Start-up (grant number: 17H07390) from the Japan Society for the Promotion of Science (JSPS). Appendix A. PCA embedding of the latent representations Here, we show that the VAE can extract the cluster structure hidden behind data. Activity patterns of the latent variable using principal component analysis (PCA) are shown in fig. A.1. The results of principal component analysis using all of the data are shown in fig. A.1a, and the result of using only three labels is shown in fig. A.1b. The color of each point represents the label of the data. MNIST is considered to have a cluster structure consisting of 10 types of labels, i.e., ‘0’-‘9’. The figures show that latent variables of VAE can extract this cluster structure. (a) (b) Figure A.1: (a) Two-dimensional PCA embedding of the representations in the latent variable space. The color of each point represents the label of the data. (b) PCA embedding with only data corresponding to three labels. 15 Appendix B. Verifying the effect of moving to other numbers The possibility that other labels confused the repeated inferences in the VAE was numerically tested. It was showed that the trajectory of inference approaches the concept by the orthogonality of the representation in the latent space. On the other hand, it is also conceivable that the escape from the concept is caused by the attraction of another cluster. To eliminate this possibility, a discriminative neural network was constructed separately from the VAE, and the final state of inference of the VAE was classified. In the following analysis, a model having a structure of Input-ConvolutionConvolution-Pooling-Dropout1-FullyConnected-Dropout2-SoftMax was constructed as the discriminative neural network. Kernel size of the convolution was set to three, the size of the pooling was to two, and the probability of dropout was set to 0.25, 0.5 in order from the input side. ReLU was used as the activation function. This model recorded a discrimination ability of 99.25% against test data included in the MNIST dataset. The result of classifying the final state of inference using the above-mentioned discriminative neural network is shown in fig. B.2a. The x-axis represents a trial of each inference with various initial images, and the y-axis represents the number label. The heat map indicates the classification probability for each number label. (b) 1.0 9 8 7 6 5 4 3 2 1 0 0 100 Trials 200 300 Only 6 All 3.5 0.6 6 \|2 0.4 \|z 0.8 0.2 2.0 Prob. Number (a) 0.0 0 Step 80 Figure B.2: (a) The result of classification of the final state T = 80 of the inference for the image of ‘6’. (b) The time evolution of the distance from the concept of ‘6’. The condition excluding trials that the activity pattern switched to different numbers is expressed in red, and condition containing all trials is expressed in gray. 16 An image of ‘6’ was used as the initial value of the inference. The discriminator classified the final state of 193 trials out of 300 trials as ‘6’. We consider the effect of other labels as the cause of the neural activity patterns approaching mismatched concepts. Taking an example of the label ‘6’, we first measured the distances between each neural activity pattern and the concept of ‘6’ in two conditions. We included only neural activity patterns reminded inside the cluster of ‘6’ in one condition, and all neural activity patterns in the latent space in the other. Then, we averaged these distances in each condition and compared their means. The average trajectories are compared in fig. B.2b. Red shows the average of only the trial with the final state identified as ‘6’, and gray shows the average of all trials. As shown in the figure above, the neural activity patterns in both conditions approached to the concept before moving to the corresponding memories. This result suggests that the presence of other labels doesn’t cause the neural activity patterns to move away from the concept. References [1] Amari, S. (1977). Neural theory of association and concept-formation. Biological Cybernetics, 26, 175–185. [2] Amit, D. J., Gutfreund, H., & Sompolinsky, H. (1985). Spin-glass models of neural networks. Physical Review A, 32, 1007. [3] Arulkumaran, K., Creswell, A., & Bharath, A. A. (2016). Improving sampling from generative autoencoders with markov chains. arXiv preprint arXiv:1610.09296, . [4] Bengio, Y., Mesnil, G., Dauphin, Y., & Rifai, S. (2013). Better mixing via deep representations. In Proceedings of the 30th International Conference on Machine Learning (pp. 552–560). [5] Brincat, S. L., & Connor, C. E. (2006). Dynamic shape synthesis in posterior inferotemporal cortex. Neuron, 49, 17–24. [6] Chollet, F. et al. (2015). Keras. https://github.com/fchollet/keras. [7] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Generative adversarial nets. In Advances in Neural Information Processing Systems 27 (pp. 2672–2680). 17 [8] Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., & Lerchner, A. (2017). β-vae: Learning basic visual concepts with a constrained variational framework. In Proceedings of the 6th International Conference on Learning Representations. [9] Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79, 2554–2558. [10] Karakida, R., Igarashi, Y., Nagata, K., & Okada, M. (2013). Inter-layer correlation in a feed-forward network with intra-layer common noise. Journal of the Physical Society of Japan, 82, 064007. [11] Katori, Y., Otsubo, Y., Okada, M., & Aihara, K. (2013). Stability analysis of associative memory network composed of stochastic neurons and dynamic synapses. Frontiers in Computational Neuroscience, 7. [12] Kingma, D., & Ba, J. (2014). Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations. [13] Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. In Proceedings of the 2nd International Conference on Learning Representations. [14] Kullback, S., & Leibler, R. A. (1951). On information and sufficiency. The Annals of Mathematical Statistics, 22, 79–86. [15] Lee, H., Grosse, R., Ranganath, R., & Ng, A. Y. (2009). Convolutional deep belief networks for scalable unsupervised learning of hierarchical representations. In Proceedings of the 26th Annual International Conference on Machine Learning (pp. 609–616). [16] Li, J., Monroe, W., Shi, T., Ritter, A., & Jurafsky, D. (2017). Adversarial learning for neural dialogue generation. In The 2017 Conference on Empirical Methods on Natural Language Processing.. [17] Liu, J., Harris, A., & Kanwisher, N. (2002). Stages of processing in face perception: an meg study. Nature Neuroscience, 5, 910–916. 18 [18] Matsumoto, N., Okada, M., Sugase-Miyamoto, Y., & Yamane, S. (2005). Neuronal mechanisms encoding global-to-fine information in inferiortemporal cortex. Journal of Computational Neuroscience, 18, 85–103. [19] Murata, S., Otsubo, Y., Nagata, K., & Okada, M. (2014). Oscillations in spurious states of the associative memory model with synaptic depression. Journal of the Physical Society of Japan, 83, 124004. [20] Nagano, Y., Karakida, R., Watanabe, N., Aoyama, A., & Okada, M. (2016). Input response of neural network model with lognormally distributed synaptic weights. Journal of the Physical Society of Japan, 85, 074001. [21] Nickel, M., & Kiela, D. (2017). Poincaré embeddings for learning hierarchical representations. In Advances in Neural Information Processing Systems 30. [22] Okada, M. (1996). Notions of associative memory and sparse coding. Neural Networks, 9, 1429–1458. [23] Radford, A., Metz, L., & Chintala, S. (2015). Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, . [24] Rezende, D. J., Mohamed, S., & Wierstra, D. (2014). Stochastic backpropagation and approximate inference in deep generative models. arXiv preprint arXiv:1401.4082, . [25] Rifai, S., Dauphin, Y. N., Vincent, P., Bengio, Y., & Muller, X. (2011). The manifold tangent classifier. In Advances in Neural Information Processing Systems 24 (pp. 2294–2302). [26] Saito, M., Matsumoto, E., & Saito, S. (2017). Temporal generative adversarial nets with singular value clipping. In 2017 IEEE International Conference on Computer Vision. [27] Saxe, A. M., McClelland, J. L., & Ganguli, S. (2014). Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. In Proceedings of the 3nd International Conference on Learning Representations. [28] Serban, I. V., Sordoni, A., Lowe, R., Charlin, L., Pineau, J., Courville, A. C., & Bengio, Y. (2017). A hierarchical latent variable encoder-decoder model for generating dialogues. In AAAI (pp. 3295–3301). 19 [29] Sugase, Y., Yamane, S., Ueno, S., & Kawano, K. (1999). Global and fine information coded by single neurons in the temporal visual cortex. Nature, 400, 869. [30] Theano Development Team (2016). Theano: A Python framework for fast computation of mathematical expressions. arXiv e-prints, abs/1605.02688. URL: http://arxiv.org/abs/1605.02688. [31] Tomczak, J. M., & Welling, M. (2017). VAE with a VampPrior. arXiv preprint arXiv:1705.07120, . [32] Vondrick, C., Pirsiavash, H., & Torralba, A. (2016). Generating videos with scene dynamics. In Advances In Neural Information Processing Systems 29 (pp. 613–621). [33] Walker, J., Doersch, C., Gupta, A., & Hebert, M. (2016). An uncertain future: Forecasting from static images using variational autoencoders. In The 14th European Conference on Computer Vision (pp. 835–851). [34] Yu, L., Zhang, W., Wang, J., & Yu, Y. (2017). Seqgan: Sequence generative adversarial nets with policy gradient. In AAAI (pp. 2852–2858). 20	9cs.NE
Optimal choice among a class of nonparametric estimators of the jump rate for piecewise-deterministic Markov processes arXiv:1506.07722v2 [math.ST] 23 May 2016 Romain Azaı̈s∗ and Aurélie Muller-Gueudin† Inria Nancy-Grand Est, Team BIGS, Institut Élie Cartan de Lorraine, Vandoeuvre-lès-Nancy, France Abstract A piecewise-deterministic Markov process is a stochastic process whose behavior is governed by an ordinary differential equation punctuated by random jumps occurring at random times. We focus on the nonparametric estimation problem of the jump rate for such a stochastic model observed within a long time interval under an ergodicity condition. We introduce an uncountable class (indexed by the deterministic flow) of recursive kernel estimates of the jump rate and we establish their strong pointwise consistency as well as their asymptotic normality. We propose to choose among this class the estimator with the minimal variance, which is unfortunately unknown and thus remains to be estimated. We also discuss the choice of the bandwidth parameters by cross-validation methods. Keywords: Cross-validation · Jump rate · Kernel method · Nonparametric estimation · Piecewisedeterministic Markov process Mathematics Subject Classification (2010): 62M05 · 62G20 · 60J25 Contents 1 Introduction 2 2 Problem formulation 2.1 Definition and notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 4 6 3 Estimation procedure 3.1 Inference for the inter-jumping times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Optimal estimation of the jump rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 How to choose bandwidth parameters α and β? . . . . . . . . . . . . . . . . . . . . . . . . 7 7 9 11 4 Simulation study 4.1 Estimation algorithm . . . 4.2 TCP-like process . . . . . 4.3 Bacterial motility . . . . . 4.4 Fatigue crack propagation 13 13 14 15 17 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A Ergodicity and invariant measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 B Proof of Theorem 3.3 19 B.1 Sketch of the proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 B.2 Remainder term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 B.3 Martingale term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 C Proof of Proposition 3.14 ∗ Electronic † Electronic 26 address: romain.azais@inria.fr; Corresponding author address: aurelie.gueudin@univ-lorraine.fr 1 1 Introduction Piecewise-deterministic Markov processes (PDMP’s in abbreviated form) have been introduced in the literature by Davis in [17] as a general class of continuous-time non-diffusion stochastic models, suitable for modeling deterministic phenomena in which the randomness appears as point events. The motion of a PDMP may be defined from three local characteristics: the flow Φ(x, t), the jump rate λ(x) and the transition measure Q(x, dy). Starting from some initial value X0 , the process evolves in a deterministic way following Φ(X0 , t) until the first jump time T1 which occurs either when the flow reaches the boundary of the state space or before, in a Poisson-like fashion with non homogenous rate λ(Φ(X0 , t)). In both cases, the post-jump location of the process at time T1 is governed by the transition distribution Q(Φ(X0 , T1 ), dy) and the motion restarts from this new point as before. This family of stochastic models is well-adapted for tackling various problems arising for example in biology [9, 16, 30, 31, 32, 33, 34], in neuroscience [22] or in reliability [7, 14, 13, 18]. Indeed, most of applications involving both deterministic motion and punctual random events may be modeled by a PDMP. Typical examples are growth-fragmentation models composed of deterministic growths followed by random losses. For example [33], the size of a cell grows exponentially in time, next the cell divides into two offsprings whose size is about half size of the parent cell, and so on. Proposing efficient statistical methods for this class of stochastic models is therefore of a great interest. Nevertheless, the very particular framework involving both deterministic motion and punctual random jumps imposes to consider specific methods. For instance, the authors of [5] have shown that the wellknown multiplicative intensity model developed by Aalen [1] for estimating the jump rate function does not directly apply to PDMP’s. Alternative approaches should be thus proposed. In the present paper, we focus on the recursive nonparametric estimation of the jump rate of a PDMP from the observation of only one trajectory within a long time interval. More precisely, the purpose of this work is to show how one may obtain by kernel methods a class of consistent estimators for the jump rate, and how one may choose among this class in an optimal way. To the best of our knowledge, the nonparametric estimation of the jump rate in a general framework has never been investigated. As PDMP’s may model a large variety of problems, some methods have been developed by many authors for their statistical inference. As presented before, the randomness of a PDMP is governed by two characteristics: the transition kernel Q(x, dy) and the jump rate λ(x). As a consequence, two main questions arise in the estimation problem for such a process, namely the statistical inference for both these features. On the one hand, a few papers investigate some nonparametric methods for estimating the transition function of a PDMP either for a specific model [13] or in a more general setting for a d-dimensional process [3]. On the other hand the estimation of the jump rate λ(x) or of the associated density function has been more extensively studied by several authors. Without attempting to give an exhaustive survey of the literature on this topic, one may refer the reader to [5, 19, 20, 25, 28] and the references therein. In the book [25], the author studies likelihood processes for observation of PDMP’s which could lead to inference methods in a parametric or semi-parametric setting. The papers [19, 20] deal with the nonparametric estimation for some PDMP’s used in the modeling of a size-structured population observed along a lineage tree. In both these articles, the authors rely on the specific form of the features of the process of interest in order to derive the asymptotic behavior of their estimation procedure. These techniques have been generalized in [28] to introduce a nonparametric method for estimating the jump rate in a specific class of one-dimensional PDMP’s with monotonic motion and deterministic breaks, that is to say when the transition measure Q(x, dy) is a Dirac mass at some location depending on x. The procedures developed in these papers [19, 20, 28] are obviously of a great interest but strongly use the particular framework involved in the investigated models and are thus not well adapted in a more general setting. In [5], the authors show that the famous multiplicative intensity model only applies for estimating the jump rate of a modified version of the underlying PDMP. This leads to a statistical method for approximating the conditional density associated with the jump rate for a process defined on a bounded metric state space. A main difficulty throughout the present paper and the articles [3, 5] is related to the presence of deterministic jumps when the path reaches the boundary of the state space. This feature is often used for modeling a deterministic switching when the quantitative variable rises over a certain threshold [2, 24]. In a statistical point of view, the interarrival times are therefore right-censored by a deterministic clock depending on the state space, which leads to some technical difficulties. We would like to emphasize that the techniques developed in the references [19, 20, 25, 28] do not take into account the likely presence of 2 forced jumps in the dynamic. One may also find in the literature a few papers [6, 8, 10, 11] which focus on the estimation of various functionals for this family of stochastic models. More precisely, the authors of [10, 11] provide numerical methods for the expectations and for the exit times of PDMP’s. In addition, the article [8] deals with a PDMP introduced for modeling the temporal evolution of exposure to a food contaminant and consider a simulation-based statistical inference procedure for estimating some functionals such as first passage times. Plug-in methods have been studied in [6] for a non-ergodic growth-fragmentation model which is absorbed under a certain threshold. In many aspects our approach and these papers are different and complementary. Indeed they are devoted to the estimation of some functionals of PDMP’s while we focus on the direct estimation of the primitive data of such a process. In this article, we introduce a three-dimensional kernel estimator computed from the observation of the embedded Markov chain of a PDMP composed of the post-jump locations Zn and the travel times Sn+1 along the path Φ(Zn , t). We establish its pointwise consistency as well as its asymptotic normality in Theorem 3.3. The estimate that we consider is recursive: it may be computed in real-time from sequential data, which may be relevant in many applications. We deduce from this result two first corollaries about the nonparametric estimation of the conditional density f (x, t) of the interarrival time Sn+1 at time t conditionally on the event {Zn = x} (see Corollary 3.8) and its survival function G(x, t) = P(Sn+1 > t \| Zn = x) (see Corollary 3.9). We also investigate in Corollary 3.10 the asymptotic behavior of an estimator for the composed function λ ◦ Φ(x, t) obtained as the ratio f (x, t)/G(x, t). We derive in (18) an uncountable class (indexed by the states ξ hitted by the reverse flow Φ(x, −t) for some time t) of consistent estimates of the jump rate λ(x). In other words, for each ξ = Φ(x, −t), we get a good estimate of λ(x). We show how one may choose among this class of estimators by minimizing their asymptotic variance. We state in (19) that this procedure is equivalent to maximize the criterion ν∞ (ξ)G(ξ, τx (ξ)) along the curve Φ(x, −t), i.e. ξ = Φ(x, −τx (ξ)), where ν∞ (ξ) denotes the invariant measure of the post-jump locations Zn and τx (ξ) is the only deterministic time to reach x following Φ(ξ, t). The choice of this criterion is far to be obvious without precisely computing the limit variance in the central limit theorem presented in Corollary 3.10. Indeed, a naive criterion to maximize is the invariant distribution ν∞ (ξ) along Φ(x, −t): the larger ν∞ (ξ) is, the larger the number of data around ξ is and the higher the quality of the estimation should be. Nonetheless, this simple criterion does not take into account that the estimate also depends on the time of interest τx (ξ) (see Remark 3.11). This question is also investigated from a numerical point of view in Subsection 4.2 where we show on synthetic data that the choice of the criterion ν∞ (ξ)G(ξ, τx (ξ)) is better than the naive one ν∞ (ξ). The bandwidths in kernel methods are free parameters that exhibit a strong influence on the quality of the estimation. We discuss the choice of the bandwidth parameters by a classic procedure that consists in minimizing the Integrated Square Error, computed here along the reverse flow Φ(x, −t): we introduce a cross-validation procedure in this Markov setting and we prove its convergence in Propositions 3.14 and 3.15. Finally, we would like to highlight that the regularity conditions that we impose are non restrictive. In particular, neither the deterministic exit time from the state space is assumed to be a bounded function, nor the transition kernel is supposed to be lower-bounded, as is the case in [5] (see eq. (2) and Assumptions 2.4). In addition, the forms of the transition measure and of the deterministic flow are not specified. The sequel of the paper is organized as follows. We begin in Section 2 with the precise formulation of our framework (see Subsection 2.1) and the main assumptions that we need in this article (see Subsection 2.2). Section 3 is devoted to the presentation of the statistical procedure and the related results of convergence. More precisely, a three-dimensional kernel estimator for the inter-jumping times is introduced and investigated in Subsection 3.1. We derive a class of estimators of the jump rate and propose how to choose among it in Subsection 3.2. The crucial choice of the bandwidth parameters is studied in Subsection 3.3. Finally, a self-contained presentation of the whole estimation procedure is provided in Subsection 4.1, then illustrated in the sequel of Section 4 on three different application scenarios, with various sample sizes and state space dimensions, involving both simulated and real datasets. More precisely, we focus on the TCP window size process used for modeling data transmission over the Internet in Subsection 4.2. Estimation of bacterial motility is tackled in Subsection 4.3, while acceleration of fatigue crack propagation is considered in Subsection 4.4. The proofs and the technicalities are postponed in Appendix A, B and C at the end of the paper. 3 2 Problem formulation This section is devoted to the definition of the PDMP’s and to the presentation of the main assumptions that we impose in the paper. List of notations: In this paper B(Rd ) denotes the Borel algebra of Rd endowed with the Euclidean norm \| · \|. In addition, the Lebesgue measure on B(Rd ) is denoted by λd (dx), with the particular notation in the one-dimensional case λ1 (dx) = dx. The ball of Rd with radius r and center x is denoted by Bd (x, r). The closure of a set E is denoted by E while ∂E stands for its boundary. 2.1 Definition and notation The motion of a PDMP may be described as the solution of an ordinary differential equation Φ punctuated at random times by random jumps governed by a transition kernel Q (see Figure 1). The random jumps occur either when the deterministic motion hits the boundary of its state space, or before, with non homogeneous rate λ taken along the curve defined by the differential equation Φ. Figure 1: Schematic path of a PDMP. This paper is devoted to the nonparametric estimation of the statedependent rate λ, that governs the spontaneous generation of jumps, from a long-time trajectory. More precisely, a PDMP (Xt )t≥0 on (Rd , B(Rd )) is defined from its three local characteristics (λ, Q, Φ): • Φ : Rd × R → Rd is the deterministic flow. It satisfies the semigroup property, ∀ x ∈ Rd , ∀ t, s ∈ R, Φ(x, t + s) = Φ(Φ(x, t), s). • λ : Rd → R+ is the jump rate. • Q : (Rd , B(Rd )) → [0, 1] is the transition kernel. We define the deterministic exit times of E for the flow and for the reverse flow as, for any x ∈ E, t+ (x) = inf {t > 0 : Φ(x, t) ∈ ∂E} and t− (x) = inf {t > 0 : Φ(x, −t) ∈ ∂E} . (1) In all the sequel, we consider a PDMP (Xt )t≥0 evolving on an open subset E of Rd . In this context, we impose as usual [17, (24.8) Standard conditions] that, Z ε ∀ x ∈ E, ∃ ε > 0, λ(Φ(x, t)) dt < +∞, 0 and ∀ x ∈ E, Q(x, E \ {x}) = 1. (2) In addition, we restrict ourselves to the case where the transition kernel Q admits a density with respect to the Lebesgue measure, Z ∀ x ∈ Rd , ∀ A ∈ B(Rd ), Q(x, A) = Q(x, y) λd (dy). (3) A This assumption is natural when one considers multivariate real-valued PDMP’s, and is satisfied in various problems arising in biology [9], population dynamics [6] or in insurance [17, (21.11) An insurance model]. 4 Starting from any initial condition X0 = x, the motion of (Xt )t≥0 may be described as follows. The distribution of the first jump time T1 is given by, R ( t exp − 0 λ(Φ(x, s)) ds if t < t+ (x), ∀ t ≥ 0, P(T1 > t \| X0 = x) = (4) 0 else. In other words, the process jumps either when the flow hits the boundary of the state space at time t+ (x) or in a Poisson-like fashion with rate λ ◦ Φ before. Next the post-jump location Z1 at time T1 is defined through the transition kernel Q: for any test function ϕ, we have Z E ϕ(Z1 ) T1 , X0 = x = ϕ(u) Q(Φ(x, T1 ), du). The path between 0 and the first jump time T1 is given by, Φ(x, t) ∀ 0 ≤ t ≤ T1 , Xt = Z1 if t < T1 , else. Now starting from the post-jump location XT1 , one chooses the next inter-jumping time S2 = T2 − T1 and the future post-jump location Z2 in a similar way as before, and so on. One obtains a strong Markov process with (Tn )n≥0 as the sequence of the jump times (where T0 = 0 by convention). The inter-jumping times are defined by S0 = 0 and, for any integer n ≥ 1, Sn = Tn − Tn−1 . Finally (Zn )n≥0 denotes the stochastic sequence of the post-jump locations of (Xt )t≥0 , with for any n, Zn = XTn . All the randomness of the PDMP (Xt )t≥0 is contained in the stochastic sequence (Zn , Sn+1 )n≥0 which is a Markov chain. In addition, the post-jump locations (Zn )n≥0 also form a discrete-time Markov process on the state space E because of the condition (2). In this paper, νn denotes the distribution of the nth post-jump location Zn for any integer n ≥ 0, while P denotes its Markov kernel, ∀ x ∈ E, ∀ A ∈ B(Rd ), P(x, A) = = P(Zn+1 ∈ A \| Zn = x) Z S(x, dt) Q(Φ(x, t), A), R+ (5) where S stands for the conditional distribution of Sn+1 given Zn for any n, ∀ x ∈ E, ∀ t ≥ 0, S(x, (t, +∞)) = = P(Sn+1 > t \| Zn = x) R ( t exp − 0 λ(Φ(x, s))ds 0 if t < t+ (x), (6) else, in light of (4). We would like to highlight that the conditional distribution S(x, ·) is absolutely continuous with respect to the unidimensional Lebesgue measure on (0, t+ (x)) with sometimes a singular component at t+ (x), Z f (x, t) dt + S(x, A ∩ {t+ (x)}), ∀ A ∈ B(R+ ), S(x, A) = (7) A∩(0,t+ (x)) where the conditional density f may be obtained by deriving (6), Z ∀ x ∈ E, ∀ 0 < t < t (x), f (x, t) = λ(Φ(x, t)) exp − + t λ(Φ(x, s))ds . (8) 0 In all the sequel G stands for the conditional survival function associated with f , that is, ∀ x ∈ E, ∀ 0 < t < t+ (x), G(x, t) = S(x, (t, +∞)), (9) where S is defined in (6). As highlighted before, the process (Zn , Sn+1 )n≥0 forms a Markov chain on the set F defined by [ F = {x} × [0, t+ (x)]. (10) x∈E R denotes the transition kernel of this process, ∀ (x, t) ∈ F, ∀ A × B ∈ B(Rd × R+ ), R((x, t), A × B) = = P(Zn+1 ∈ A, Sn+2 ∈ B \| Zn = x, Sn+1 = t) Z Q(Φ(x, t), dξ) S(ξ, B), (11) A and, for any n, µn denotes the distribution of the couple (Zn , Sn+1 ). 5 2.2 Assumptions The main assumption that we impose in the present paper is a condition of ergodicity on the Markov chain (Zn )n≥0 . This property is often a keystone in statistical inference for Markov processes and may be directly imposed [3, 5] or established [28] from the primitive features of the data. Assumption 2.1 There exists a distribution ν∞ on E such that, for any initial distribution ν0 = δ{x} , x ∈ E, lim kνn − ν∞ kT V = 0, n→+∞ where k · kT V stands for the total variation norm. This assumption may be checked directly on the Markov kernel P of (Zn )n≥0 from the existence of a Foster-Lyapunov’s function or Doeblin’s condition for instance [29, Theorem 16.0.2]. In the following remark we establish a first property of the sequence (νn )n≥0 and of its limit ν∞ . Remark 2.2 Since the transition kernel Q is assumed to be absolutely continuous with respect to the Lebesgue measure (3), the kernel P given by (5) of the post-jump locations (Zn )n≥0 also admits a density. As a consequence, for any integer n, the distribution νn of Zn and thus the invariant measure ν∞ introduced in Assumption 2.1 admit a density on the state space E. For the sake of clarity, we write ν∞ (dx) = ν∞ (x)λd (dx) with a slight abuse of notation. We add some regularity conditions on the main features of the process to show the convergence of the estimates in Theorem 3.3. Assumptions 2.3 • The sup-norms kQk∞ and kf k∞ are finite. These conditions are used in the proof of Theorem 3.3 to find an upper bound of the non diagonal terms of some square variation process of interest. • The functions Q and f are Lipschitz, ∀ x, y ∈ E, ∀ 0 < s < t < t+ (x) ∧ t+ (y), \|f (x, t) − f (y, s)\| ≤ [f ]Lip (\|t − s\| + \|y − x\|), ∀ x ∈ E, ∀ y, z ∈ E, \|Q(x, y) − Q(x, z)\| ≤ [Q]Lip \|y − z\|. These conditions are used in the proof of Theorem 3.3 to control the diagonal terms of the same variation process and to study the convergence of some remainder terms. • The survival function G is Lipschitz, ∀ x, y ∈ E, ∀ 0 < t < t+ (x) ∧ t+ (y), \|G(x, t) − G(y, t)\| ≤ [G]Lip \|x − y\|. This condition is used in the proof of Theorem 3.3 to investigate the convergence of the remainder terms. • The deterministic exit time t+ is continuous. This condition is used to find some admissible initial bandwidths v0 and w0 . Finally, we consider an additional condition on both the transition kernels P and Q and the flow Φ in order to ensure the Lipschitz mixing property of the Markov chain (Zn )n≥0 . This will be sufficient to establish the almost sure convergence to 0 of the remainder term with the adequate rate in the proof of Theorem 3.3. Assumptions 2.4 The transition kernel P of the Markov chain (Zn )n≥0 satisfies, for some a1 ≥ 1 and a2 < 1, Z 2 ∀ (x, y) ∈ E , \|u − v\|a1 P(x, du)P(y, dv) ≤ a2 \|x − y\|a1 . Rd ×Rd In addition, the composed function Q(Φ(·, ·), ·) belongs to the regularity class Li(r1 , r2 ) defined by \|Q(Φ(x1 , t1 ), y1 )−Q(Φ(x2 , t2 ), y2 )\| = O (\|(x1 , t1 , y1 ) − (x2 , t2 , y2 )\|r2 (\|(x1 , t1 , y1 )\|r1 + \|(x2 , t2 , y2 )\|r1 + 1)) , for some positive numbers r1 and r2 satisfying 2(r1 + r2 ) ≤ a1 . 6 3 Estimation procedure 3.1 Inference for the inter-jumping times n For any integer n, we introduce the σ(Z0 , S1 , . . . , Zn−1 , Sn )-measurable functions νb∞ defined on E, and n n b b F and G defined on the interior F̊ (let us recall that the set F is given in (10)), by, ∀ (x, t) ∈ F̊ , Fbn (x, t) = ∀ (x, t) ∈ F̊ , Gbn (x, t) = n ∀ x ∈ E, νb∞ (x) = n−1 Si+1 − t 1X 1 Zi − x K , K 1 d n i=0 vid wi vi wi n−1 1X 1 Zi − x 1{Si+1 >t} , K d n i=0 vid vi n−1 1X 1 Zi − x , Kd n i=0 vid vi (12) (13) where Kp denotes a kernel function on Rp , p ∈ {1, d}, and the bandwidths are defined for any integer k by vk = v0 (k + 1)−α and wk = w0 (k + 1)−β for some α, β > 0 and initial positive values v0 and w0 . It should be already noted that these quantities are of a great interest in the statistical study of the inter-jumping times of the PDMP (Xt )t≥0 . Indeed, we will see that: • The ratio b n (x,t) F n (x) ν b∞ estimates the conditional density f (x, t) defined in (8). • The ratio bn (x,t) G n (x) ν b∞ estimates the conditional survival function G(x, t) defined in (9). • The ratio b n (x,t) F bn (x,t) G estimates the composed function λ(Φ(x, t)). n In addition, as the name suggests, we will state that νb∞ (x) is a good estimate of the density ν∞ (x) of the unique invariant distribution of the post-jump locations, which is relevant in the estimation problem for PDMP’s but has already been investigated in [3, Proposition A.11]. In all the sequel, we impose a few assumptions on both the kernel functions K1 and Kd . Assumptions 3.1 For any p ∈ {1, d}, the kernel function Kp is assumed to be a nonnegative smooth function satisfying the following conditions: • The sup-norm kKp k∞ is finite. R • Rp Kp dλp = 1. • supp Kp ⊂ Bp (0p , δ). Together with well-chosen initial bandwidths, this condition avoids to compute the kernel estimator from data located at the boundary of the state space F (see also Remark 3.4). • Only for p = d. The function Kd is Lipschitz, ∀ x, y ∈ Rd , \|Kd (x) − Kd (y)\| ≤ [K]Lip \|x − y\|. This condition is used to show eq. (57) in the proof of Proposition 3.14. R Remark 3.2 In particular, Assumptions 3.1 ensure that, for any p ∈ {1, d}, τp2 = Rp K2p dλp is finite. This is used to find an upper bound ofRthe non diagonal terms of the variation process in the proof of Theorem 3.3. In addition, the integral Rp \|u\|Kp (u)dλp (u) is also finite, which is needed to establish the almost sure convergence of the remainder terms in the proof of the same result. In the sequel the admissible set for the bandwidth parameters α and β is given by A = (α, β) ∈ R2 : α > 0, β > 0, αd + β < 1, αd + β + 2 min(α, β) > 1 . In this part, our main result is obtained from the use of vector martingales and is stated in the following theorem. 7 Theorem 3.3 For any couple (x, t) ∈ F̊ such that ν∞ (x)f (x, t) > 0, for any (α, β) ∈ A and (v0 , w0 ) such that t+ (ξ), (14) t + w0 δ < inf ξ∈Bd (x,v0 δ) where δ appears in the third item of Assumptions 3.1, we have the almost sure convergence,   n   Fb (x, t) ν∞ (x)f (x, t) a.s n  Gb (x, t)  −→  ν∞ (x)G(x, t)  n ν∞ (x) νb∞ (x) and the asymptotic normality,  n    Fb (x, t) ν∞ (x)f (x, t) 1−αd−β d  Gbn (x, t)  −  ν∞ (x)G(x, t)  −→ n 2 N (03 , Σ(x, t, α, β)), n ν∞ (x) νb∞ (x) where the variance-covariance matrix Σ(x, t, α, β) is degenerate with only one positive term at position (1, 1). Σ(x, t, α, β) is defined in (52). 2 Proof. The proof is stated in Appendix B. Remark 3.4 The existence of a couple (v0 , w0 ) satisfying (14) is obvious whenever the exit time t+ is continuous (see Assumptions 2.3). This condition ensures that all the inter-jumping times used in the calculus of Fbn and Gbn are not obtained from forced jumps when the process reaches the boundary of the state space. In the case where t+ (x) = ∞, it is obvious that no interarrival times are right-censored. The consistency and the asymptotic normality are therefore still accurate without any condition on (v0 , w0 ). Remark 3.5 In Theorem 3.3, the choice (14) of the initial bandwidths v0 and w0 is locally dependent on the point of interest. This may appear restrictive but may be avoided by considering the elements of n o C = C × [0, T ] : C is a compact subset of E and T < inf t+ (x) . (15) x∈C Indeed, for any C × [0, T ] ∈ C, there always exists a couple (v0 , w0 ) such that T + w0 δ < inf inf x∈C ξ∈Bd (x,v0 δ) t+ (ξ). Thus, (v0 , w0 ) satisfies (14) for any point (x, t) ∈ C × [0, T ]. Remark 3.6 The variance-covariance matrix appearing in the asymptotic normality presented in Theorem 3.3 is degenerate with only the component (1, 1) positive. It means that the rate of the estimators n νb∞ and Gbn is faster than the one of Fbn . This is straightforward because Fbn is obtained by smoothing n the empirical distribution of the data both in the spatial and temporal directions contrary to νb∞ and Gbn . The proof of the previous result may be adapted to show the two-dimensional central limit theorem, with α such that αd < 1 and α(d + 2) > 1: n 1−αd ν∞ (x)G(x, t) d Gb (x, t) − −→ N (02 , Σ0 (x, t, α)), n 2 n ν∞ (x) νb∞ (x) where Σ0 (x, t, α) is a diagonal 2 × 2-matrix. The keystone to state this convergence is the behavior given in (50) of the hook of a vector martingale. If we assume the geometric ergodicity of the Markov chain (Zn )n≥0 , one may also obtain the rate of n convergence of the variances of the estimates Fbn , Gbn and νb∞ uniformly on any compact subset of F̊ such that the parameters v0 and w0 may be uniformly chosen (see Remark 3.5). 8 Proposition 3.7 Let us assume that there exists b > 1 such that kνn −ν∞ kT V = O(b−n ). The geometric ergodicity is in particular ensured by Doeblin’s condition (see [29, Theorem 16.0.2]). Then, for any set C × [0, T ] ∈ C and for any couple (α, β) such that 2(αd + β) < 1, we have sup Var(Fbn (x, t)) = O(n2(αd+β)−1 ), = O(n2αd−1 ). (x,t)∈C sup n Var(Gbn (x, t)) + Var(b ν∞ (x)) (x,t)∈C Let us recall that C has been defined in (15). The rate of convergence for Fbn is faster than the one given in Theorem 3.3 whenever 3(αd + β) < 1. Proof. The proof is similar to the demonstrations of Proposition B.5 and Corollary B.6 of [3] and relies on the control of the covariance process of functionals of a geometrically ergodic Markov chain (see [29, Theorem 16.1.5]). 2 We present in the sequel some corollaries of Theorem 3.3 that are of interest in the estimation problem for the inter-jumping times. First, we define the estimator fbn (x, t) of f (x, t) by, Fbn (x, t) ∀ (x, t) ∈ F̊ , fbn (x, t) = n , νb∞ (x) with the usual convention 0/0 = 0. We have the following result of convergence. Corollary 3.8 For any couple (x, t) ∈ F̊ such that ν∞ (x)f (x, t) > 0, for any (α, β) ∈ A and (v0 , w0 ) satisfying (14), we have 1−αd−β τ12 τd2 f (x, t) a.s. d and n 2 fbn (x, t) −→ f (x, t) fbn (x, t) − f (x, t) −→ N 0, . (1 + αd + β)ν∞ (x) Proof. This result is a direct application of Theorem 3.3 and Slutsky’s lemma. 2 Another feature of interest for the inter-jumping times is the survival function G. One may estimate this quantity by, bn b n (x, t) = G (x, t) , ∀ (x, t) ∈ F̊ , G n (x) νb∞ with the convention 0/0 = 0. Some properties of convergence are stated in the following corollary. Corollary 3.9 For any couple (x, t) ∈ F̊ such that ν∞ (x)f (x, t) > 0, for any α such that αd < 1 and α(d + 2) > 1, and any v0 , we have 1−αd τd2 G(x, t) a.s. d n n b b 2 and n G (x, t) −→ G(x, t) . G (x, t) − G(x, t) −→ N 0, (1 + αd)ν∞ (x) Proof. The almost sure convergence is a direct application of Theorem 3.3. The central limit theorem is a consequence of the asymptotic normality established in Remark 3.6 and Slutsky’s lemma. 2 3.2 Optimal estimation of the jump rate n We propose to estimate the composed function λ ◦ Φ by the ratio λ[ ◦ Φ defined by, n Fbn (x, t) ∀ (x, t) ∈ F̊ , λ[ ◦ Φ (x, t) = , Gbn (x, t) (16) again with the convention 0/0 = 0. Pointwise convergence and asymptotic normality are again a consequence of Theorem 3.3. 9 Corollary 3.10 For any couple (x, t) ∈ F̊ such that ν∞ (x)f (x, t) > 0, for any (α, β) ∈ A and (v0 , w0 ) satisfying (14), we have n a.s. λ[ ◦ Φ (x, t) −→ λ ◦ Φ(x, t), and n 1−αd−β 2 n d λ[ ◦ Φ (x, t) − λ ◦ Φ(x, t) −→ N 0, τ12 τd2 λ ◦ Φ(x, t) (1 + αd + β)ν∞ (x)G(x, t) . Proof. This result is a direct application of Theorem 3.3 and Slutsky’s lemma together with (6), (8) and (9). 2 In all the sequel, we focus on the estimation of the rate λ (and not on the composed function λ ◦ Φ) n from the estimate λ[ ◦ Φ defined in (16). In particular, for some fixed value x ∈ E, we introduce a n bξ (x) of λ(x) indexed by the elements ξ of the curve Cx described by the reverse class of estimators λ flow Φ(x, −t), t ≥ 0, and we propose to choose an estimate from this class in an optimal way. Before proceeding further we define the notation Cx as Cx = {Φ(x, −t) : 0 ≤ t < t− (x)}. By definition (1) of t− (x) we have Cx ⊂ E. In addition, for any ξ ∈ Cx , we define τx (ξ) as the unique time satisfying Φ(ξ, τx (ξ)) = x, ξ ∈ Cx . Thus we have the following trivial result, ∀ ξ ∈ Cx , λ ◦ Φ(ξ, τx (ξ)) = λ(x). As a consequence, we propose to define a class of estimators of λ(x) by n o n bn [ Λn (x) = λ ξ (x) = λ ◦ Φ (ξ, τx (ξ)) : ξ ∈ Cx , (17) (18) n where λ[ ◦ Φ estimates λ ◦ Φ (see Corollary 3.10) and has already been defined in (16). By virtue of Corollary 3.10 and only using (17), one has, as n goes to infinity, 1−αd−β a.s. d 2 bn bn and n 2 ∀ ξ ∈ Cx , λ λ ξ (x) −→ λ(x) ξ (x) − λ(x) −→ N 0, σξ,α,β (x) , where the asymptotic variance is given by 2 σξ,α,β (x) = τ12 τd2 λ(x) , (1 + αd + β)κx (ξ) with κx (ξ) = ν∞ (ξ)G(ξ, τx (ξ)). (19) n bn In this paper, we choose to approximate λ(x) by the element λ ξ (x) ∈ Λ (x) minimizing the asymptotic 2 variance σξ,α,β (x). In other words, our optimal estimator of the jump rate λ(x) is obtained as bn bn λ ∗ (x) = λξ∗ (x), where ξ∗ = arg max κx (ξ). ξ∈Cx Remark 3.11 A good criterion should be to maximize the invariant measure ν∞ (ξ), ξ ∈ Cx . Indeed, if ν∞ (ξ) is large, a large frequency of post-jump locations around ξ may be available in the dataset. Nevertheless, roughly speaking, the quantities of interest f (ξ, τx (ξ)) and G(ξ, τx (ξ)) are well estimated if a large number of post-jump locations are around ξ together with inter-jumping times around τx (ξ). This naive criterion may be corrected by including the quality of the estimation at time τx (ξ), that is to say, by maximizing the product ν∞ (ξ)G(ξ, τx (ξ)). We also refer the reader to the simulation study presented in Section 4 and more precisely to Figure 8. The criterion κx (ξ), ξ ∈ Cx , is generally uncomputable from the known features of the PDMP and thus remains to be estimated. In light of Theorem 3.3 and by definition (19) of κx (ξ), we naturally propose to approximate this quantity by, bn ∀ ξ ∈ Cx , κ bn x (ξ) = G (ξ, τx (ξ)). 10 (20) As a consequence, we propose to estimate the jump rate λ(x) by its statistical approximation in Λn (x) maximizing the estimated criterion κ bx (ξ). More precisely, n b b∗ (x) = λ bnbn (x), λ ξ ∗ where ξb∗n = arg max κ bn x (ξ). (21) ξ∈Cx The high oscillations or alternatively the high smoothness of a kernel estimator with a ill-chosen bandwidth suggest that the choice of the parameter α appearing in κ bn x (see eq. (12) and (20)) is crucial at this maximization step. 3.3 How to choose bandwidth parameters α and β? This part is devoted to the choice of the bandwidth parameters α and β. The criteria that we introduce in this part to choose these features are defined as line integrals along the curve Cx . As a consequence, we need to ensure that this kind of quantity is well defined in our setting. Assumption 3.12 For any starting condition ξ ∈ E, the reverse flow t 7→ Φ(ξ, −t) defines a change of variable, that is to say, is a diffeomorphic mapping. It is common in the literature to minimize the Integrated Square Error (ISE) to choose the optimal bandwidths of a kernel estimator, in particular from dependent data: one may refer the reader to [23, 26, 27]. Another classical solution is to investigate the behavior of the Mean Integrated Square Error (MISE). In the framework of Gaussian dependent data, the authors of [15] have shown that the optimal bandwidths obtained by minimizing the ISE and the MISE are very close if the dependence is of short range, that is to say, if the covariance function is integrable (see [15, Theorem 2.2]). It should be noted that a geometric ergodic Markov chain satisfies this kind of condition (see [29, Theorem 16.1.5]). Let us recall that α appears at first in the computation of the estimated criterion κ bn x which we need n n b to maximize along the curve Cx . Indeed κ bx is computed (20) from the estimate G which implicitly depends on α. We propose to choose the bandwidth parameter α by minimizing the ISE associated with κ bx and defined by Z 2 ISEn = (b κn κ (α) x (ξ) − κx (ξ)) dξ Cx Z = κx (ξ)2 dξ + εn (22) κ (α), Cx where the function εn κ is given by Z εn (α) = κ Gbn (ξ, τx (ξ))2 dξ − 2 Z Gbn (ξ, τx (ξ)) κx (ξ)dξ. (23) Cx Cx One may remark that here the ISE is unusually computed along a curve of interest. In (22) the dependency on α only holds through the function εn κ . As a consequence the optimal parameter α minimizing this stochastic function (23) also minimizes the ISE. The function εn κ is generally not computable since the unknown quantity κx appears in its definition. As a consequence, we propose to estimate εn κ by cross-validation which is a popular technique for selecting the bandwidth that minimizes the ISE. The authors would like to highlight that cross-validation involves here two main difficulties. First the estimators are computed from dependent data which are not identically distributed. In addition, there is almost surely no data on the set of integration Cx whenever the dimension d is larger than 2. That is why we propose a specific procedure adapted to this framework. Before defining our cross-validation estimate of εn κ , we need to introduce the quantities Tx,ρ and θx . First Hx denotes the hyperplane orthogonal to Cx at x, that is, n o Hx = y ∈ Rd : y − x ⊥ ∇t Φ(x, 0) . In addition, for any ρ > 0, we introduce the notation Dx,ρ for Dx,ρ = Bd (x, ρ) ∩ Hx . 11 Furthermore Tx,ρ denotes the tube around Cx with radius ρ, [ Tx,ρ = Cy . (24) D y∈ x,ρ Finally, for any ξ ∈ Tx,ρ , θx (ξ) denotes the unique time such that Φ(ξ, θx (ξ)) ∈ Dx,ρ . In particular, ∀ ξ ∈ Cx , θx (ξ) = τx (ξ). (25) It should be noted that, for ρ small enough, Tx,ρ ⊂ E. We focus now on a cross-validation method for estimating the quantity εn bκn,en,ρ of εκn κ . The estimate ε e e et )t≥0 is defined from the observation of the embedded Markov chain (Zk , Sk+1 )k≥0 of another PDMP (X (independent on the first one (Xt )t≥0 and distributed according to the same parameters), by εbκn,en,ρ (α) = Z Cx 2 κ bn x (ξ) dξ − ne −1 + 1 X bn e ek ) 1Tx,ρ (Z ek ) 1 e G Z k , θ x (Z ek ),+∞) (Sk+1 ), d−1 (θx (Z n e π 2 ρd−1 k=0 2Γ d−1 2 (26) where Γ denotes as usually the Euler function. Some regularity conditions are necessary to investigate the asymptotic behavior of the cross-validation estimate εbκn,en,ρ in Proposition 3.14. Assumptions 3.13 • The sup-norm kν∞ k∞ is finite and ν∞ is Lipschitz, ∀ x, y ∈ E, \|ν∞ (x) − ν∞ (y)\| ≤ [ν∞ ]Lip \|x − y\|. • The deterministic exit time for the reverse flow t− is Lipschitz, ∀ x, y ∈ E, \|t− (x) − t− (y)\| ≤ [t− ]Lip \|x − y\|. • The flow Φ is Lipschitz, ∀ x, y ∈ E, ∀ t ∈ R, \|Φ(x, t) − Φ(y, t)\| ≤ [Φ]Lip \|x − y\|. • The sup-norm k∇t Φk∞ is finite and ∇t Φ is Lipschitz, ∀ x, y ∈ E, ∀ t ∈ R, \|∇t Φ(x, t) − ∇t Φ(y, t)\| ≤ [∇t Φ]Lip \|x − y\|. Proposition 3.14 Conditionally to σ(Z0 , S1 , . . . , Zn−1 , Sn ), we have lim εbκn,en,ρ (α) = εn κ (α) n e →∞ ρ→0 a.s. 2 Proof. The proof is stated in Appendix C. By virtue of Proposition 3.14, one may obtain an estimate of the optimal bandwidth parameter α arising in κ bn bκn,en,ρ for some small enough ρ and large enough n e. In addition, x by minimizing the quantity ε n bn by (16), (18) and (20), the quantity κ bx also appears in the calculus of the estimator λ ξ (x). In particular, bn the same choice of α may be done for computing the denominator κ bn x (ξ) of λξ (x). As a consequence it remains to choose in an optimal way the bandwidth parameters α and β arising in the formula (16) of bn the numerator Fbn (ξ, τx (ξ)) of λ ξ (x) (18). In a similar way as before, we propose to choose α and β by minimizing the ISE associated with Fbn (·, τx (·)) and computed along the curve Cx , Z 2 ISEn = Fbn (ξ, τx (ξ)) − F (ξ, τx (ξ)) dξ F (α, β) Cx Z F(ξ, τx (ξ))2 dξ + εn F (α, β), = Cx 12 (27) where Fbn implicitly depends on α and β, F(ξ, τx (ξ)) stands for ν∞ (ξ)f (ξ, τx (ξ)) and εn F is given by Z Z εn Fbn (ξ, τx (ξ))2 dξ − 2 Fbn (ξ, τx (ξ))F(ξ, τx (ξ))dξ. F (α, β) = Cx Cx As in the previous part we propose to estimate εn F by cross-validation from the observation of the ek , Sek+1 )k≥0 of another PDMP (X et )t≥0 . We define our estimate εbn,en,ρ1 ,ρ2 of εn embedded chain (Z F by F Z εbFn,en,ρ1 ,ρ2 (α, β) = Fbn (ξ, τx (ξ))2 dξ Cx − 2Γ d−1 2 n e ρ2 π d−1 2 n e −1 X ρd−1 1 k=0 +1 e ek , θx (Z ek ) 1Tx,ρ (Z ek ) 1 Fbn Z 1 (θx (Zek )− ρ2 ,θx (Zek )+ ρ2 ) (Sk+1 ). (28) 2 2 The convergence of εbFn,en,ρ,δ is investigated in Proposition 3.15. Proposition 3.15 Conditionally to σ(Z0 , S1 , . . . , Zn−1 , Sn ), we have lim εbFn,en,ρ1 ,ρ2 (α, β) = εn F (α, β) n e →∞ ρ1,2 →0 a.s. Proof. The proof is similar to the demonstration of Proposition 3.14 stated in Appendix C. 4 2 Simulation study In this section, we provide a self-contained presentation of the estimation procedure, as well as three application scenarios on both simulated and real datasets. 4.1 Estimation algorithm The sequel is devoted to the self-contained presentation of the estimation procedure provided in this paper. Precisely, we are interested in the estimation of λ(x) for some x ∈ E ⊂ Rd . 4.1.1 Preliminary computations These preliminary computations only require to manipulate the flow Φ and the state space E. • Compute the curve Cx = {Φ(x, −t) : t ≥ 0} ∩ E. • Choose ρ > 0 and compute Dx,ρ = Tx,ρ = Bd (x, ρ) ∩ {y ∈ Rd : y − x ⊥ ∇t Φ(x, 0)}, [ Cy . D y∈ x,ρ • For any ξ ∈ Tx,ρ compute θx (ξ) as the unique solution of Φ(ξ, θx (ξ)) ∈ Dx,ρ . The mapping τx is only defined on Cx ⊂ Tx,ρ by τx (ξ) = θx (ξ). 4.1.2 Preliminary estimates These preliminary computations require to choose kernel functions K1 and Kd on R and Rd , as well as two positive values v0 and w0 . • For any couple (α, β), compute n Fbα,β (ξ, t) = n−1 X Zi − ξ 1 Si+1 − t αd+β (i + 1) K K . 1 d vi wi nv0d w0 i=0 • Compute n−1 1 X Zi − ξ αd Gbαn (ξ, t) = (i + 1) K 1{Si+1 >t} . d vi nv0d i=0 13 4.1.3 Choice of bandwidth parameters by cross-validation ei , Sei+1 )0≤i≤en−1 , From the observation of two independent embedded chains (Zi , Si+1 )0≤i≤n−1 and (Z with n e n, one determines the optimal bandwidth parameters appearing in the preceding estimates. In practice, from only one trajectory of the underlying PDMP, one may divide the data into two categories: the largest one is used for the estimation, while the cross-validation step relies on the other one. • Compute αG = arg max α>0 Z ne −1 2 Γ d−1 + 1 X bn e 2 ek ) 1Tx,ρ (Z ek ) 1 e Gbαn (ξ, τx (ξ))2 dξ− Gα Zk , θx (Z ek ),+∞) (Sk+1 ). d−1 (θx (Z Cx n eπ 2 ρd−1 k=0 • Choose ρ2 > 0 and compute Z n (αF , β F ) = arg max Fbα,β (ξ, τx (ξ))2 dξ α,β>0 − Cx 2Γ d−1 2 n e ρ2 π 4.1.4 d−1 2 n e −1 X ρd−1 k=0 +1 n ek )1 ek , θx (Z ek ) 1Tx,ρ (Z e Fbα,β Z (θx (Zek )− ρ2 ,θx (Zek )+ ρ2 ) (Sk+1 ). 2 2 Estimation of λ(x) Finally, we compute the best estimate of λ(x) from all the preceding computations. • Compute ξb∗n = arg max GbαnG (ξ, τx (ξ)). ξ∈Cx F F G • From (α , β ), α and ξb∗n , compute bnF F ξb∗n , τx (ξb∗n ) F α ,β b b∗ (x) = . λ GbnG ξb∗n , τx (ξb∗n ) n α Remark 4.1 The time complexity of this algorithm depends on several parameters, namely the number n of observed jumps, the number n e of observed data for the cross-validation steps, and also the numbers Nξ , Nα and Nβ for discretizing the state spaces of ξ, α and β at each maximization procedure. It is easy to see that the cross-validation step is the most complex but remains polynomial, precisely in O(nNα Nβ (Nξ + n e)). 4.2 TCP-like process The application on which we focus in this part is a variant of the famous TCP window size process appearing in the modeling of the Transmission Control Protocol used for data transmission over the Internet and presented in [12]. This protocol has been designed to adapt to the traffic conditions of the network: for a connection, the maximum number of packets that can be sent is given by a random variable called the congestion window size. At each time step, if all the packets are successfully transmitted, then one tries to transmit one more packet until a congestion appears. The model presented in this part is two dimensional. For the sake of clarity, we will use the following notation: for any x ∈ R2 , x1 and x2 denote the components of x. We consider a PDMP (Xt )t≥0 evolving on the state space E = (0, 1)2 . The deterministic part of the model is defined from the flow Φ given by, ∀ x ∈ R2 , ∀ t ∈ R, Φ(x, t) = (x1 + t, x2 ). The jump rate λ is defined by, ∀ x ∈ R2 , λ(x) = x1 + x2 . The transition kernel Q is defined by, Z ∀ x ∈ E, ∀ A ∈ B(E), Q(x, A) ∝ A 14 u(1 − u)2/x1 −1 v(1 − v) du dv. Starting from (x1 , x2 ), the process evolves in the unit square, always to the right, until a jump appears either when the motion hits the boundary {(ξ1 , ξ2 ) : ξ1 = 1} or with the non homogeneous rate x1 +x2 +t before, that is according to a Weibull distribution. The two components of the post-jump location are independent and both governed by a Beta-distribution, in such a way that the process tends with a high probability to jump to the left of the location just before the jump. One obtains a TCP-like process (see Figure 2) for which the second dimension models the quality of the network (upper the second component is, higher the probability of a congestion is). [ Figure 2 approximately here ] The asymptotic behavior of the process may be represented by the invariant distribution ν∞ of the n post-jump locations. Since this quantity is unknown, we propose to show in Figure 3 its estimate νb∞ defined by (13) and computed from n = 20 000 observed jumps. [ Figure 3 approximately here ] We present here all the procedure for estimating the jump rate λ at the location x = (0.75, 0.5) for which the quality of the network is average. In this context, the class of estimators of λ(x) is indexed by the elements ξ ∈ Cx = (0, 0.75] × {0.5}. It should be noted that the invariant distribution at x is quite low (see Figure 3). Nevertheless our method is expected to work pretty well even in this unfavorable framework. In this simulation study, we assume that we observe the embedded Markov chain (Zn , Sn+1 )n≥0 until the 10 000th jump. The cross-validation procedure is computed from an additional chain, independent on the first one, and observed until the 1 000th jump. When boxplots are presented, they have been computed over 100 replicates. bn We begin with the choice of the bandwidth parameter α appearing in κ bn x (ξ) = G (ξ, τx (ξ)). The cross-validation procedure relies on the minimization of the estimate εbκn,en,ρ (α) which depends on the positive parameter ρ. We present in Figure 4 this quantity as a function of α and from different values of ρ. Fortunately, this new parameter seems to have little influence over the behavior of the estimation of the ISE along Cx . [ Figure 4 approximately here ] Now, from this α (denoted in the sequel αG ), we maximize the estimated criterion κ bn x (ξ) along the n b curve Cx (see Figure 5): we obtain the optimal point ξ∗ at which we will compute our estimator of λ(x). The crucial role of α at this maximization step is illustrated in Figure 6. [ Figure 6 approximately here ] We continue with the choice of the couple (α, β) implicitly appearing in the estimator Fbn (ξ, τx (ξ)). The optimal parameters (denoted in the sequel (αF , β F )) are obtained by minimizing the estimate εbFn,en,ρ1 ,ρ2 (α, β) of the related ISE (see Figure 7). [ Figure 7 approximately here ] We compute the estimators for different values of ξ ∈ Cx and with the optimal bandwidths αG F F and (α , β ). The related boxplots are presented in Figure 8. The procedure makes us able to choose the best index ξb∗n which most of the time corresponds with the estimate with least bias and variance. This proves the strong interest of the estimation algorithm developed in this paper. bn λ ξ (x) [ Figure 8 approximately here ] 4.3 Bacterial motility We present here a model of bacterial motility. Most motile bacteria move by the use of a flagellum or several flagella. The bacteria moves in the direction of the flagellum. A flagellum behaves like a rotary motor. Periodically, the flagellum changes its direction and results in reorientation of the bacteria. This allows bacteria to change direction. Bacteria can sense nutrients and move towards them. Additionally, they can move in response to temperature, light, etc. If the bacteria is in a favorable environment, the frequency of changes in direction is low. This intelligent behavior is allowed by the fact that the jump rate of the direction depends on the environment. For example, [30, 31, 34] propose models for 15 the trajectory of the bacteria E. Coli. In particular, the author of [30] uses PDMP’s to describe the movement of bacteria under the influence of an external attractive chemical signal. We present here a variant of the model presented in [30]. The path of the bacteria is described by a PDMP (Xt )t≥0 evolving in the three-dimensional space state E = D × [0, 2π) with D the unit disk of R2 . For the sake of clarity, we will use the following notation: for any x ∈ R3 , x1 , x2 and x3 denote the components of x. In our case, (x1 , x2 ) is the position of the bacteria in the unit disk, and x3 is the direction of its flagellum. The bacteria moves in the direction of its flagellum. In other words, the flow is given by, ∀ x ∈ R3 , ∀ t ∈ R, Φ(x, t) = (x1 + t cos x3 , x2 + t sin x3 , x3 ). The bacteria changes its direction according to a jump rate function λ(x) = λ(x1 , x2 ) which depends on the environment only through the position (x1 , x2 ) and not on the current angle. The jump rate describes the interaction between the bacteria and its environment. When the bacteria changes its direction, the new direction is chosen preferentially in the direction of a more favorable environment. The transition kernel Q(x, ·) models this change of direction. In our case, we suppose the bacteria has no information a priori on the quality of the environment around itself. Thus, Q can be defined by, Z 1 ∀ x ∈ E, ∀A ∈ B([0, 2π[), Q(x, (x1 , x2 ) × A) = du. 2π A Starting from the position (x1 , x2 ) in D, the bacteria evolves in the direction x3 ∈ [0, 2π), until a jump appears either when the bacteria hits the boundary ∂D of its environment or with the rate λ (Φ(x, t)) = λ(x1 + t cos x3 , x2 + t sin x3 ) before. The post-jump direction is next randomly chosen in [0, 2π), and the bacteria continues its path in this new direction, and so on. A possible path of this model is presented in Figure 9. [ Figure 9 approximately here ] In this simulation study, the jump rate is taken constant equal to λ = 1. In other words, there is no interaction between the bacteria and its environment. Of course, this assumption is not taken into account in the estimation. Our goal is to estimate the jump rate λ in different points of E in order to check if it depends on the position. Actually, this provides an estimate of the influence of a likely external attractive signal on the bacteria. The estimation algorithm presented in this paper yields an estimate of the jump rate at the threedimensional state (x1 , x2 , θ). Nevertheless, one should take into account that the jump rate do not depend on the angle θ. We explain here the procedure to reduce the dimension and estimate the jump rate at the location (x1 , x2 ) ∈ D. bnbn (xθ ) at the • Let θ ∈ [0, 2π) be a fixed angle. Estimate the best index ξb∗n (θ) and the jump rate λ ξ∗ (θ) location xθ = (x1 , x2 , θ). bnbn (xθ ) estimates λ(x1 , x2 ). We reproduce the preceding step for any θ ∈ [0, 2π). We • For any θ, λ ξ∗ (θ) obtain both a connected trajectory around (x1 , x2 ) Cx1 ,x2 = ξb∗n (θ) : θ ∈ [0, 2π) , (29) and a class of optimal estimators n b bn Lx ,x = λ 1 2 ξ∗ (θ) n bξ : ξ ∈ Cx ,x (xθ ) : θ ∈ [0, 2π) = λ 1 2 indexed by elements of Cx1 ,x2 . • As the jump rate does not depend on the direction θ, we aggregate the estimators in Lx1 ,x2 to obtain the following jump rate estimate Z 2π bnbn (xθ ) dθ. b 1 , x2 ) = 1 λ (30) λ(x ξ∗ (θ) 2π 0 16 We investigate the estimation of the jump rate at the following 9 target points, (x1 , x2 ) ∈ {(0, 0), (−0.5, 0), (−0.5, 0.5), (−0.5, −0.5), (0, 0.5), (0, −0.5), (0.5, 0), (0.5, 0.5), (0.5, −0.5)}, from a trajectory of n = 100 000 changes of direction. In the sequel, the aggregated estimator (30) is approximated from a discretization of the interval [0, 2π) with step π/8. For each of the 9 target points, bnbn (xθ ) for θ in the discretization grid. we present in Figure 10 the boxplot of the estimates λ ξ (θ) ∗ [ Figure 10 approximately here ] As shown in Figure 11, the trajectories Cx1 ,x2 defined in (29) are very similar and close to the boundary ∂D. We explain this fact as follows. For each target state x = (x1 , x2 , θ), the optimal point bn ξb∗n (θ) maximizes κ bn x (ξ) = G (ξ, τx (ξ)) along Cx = {(x1 − t cos θ, x2 − t sin θ, θ) : t ≥ 0} ∩ D, where τx (ξ) satisfies x1 −τx (ξ) cos θ = ξ1 and x2 −τx (ξ) sin θ = ξ2 . Recall that Gbn (ξ, τx (ξ)) is an estimator of κx (ξ) = ν∞ (ξ)G(ξ, τx (ξ)). On the first hand, G(ξ, t) = exp(−t) for t small enough, then the only point that maximizes G(ξ, τx (ξ)) is ξ = x itself. On the other hand, there is an important number of data that are very close to the boundary ∂D (due to the fact that the bacteria jumps when it hits the boundary ∂D). Then the maximization of Gbn (ξ, τx (ξ)) is a compromise between ξ = x and ξ near the boundary. b 1 , x2 ) at the 9 target points are close together, and close to the true value Finally, the estimates λ(x λ = 1 with a small variance, which is a good indicator of a constant jump rate (see Figure 11). [ Figure 11 approximately here ] Now, we investigate the effect of the size n of the dataset on the quality of the estimation. We compare results of simulations from n = 100 000, n = 50 000 and n = 20 000 observed data. When n = 20 000, the trajectories Cx1 ,x2 are not all the time close to the boundary. It is due to the fact that, in this case, some areas of the state space are not well-covered by the observed post-jump locations. Then, if the target point is x = (x1 , x2 , θ), the optimal point ξb∗n (θ) ∈ Cx is obtained at the location x itself (see Figure 12). In addition, the estimated curves Cx1 ,x2 obtained from n = 50 000 data are very close to the ones computed from n = 100 000 observed jumps. [ Figure 12 approximately here ] Of course, the larger the number of data n is, the better the estimations (see Figures 13 and 14). In particular, it is difficult to conclude that the jump rate is constant from only n = 20 000 observed jumps because the variance of the estimates is too large. Nevertheless, a dataset of size 50 000 may be sufficient to conclude that the jump rate does not depend on the location. [ Figures 13 and 14 approximately here ] 4.4 Fatigue crack propagation Fatigue crack propagation is a stochastic phenomenon due to the inherent uncertainties originating from material properties and cyclic mechanical loads. As a consequence, stochastic processes offer an appropriate framework for modeling crack propagation. Fatigue life may be divided into crack initiation and two crack growth periods, namely the linear or stable regime described by Paris’ equation da = C (∆K(a))m , dN and the acceleration or unstable regime modeled by Forman’s equation C (∆K(a))m da = , dN (1 − R)Kc − ∆K(a) where a is the crack length, N is the time measured in number of loading cycles, R is the stress ratio, Kc represents the maximal value of the stress intensify factor required to induce failure and ∆K(a) is the range of the stress intensity factor given by πa −1/2 √ ∆K(a) = ∆σ cos πa, ω 17 where ω is the size of the test specimen and ∆σ is the stress range. In addition, m and C are two unknown material parameters. In this context, regime-switching models have been proposed to analyze fatigue crack growth data [7, 14]. The authors of [7] propose to estimate the transition time between Paris’ and Forman’s regimes assuming that the crack propagation follows the trajectory of some piecewisedeterministic Markov process. The estimation procedure is performed on Virkler’s dataset [35] in which 62 identical centre-cracked aluminium alloy specimens (ω = 152 mm) were tested under constant amplitude loading ∆σ = 48.28 MPa at a stress ratio R = 0.2. The number of loading cycles for the crack tip to advance a predetermined increment ∆a was recorded from an initial crack length of a0 = 9 mm to a final length of 49.8 mm. 68 crack growth histories were obtained from these tests (see Figure 15). [ Figure 15 approximately here ] In this paper, we assume that the crack growth propagation follows Paris’ equation with random parameters m and C before switching to Forman’s equation with another set of parameters m and C distributed according to some transition measure. The transition occurs at a random time T1 given by its survival function Z t λ ΦP P(T1 > t \| a0 = 9 mm, m, C) = exp − m,C (a0 , s) ds , 0 ΦP m,C where is the deterministic flow of Paris’ equation with parameters m and C. In the sequel, ΦP m,C will be computed from the Runge-Kutta method because there is no explicit solution to this differential equation. We propose to estimate the jump rate of this stochastic model from real crack growth data. It should be noted that the jump times are not directly observed in Virkler’s dataset but have been estimated in [7]. The jump rate is of significant importance in understanding the transition between the stable and unstable regimes of crack propagation, and has never been estimated in the framework of PDMP’s. We emphasize that Virkler’s dataset is composed of 68 independent experiments and thus does not directly follow the theoretical framework developed in this manuscript. Nevertheless, our estimation procedure is expected to perform pretty well in this more favorable context (independent curves instead of only one Markov path). Material parameters m and C are unobserved but have also been estimated from Virkler’s experiments in [7]. It is well-known in the literature that there is a strong linear relationship between these features. This characteristic has again been highlighted in [7] and is shown in Figure 16. In order to reduce the dimension and thus simplify the model, we propose to parametrize Paris’ equation with only one parameter, say m, while log(C) is obtained from this linear relationship. [ Figure 16 approximately here ] We focus on the estimation of λ(a) for some crack length a. For each parameter m of Paris’ equation there exists a unique deterministic time τm (a) such that Paris’ flow reaches a at time τm (a) starting from a0 = 9 mm. Criterion κa (m) and its maximum have been estimated for different values of a (see Figure 17). In particular, one may observe that the larger the target length a is, the larger the optimal parameter m. [ Figure 17 approximately here ] For each crack length a, we estimate the jump rate λ(a) from the parameter m maximizing the estimated criterion κa (m). We obtain the estimated function displayed in Figure 18. This curve is increasing as expected and describes the transition rate from the stable region of propagation towards the acceleration regime that leads to fracture. This makes us able to detect conditions of crack growth instability and could be used to predict the critical length in fatigue crack propagation with a given level of confidence. [ Figure 18 approximately here ] A Ergodicity and invariant measures In this section we present some preliminary and technical results about the invariant distributions of the underlying Markov chains. We begin with some properties of the Markov chain (Zn )n≥0 of the post-jump locations of the PDMP of interest. In particular, under Assumption 2.1, one may state the following result. 18 Proposition A.1 We have the following statements: • (Zn )n≥0 is ν∞ -irreducible, positive Harris-recurrent and aperiodic. • ν∞ is the unique invariant distribution of (Zn )n≥0 . Proof. The proof is similar to the demonstration of Proposition 4.2 of [4]. 2 These properties make us able to apply the law of large numbers to the Markov chain (Zn )n≥0 (see [29, Theorem 17.1.7]). Now we propose to focus on the sequence (Zn , Sn+1 )n≥0 which also forms a Markov process whose transition kernel is given by R in (11). Let us recall that µn denotes the distribution on the state space F defined by (10) of the couple (Zn , Sn+1 ), n ≥ 0. We define the measure µ∞ by, Z ∀ A ∈ B(Rd ), ∀ t ≥ 0, µ∞ (A × (t, +∞)) = ν∞ (dx) S(x, (t, +∞)), (31) A where the conditional distribution S is given by (6). In particular, µ∞ is the unique invariant distribution of (Zn , Sn+1 )n≥0 . Proposition A.2 We have the following statements: • For any initial distribution µ0 = δ{x,t} , (x, t) ∈ F , we have lim kµn − µ∞ kT V = 0. n→∞ • (Zn , Sn+1 )n≥0 is µ∞ -irreducible, positive Harris-recurrent and aperiodic. • µ∞ is the unique invariant distribution of (Zn , Sn+1 )n≥0 . Proof. The proof is similar to the demonstrations of Lemma 4.4 and Proposition 4.5 of [4]. 2 We are able again to apply the law of large numbers to the Markov chain (Zn , Sn+1 )n≥0 by virtue of this result. It should be noted that the measures µn , n ≥ 0, and µ∞ do not share with νn , n ≥ 0, and ν∞ the property of absolute continuity with respect to the Lebesgue measure presented in Remark 2.2. Remark A.3 For any n, µn and thus the limit µ∞ admit a density only on the interior F̊ of the state space F because of the expression (7) of S and thanks to Remark 2.2. Finally, we would like to highlight that the link between the measures ν∞ and µ∞ may be expressed in another way than (31). Indeed, for any x ∈ E, we have Z Q(Φ(y, s), x)µ∞ (dy × ds) = ν∞ (x), (32) F by the expression (5) of the transition kernel P of (Zn )n≥0 and because ν∞ P = ν∞ by virtue of Proposition A.1. The formula (32) of ν∞ will be useful in our investigations. B Proof of Theorem 3.3 In all this section we use the notation Fn for the σ-algebra σ(Z0 , S1 , . . . , Zn , Sn+1 ), n ≥ 0. In addition, the classical symbols ∼, o and O must be understood to hold here in the almost sure sense. B.1 Sketch of the proof The proof of Theorem 3.3 relies on the following decomposition. For any integer n ≥ 1, we have  n    Fb (x, t) ν∞ (x)f (x, t) n  Gb (x, t)  −  ν∞ (x)G(x, t)  = Mn−1 + Rn , n n ν∞ (x) νb∞ (x) (33) where the sequence (Mn )n≥0 is a (Fn )n≥0 -vector martingale defined by (37) and studied in Appendix B.3, and Rn is a remainder term defined by (34) and studied in Appendix B.2. 19 In Appendix B.2, we establish that the remainder term almost surely goes to 0 when n tends to infinity. This is a first step to show the almost sure convergence presented in Theorem 3.3. In addition we investigate the rate of convergence of the remainder term in Appendix B.2 under an additional Lipschitz mixing condition stated in Assumption 2.4. This is enough to prove the asymptotic normality given in Theorem 3.3. The rest of the proof deals with the study of the martingale term. We prove in Appendix B.3 that the process (Mn )n≥0 is a vector martingale. We investigate its asymptotic behavior by studying its square variation process (hMin )n≥0 in Appendix B.3. Thanks to these results we state the law of large numbers and the central limit theorem for (Mn )n≥0 in Appendix B.3. Finally the almost sure convergence presented in Theorem 3.3 is a direct application of (33) and (35) together with (51), while the asymptotic normality is obtained from (33), (36) and (54). B.2 Remainder term This part of the paper is devoted to the asymptotic study of the remainder term sequence (Rn )n≥1 appearing in (33). B.2.1 Definition of the remainder term i h (k) For any n, one may write Rn = Rn , where each of the components is defined by 1≤k≤3 (1) Rn = 4 X (1,j) Rn , j=0 (2) Rn = 3 X (2,j) Rn j=0 and (3) Rn = 2 X (3,j) Rn . (34) j=0 (1) The five terms that define Rn are given by 1 Z0 − x S1 − t (1,0) Rn = K K , 1 d v0 w0 nv0d w0 n−1 Z 1X (1,1) Rn = f (x + uvj , t + vwj )Kd (u)K1 (v) (Q (Φ (Zj−1 , Sj ) , x + uvj )−Q (Φ (Zj−1 , Sj ) , x)) du dv, n j=1 Rd ×R (1,2) Rn = (1,3) Rn = (1,4) Rn = Z n−1 1X Q (Φ (Zj−1 , Sj ) , x) [f (x + uvj , t + vwj ) − f (x, t + vwj )] Kd (u)K1 (v) du dv, n j=1 Rd ×R Z n−1 1X Q (Φ (Zj−1 , Sj ) , x) [f (x, t + vwj ) − f (x, t)] Kd (u)K1 (v) du dv, n j=1 Rd ×R " n−1 # 1X Q (Φ (Zj−1 , Sj ) , x) − ν∞ (x) f (x, t). n j=1 (2) The four terms that define Rn are given by 1 1 Z0 − x (2,0) Rn = K 1{S1 >t} , d n v0d v0 n−1 Z 1X (2,1) Rn = [Q (Φ (Zj−1 , Sj ) , x + uvj ) − Q (Φ (Zj−1 , Sj ) , x)] G(x + uvj , t)Kd (u) du, n j=1 Rd (2,2) Rn = (2,3) Rn = Z n−1 1X Q (Φ (Zj−1 , Sj ) , x) [G(x + uvj , t) − G(x, t)] Kd (u) du, n j=1 Rd " n−1 # 1X Q (Φ (Zj−1 , Sj ) , x) − ν∞ (x) G(x, t). n j=1 20 (3) Finally, the three terms that define Rn are given by Z0 − x 1 1 (3,0) K , Rn = d n v0d v0 n−1 Z 1X (3,1) Rn = [Q (Φ (Zj−1 , Sj ) , x + uvj ) − Q (Φ (Zj−1 , Sj ) , x)] Kd (u) du, n j=1 Rd (3,2) Rn B.2.2 = n−1 1X Q (Φ (Zj−1 , Sj ) , x) − ν∞ (x). n j=1 Almost sure convergence (1) We only investigate the first component (Rn )n≥1 since the other terms may be treated with similar (1,0) arguments. First, it is obvious that the sequence (Rn )n≥1 almost surely tends to 0. In addition, (1,4) the term (Rn )n≥1 converges to 0 by virtue of the ergodic theorem applied to the Markov chain (Zn , Sn+1 )n≥0 (thanks to Proposition A.2) together with (32). Because both the functions Q and f are Lipschitz and bounded (see Assumptions 2.3), we have Z n−1 [Q]Lip kf k∞ X (1,1) Rn ≤ vj \|u\|Kd (u)du, n Rd j=1 (1,2) Rn ≤ Z n−1 kQk∞ [f ]Lip X \|u\|Kd (u)du, vj n Rd j=1 (1,3) Rn ≤ Z n−1 kQk∞ [f ]Lip X wj \|v\|K1 (v)dv, n R j=1 which are some Cesàro means because both the sequences (vn )n≥0 and (wn )n≥0 tend to 0. Thus, a.s. Rn −→ 03 . B.2.3 (35) Rate of convergence Under the first item of Assumptions 2.4, the Markov chain (Zn )n≥0 and thus the two-dimensional process (Zn , Sn+1 )n≥0 satisfy the contraction assumption given in Theorem 6.3.17 of [21]. By applying this theorem to the function ϕ = Q(Φ(·, ·), ·) ∈ Li(r1 , r2 ) (see second item of Assumptions 2.4), we obtain that, for any γ ∈ (0, 1) and x ∈ E, n− 1+γ 2 n−1 X a.s [Q (Φ (Zj−1 , Sj ) , x) − ν∞ (x)] −→ 0, j=1 because 2(r1 + r2 ) ≤ a1 . Therefore, we have for any couple (k, j) in the set {(1, 4), (2, 3), (3, 2)}, n 1−αd−β 2 a.s. (k,j) Rn −→ 0. The same result is obvious for the other couples (k, j) with the condition αd + β + 2 min(α, β) > 1. Finally we obtain, when n goes to infinity, n B.3 B.3.1 1−αd−β 2 a.s. Rn −→ 03 . (36) Martingale term (Mn )n≥0 is a vector martingale h i (k) Let (Mn )n≥0 be the three-dimensional process defined for any n by Mn = Mn , where each 1≤k≤3 component is defined as Mn(k) = n X (k) Aj (k) − Bj , k ∈ {1, 2, 3}, j=1 21 (37) (k) and the terms Aj (1) = (1) = Aj Bj (2) = (2) = (3) = (3) = Aj Bj Aj Bj (k) and Bj 1 Kd vjd Z are given by 1 Sj+1 − t Zj − x , K1 vj wj wj Q(Φ(Zj−1 , Sj ), x + uvj )f (x + uvj , t + vwj ) Kd (u)K1 (v) λd (du) dv, Zj − x 1 1{Sj+1 >t} , K d vj vjd Z Q(Φ(Zj−1 , Sj ), x + uvj )G(x + uvj , t) Kd (u) λd (du), Rd Zj − x 1 , K d vj vjd Z Q(Φ(Zj−1 , Sj ), x + uvj ) Kd (u) λd (du). Rd ×R Rd We claim that the process (Mn )n≥0 is a (Fn )n≥0 -martingale. The keystone is to show that, for any n and k, we have h i (k) Fn−1 = Bn−1 . E A(k) n The proof presents no particular difficulty except for the first component k = 1 for which we provide some details. Let us recall that R denotes the transition kernel of the Markov chain (Zn , Sn+1 )n≥0 (see Subsection 2.1). We have Z i h 1 s−t ξ−x F = E A(1) K R ((Zn−1 , Sn ), dξ × ds) K 1 n−1 d n vnd wn Rd ×R+ vn wn Z s−t 1 ξ−x = K Q (Φ(Zn−1 , Sn ), dξ) S(ξ, ds), (38) K 1 d vnd wn Rd ×R+ vn wn with (11). Let us recall that the bandwidth sequence (vn )n≥0 is decreasing. Together with the third item of Assumptions 3.1, we obtain ·−x ·−x ⊂ supp Kd ⊂ Bd (x, v0 δ). supp Kd vn v0 With similar arguments, we obtain ·−t ·−t supp K1 ⊂ supp K1 ⊂ B1 (t, w0 δ) ⊂ −∞, inf t+ (ξ) , ξ∈Bd (x,v0 δ) wn w0 (39) with the condition (14) on the couple (v0 , w0 ). By (38), (39) together with the expression (7) of S we obtain Z h i 1 ξ−x s−t E A(1) Fn−1 = d Kd K1 Q (Φ(Zn−1 , Sn ), dξ) f (ξ, s) ds. n vn wn Rd ×R+ vn wn One may conclude by a change of variable and (3). B.3.2 Predictable square variation process The asymptotic behavior of the martingale (Mn )n≥0 may be investigated by studying its predictable square variation process that we denote as usual (hMin )n≥0 . At any time n, hMin is the 3×3 symmetric matrix defined by n h h i i X (i) (j) (i) (j) hMin = E Mk Mk Fk−1 − Mk−1 Mk−1 . 1≤i,j≤3 k=1 Now, we calculate each coefficient of this matrix by beginning with the diagonal terms and the first of (1,1) these hMin . One has hMi(1,1) = n Tn(1) − Tn(2) , n 22 (1) (2) where the terms Tn and Tn are defined by n 1X 1 Q(Φ(Zj−1 , Sj ), x + uvj )f (x + uvj , t + vwj )K2d (u)K21 (v)du dv, n j=1 vjd wj Rd ×R 2 n Z X Q(Φ(Zj−1 , Sj ), x + uvj )f (x + uvj , t + vwj )Kd (u)K1 (v)du dv . Z Tn(1) = Tn(2) = j=1 Rd ×R Since the functions Q and f are bounded (see Assumptions 2.3), we easily obtain that Tn(2) = O(n). (40) Let us introduce the additional notation, Ten(1) = Ťn(1) = = n Z 1X Q(Φ(Zj−1 , Sj ), x + uvj )f (x + uvj , t + vwj )K2d (u)K21 (v)du dv, n j=1 Rd ×R Z n 1X Q(Φ(Zj−1 , Sj ), x)f (x, t) K2d (u)K21 (v)du dv n j=1 Rd ×R n τd2 τ12 X Q(Φ(Zj−1 , Sj ), x)f (x, t), n j=1 by Remark 3.2 on the kernel functions K1 and Kd . Since both Q and f are Lipschitz and bounded (see Assumptions 2.3), together with supp Kp ⊂ Bp (0p , δ) (see Assumptions 3.1), the stochastic sequences (1) (1) (Ten )n≥0 and (Ťn )n≥0 have the same limit thanks to n (kQk∞ [f ]Lip + [Q]Lip kf k∞ ) δτd2 τ12 X Ten(1) − Ťn(1) ≤ (vj + wj ), n j=1 (41) which is a Cesàro mean. In addition, by the almost sure ergodic theorem (see Proposition A.2) together with (32), we obtain a.s. Ťn(1) −→ τd2 τ12 ν∞ (x)f (x, t). (42) By (41) and (42) and by virtue of Lemma 7.1.5 of [21], we obtain (1) 2 2 Tn a.s. τ τ ν∞ (x)f (x, t) −→ d 1 . αd+β n 1 + αd + β Together with (40), as n goes to infinity, (1,1) 2 2 hMin a.s. τ τ ν∞ (x)f (x, t) −→ d 1 = 41 . 1+αd+β n 1 + αd + β (43) The second diagonal term of the predictable variation process may be studied in a similar way as before. We obtain " Z n X 1 (2,2) hMin = Q(Φ(Zj−1 , Sj ), x + uvj )G(x + uvj , t)K2d (u)du d v d R j j=1 Z 2 # − Q(Φ(Zj−1 , Sj ), x + uvj )G(x + uvj , t)Kd (u)du , Rd and, as n goes to infinity, (2,2) hMin n1+αd a.s. −→ τd2 ν∞ (x)G(x, t) = 42 . 1 + αd 23 (44) The third and last diagonal term may also be investigated in the same way. We have " Z n X 1 (3,3) Q(Φ(Zj−1 , Sj ), x + uvj )K2d (u)du hMin = d v Rd j j=1 2 # Z Q(Φ(Zj−1 , Sj ), x + uvj )Kd (u)du , − Rd and, when n goes to infinity, (3,3) 2 hMin a.s. τd ν∞ (x) −→ = 43 . (45) n1+αd 1 + αd Now we focus on the non diagonal terms. For any integer n, we have "Z n X (1,2) hMin = Q(Φ(Zj−1 , Sj ), x + uvj )2 f (x + uvj , t + vwj )G(x + uvj , t)K2d (u)K1 (v)du dv Rd ×R j=1 Z × hMi(1,3) n = n X Q(Φ(Zj−1 , Sj ), x + uvj )f (x + uvj , t + vwj )Kd (u)K1 (v)du dv Rd ×R Z # − Rd "Z Rd ×R j=1 Q(Φ(Zj−1 , Sj ), x + uvj )G(x + uvj , t)Kd (u)du , Q(Φ(Zj−1 , Sj ), x + uvj )2 f (x + uvj , t + vwj )K2d (u)K1 (v)du dv Z − = Rd "Z n X R j=1 Q(Φ(Zj−1 , Sj ), x + uvj )f (x + uvj , t + vwj )Kd (u)K1 (v)du dv Rd ×R Z # × hMi(2,3) n Q(Φ(Zj−1 , Sj ), x + uvj )Kd (u)du , × (47) Q(Φ(Zj−1 , Sj ), x + uvj )2 G(x + uvj , t)K2d (u)du d Z − (46) Q(Φ(Zj−1 , Sj ), x + uvj )G(x + uvj , t)Kd (u)du Rd Z # Rd Q(Φ(Zj−1 , Sj ), x + uvj )Kd (u)du . (48) From (46), (47) and (48), and because Q and f are bounded (see Assumptions 2.3) together with the R fact that the integral K2p dλp is finite (see Remark 3.2), we easily obtain that, for any i 6= j, hMi(i,j) = O(n). n (49) As a conclusion, by (43), (44), (45) and (49), one may sum up the asymptotic behavior of the predictable variation process (hMin )n≥0 by the following formula,  β  n 41 0 0 hMin ∼ (50) 0 42 0 . n1+αd 0 0 43 It should be noted that the coefficients 4i , 1 ≤ i ≤ 3, are positive because we assume that ν∞ (x) > 0 and f (x, t) > 0 in the statement of Theorem 3.3. B.3.3 Limit theorems for the vector martingale Law of large numbers. We propose to apply the law of large numbers for vector martingales (see [21, Theorem 4.3.15]) to the process of interest (Mn )n≥0 . In the sequel, for any n, Tn denotes the trace of the matrix hMin , while En stands for its minimum eigenvalue. First, in light of (50), the trace (Tn )n≥0 24 almost surely tends to infinity. Thus we are able to apply the third item of Theorem 4.3.15 of [21] with any function a(t) = t1+η , η > 0, and we obtain, 2 log(Tn )1+η hMi−1/2 Mn = o . n En By (50) again, we have, when n goes to infinity, Tn = O n1+αd+β and En ∼ min(42 , 43 ) n1+αd . As a consequence, using that n1+αd+β = o(n2 ) whenever αd + β < 1 together with (50), we have the law of large numbers, as n tends to infinity, kMn k a.s. −→ 0. (51) n Central limit theorem. We now investigate the asymptotic normality of the vector martingale (Mn )n≥0 . We apply Corollary 2.1.10 of [21] with the sequence (an )n≥0 defined by an = n1+αd+β . The first assumption of this result is obviously satisfied: the sequence (a−1 n hMin )n≥0 almost surely converges to some positive semi-definite matrix. Indeed, by (50),  2 2  τd τ1 ν∞ (x)f (x,t) 0 0 1+αd+β a.s.   n−1−αd−β hMin −→  (52) 0 0 0  = Σ(x, t, α, β), 0 0 0 where Σ(x, t, α, β) is a degenerate variance-covariance matrix. As a consequence, we only have to check Lindeberg’s condition in order to establish the central limit theorem for (Mn )n≥0 . In other words, we have to prove that, for any > 0, # " n X 1 P 2 (53) E \|Mj − Mj−1 \| 1 Fj−1 −→ 0. 1+αd+β 2 n1+αd+β j=1 \|Mj −Mj−1 \|≥n (k) (k) For any 1 ≤ j ≤ n and 1 ≤ k ≤ 3, we study the three components Mj − Mj−1 . We have (j + 1)αd+β Zj − x Sj+1 − t (1) (1) Mj − Mj−1 = K K 1 d vj wj v d w0 Z 0 − Q(Φ(Zj−1 , Sj ), x + uvj )f (x + uvj , t + vwj ) Kd (u)K1 (v) λd (du) dv, Rd ×R (2) = (3) = (2) − Mj−1 (3) − Mj−1 Mj Mj (j + 1)αd Zj − x K 1{Sj+1 >t} d vj v0d Z − Q(Φ(Zj−1 , Sj ), x + uvj )G(x + uvj , t) Kd (u) λd (du), Rd Z (j + 1)αd Zj − x − Q(Φ(Zj−1 , Sj ), x + uvj ) Kd (u) λd (du). K d vj v0d Rd Thus, we obtain (1) ≤ (n + 1)αd+β kKd k∞ kK1 k∞ + kQk∞ kf k∞ , v0d w0 (2) ≤ (n + 1)αd kKd k∞ + kQk∞ , v0d (1) − Mj−1 \| (2) − Mj−1 \| \|Mj \|Mj (n + 1)αd kKd k∞ + kQk∞ . v0d 1+αd+β Together with the condition αd + β < 1, \|Mj − Mj−1 \| = o n 2 . As a consequence, there exists n o 1+αd+β an integer N such that for any n ≥ N , the event \|Mj − Mj−1 \| ≥ n 2 is almost surely empty. (3) \|Mj (3) − Mj−1 \| ≤ This shows Lindeberg’s condition (53). Finally, by (52) and (53), we obtain, as n tends to infinity, n− 1+αd+β 2 d Mn −→ N (03 , Σ(x, t, α, β)) . 25 (54) C Proof of Proposition 3.14 en , Sen+1 )n≥0 , we By virtue of the ergodic theorem (see Proposition A.2) applied to the Markov chain (Z have, as ñ tends to infinity, Z Z 2 Γ d−1 +1 a.s. n,e n,ρ n 2 2 Gbn (ξ, θx (ξ)) 1(θx (ξ),+∞) (t)µ∞ (dξ × dt), εbκ (α) −→ κ bx (ξ) dξ − d−1 d−1 2 Cx Tx,ρ ×R+ π ρ conditionally to σ(Z0 , S1 , . . . , Zn−1 , Sn ). Together with the definition of µ∞ (31), the expression (9) of S(x, ·) and Remark 2.2, we obtain Z Z 2 Γ d−1 +1 a.s. n,e n,ρ n 2 2 εbκ (α) −→ κ bx (ξ) dξ − Gbn (ξ, θx (ξ)) G(ξ, θx (ξ))ν∞ (ξ)λd (dξ), (55) d−1 d−1 2 Cx Tx,ρ ρ π conditionally to σ(Z0 , S1 , . . . , Zn−1 , Sn ). By definition (24) of Tx,ρ we have Z Tx,ρ Gbn (ξ, θx (ξ)) G(ξ, θx (ξ))ν∞ (ξ)λd (dξ) = ! Z Z Dx,ρ Gbn (ξ, θy (ξ)) G(ξ, θy (ξ))ν∞ (ξ)dξ λd−1 (dy). (56) Cy In addition, one has Z Gbn (ξ, θy (ξ)) G(ξ, θy (ξ))ν∞ (ξ)dξ = Cy t− (y) Z Gbn (Φ(y, −t), t)G(Φ(y, −t), t)ν∞ (Φ(y, −t)) \|∇t Φ(y, −t)\| dt. 0 As a consequence, under Assumptions 3.13, a conscientious calculus together with eq. (25) shows that Z Gbn (ξ, θy (ξ)) G(ξ, θy (ξ))ν∞ (ξ)dξ − Cy Z Gbn (ξ, τx (ξ)) G(ξ, τx (ξ))ν∞ (ξ)dξ = O(\|x − y\|). (57) Cx We obtain the expected result from (55), (56) and (57) and by remarking that the normalizing constant is λd−1 (Dx,ρ ). Remark C.1 In order to prove (57), one may split the integral of interest into two terms: an integral between 0 and t− (x)∧t− (y) and a remainder term integrated between t− (x)∧t− (y) and t− (x)∨t− (y). The first one is clearly upper bounded by the integral of a Lipschitz function between 0 and t− (x)+\|x−y\|[t− ]Lip , thus bounded by an integral between 0 and t− (x) + ρ[t− ]Lip . The second integral is obviously bounded by kGbn k∞ kν∞ k∞ k∇t Φk∞ [t− ]Lip \|x − y\|. References [1] Odd Olai Aalen. Statistical inference for a family of counting processes. ProQuest LLC, Ann Arbor, MI, 1975. Thesis (Ph.D.)–University of California, Berkeley. [2] Terje Aven and Uwe Jensen. Stochastic models in reliability, volume 41 of Applications of Mathematics (New York). Springer-Verlag, New York, 1999. [3] Romain Azaı̈s. A recursive nonparametric estimator for the transition kernel of a piecewisedeterministic Markov process. ESAIM: Probability and Statistics, 18:726–749, January 2014. [4] Romain Azaı̈s, François Dufour, and Anne Gégout-Petit. Nonparametric estimation of the jump rate for non-homogeneous marked renewal processes. Ann. Inst. H. Poincaré Probab. Statist., 49(4):1204– 1231, 11 2013. [5] Romain Azaı̈s, François Dufour, and Anne Gégout-Petit. Non-parametric estimation of the conditional distribution of the interjumping times for piecewise-deterministic markov processes. Scandinavian Journal of Statistics, 41(4):950–969, 2014. [6] Romain Azaı̈s and Alexandre Genadot. Semi-parametric inference for the absorption features of a growth-fragmentation model. TEST, 24(2):341–360, 2015. 26 [7] Anis Ben Abdessalem, Romain Azaı̈s, Marie Touzet-Cortina, Anne Gégout-Petit, and Monique Puiggali. Stochastic modelling and prediction of fatigue crack propagation using piecewise-deterministic markov processes. Accepted for publication in Journal of Risk and Reliability, 2016. [8] Patrice Bertail, Stephan Clemençon, and Jessica Tressou. A storage model with random release rate for modelling exposure to food contaminants. Mathematical Bioscience and Engineering, 5(1):35–60, 2008. [9] Patrice Bertail, Stephan Clemençon, and Jessica Tressou. Statistical analysis of a dynamic model for dietary contaminant exposure. Journal of Biological Dynamics, 4(2):212–234, 2010. [10] Adrien Brandejsky, Benoı̂te de Saporta, and François Dufour. Numerical method for expectations of piecewise-determistic markov processes. CAMCoS, 7(1):63–104, 2012. [11] Adrien Brandejsky, Benoı̂te De Saporta, and François Dufour. Numerical methods for the exit time of a piecewise-deterministic markov process. Adv. in Appl. Probab., 44(1):196–225, 03 2012. [12] Djalil Chafai, Florent Malrieu, and Katy Paroux. On the long time behavior of the TCP window size process. Stochastic Processes and their Applications, 120(8):1518–1534, April 2010. [13] Julien Chiquet and Nikolaos Limnios. A method to compute the transition function of a piecewise deterministic Markov process with application to reliability. Statist. Probab. Lett., 78(12):1397–1403, 2008. [14] Julien Chiquet, Nikolaos Limnios, and Mohamed Eid. Piecewise deterministic markov processes applied to fatigue crack growth modelling. Journal of Statistical Planning and Inference, 139(5):1657 – 1667, 2009. [15] Gerda Claeskens and Peter Hall. Effect of dependence on stochastic measures of accuracy of density estimations. The Annals of Statistics, 30(2):431–454, 2002. [16] Alina Crudu, Arnaud Debussche, Aurélie Muller, Ovidiu Radulescu, et al. Convergence of stochastic gene networks to hybrid piecewise deterministic processes. The Annals of Applied Probability, 22(5):1822–1859, 2012. [17] Mark H. A. Davis. Markov models and optimization, volume 49 of Monographs on Statistics and Applied Probability. Chapman & Hall, London, 1993. [18] Benoı̂te de Saporta, François Dufour, Huilong Zhang, and Charles Elegbede. Optimal stopping for the predictive maintenance of a structure subject to corrosion. Journal of Risk and Reliability, 226 (2):169–181, 2012. [19] Marie Doumic, Marc Hoffmann, Nathalie Krell, and Lydia Robert. Statistical estimation of a growth-fragmentation model observed on a genealogical tree. Bernoulli (to appear), 2015. [20] Marie Doumic, Marc Hoffmann, Patricia Reynaud-Bouret, and Vincent Rivoirard. Nonparametric estimation of the division rate of a size-structured population. SIAM Journal on Numerical Analysis, 50(2):925–950, 2012. [21] Marie Duflo. Random iterative models. Applications of Mathematics. Springer-Verlag, Berlin, 1997. [22] Alexandre Genadot and Michèle Thieullen. Averaging for a fully coupled piecewise-deterministic markov process in infinite dimensions. Adv. in Appl. Probab., 44(3):749–773, 09 2012. [23] Jeffrey D. Hart and Philippe Vieu. Data-driven bandwidth choice for density estimation based on dependent data. The Annals of Statistics, pages 873–890, 1990. [24] Jianghai Hu, Wei-Chung Wu, and Shankar Sastry. Modeling subtilin production in bacillus subtilis using stochastic hybrid systems. In R. Alur and G.J. Pappas, editors, Hybrid Systems: Computation and Control , number 2993 in LNCS, Springer-Verlag, Berlin, 2004. [25] Martin Jacobsen. Point Process Theory and Applications: Marked Point and Piecewise Deterministic Processes. Probability and its Applications. Birkhäuser, Boston-Basel-Berlin, 2006. [26] Tae Yoon Kim. Asymptotically optimal bandwidth selection rules for the kernel density estimator with dependent observations. Journal of statistical planning and inference, 59(2):321–336, 1997. [27] Tae Yoon Kim and Dennis D. Cox. Bandwidth selection in kernel smoothing of time series. Journal of Time Series Analysis, 17(1):49–63, 1996. 27 [28] Nathalie Krell. Statistical estimation of jump rates for a specific class of piecewise deterministic markov processes. arXiv:1406.2845, March 2015. [29] Sean Meyn and Richard L. Tweedie. Markov chains and stochastic stability. Cambridge University Press, Cambridge, second edition, 2009. [30] N.R. Norris. Exploring the optimality of various bacterial motility strategies: a stochastic hybrid systems approach. PhD thesis, Massachusetts Institute of Technology, 2013. [31] H. G. Othmer, X. Xin, and C. Xue. Excitation and adaptation in bacteria. a model signal transduction system that controls taxis and spatial pattern formation. International journal of molecular sciences, 14(5):9205–9248, 2013. [32] Ovidiu Radulescu, Aurélie Muller, and Alina Crudu. Théorèmes limites pour les processus de markov à sauts. Technique et Science Informatiques, 26(3-4):443–469, 2007. [33] Lydia Robert, Marc Hoffmann, Nathalie Krell, Stephane Aymerich, Jerome Robert, and Marie Doumic. Division in escherichia coli is triggered by a size-sensing rather than a timing mechanism. BMC Biology, 12(1):17, 2014. [34] M. J. Tindall, P. K. Maini, S. L. Porter, and J. P. Armitage. Overview of mathematical approaches used to model bacterial chemotaxis ii: bacterial populations. Bulletin of mathematical biology, 70(6):1570–1607, 2008. [35] D. Virkler, B. Hillberry, and P. Goel. The statistical nature of fatigue crack propagation. J Engng Mater Technol, 101(2):148 – 153, 1979. 28 1.0 0.6 0.0 0.2 0.4 First dimension 0.8 1.0 0.8 0.6 0.4 Second dimension 0.2 0.0 0.0 0.2 0.4 0.6 0.8 1.0 0 1 2 First dimension 3 4 5 6 Time on 0.6 1.0 0.5 0.2 Secon asure d dime nsion e iant m t dim ensi 1.5 0.2 Invar Firs Second dimension Figure 2: Two representations of the same simulated path of the TCP-like model of interest until the 10th jump. A vector field graph is given on the left, while we observe the trajectory of the first component versus time on the right. 0.4 0.6 0.8 First dimension Figure 3: Estimation of the invariant distribution of the post-jump locations computed from the 20 000 first jumps of the TCP-like model. 29 0.1 0.2 0.3 0.4 0.5 −0.2 −0.3 −0.4 −0.5 −0.7 −0.6 Estimated ISE (up to an additive constant) −0.6 −0.7 −0.9 −0.8 Estimated ISE (up to an additive constant) −0.5 −0.6 −0.7 −0.8 −0.9 −1.0 Estimated ISE (up to an additive constant) −1.1 0.0 0.0 Parameter α 0.1 0.2 0.3 0.4 0.5 0.0 0.1 Parameter α 0.2 0.3 0.4 0.5 Parameter α 1.5 0.5 1.0 Estimated κ criterion 1.2 1.0 0.8 0.6 0.4 Estimated κ criterion 1.4 2.0 Figure 4: The choice of the bandwidth parameter α appearing in the criterion κ bnx is obtained by minimizing the cross-validation estimate εbκn,en,ρ (α) of the related ISE. The parameter ρ seems to have only a small influence over the minimization: the estimated error is computed from ρ = 0.005, ρ = 0.01 and ρ = 0.02 from left to right. 0.1 0.2 0.3 0.4 0.5 0.6 0.05 0.7 0.15 0.25 0.35 0.45 0.55 0.65 0.75 Deterministic path Deterministic path Figure 5: The optimal index ξb∗n is calculated by maximizing the criterion κ bnx (ξ), ξ ∈ Cx , computed with the optimal parameter α. On the left side of the figure, we compare the estimate κ bnx (ξ) (full line) with n νb∞ (ξ)G(ξ, τx (ξ)) (dashed line) both computed from α = 0.1, which confirms that the estimation performs pretty well. This quantity seems to admit one and only one absolute maximum. 30 6 5 4 3 2 Estimated κ criterion 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Deterministic path ha alp 0.25 −3.20 −3.25 −3.30 −3.35 −3.40 −3.45 −3.50 0.15 Parameter β ISE ta be 0.35 Figure 6: The parameter α plays a crucial role in the estimation of κ bnx . We compare here the curves computed from α = 0.1 and already presented in Figure 5 (full and dashed lines) with the too oscillating estimate obtained from α = 0.4 (dotted line). 0.10 0.12 0.14 0.16 0.18 0.20 Parameter α Figure 7: The optimal bandwidth parameters α and β implicitly appearing in Fbn (ξ, τx (ξ)), ξ ∈ Cx , are obtained by minimizing the cross-validation estimate εbFn,en,ρ1 ,ρ2 (α, β), computed from ρ1 = ρ2 = 0.1 here, of the related ISE. The parameter β seems to have only little influence over the estimation error in comparison to α. 31 3.0 2.5 2.0 1.5 1.0 Estimation of λ(0.75,0.5) 0.5 0.0 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 Deterministic path 0.0 −1.0 x2 1.0 bn (x) is computed from different values of the index ξ ∈ Cx , and from the optimal Figure 8: The estimator λ ξ bandwidth parameters αG and (αF , β F ). Over the 100 replicates, the optimal points ξb∗n are most of the time located at 0.55 and more than 90 times between 0.5 and 0.6 (enhanced by gray colors), which seems to correspond with the estimators with least bias and variance. In particular, we obtain a better result than n (index around 0.35), see Remark 3.11. by maximizing the estimated invariant measure νb∞ −1.0 −0.5 0.0 0.5 1.0 x1 Figure 9: Simulated trajectory of the bacteria in the unit disk until the 100th jump. 32 1.5 0.5 Estimated jump rate 1 2 3 4 5 6 7 8 9 bn (xθ ) for θ Figure 10: For each of the 9 target points (x1 , x2 ), the boxplot of the optimal estimators λ ξbn (θ) ∗ 1.05 0.95 1.02 0.95 1.07 0.96 1.06 −1.0 −0.5 0.0 0.5 1.4 1.0 1.03 0.6 0.0 1.06 −1.0 x2 1.0 Estimated jump rate taken in a uniform discretized grid of [0, 2π] with step π/8 is presented. The black thick lines in the boxplots b 1 , x2 ) defined in (30), while the dashed line represents the true correspond to the aggregated estimate λ(x jump rate. 1.0 x1 b 1 , x2 ) have been Figure 11: For each of the 9 target points, the curve Cx1 ,x2 as well as the estimate λ(x computed from n = 100 000 observed data and are provided (left). The boxplot of these 9 estimates is also given (right) and presents no bias and a small dispersion. 33 0.98 0.96 1.2 1.1 1.43 −1.0 −0.5 0.0 0.5 1.4 1.24 1.0 1.4 0.6 Estimated jump rate 1.0 0.0 0.98 −1.0 x2 1.37 1.0 1.24 0.98 0.96 1.2 1.1 1.43 −1.0 −0.5 0.0 0.5 1.4 1.4 1.0 0.98 0.6 0.0 1.37 −1.0 x2 1.0 Estimated jump rate x1 1.0 x1 1.4 1.0 0.6 Estimated jump rate b 1 , x2 ) have been Figure 12: For each of the 9 target points, the curve Cx1 ,x2 as well as the estimate λ(x computed from n = 20 000 (top) and n = 50 000 (bottom) observed data and are provided (left). The boxplots of these 9 estimates are also given in both cases (right). n=100 000 n=50 000 n=20 000 2 4 6 8 b 1 , x2 ) Figure 13: For each of the 9 target point (x1 , x2 ) indexed from 1 to 9, estimated jump rates λ(x computed from different datasets. 34 1.4 1.0 0.6 Estimated jump rate n=20000 n=50000 n=100000 30 20 0 10 Crack length (mm) 40 50 b 1 , x2 ) computed from datasets of different sizes. Figure 14: Boxplots of the 9 estimated jump rates λ(x 0 50000 100000 200000 300000 Time (number of cycles) −25 −26 −27 −28 Parameter log(C) −24 −23 Figure 15: Virkler’s dataset contains 68 independent crack growth histories starting from a0 = 9 mm. 2.4 2.6 2.8 3.0 3.2 Parameter m Figure 16: Material parameters m and log(C) are strongly linked by a linear relationship used to reduce the dimension of the underlying model: log(C) = −9.25 − 5.89 × m + ε. 35 0.8 0.7 0.6 0.5 0.3 0.4 Estimated criterion at 35mm 0.9 1.0 1.4 1.2 1.0 0.8 Estimated criterion at 25mm 0.6 2.4 2.6 2.8 3.0 3.2 2.4 2.6 2.8 3.0 3.2 40 45 2.95 2.90 2.80 2.85 0.15 0.20 0.25 Best parameter m in the estimation 0.30 Parameter m 0.10 Estimated criterion at 45mm Parameter m 2.4 2.6 2.8 3.0 3.2 25 30 Parameter m 35 Crack length 8e−05 6e−05 4e−05 2e−05 0e+00 Jump rate to unstable region Figure 17: Estimation of the criterion κa (m) for different values of the target length a: a = 25 mm (top, left), a = 35 mm (top, right) and a = 45 mm (bottom, left). The relationship between the optimal parameter m maximizing κa (m) and the target crack length a is also presented (bottom, right). 25 30 35 40 45 Crack length (mm) Figure 18: Estimation of the jump rate λ(a) for different crack lengths a between 25 mm and 45 mm in the stochastic model of fatigue crack propagation from Virkler’s dataset. 36	10math.ST
Classification of Major Depressive Disorder via Multi-Site Weighted LASSO Model Dajiang Zhu1, Brandalyn C. Riedel1, Neda Jahanshad1, Nynke A. Groenewold2,3, Dan J. Stein3, Ian H. Gotlib4, Matthew D. Sacchet5, Danai Dima6,7, James H. Cole8, Cynthia H.Y. Fu9, Henrik Walter10, Ilya M. Veer11, Thomas Frodl11,12, Lianne Schmaal13.14,15, Dick J. Veltman15, Paul M. Thompson1 1 Imaging Genetics Center, USC Stevens Neuroimaging and Informatics Institute, Keck School of Medicine of the University of Southern California, CA, USA; 2 BCN NeuroImaging Center and Department of Neuroscience of the University of Groningen, University Medical Center Groningen, The Netherlands; 3 Dept of Psychiatry and Mental Health, University of Cape Town, South Africa; 4 Neurosciences Program and Department of Psychology, Stanford University, CA, USA; 5 Department of Psychiatry and Behavioral Sciences, Stanford University, CA, USA; 6 Dept of Neuroimaging, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, UK; 7 Dept of Psychology, School of Arts and Social Science, City, University of London, UK; 8 Department of Medicine, Imperial College London, UK; 9 Department of Psychological Medicine, King’s College London, UK; 10 Dept of Psychiatry and Psychotherapy, Charité Universitätsmedizin Berlin, Germany; 11 Department of Psychiatry, Trinity College Dublin, Ireland; 12 Dept of Psychiatry and Psychotherapy, Otto von Guericke University Magdeburg, Germany; 13 Dept of Psychiatry and Neuroscience Campus Amsterdam, VU University Medical Center, The Netherlands; 14 Orygen, The National Centre of Excellence in Youth Mental Health, Australia; 15 Center for Youth Mental Health, The University of Melbourne, Australia Abstract. Large-scale collaborative analysis of brain imaging data, in psychiatry and neurology, offers a new source of statistical power to discover features that boost accuracy in disease classification, differential diagnosis, and outcome prediction. However, due to data privacy regulations or limited accessibility to large datasets across the world, it is challenging to efficiently integrate distributed information. Here we propose a novel classification framework through multi-site weighted LASSO: each site performs an iterative weighted LASSO for feature selection separately. Within each iteration, the classification result and the selected features are collected to update the weighting parameters for each feature. This new weight is used to guide the LASSO process at the next iteration. Only the features that help to improve the classification accuracy are preserved. In tests on data from five sites (299 patients with major depressive disorder (MDD) and 258 normal controls), our method boosted classification accuracy for MDD by 4.9% on average. This result shows the potential of the proposed new strategy as an effective and practical collaborative platform for machine learning on large scale distributed imaging and biobank data. Keywords: MDD, weighted LASSO 1 Introduction Major depressive disorder (MDD) affects over 350 million people worldwide [1] and takes an immense personal toll on patients and their families, placing a vast economic burden on society. MDD involves a wide spectrum of symptoms, varying risk factors, and varying response to treatment [2]. Unfortunately, early diagnosis of MDD is challenging and is based on behavioral criteria; consistent structural and functional brain abnormalities in MDD are just beginning to be understood. Neuroimaging of large cohorts can identify characteristic correlates of depression, and may also help to detect modulatory effects of interventions, and environmental and genetic risk factors. Recent advances in brain imaging, such as magnetic resonance imaging (MRI) and its variants, allow researchers to investigate brain abnormalities and identify statistical factors that influence them, and how they relate to diagnosis and outcomes [12]. Researchers have reported brain structural and functional alterations in MDD using different modalities of MRI. Recently, the ENIGMA-MDD Working Group found that adults with MDD have thinner cortical gray matter in the orbitofrontal cortices, insula, anterior/posterior cingulate and temporal lobes compared to healthy adults without a diagnosis of MDD [3]. A subcortical study – the largest to date – showed that MDD patients tend to have smaller hippocampal volumes than controls [4]. Diffusion tensor imaging (DTI) [5] reveals, on average, lower fractional anisotropy in the frontal lobe and right occipital lobe of MDD patients. MDD patients may also show aberrant functional connectivity in the default mode network (DMN) and other task-related functional brain networks [6]. Fig. 1. Overview of our proposed framework. Even so, classification of MDD is still challenging. There are three major barriers: first, though significant differences have been found, these previously identified brain regions or brain measures are not always consistent markers for MDD classification [7]; second, besides T1 imaging, other modalities including DTI and functional magnetic resonance imaging (fMRI) are not commonly acquired in a clinical setting; last, it is not always easy for collaborating medical centers to perform an integrated data analysis due to data privacy regulations that limit the exchange of individual raw data and due to large transfer times and storage requirements for thousands of images. As biobanks grow, we need an efficient platform to integrate predictive information from multiple centers; as the available datasets increase, this effort should increase the statistical power to identify predictors of disease diagnosis and future outcomes, beyond what each site could identify on its own. In this study, we introduce a multi-site weighted LASSO (MSW-LASSO) model to boost classification performance for each individual participating site, by integrating their knowledge for feature selection and results from classification. As shown in Fig. 1, our proposed framework features the following characteristics: (1) each site retains their own data and performs weighted LASSO regression, for feature selection, locally; (2) only the selected brain measures and the classification results are shared to other sites; (3) information on the selected brain measures and the corresponding classification results are integrated to generate a unified weight vector across features; this is then sent to each site. This weight vector will be applied to the weighted LASSO in the next iteration; (4) if the new weight vector leads to a new set of brain measures and better classification performance, the new set of brain measures will be sent to other sites. Otherwise, it is discarded and the old one is recovered. 2 Methods 2.1 Data and demographics For this study, we used data from five sites across the world. The total number of participants is 557; all of them were older than 21 years old. Demographic information for each site’s participants is summarized in Table 1. Sites 1 2 3 4 5 Total Total N of Total N of Age of Controls N MDD Controls (%) (Mean ± SD; y) patients (%) Groningen 45 22 (48.89%) 23 (51.11%) 42.78 ± 14.36 38.17 ± 9.97 Stanford 110 54 (49.09%) 56 (50.91) 51.72 ± 7.94 BRCDECC 130 69 (53.08%) 61 (46.92%) 172 101 (58.72%) 71 (41.28%) 41.09 ± 12.85 Berlin 100 53 (53%) 47 (47%) 38.49 ± 12.37 Dublin Combined 557 299 (53.68%) 258 (46.32$) Age of MDD (Mean ± SD; y) 43.14 ± 13.8 37.75 ± 9.78 47.85 ± 8.91 41.21 ± 11.82 41.81 ± 10.76 % Female MDD 72.73 57.41 68.12 64.36 62.26 % Female Total 73.33 60.00 60.77 60.47 57.00 Table 1. Demographics for the five sites participating in the current study. 2.2 Data preprocessing As in most common clinical settings, only T1-weighted MRI brain scans were acquired at each site; quality control and analyses were performed locally. Sixty-eight (34 left/34 right) cortical gray matter regions, 7 subcortical gray matter regions and the lateral ventricles were segmented with FreeSurfer [8]. Detailed image acquisition, pre-processing, brain segmentation and quality control methods may be found in [3, 9]. Brain measures include cortical thickness and surface area for cortical regions and volume for subcortical regions and lateral ventricles. In total, 152 brain measures were considered in this study. 2.3 Algorithm overview To better illustrate the algorithms, we define the following notations: 1. 𝐹𝑖 : The selected brain measures (features) of Site-i; 2. 𝐴𝑖 : The classification performance of Site-i; 3. W: The weight vector; 4. w-LASSO (W, 𝐷𝑖 ): Performing weighted LASSO on 𝐷𝑖 with weight vector – W; 5. SVM (𝐹𝑖 , 𝐷𝑖 ): Performing SVM classifier on 𝐷𝑖 using the feature set - 𝐹𝑖 ; The algorithms have two parts that are run at each site, and an integration server. At first, the integration server initializes a weight vector with all ones and sends it to all sites. Each site use this weight vector to conduct weighted LASSO (Section 2.6) with their own data locally. If the selected features have better classification performance, it will send the new features and the corresponding classification result to the integration server. If there is no improvement in classification accuracy, it will send the old ones. After the integration server receives the updates from all sites, it generates a new weight vector (Section 2.5) according to different feature sets and their classification performance. The detailed strategy is discussed in Section 2.5. Algorithm 1 (Integration Server) Initialize W (with all features weighted as one) Send W to all sites while at least one site has improvement on A update W (Section 2.5) Send W to all sites end while Send W with null to all sites 1. 2. 3. 4. 5. 6. 7. Table 2. Main steps of Algorithm 1. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. Algorithm 2 (Site-i) 𝐹𝑖 ← ∅, 𝐴𝑖 ← 0 while received W is not null 𝐹𝑖′ ← w-LASSO (W, 𝐷𝑖 ) (Section 2.6) if 𝐹𝑖′ ≠ 𝐹𝑖 𝐴′𝑖 ← SVM (𝐹𝑖′, 𝐷𝑖 ) if 𝐴′𝑖 > 𝐴𝑖 send 𝐹𝑖′ and 𝐴′𝑖 to Integration Server 𝐹𝑖 ← 𝐹𝑖′, 𝐴𝑖 ← 𝐴′𝑖 else send 𝐹𝑖 and 𝐴𝑖 to Integration Server end if end if end while Table 3. Main steps of Algorithm 2. 2.4 Ordinary LASSO and weighted LASSO LASSO [11] is a shrinkage method for linear regression. The ordinary LASSO is defined as: β̂(LASSO) = arg min ‖y − ∑ni=1 xi βi ‖2 + λ∑ni=1\|βi \| (1) Y and x are the observations and predictors. λ is known as the sparsity parameter. It minimizes the sum of squared errors while penalizing the sum of the absolute values of the coefficients - β. As LASSO regression will force many coefficients to be zero, it is widely used for variable selection [11]. However, the classical LASSO shrinkage procedure might be biased when estimating large coefficients [12]. To alleviate this risk, adaptive LASSO [12] was developed and it tends to assign each predictor with different penalty parameters. Thus it can avoid having larger coefficients penalized more heavily than small coefficients. Similarly, the motivation of multi-site weighted LASSO (MSW-LASSO) is to penalize different predictors (brain measures), by assigning different weights, according to its classification performance across all sites. Generating the weights for each brain measure (feature) and the MSW-LASSO model are discussed in Section 2.5 and 2.6. 2.5 Generation of a Multi-Site Weight In Algorithm 1, after the integration server receives the information on selected features (brain measures) and the corresponding classification performance of each site, it generates a new weight for each feature. The new weight for the 𝑓 𝑡ℎ feature is: 𝑊𝑓 = ∑𝑚 𝑠=1 𝛹𝑠,𝑓 𝐴𝑠 𝑃𝑠 ⁄𝑚 (2) 1, 𝑖𝑓 the 𝑓 𝑡ℎ feature was selected in site − s 𝛹𝑠,𝑓 = { 0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 (3) Here m is the number of sites. 𝐴𝑠 is the classification accuracy of site-s. 𝑃𝑠 is the proportion of participants in site-s relative to the total number of participants at all sites. Eq. (3) penalizes the features that only “survived” in a small number of sites. On the contrary, if a specific feature was selected by all sites, meaning all sites agree that this feature is important, it tends to have a larger weight. In Eq. (2) we consider both the classification performance and the proportion of samples. If a site has achieved very high classification accuracy and it has a relatively small sample size compared to other sites, the features selected will be conservatively “recommended” to other sites. In general, if the feature was selected by more sites and resulted in higher classification accuracy, it has larger weights. 2.6 Multi-Site weight LASSO In this section, we define the multi-site weighted LASSO (MSW-LASSO) model: β̂MSW−Lasso = arg min ‖y − ∑ni=1 xi βi ‖2 + λ∑ni=1(1 − ∑𝑚 𝑠=1 𝛹𝑠,𝑖 𝐴𝑠 𝑃𝑠 ⁄𝑚 )\|βi \| (4) Here xi represents the MRI measures after controlling the effects of age, sex and intracranial volume (ICV), which are managed within different sites. y is the label indicating MDD patient or control. n is the 152 brain measures (features) in this study. In our MSW-LASSO model, a feature with larger weights implies higher classification performance and/or recognition by multiple sites. Hence it will be penalized less and has a greater chance of being selected by the sites that did not consider this feature in the previous iteration. 3 Results 3.1 Classification improvements through the MSW-LASSO model In this study, we applied Algorithm 1 and Algorithm 2 on data from five sites across the world. In the first iteration, the integration server initialized a weight vector with all ones and sent it to all sites. Therefore, these five sites conducted regular LASSO regression in the first round. After a small set of features was selected using similar strategy in [9] within each site, they performed classification locally using a support vector machine (SVM) and shared the best classification accuracy to the integration server, as well as the set of selected features. Then the integration server generated the new weight according to Eq. (2) and sent it back to all sites. From the second iteration, each site performed MSW-LASSO until none of them has improvement on the classification result. In total, these five sites ran MSW-LASSO for six iterations; the classification performance for each round is summarized in Fig. 2 (a-e). Fig. 2. Applying MSW-LASSO to the data coming from five sites (a-e). Each subfigure shows the classification accuracy (ACC), specificity (SPE) and sensitivity (SEN) at each iteration. (f) shows the improvement in classification accuracy at each site after performing MSW-LASSO. Though the Stanford and Berlin sites did not show any improvements after the second iteration, the classification performance at the BRCDECC site and Dublin continued improving until the sixth iteration. Hence our MSW-LASSO terminated at the sixth round. Fig. 2f shows the improvements of classification accuracy for all five sites the average improvement is 4.9%. The sparsity level of the LASSO is set as 16% which means that 16% of 152 features tend to be selected in the LASSO process. Section 3.3 shows the reproducibility of results with different sparsity levels. When conducing SVM classification, the same kernel (RBF) was used, and we performed a grid search for possible parameters. Only the best classification results are adopted. 3.2 Analysis of MSW-LASSO features In the process of MSW-LASSO, only the new set of features resulting in improvements in classification are accepted. Otherwise, the prior set of features is preserved. The new features are also “recommended” to other sites by increasing the correspond- ing weights of the new features. Fig. 3 displays the changes of the involved features through six iterations and the top 5 features selected by the majority of sites. Fig. 3. (a) Number of involved features through six iterations. (b-f) The top five consistently selected features across sites. Within each subfigure, the top showed the locations of the corresponding features and the bottom indicated how many sites selected this feature through the MSW-LASSO process. (b-c) are cortical thickness and (d-f) are surface area measures. At the first iteration, there are 88 features selected by five sites. This number decreases over MSW-LASSO iterations. Only 73 features are preserved after six iterations but the average classification accuracy increased by 4.9%. Moreover, if a feature is originally selected by the majority of sites, it tends to be continually selected after multiple iterations (Fig. 3d-e). For those “promising” features that are accepted by fewer sites at first, they might be incorporated by more sites as the iteration increased (Fig. 2b-c, f). 3.3 Reproducibility of the MSW-LASSO Selected Improvement, in % Selected Improvement, in % Features ACC SPE SEN features ACC SPE SEN 3.1 1.8 4.4 2.6 3.1 2.5 13% 33% 3.9 1.4 6.0 1.7 2.1 1.5 20% 36% 3.8 2.9 4.4 2.5 4.1 1.4 23% 40% 4.3 3.4 5.2 3.1 1.1 5.0 26% 43% 2.9 3.0 2.9 2.8 3.9 1.9 30% 46% Table 4. Reproducibility results with different sparsity levels. The column of selected features represents the percentage of features preserved during the LASSO procedure, and the average improvement in accuracy, sensitivity, and specificity by sparsity. For LASSO-related problems, there is no closed-form solution for the selection of sparsity level; this is highly data dependent. To validate our MSW-LASSO model, we repeated Algorithm 1 and Algorithm 2 at different sparsity levels, which leads to preservation of different proportions of the features. The reproducibility performance of our proposed MSW-LASSO is summarized in Table 4. 4 Conclusion and Discussion Here we proposed a novel multi-site weighted LASSO model to heuristically improve classification performance for multiple sites. By sharing the knowledge of features that might help to improve classification accuracy with other sites, each site has multiple opportunities to reconsider its own set of selected features and strive to increase the accuracy at each iteration. In this study, the average improvement in classification accuracy is 4.9% for five sites. We offer a proof of concept for distributed machine learning that may be scaled up to other disorders, modalities, and feature sets. 5 References 1. World Health Organization. World Health Organization Depression Fact sheet, No. 369. (2012). Available from: http://www.who.int/mediacentre/factsheets/fs369/en/. 2. Fried, E.I., et al. "Depression is more than the sum score of its parts: individual DSM symptoms have different risk factors." Psych Med. 2067-2076 (2014). 3. Schmaal, L., et al. "Cortical abnormalities in adults and adolescents with major depression based on brain scans from 20 cohorts worldwide in the ENIGMA Major Depressive Disorder Working Group." Mol Psych. doi: 10.1038/mp.2016.60 (2016). 4. Schmaal, L., et al. "Subcortical brain alterations in major depressive disorder: findings from the ENIGMA Major Depressive Disorder working group." Mol Psych 806-812 (2016). 5. Liao, Y., et al. "Is depression a disconnection syndrome? Meta-analysis of diffusion tensor imaging studies in patients with MDD." J Psych & Neurosci. 49 (2013). 6. Sambataro, F., et al. "Revisiting default mode network function in major depression: evidence for disrupted subsystem connectivity." Psychl Med. 2041-2051 (2014). 7. Lo, A., et al. "Why significant variables aren’t automatically good predictors." PNAS. 13892-13897 (2015). 8. https://surfer.nmr.mgh.harvard.edu/ 9. Zhu, D., et al. Large-scale classification of major depressive disorder via distributed Lasso. Proc. of SPIE, 10160 (2017). 10. Tibshirani, R., “Regression shrinkage and selection via the LASSO.” Journal of the Royal Statistical Society. 58: 267–288 (1996). 11. Li, Qingyang, et al., "Large-Scale Collaborative Imaging Genetics Studies of Risk Genetic Factors for Alzheimer’s Disease Across Multiple Institutions." MICCAI. 335-343 (2016). 12. Zou, H., “The adaptive LASSO and its oracle properties.” J. Amer. Statist. Assoc 101(476):1418-1429 (2006). 13. Koutsouleris, N., et al. Individualized differential diagnosis of schizophrenia and mood disorders using neuroanatomical biomarkers. Brain, 138(7), 2059-2073 (2015). * * Supported in part by NIH grant U54 EB020403; see ref. 3 for additional support to coauthors for cohort recruitment.	5cs.CE
A CRITERION FOR I-ADIC COMPLETENESS arXiv:1312.3908v1 [math.AC] 13 Dec 2013 PETER SCHENZEL A BSTRACT. Let I denote an ideal in a commutative Noetherian ring R. Let M be an R-module. The I-adic completion is defined by M̂ I = lim α M/I α M. Then M is called I-adic complete ←− whenever the natural homomorphism M → M̂ I is an isomorphism. Let M be I-separated, i.e. ∩α I α M = 0. In the main result of the paper it is shown that M is I-adic complete if and only if Ext1R ( F, M ) = 0 for the flat test module F = ⊕ri=1 R xi where { x1 , . . . , xr } is a system of elements such that Rad I = Rad xR. This result extends several known statements starting with C. U. Jensen’s result (see [9, Proposition 3]) that a finitely generated R-module M over a local ring R is complete if and only if Ext1R ( F, M ) = 0 for any flat R-module F. 1. I NTRODUCTION Let R denote a commutative Noetherian ring. For an ideal I ⊂ R we consider the I-adic completion. For an R-module M it is defined by lim α M/I α M = M̂ I . In the case of ( R, m) ←− a local ring and a finitely generated R-module M it was shown (see [13]) that the following conditions are equivalent: (i) M is m-adic complete. (ii) Ext1R ( F, M ) = 0 for any flat R-module F. (iii) Ext1R ( F, M ) = 0 for F = ⊕ri=1 R xi , where x1 , . . . , xr ∈ m are elements that generate an m-primary ideal. Here R x denotes the localization of R with respect to { xα \|α ∈ N ≥0 }. This is an extension (in the local case) of a result of C. U. Jensen (see [9, Proposition 3]) who proved the equivalence of the first two conditions. The main result of the present paper is an extension to the case of I-adic completion. More precisely we prove the following result: Theorem 1.1. Let I be an ideal of a commutative Noetherian ring R. Let M denote an arbitrary Rmodule that is I-separated, i.e. ∩α I α M = 0. Then the following conditions are equivalent: (i) (ii) (iii) (iv) M is I-adic complete. ExtiR ( F, M ) = 0 for all i ≥ 0 and any flat R-module F with F ⊗ R R/I = 0. Ext1R ( R x , M ) = 0 for all elements x ∈ I. Ext1R (⊕ri=0 R xi , M ) = 0 for a system of elements x = { x1 , . . . , xr } such that Rad xR = Rad I. Note that in his paper (see [9, Proposition 3]) C. U. Jensen proved the following: Let R denote a semi local ring and M a finitely generated R-module. Then M is complete if and only if Ext1R ( F, M ) = 0 for all flat (resp. countably generated flat) R-modules F. A different proof for the vanishing of ExtiR ( F, M ) for all i ≥ 1, all flat R-modules and M a complete R-module follows by results of R.-O. Buchweitz and H. Flenner (see [5, Theorem 2]). 2010 Mathematics Subject Classification. Primary: 13J10 ; Secondary: 13C11, 13D07. Key words and phrases. adic completion, flat module, Ext. 1 2 PETER SCHENZEL Moreover another criterion for I-adic completeness was shown by A. Frankild and S. SatherWagstaff. Let I ⊂ R an ideal contained in the Jacobson radical of R. Let M be a finitely generated R-module. In their paper (see [7]) they proved that M is I-adically complete if and only if ExtiR ( R̂ I , M ) = 0 for all 1 ≤ i ≤ dimR M, where R̂ I denotes the I-adic completion of R. Whence, the main result of the present paper is the construction of a simple flat test module F for the completeness of M in the I-adic topology in terms of the vanishing of Ext1R ( F, M ). This generalizes the case of the maximal ideal proved in [13]. Instead of Matlis Duality used in [13] (not available in the case of I-adic topology) here we use some homological techniques. 2. P RELIMINARY R ESULTS In the following R denotes a commutative Noetherian ring. For an ideal I ⊂ R we consider the I-adic completion. For an R-module M it is defined by lim α M/I α M = M̂ I . We say that M ←− I : M → M̂ I (induced by is I-adically complete whenever the natural homomorphism τ = τM α the natural surjections M → M/I M) is an isomorphism. It implies that M is I-separated, i.e. ∩α I α M = 0. We begin with a few preparatory results needed later on. Proposition 2.1. Let I ⊂ R denote an ideal. Let M denote an R-module such that M ⊗ R R/I = 0. Then M ⊗ R X = 0 for any R-module X with SuppR X ⊆ V ( I ). Proof. The assumption M ⊗ R R/I = 0 implies that M = I α M for all α ≥ 1, or in other words M ⊗ R R/I α = 0. First of all let X be a finitely generated R-module. Then I α X = 0 for a certain α ∈ N. Therefore M ⊗ R X = M ⊗ R X/I α X = ( M ⊗ R R/I α ) ⊗ R X = 0. If X is not necessarily a finitely generated R-module, then X ≃ lim Xα for a direct system of −→ finitely generated submodules Xα ⊂ X ordered by inclusions. Whence SuppR Xα ⊆ V ( I ) and M ⊗ R X = lim( M ⊗ R Xα ) = 0, −→ as required. Let I ⊂ R denote an ideal. For an R-module X the left derived functors ΛiI ( X ), i ∈ Z, are defined by Hi (lim α ( R/I α ⊗ R F· )), where F· denotes a flat resolution of X. The functors −→ ΛiI (·) were first systematically studied by Greenlees and May (see [8]) and by A.-M. Simon (see [14]) and more recently in [1] and [11]. Note that there is a natural surjective homomorphism Λ0I ( X ) → X̂ I . Since R is Noetherian it follows that M/I α M ≃ M̂ I /I α M̂ I for all α ≥ 1 for any R-module M (see [16, Section 2.2]). Whence the completion M̂ I is I-adically complete. In the following we will discuss the assumption on the R-module M in Proposition 2.1. It provides a certain kind of a Nakayama Lemma. Its proof is known (see [14, 5.1] and also [15, 1.3, Lemma (ii)]). Here we add these arguments. Proposition 2.2. Let I denote an ideal in a commutative Noetherian ring R. Let M denote an arbitrary R-module. Then M ⊗ R R/I = 0 if and only if M̂ I = 0 and if and only if Λ0I ( M ) = 0. Proof. If M ⊗ R R/I = 0, then M/I α M = 0 for all α ≥ 1 and M̂ I = lim α M/I α M = 0. Con←− versely, suppose that M̂ I = 0. Then 0 = M̂ I /I M̂ I ≃ M/I M, as required. If Λ0I ( M ) = 0, then clearly M̂ I = 0. Let F1 → F0 → M → 0 denote a free presentation of M. If Λ0I ( M ) = 0, then F̂1I → F̂0I is onto. Therefore F1 /IF1 → F0 /IF0 is onto too, i.e. M/I M = 0 as required. A CRITERION FOR I-ADIC COMPLETENESS 3 In the following let x = { x1 , . . . , xr } denote a system of elements of R such that Rad xR = Rad I. Then we consider the Čech complex Čx of the system x (see e.g. [11, Section 3]). Note that this is the same as the complex K∞ ( x) introduced in [1] as the direct limit of the Koszul complexes K· ( xn ). There is a natural morphism Čx → R. In our context the importance of the Čech complex is the following result. It allows the expression of the right derived functors ΛiI (·) in different terms. Theorem 2.3. Let I ⊂ R denote an ideal of a commutative Noetherian ring R. Let M denote an R-module. Then there are natural isomorphisms ΛiI ( M ) ≃ Hi (HomR (Čx , ER ( M )· )) for all i ∈ Z, where ER ( M )· denotes an injective resolution of M. Proof. This statement is one of the main results of [11, Section 4]. In the formulation here we do not use the technique of derived functors and derived categories. (For a more advanced exposition based on derived categories and derived functors the interested reader might also consult [1].) Let us summarize a few basic results on completions used in the sequel. The following results are well-known. Proposition 2.4. Let J ⊂ I denote ideals of a commutative Noetherian ring. Let x = { x1 , . . . , xr } denote a system of elements of R. Let M denote an R-module. (a) Suppose that M is I-adic complete. Then it is also J-adic complete. (b) Suppose that Rad I = Rad xR. Assume that M is complete in the xi R-adic topology for i = 1, . . . , r and M is I-separated. Then M is I-adic complete. Proof. For the proof of (a) see [16, 2.2.6]. Let M̂ I denote the I-adic completion of M. Because the I-adic topology and the topology defined by {( x1α , . . . , xrα ) R}α≥0 are equivalent it follows M̂ I ≃ lim α M/( x1α , . . . , xrα ) M. Therefore an element m ∈ M̂ I has the form m = ∑α≥0 ∑ri=1 xiα mα,i ←− with mα,i ∈ M. Since M is xi R-adically complete M ≃ lim α M/xiα M and mi = ∑α xiα mα,i for ←− some mi ∈ M. Then m′ = ∑ri=1 mi ∈ M maps to m under τ : M → M̂ I . Because M is I-separated it is I-adically complete. This proves (b). The following result is helpful in order to compute projective limits of Ext-modules. Here lim 1 denotes the first right derived functor of the projective limit (see [17]). ←− Let M denote an R-module and x ∈ R. We consider the following projective system { M, x}, x where Mα = M for all α ∈ N and the transition map Mα+1 → Mα is the multiplication by x. Lemma 2.5. With the previous notation it follows that lim i { M, x} ≃ ExtiR ( R x , M ) for i = 0, 1 and ExtiR ( R x , M ) = 0 for all i ≥ 2. ←− If M is xR-separated, then lim { M, x} = HomR ( R x , M ) = 0. ←− Proof. Let { R, x} denote the direct system with Rα = R for all α ∈ N and the transition map x Rα → Rα+1 is the multiplication by x. Then R x ≃ lim α { R, x} as it is well known. Therefore, −→ HomR ({ R, x}, M ) ≃ { M, x}. Now, by [12, Lemma 2.6] (see also Lemma 4.1) there are short exact sequences 0 → lim 1 {ExtiR−1 ( R, M ), x} → ExtiR ( R x , M ) → lim{ExtiR ( R, M ), x} → 0 ←− ←− for all i ∈ Z. Whence the results of the first part follow. If M is xR-separated, the vanishing of HomR ( R x , M ) is easy to see. 4 PETER SCHENZEL The previous result is a slight modification of Lemma [13, Lemma 2.7]. Remark 2.6. A morphism of two complexes X · → Y · of R-modules is called a quasi-isomorphism (or homology isomorphism) whenever the induced homomorphisms on the cohomology modules H i ( X · ) → H i (Y · ) are isomorphisms for all i ∈ Z. For the definitions of HomR ( X · , Y · ) and X · ⊗ R Y · we refer to [17] and [2]. For the details on homological algebra of complexes of modules we refer to [2]. 3. O N I- ADIC C OMPLETIONS Let I ⊂ R denote an ideal of a commutative Noetherian ring R. In the first main result we shall prove a vanishing result for a certain Ext-module for an I-adic complete R-module. Theorem 3.1. Let M denote an arbitrary R-module. Let F denote a flat R-module satisfying F ⊗ R R/I = 0. Then ExtiR ( F, M̂ I ) = 0 for all i ∈ Z. Proof. For the proof we use the techniques summarized in Theorem 2.3. To this end we fix a few notation. Let x = { x1 , . . . , xr } denote a system of elements of R such that Rad xR = Rad I. Let Čx denote the Čech complex with respect to x as defined in [11, Section 3]. It is a bounded complex of flat R-modules with a natural morphism Čx → R of complexes. Let E· denote an injective resolution of M̂ I as an R-module. By applying HomR (·, E· ) it induces a morphism of complexes E· → HomR (Čx , E· ). Next we investigate the complex HomR (Čx , E· ). It is a left bounded complex of injective Rmodules. Moreover, by virtue of Theorem 2.3 it follows that ΛiI ( M̂ I ) ≃ Hi (HomR (Čx , E· )) for all i ∈ Z. Therefore Hi (HomR (Čx , E· )) = 0 for all i 6= 0 and H0 (HomR (Čx , E· )) ≃ M̂ I since M̂ I is I-adic complete and ΛiI ( M̂ I ) = 0 for all i > 0 (see [14, 5.2]). Therefore, the morphism E· → HomR (Čx , E· ) is a quasi-isomorphism. Now we consider the complexes HomR ( F, HomR (Čx , E· )) ≃ HomR ( F ⊗ R Čx , E· ). We claim that they are homologically trivial complexes. For that reason it will be enough to show that F ⊗ R Čx is homologically trivial since E· is a bounded below complex of injective R-modules. Since F is a flat R-module it yields that H i ( F ⊗ R Čx ) ≃ F ⊗ R H i (Čx ) = 0 for all i ∈ Z. For the vanishing apply Proposition 2.1 with H i (Čx ) ≃ H Ii ( R) for all i ∈ Z and SuppR H Ii ( R) ⊆ V ( I ). As shown above E· → HomR (Čx , E· ) is a quasi-isomorphism of left bounded complexes of injective R-modules. Applying the functor HomR ( F, ·) induces a quasi-isomorphism HomR ( F, E· ) → HomR ( F, HomR (Čx , E· )). The complex at the right is cohomologically trivial, while the cohomology of the complex at the left is ExtiR ( F, M̂ I ). Therefore, the Ext-modules vanish, as required. In the following result we prove another behaviour of certain Ext-modules in respect to the I-adic completion. Theorem 3.2. Let I ⊂ R denote an ideal. Let M, X be arbitrary R-modules with τ : M → M̂ I the natural map. Suppose that SuppR X ⊆ V ( I ). Then (a) ExtiR ( X, M̂ I /τ ( M )) = 0 and (b) the natural homomorphism ExtiR ( X, M ) → ExtiR ( X, M̂ I ) is an isomorphism for all i ∈ N . A CRITERION FOR I-ADIC COMPLETENESS 5 Proof. Let x = { x1 , . . . , xr } denote elements of R such that Rad xR = Rad I. Let Čx denote the Čech complex with respect to x. Then there is a short exact sequence of complexes 0 → Dx [−1] → Čx → R → 0, where Dx is the global Čech complex. That is Dxi = Čxi+1 for i ≥ 0 and Dxi = 0 for i < 0. Now let E· denote an injective resolution of M. By applying the functor HomR (·, E· ) to the short exact sequence of complexes it provides a short exact sequence 0 → E· → HomR (Čx , E· ) → HomR ( Dx [−1], E· ) → 0 of left bounded complexes of injective R-modules. The complex in the middle is an injective resolution of M̂ I as follows by 2.3 and the fact that M̂ I is I-adic complete (see [14, 5.2]). Therefore the complex at the right is quasi-isomorphic to M̂ I /τ ( M ) considered as a complex concentrated in homological degree zero. In order to prove the statement in (a) let L· denote a projective resolution of X. Then ExtiR ( X, M̂ I /τ ( M )) ≃ H i (HomR ( L· , HomR ( Dx [−1], E· ))) (see [2]) and it will be enough to show that the last modules vanish for all i ∈ Z. To this end consider the isomorphism of complexes HomR ( L· , HomR ( Dx [−1], E· )) ≃ HomR ( L· ⊗ R Dx , E· )[1]. Now Dx is a bounded complex of flat R-modules. Therefore there is the quasi-isomorphism L· ⊗ R Dx → X ⊗ R Dx . It will be enough to show that the complex X ⊗ R Dx is homologically trivial. But this is true since X ⊗ R Dxi = 0 for all i ∈ Z because of SuppR X ⊆ V ( I ). This proves the statement in (a). By view of the above investigations there is a quasi-isomorphism E· → HomR (Čx , E· ) where the second complex is an injective resolution of M̂ I . This proves the statement in (b). We will continue with a result on the vanishing of certain Ext-modules. As a step towards to our main result we shall prove a partial result in order to characterize the completion. Theorem 3.3. Let R denote a commutative Noetherian ring. Let M denote an arbitrary R-module. Let x ∈ R be an element such that 0 : M xα = 0 : M x β for all α ≥ β. Then the following conditions are equivalent: (i) M is xR-adic complete. (ii) ExtiR ( R x , M ) = 0 for i = 0, 1. Proof. Let N = ∪α≥1 0 : M xα . Then N = 0 : M x β as follows by the definition of β. That is, in both of the statements we may replace x by x β without loss of generality. That is, we may assume that xN = 0. Then there is a commutative diagram with exact rows: x α +1 0 → M/N −→ M → M/xα+1 M → 0 ↓x k ↓ xα 0 → M/N −→ M → M/xα M → 0. By passing to the inverse limit it provides an exact sequence 0 → lim { M/N, x} → M → M̂ x → lim 1 { M/N, x} → 0. ←− ←− By view of Lemma 2.4 it yields that lim i { M/N, x} ≃ ExtiR ( R x , M/N ) for i = 0, 1. Furthermore, ←− the short exact sequence 0 → N → M → M/N → 0 induces an isomorphism ExtiR ( R x , M ) ≃ ExtiR ( R x , M/N ). This follows by the long exact cohomology sequence and ExtiR ( R x , N ) = 0 for 6 PETER SCHENZEL all i ∈ Z. For this vanishing note that the multiplication by x acts on R x as an isomorphism and on N as the zero map. That is, the homomorphism M → M̂ x is an isomorphism, if and only if ExtiR ( R x , M ) = 0 for i = 0, 1 which proves (i) ⇐⇒ (ii). Proof of Theorem 1.1. The implication (i) =⇒ (ii) is a consequence of Theorem 3.1. Moreover, the implications (ii) =⇒ (iii) is trivial and (iii) =⇒ (iv) is easy to see. In order to prove (iv) =⇒ (i) we first note that Ext1R ( R xi , R) = 0 for all i = 1, . . . , r. By Theorem 3.3 this implies that M is xi Radically complete for i = 1, . . . , r. Since x = { x1 , . . . , xr } generates I up to the radical it follows that M is I-adic complete by Proposition 2.4 since M is I-adic separated. This completes the proof. 4. A N ALTERNATIVE PROOF OF T HEOREM 3.1 In this section there is an alternative proof of Theorem 3.1 based on some results on inverse limits. The results of the following two lemmas might be also of some independent interest. These statements are particular cases of certain spectral sequences on inverse limits (see [10]). Here we give an elementary proof based on the description of lim 1 as discussed in [17]. ←− Lemma 4.1. Let R denote a commutative ring. Let { Mα }α∈N be a direct system of R-modules. Let N denote an arbitrary R-module. Then there is a short exact sequence 0 → lim 1 ExtiR−1 ( Mα , N ) → ExtiR (lim Mα , N ) → lim ExtiR ( Mα , N ) → 0 ←− −→ ←− for all i ∈ Z. Proof. This result was proved in ([12, Lemma 2.6]) for a Noetherian ring. The same arguments work in the general case. In the following we need a certain dual statement of Lemma 4.1. In a certain sense it will be the key argument for the second proof of Theorem 3.1. Lemma 4.2. Let R denote a commutative ring. Let M denote an arbitrary R-module. Let { Nα }α∈N be an inverse system of R-modules with lim 1 Nα = 0. Then there is a short exact sequence ←− 1 i−1 0 → lim ExtR ( M, Nα ) → ExtiR ( M, lim Nα ) → lim ExtiR ( M, Nα ) → 0 ←− ←− ←− for all i ∈ Z. Proof. Because of lim 1 Nα = 0 there is a short exact sequence ←− 0 → lim Nα → ∏ Nα → ∏ Nα → 0, ←− where the third homomorphism is the transition map (see [17]). It induces a long exact cohomology sequence f · · · → ∏ ExtiR−1 ( M, Nα ) → ∏ ExtiR−1 ( M, Nα ) → ExtiR ( M, lim Nα ) ←− g i i → ∏ ExtR ( M, Nα ) → ∏ ExtR ( M, Nα ) → · · · . To this end recall that Ext transforms direct products into direct products in the second variable and cohomology commutes with direct products (see e. g. [6]). Now it is known (see e.g. [17]) that Coker f ≃ lim 1 ExtiR−1 ( M, Nα ) and Ker g = lim ExtiR ( N, Nα ). ←− ←− This completes the proof. A CRITERION FOR I-ADIC COMPLETENESS 7 Remark 4.3. The assumption that lim 1 Nα = 0 is fulfilled whenever the projective system { Nα } ←− satisfies the Mittag-Leffler condition. That is, for instance when the transition map Nα+1 → Nα is surjective for all α ≥ 1. Note that the proof of the following Theorem is motivated by some arguments done by Buchweitz and Flenner (see [5]). Theorem 4.4. Let R denote a commutative ring. Let M denote an R-module. Let F denote a flat R-module satisfying F ⊗ R R/I = 0. Then ExtiR ( F, M̂ I ) = 0 for all i ∈ Z. Proof. By definition we have M̂ I = lim α M/I α M. By Lemma 4.2 there is a short exact sequence ←− 0 → lim 1 ExtiR−1 ( F, M/I α M ) → ExtiR ( F, M̂ I ) → lim ExtiR ( F, M/I α M ) → 0. ←− ←− i I In order to show the vanishing of ExtR ( F, M̂ ) for all i ∈ Z it will be enough to show the vanishing of Exti ( F, M/I α M ) for all i ∈ Z and all α ≥ 1. We claim that ExtiR ( F, M/I α M ) ≃ ExtiR/I α ( F/I α F, M/I α M ) for all i ∈ Z and all α ≥ 1. In order to show these isomorphisms let L· be a projective resolution of F as an R-module. As F is flat as an R-module ToriR ( F, R/I α ) = 0 for all i > 0 and L· ⊗ R R/I α is a projective resolution of F/I α F as an R/I α -module. By adjunction there are isomorphisms of complexes HomR/I α ( L· ⊗ R R/I α , M/I α M ) ≃ HomR ( L· , M/I α M ). By taking cohomology it proves the above claim. The vanishing follows now because of F/I α F = 0 as a consequence of the assumption F ⊗ R R/I = 0. Remark 4.5. In fact Theorem 4.4 is a slight sharpening of Theorem 3.1. To this end recall that for an R-module M and an ideal I ⊂ R the I-adic completion M̂ I is not necessarily I-adic complete (as used in the proof of Theorem 3.1). For an explicit example see J. Bartijn’s Thesis [3, I, §3, page 19]. Note that in this example the ring is not Noetherian. It grows out of [4, III, §2, Exerc. 12]. See also A. Yekutieli (see [18, Example 1.8]). A CKNOWLEDGEMENTS The author thanks Anne-Marie Simon and the reviewer for a careful reading of the manuscript and suggesting some comments. R EFERENCES [1] L. A LONSO TARR ÎO , A. J EREMIAS L ÓPEZ , J. L IPMAN : ‘Local homology and cohomology of schemes’, Ann. scient. Éc. Norm. Sup., 4e sér. 30 (1997), 1-39. 2, 3 [2] L. L. AVRAMOV, H.-B. F OXBY: ‘Homological dimensions of unbounded complexes’, J. Pure Appl. Algebra 71 (1991), 129-155. 4, 5 [3] J. B ARTIJN : ‘Flatness, completions, regular sequences, un ménage à trois, Thesis, Utrecht, 1985. 7 [4] N. B OURBAKI : ‘Algebrè commutative’, Hermann, Paris, 1961. 7 [5] R.-O. B UCHWEITZ , H. F LENNER : ‘Power series rings and projectivity’, Manuscripta math. 119 (2006), 107114. 1, 7 [6] E. E. E NOCHS , O. J ENDA : ‘Relative Homological Algebra’, 2nd revised ed., Walter de Gruyter, Berlin, 2011. 6 [7] A. F RANKILD , S. S ATHER -WAGSTAFF : ‘Detecting completeness from Ext-Vanishing’, Proc. Amer. Math. Soc. 136 (2008), 2303-2312. 2 [8] J. P. C. G REENLEES , J. P. M AY: ‘Derived functors of the I-adic completion and local homology’, J. Algebra 149 (1992), 438-453. 2 [9] C. U. J ENSEN : ‘On the vanishing of lim ( i) ’, J. Algebra 15 (1970), 151-166. 1 ←− 8 PETER SCHENZEL [10] C. U. J ENSEN : ‘Les foncteurs dérivés de lim et leurs applications en théorie des modules’, Lecture Notes in ←− Math. 254, Springer, 1972. 6 [11] P. S CHENZEL : ‘Proregular sequences, local cohomology, and completion’, Math. Scand. 92 (2003), 181-180. 2, 3, 4 [12] P. S CHENZEL : ‘On the structure of the endomorphism ring of a certain local cohomology module’, J. Algebra 344 (2011), 229-245. 3, 6 [13] P. S CHENZEL : ‘A criterion for completeness’, to appear. 1, 2, 4 [14] A.-M. S IMON : ‘Some homological properties of complete modules’, Math. Proc. Cambr. Phil. Soc. 108 1990, 231-246. 2, 4, 5 [15] A.-M. S IMON : ‘Adic - completion and some dual homological results’, Publ. Matematiques Univ. Autonoma de Barcelona 36, No. 2, 1992, 965-979. 2 [16] J. S TROOKER : ‘Homological questions in local algebra’, London Math. Lect. Note Ser. 145, Cambridge Univ. Press, 1990. 2, 3 [17] C. W EIBEL : ‘An introduction to homological algebra’, Cambridge Univ. Press, 1994. 3, 4, 6 [18] A. Y EKUTIELI : ‘On flatness and completion for infinitely generated modules over noetherian rings’, Comm. Algebra 39, No. 11, 2011, 4221-4245. 7 M ARTIN -L UTHER -U NIVERSIT ÄT H ALLE -W ITTENBERG , I NSTITUT (S AALE ), G ERMANY E-mail address: peter.schenzel@informatik.uni-halle.de F ÜR I NFORMATIK , D — 06 099 H ALLE	0math.AC
Real-Time Localization and Tracking of Multiple Radio-Tagged Animals with an Autonomous UAV arXiv:1712.01491v4 [cs.SY] 14 Feb 2018 Hoa Van Nguyen, Michael Chesser, Fei Chen, S. Hamid Rezatofighi and Damith C. Ranasinghe School of Computer Science The University of Adelaide, SA 5005, Australia Email:{hoavan.nguyen,michael.chesser,fei.chen,hamid.rezatofighi,damith.ranasinghe}@adelaide.edu.au Abstract—Autonomous aerial robots provide new possibilities to study the habitats and behaviors of endangered species through the efficient gathering of location information at temporal and spatial granularities not possible with traditional manual survey methods. We present a novel autonomous aerial vehicle system to track and localize multiple radio-tagged animals. The simplicity of measuring the received signal strength indicator (RSSI) values of very high frequency (VHF) radio-collars commonly used in the field is exploited to realize a low cost and lightweight tracking platform suitable for integration with unmanned aerial vehicles (UAVs). Due to uncertainty and the nonlinearity of the system based on RSSI measurements, our tracking and planning approaches integrate a particle filter for tracking and localizing; a partially observable Markov decision process (POMDP) for dynamic path planning. This approach allows autonomous navigation of a UAV in a direction of maximum information gain to locate multiple mobile animals and reduce exploration time; and, consequently, conserve on-board battery power. We also employ the concept of a search termination criteria to maximize the number of located animals within power constraints of the aerial system. We validated our real-time and online approach through both extensive simulations and field experiments with two mobile VHF radio-tags. I. I NTRODUCTION Understanding basic questions of ecology such as how animals use their habitat, their movements and activities are necessary for addressing numerous environmental challenges ranging from invasive species to diseases spread by animals and saving endangered species from extinction. Conservation biologists, ecologists as well as natural resource management agencies around the world rely on numerous methods to monitor animals. The most predominant, cost-effective tool suitable for a wide range of animal sizes is very high frequency (VHF) radio collar tagging and subsequent radio telemetry measurements to locate spatiotemporal information of concerned species [10, 20]. However, the traditional method of radio tracking typically requires researchers to trek long distances in the field, armed with cumbersome VHF radio receivers with handheld antennas and battery packs to manually home in on radio signals emitted from radio-tagged or collared animals. Consequently, the precious spatial data acquired through radio tracking comes at a significant cost to researchers in terms of manpower, time and funding. The problem is often compounded by other challenges, such as low animal recapture rates, equipment failures, and the inability to track animals that move into inaccessible terrain. Furthermore, many of our Fig. 1. An overview of the UAV tracking platform with its sensor system. most endangered species also happen to be the most difficult to track due to their small size, inconspicuousness, and location in remote habitats. Automated tracking and location of wildlife with autonomous unmanned aerial vehicles (UAVs) can provide new possibilities to better understand ecology and our native wildlife to safeguard biodiversity and manage our natural resources cost-effectively. We present a low-cost approach capable of realization in a lightweight payload for transforming existing commodity drone platforms into autonomous aerial vehicle systems as shown in Fig. 1 to empower conservation biologists to track and localize multiple radio-tagged animals. The main contribution of our work is a new autonomous aerial vehicle system for simultaneously tracking and localizing multiple mobile radio-tagged animals using VHF radiocollars, commonly used in the field. In particular: i) our system is realized in a 260 g payload suitable for a multitude of low-cost, versatile, easy to operate multi-rotor UAVs without a remote pilot license. Our lightweight realization—of less than 2 kg system mass—is achieved through a new sensor design that exploits the simplicity of a software defined radio architecture for capturing received signal strength indicator (RSSI) value from multiple VHF radio tags and a compact, lightweight VHF antenna geometry; ii) due to the noisy, complex and nonlinear characteristics of RSSI data we integrate a sequential Monte Carlo implementation of a Bayesian filter, also known as particle filter (PF), for real-time tracking and localization jointly with a partially observable Markov decision process (POMDP) with Réyni divergence between prior and posterior estimates of animal locations for autonomy and dynamic online path planning to minimize flight time while maximizing number of located animals. Further, our formulation considers the trade-off between location accuracy and resource constraints of the UAV, its maneuverability, and power constraints to develop a practical solution; and iii) we validate our method through extensive simulations and field experiments with mobile VHF radio-tags. II. R ELATED W ORK Since this application is related to locating animals, we will focus on progress made towards the autonomous localization and tracking of multiple radio-tagged animals here. Off-line estimations of a radio beacon from data logged from a UAV have been demonstrated in [8, 23]. Pioneering achievements in autonomous wildlife tracking have been made through simulation studies [14] and experimentally demonstrated systems [4, 11, 21, 22] in recent years. In particular, the first demonstration of a UAV for proved in [4]. The recent approaches [4, 22] for real-time localization of a static target (assuming stationary wildlife) used wireless signal characteristics captured by a receiver to estimate location; in particular, the angle-of-arrival (AoA) of a radio beacon was determined using an array of antennas with the information related to a ground-based receiver for location estimations. Although the approach can conveniently manage topological variations in terrain, AoA systems require a large bulky receiver system and multiple antenna elements as well as long observation times; 45 seconds per observation as reported in [4]. Moreover, the antenna systems being mounted on top of the UAV [4] is likely to lead to difficulty in tracking terrestrial animals although being suitable for locating avian species dwelling in trees. We can see that there are few investigations that have studied the problem of locating radio-collared animals using autonomous robots. Although a system based on angle-ofarrival was recently evaluated to locate a stationary animal, the development of a low-cost and lightweight autonomous system capable of long-range flights and localization of multiple mobile radio-collared animals still remains. This is especially significant in the realization of a system that is widely accessible to conservation biologists in the field where a very small UAV—of less than 2 kg—can be flown without a formal pilot license and with fewer restrictions given the exclusion of this category of UAVs from regulatory regimes [3, p.18]. We present an alternative approach exploiting RSSI based range only measurements because of the ability to use a simpler sensing system on board commodity UAVs to realize lower cost and longer flight range UAVs for tracking and localizing multiple animals. Together with a strongly principled approach for joint tracking and planning, our lightweight autonomous aerial robot platform provides a cost-effective method for wildlife conservation and management. III. T RACKING AND P LANNING P ROBLEM F ORMULATION Real-time tracking requires an online estimator and a dynamic planning method. This section presents our tracking and localizing formulation under a principled framework of a Bayesian filter for tracking and POMDP for planning strategy. A. Tracking and localizing For tracking, we use a Bayesian filter. It is an online estimation technique which deals with the problem of inferring knowledge about the unobserved state of a dynamic system— in our problem, wildlife—which changes over time, from a sequence of noisy measurements. Suppose x ∈ X and z ∈ Z are respectively the system (kinematic) state vector in the state space X and the measurement (observation) vector in the observation space Z. The problem is estimating the state x ∈ X from the measurement z ∈ Z or calculating the marginal posterior distribution p(xk \|z1 :k ) sequentially through prediction Z(1) and update (2) steps. p(xk \|z1 :k −1 ) = p(xk \|xk −1 )p(xk −1 \|z1 :k −1 )d xk −1 (1) p(xk \|z1 :k ) = R p(zk \|xk )p(xk \|z1 :k −1 ) p(zk \|xk )p(xk \|z1 :k −1 )d xk (2) In the case of a nonlinear system or non-Gaussian noise, there is no general closed-form solution for the Bayesian recursion and p(xk \|z1 :k ) generally has a non-parametric form. Therefore, in our problem, we use a particle filter implementation as an approximate solution for the Bayesian filtering problem due to our highly nonlinear measurement model. Particle Filter (PF): A particle filter uses a sampling approach to represent the non-parametric form of the posterior density p(xk \|z1 :k ). The samples from the distribution are represented by a set of particles; each particle has a weight assigned to represent the probability of that particle being sampled from the probability density function. Then, these particles representing the nonparametric form of p(xk \|z1 :k ) are propagated over time. In the simplest version of the particle filter, known as the bootstrap filter first introduced by Gordon in [5], the samples are directly generated from the transitional dynamic model. Then, to reduce the particle degeneracy, resampling and injection techniques are implemented; a detailed algorithm can be found in [17]. Measurement model: The update process of a PF requires the derivation of a likelihood of measurements. In our problem, based on estimating a target’s—VHF radio tag’s—range from the receiver, we require a realistic signal propagation model to obtain the likelihood of receiving a given measurement. We employ two VHF signal propagation models suitable for describing RSSI measurements in non-urban outdoor environments. Denoting h(x, u) as the RSSI measurement function between target x and observer (UAV) state u, we have: i) Log Distance Path Loss Model (LogPath): The received power is the only line of sight power component transmitted from a transmitter subjected to signal attenuation such as through absorption and propagation loss: h(x, u) = Prd0 − 10n log10 (d(x, up )/d0 ) + Gr (x, u) (3) where t t t T is the target’s position; up = • x = [px , py , pz ] u u u T [px , py , pz ] is the observer’s (UAV) position in Cartesian coordinates; u = [up ; θu ] is the UAV’s state which includes its heading angle θu . • • • • d(x, up ) is the euclidean distance between the target’s position and UAV’s position. Gr (x, u) is the UAV receiver antenna gain which depends on its heading, its position, and target’s position. Prd0 is received power at distance d0 . n is the path-loss exponent that characterizes the signal losses such as absorption and propagation losses and this parameter depends on the environment with typical values range from 2 to 4. ii) Log Distance Path Loss Model with Multi-Path Fading (MultiPath): The received power is composed of both line of sight power component transmitted from a transmitter and the multi-path power component reflected from the ground plane subjected to signal attenuation such as through absorption and propagation loss: [24, p. 81]: h(x, u) = Prd0 − 10n log10 (d(x, up )/d0 ) + Gr (x, u) + 10n log10 (\|1 + Γ(ψ)e−j4ϕ \|) (4) where, in addition to terms in 3 • ψ is the contact angle of the reflected path and ground. p εg − cos2 (ψ)]/[sin(ψ) + • Γ(ψ) = [sin(ψ) − p 2 εg − cos (ψ)] is the ground reflection coefficient with εg is the relative permittivity of the ground. • 4ϕ = 2π4d/λ is the phase difference between two waves where λ is the wave length and 4d = ((ptx − pux )2 + (pty − puy )2 + (ptz + puz )2 )1/2 − d(x, u). In non-urban environments, received power is usually corrupted by environmental noise, with the assumption that the noise is white, the total received power z = Pr (x, u) [dB] is: z = h(x, u) + ηP (5) N (0, σP2 ) where ηP ∼ is Gaussian white noise with covariance σP2 . Notably, even if RSSI noise is not completely characterized by a white noise model, we can still model it as Gaussian noise by using a higher covariance σP2 to account for the unknown uncertainty. We use data captured in experiments using our sensor system to validate the physical sensor characteristics Gr (x, u) (see Sec. IV) and n defined by environmental characteristics, as well as estimate the propagation model reference power parameter Prd0 and noise σP (see Sec. VI-B). Measurement likelihood: Based on (5) with Gaussian noise ηP , the likelihood of measurement zk , given target and sensor position are xk and uk , respectively, at time k is p(zk \|xk , uk ) ∼ N (zk ; h(xk , uk ), σP2 ) where N (z; µ, Σ) is a normal distribution with mean µ and covariance Σ. B. Path Planning The UAV planning problem is similar to the problem of an agent computing optimal actions under partially observable Markov decision process (POMDP) to maximize its reward. Kaelbling et al. [9] have shown that a POMDP framework implements an efficient and optimal approach based on previous actions and observations to determine the true world states. POMDP in conjunction with a particle filter provides a principled approach for evaluating planning decision to realize an autonomous system for tracking. In general, a POMDP is described by the 6-tuple (S, A, T , R, O, Z) where S is set of both UAV and target states (s = {x, u} ∈ S), A is set UAV of actions, T is state-transition function T (s, a, s0 ) = p(s0 \|s, a), R(s, a) is reward function, O is set of observations and Z is observation function Z(s, a, o) = p(o\|s, a) with s, s0 ∈ S is current state and next state respectively, a ∈ A is taken action and o ∈ O is the observation—i.e measurement. The goal of a POMDP is to find P an optimal policy to maximize the total k+H expected reward E[ t=k γ k−t R(st , at )] where H is lookahead horizon steps, γ ∈ (0, 1) is the discount factor which serves as the value difference between the current reward versus the future reward; st , at is state and action at time step t and E[·] is the expectation operator [7]. The reward function can be calculated using different methods such as task-driven or information-driven strategies. When uncertainty is high, the information gain approach is preferable to reduce a target’s location uncertainty [2]; hence, we used this method to calculate our reward function. There are several approaches to evaluate information gain in robotic path planning such as Shannon entropy [4], KullbackLeibler (KL) divergence or Rényi divergence [6]. We adapted the approach in [15, 16] to implement Rényi divergence as our reward function since it fits naturally with our MonteCarlo sampling method. Here, Rényi divergence calculates the distance between prior and posterior densities [2, 16]: Z α 1 (i) log Rk+H (a) = p(xk+H \|z1:k ) α−1 (6) h i1−α (i) × p(xk+H \|z1:k , zk+1:k+H (a)) dx where the prior density p(xk+H \|z1:k ) is calculated by propagating current posterior particles sampled from p(xk \|z1:k ) to time k +H using the prediction step (1). The posterior density (m) (m) p(xk+H \|z1:k , zk+1:k+H (a)) where zk+1:k+H (a) is the future measurement set that will be observed if action a ∈ Ak is taken; this is calculated by applying both prediction (1) and update steps (2) up to time k + H. However, using Bayes update procedure is computationally expensive and prohibitive in a real-time setting. Instead, we implement a computationally efficient approach using a black box simulator proposed in [19] along with the Monte Carlo sampling approach. Hence, the problem transforms to find an optimal action a ∈ Ak to maximize total expected reward: k+H Ms 1 X X (m) γ k−t Rt (a) (7) ak ≈ arg max a∈Ak Ms m=1 t=k where Ms is the number of future measurements. C. Multi-targets Tracking The particle filter proposed in Sec. III-A can be extended to multi-target tracking (MTT). However, MTT normally deals with the complex data association problem where it is difficult to determine which measurement belongs to which target. In contrast, for our system, each target can be estimated from the measurement based on the signal frequency and tracked independently. Thus, we do not need to solve the data association problem. Notably, not all of the targets are detected due to, for example, the UAV movements, the measurement range limits imposed by propagations losses and receiver sensitivity. Therefore, if the target is not detected, the solver does not update its estimated position. Besides maximizing the number of targets localized and tracked, we formulated a termination condition for each target to conserve UAV battery power; a target is considered localized if its location uncertainty—as determined by its particle covariance—is sufficiently small (<NT h ). Then, those found targets are forgotten to aid the solver prioritize its computing resources on those targets with high uncertainty. IV. S YSTEM I MPLEMENTATION We implemented an experimental aerial robot system based on our tracking and planning formulation. An overview of the complete system is described in Fig. 2a. Our experimental system used a 3DR IRIS+ UAV platform and a new sensor system built with: i) a compact directional VHF antenna design; and ii) a software defined signal processing module capable of simultaneously processing signals from multiple targets and remotely communication with a ground control system for tracking and planning. In our system, the ArduPilotMega (APM) on the IRIS+ UAV transmits back its global positioning system (GPS) location to the Telemetry Host tool developed by our group to communicate with the APM module using MAVLink protocol over a 915 MHz full duplex radio channel. The sensor system together with the Antenna, SDR receiver and the embedded compute delivers targets’ RSSI data through a 2.40 GHz radio channel to the ground control system. GPS and targets’ RSSI data are delivered to our tracking and planning algorithm—solver—through the Telemetry Host using a RESTful web service. The solver estimates target locations and calculates new control actions per each POMDP cycle to command the UAV through MAVLink to fly to a new location. In order to ensure safety and meet University regulatory requirements we also employ QGroundControl— a popular cross-platform flight control and mission planning software—to monitor and abort autonomous navigation. We detail our Sensor System below. Signal Processing Module: In contrast to previous work [4], we propose using a software defined radio (SDR) receiver to implement the signal processing components. The key advantages of our choice are the ability to: i) reduce the weight of the receiver; ii) rapidly scan a large frequency spectrum to track multiple animals beaconing on different VHF frequency channels; and iii) because the signal processing chain is defined in software, we have the ability to reconfigure the system on the fly. In this work, we use the HackRF One SDR—an open source platform developed by Ossmann [13] capable of directly converting radio frequency (RF) signals to digital signals using an analog-to-digital converter (ADC)—together with an Intel Edison board as our embedded compute module. We implemented a Discrete Fourier Transform (DFT) filter to isolate, from multiple signals, each unique VHF frequency channel associated with an animal radio collar and measure the signal strength of the received signal. Antenna: A lightweight folded 2-element Yagi antenna was specially designed for our sensor system. Our design achieves a low profile antenna capable of being within the form factor of low-cost commodity UAVs suitable for easy operation in the field. Similar to a standard 2-element Yagi antenna, the folded design has one reflector and one driven element as shown in Fig. 2b. The antenna operates in the frequency range from 145 to 155 MHz (a typical range for wildlife radio tags), and a center frequency of f = 150 MHz. The length of driven and reflector elements are Dd = 0.3975λ and Dr = 0.402λ, respectively, while d1 = 0.1λ, d2 = 0.03λ and the inductive loading ring diameter is d3 = 0.015λ. Here, the wavelength λ = c/f = 2 (m) with c = 3·108 (m/s). The antenna gain model calculated for the the design is shown in Fig. 5b. A. Planning implementation for a real-time system Implementing planning algorithms on any real-time systems is always challenging because of its high computational demand. Thus, in this section, we present a few approaches to minimize the planning computational time while not scarifying the overall localization performance. i) Notably, for RSSI data, the uncertainty in the estimation of a target’s location is reduced when the maximum gain of the directional antenna mounted on the UAV points toward the target position. Hence, to increase the localization accuracy, the UAV heading angle θku must be controlled during the path planning process, although the multi-rotor UAV can be maneuvered without changing its heading. We adopted a set of discrete UAV rotation angles for the control actions Ak in [16] as this helps reduce the computational complexity of the POMDP planning process by limiting the number of possible actions to evaluate. ii) The solver performs planning in every Np observation cycles with Np > 1 instead of every observation. This approach helps to ensure that the solver prioritizes its limited computational resource on tracking targets instead of only performing planning steps. iii) A coarse planning interval tp in the planning procedure is implemented to minimize the computational time by reducing number of look-ahead steps while still having the same look-ahead horizon. For example, we want to estimate the target’s state in a 10 second horizon, we can use the normal interval tp = 1(s) and estimate the target’s state 10 times or use the coarser interval tp = 5(s) and perform the estimation twice; the latter approach is computationally less expensive. iv) Instead of selecting the best action from the possible action space Ak , the domain knowledge of the receiver antenna gain is used to select a subset of actions that gives the highest received gain using Alg. 1. Fig. 2. a) The full communication channels between UAV and the ground control system with its main softwares and protocols. b) Folded 2-element Yagi antenna used in our sensor system for observations Algorithm 1 Calculate the control action subset Input: Number of preferred actions NA,s , Ak , the antenna gain Gr , the target’s position xk+H Output: Ask 1: for l = 1 : NA,k do 2: Get ulk+H ∈ Ak (l) 3: Calculate Glr = Gr (xk+H , ulk+H ) 4: end for 5: Ask = Ak (Glr ≥ Top NA,s of Gr ) Following the above implementation approach, UAV motion includes two modes: i) changing its heading angle while hovering; and ii) moving forward to its direct location. In one planning procedure with Np cycles, the UAV needs b\|4θ\|/θmax c cycles to rotate, and spends the remaining cycles Np − b\|4θ\|/θmax c to move forward without changing its heading. Here b·c and \| · \| are the floor and absolute operator respectively, and θmax is the UAV maximum rotation angle in one cycle . The sign of 4θ decides the rotation direction (+ for the clockwise, and − for the counter-clockwise). V. S IMULATION E XPERIMENTS Implementing on a real system is time-consuming and difficult. Hence, we want to validate our systems firstly through several simulation experiments to: i) verify our tracking and planning algorithms; ii) investigate how our planning parameters such as different α values of the Rényi divergence or number of discrete actions NA,s = \|Ask \| created in Alg. 1 contribute to the overall algorithm performance; and iii) compare our proposed Rényi divergence based planning technique with other well-known methods, and the impact of the look-ahead horizon parameters on computational time and localization accuracy. All of the simulation experiments were processed on a PC with an Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz, 32GB RAM and MATLAB-2016b. A. Tracking and Planning Simulation This simulation was implemented to validate our approach. In this experiment, the UAV attempted to search and localize 10 moving targets randomly located in an area of 500 m × 500 m. The following are list of parameters used in this simulation: the sampling time step is 1 second since the tag emits pulse signals every 1 second. The solver performed a planning procedure every Np = 5 s, and the look-ahead horizon parameters: H = NH tp = 5 s with number of horizon NH = 1 and the planning interval tp = 5 s. The UAV started from its home location at u1 = [0, 0, 20, 0]T , moved under the constant velocity vu = 5 m/s with its maximum heading rotation angle θmax = π/6 rad/s. Number of particles for each target were capped at Ns = 10, 000, with the future sample measurement Ms = 50, Rényi divergence parameter α = 0.5, number of actions NA,s = 5. In addition, a target is considered localized if its location uncertainty determined by its particles covariance is small enough—NT h = 10, 000 was chosen as the limit. The LogPath measurement model with Prd0 = −22.3 dB, n = 3.1, σP = 4.22 was used to verify our proposed algorithm. To demonstrate that our algorithm was able to localize mobile targets, a wombat—an animal that usually wanders around its area was considered. Hence, a random walk model was used to to describe its behavior with a single target’s transitional density: p(xk \|xk−1 ) = N (xk ; Fxk−1 , Q) (8) where F = I3 with In is n × n identity matrix , Q = 2 σQ diag([1, 1, 0]T ), σQ = 2 m/s. Fig. 3 shows localization results for 10 mobile targets where the estimation details are annotated next to the target’s position with two indicators: RMS and flight time—see Sec. V-B for definitions. In summary, for this scenario, it took the UAV 587 seconds to localize all ten moving targets with the maximum error distance less than 15 m, except for the outlier, target #2 (RMS = 26.3 m). At flight time 587 s, after finishing localizing the last target (target #7), the UAV was sent a command to fly back to its original station. In this case, the total UAV travel distance was 1.93 km. The results demonstrate that our algorithm can search and accurately localize multiple numbers of targets in real time (about 10 minutes) at the travel distance 1.93 km that most commercial off the shelf drones can do. B. Monte Carlo simulations For this experiment, all of Monte Carlo setup parameters were kept the same as in Sec. V-A, except for the investigated ones. In addition, to ensure the results were not random, all of the conducted experiments were performed over 100 Monte Carlo runs. The tracking algorithm was evaluated based on the following criterion: • Estimation Error is the absolute distance between ground truth and estimated target location Drms = P Ntg j j j j 2 j=1 drms /Ntg with drms = [(xtruth − xest ) + j j (ytruth − yest )2 ]1/2 . • Flight time (s) for UAV to localize all of the targets which includes hovering time when the UAV waits for commands from the solver to take an action. • UAV travel distance distance to find all of target positions y (m) 400 5 RMS: 8.1 Flight time: 244 300 RMS: 11.2 10 Flight time: 101 RMS: 4.8 1 200 Flight time: 88 6 2 RMS: 26.3 RMS: 7.3 Flight time: 155 Flight time: 129 100 RMS: 10.8 Flight time: 419 RMS: 13.6 Flight time: 517 3 7 RMS: 14.2 Flight time: 587 0 0 100 200 300 400 500 x (m) Fig. 3. Simulation results with 10 moving targets localized using a single UAV by Particle Filter and POMDP. TABLE I L OCALIZATION PERFORMANCE FOR DIFFERENT ALPHA VALUES RMS (m) Flight time (s) UAV travel distance (km) α = 0.1 12.35 724 2.38 α = 0.5 12.77 741 2.41 α = 0.9999 12.96 727 2.34 Computational cost: calculated for 2 scenarios: i) execution time for the solver to perform the tracking algorithm only (called non-planning time), and ii) the execution time for the solver to select the best action (planning step) as well as conduct tracking (called planning time). Firstly, our search and localization algorithms were evaluated using different α values for Réyni reward function in (6). Table I presents the Monte Carlo results for α = {0.1, 0.5, 0.9999}. In general, the α values do not significant impact the overall performance. However, applying α = 0.1 provides the best localization results in terms of estimation error and search duration. Applying α = 0.5 proposed in [16, 18] results in the worst performance, it increases flight time and travel distance necessary to complete the localization task. Using α = 0.9999 (considered as using KL divergence) which is popular in information gain helps to save UAV travel distance while sacrificing location accuracy. One explanation for this scenario is that our noisy measurement causes the (m) predicted posterior p(xk+H \|z1:k , zk+1:k+H (a)) in (6) to be less informative due to high uncertainty. Therefore, the reward function should pace more emphasis more on the current posterior instead by using a small α value or setting α → 1 to completely ignore the future posterior. It also explains why selecting α = 0.5 (equally weighting the current and the future posterior) result in the worst localization performance. Equally important, we conducted experiments to understand how the action space set NA,s created by Alg. 1 affects to our tracking performance in term of planning time and localization error. Table II shows Monte Carlo results for NA,s = {2, 3, 4, 5, 6, 7}. It presents an interesting result that applying NA,s = 4 provides the best localization performance • Number of actions NA,s RMS (m) Flight time (s) UAV travel distance (km) Planning time (s) 15 14 13 12 11 10 9 8 7 6 5 1 2 3 4 5 6 7 14.18 840 12.64 781 12.17 693 12.27 723 12.83 756 12.63 799 2.62 2.53 2.39 2.50 2.52 2.70 1.16 1.19 1.23 1.27 1.36 1.47 RMS (m) Flight time (s) Travel Distance (km) 2 3 4 5 6 7 8 Number of Targets 9 944 844 744 644 544 444 344 244 144 10 2.5 2 1.5 1 Travel Distance (km) 4 RMS: 9.5 Flight time: 393 Flight time (s) 8 9 RMS (m) 500 TABLE II L OCALIZATION PERFORMANCE FOR DIFFERENT NUMBER OF ACTIONS Real position of Target Estimated position of Target UAV Trajectory RMS: 9.5 Flight time: 288 0.5 Fig. 4. Localization performance for different number of targets Ntg increase from 1 to 10. in terms of estimation error, flight duration and travel distance; a desirable result for realizing real time planning with limited computational resources. Another aspect that we want to examine is how our proposed algorithm performs when a maximum number of targets Ntg change from 1 to 10. As depicted in the Fig. 4, our algorithm’s estimation error was stable and invariant to the number of targets. Moreover, it is reasonable that the flight time and the travel distance increased linearly with target numbers because it took more time and power to track more targets. Lastly, we examined how Rényi divergence performs in different look-ahead horizons H = NH tp compared to: i) the Shannon entropy [4]; ii) a naive approach that moves UAV to the closest estimated target location; and iii) a uniform search with predefined path used in [18]. Table III shows the Monte Carlo comparison results among various planning algorithms. All of parameters were reused from the Sec. V-A, except for α = 0.1 and NA,s = 4 were updated based on the previous experimental results. The result has demonstrated that Rényi divergence reward function is superior to other planning strategies in term of localization accuracy, including the Shannon entropy with the same horizon settings. For Rényi reward function itself, the large look ahead horizon number NH > 1 helps to improve the localization accuracy; however, it requires higher computational power (planning) and causes the UAV to travel further. Using NH = 1; tp = 5 provides the best trade-off between computational time and accuracy. Summary: According to the above simulation results, we select α = 0.1, NA,s = 4, and NH = 1, tp = 5 s as the planning parameters for the field experiment since these parameters provides the lowest computational cost, best performance in term of location estimation error, travel distance and flight time. Fig. 5. a) Waterfall plot for rotor noise experiment when four motors at full rotation speed; b) Normalized antenna gain in E-plane G(φ). The red line is gain modeled pattern and black line is the normalized measured gain pattern from 30 measurements collected by rotating the UAV heading at 15◦ intervals; c) Plot of measured RSSI data points and its estimated models over a distance from 10 m to 320 m at 10 m intervals. TABLE III L OCALIZATION PERFORMANCE FOR DIFFERENT PLANNING ALGORITHM NH tp (s) RMS (m) Flight time (s) UAV travel distance (km) Plan Time (s) Non Plan Time (s) Uniform N/A N/A 18.8 921 Closest Target N/A N/A 13.4 799 Shannon [4] 1 5 12.6 774 1 5 12.5 699 Rényi 3 1 12.4 889 5 1 12.0 811 10 1 11.6 822 3.72 2.29 2.54 2.27 2.99 2.82 2.42 1.58 1.11 1.38 1.28 1.53 1.65 2.71 1.58 1.03 0.99 0.97 0.96 0.97 0.96 VI. F IELD E XPERIMENTS We describe here our extensive experiments regime to validate our approach and evaluate the performance of our aerial robot system in the field. Our aim is to: i) investigate the possibility for signal interference from spinning motors of a UAV on RSSI measurements; ii) estimate the model parameters in the sensor model and validate the proposed model; and iii) conduct field trials to demonstrate and evaluate our system capabilities. A. Rotors noise We investigated the rotor noise to confirm that our system is not affected by the electromagnetic interference from the UAV’s motors. It also helps to clear the concern raised in [4] that the rotor noise may affect the RSSI measurements. Four motors of the 3DR IRIS+ quad-copter shown in Fig.1 were used in this experiment. The RSSI data of a radio collar were measured across 149 MHz to 151 MHz frequency spectrum when four motors were operating at 20%,50%,100% of its maximum speed of 10, 212 rounds per minute. Fig 5a shows the frequency spectrum of the received signal. We can see that there was no difference in the frequency characteristics when the rotors were in ON and OFF states. This result confirms that the rotors do not spin fast enough to generate high-frequency interference to impact to our RSSI measurements. B. Sensor model validation and parameter estimation Antenna Gain: The antenna gain pattern was measured to verify its directivity compared to the antenna gain model Gr (x, u) = Gr (φ) calculated—following [12, p.1252]—based on the physical design as discussed in Sec. IV. Fig. 5b shows the measured and modeled radiation patterns Gr (φ) , here φ is the angle between the UAV heading and the direction from its position to target’s position on 0xy plane. The result shows that the front-to-back ratio is smaller (2 dB) than expected and this is an artifact of folding the reflector on our design. Signal propagation model parameter validation: We collected RSSI data points over a range from 10 m to 320 m between the UAV and a VHF radio tag. The tag and UAV were kept at a height of 5 m above ground during this experiment. The tag was stationary at all times, while the UAV was directed to move away in a straight line from the tag at 10 m intervals whilst hovering at each location to allow the collection of approximately 30 measurements. The UAV heading was maintained to ensure consistent antenna gain during the experiment. Since we operated in an open terrain over a grassland, we selected the path loss exponent n = 2 suitable for modeling free space path loss. Fig. 5c shows the measured RSSI and the propagation models obtained using a nonlinear regression algorithm to estimate model parameters; we have the following results for reference power Prd0 in (3), (4) at the reference distance d0 = 1 m, and measurement noise variance σP in (5): d0 • LogPath model : Pr = −45.69 ; σP = 4.21 (dB). d0 • MultiPath model : Pr = −45.28 ; σP = 2.31 (dB). The results show that both models, as expected, derived a similar reference power Prd0 whilst providing a reasonable fit to measurement data and this affirms the validity of our propagation model. Although LogPath model is reasonable, MultiPath model is more accurate and yields a smaller measurement noise variance. The results confirms the impact of ground reflections, especially close to the signal source. C. Field Trials We present two sets of field experiments to validate the two measurement models and conducted a total of 16 autonomous flights to demonstrate our system capabilities. Our experiments were designed around Anonymous University regulations governing the conduct of experimental UAV research. Given the need to operate in autonomous mode, our flight zone as well as the scope of the experiment was restricted to University owned property designated for UAV flight tests. Therefore: i) the UAV task was set to search and localize two mobile targets in a search area 75 m × 300 m; and ii) instead of wildlife, we relied on two people, each wearing a VHF radio tag on their forearm, and a mobile phone based GPS data logger on TABLE IV C OMPARISON OF L OCALIZATION PERFORMANCE Model LogPath MultiPath Cliff [4] Target Type Mobile Mobile Stationary Trials RMS (m) 8 8 6 30.1 ± 12.8 22.7 ± 13.9 23.8 ± 14.0 Total Flight Time (s) 255 ± 104 138 ± 53 8381 Travel Distance (m) 549 ± 167 286 ± 121 N/A their hands to obtain ground truth; with two extra personnel stationed to maintain constant sight of the UAV and people in the field and abort the autonomous mode to transfer control to manual operations by a pilot. The volunteers with the radio tags were asked to walk randomly and no other instructions were given. Fig. 6 shows the tracking and localization results along with UAV trajectories based on the two different measurement models. Table IV presents the summary comparison results of location estimates between the two measurement models. Smaller RMS (root mean square) estimation error values suggest a higher accuracy, while shorter flight times and travel distance to localize all targets are highly desirable for a practicable system given the power constrained nature of commodity UAVs. The result confirm that the MultiPath model is superior to the standard LogPath model since it has been able to account for ground reflections and the UAV is not required to approach the target as closely when using the LogPath model to reduce its measurement uncertainty. D. Discussion We also summarise results from Cliff et al. [4] for comparison in Table IV while Table V presents a complete comparison between our proposed system and Cliff et al. [4] system. Notably, our search area is smaller compared to Cliff et al. [4] (75 m × 300 m v.s 1000 m × 1000 m) due to our test flight zone restrictions, however, we have set up our initial distance from the UAV home position to its farthest target’s position (target #2 in this case) to be equivalent to the distance of the stationary target in [4]; approximately 300 m. The results in Table IV demonstrate that our proposed method can localize two mobile targets in a shorter flight time (the flight time of MultiPath model is one-sixth of that in Cliff et al. [4]) with better accuracy. Moreover, we search and locate two mobile targets; in contrast, Cliff et al. [4] method was implemented to locate a single and stationary target. In general, as shown in Table V, our system is more compact, lighter, and has a payload that is one-third of that in [4] and consequently capable of longer flight times on any given UAV. Although, our reliance on an SDR without a preamplifier has resulted in a shorter detection range, our total system mass being under 2 kg is significant since it enables ecologists to operate our system without a remote pilot licenses (RePL). Moreover, the ability to instantly collect range-only measurements also helps reduce flight time significantly compared to the bearing-only method, requiring full rotations of 1 Information regarding the total flight is not reported in Cliff et al. [4], however, as shown in Fig. 9 in Cliff et al. [4], one observation took 76.21s and one trial needed 11 observations, hence total flight time is 11 × 76.21 = 838.31s Fig. 6. Field experiment results to search, track and localize two mobile tags for the two different measurement models. a) Standard LogPath. b) MultiPath. TABLE V C OMPARISON BETWEEN OUR SYSTEM AND C LIFF ET AL . [4] SYSTEM . Ours 260 1,280 Quadcopters (smaller drone) Software defined radio (digital-based,rapidly scan multiple frequencies to support multiple frequencies) Cliff et al. [4] 750 2,200 Octocopters (relatively larger drone) Detection range (m) 320 500 Measurement model Range-only (exploiting the simplicity of a range-only measurement system) Payload (g) Total mass (g) Drone type Receiver Architecture Filtering method Planning Nature of targets Particle filter (O(N ) operations per iteration) Rényi divergence multiple mobile target tracking Analog filtering circuit (analog-based, difficult to re-configure for a new frequency) Bearing-only (antenna array, and UAV rotation at grid points with a phase difference measurement system) Grid-based filter (O(N 2 ) operations per iteration) Shannon entropy A single stationary target localization a UAV at each observation point, as shown in the Table IV. Furthermore, as discussed in [1], the computational cost for grid-based methods used in [4] increases dramatically, with the number of cells whilst the grid must be dense enough to achieve accurate estimations; e.g. , a grid-based filter with N cells conducts O(N 2 ) operations per iteration, while a similar particle filter with N particles only requires O(N ) operations. Hence, the grid-based filter method only works in case of stationary targets as in [4] where the most expensive computational step, the prediction step, is skipped. Moreover, as shown in Table. III, our planning algorithm based on Rényi divergence is superior to the Shannon entropy approach in [4] in both important metrics: accuracy and flight time. VII. C ONCLUSION We have developed and demonstrated an autonomous aerial vehicle system for range only tracking and localization of VHF radio-tagged animals under RSSI based measurement uncertainty and mobility of targets during their discovery in the field. The joint particle filter and POMDP with Rényi divergence based reward function provided an accurate method to explore, track and locate multiple animals while considering the resource constraints of the underlying UAV platform. In addition, we have realized a UAV system under 2 kg to ensure both the practicability and the accessibility of the technology to conservation biologists. While we have demonstrated a successful system, we have only formulated our approach as a two-dimensional tracking problem that is ideally suitable for tracking endangered species in largely flat terrains and grasslands. Consequently, the current approach is not suitable for tackle tracking wildlife in hills or mountainous areas and it would require: i) a UAV capability to maintain a fixed relative altitude above the ground; or ii) formulating a 3D tracking problem to extend our method to all topographical conditions. We leave the latter for future work. VIII. ACKNOWLEDGMENTS This work was jointly supported by the Western Australia Parks and Wildlife, the Australian Research Council (LP160101177), the Defense Science and Technology Group, and the University of Adelaide’s Unmanned Research Aircraft Facility. R EFERENCES [1] M Sanjeev Arulampalam, Simon Maskell, Neil Gordon, and Tim Clapp. A tutorial on particle filters for online nonlinear/non-Gaussian Bayesian tracking. IEEE Transactions on signal processing, 50(2):174–188, 2002. [2] Michael Anthony Beard, Ba-Tuong Vo, Ba Ngu Vo, and Sanjeev Arulampalam. Void probabilities and cauchy-schwarz divergence for generalized labeled multibernoulli models. IEEE Transactions on Signal Processing, 65, 2017. [3] CASA. AC 101-01 v2.0 - Remotely piloted aircraft systems, 2018. [Online; accessed 14-February-2018]. [4] Oliver M Cliff, Robert Fitch, Salah Sukkarieh, Debbie Saunders, and Robert Heinsohn. Online Localization of Radio-Tagged Wildlife with an Autonomous Aerial Robot System. In Robotics: Science and Systems, 2015. [5] Neil J Gordon, David J Salmond, and Adrian FM Smith. Novel approach to nonlinear/non-Gaussian Bayesian state estimation. IEE Proceedings F - Radar and Signal Processing, 140(2):107–113, 1993. [6] Alfred O Hero, Christopher M Kreucher, and Doron Blatt. Information theoretic approaches to sensor management. Springer US, 2008. [7] David Hsu, Wee Sun Lee, and Nan Rong. A point-based POMDP planner for target tracking. In Proc. of IEEE ICRA, pages 2644–2650, 2008. [8] Austin M Jensen, David K Geller, and YangQuan Chen. Monte Carlo simulation analysis of tagged fish radio tracking performance by swarming unmanned aerial vehicles in fractional order potential fields. Journal of Intelligent & Robotic Systems, 74(1-2):287–307, 2014. [9] Leslie Pack Kaelbling, Michael L. Littman, and Anthony R. Cassandra. Planning and acting in partially observable stochastic domains. Artificial Intelligence, 101(1):99 –134, 1998. [10] Roland Kays, Sameer Tilak, Margaret Crofoot, Tony Fountain, Daniel Obando, Alejandro Ortega, Franz Kuemmeth, Jamie Mandel, George Swenson, Thomas Lambert, et al. Tracking animal location and activity with an automated radio telemetry system in a tropical rainforest. The Computer Journal, pages 1931–1948, 2011. [11] Fabian Körner, Raphael Speck, Ali Haydar Göktogan, and Salah Sukkarieh. Autonomous airborne wildlife tracking using radio signal strength. In Proc. of IEEE/RSJ IROS, pages 107–112, 2010. [12] Sophocles J Orfanidis. Electromagnetic waves and antennas. Rutgers University New Brunswick, NJ, 2002. [13] Michael Ossmann. Software Defined Radio with HackRF, 2015. [14] André Posch and Salah Sukkarieh. UAV based search for a radio tagged animal using particle filters. In Australasian Conference on Robotics and Automation (ACRA), Sydney, Australia, Dec, pages 2–4, 2009. [15] Branko Ristic. Particle Filters for Random Set Models. Springer-Verlag New York, 2013. [16] Branko Ristic and Ba-Ngu Vo. Sensor control for multiobject state-space estimation using random finite sets . Automatica, 46(11):1812 – 1818, 2010. [17] Branko Ristic, Sanjeev Arulampalam, and Neil (chercheur) Gordon. Beyond the Kalman filter : particle filters for tracking applications. Artech House, 2004. [18] Branko Ristic, Mark Morelande, and Ajith Gunatilaka. Information driven search for point sources of gamma radiation. Signal Processing, 90(4):1225–1239, 2010. [19] David Silver and Joel Veness. Monte-Carlo planning in large POMDPs. In Advances in neural information processing systems, pages 2164–2172, 2010. [20] Bindi Thomas, John D Holland, and Edward O Minot. Wildlife tracking technology options and cost considerations. Wildlife Research, 38(8):653–663, 2012. [21] Pratap Tokekar, Deepak Bhadauria, Andrew Studenski, and Volkan Isler. A robotic system for monitoring carp in Minnesota lakes. Journal of Field Robotics, 27(6): 779–789, 2010. [22] Joshua Vander Hook, Pratap Tokekar, and Volkan Isler. Cautious Greedy Strategy for Bearing-only Active Localization: Analysis and Field Experiments. Journal of Field Robotics, 31(2):296–318, 2014. [23] Neeti Wagle and Eric Frew. Spatio-temporal characterization of airborne radio frequency environments. In IEEE GLOBECOM Workshops, pages 1269–1273, 2011. [24] Jakes WC Jr. Microwave mobile communications. Wiley, New York, 1974.	3cs.SY
Moonshine: Distilling with Cheap Convolutions arXiv:1711.02613v1 [stat.ML] 7 Nov 2017 Elliot J. Crowley, Gavin Gray, Amos Storkey School of Informatics The University of Edinburgh Abstract Model distillation compresses a trained machine learning model, such as a neural network, into a smaller alternative such that it could be easily deployed in a resource limited setting. Unfortunately, this requires engineering two architectures: a student architecture smaller than the first teacher architecture but trained to emulate it. In this paper, we present a distillation strategy that produces a student architecture that is a simple transformation of the teacher architecture. Recent model distillation methods allow us to preserve most of the performance of the trained model after replacing convolutional blocks with a cheap alternative. In addition, distillation by attention transfer provides student network performance that is better than training that student architecture directly on data. 1 Introduction Despite advances in deep learning for a variety of tasks [LeCun et al., 2015], deployment of deep learning into embedded devices e.g. mobile phones, digital cameras, vehicle navigation systems, has been relatively slow due to resource constraints under which these devices operate. Big, memory-intensive neural networks don’t fit on these devices, but do these networks have to be big and expensive? The dominant run-time memory cost of neural networks is the number of parameters that need to be stored. Can we have networks with substantially fewer parameters, without the commensurate loss of performance? It is possible to take a large pre-trained teacher network, and use its outputs to aid in the training of a smaller student network [Ba and Caruana, 2014] through some distillation process. By doing this the student network is more powerful than if it was trained solely on the training data, and is closer in performance to the larger teacher. The lower-parameter student network typically has an architecture that is more shallow, or thinner – by which we mean its filters have less channels [Romero et al., 2014] – than the teacher. While it is not possible to arbitrarily approximate any network with another [Urban et al., 2017], the limit in neural network performance is at least in part due to the training algorithm, rather that its representational power. In this paper, we take an alternative approach in designing our student networks. Instead of making networks thinner, or more shallow, we take the standard convolutional block such networks possess and replace it with a cheaper convolution block, keeping the original architecture. For example, in a ResNet [He et al., 2016a] this standard block is a pair of sequential 3×3 convolutions. We show that for a comparable number of parameters, student networks that retain the architecture of their teacher but with cheaper convolutional blocks outperform student networks with the original blocks and smaller architectures. As a model compression strategy, this is very effective. At the same time this transformation is easy to implement in any deep learning framework; replacing convolutional blocks is a simple substitution into any existing architecture. Furthermore, the optimisation scheme used on the teacher network can be repeated on the student, making another round of hyperparameter optimisation unnecessary. The cheap convolutional blocks we suggest are described in Section 3 as well as an overview of the methods we employ for distillation. In Section 4 we train a number of student networks for the task of image classification on the CIFAR-10 and CIFAR-100 [Krizhevsky, 2009] datasets and demonstrate that those with cheap convolutions perform better than traditional student networks for a given parameter cost. This level of parameter reduction is competitive with much more complicated methods in the literature [Howard et al., 2017]; and some of these methods may be complementary [Han et al., 2015]. Though it is possible to train the resulting architectures directly, this is demonstrably less effective than distilling them from the larger teacher model. 2 Related Work The parameters in deep networks have a great deal of redundancy; it has been shown that many of them can be predicted from a subset of parameters [Denil et al., 2013]. However the challenge remains to find good ways to exploit this redundancy without losing model accuracy. This observation, along with a desire for efficiency improvements has driven the development of smaller, and less computationally-intensive convolutions. One of the most prominent examples is the depthwise separable convolution [Sifre, 2014] which applies a separate convolutional kernel to each channel, followed by a pointwise convolution [Lin et al., 2014] over all channels; depthwise separable convolutions have been used in several architectures [Ioffe and Szegedy, 2015, Chollet, 2016, Xie et al., 2017], and were explicitly adapted to mobile devices in Howard et al. [2017]. However, separating the spatial and channel-wise elements is not the only way to simplify a convolution. In Jin et al. [2015] the authors propose breaking up the general 3D convolution into a set of 3 pointwise convolutions along different axes. The authors of Wang et al. [2016] start with separable convolutions and add topological subdivisioning, a way to treat sections of tensors separately, and a bottleneck of the spatial dimensions. Both of these methods demonstrate models that are several times smaller than the original model while maintaining accuracy. In a separable convolution, the most expensive part is the pointwise convolution, so it has been proposed that this operation could also be grouped over sets of channels. However, to maintain some connections between channels, it is helpful to add an operation mixing the channels together [Zhang et al., 2017]. More simply, a squared reduction can be achieved by applying a bottleneck on the channels before the spatial convolution [Xie et al., 2017, Iandola et al., 2016]. In this paper we examine the potency of a separable bottleneck structure. The work discussed thus far in this section involves learning a compressed network from scratch. There are clear alternatives to this such as retraining after reducing the number of parameters [Han et al., 2015, Li et al., 2017]. We are interested in learning our smaller network as a student through distillation [Ba and Caruana, 2014, Buciluǎ et al., 2006] in conjunction with a large pre-trained teacher network. How small can our student be? The complex function of a large, deep teacher network can, theoretically, be approximated by a network with a single hidden layer with enough units [Cybenko, 1989]. The difficulty in practice is learning that function. Knowledge distillation [Ba and Caruana, 2014, Hinton et al., 2016] proposes to use the information in the logits of a learnt network to train the smaller student network. In early experiments, this 2 was shown to be effective; networks much smaller than the original could be trained with small increases in error. However, modern deep architectures prove harder to compress. For example, a deep convolutional network cannot be trivially replaced by a feedforward architecture [Urban et al., 2017]. Two methods have been proposed to deal with this. First, in Romero et al. [2014] the authors use a linear map between activations at intermediate points to produce an extra loss function. Second, in attention transfer [Zagoruyko and Komodakis, 2017], the authors choose instead to match the activations after taking the mean over the channels. In the context of this paper, we found attention transfer to be effective in our experiments, as described in Section 4. 3 Model Compression with Cheap Convolutions Given a large, deep network that performs well on a given task, we are interested in compressing that network so that it uses fewer parameters. A flexible and widely applicable way to reduce the number of parameters in a model is to replace all its convolutional layers with a cheaper alternative. Doing this replacement invariably impairs performance when this reduced network is trained directly on the data. Fortunately, we are able to demonstrate that modern distillation methods enable the cheaper model to have performance closer to the original large fully-convolutional network. 3.1 Distillation For this paper, we utilise and compare two different distillation methods for learning a smaller student network from a large, pre-trained teacher network: knowledge distillation [Hinton et al., 2016, Ba and Caruana, 2014] and attention transfer [Zagoruyko and Komodakis, 2017]. We briefly explain these methods below: Knowledge Distillation P Let us denote the cross entropy of two probability vectors p and q as LCE (p, q) = − k pk log qk . Assume we have a dataset of elements, with one such element denoted x, where each element has a corresponding one-hot class label: denote the one-hot vector corresponding to x by y. Given x, we have a trained teacher network t = teacher(x) that outputs the corresponding logits, denoted by t; likewise we have a student network that outputs logits s = student(x). To perform knowledge distillation we train the student network to minimise the following loss function (averaged across all data items): t s 2 LKD = (1 − α)LCE (y, σ(s)) + 2T αLCE σ ,σ , (1) T T where σ(.) is the softmax function, T is a temperature parameter and α is a parameter controlling the ratio of the two terms. The first term is a standard cross entropy loss penalising the student network for incorrect classifications. The second term is minimised if the student network produces outputs similar to that of the teacher network. The idea being that the outputs of the teacher network contain additional, beneficial information beyond just a class prediction. Attention Transfer Consider some choice of layers i = 1, 2, ..., NL in a teacher network, and the corresponding layers in the student network. At each chosen layer i of the teacher network, collect the spatial map of the activations for channel j into the vector atij . Let Ati collect atij for all j. Likewise for the student network we correspondingly collect into asij and Asi . 3 Now given some choice of mapping f (Ai ) that maps each collection of the form Ai into a vector, attention transfer involves learning the student network by minimising: LAT = LCE (y, σ(s)) + β NL X i=1 f (Ati ) f (Asi ) − t \|\|f (Ai )\|\|2 \|\|f (Asi )\|\|2 , (2) 2 where β is a hyperparameter. Zagoruyko and Komodakis [2017] recommended using PNA f (Ai ) = (1/NAi ) j=1i a2ij , where NAi is the number of channels at layer i. In other words, the loss targeted the difference in the spatial map of average squared activations, where each spatial map is normalised by the overall activation norm. Let us examine the loss (2) further. The first term is again a standard cross entropy loss. The second term, however, ensures the spatial distribution of the student and teacher activations are similar at selected layers in the network, the explanation being that both networks are then paying attention to the same things at those layers. 3.2 Cheap Convolutions As large fully-connected layers are no longer commonplace, convolutions make up almost all of the parameters in modern networks.1 It is therefore desirable to make them smaller. Here, we present several convolutional blocks that may be introduced in place of a standard block in a network to substantially reduce its parameter cost. First, let us consider a standard two dimensional convolutional layer that contains Nout filters, each of size Nin × k × k (assuming square convolutions). Nout is the number of channels of the layer output, Nin is the number of channels of the input, and k × k is the kernel size of each convolution. In modern neural networks it is almost always the case that Nin 6 Nout . Let N = max(Nin , Nout ). Then the parameter cost of this layer is Nin Nout k 2 , and is bounded by N 2 k 2 . In a typical residual network, a block contains two such convolutions. We will refer to this as a Standard block S, and it is outlined in Table 1. One alternative to full convolutions, with parameters that scale approximately as N 2 , is to break each convolution into g groups, as shown in Figure 1. By restricting the convolutions to only mix channels within each group and not between groups, we obtain a substantial reduction in the number of parameters for a grouped computation: for example, for Nin = Nout = N the cost changes from N 2 k 2 for a standard layer to g groups of (N/g)2 k 2 parameter convolutions, hence reducing the parameter cost by a factor of g. We can then provide some cross-group mixing by following each grouped convolution with a pointwise convolution, with a N 2 parameter cost (when Nin 6= Nout the change in channel size occurs across this pointwise convolution). We refer to this substitution operator as G(g) (grouped convolution with g groups), and illustrate it in Figure 2. In the original ResNet paper [He et al., 2016a] the authors introduce a bottleneck block which we have parameterised, and denoted as B(b) in Table 1: the input first has its channels decreased by a factor of b via a pointwise convolution, before a full convolution is carried out. Finally, another pointwise convolution brings the representation back up to the desired Nout . We can reduce the parameter cost of this block even further by replacing the full convolution with a grouped one; the Bottleneck Grouped + Pointwise block is referred to as BG(b, g) and is illustrated in Figure 2. These substitute blocks are compared in Table 1 and their computational costs (for simplicity we take the case where Nin = Nout = N ) are given. In practice, by varying the bottleneck size and the number of groups, network parameter numbers may vary over two orders of magnitude; enumerated examples are given in Tables 3 and 4. 1 The parameters introduced by batch normalisation are negligible compared to those in the convolutions. However, they are included for completeness in Table 1. 4 N/g Figure 1: Grouped convolutions operate by passing independent filters over the tensor after it is separated into g groups over the channel dimension. Consider a grouped convolution where the input and and output both have N channels: each of the g filters only needs to operate over N/g channels. This reduces the parameter cost of the convolution by a factor of g. Standard G(g) BG(b,g) Figure 2: The Grouped + Pointwise (G(g)) block substitutes all full k × k convolutions in the Standard block with a grouped convolution followed by a pointwise (1 × 1) convolution. To reduce parameters further, a pointwise Bottleneck can be used before the Grouped + Pointwise convolution (BG(b, g)). 5 Using grouped convolutions and bottlenecks are common methods for parameter reduction when designing a network architecture. Both are easy to implement in any deep learning framework. Sparsity inducing methods [Han et al., 2015], or approximate layers [Yang et al., 2015], may also provide advantages, but these are complementary to the approaches here. More structured reductions such as grouped convolutions and bottlenecks can be advantageous over sparsity methods in that the sparsity structure does not need to be stored. The claim of this paper is that structured parameter reductions are sufficient to achieve model compression results in line with state of the art when using effective model distillation. Block S G(g) B(b) BG(b, g) Structure BN+ReLU Conv BN+ReLU Conv BN+ReLU GConv (g) BN+ReLU Conv1x1 BN+ReLU GConv (g) BN+ReLU Conv1x1 BN+ReLU Conv1x1(N → Nb ) BN+ReLU Conv BN+ReLU Conv1x1( Nb → N ) BN+ReLU Conv1x1(N → Nb ) BN+ReLU GConv(g) BN+ReLU Conv1x1( Nb → N ) 2N 2 k2 2N 2 ( kg + 1) N 2 ( kb2 + 2b ) k 2 N 2 ( gb 2 + b) 8N 16N N (4 + 8b ) N (4 + 8b ) Conv Params BN Params 2 2 2 Table 1: Convolutional Blocks used in this paper: a standard block S, a grouped + pointwise block G, a bottleneck block B, and a bottleneck grouped + pointwise block BG. All blocks use pre-activations [He et al., 2016b]. Conv refers to a k×k convolution. GConv is a grouped k × k convolution and Conv1x1 is a pointwise convolution. BN+ReLU refers to a batchnorm layer followed by a ReLU activation. We assume that the input and output to each block has N channels and that channel size does not change over a particular convolution unless written out explicitly as (x → y). Where applicable, g is the number of groups in a grouped convolution and b is the bottleneck contraction. We give the parameter cost of the convolutions in each block in terms of these parameters. The batch-norm parameter cost (assuming running mean/variance are kept for normalisation) is also given, but is markedly smaller. 4 Experiments In this section we train and evaluate a number of student networks, each distilled from the same large teacher network. We distil with (i) knowledge distillation and (ii) attention transfer. We also train the networks without any form of distillation to observe whether the the distillation process is necessary to obtain good performance. In this way we demonstrate that the high performance comes from the distillation, and cannot be achieved by directly training the student networks using the data. For comparison we also study student networks with smaller architectures (i.e. fewer layers/filters) than the teacher. This enables us to test if the block transformations we propose are key, or it is simply a matter of distilling networks with smaller numbers of parameters. We compare the smaller student architectures with student architectures implementing cheap, substitute convolutional blocks (but with the same architecture as the teacher). The different convolutional blocks are summarised in Table 1 and the student networks are described in detail in Section 4.1. Experiments are conducted for both the CIFAR-10 and CIFAR-100 datasets. Results for 6 group output size structure conv1 conv2 conv3 conv4 pool fc 16 × 32 × 32 16k × 32 × 32 32k × 16 × 16 64k × 8 × 8 64k × 1 × 1 classes 1× Conv3x3(N = 16) n× Block(N = 16k) n× Block(N = 32k) n× Block(N = 64k) 8 × 8 avg-pool 64k × classes fully connected Table 2: Summary of the Wide ResNet structures used in experiments; matching those in Zagoruyko and Komodakis [2017]. The bulk of the parameters are in {conv2, conv3, conv4} which each consist of n blocks with channel width N controlled by k. We explore the effect of substituting these blocks with cheaper alternatives. classes refers to the number of object classes which is, perhaps unsurprisingly, 10 for CIFAR-10 and 100 for CIFAR-100. CIFAR-10 are given in Table 3 and Figure 3. Results for CIFAR-100 can be found in Table 4 and Figure 4. These results are discussed in detail in Section 4.2. 4.1 Network Descriptions For our experiments we utilise the competitive Wide Residual Network (WRN) architecture [Zagoruyko and Komodakis, 2016], which is briefly summarised in Table 2. The bulk of the network lies in the {conv2, conv3, conv4} groups and the network depth d determines the number of convolutional blocks n in these groups as n = (d − 4)/6. The network width, denoted by k, affects the channel size of the filters in these blocks. Note that when we employ attention transfer the student and teacher outputs of groups {conv2, conv3, conv4} are used as {A1 , A2 , A3 } in the second term of Equation (2) with NL = 3. For our teacher network we use WRN-40-2 (a WRN with depth 40 and width 2) with standard (S) blocks. 3 × 3 kernels are used for all non-pointwise convolutions in our student and teacher networks unless stated otherwise. For our student networks we use: • WRN-40-1, 16-2, and 16-1 with S blocks. These are student networks that are thinner and/or more shallow than the teacher and represent typical student networks most works employ. • WRN-40-2 with S blocks where the 3 × 3 kernels have been replaced with 2 × 2 dilated kernels (as described in [Yu and Koltun, 2016]). This allows us to see if it possible to naively reduce parameters by effectively zeroing out elements of each kernel. • WRN-40-2 using a bottleneck block B with 2× and 4× channel contraction (b). • WRN-40-2 using a grouped + pointwise block G for group sizes (g) {2, 4, 8, 16, N/16, N/8, N/4, N/2, N} where N is the number of channels in that block. This allows us to explore the spectrum between full convolutions (g = 1) and fully separable convolutions (g = N ). • WRN-40-2 with a bottlenecked grouped + pointwise block BG. We use b = 2 with groups sizes of {2, 4, 8, 16, M/16, M/8, M/4, M/2, M} where M = N/b is the number of channels after the bottleneck. We use this notation so that (g = M ) represents fully separable convolutions and we can easily denote divisions thereof. BG(4, M ) is also used to observe the effect of extreme compression. Implementation Details Experiments were conducted in PyTorch [Paszke et al.]. For training we used minibatches of size 128. Before each minibatch, the images were padded 7 by 4 × 4 zeros, and then a random 32 × 32 crop was taken. Each image was left-right flipped with a probability of a half. Training was conducted for 200 epochs using SGD with standard momentum fixed at 0.9 with an initial learning rate of 0.1. The learning rate was reduced by a factor of 0.2 at the start of epochs 60, 120, and 160. For knowledge distillation we set α to 0.9 and used a temperature of 4. For attention transfer β was set to 1000. 4.2 Analysis and Observations Figure 3 compares the parameter cost of each student network (on a log scale) against the test error on CIFAR-10 obtained with attention transfer. On this plot, the ideal network would lie in the bottom-left corner (few parameters, low error). What is fascinating is that almost every network with the same architecture as the teacher, but with cheap convolutional blocks (those on the blue, green, and cyan lines) performs better for a given parameter budget than the reduced architecture networks with standard blocks (the red line). BG(2, 2) outperforms 16 − 2 (5.57% vs. 5.66%) despite having considerably fewer parameters (292K vs. 693K). Several of the networks with BG blocks both significantly outperform 16 − 1 and use less parameters. Figure 3: Test Error vs. No. parameters for student networks learnt with attention transfer on CIFAR-10. Note that the x-axis is logarithmically scaled. Points on the red curve correspond to networks with S convolutional blocks and reduced architectures. All other networks have the same WRN-40-2 architecture as the teacher but with cheap convolutional blocks: G (green), B (blue), and BG (cyan). The blocks are described in detail in Table 1. Notice that the student networks with cheap blocks outperform those with smaller architectures and standard convolutions for a given parameter budget. 8 It is encouraging that significant compression is possible with only small losses; several networks perform almost as well as the teacher with considerably less parameters – G(N/8) (blue) has an error of 5.06%, close to that of the teacher, but has just over a fifth of the parameters. BG(2, M/8) has less than a tenth of the parameters of the teacher, for a cost of 1.15% increase in error. With a similar change in error, these compression rates exceed those found in contemporary papers, such as Cohen and Welling [2016]. Even simply switching all convolutions with smaller, dilated equivalents S − 2 × 2 allows one to use half the parameters for a similar performance. An important lesson can be learnt regarding grouped + pointwise convolutions. They are often used in their depthwise-separable [Chollet, 2016] form (when the number of groups is equal to N , the total number of channels in the convolution). However, the networks with half, or quarter that number of groups perform substantially better for a modest increase in parameters. G(N/4) has 374K parameters compared to the 304K of G(N ) but has an error that is 1.26% lower. As we make the number of groups smaller the performance gets close to that of the teacher as the network structure is getting closer and closer to the original convolutions in the teacher (neglecting the pointwise component). The number of groups is an easy parameter to tune to trade some performance for a smaller network. Grouped + pointwise convolutions also work well in conjunction with a bottleneck of size 2, although for large bottlenecks the error increases rather significantly – as can be seen for BG(4, M ). Despite this, it is still of comparable performance to 16 − 1 with half the parameters. We observe similar trends for CIFAR-100 in Figure 4. Figure 4: Test Error vs. No. parameters for student networks learnt with attention transfer on CIFAR-100. Points on the red curve correspond to networks with S convolutional blocks and reduced architectures. All other networks have the same WRN-40-2 architecture as the teacher but with cheap convolutional blocks: G (green), B (blue), and BG (cyan). 9 We also observe that attention transfer from the teacher network is substantially better than knowledge distillation or than training a network structure directly on the data. Consider Table 3, which shows the attention transfer errors of Figure 3 (the AT Error column) alongside those of networks trained with knowledge distillation (KD Error), and no distillation (Error) for CIFAR-10. In all cases, the student network trained with attention transfer is better than the student network trained by itself – a distillation process appears to be necessary. Some performances are particularly impressive – for G(2) blocks the error is only 0.08% higher than the teacher despite the network having under half as many parameters. It is also noticeable that knowledge distillation gives similar, or even worse performance than if the student network was trained by itself. These conclusion are mirrored for CIFAR-100 training (Table 4). Romero et al. [2014] note that the performance issues with knowledge distillation occur in networks with a depth of more than 7 layers. Zagoruyko and Komodakis [2017] also observe that experiments training a CNN on Imagenet would not converge when using knowledge distillation. 5 Conclusion After training a large, deep model it may be prohibitively time consuming to design a model compression strategy in order to deploy it. On many problems, it may also be more difficult to achieve the desired performance with a smaller model. We have demonstrated a model compression strategy that is fast to apply, and doesn’t require any additional engineering. Furthermore, the optimisation algorithm of the larger model is sufficient to train the cheaper student model. The cheap convolutions used in this paper were chosen for their ease of implementation. Future work could investigate more complicated approximate operations, such as those described in Moczulski et al. [2015]; which could make a difference for the 1×1 convolutions in the final layers of a network. One could also make use of custom blocks generated through a large scale black box optimisation as in [Zoph et al., 2017]. Equally, there are many methods for low rank approximations that could be applicable [Jaderberg et al., 2014, Garipov et al., 2016, Sainath et al., 2013]. We hope that this work encourages others to consider cheapening their convolutions as a compression strategy. Acknowledgements. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 732204 (Bonseyes). This work is supported by the Swiss State Secretariat for Education‚ Research and Innovation (SERI) under contract number 16.0159. The opinions expressed and arguments employed herein do not necessarily reflect the official views of these funding bodies. 10 Depth-Width Block Teacher 40-2 16-2 40-1 16-1 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 S S S S S-2x2 G(2) G(4) G(8) G(16) G(N/16) G(N/8) G(N/4) G(N/2) G(N) B(2) B(4) BG(2,2) BG(2,4) BG(2,8) BG(2,16) BG(2,M/16) BG(2,M/8) BG(2,M/4) BG(2,M/2) BG(2,M) BG(4,M) No. Params Error KD Error AT Error 2248954 693498 566650 175994 1012474 1369530 825210 553050 416970 651834 466362 373626 327258 304074 437242 155002 292090 219514 183226 165082 234704 195322 171130 159034 152986 85450 4.79 6.53 6.48 8.81 5.89 5.30 5.50 5.92 6.65 5.72 6.07 6.93 7.12 8.51 6.36 7.94 6.12 6.75 6.94 6.77 6.26 6.75 7.06 7.45 7.95 9.04 – 6.03 6.39 8.75 6.03 5.37 5.81 5.72 6.38 5.72 5.61 6.45 6.83 8.01 6.28 7.83 6.25 6.75 6.98 6.97 6.50 6.49 7.15 7.47 7.99 8.61 – 5.66 5.50 7.72 5.09 4.87 5.00 5.05 5.13 5.12 5.06 5.31 5.98 6.57 5.37 6.93 5.57 6.05 6.09 6.19 6.02 5.94 6.03 6.17 6.67 7.87 Table 3: Student Network test error on CIFAR-10. Each network is a Wide ResNet with its depth-width given in the first column, and with its block type in the second. S is a standard convolutional block. S − 2 × 2 is a standard block with dilated 2x2 kernels. G(g) is a grouped + pointwise block with g groups. B(b) is a bottleneck block with contraction b. BG(b, g) is a bottleneck block with contraction b and a grouped convolution with g groups. N refers to the channel width of each block, and M refers to the channel width after the bottleneck where applicable. The total parameter cost of the network is given in the third column. Errors are reported for (i) learning with no distillation (Error), (ii) knowledge distillation with a teacher (KD Error), and attention transfer with a teacher (AT Error). The same teacher is used for training, and is given in the first row. This table shows that (i) through attention transfer it is possible to cut the number of parameters of a network, but retain high performance and (ii) for a similar number of parameters, students with cheap convolutional blocks, outperform those with expensive convolutions and smaller architectures. 11 Depth-Width Block Teacher 40-2 16-2 40-1 16-1 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 40-2 S S S S S-2x2 G(2) G(4) G(8) G(16) G(N/16) G(N/8) G(N/4) G(N/2) G(N) B(2) B(4) BG(2,2) BG(2,4) BG(2,8) BG(2,16) BG(2,M/16) BG(2,M/8) BG(2,M/4) BG(2,M/2) BG(2,M) BG(4,M) No. Params Error KD Error AT Error 2260564 705108 572500 181844 1024084 1381140 836820 564660 428580 663444 477972 385236 338868 315684 448852 166612 303700 231124 194836 176692 255316 206932 182740 170644 164596 97060 23.85 27.63 29.64 34 27.2 25.94 26.20 26.49 28.85 27.08 27.85 28.91 30.24 31.84 28.27 31.63 28.51 29.39 30.21 30.57 29.69 29.09 30.42 30.44 30.90 33.64 – 27.97 30.21 37.28 26.98 24.92 25.48 26.64 27.10 26.11 27.05 27.93 28.89 29.99 28.08 33.63 28.82 29.25 29.34 30.54 28.69 29.13 30.28 30.66 31.18 37.34 – 27.24 28.24 33.74 26.09 24.45 25.30 25.71 26.34 25.78 26.15 26.85 28.54 30.06 26.68 30.56 28.28 28.54 28.89 29.46 29.05 28.16 28.60 29.51 30.03 32.89 Table 4: Student Network test error on CIFAR-100. Each network is a Wide ResNet with its depth-width given in the first column, and with its block type in the second. These blocks are described in detail in Section 3.2. The total parameter cost of the network is given in the third column. Errors are reported for (i) learning with no distillation (Error), (ii) knowledge distillation with a teacher (KD Error), and attention transfer with a teacher (AT Error). The same teacher is used for training, and is given in the first row. 12 References Lei Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In Advances in Neural Information Processing Systems, 2014. Cristian Buciluǎ, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2006. François Chollet. Xception: Deep learning with depthwise separable convolutions. CoRR, abs/1610.02357, 2016. URL http://arxiv.org/abs/1610.02357. Taco Cohen and Max Welling. Group equivariant convolutional networks. In Maria Florina Balcan and Kilian Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 2990–2999, New York, New York, USA, 20–22 Jun 2016. PMLR. URL http://proceedings.mlr.press/v48/cohenc16.html. George Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control, Signals, and Systems (MCSS), 2(4):303–314, 1989. Misha Denil, Babak Shakibi, Laurent Dinh, Marc’Aurelio Ranzato, and Nando de Freitas. Predicting parameters in deep learning. In Advances in Neural Information Processing Systems, 2013. Timur Garipov, Dmitry Podoprikhin, Alexander Novikov, and Dmitry P. Vetrov. Ultimate tensorization: compressing convolutional and FC layers alike. CoRR, abs/1611.03214, 2016. URL http://arxiv.org/abs/1611.03214. Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and Huffman coding. CoRR, abs/1510.00149, 2015. URL http://arxiv.org/abs/1510.00149. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016a. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European Conference on Computer Vision, 2016b. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. CoRR, abs/1503.02531, 2016. URL http://arxiv.org/abs/1503.02531. Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017. URL http://arxiv.org/abs/1704.04861. Forrest N. Iandola, Matthew W. Moskewicz, Khalid Ashraf, Song Han, William J. Dally, and Kurt Keutzer. SqueezeNet: AlexNet-level accuracy with 50× fewer parameters and < 1MB model size. CoRR, abs/1602.07360, 2016. URL http://arxiv.org/abs/ 1602.07360. Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pages 448–456, 2015. Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions. In British Machine Vision Conference, 2014. Jonghoon Jin, Aysegul Dundar, and Eugenio Culurciello. Flattened convolutional neural networks for feedforward acceleration. In International Conference on Learning Representations, 2015. 13 Alex Krizhevsky. Learning multiple layers of features from tiny images. Master’s thesis, University of Toronto, 2009. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553): 436–444, 2015. Zhe Li, Xiaoyu Wang, Xutao Lv, and Tianbao Yang. SEP-Nets: Small and effective pattern networks. CoRR, abs/1706.03912, 2017. URL http://arxiv.org/abs/1706. 03912. Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. In International Conference on Learning Representations, 2014. Marcin Moczulski, Misha Denil, Jeremy Appleyard, and Nando de Freitas. ACDC: a structured efficient linear layer. CoRR, abs/1511.05946, 2015. URL http://arxiv. org/abs/1511.05946. Adam Paszke, Sam Gross, Soumith Chintala, and Gregory Chanan. PyTorch: Tensors and dynamic neural networks in Python with strong GPU acceleration. https://github. com/pytorch/pytorch. Accessed: 31st October 2017. Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. FitNets: Hints for thin deep nets. CoRR, abs/1412.6550, 2014. URL http://arxiv.org/abs/1412.6550. Tara N. Sainath, Brian Kingsbury, Vikas Sindhwani, Ebru Arisoy, and Bhuvana Ramabhadran. Low-rank matrix factorization for deep neural network training with high-dimensional output targets. In IEEE International Conference on Acoustics, Speech and Signal Processing, 2013. Laurent Sifre. Rigid-Motion Scattering for Image Classification. PhD thesis, École Polytechnique, 2014. Gregor Urban, Krzysztof J. Geras, Samira Ebrahimi Kahou, Ozlem Aslan, Shengjie Wang, Rich Caruana, Abdelrahman Mohamed, Matthai Philipose, and Matt Richardson. Do deep convolutional nets really need to be deep and convolutional? In International Conference on Learning Representations, 2017. Min Wang, Baoyuan Liu, and Hassan Foroosh. Factorized convolutional neural networks. CoRR, abs/1608.04337, 2016. URL http://arxiv.org/abs/1608.04337. Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. Zichao Yang, Marcin Moczulski, Misha Denil, Nando de Freitas, Alex Smola, Le Song, and Ziyu Wang. Deep fried convnets. In Proceedings of the IEEE International Conference on Computer Vision, 2015. Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. In International Conference on Learning Representations, 2016. Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference, 2016. Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In International Conference on Learning Representations, 2017. X. Zhang, X. Zhou, M. Lin, and J. Sun. ShuffleNet: An extremely efficient convolutional neural network for mobile devices. CoRR, abs/1707.01083, 2017. URL http:// arxiv.org/abs/1707.01083. Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. Learning transferable architectures for scalable image recognition. CoRR, abs/1707.07012, 2017. URL http: //arxiv.org/abs/1707.07012. 14	1cs.CV
1 R. Serra et al. arXiv:1407.8508v1 [q-bio.MN] 30 Jul 2014 A stochastic model of catalytic reaction networks in protocells Roberto Serra1,2 , Alessandro Filisetti2 , Marco Villani1,2 , Alex Graudenzi3 , Chiara Damiani3,4 , and Tommaso Panini5 1 Dept. of Physics, Informatics and Mathematics, Modena and Reggio Emilia University 2 European Centre for Living Technology, ECLT, University Ca’ Foscari of Venice, Italy 3 Dept. of Informatics, Systems and Communication, University of Milan Bicocca, Italy 4 SYSBIO - Centre for Systems Biology, Piazza della Scienza 2, 20126 Milano, Italy 5 University of Turin, Department of Economics and Statistics "Cognetti de Martiis", Torino February 27, 2018 Abstract Protocells are supposed to have played a key role in the self-organizing processes leading to the emergence of life. Existing models either (i) describe protocell architecture and dynamics, given the existence of sets of collectively self-replicating molecules for granted, or (ii) describe the emergence of the aforementioned sets from an ensemble of random molecules in a simple experimental setting (e.g. a closed system or a steady-state flow reactor) that does not properly describe a protocell. In this paper we present a model that goes beyond these limitations by describing the dynamics of sets of replicating molecules within a lipid vesicle. We adopt the simplest possible protocell architecture, by considering a semi-permeable membrane that selects the molecular types that are allowed to enter or exit the protocell and by assuming that the reactions take place in the aqueous phase in the internal compartment. As a first approximation, we ignore the protocell growth and division dynamics. The behavior of catalytic reaction networks is then simulated by means of a stochastic model that accounts for the creation and the extinction of species and reactions. While this is not yet an exhaustive protocell model, it already provides clues regarding some processes that are relevant for understanding the conditions that can enable a population of protocells to undergo evolution and selection. 1 2 R. Serra et al. Keywords: Autocatalytic Sets of Molecules, Catalytic Reaction Sets, Origin of Life, Stochastic Simulations, Protocell 1 Introduction It is widely believed that the origin of life required the formation of sets of molecules able to collectively self-replicate, as well as of compartments able to undergo fission and proliferate, i.e.protocells [36, 37, 49]. In particular, in order to observe a lifelike behavior it was necessary that some of the chemical reactions were coupled to the rate of proliferation of the compartments. Several protocell architectures have been proposed [3, 12, 15, 27, 29, 30, 34, 37, 38, 39, 41, 44, 48, 49], most of them identifying the compartment with a lipid vesicle that may spontaneously fission under suitable circumstances. On the other hand, many distinct models were proposed to describe sets of reactions involving randomly generated molecules [1, 10, 20, 26]. In many cases, although this is not in principle required, it is assumed that only catalyzed reactions take place at a significant rate, therefore these sets are also termed catalytic reaction sets (briefly, CRSs). It is worth noting that the appearance of new molecules implies the appearance of new reactions involving those new molecules, so that both the set of molecular types and the set of reactions change in time. Hence, it is possible that at a certain time a set of molecules able to catalyze each other’s formation emerges [10, 20, 26], and we will refer to it as an autocatalytic set (ACS). It can be noticed that a CRS can contain one or more ACSs, or none. Even though some models of protocell actually describe the coupling between reaction networks and the dynamics of a lipid container, they consider only a fixed set of molecular species and reactions [3, 28, 40, 52], hence providing an incomplete representation of this complex interplay. Conversely, while there are several studies on collectively self replicating sets of molecules in a continuously stirred open-flow tank reactor, CSTR [1, 19, 23, 45, 46, 47] including our own [11, 13, 42], they provide only limited information about the behavior of a protocell. Therefore, in order to develop a framework that may unify the CRSs and the protocell modeling approaches, it is necessary i) to analyze the behavior of CRSs in a vesicle, and ii) to investigate the coupling of the evolving chemical population with the growth of the lipid container and its fission. In this paper we propose a step towards the first goal, while deferring the second one to a further work. In particular, we here analyze the behavior of a dynamical model of CRSs in a simplified model of a non-growing vesicle. To the best of our knowledge this is a novel approach. A few important remarks. Let us first observe that the CSTR is not an a-priori good model of a protocell for at least two reasons: i) in general, in protocells there is no constant inflow and ii) protocells have semipermeable membranes, which allow the inflow/outflow of some, but not all, molecular types. 2 3 R. Serra et al. On the contrary, in open flow reactors all that is contained in the inflow enters the reactor and all that is dissolved in the reactor can be washed out in the outflow.1 Another important limit of the CSTR concerns its evolvability. It has been argued [31, 51] that the presence of different asymptotic dynamical states and the ability to shift between them may be essential to achieve the viable evolution of the first forms of life. Recent works [50] have found that, in models of catalytic reaction networks in CSTRs, generally only one of these states is found, apart from fluctuations [5]. Furthermore, in order to accomplish the goal of this work, we need to better specify both the model of catalytic reactions sets and that of the protocell. As far as the former is concerned, we have studied the dynamics of random sets of molecules by revisiting a model by Kauffman, who proposed an interesting way to build new molecular species from the existing ones (see section 2 for a description). The original version of the model relied on purely graph-theoretical arguments [26], which are important, but fail to appreciate the effects of the dynamics, including noise, fluctuations and small-number effects. The dynamics has been later introduced by Farmer et al. [1, 10], who described the kinetics by using ordinary differential equations. However, this formalism does not account for the chance of a species to become extinct in a finite amount of time, as it may instead well happen (so the reaction graph may grow but never shrinks). In order to overcome these limitations, Bagley proposed an empirical correction by setting to zero the concentration values that happen to fall below a certain threshold [1]. In our works we rather use from the very beginning a stochastic approach to analyze the dynamics, the well-known Gillespie algorithm [16, 17], in order to deal in a rigorous way with low concentrations and with their fluctuations. Note that the Kauffman model largely relies upon randomness. In particular, every polymer in the system has a fixed probability (that may vanish) to catalyze any possible reaction. Therefore, in different simulations the same species can catalyze different reactions leading to the formation of different “chemistries”. Thus, this is exactly the language we choose: a set of tuples {species, catalysis, reaction}, where the species catalyzes the reaction, will be called “chemistry”, because it describes a possible artificial world.2 We can then simulate different chemistries and look for generic properties of the set of chemistries; but in a different series of experiments we can also keep the chemistry fixed, and simulate various time histories. In principle, these may differ, since the discovery of a given catalyst at an early phase in a finite system might channel the following evolution in a way or another. Since the number of molecules of some species may be very small, it is not in principle legitimate 1 Limitations on the outflow can be modeled in a chemostat e.g. by supposing that all the molecules that are larger than a certain size precipitate and cannot be washed away. 2 It is worthwhile to notice that the presence of the “catalysis” within the tuple allows the possibility for a species to catalyze more than one reaction and for a reaction to be catalyzed by more than one species. 3 4 R. Serra et al. to ignore this aspect, and our stochastic model is particularly well suited to analyze it, as it will be shown in Section 2. Of course, there can be conditions where all the simulations of a given chemistry converge asymptotically to the same chemical mixture. Moving now to the protocell model, note that they are usually based on lipid vesicles, i.e. approximately spherical structures with an aqueous interior and a membrane composed by a lipid bilayer, which spontaneously form when lipids are mixed with water under certain conditions [2, 7, 18, 27, 32, 34, 35]. Even though different protocell architectures have been proposed, we will here consider the simplest model, namely that in which all the key reactions take place in the aqueous phase inside the protocell. It would be indeed straightforward to model the coupling between some of these molecules and the growth of the protocell following an approach similar to that of our previous studies [3, 12, 40]. Yet, the main objective of the present work is that of studying the dynamics of CRSs embedded in a vesicle, so we will simplify our treatment by ignoring the growth dynamics of the protocell, and keeping its volume fixed. This implies that our study will be limited to time intervals that are short with respect to those describing the growth of the whole protocell. The selective character of membranes is a key ingredient of our model: we will suppose for simplicity that all (and only) the molecules that are shorter than a certain length can cross the membrane. The transmembrane motion of the permeable species is here supposed to be driven by the difference of their concentrations in the internal aqueous volume of the protocell and in the external aqueous environment. We will assume that transmembrane diffusion is extremely fast, so that there is always equilibrium between the concentrations of the species that can cross the membrane; this adiabatic hypothesis could be easily relaxed in the future. Furthermore, we assume that protocells are turgid, so that the constant-volume approximation implies that we will also neglect issues related to osmotic pressure. Another related aspect of the model is that, since it is assumed that the permeable species are at equilibrium, while the non-permeable ones never cross the barriers, infinite concentration growth is possible; this is obviously a nonphysical behavior, so the model validity is limited in time. All these simplifications, which will be removed in subsequent studies, are also justified by the fact that our main goal is that of studying how the dynamics of CRSs are affected by being embedded in a vesicle. This model can be used in order to investigate the behavior of the system in different conditions and to address some important questions. The first and perhaps most important one is the reason why compartments seem to be necessary for life. Indeed, the very first studies on self-replicating molecules [8, 9, 24, 26] were not interested in this aspect, so the CRSs were supposed to exist, e.g., in a pond or in a beaker. Yet life seems to require compartments, that are ubiquitous. It is then important to understand whether there are major differences between what may happen in a protocell and what happens in the bulk phase. 4 5 R. Serra et al. It would be unconvincing to postulate a priori that the internal and external environments are different. It is indeed more likely to assume that the vesicles form in a pre-existing aqueous environment, so the average internal milieu is essentially the same as the external one.3 Then, if a membrane surrounds a portion of the fluid, what can happen that makes a difference? Let us first observe that protocells are small (their typical linear dimensions ranging from 0.1 to 10µm). If we imagine that a population of protocells exists, and they are not “overcrowed”, their total internal volume will typically be much smaller than the total external volume (this is a fortiori true for an isolated one). Moreover, every point in the interior of a protocell is not allowed to be far away from the surface of the protocell that contains it. These observations imply that the effect of surfaces will be much larger within protocells then outside them. Suppose for example that the membrane hosts some catalytic activities, so that important molecules are synthesized close to its boundaries, both inside and outside, and diffuse freely. If the membrane width is much smaller than the protocell radius, then the internal and external surface areas are very close to each other, but the external volume is much larger than the external one: therefore the internal concentrations will be much higher than those in the external environment. In this case, the system behavior in the interior can be significantly different from the external one [43]. Note also that this effect may be different for different molecules: the formation of some of them might be catalyzed by the membrane, while others might be unaffected: so even the relative concentrations of different chemicals may differ in the two cases. Indeed, there are important protocell models that are based on such an active catalytic role of the membrane [37]. In these cases it is easy to understand what the role of the protocell is, since it provides essential catalysts and a way to keep their products closer. But protocells might be able to give rise to an internal environment different from the bulk even if the catalytic activity is absent. The reason for this seemingly counterintuitive behavior is, once again, the smallness of the protocells. Note that we are considering a case where new molecules are formed (from those that are already in the interior of the protocell plus those that can cross the semipermeable membrane). If the concentrations are not too high, it is likely that the total numbers of newly formed molecules are quite low, so that different protocells might host different groups of molecules. It might even happen that a molecular type is present in some protocells and not in others. In order to get a feeling for this possibility, let us provide some realistic estimates of the number of molecules of different types that can be present in a protocell. Let us consider typical vesicles (linear dimension about 1µm) and small ones (0.1µm). Typical concentrations of macromolecules may be in the 3 It has been observed that some superconcentration phenomena can take place under particular circumstances [6, 43] but we will neglect them here. 5 6 R. Serra et al. Typical (1µ3 ) Small (10−3 µ3 ) 1M 108 105 1mM 105 102 0.1mM 104 10 1µM 102 0.1 1nM 0.1 10−4 Table 1: excepted number of molecules of a given species in a given protocell; rows refer to protocell volumes, columns to concentrations. millimolar to nanomolar range; the excepted numbers of molecules in a single protocell are therefore given in table 1. Let us recall that these numbers refer to the excepted values, but there are fluctuations that may be relevant when small numbers are involved. For example, in the case of a 1µM concentration in small vesicles, there will be on average 1 molecule every 10 cells: it is apparent that different protocells will have widely different initial compositions. We therefore come to the conclusion that the creation of small compartments can allow the formation of a population of different individuals out of the fluctuations of an environment that looks macroscopically homogeneous. While not yet sufficient, this is definitely a necessary condition for darwinian evolution to take place (obviously supposing that the compartments can divide and that their division rate depends upon composition). Moreover, in small stochastic systems it may also happen that there are different trajectories stemming from the same initial conditions, due to the order in which new molecules are synthesized. These aspects of protocell dynamics are very important and our model, in spite of its current limitations, is well suited to explore the related phenomena. It is indeed possible to analyze the different possible stochastic effects, that include: (i) the path dependency induced by the random order in which new molecules are generated, with particular regard to the low concentration effects: if a catalyst is discovered in advance with respect to another, the system evolution may be different; this can be studied by comparing different simulations referring to the same chemistry, starting from the same initial conditions; (ii) the differences induced by different initial conditions, randomly generated from the same distribution; (iii) the different behaviors of distinct chemistries: in the real world the rules of the chemistry are given, but in the kind of analysis performed here it is also interesting to understand how different chemistries may affect the behavior of the system and the diversity of a population of protocells. As, for example, the role of RAF sets [20, 21, 22] in the overall dynamics. In Section 3 we present the results of the simulations on the model. In the concluding Section 4 we will discuss the main findings of this paper, and we will propose further analysis and refinement of the models. 6 7 2 R. Serra et al. Main features of the model Entities and interactions. The model describes an open system in which simple molecules interact with each other through elementary catalyzed reactions. The basic entities of the system are monomers and polymers, identified by ordered strings of letters oriented from left to right taken from a finite alphabet (e.g. A, B, C, . . .). We will refer to the “letters” also as bricks, while the term monomer will be reserved to those molecular types that are composed of a single brick. Every species xi , i = 1, 2, . . . , N composing the entire set of species X = [x1 , x2 , . . . , xN ] is characterized by its specific amount (either quantity or concentration), denoted by x̂i . The number of copies of a specific species defines its number of molecules. The two basic reactions are: i) cleavage, i.e. the cutting of a species composed of more than one brick into two shorter species (e.g. ABBA → A+BBA) and ii) condensation, i.e. the concatenation of two species in a longer one (e.g. BBA + AAB → BBAAAB). A key assumption of the model is that any reaction occurs only if catalyzed by one of its specific catalysts; hence we exclude the presence of spontaneous reactions by assuming sufficiently high activation energy for each reaction. In particular, condensation requires an intermediate reaction in which a temporary complex between the catalyst and the substrate is formed. A further restriction regarding catalysis imposes that only species that are composed of at least a minimum number of bricks can be catalysts. Additionally, we neglect the presence of backward reactions (exception made for the dissociation reaction of intermediate complexes) by hypothesizing that the Gibbs energy ∆G for any reaction is sufficiently large. A schematic representation of the various reaction types is given in Fig. 1. The reaction network. Provided that a catalyst for the reaction exists, each species in the system could condensate with any other species in the system, or be split at any cutting point into smaller species. The cardinality of the set R of all the conceivable reactions for the system, at a given time, is therefore given by: N X R̂ = (L(xi ) − 1) + N 2 (1) i=1 where N is the cardinality of X and L(xi ) is the length of xi (i.e. the number of bricks of that specific species). The first term of Eq. 1 refers to the conceivable cleavages and the second one to the conceivable condensations. Hence, the set of reactions that are actually allowed, i.e. the “chemistry” of the system, is determined by the catalysts involved in the various reactions. Following Kauffman, we define the chemistry of the system at random. In this regard, any species xi has a finite fixed probability pi of being chosen as catalyst 7 8 R. Serra et al. CONDENSATION K comp S1 + C → S1C A K decomp K cond S1C → S1 + C S1C + S2 → P + C A B B A A A B B B B A B B A A A B B A B B A B B A B B B A A B B A A B A A A B B B A B CLEAVAGE K cleav S + C → P1 + P2 + C A A B B A B B B B A A A B B B A B B B A A B A B Figure 1: Schematic representation of the two possible reaction types, namely condensations and cleavages. S, S1 and S2 represent the substrates of a reaction, C is a species with catalytic function, S1 C is the transient complex that is formed in the first step of the condensation and P, P1 and P2 stand for the products of a reaction. Kcleav , Kcomp , Kdiss and Kcond respectively stand for the kinetic rates of cleavage, complex formation, complex dissociation and final condensation. of a randomly chosen reaction among those belonging to R̂. It is worth stressing that, although the reaction network is built probabilistically, it remains invariant in time; in other words, once a species is chosen to be the catalyst of a given reaction, that species will always be catalyst for that reaction. Further details concerning the structures of the model can be found in [11, 14]. The dynamics. The system’s dynamics is simulated by means of an extension of the well-known Gillespie algorithm [16, 17] for the stochastic simulation of chemical reaction systems, in which we allow the appearance of novel species and reactions that are not present in the system in the initial conditions. In fact, cleavage and condensation of elements either present within the protocell or entering it from the external environment can generate new species, which, in turn, can be involved in new reactions, both as catalysts and substrates. The container: the introduction of the protocell. In its classical formulation [1], the catalytic reaction network is modeled within a continuous stirred-tank reactor (CSTR), in which the ingoing and outgoing fluxes could be adjusted according to the experimental needs. In this work, we introduce a major modification of the model by introducing a semi-permeable membrane that separates the catalytic reaction network from the external environment (see Fig. 2). The semi-permeable membrane is here modeled by allowing only some species to enter and leave the protocell, namely those that are shorter than an arbi- 8 9 R. Serra et al. Figure 2: Schematic representation of the semi-permeable membrane as conceived in our model of protocell. The membrane is here represented as a lipid bilayer that shapes a spherical vesicle. In this example, only the species shorter than 3 letters can cross the membrane, entering or leaving the internal compartment, the dynamics of which is actually simulated in our model of catalytic reaction networks. On the opposite, longer species are confined either inside or outside the vesicle. trary length Lperm . All the species that are longer than Lperm either remain entrapped within the protocell or never enter it from the outside. Another important assumption is that the concentration of the permeable molecules is homogeneous both inside and outside the protocell, and that they take the same value. That is, we assume infinitely fast diffusion both in the bulk phases and across the membrane, so that the chemical potentials of the permeable species are the same. The species that can cross the membrane will be also defined, from now on, as the buffered species. In this work, we consider the volume of the protocell to be constant, while it is planned to introduce protocell size and duplication dynamics in further developments of the model. We also remark that in this model the transport processes in a vesicle are treated in a way that, albeit simplified, parallels their actual dynamics. While small flow reactors have also been proposed as protocell models [22, 50], they are indeed not well suited, since they require a constant inflow that has no physical analogue in a vesicle, and usually also allow the outflow of all the solutes. 9 10 R. Serra et al. Algorithm 1 Formal definition of the model as multiset-rewriting system. 1: let Ψ be the alphabet of symbols denoting monomers, A, B, C...; 2: the Θ∗ and Ξ∗ be the (infinite) set of all finite-length non-empty polymers and substrate-catalyst complexes over Ψ, and let us denote by P · Q (resp. P _Q, substrate-catalyst) the polymer (resp. complex) obtained by concatenating polymers P and Q; 3: let the state of the system be multiset Φ ≡ P1 \| P2 \| ... \| Pn counting the occurrences of each polymer/complex Pi ∈ {Θ∗ ∪ Ξ∗ }; 4: consider the following reaction schema with variables X, Y and Z, in a multiset-rewriting notation: • [A: Cleavage] Rcl : X · Y \| Z 7→ X \| Y \| Z {Z is the catalyst, X · Y is the polymer breaking to X and Y } • [B1: Complex formation] Rcf : X\|Y 7→ X_Y • [B2: Complex dissociation] Rcd : X_Y 7→ X \| Y • [B3: Final condensation] Rf c : X_Y \| Z 7→ X · Z \| Y {X and Z are substrates, Y is the catalyst, X_Y is the substratecatalyst complex, X · Z the polymer composed by X and Z} Variable polymers (X · Y ) take values in Θ∗ , whereas complexes (X_Y ) in Ξ∗ . 5: (Reactions generation) an instantiation map ϕ is evaluated via standard pattern matching to transform the above schema in all the possible rewriting rules Φ Φ Φ Rϕ = Rcl ∪ Rcf ∪ Rcd ∪ RfΦc for multiset Φ. The map ϕ, which assigns values to X, Y and Z, is required to be maximal and univocal, i.e. it must yield all the possible rewriting rules applicable in Φ and instantiated consistently (via deterministic association) for any of its evaluation. 6: Φ and Rϕ yield a Continuous-time Markov Chain. By firing a reaction in Rϕ via the classical Gillespie algorithm a new state of the system is generated. In this particular implementation of the model, polymers shorter than 3 are fed and cannot be instantiated as catalysts (see Appendix A). The simulation proceeds with the new state from previous step. 3 Results of the simulations The preliminary experiments on the protocell model were performed by keeping fixed some key structural parameters (see the Appendix for the details) and 10 11 R. Serra et al. creating different random chemistries. In particular, we here present the details of two specific chemistries, which were specifically selected among many (on a total of 10 simulated chemistries) because they exhibit very specific dynamical properties. In this regard, it is first important to introduce the concept of RAF set, which be fundamental in the description of these systems. Following Hordijk et al. [20, 21, 22], given the entire chemistry, a subset of reactions R is: (i) Reflexively autocatalytic (RA) if every reaction in R is catalyzed by at least one other reaction belonging to R, (ii) Food-generated (F) if every substrate in R can be produced from the food set F by means of the reactions belonging to R, (iii) Reflexively autocatalytic and F-generated (RAF) if both the previous conditions are satisfied. The chemistries of the two presented protocells differ for the presence/absence of a RAF set (RAF in short from now on): in particular, in the first protocell (CH1) no RAFs are present, whereas in the second protocell (CH2) a RAF formed by an autocatalysis consuming molecules from the food set is present. In order to assess the behavior of the system, a method to measure the similarity between two different states of the system is that proposed by Kaneko [25] and is based on the comparison of the vectors that describe their chemical compositions. Let us define the k-dimensional vectors Cj (t) = [cj,1 (t), cj,2 (t), ..., cj,N (t)] and Ck = [ck,1 (t), ck,2 (t), ..., ck,N (t)] whose components are the concentrations of the species present at time t in systems j and k respectively. Same positions in vectors refer to same species, hence species present in system j that are not present in k have concentration equal to 0 and viceversa. The similarity between the two systems is then computed by means of the normalized inner product: ! C~j (t) · C~k (t) −1 (2) Θt = cos \|\|Cj (t)\|\| · \|\|Ck (t)\|\| where Θt is the angle between the two vectors measured at the time t. Throughout this paper, angles will be measured in degrees. 3.1 Low concentration effects In this section we investigate the low concentration effects in protocells. To this end, the selected chemistries are tested with respect to 4 different initial uniform concentrations of the chemical species present inside the vesicle: in detail, the initial concentration of all the non-buffered species is equal to: 1mM , 0.1mM , 0.01mM and 1µM . Conversely, the amount of each buffered species is always fixed to 1mM . Notice that, in general, the protocell model presented here does not lead to stationary chemical distributions. Indeed, situations are possible where some particular species indefinitely increases its concentration: an example is the RAF set in CH2, where a particular species catalyses its own formation using two buffered species as substrates: in this case an exponential growth is achieved. Another simple case could be the linear growth of a species produced directly from the buffered species by means of a catalyst not involved in other reactions, 11 12 R. Serra et al. Conc Molarity (Cond.1) 1mM (Cond.2) 0.1mM (Cond.3) 0.01mM (Cond.4) 1µM Molecules per species Average 600 60 6 1 CH1 Θ3000 (mean) 0.41 2.34 7.71 11.15 CH2 Θ3000 (max.) 0.68 5.86 15.28 19.35 Θ3000 (mean) 0.06 0.18 9.69 3.67 CH2 (no RAF) Θ3000 (max.) 0.19 0.52 21.86 11.91 Θ3000 (mean) 0.57 1.69 6.48 9.44 Θ3000 (max.) 0.96 2.78 11.21 15.35 Table 2: In the table the average and the maximum values of Θ3000 relative to 10 distinct simulations of each of the 4 different initial conditions (rows in the table) are shown. The measures are reported for the 2 different chemistries: the one without RAFs and the one with RAFs, for which they are also computed without taking into account the species belonging to the RAF, column “CH2 (no RAFs)”. The 4 conditions differ in the average magnitude of the initial concentrations of the initial set of molecular species not belonging to the buffered flux (the food set). A realization of each of the 4 initial concentration is drawn at random from a Poisson distribution, according to the given parameters, and is maintained invariant across the 10 different runs. i.e. present in constant concentration. Of course, more complex situations are possible, even not directly originating from the food set. Yet, many simulations show a transient in which the chemicals distribution rapidly change followed by a long quasi-equilibrium, where changes are limited to small adjustments. By applying the parameters adopted in the present work, such a state is already achieved in 2500 seconds, except for the case of very low concentrations (1µM ). Diversity. In table 2 some statistics on how Θ varies according to the different cases are shown and, in particular, it is possible to appreciate an increase of the dissimilarity among protocells in correspondence with a decrease on the initial concentration, as indicated by the average and the maximum value of Θ between all the runs of each specific chemistry and specific initial condition, computed at time t = 3000.4 Since these runs are characterized by identical chemistries and identical initial conditions, the angle reported here is indeed the result of dynamical evolutions that differ only for the simulation random seed. It is worth remarking that the lower is the concentration profile, the higher is the distance among the final distributions,5 i.e. the angle Θ reaches its maximum 4 Exceptionally for the case 1µM , Θ is not computed after 3000 second but when at least 5000 reactions have occurred within the simulation. The reason for this is that the low concentrations involve a so slow dynamics that 3000 seconds are not enough in order to observe significant chemical changes. 5 In regard to CH2 we are here considering the value of Θ excluding the species belonging to the RAF set (last columns of the table). Since the molecules belonging to the RAF set reach a concentration much greater with respect to the other molecules, considering them in the angle computation would misrepresent the distance among the simulations. 12 13 R. Serra et al. value with regard to very low initial concentrations (e.g. 1 molecule per species on average). It is interesting to notice that the displayed low concentration effects with respect to the overall similarity of the system are found in both chemistries, hinting at a generic property of such systems, independently of the possible presence of RAF sets. The influence on RAF set dynamics. The asynchronous framework implies that only one reaction occurs at the time. Given that RAFs in general involve more than one reaction, in order to detect their presence we analyze the system in a time interval (i) sufficiently large to let a significant number of reactions to occur, yet (ii) not embracing the whole simulation (in order to avoid the presence of too rare reactions); It is worth stressing that since the analysis are made ex post with respect to the simulation of the system, they do not affect the simulated dynamics, but just the way of representing it. In fig. 3, we show the presence (or the absence) of the RAF of CH2 every 10 seconds, by using a time window of 5 seconds,6 with respect to the 4 aforementioned different average concentrations. Note that in all the initial conditions the autocatalytic species has a concentration higher than zero, thus the corresponding RAF is always able to achieve its own growth viability. From Fig. 3 it is possible to observe the strong correlation between the average concentration of the species within the protocell and the presence of RAFs. As long as the concentration decreases, so does the probability of detecting a RAF, yet in the long run in most of the simulations RAFs eventually emerge. 3.2 Sensitivity to initial conditions We here analyze the effects of variations in the initial concentrations of the single species present within the protocells, while maintaining their average concentration fixed. We simulate 10 different variations of the single species concentrations for both the case of average concentration equal to 0.01mM (condition 3) and average concentration 1µM (condition 4). In Fig. 4 we display the variation of the similarity in time for each couple of simulations, with respect to both conditions, in order to provide a picture of the sensitivity of the overall diversity to the initial conditions. Note that the very low concentrations of condition 4 (in average, only one molecule for each species) imply that in certain protocell realizations some species are missing: in particular, each simulation starts from a different set of species composed of, on average, 40 species, on the possible 62. This may explain the very high values of Θ0 with respect to condition 4 (Fig. 4). Besides, one can see that condition 4 shows an evident bifurcation (Fig. 4b). Given that in some simulations, because of the low concentration effects, the 6 We set the sampling frequency and the time threshold of the windows by taking advantage from several initial model threads, not essential to the comprehension of this article. 13 14 R. Serra et al. Figure 3: From the top to the bottom the four traces, accounting for conditions 1, 2, 3 and 4 of the chemistry 2, i.e. the one containing the RAF set, depicting the presence of the RAF during the simulation (time flows from left to right, each row represents a different run of the same condition). The black zones stands for the absence of the RAF set, while gray zones denote the presence of the RAF. It is important to remark that RAFs are searched in a dynamic network created with the reactions occurred in a mobile time window of 5 seconds. autocatalytic species (and so the RAF) can be present or absent in the initial condition, the system can indeed reach different regions of the state space, leading to deeply different histories. On the opposite, condition 3 shows the aforementioned regulatory activity of the always active RAF. (a) (b) Figure 4: The angles between each couple of different simulations in time, for condition 3 (0.01mM - left panel) and condition 4 (1µM - right panel). Finally, it is important to remark that during each simulation many species disappear and many other appear, so a compact way to follow the process is that of monitoring, during each run, the variation of the angle ΘT 0 →T t between the initial and the current chemical setting (Fig. 5a). The graph indicates that Θ0→3000 approaches a value close to 80 degrees, which means a quasiorthogonality of the system with respect to the initial condition. 14 15 R. Serra et al. The dynamics of the system, in term of molecular concentrations, leading to such a divergence from the beginning is shown in Fig. 5b. It is reasonable to hypothesize that the clear inversions in the trajectories of the concentrations of certain species may be responsible for this phenomenon. (a) (b) Figure 5: On the left panel the average angle (and the standard deviation) ΘT0 →Tt measured from T0 to Tt for each chemical distribution, within each single run of conditions 1 (average molarity 1mM ) and 2 (average molarity 0.1mM ), is shown. It is worthwhile to remark that Θ = 90 stands for a complete orthogonality between the chemical distributions. In order to appreciate the convergence towards a quasi-orthogonality, on the right panel the concentration of the species in time of a particular simulation are shown (Simulation: CH2, condition 1, run 1). The exponential growth concentration of the autocatalytic species is shown with a limit of 3mM . 4 Conclusion and further developments In this paper we introduced a simplified model of a non-growing protocell and we investigated the behavior of a stochastic model of catalytic reaction networks in such an environment. To the best of our knowledge this is a novel approach. The crucial importance of the small size of the protocell has been stressed, and the effects of the fact that some chemicals can be present in low numbers have been investigated. While a broader analysis is ongoing, we have here shown that it is possible to reach different compositions of the chemical species, in the particular case in which some species are present in the bulk at low concentrations. We have also shown that there are two different, possibly overlapping reasons for this diversity: (i) the random sequence of molecular events involving those species and (ii) the random differences in their initial concentrations. We have also stressed the importance of RAF sets in influencing the overall dynamics. There are several ways in which this work might seed further research. The most obvious is that of relaxing the physical limitations that have been considered, e.g. infinitely fast diffusion, yet we do not except that this may change the major conclusions summarized above. 15 16 R. Serra et al. Obviously, a very interesting direction is that of considering a protocell that is able to grow and divide. The processes involved in protocell growth and replication are indeed complex and, in particular, a necessary condition for its existence and replication is the coupling between the rates of molecules replication and cell growth. We have shown elsewhere [3, 12, 40] that the very existence of this coupling suffices to guarantee (under very general conditions) that, in the long run, the rate of cell division and that of duplication of the replicating molecules converge to the same value, thereby allowing sustainable growth of a population of protocells.7 However, those results were achieved by supposing a fixed set of genetic memory molecules, with some possible extinction. It could be sound to extend this approach to the case where there are evolving chemical reaction sets and to verify whether synchronization occurs. An important aspect to be addressed in the case of growing vesicles is also the effect of volume growth on the concentrations of the various chemicals: a preliminary investigation can be found in [4]. Besides, we have not here explicitly considered the possibly catalytic role of membranes that, as it has been discussed in section 1, might be a major cause of the difference between the intracellular environment(s) and the bulk. In a fixed volume model this effect can be lumped in the effective reaction rates, but if we consider a growing protocell we have to take into account the differences between surfaces and volumes. This might also lead to interesting phenomena that will be analyzed in future developments. To conclude, different protocells may host different mixtures of molecular species, even if they share the same chemistry (i.e., they “inhabit the same world”). It might be extremely interesting to model the behavior of populations of different protocells of this kind, which may show different growth rates, but may also undergo phenomena like coalescence, exchange of material, etc. Thus, further investigations will be indeed necessary to assess different generations of protocell populations and their possible evolution pathways. Last but not least, it will be interesting to extend these studies to other protocell architectures like e.g. the Los Alamos bug, Gard models, and others. 5 Acknowledgements Stuart Kauffman, Norman Packard and Wim Hordijk kindly shared with us their deep understanding of autocatalytic sets in several useful discussions. Useful discussions with Ruedi Füchslin, Davide De Lucrezia, Timoteo Carletti, Andrea Roli and Giulio Caravagna are also gratefully acknowledged. The authors are also grateful to Giulia Begal for kindly drawing the image of Fig. 2. C.D. wishes to acknowledge the project SysBionet (12-4-5148000-15; Imp. 611/12; CUP: H41J12000060001; U.A. 53) for the financial support of the work. The final publication is available at Springer via http://dx.doi.org/10.1007/s11047014-9445-6 7 This property was proved earlier by Munteanu et al. for the Los Alamos bug mode [33]. 16 17 R. Serra et al. A Appendix 1 - Simulation environment and parameter settings Simulations were performed with the CaRNeSS simulator8 developed by the research group. In the following, the baseline setting of the system used in the simulations is reported (for the parameters that were variated in the different experiments please refer to the text): (•) Alphabet: A, B, (•) Volume = = 1e − 18dm3 = 1µ3 , (•) Average catalysis probability = 1 catalyzed reaction for species, (•) Maximum length of the species, Lmax = 6, (•) Lperm = 2, (•) Monomers and dimers do NOT catalyze, (•) Kcleav = 25M −1 sec−1 , (•) Kcomp = 50M −1 sec−1 , (•) Kdiss = 1M −1 sec−1 , (•) Kcond = 50M −1 sec−1 . References [1] R.J. Bagley and J D Farmer. Spontaneous emergence of a metabolism. Artificial Life II. Santa Fe Institute Studies in the Sciences of Complexity, X:93–141, 1992. [2] Ann-Sofie Cans, Meghan Andes-Koback, and Christine D Keating. Positioning lipid membrane domains in giant vesicles by micro-organization of aqueous cytoplasm mimic. J Am Chem Soc, 130(23):7400–7406, 2008. [3] T Carletti, R Serra, M Villani, I Poli, and A Filisetti. Sufficient conditions for emergent synchronization in protocell models. J Theor Biol, 254(4):741–751, 2008. [4] Timoteo Carletti and Alessandro Filisetti. The Stochastic Evolution of a Protocell: The Gillespie Algorithm in a Dynamically Varying Volume. Computational and Mathematical Methods in Medicine, 2012(Article ID 423627):1–13, 2012. [5] Xiaofeng Dai, Olli Yli-Harja, and Andre S Ribeiro. Determining noisy attractors of delayed stochastic gene regulatory networks from multiple data sources. Bioinformatics (Oxford, England), 25(18):2362–8, September 2009. [6] Tereza Pereira de Souza, Pasquale Stano, Frank Steiniger, Erica D’Aguanno, Emiliano Altamura, Alfred Fahr, and Pier Luigi Luisi. Encapsulation of Ferritin, Ribosomes, and Ribo-Peptidic Complexes Inside Liposomes: Insights Into the Origin of Metabolism. Origins of life and evolution of the biosphere : the journal of the International Society for the Study of the Origin of Life, October 2012. [7] Lisa M Dominak and Christine D Keating. Polymer encapsulation within giant lipid vesicles. Langmuir, 23(13):7148–7154, 2007. [8] Freeman J Dyson. Origins of life. Cambridge: Cambridge University Press, 1985. [9] M Eigen and P Schuster. The hypercycle. A principle of natural self-organization. Part A: Emergence of the hypercycle. Die Naturwissenschaften, 64(11):541–65, November 1977. [10] JD Farmer and SA Kauffman. Autocatalytic replication of polymers. Physica D: Nonlinear Phenomena, 220:50–67, 1986. 8 https://github.com/paxelito/carness 17 18 R. Serra et al. [11] A Filisetti, A Graudenzi, R Serra, M Villani, D De Lucrezia, and I Poli. The role of energy in a stochastic model of the emergence of autocatalytic sets. In Lenaerts T, Giacobini M, H Bersini, P Bourgine, M Dorigo, and R Doursat, editors, Advances in Artificial Life, ECAL 2011 Proceedings of the Eleventh European Conference on the Synthesis and Simulation of Living Systems, pages 227–234. MIT Press, Cambridge, MA, 2011. [12] A. Filisetti, R. Serra, T. Carletti, M. Villani, and I. Poli. Non-linear protocell models: synchronization and chaos. The European Physical Journal B, 77(2):249–256, June 2010. [13] Alessandro Filisetti, Alex Graudenzi, Roberto Serra, Marco Villani, Davide De Lucrezia, Rudolf M Fuchslin, Stuart A Kauffman, Norman Packard, and Irene Poli. A stochastic model of the emergence of autocatalytic cycles. Journal of Systems Chemistry, 2(1):2, 2011. [14] Alessandro Filisetti, Alex Graudenzi, Roberto Serra, Marco Villani, Rudolf M Füchslin, Norman Packard, Stuart a Kauffman, and Irene Poli. A stochastic model of autocatalytic reaction networks. Theory in biosciences = Theorie in den Biowissenschaften, pages 1–9, October 2011. [15] T Ganti. Chemoton Theory, Vol. I: Theory of fluyd machineries: Vol. II: Theory of livin system. New York: Kluwer Academic/Plenum, 2003. [16] Daniel T Gillespie. Exact Stochastic Simulation of Coupled Chemical Reactions. The Journal of Physical Chemistry, 81(25):2340–2361, 1977. [17] Daniel T Gillespie. Stochastic Simulation of Chemical Kinetics. Annual Review of Physical Chemistry, 58(1):35–55, 2007. [18] Martin M Hanczyc, Shelly M Fujikawa, and Jack W Szostak. Experimental models of primitive cellular compartments: encapsulation, growth, and division. Science (New York, N.Y.), 302(5645):618–622, October 2003. [19] Wim Hordijk and Jose F Fontanari. Catalytic Reaction Sets, Decay, and the Preservation of Information, 2002. [20] Wim Hordijk, Jotun Hein, and Mike Steel. Autocatalytic Sets and the Origin of Life. Entropy, 12(7):1733–1742, June 2010. [21] Wim Hordijk and Mike Steel. Detecting autocatalytic, self-sustaining sets in chemical reaction systems. Journal of theoretical biology, 227(4):451–461, April 2004. [22] Wim Hordijk, Mike Steel, and Stuart Kauffman. The Structure of Autocatalytic Sets: Evolvability, Enablement, and Emergence. page 12, May 2012. [23] S Jain and S Krishna. A model for the emergence of cooperation, interdependence, and structure in evolving networks. Proc Natl Acad Sci U S A, 98(2):543–547, 2001. [24] Sanjai Jain and Sandeep Krishna. Autocatalytic set and the growth of complexity in an evolutionary model. Phys Rev Lett, 81:5684–5687, 1998. [25] Kunihiko Kaneko. Life: An Introduction to Complex Systems Biology (Understanding Complex Systems). Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2006. [26] S A Kauffman. Autocatalytic sets of proteins. J Theor Biol, 119(1):1–24, 1986. [27] P.L. Luisi, F. Ferri, and P. Stano. Approaches to semi-synthetic minimal cells: a review. Naturwissenschaften, 93(1):1–13, 2006. [28] Sheref S Mansy. Model Protocells from Single-Chain Lipids. International Journal of Molecular Sciences, 10(3):835–843, 2009. 18 19 R. Serra et al. [29] Sheref S Mansy, Jason P Schrum, Mathangi Krishnamurthy, Sylvia Tobé, Douglas A Treco, and Jack W Szostak. Template-directed synthesis of a genetic polymer in a model protocell. Nature, 454(7200):122–125, 2008. [30] Harold J. Morowitz, Bettina Heinz, and David W. Deamer. The chemical logic of a minimum protocell. Origins of Life and Evolution of the Biosphere, 18(3):281–287, September 1988. [31] Simon Conway Morris. Life’s Solution, Inevitable Humans in a Lonely Universe. University of Cambridge, 2003. [32] O G Mouritsen. Life - As a Matter of Fat. The Emerging Science of Lipidomics. Springer, Berlin, first edition, 2005. [33] A Munteanu, C S Attolini, Steen Rasmussen, H Ziock, and R V Solé. Generic Darwinian selection in protocell assemblies. DOI: SFI-WP 06-09-032, SFI Working Papers, Santa Fe Institute, 2006. [34] Andreea Munteanu and Ricard V Solé. Phenotypic diversity and chaos in a minimal cell model. J Theor Biol, 240(3):434–442, 2006. [35] Alla Polozova, Xingong Li, Tong Shangguan, Paul Meers, Daniel R Schuette, Nozomi Ando, Sol M Gruner, and Walter R Perkins. Formation of homogeneous unilamellar liposomes from an interdigitated matrix. Biochimica et Biophysica Acta (BBA) Biomembranes, 1668(1):117–125, 2005. [36] Steen Rasmussen, Liaohai Chen, David Deamer, David C.Krakauer, Norman H Packard, Peter F Stadler, and Mark A Bedau. Transitions from Nonliving to Living Matter. Science, 303, 963-965, 2004. [37] Steen Rasmussen, Liaohai Chen, Martin Nilsson, and Shigeaki Abe. Bridging nonliving and living matter. Artificial life, 9(3):269–316, January 2003. [38] Tristan Rocheleau, Steen Rasmussen, Peter E Nielsen, Martin N Jacobi, and Hans Ziock. Emergence of protocellular growth laws. Philos Trans R Soc Lond B Biol Sci, 362(1486):1841–1845, 2007. [39] D Segré and D Lancet. Composing life. EMBO reports, 1(3):217–22, September 2000. [40] Roberto Serra, Timoteo Carletti, and Irene Poli. Syncronization phenomena in surfacereaction models of protocells. Artificial life, 123(2):123–38, 2007. [41] Roberto Serra, Timoteo Carletti, Irene Poli, Marco Villani, and Alessandro Filisetti. Conditions for emergent synchronization in protocell. In In J. Jost and D. Helbing (eds): Proceedings of ECCS07: European Conference on Complex Systems. CD-Rom, paper n.68, 2007. [42] Roberto Serra, Alessandro Filisetti, Marco Villani, Chiara Damiani, Alex Graudenzi, and Tommaso Panini. A stochastic model of catalytic reaction networks in protocells. Submitted to Artificial Life, 2013. [43] Roberto Serra and Marco Villani. Mechanism for the formation of density gradients through semipermeable membranes. Physical Review E, 87(6):062814, June 2013. [44] R.V. Ricard V Solé, Andreea Munteanu, Carlos Rodriguez-Caso, Javier Macía, and J. Mac\’\ia. Synthetic protocell biology: from reproduction to computation. Philos Trans R Soc Lond B Biol Sci, 362(1486):1727–1739, 2007. [45] P F Stadler. Dynamics of autocatalytic reaction networks. IV: Inhomogeneous replicator networks. Bio Systems, 26(1):1–19, January 1991. 19 20 R. Serra et al. [46] P F Stadler and P Schuster. Dynamics of small autocatalytic reaction networks–I. Bifurcations, permanence and exclusion. Bulletin of mathematical biology, 52(4):485–508, January 1990. [47] Peter F. Stadler, Wolfgang Schnabl, Christian Forst, and Peter Schuster. Dynamics of Small Autocatalytic Reaction Networks II : Replication, Mutation and Catalysis, 1995. [48] Pasquale Stano and Pier Luigi Luisi. Achievements and open questions in the selfreproduction of vesicles and synthetic minimal cells. Chemical communications (Cambridge, England), 46(21):3639–53, June 2010. [49] J W Szostak, D P Bartel, and P L Luisi. Synthesizing life. Nature, 409(6818):387–390, 2001. [50] Vera Vasas, Chrisantha Fernando, Mauro Santos, Stuart Kauffman, and Eors Szathmary. Evolution before genes. Biology Direct, 7(1):1, January 2012. [51] R. Wesson. Beyond Natural Selection. MIT Press, Cambridge, MA, 1991. [52] Ting F Zhu and Jack W Szostak. Coupled Growth and Division of Model Protocell Membranes. Journal of the American Chemical Society, 131(15):5705–5713, 2009. 20	5cs.CE
1 Adaptive Delivery in Caching Networks Seyed Ali Saberali, Hamidreza Ebrahimzadeh Saffar, Lutz Lampe, and Ian Blake arXiv:1707.09662v1 [cs.IT] 30 Jul 2017 Abstract The problem of content delivery in caching networks is investigated for scenarios where multiple users request identical files. Redundant user demands are likely when the file popularity distribution is highly non-uniform or the user demands are positively correlated. An adaptive method is proposed for the delivery of redundant demands in caching networks. Based on the redundancy pattern in the current demand vector, the proposed method decides between the transmission of uncoded messages or the coded messages of [1] for delivery. Moreover, a lower bound on the delivery rate of redundant requests is derived based on a cutset bound argument. The performance of the adaptive method is investigated through numerical examples of the delivery rate of several specific demand vectors as well as the average delivery rate of a caching network with correlated requests. The adaptive method is shown to considerably reduce the gap between the non-adaptive delivery rate and the lower bound. In some specific cases, using the adaptive method, this gap shrinks by almost 50% for the average rate. Index Terms Adaptive delivery algorithm, average delivery rate, coded caching, correlated requests, placement optimization, redundant demands. I. I NTRODUCTION Local content caching is a promising technique to meet the unprecedented traffic demands in the next generation communication networks [1]–[6]. Caching networks take advantage of the users’ contextual information to predict the future user demands. This enables the network to store the popular content at storage nodes, also known as caches, close to the end users and satisfy the user requests locally [3]. 2 Caching networks operate in two phases, which are commonly referred to as placement and delivery phases. In the placement phase, the caches fill their memories with parts of the popular files up to their storage capacity. This phase takes place when the network traffic is low. In contrast, the delivery phase is performed when the network is congested. In this phase, upon the users’ requests, each cache provides its users with the parts of the files that it has available. The remaining parts of the files are conventionally delivered to the users through separate unicast transmissions performed by a central server on a channel that is shared by the users. In a more recent caching approach [2], known as coded caching, the central server uses simultaneous coded-multicasting to deliver the requested content to the users to further reduce the network congestion. An information-theoretic formulation of coded caching was developed in [2]. The authors defined the delivery rate as the total traffic on the shared communication link due to the server’s messages, such that the users’ requests are satisfied. Moreover, they proposed a centralized coded-caching scheme to reduce the delivery rate. In a later work [1], a decentralized caching scheme was proposed that did not require any coordination between the caches to operate. The decentralized nature of this method made it the building block of several caching schemes that were designed later for more complicated scenarios [6]–[11]. Both [1], [2] used the peak delivery rate as the figure of merit of the caching network. The peak rate results from the worst-case demand vector, where all the users request distinct files. However, the average delivery rate is also a significant performance metric of a caching network. Average delivery rate depends on the statistics of the user requests. Thus, the statistical patterns in the user demands can significantly affect the design of the caching scheme. One statistical property of user demands is the popularity distribution of the files. The caching schemes of [1], [2] can be used if the popularity distribution is uniform. On the other hand, [5]–[8] have proposed different caching schemes to account for non-uniform popularities. In particular, the caching schemes of [6], [7] are designed based on grouping of the files into several popularity 3 groups, with the files in each group having relatively close popularity levels. They provide more storage resources to the files in the more popular groups. Then, they use the decentralized caching scheme of [1] within each group separately. Also, [8] groups the library of files into two groups of popular and unpopular files. The requests for popular files are delivered through the delivery algorithm of [1], while the requests of unpopular files are delivered through uncoded messages. The same problem is investigated in [5], assuming a Zipf popularity distribution and independent and identically distributed user requests. The placement of [5] is based on the partitioning of each file into equal length packets and randomly distributing the packets (not bits) over the caches. Unlike the other schemes, the delivery of [5] is not based on the delivery algorithm of [1], but on chromatic number index coding. In contrast to [6]–[8], this scheme does not restrict the coding opportunities to the requests within each popularity group. However, its implementation is more complicated as it requires vertex coloring of a conflict graph. The statistics of the users’ requests can further affect the design of caching networks through increasing the chance of multiple identical requests. In such a scenario, one might be able to modify the delivery algorithm to benefit from the redundancies in the user demands, to further reduce the average delivery rate. Redundant demands are likely to be made when the files have significantly different popularity levels or when there are positive correlations among the requests of different users. For the case of non-uniform file popularities, the schemes in [6]–[8] do not take the effect of identical requests into account during the delivery phase. This is because the delivery in all these schemes is based on the delivery of [1], which is designed for the demand vectors with distinct requests. In addition to non-uniform popularity levels, correlated user requests are likely in many practical scenarios. A considerable amount of multimedia requests are made through the social networks like Facebook, Twitter and Instagram and movie providing websites like Netflix. In such scenarios, the users with overlapping circles of friends, the ones who follow the same people or pages, and those who live in the same geographical area or have common personal, social and professional interests are likely to get suggestions for the same content in 4 their media feeds, and therefore, request the same files. In this paper, we investigate the delivery of redundant demands in caching networks. We study a model where placement is fixed, yet the requests are changing by the time and the delivery adapts to the requests. We propose an adaptive delivery scheme based on message selection to minimize the delivery traffic. Specifically, upon receiving a demand vector from the users, the server exploits the redundancy pattern in the user demands to decide whether to use uncoded messages or the coded messages of [1] to deliver each part of the files requested. We assume that the placement phase is accomplished through the placement schemes of [1], [2]. This ensures that the peak delivery rate does not exceed the delivery rates of [1], [2], so the link capacity constraints are satisfied. Further, if the file popularities are relatively uniform or little prior knowledge about the popularity distribution is available during the placement time, it is natural to accomplish the placement as in [1], [2]. In the delivery phase, however, the users reveal their demands to the server. The server can use this knowledge as a side information and adapt its choice of coded and uncoded messages accordingly to benefit from the possible redundancies in the requests. To the best of the authors’ knowledge, this paper is the first work in the literature to consider this scenario and to specifically design a scheme for the delivery of redundant requests. Although we use the placement schemes in [1], [2], our proposed delivery method is based on an optimization formulation of the content placement problem. Namely, we use a modified version of this problem to optimize the choice of coded and uncoded messages in our proposed delivery scheme. A side result of the placement optimization problem is the generalization of the centralized placement of [2] to arbitrary cache sizes. In particular, we derive the parameters of the centralized caching analytically for cases where the total cache capacity is not an integer multiple of the total size of the files in the library. We show the superiority of our adaptive method through numerical examples for several specific demand vectors. We derive a lower bound on the delivery rate of the redundant requests 5 based on a cutset bound argument, and compare the rate of the proposed delivery method with the lower bound. Moreover, we study the dynamics of a caching system with correlated user demands. We apply Gibbs sampling [12], [13], to generate sample demand vectors based on a stochastic modeling of the dependencies among the user requests. It is shown that the proposed method is superior to the conventional non-adaptive method in terms of the average delivery rate. In some specific cases, the adaptive method decreases the gap between the average rate of the non-adaptive scheme and the lower bound by almost 50%. The remainder of this paper is organized as follows. In Sec. II, we present the network model and review the caching schemes of [1], [2]. We formulate the rate minimization problem in Sec. III. In Sec. IV, we propose the adaptive delivery scheme and derive a lower bound on the delivery rate. Sec. V presents numerical examples and simulation results. Finally, we conclude the paper in Sec. VI. II. P ROBLEM M ODEL AND R EVIEW In this section, we explain the problem model and briefly review the caching schemes of [1], [2]. Assume a network with a central server and K caches, where the server is able to communicate with the caches through a broadcast link (see Fig. 1). We denote the set of all caches in the network by K. A library of N ≥ K popular files is given, where each file is F bits long. We assume that all files are available at the central server and that each cache has a memory capacity of M × F bits. q , M/N represents the ratio of the cache size to the library size. Placement Phase: In the placement phase, the caches fill their memories with parts of the popular files based on a placement algorithm. We assume that placement takes place only once and remains unchanged during the delivery phase. The resulting distribution of bits in the caches can be described as follows. For a given file n and a given subset of caches S ⊂ K, denote by VSn the subset of bits of file n that are exclusively 6 Server Cache 1 Cache 2 Cache K −1 Cache K ··· Fig. 1. A network with K caches and a central server. stored at the caches in S. Note that the resulting subsets of bits partition the set of all the bits of every file into 2K partitions. Define s , \|S\| and xs , \|VSn \|/F (1) as the portion of the bits of file n that are exclusively stored at each subset S of caches with cardinality s. Here, we have assumed that \|VSn \| only depends on s. In particular, it neither depends on n nor on the particular choice of caches in S as long as the cardinality of S is s. This holds because of symmetry, as we assume a uniform distribution over file popularities. The placement phase can be performed through either the centralized scheme of [2] or the decentralized scheme of [1]. The centralized caching scheme of [2] can be used only when t , KM is an integer. For the centralized placement, split each file into Kt non-overlapping N subfiles of the same length F/ Kt . Assign each one of these subfiles to a subset of caches S : s = t, in a one-to-one manner. Store the bits belonging to each subfile in all the caches in 7 the corresponding S. This results in xcen s =    1/ K s   0 For the decentralized placement, each cache stores s=t . (2) s 6= t M F N bits of each file uniformly at random. It can be shown that for large F [1] xdecen ≈ q s−1 (1 − q)K−s+1 , s s = 0, ..., K (3) with high probability. Delivery Phase: In the delivery phase, the network serves one user of every cache at a time. Denote the requests of the users of caches 1, ..., K with d1 , ..., dK , respectively. We refer to the vector [d1 , ..., dK ] as the demand vector. Note that the demand vector evolves with time during the delivery phase. We represent the number of distinct files in the demand vector by L, where 1 ≤ L ≤ K. We call the demand vector redundant if L < K. In addition, denote by ki , the number of requests for the i-th most requested file in the current demand vector. Thus ki ≥ kj for i > j and i, j ∈ {1, ..., L}. We call (k1 , ..., kL ) the redundancy pattern of the demand vector. For a demand vector [d1 , ..., dK ], we define the delivery rate R(M, [d1 , ..., dK ]) as the traffic on the shared broadcast link due to the server’s messages, such that all the caches successfully recover the files they requested. We express the rate in terms of the equivalent total number of files that must be transferred on the shared link. So, a rate of R files is equivalent to R × F bits. To construct file dk , cache k needs to receive VSdk for all S ⊂ K\{k}. The server, delivers these bits to the caches through the coded delivery messages given by Algorithm 1 proposed in [1]. Notice that the delivery method for the centralized caching in [2] is a special case of Algorithm 1. n Note that if file n is requested by multiple users, including user k, Algorithm 1 embeds VS\{k} into several messages. If s > 1, user k has the side information to decode only one of those messages. As a result, the server needs to send all the messages with s > 1, even though the 8 Algorithm 1 Delivery algorithm of [1] Require: {VSn }n=1,...,N, S⊂K # From the placement phase Procedure Delivery(d1 , ..., dK ) for s = K, K − 1, ..., 1 do for S ⊂ K : \|S\| = s do dk server sends ⊕k∈S VS\{k} end for end for demand vector is redundant. This is not the case for the messages with s = 1, i.e., S = {k}. In dk these cases, ⊕k∈S VS\{k} = V∅dk . Such uncoded messages deliver the bits that are not stored at any cache in the system. All the users that request file n can decode V∅n , so it needs to be sent only once. As a result, the traffic due to the uncoded messages is Lx0 instead of Kx0 . Thus, the total delivery rate will be Lx0 + K−1 X s=1 K xs . s+1 Note that when L = K, substitution of (2) and (3) in (4) gives M 1 K 1− N 1+KM N as the peak rate of the centralized caching scheme [2], and M (1 − (1 − M/N ))K K 1− N KM/N (4) (5) (6) as the peak rate of the decentralized caching scheme [1]. From (4), one notes that for the redundant demand vectors, the actual rate of Algorithm 1 is smaller than (5) and (6) for the centralized and decentralized caching schemes, respectively. This observation is the basis of our analysis in Sec. IV. 9 III. O PTIMALITY OF THE C ENTRALIZED P LACEMENT WITH DELIVERY BASED ON A LGORITHM 1 As was formerly discussed, we use either of the methods in [1] or [2] for the placement phase of our caching scheme. In this section, we show that the centralized placement scheme of [2] is the optimal placement to minimize the peak rate of delivery Algorithm 1. We also generalize the application of the centralized placement to the cases that t = KM/N is not an integer. The optimal placement is characterized by the optimal parameters xs that lead to the smallest peak delivery rate of Algorithm 1. Based on (4), the peak rate minimization problem can be formulated as K−1 X K minimize xs xs s+1 s=0 K X K subject to xs = 1 s s=0 K X K −1 M xs ≤ s−1 N s=1 (7) xs ≥ 0, s = 0, 1, ..., K. The first constraint in (7) ensures that the resulting subsets partition the bits of each file. It also guarantees that xs ≤ 1. The second constraint represents the storage capacity constraint. The objective function of (7) is the the worst-case delivery rate of Algorithm 1. We now present the analytical solution of (7) in Proposition 1. Proposition 1 (Optimal Placement for the Worst-Case Demand): Let t = K M . The solution N of (7) is x∗s =    1/   0, K t , s=t (8a) otherwise 10 M/N x∗0 x∗1 x∗2 x∗3 x∗4 x∗5 t = KM/N 0.1 0.5 0.1 0 0 0 0 0.5 0.2 0 0.2 0 0 0 0 1 0.3 0 0.1 0.05 0 0 0 1.5 0.5 0 0 0.05 0.05 0 0 2.5 0.8 0 0 0 0 0.2 0 4 0.9 0 0 0 0 0.1 0.5 4.5 TABLE I. Optimal file placement parameters for K = 5. if t is an integer, and is x∗s =    (dte − t)/     (t − btc)/       0, K btc , s = btc K dte , s = dte (8b) otherwise if t is not an integer. btc and dte denote the largest integer smaller than t and the smallest integer larger than t, respectively. Proof. See Appendix A. Proposition 1 shows that the centralized placement of [2] is optimal for Algorithm 1 when t is an integer. Further, it generalizes the centralized placement scheme to the caching systems with non-integer t. Table I shows the optimal placement parameters for a system of K = 5 caches and a library of N = 1000 files for various storage capacities. Note that two x∗s values are non-zero when t is non-integer. IV. A DAPTIVE C ACHING S CHEME We now design an adaptive delivery method that benefits from the redundancies in the user requests without changing the cache content. Further, we derive a lower bound on the delivery rate of the redundant demand vectors. 11 A. Adaptive Delivery Method For the adaptive method, we introduce an extra step to the delivery phase, which takes place after receiving each request vector and before the transmission of the server messages to the users. In this step, the server decides whether to send each part of the requested files through the corresponding coded message in Algorithm 1 or through an uncoded message. The use of uncoded messages instead of coded messages to deliver file n is equivalent to transferring bits from VSn : s > 0 to V∅n . Notice that by such a transfer, the cache only ignores parts of its content and it does not change the actual placement of files. Let V̂Sn represent the subset of the bits of file n exclusively cached at S after the transfer is done, and ySn , \|V̂Sn \|/F. (9) In our delivery method, the server first optimizes ySn . Then, it arbitrarily picks ySn F bits of VSn to form V̂Sn , and adds the rest of the bits to V̂∅n . Finally, it uses Algorithm 1 for delivery based on the resulting subsets V̂Sn instead of VSn . We now find the optimal lengths of the updated partition sets V̂Sn to minimize the sum of dk the lengths of messages ⊕k∈S V̂S\{k} over all the subsets S ⊂ K. Assume that the caches have requested L ≤ K distinct files in the current demand vector. Denote by D, the set of the distinct files requested in the current demand vector. Note that \|D\| = L, and both D and L evolve with time. For a fixed demand vector (the current demand vector), the rate minimization problem is given by minimize d ySk subject to X S:S⊂K X dk max yS\{k} k∈S ySdk = 1, ∀ dk ∈ D (10) S:S⊂K 0 ≤ ySdk ≤ x\|S\| , ∀ dk ∈ D, ∀S ⊂ K : \|S\| > 0 0 ≤ y∅dk ≤ 1, ∀ dk ∈ D. 12 In (10), x\|S\| = \|VSn \| are known from the placement phase, and are given by (3) and (8) for dk is the length of the the decentralized and centralized placements, respectively. maxk∈S yS\{k} dk . Thus, the objective function is the rate of Algorithm 1 operating based on message ⊕k∈S V̂S\{k} the adjusted subsets V̂Sn . Similar to (7), the equality constraint of (10) is the partition constraint. Also, the constraints on the ranges of the parameters let the server to use uncoded messages instead of coded messages, but not vice versa. Problem (10) can be posed as a linear programming problem by the standard technique of defining ancillary variables dk zS = max yS\{k} (11) k∈S and adding the extra constraints dk dk , , zS ≤ −yS\{k} zS ≥ yS\{k} k∈S (12) for all S ∈ K : \|S\| > 0 [14, Sec. 4.3]. The resulting linear programming problem can be solved numerically for yS∗dk . Algorithm 2 shows the adaptive delivery scheme. B. Simplified Adaptive Delivery A simplified version of the message selection step can be formulated by only taking the number of distinct requests L into account, and ignoring the redundancy pattern of the demand vector. Then, because of the symmetry, we set ySn = ys for all n and all S : \|S\| = s. This leads to minimize ys Ly0 + K−1 X s=1 K ys s+1 K X K subject to ys = 1 s s=0 0 ≤ ys ≤ xs , s = 1, ..., K 0 ≤ y0 ≤ 1 as the simplified message selection problem. (13) 13 Algorithm 2 Original Adaptive Delivery Algorithm Require: {VSn }n=1,...,N, ,S⊂K 1: # From the placement phase Procedure AdaptiveDelivery(d1 , ..., dK ) # Message Selection Step 2: D ← unique(d1 , ..., dK ) # set of distinct files requested 3: {yS∗ dk }dk ∈D,S⊂K ← Solution of Problem (10) 4: for dk ∈ D do 5: V̂∅dk ← ∅ 6: for S ⊂ K do # initialization of V̂∅dk 7: V̂Sdk ← {first yS∗ dk F bits of VSdk } 8: V̂∅dk ← V̂∅dk ∪ { last (1 − yS∗ dk )F bits of VSdk } 9: 10: end for end for # Message Construction Step 11: 12: 13: 14: 15: for s = K, K − 1, ..., 1 do for S ⊂ K : \|S\| = s do dk server sends ⊕k∈S V̂S\{k} end for end for c. Optimal parameters for the simplified message selection Proposition 2: Let ŝ = b K−L L+1 problem of (13) are given by ys∗ =  P     i=1,...,ŝ    0,       x s , K i xi , s=0 s = 1, ..., ŝ . (14) s = ŝ + 1, ..., K Proof: If we transfer bits from the subsets VSn : \|S\| = s to V∅n , the resulting change in the rate 14 will be L K s xs − K s+1 xs . We transfer the bits only if this difference is negative. This is the case when s ≤ ŝ. This results to the parameters of (14). Algorithm 3 shows the simplified adaptive delivery scheme. Algorithm 3 Simplified Adaptive Delivery Algorithm Require: {VSn }n=1,...,N, S⊂K 1: # From the placement phase Procedure SimplifiedAdaptiveDelivery(d1 , ..., dK ) # Message Selection Step 2: L = size(unique(d1 , ..., dK )) # number of distinct requests 3: c ŝ ← b K−L L+1 4: for dk ∈ D do 5: V̂∅dk ← ∪S:s≤ŝ VSdk # corresponds to the first rule of (14) 6: for S ⊂ K : \|S\| > 0 do 7: 8: 9: 10: 11: 12: 13: if \|S\| ≤ ŝ then V̂∅dk ← ∅ # corresponds to the second rule of (14) else V̂Sdk ← VSdk # corresponds to the third rule of (14) end if end for end for # Message Construction Step 14: 15: 16: 17: 18: for s = K, K − 1, ..., 1 do for S ⊂ K : \|S\| = s do dk server sends ⊕k∈S V̂S\{k} end for end for 15 C. Lower Bound Let RL∗ (M ) denote the smallest rate that is achievable for every possible demand vector with L distinct requests. Proposition 3 gives a lower bound on RL∗ (M ) based on a cutset bound argument. Proposition 3 (Cutset Bound): Assume that K caches request L ≤ K distinct files. Then, RL∗ (M ) must satisfy RL∗ (M ) ≥ max s∈{1,...,L} s s− M bN/sc . (15) Proof. See Appendix B. V. N UMERICAL E XAMPLES AND S IMULATION R ESULTS In this section, we investigate the performance of the proposed adaptive delivery method through numerical examples and computer simulations. A. Numerical Examples for Specific Demand Vectors We first consider the performance of the adaptive methods for specific instances of the demand vector. Fig. 2 shows the delivery rates of the non-adaptive delivery scheme of Algorithm 1, the simplified and the original adaptive schemes, and the lower bound in Proposition 3, for a network of K = 9 caches. The placement in all cases is identical and is accomplished through the centralized scheme with the parameters in (8). Also, we calculate the rate of the non-adaptive scheme by (4). In this example, we have considered four redundancy patterns for the demand vector, all with L = 3 distinct file requests. As shown in Fig. 2, the rate of the non-adaptive scheme, the simplified adaptive scheme and the lower bound only depend on L and not the specific redundancy pattern. In contrast, the rate of the original adaptive method depends on the redundancy pattern which has led to different rates for the different patterns. In Fig. 2, we observe a considerable improvement in the delivery rate for M N ≤ 0.25 when the adaptive methods are used. Table II shows the reduction in the gap between the nonadaptive delivery rate and the lower bound when the adaptive schemes are used. At some storage 16 4 Non-Adaptive Simplified Adaptive Adaptive: (3,3,3) Adaptive: (5,2,2) Adaptive: (4,4,1) Adaptive: (7,1,1) Lower Bound * ▲ 3.5 ■ ● * ▲◆● Rate (files) ■ 3▲ ● ◆ ◆ ▲ ■ ▲ ● ■ * ◆ ● ■ 2.5 ▲ ■ ● ◆ ▲● ■ ◆ 2 ▲● ◆ ■ ◆ ▲◆● ■ 1.5 1 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 q=M/N Fig. 2. Comparison of the rate of different delivery schemes for a system with K = 9 caches. In all cases, centralized placement is used and the users only request L = 3 distinct files. (k1 , k2 , k3 ) shows the number of users requesting each file. M/N Delivery Redundancy Method Pattern 0.025 0.1 0.15 0.2 Simplified Adaptive All 49% 52% 37% 13% Adaptive (3, 3, 3) 49% 52% 37% 13% Adaptive (5, 2, 2) 61% 61% 45% 17% Adaptive (4, 4, 1) 66% 66% 51% 25% Adaptive (7, 1, 1) 78% 76% 64% 43% TABLE II. Improvement of the performance gap to the lower bound in Fig. 2. capacities, we observe 50% and 75% reduction in the gap for redundancy patterns (3, 3, 3) and (7, 1, 1), respectively. Also, we notice that for the symmetric redundancy pattern (3, 3, 3), both adaptive methods led to the same delivery rate. As the redundancy pattern gets more asymmetric, the gap between the rate of the original and the simplified adaptive methods increases. Further, we observe that unlike the adaptive schemes, the delivery rate of the non-adaptive method increases 17 4.5 ◊ 4◊ ▲ ◊ * * Rate (files) 3◊ ▲ * ▲ ▲ ◊ 3.5 * ◊ * ◊ ▲ ▲ Non-Adaptive Simplified Adaptive Adaptive Lower Bound ◊ ● ● ▲ ▲ ▲ L = 4: (3, 2, 2, 2) L = 3: (4, 3, 2) ◊ * ▲ 2.5 ▲ ◊ ▲ 2 ◊ ▲ 1.5 1 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 q = M/N (a) Centralized Placement 4◊ ▲◊ ▲ * ◊ * ◊ ▲ 3.5 ◊ 3◊ ▲▲ Rate (files) ** ◊ ▲ ▲ ▲ ◊ ▲ ◊ ▲ ◊ ▲ 2.5 Non-Adaptive Simplified Adaptive Adaptive Lower Bound * ◊ * ● ● ◊ ▲ ◊ ▲ ◊ ▲ ◊ ▲ 2 L = 4: (3, 2, 2, 2) L = 3: (4, 3, 2) ◊ ◊ ▲ ▲ ◊ ▲ ▲ 1.5 ▲ ◊ ▲ 1 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 ▲ ◊ ▲ 0.45 0.5 q=M/N (b) Decentralized Placement Fig. 3. Comparison of the rate of different delivery schemes for a system with K = 9 caches. 18 4.5 ▲ ◊ 4 Rate (files) 3.5 3◊ ◊ ◊ ◊ * ▲ ◊ * ▲ ◊ * ▲ ◊ * ▲ * ▲ ◊ 2.5 ◊ * ▲ 2 ▲ * ▲ 1.5 ° 1▲ * °° 0.5 ◊ 1 ° ◊ * ▲ ◊ ▲ * ▲ ◊ * ° ° ° ▲ ◊ * ° ▲ ◊ * ° M/N = 0.1 M/N = 0.2 ° ° ° 2 3 4 ◊ * ▲ ° Non-adaptive Simplified Adaptive Adaptive (Average) Lower Bound 5 6 7 ° 8 L (a) Centralized Placement 5.5 ▲ ◊ * 5 ▲ ◊ * 4.5 ▲ ◊ * Rate (files) 4 ▲ ◊ * 3.5 ◊ * ▲ ◊ 3 2.5 ◊ 2 ◊ ◊ * ▲ ◊ * ▲ ● ● ▲ * * ▲ ◊ 1.5 * 1▲ * 0.5 ▲ ◊ * ▲ ◊ * ◊ * ▲ 1 ▲ 2 3 4 5 ▲ ◊ * ▲ ◊ * M/N = 0.1 M/N = 0.2 Non-adaptive Simplified Adaptive Adaptive (Average) Lower Bound 6 7 8 L (b) Decentralized Placement Fig. 4. The effect of the number of distinct files requested on the delivery rate. Here K = 8 and N = 103 . with the storage capacity for small M/N . This shows the inefficiency of Algorithm 1 to deliver the redundant requests. Fig. 3 compares the performance of the delivery methods for two different redundancy levels 19 L = 3 and L = 4. The results are shown in Figs. 3a and 3b for the cases where the centralized and decentralized placement schemes are used, respectively. For L = 4, the rate is improved only by the original adaptive method. In general, the simplified method requires higher redundancy levels (smaller L) compared to the original adaptive method, to be able to improve the rate. This fact is shown in Fig. 4, where the delivery rates are plotted versus L. For the original adaptive method, the delivery rate is averaged over all the redundancy patterns with L distinct requests. One notices that the reduction in the delivery rate of the non-adaptive method is more considerable for smaller M/N . This is because when M/N is small, a large number of bits are in V∅n subsets and need to be delivered through uncoded messages. Based on Algorithm 1, the number of uncoded messages decreases by decreasing L. So, the reduction in the rate is larger when M/N is small. B. Simulation of the Network Dynamics We now investigate the average rates of the different delivery methods through a stochastic modeling of the dynamics of a caching network. Consider a graph representation of the network where vertices represent the caches. An (undirected) edge between two vertices shows that the requests of the corresponding caches are correlated. To model the correlation between the requests, we assume that each cache k requests a file, either based on its neighbours previous requests with probability r or independently with probability 1 − r. In the former case, k chooses a file from the set of the last files requested by its neighbours uniformly at random. However, when choosing independently, cache k picks a file n from the library of N files based on the popularity distribution of the files pn . In our simulations, we mainly use uniform popularity distribution, as it is the focus of this paper. We also consider a scenario where the file popularities are assumed to be uniform during the placement phase, but the actual demands in the delivery phase follow a non-uniform distribution. We use Zipf distribution with parameter θ to model the non-uniform file popularities. This gives 20 r = 0.7 r = 0.9 r = 0.9 θ=0 θ=0 θ = 0.75 Maximum ρij 0.19 0.34 0.34 Average ρij 0.16 0.32 0.31 Average L 4.80 3.41 3.18 TABLE III. Empirical correlation coefficients and the resulting number of distinct files per demand vector for the simulations in Fig. 5. ρij is the correlation coefficient between the requests of caches i and j, i 6= j. pn = (1/n)θ PN θ m=1 (1/m) [15]. The larger θ is, the more non-uniform is the popularity distribution. Typical values of θ are between 0.5 and 2 [7]. θ = 0 corresponds to uniform distribution. The model described above completely determines the conditional probabilities of the users’ requests. The chance of requesting file n by cache k can be written as     \|N 1(k)\| r + pn (1 − r), n ∈ N (k) p̂n,k =   pn (1 − r), otherwise (16) where N (k) is the set of the last files requested by the neighbour caches. We use Gibbs sampling [12, Sec. 24.2], [13, Sec. 3] to generate sample vectors from the joint distribution of the user demands based on the network graph and (16). In our simulations, we set K = 8 and N = 103 . We assume a complete graph for the network, i.e., each vertex is of degree K − 1. We use r to control the dependency level of the users’ requests. We also control the popularity distribution by θ. To use Gibbs sampling, we need to give the underlying Markov chain enough burn-in time to reach its stationary distribution. We use the estimated potential scale reduction (ESPR) convergence criterion in [12, Sec. 24.4.3.1] with 5 chains, to determine the burn-in time required. Ignoring the first 150 sample vectors, i.e., 8 × 150 samples, suffices to get \|ESPR − 1\| ≤ 0.01, which shows that the stationary distribution is reached. We use 103 sample vectors after the burn-in time to evaluate the average rate of the different delivery schemes. Table III presents more details about the correlation coefficients and the redundancy levels obtained empirically for each simulation. 21 5 ▲ ◊ * 4.5 ° Average Rate (files) 3 * ▲ ° 4 3.5 ◊ ◊ ▲ * ▲ ◊ ° ▲ ◊ ° ° ◊ * ▲ * ▲ * ° ° ▲ * r = 0.7, θ = 0 r = 0.9, θ = 0 r = 0.9, θ = 0.75 ◊ ◊ ◊ ◊ ▲ * ◊ ▲ * ▲ * ° ◊ * * ▲ ▲ ▲ * °° 2.5 Non-Adaptive Simplified Adaptive Adaptive Lower Bound * ▲ ▲ ◊ ▲ * ▲ * ▲ °° 2 ▲ ◊ ▲ * 1.5 ° ° 1 0.5 0 0.05 0.1 0.15 0.2 ° 0.25 ▲ ◊ ▲ * ° 0.3 M/N Fig. 5. Performance of the different delivery schemes in terms of their avergae delivery rates. Here, K = 8, N = 1000 and the central placement is used. Fig. 5 shows the resulting average delivery rates. It also shows a lower bound on the average rate that is calculated by taking the average of the lower bounds of (15) for the sample demand vectors used. We observe that as requests become more correlated (larger r) and the file popularities get more non-uniform (larger θ), the adaptive method makes larger improvements in the rate. Also, the adaptive schemes are effective in decreasing the average delivery rate for M/N < 0.25. The improvement in the performance gap to the lower bound is shown in Table IV. VI. C ONCLUSION We proposed a new delivery scheme for caching networks that exploits the redundancies in the users’ demand vector to reduce the delivery traffic. The proposed scheme allows the server to decide between the use of coded messages of [1] or uncoded messages for the delivery of each part of the files requested. This choice is made based on the redundancy pattern of the requests in the current demand vector. The server’s decision making process is formulated as a linear programming problem which must be solved numerically. To facilitate the decision 22 a) M/N = 0.075 r = 0.7 r = 0.9 r = 0.9 Delivery Method θ=0 θ=0 θ = 0.75 Adaptive 14% 41% 47% Simplified Adaptive 5% 28% 36% b) M/N = 0.125 r = 0.7 r = 0.9 r = 0.9 Delivery Method θ=0 θ=0 θ = 0.75 Adaptive 16% 41% 48% Simplified Adaptive 5% 28% 36% TABLE IV. Improvement in the performance gap to the lower bound in Fig. 5. making process, a simplified decision rule is also derived analytically. Further, we derived a lower bound on the delivery rate of redundant demands based on a cutset bound argument. The proposed adaptive schemes are shown to significantly improve the delivery rate through several numerical examples. They decreased the performance gap of the non-adaptive method to the lower bound by up to 80% for highly redundant demand vectors. We also investigated the dynamics of a caching network through Markov chain Monte-Carlo simulations and reported the average delivery rate of the adaptive schemes. The adaptive methods considerably outperform the non-adaptive methods in terms of the resulting average delivery rates when M/N < 0.25. We also generalized the application of the centralized placement scheme of [2] to caching networks with non-integer KM/N . A PPENDIX A P ROOF OF P ROPOSITION 1 Proof: By the Karush-Kuhn-Tucker (KKT) conditions [14, Sec. 5.5.3] for optimization problem (7), we get K K −1 0 K − λs + λ +ν = 0, s = 0, ..., K s+1 s−1 s (17) 23 where λs ≥ 0 is the Lagrange multiplier for the inequality constraint xs ≥ 0, and λ0 ≥ 0 and ν are the Lagrange multipliers for the capacity inequality constraint and the partition equality constraint, respectively. KKT conditions require λs x∗s = 0. As a result, x∗s > 0 requires λs = 0. From (17), we can have λs = 0 for at most two s indices. This is because λ0 and ν provide only two degrees of 0 K K freedom to set s+1 + K−1 λ + ν = 0 and the coefficient matrix given by these linear s−1 s equations is rank 2. As a result, we have either one or two x∗s values greater than zero. We consider each case separately. First, assume that only x∗s0 is non-zero. Then, the equality constraint and the capacity constraint require x∗s0 = 1/ sK0 and s0 = K M , respectively. If t = K M is an integer, the optimal solution N N is achieved, as is given by (8a). Otherwise, not all the storage capacity is used and the solution is not optimal, i.e., the optimal solution has more than one non-zero x∗s . So, as the second case, let x∗s > 0 for exactly two values of s, namely s1 and s2 > s1 . From the storage and partition constraints we get K−1 K − s2 M/N s2 −1 K−1 K K−1 K − s1 s2 s1 −1 s2−1 K−1 K − M/N s1 −1 s1 K K−1 . K K−1 − s1 s2−1 s2 s1 −1 x∗s1 = x∗s2 = (18a) (18b) Since x∗s1 , x∗s2 > 0, (18) requires s1 ≤ K Given (18), the objective function K s1 +1 M , N x∗s1 + K − (K + 1) s2 ≥ K K s2 +1 M . N (19) x∗s2 simplifies to KM/N + s1 s2 . (s1 + 1)(s2 + 1) (20) The function in (20) is decreasing in s1 and increasing in s2 in the region specified by (19). Therefore, to minimize the objective function, s1 must take its largest value bK M c, and s2 must N take its smallest value dK M e. Substitution of these values in (18) gives the optimal parameters N in (8b). This completes the proof of Proposition 1. 24 A PPENDIX B P ROOF OF P ROPOSITION 3 Proof: We modify the cutset bound argument of [2, Sec. VI] to bound the minimum delivery rate of the demand vectors with L ≤ K distinct requests. Let S be a subset of caches with \|S\| = s, such that there are no two caches in S with identical user requests. Assume that these caches have requested files 1, ... , s from the library of N files. Let X1 denote the server’s input to the shared link which determines files 1, .., s. Similarly, assume that the same users request files (i − 1)s + 1, ..., is and the server input Xi determines the files requested. Let i = 1, ..., bN/sc. Consider the cut separating X1 , ..., XbN/sc and the caches in S from the corresponding users (see Fig. 6). Since we assume that the coded caching scheme works and all files are perfectly decoded, the total information available to the users in the cut should be more than or equal to the total information requested by them. In other words, bN/scRL∗ (M ) + sM ≥ sbN/sc. Since s can accept any value between 1 and L, (15) results. R EFERENCES [1] M. A. Maddah-Ali and U. Niesen, “Decentralized coded caching attains order-optimal memory-rate tradeoff,” IEEE/ACM Trans. Networking, vol. 23, pp. 1029–1040, Aug. 2015. [2] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” IEEE Trans. Inf. Theory, vol. 60, pp. 2856–2867, May 2014. [3] E. Bastug, M. Bennis, and M. Debbah, “Living on the edge: The role of proactive caching in 5G wireless networks,” IEEE Commun. Mag., vol. 52, pp. 82–89, Aug. 2014. [4] K. Shanmugam, N. Golrezaei, A. G. Dimakis, A. F. Molisch, and G. Caire, “Femtocaching: Wireless content delivery through distributed caching helpers,” IEEE Trans. Inf. Theory, vol. 59, pp. 8402–8413, Dec. 2013. [5] M. Ji, A. M. Tulino, J. Llorca, and G. Caire, “On the average performance of caching and coded multicasting with random demands,” in Proc. 11th International Symposium on Wireless Communications Systems (ISWCS), pp. 922–926, Aug. 2014. [6] J. Hachem, N. Karamchandani, and S. Diggavi, “Content caching and delivery over heterogeneous wireless networks,” in IEEE Conference on Computer Communications (INFOCOM), pp. 756–764, Apr. 2015. 25 Server X1 X2 ··· XbN/sc cache cache cache cache cache 1 2 3 4 5 Fig. 6. An example of a cutset separating the caches and the server from the users of the caches in S = {1, 2, 4}. Solid (dashed) lines represent the information flow to the users selected (not selected) in the cutset. Here s = 3 and there are b Ns c server messages. Users with the same color have identical requests. Notice that no two users with the same request are picked in S. Here K = 5. [7] U. Niesen and M. A. Maddah-Ali, “Coded caching with nonuniform demands,” in IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), pp. 221–226, Apr. 2014. [8] J. Zhang, X. Lin, and X. Wang, “Coded caching under arbitrary popularity distributions,” in Proc. Information Theory and Applications Workshop (ITA), Feb. 2015. [9] R. Pedarsani, M. A. Maddah-Ali, and U. Niesen, “Online coded caching,” in Proc. IEEE Int. Conf. Communications, pp. 1878–1883, June 2014. [10] J. Hachem, N. Karamchandani, and S. N. Diggavi, “Effect of number of users in multi-level coded caching,” in Proc. IEEE Int. Symp. Information Theory, pp. 1701–1705, June 2015. [11] N. Karamchandani, U. Niesen, M. A. Maddah-Ali, and S. Diggavi, “Hierarchical coded caching,” in Proc. IEEE Int. Symp. Information Theory, pp. 2142–2146, June 2014. [12] K. P. Murphy, Machine Learning: A Probabilistic Perspective. The MIT Press, 2012. [13] A. Fischer and C. Igel, “An introduction to restricted Boltzmann machines,” in Progress in Pattern Recognition, Image Analysis, Computer Vision, and Applications, pp. 14–36, Springer, 2012. [14] S. Boyd and L. Vandenberghe, Convex Optimization. New York, NY, USA: Cambridge University Press, 2004. [15] L. Breslau, P. Cao, L. Fan, G. Phillips, and S. Shenker, “Web caching and Zipf-like distributions: evidence and implications,” in IEEE Conference on Computer Communications (INFOCOM), pp. 126–134, Mar. 1999.	7cs.IT
1 A High-Gain Nonlinear Observer with Limited Gain Power arXiv:1501.04330v3 [cs.SY] 7 Oct 2016 Daniele Astolfi and Lorenzo Marconi Abstract—In this note we deal with a new observer for nonlinear systems of dimension n in canonical observability form. We follow the standard high-gain paradigm, but instead of having an observer of dimension n with a gain that grows up to power n, we design an observer of dimension 2n − 2 with a gain that grows up only to power 2. ν(t) ≡ 0, estimating the state x can be addressed by means of a high-gain nonlinear observer of the form Index Terms—Observability, nonlinear observers, high-gain observers. Dn (ℓ) = diag ℓ , . . . , ℓn , I. I NTRODUCTION In this note we consider the problem of state observation for nonlinear systems of the form ¯ , ż = f (z) + d(t) y = h(z) + ν(t) (1) where z ∈ Z ⊆ Rn is the state, y ∈ R is the measured output, ¯ ∈ Rn is f (·) and h(·) are sufficiently smooth functions, d(t) a bounded disturbance and ν ∈ R is the measurement noise. Among the different techniques for observer design available in literature (see [12], [6]) we are particularly interested to the so-called high-gain methods that have been shown to be effective in many control scenarios. In this respect we assume that the pair (f (·), h(·)) fulfils an uniform observability assumption (see Definition 1.2 in [12]), which implies the existence of a diffeomorphism φ : Z → Rn such that the dynamic of the new state variable x = φ(z) is described by the canonical observability form (see Theorem 4.1 in [12]) ẋ = An x + Bn ϕ(x) + d(x, t) , y = Cn x + ν(t) , (2) where ϕ(·) is a locally Lipschitz function, d(x, t) := dφ(z) dz ¯ , d(t) z=φinv (x) with φinv (·) the inverse of φ(·) (namely φinv ◦ φ(z) = z for all z ∈ Z), and (An , Bn , Cn ) is a triplet in “prime form” of dimension n, that is 0(n−1)×1 0(n−1)×1 In−1 , Bn = An = , 0 01×(n−1) 1 1 01×(n−1) . Cn = (3) System (2) is defined on a set X := φ(Z) ⊆ Rn . For the class of systems (2), it is a well-known fact ([7]) that the problem of asymptotically, in case d(x, t) ≡ 0 and D. Astolfi is with CASY-DEI, University of Bologna, Italy and with MINES ParisTech, PSL Research University, CAS, Paris, France, (daniele.astolfi@unibo.it). L. Marconi is with the CASY-DEI, University of Bologna, Italy, (lorenzo.marconi@unibo.it). Work supported by the European project SHERPA (ICT 600958). x̂˙ = An x̂ + Bn ϕs (x̂) + Dn (ℓ) Kn (Cn x̂ − y) (4) with Kn = c1 · · · cn ⊤ , where ℓ is a high-gain design parameter taken sufficiently large (i.e. ℓ ≥ ℓ⋆ with ℓ⋆ ≥ 1), the ci ’s are chosen so that the matrix (A + KC) is Hurwitz (i.e. all its eigenvalues are on the lefthalf complex plane), and ϕs (·) : Rn → R is an appropriate ”saturated” version of ϕ(·). As a matter of fact, it can be proved that if d(x, t) ≡ 0 and ν(t) ≡ 0, if ϕ(·) is uniformly Lipschitz in X , namely there exists a ϕ̄ > 0 such that kϕ(x′ ) − ϕ(x′′ )k ≤ ϕ̄kx′ − x′′ k ∀ x′ , x′′ ∈ X , (5) and ϕs (·) is chosen bounded and to agree with ϕ(·) on X , the observation error e(t) = x(t) − x̂(t) originating from (2) and (4) exponentially converges to the origin with an exponential decay rate of the form kx(t) − x̂(t)k ≤ α ℓn−1 exp(−βℓt)kx(0) − x̂(0)k , where α and β are positive constants, for all possible initial condition x̂(0) ∈ Rn as long as x(t) ∈ X . In particular, note that the exponential decay rate may be arbitrarily assigned by the value of ℓ with a polynomial ”peaking” in ℓ of order n−1. It is worth noting that the uniform Lipschitz condition (5) is automatically fulfilled if X is a compact set. In case d(x, t) or ν(t) are not identically zero, as long as they are bounded for all t ≥ 0 and for all1 x ∈ X , the observer (4) guarantees a bound on the estimation error that depends on the bound of d(·, ·), of ν(·) and on the value of ℓ. In particular, the following asymptotic bounds can be proved lim sup kx(t) − x̂(t)k ≤ 1 γ max lim sup k Γ(ℓ) d(x(t), t)k , lim sup kℓn−1 ν(t)k t→∞ t→∞ ℓ t→∞ where γ is a positive constant and Γ(ℓ) = diag(ℓn−1 , . . . , ℓ, 1) . (6) As above the previous asymptotic bound holds for all possible x̂(0) ∈ Rn as long as x(t) ∈ X . Note that a high value of ℓ leads to an arbitrarily small asymptotic gain on the n-th component of the disturbance d(·, ·). On the other hand, a 1 boundedness of d(x, t) is automatically guaranteed if X is compact. This, in turn, is the typical case when such observers are used in semiglobal output feedback stabilisation problems, [17]. 2 large value of ℓ is, in general, detrimental for the sensitivity of the asymptotic estimate to the sensor noise and to the first n − 2 disturbance components. Observers of the form (4) are routinely used in many observation and control problems. For instance, the feature of having an exponential decay rate and an asymptotic bound on the last component of d that can be arbitrarily imposed by the value of ℓ is the main reason why the above observer plays a fundamental role in output feedback stabilisation and in setting up semiglobal nonlinear separation principles ([17], [4]). In that case the set X is an arbitrarily large compact set which is made invariant by the design of the state feedback stabilisation law and of the high-gain observer. We observe that, although the asymptotic gain with respect to ν increase with ℓn−1 , the observer is anyway able to guarantee ISS with respect to the sensor noise ([14]). The main drawback of observers of the form (4), though, is related to the increasing power (up to the order n) of the high-gain parameter ℓ, which makes the practical numerical implementation an hard task when n or ℓ are very large. Motivated by these considerations, in this note we propose a new observer for the class of systems (2) that preserves the same high-gain features of (4) but which substantially overtakes the implementation problems due to the high-gain powered up to the order n. Specifically, we present a highgain observer structure with a gain which grows only up to power 2 (instead of n), at the price of having the observer state dimension 2n − 2 instead of n. II. M AIN R ESULT We start by presenting a technical lemma instrumental to the proof of the main result presented in Proposition 1. Let Ei ∈ R2×2 , Qi ∈ R2×2 , i = 1, . . . , n − 1, and N ∈ R2×2 be matrices defined as −ki1 1 0 0 0 ki1 Ei = , N= , Qi = −ki2 0 0 ki2 0 1 where (ki1 , ki2 ) are positive coefficients, and let M ∈ R(2n−2)×(2n−2) be the block-tridiagonal matrix defined as   E1 N 0 ... ... 0 ..   ..  Q2 E2 N  . .     . . . . . .. .. .. .. ..  0     ..  .. .. .. M =  .  . . . Q E N . i i    .  . . . . .. .. .. ..  .. 0     .  ..  .. . Qn−2 En−2 N  0 ... ... ... 0 Qn−1 En−1 (7) It turns out that the eigenvalues of M can be arbitrarily assigned by appropriately choosing the coefficients (ki1 , ki2 ), i = 1, . . . , n − 1, as claimed in the next lemma. Lemma 1 Let P(λ) = λ2n−2 + m1 λ2n−3 + ... + m2n−3 λ + m2n−2 be an arbitrary Hurwitz polynomial. There exists a choice of (ki1 , ki2 ), i = 1, . . . , n − 1, such that the characteristic polynomial of M coincides with P(λ). The proof of this Lemma is deferred to the appendix where a constructive procedure for designing (ki1 , ki2 ) given P(λ) is presented. The structure of the proposed observer has the following form ξ˙i = A ξi + N ξi+1 + D2 (ℓ) Ki ei i = 1, . . . , n−2 .. . ξ˙n−1 = A ξn−1 + B ϕs (x̂′ ) + D2 (ℓ) K(n−1) en−1 (8) where (A, B, C) is a triplet in prime form of dimension 2, T ξi ∈ R2 , Ki = (ki1 ki2 ) , D2 (ℓ) = diag(ℓ, ℓ2 ), x̂′ = L1 ξ L1 = blkdiag ( C, . . . , C , I2 ) \| {z } (n−2) times ξ = col(ξ1 , . . . , ξn−1 ) ∈ R2n−2 , e1 = y − Cξ1 , ei = B T ξi−1 − Cξi i = 2, . . . , n − 1 , and ϕs (·) is an appropriate saturated version of ϕ(·). The variable x̂′ represents an asymptotic estimate of the state x of (2). It is obtained by “extracting” n components from the state ξ according to the matrix L1 defined above. As clarified next, the redundancy of the observer can be used to extract from ξ an extra state estimation that is x̂′′ = L2 ξ L2 = blkdiag (I2 , B T , . . . , B T ) . \| {z } (n−2) times The following proposition shows that the observer (8) recovers the same asymptotic properties for the two estimates x̂′ and x̂′′ of the “standard” high-gain observer (4). In the statement of the proposition we let x̂ = col(x̂′ , x̂′′ ) , x = col(x, x) . Proposition 1 Consider system (2) and the observer (8) with the coefficients (ki1 ki2 ) fixed so that the matrix M defined in (7) is Hurwitz (see Lemma 1). Let ϕs (·) be any bounded function that agrees with ϕ(·) on X , and assume that d(x, t) is bounded for all x ∈ X and for all t ≥ 0. Then there exist ci > 0, i = 1, . . . , 4, and ℓ⋆ ≥ 1 such that for any ℓ ≥ ℓ⋆ and for any ξ(0) ∈ R2n−2 , the following bound holds kx̂(t) − x(t)k ≤ max{c1 ℓn−1 exp(−c2 ℓ t) kx̂(0) − x(0)k , 1 c3 k Γ(ℓ) d(·)k∞ , c4 ℓn−1 k ν(·)k∞ } (9) ℓ where Γ(ℓ) is as in (6), for all t ≥ 0 such that x(t) ∈ X . Proof: Consider the change of variables ξi 7→ ξ̃i := ξi − col (xi , xi+1 ) i = 1, . . . , n − 1 , by which system (8) transforms as ξ˜˙1 = H1 ξ˜1 + N ξ˜2 − d¯1 (t) + D2 (ℓ)K1 ν(t) ξ̃˙i ξ˜˙n−1 = Hi ξ̃i + N ξ˜i+1 + D2 (ℓ) Ki B T ξ̃i−1 − d¯i (t) i = 2, . . . , n − 2 = Hn−1 ξ˜n−1 + D2 (ℓ) Kn−1 B T ξ˜n−2 + ˜ x) − d¯n−1 (t) B ∆ϕ(ξ, 3 where d¯i (t) = (di (x(t), t), di+1 (x(t), t))T , Hi = A−D2 (ℓ) Ki C , for i = 1, . . . , n − 1, with di (t) the i-th element of the vector d(t), ξ̃ = col(ξ˜1 , . . . , ξ˜n−1 ), and ˜ x) = ϕs (L1 ξ˜ + x) − ϕ(x) . ∆ϕ(ξ, Rescale now the variables ξ˜i as follows ξ˜i 7→ εi := ℓ2−i D2 (ℓ)−1 ξ̃i i = 1, . . . , n − 1 . By letting ε = col(ε1 . . . εn−1 ), an easy calculation shows that ε̇ = ℓM ε + ℓ−(n−1) (B2n−2 ∆ϕℓ (ε, x) + υℓ (t) + nℓ (t)) , (10) where B2n−2 is the zero column vector of dimension 2n − 2 with a 1 in the last position, and υℓ (t) = −col ℓn D2 (ℓ)−1 d¯1 , . . . , ℓn−i+1 D2 (ℓ)−1 d¯i , . . . , ℓ2 D2 (ℓ)−1 d¯n−1 , nℓ (t) = ℓn K̄1 ν(t) , K̄1 = col(K1 , 0, . . . , 0) , and ∆ϕℓ (ε, x) = ϕs (L1 S(ℓ)ε + x) − ϕ(x) , where S(ℓ) = diag( 1ℓ , 1 , ℓ , . . . , ℓn−3 ) ⊗ D2 (ℓ). Being ϕ(·) uniformly Lipschitz in X and ϕs (·) bounded, there exists a δ1 > 0, δ2 > 0 and δ3 > 0 such that kℓ−(n−1) ∆ϕℓ (ε, x)k ≤ δ1 kεk , kℓ−(n−1) υℓ (t)k ≤ δ2 kℓ Dn (ℓ)−1 d(x(t), t)k , kℓ−(n−1) nℓ (t)k ≤ δ3 kℓ ν(t)k , for all ε ∈ R2n−2 , x ∈ X and ℓ ≥ 1. The rest of the proof follows standard Lyapunov arguments that, for sake of completeness, are briefly recalled. Let P = P T be such that P M + M T P = −I and consider the Lyapunov function V = εT P ε. Taking derivative of V along the solutions of (10), using the previous bounds and letting ℓ⋆ = 4 δ1 kP k, one obtains that there exist positive constants a1 , a2 , a3 such that for any ℓ ≥ ℓ⋆ kεk ≥ max a1 kDn (ℓ)−1 d(x(t), t)k, a2 kν(t)k ⇒ V̇ ≤ −a3 ℓ kεk2 . As P is symmetric and definite positive, it turns out that λkεk2 ≤ V (ε) ≤ λ̄kεk2 where λ and λ̄ are respectively the smallest and the highest eigenvalue of P . By using these bounds, the previous implication leads to conclude that, as long as V ≥ λ̄ max a21 kDn (ℓ)−1 d(x(t), t)k2 , a22 kν(t)k exp(− aλ̄3 ℓ t)V (0). By using again the bound on then V (t) ≤ V in terms of λ and λ̄, the following estimate on ε(t) can be easily obtained kε(t)k ≤ max{ a4 exp(−a5 ℓ t) kε(0)k , a6 kDn (ℓ)−1 d(x(t), t)k , a7 kν(t)k} p p where p a4 = λ̄/λ, a5 = a3 /2λ̄, a6 = λ̄/λ a1 , λ̄/λ a2 . Now, using the fact that, for all ℓ > 1, a7 = ˜ ≤ kεk ≤ kξk, ˜ the previous bound leads to the ℓ−(n−1) kξk ˜ following estimate on ξ(t) ˜ ˜ kξ(t)k ≤ max{ a4 ℓn−1 exp(−a5 ℓ t) kξ(0)k , a6 ℓn−1 kDn (ℓ)−1 d(x(t), t)k , a7 ℓn−1 kν(t)k} by which the claim of the proposition immediately follows by bearing in mind the definition of Γ(ℓ), x̂, x, and by noting ˜ ≤ kx̂ − xk ≤ 2kξk. ˜ that kξk III. A BOUT THE SENSITIVITY OF THE OBSERVER TO HIGH FREQUENCY NOISE IN THE LINEAR CASE The trade-off between the speed of the state estimation and the sensitivity to measurement noise is a well-known fact in the observer theory. In this respect, high-gain observers tuned to obtain fast estimation dynamics are necessarily very sensitive to high-frequency noise. Bounds on the estimation error in presence of measurement noise for the standard high-gain observers have been studied, for instance, in [18] and [5], and different techniques have been developed in order to improve rejection, mainly based on gain adaptation (see, among others, [1], [16]). In this section we compare the properties of the standard high-gain observer (4) and the proposed observer (8) with respect to high-frequency measurement noise by specialising the analysis to linear systems. In particular we consider systems of the form (2) with ϕ(·) a linear function of the form ϕ(x) = Φ x where Φ is a row vector of dimension n. Moreover, in this contest, we consider d(t) ≡ 0 and ν(t) = aN sin(ωN t + fN ) (11) where aN > 0, ωN > 0, and fN are constants. It is shown that the ratio between the asymptotic estimation error on the i-th state variable provided by the new observer (8) and the one provided by the standard observer (4) is a strictly decreasing polynomial function of the noise frequency for i = 2, . . . , n. In this regard the new observer has better asymptotic properties with respect to high-frequency noise as far as the state estimation variables are concerned (except for the first one). This is formalised in the next proposition. Proposition 2 Let n ≥ 1. Let Kn ∈ Rn , Ki ∈ R2 , i = 1, . . . , n − 1 and ℓ ∈ R be fixed so that the error dynamics of the observers (4) and (8) are Hurwitz. Moreover, with ρ ∈ {1, 2, . . . , n} denoting the position index of the first non zero coefficient of the vector Φ (and ρ = n if Φ is the zero vector), let ri′ ≥ 1 be the constants defined as ri′ = min { i , (n − 1) , (ρ + n − i + 1)} i = 1, . . . , n . ⋆ ⋆ There exist ωN > 0 and c̄i > 0 such that for all ωN > ωN , aN > 0 and fN we have lim sup \|x̂′i (t) − xi (t)\| t→+∞ lim sup \|x̂i (t) − xi (t)\| t→+∞ −(ri′ −1) ≤ c̄i ωN ∀ i = 1, . . . , n. 4 Proof: Consider system (2) and the standard high-gain observer (4). By letting e = ℓ Dn (ℓ)−1 (x̂−x), the e-dynamics read as (12) ė = ℓ An + Kn Cn e + Bn Φ Θn (ℓ) e + ℓKn ν(t) 1 where Θn (ℓ) = n Dn (ℓ) with Dn (ℓ) defined in (4). It is a ℓ linear system that is Hurwitz by the choices of Kn and ℓ. Similarly, consider system (2) and the new observer (8). With ε defined as in the proof of Proposition 1 we obtain system (10) compactly rewritten as ε̇ = ℓM ε + B2n−2 Φ Θn (ℓ)L1 ε + ℓK̄1 ν(t) . (13) It is an Hurwitz system by the choices of Ki , i = 1, . . . , n− 1, and ℓ. We consider now the n systems given by the dynamics (12) with input ν and with outputs x̂i − xi = ℓi−1 ei , and we denote by Fi′ : R → C, i = 1, . . . , n, the harmonic transfer functions of these systems. Simple computations show that these systems have relative degree ri′ , for all i = 1, . . . , n, with ri′ defined in the statement of the proposition. By definition of harmonic transfer function and by the fact that systems (12) and (13) are Hurwitz, it turns out that = \|Fi (ωN )\| aN  −z Υ(z[0,3] ) = −ż −z̈ t→+∞ = \|Fi′ (ωN )\| aN for any ωN ≥ 0 and ∀ i = 1, . . . , n. Furthermore, by the fact that Fi (ω) and Fi′ (ω) have, respectively, relative degrees ri = 1 and ri′ , i = . . . , n, it turns out that there exist positive ⋆ > 0 such that ci , c′i and ωN \|Fi (ω)\| \|Fi′ (ω)\| ≥ ci ω −1 ′ ≤ c′i ω −ri ⋆ ∀ ω ≥ ωN , by which the result immediately follows. IV. E XAMPLE : O BSERVER FOR THE UNCERTAIN VAN D ER P OL O SCILLATOR Let consider the uncertain Van der Pol oscillator z̈ = −α2 z + β(1 − z 2 )ż , y = z + ν(t) ,  (1 − z 2 )ż  z̈ − 2y ż 2 − z 2 z̈ (3) z − 2ż 3 − 6z żz̈ − z 2 z (3) and ρ(z[0,4] ) = z (3) , z (4) (1 − z 2 ) − 12ż 2 z̈ − 6z z̈ 2 − 8z żz (3) with (see [9]) (14) where y ∈ R is the measured output and α, β are uncertain constant parameters. We let µ = (α2 , β)⊤ and we assume that µ ∈ U, with U a compact set of R2 not containing the origin. The state (z, ż) belongs to a compact invariant set W ⊂ R2 , which is the limit cycle of the Van Der Pol oscillator. We observe that W depends on µ. Following [3], system (14) extended with µ̇ = 0 can be immersed into a system in the canonical observability form (2) with n = 5. As a matter of for all z[0,4] ∈ X . Hence, by letting ϕ(z[0,4] ) = ρ(z[0,4] ) µ̂(z[0,3] ) with µ̂(z[0,4] ) = Υ† (z[0,4] ) z[2,4] , where Υ† (·) is the left-inverse of Υ(·), and letting x = z[0,4] , it is immediately seen that system (14) and µ̇ = 0 restricted to W × U is immersed into the system ẋ = A5 x + B5 ϕ(x) , y = C5 x + ν(t) , (15) where A5 , B5 , C5 is a triplet in prime form of dimension 5. By following the prescriptions of Section II, we implemented the proposed observer (8) as ξ̇1 = .. . A ξ1 + N ξ2 + D2 (ℓ1 )K1 (y − Cξ1 ) ξ̇4 = A ξ4 + B ϕs (x̂′ ) + D2 (ℓ1 )K4 (B ⊤ ξ3 − Cξ4 ) t→+∞ lim sup \|x̂′i (t) − xi (t)\| z (5) = ρ(z[0,4] ) µ where rankΥ(z[0,3] ) = 2 = ℓi−1 εi,1 i = 1, . . . n − 1, = ℓn−1 εn−1,2 , lim sup \|x̂i (t) − xi (t)\| z[2,4] = Υ(z[0,3] ) µ , i = 1, . . . n , and we denote by Fi : R → C, i = 1, . . . , n, the harmonic transfer functions of these systems. A simple computation shows that these systems have relative degree ri = 1, for all i = 1, . . . , n. Similarly, we consider the n systems given by the dynamics (13) with input ν and outputs x̂′i − xi x̂′n − xn fact, let z[i,j] := col (z (i) , . . . , z (j) ) be the vector of time derivatives of z, with 0 ≤ i < j, and let X be the compact set of R5 such that z[0,1] ∈ W ⇒ z[0,4] ∈ X . Simple computations show that (16) where ϕs (·) is any locally Lipschitz bounded function that agrees with ϕ(·) on X , ξi = (ξi1 , ξi2 ) and the coefficients of Ki are k11 = 0.6, k12 = 0.3, k21 = 0.6, k22 = 0.111, k31 = 0.6, k32 = 0.0485, k41 = 0.6, k42 = 0.0178, such that the roots of P(λ) are −0.1, −0.2, −0.2, −0.3, −0.3, −0.4, −0.4, −0.5. With the same notation of (8) we have ξ = (ξ1 , ξ2 , ξ3 , ξ4 ), x̂′ = L1 ξ and x̂′′ = L2 ξ. In the simulations we fixed α = 1, β = 0.5, gain ℓ1 = 100 and initial conditions (z, ż) = (1, 0) for (14) and ξ = 0 for (16). Figures 1 and 2 show the error state estimate \|x̂′ − x\| and \|x̂′′ − x\| of the proposed observer (16) for the first two components (namely the estimation of the state of (14)) when there is not sensor noise. By following Section III we compared the observer (16) with a standard high-gain observer in presence of high-frequency sensor noise, numerically taken as ν(t) = 10−2 sin(103 t). The high-gain observer has been implemented as x̂˙ = A5 x̂ + B5 ϕs (x̂) + D(ℓ2 )K5 (y − C5 x̂) (17) where K5 = (1.5, 0.85, 0.225, 0.0274, 0.0012)T so that the eigenvalues of (A5 + K5 C5 ) are −0.1, −0.2, −0.3, 5 −0.4, −0.5, and ℓ2 = 100. Table 1 shows the normalized asymptotic error magnitudes of the proposed observer (16) and the standard high-gain observer (17), where the normalized asymptotic error for the i-th estimate is defined as kx̂i −xi ka = lim sup kx̂i (t)−xi (t)k/aN . Although the result t→∞ in Section III is given just for linear systems, the numerical results shown in the table show a remarkable improvement of the sensitivity to high-frequency measurement noise of the new observer with respect to the standard high-gain observer. Standard High Gain Observer x̂ kx̂1 − x1 ka kx̂2 − x2 ka kx̂3 − x3 ka kx̂4 − x4 ka kx̂5 − x5 ka = 0.15 =8 = 2 · 102 = 2.5 · 103 = 104 Modified Observer x̂′ = L1 ξ Modified Observer x̂′′ = L2 ξ kx̂′1 kx̂′2 kx̂′3 kx̂′4 kx̂′5 kx̂′′ 1 − x1 ka kx̂′′ 2 − x1 ka ′′ kx̂3 − x3 ka kx̂′′ 4 − x4 ka kx̂′′ 5 − x5 ka − x1 ka − x2 ka − x3 ka − x4 ka − x5 ka = = = = = 0.06 0.2 0.2 0.1 0.3 = = = = = 0.06 3 3 2 0.3 Table 1: Normalized asymptotic errors in presence of noise. numerically simulated in the previous section. A complete characterisation of the sensitivity to sensor noise of the new observer is an interesting research topic are that is now under investigation. The peaking phenomenon due to wrong initial conditions and fast convergence that is typical of high-gain observers is not prevented by our proposed structure. However, other techniques to deal with peaking (such as saturations, timevarying gains [16], gradients techniques [2], and others) are available and can be adopted to improve the proposed observer structure. In this work we didn’t consider the multioutput case. For the specific class of multi-output systems which are diffeomorphic to a block triangular form in which each block is associated to each output and it has a triangular dependence on the states of that subsystem (see [4]), the proposed structure can be simply applied block-wise to obtain a high-gain observer. Apart this case, a complete extension to the multi-output case is not immediate and under investigation. Acknowledgment. We wish to thank Laurent Praly for suggesting the design procedure presented in Appendix A. 0.5 0 A PPENDIX A. Procedure to assign the eigenvalues of M −0.5 −1 0 0.2 0.4 0.6 0.8 Consider the matrices Mi ∈ R2i−2 × R2i−2 recursively defined as Mi−1 N̄i , i = 2, . . . , n−1 M1 = E1 , Mi = Q̄i Ei 1 t (s) Fig. 1: Error state estimate \|x̂′′1 − x1 \|, \|x̂′1 − x1 \|. where N̄i = col(02(i−2)×2 , N ), Q̄i = (02(i−2)×2 , Qi ) and Ei , i = 1, . . . , n − 1, Qi , i = 2, . . . , n − 1, and N are defined as in the definition of M . Note that M = Mn−1 and, by letting Ki = (ki1 ki2 )T , note that Qi and Ei depend on Ki , while Mi depends on K1 , . . . , Ki . We let PMi (λ) = λ2i + mi1 λ2i−1 + 2i−3 + . . . + mi2i−1 λ + mi2i and PMi−1 (λ) = λ2i−2 + mi−1 1 λ i−1 i−1 . . .+m2i−3 λ+m2i−2 the characteristic polynomials of Mi and Mi−1 , and we use the notation mi[1,j] = col(mi1 , . . . , mij ) ∈ i−1 i−1 k Rj , mi−1 [1,k] = col(m1 , . . . , mk ) ∈ R for some j ≤ 2i and k ≤ 2i − 2. The characteristic polynomial of PMi (λ) is computed as 30 20 10 0 −10 0 0.2 0.4 0.6 0.8 1 PMi (λ) t(s) Fig. 2: Error state estimate (red line). \|x̂′′2 − x2 \| (blue line) and \|x̂′2 − x2 \| V. C ONCLUSIONS We presented a new observer design based on high-gain techniques with a tunable state-estimate convergence speed. With respect to standard high-gain observers the state dimension is larger (2n − 2 instead of n) with a clear benefit in the observer implementation due to the power of the highgain which is only 2 and not n. Moreover, when specialised to linear systems, we showed the benefit of the proposed observer with respect to the standard high-gain observer in terms of high-frequency noise rejection. Benefits that are clearly confirmed also for the nonlinear Van-der Pol example = λ(λ + ki1 )PMi−1 (λ) +ki2 PMi−1 (λ) − λ(λ + k(i−1)1 )PMi−2 (λ) . Hence, simple, although lengthy, computations show that the coefficients mi[1,2i] of PMi (λ) and mi−1 [1,2i−2] of PMi−1 (λ) are related as follow mi[1,2i−2] = (I2i−2 + ki1 F )mi−1 [1,2i−2] + ki1 v1 mi2i−1 = ki1 mi−1 2i−2 mi2i ki2 mi−1 2i−2 = 2i−2 (18) where v1 ∈ R is the zero vector with a 1 in the first position, and F ∈ R(2i−2)×(2i−2) is the zero matrix with the identity matrix I2i−3 in the lower left block. Note that (I2i−2 + ki1 F ) is invertible for all ki1 . Hence, from the first of (18), one obtains i mi−1 [1,2i−2] = Λ(m[1,2i−2] , ki1 ) 6 where Λ(mi[1,2i−2] , −1 ki1 ) = (I2i−2 + ki1 F ) (mi[1,2i−2] − ki1 v1 ) , which, embedded in the second and in the third of (18), yield the relations σ1 (mi[1,2i−1] , ki1 ) = 0 , ki2 = σ2 (mi[1,2i] , ki1 ) where σ1 (mi[1,2i−1] , ki1 ) = ki1 v2T Λ(mi[1,2i−2] , ki1 ) − mi2i−1 σ2 (mi[1,2i] , ki1 ) = v2T mi2i i Λ(m[1,2i−2] , ki1 ) in which v2 ∈ R2i−2 is the zero vector with a 1 in the last position. We observe that σ1 (·, ·) is a polynomial in ki1 of odd order 2i − 1. As a consequence, for any mi[1,2i−1] there always exists at least one real ki1 fulfilling σ1 (mi[1,2i−1] , ki1 ) = 0. The previous results can be used to set up a ”basic assignment algorithm” that is then used iteratively to solve the eigenvalues assignment of the matrix M . Basic assignment algorithm. Let P̄i (λ) = λ2i + m̄i1 λ2i−1 + . . . + m̄i2i be an arbitrary polynomial. Then, there exist a real K̄i = (k̄i1 , k̄i2 )T and a polynomial P̄i−1 (λ) = λ2i−1 + 2i−2 m̄i−1 + . . . + m̄i−1 1 λ 2i−2 such that Ki PMi−1 = K̄i = P̄i−1 (λ) ⇒ PMi (λ) = P̄i (λ) . As a matter of fact, by letting m̄i[1,2i−1] the coefficients of P̄i (λ), it is possible to take k̄i1 as a real solution of σ1 (m̄i[1,2i−1] , ki1 ) = 0, k̄i2 = σ2 (m̄i[1,2i] , k̄i1 ), and to take the coefficients m̄i−1 [1,2i−2] of the polynomial P̄i−1 (λ) as i m̄i−1 = Λ( m̄ [1,2i−2] , k̄i1 ) . ⊳ [1,2i−2] With the previous algorithm in hand, the design of K1 , . . . , Kn−1 to assign an arbitrary characteristic polynomial to M , can be then immediately done by the following steps: 1) With P̄n−1 (λ) the desired characteristic polynomial of M , compute (K̄n−1 , P̄n−2 (λ)) by running the basic assignment algorithm with i = n − 1. 2) Compute iteratively (K̄i , P̄i−1 (λ)) by running the basic assignment algorithm for i = n − 2, . . . , 2. 3) Compute K̄1 = (k̄i1 , k̄i2 )T so that λ2 + ki1 λ + ki2 = P̄1 (λ). R EFERENCES [1] J. H. Ahrens and H. K. Khalil, “High-gain observer in the presence of measurement noise: a switched-gain approach”, Automatica vol. 45, pp 936-943, 2009. [2] D. Astolfi and L. Praly, “Output feedback stabilization for SISO nonlinear systems with an observer in the original coordinates”, IEEE 52nd Conference on Decision and Control, pp. 5927-5932, December 2013. [3] D. Astolfi, L. Praly and L. Marconi, “A note on observability canonical forms for nonlinear Systems”, IFAC, 9th Symposium on Nonlinear Control Systems, pp 436-438, September 2013. [4] A. N. Atassi and H. K. Khalil, “A separation principle for the stabilization of a class of nonlinear systems”, IEEE Transactions on Automatic Control, 44, pp 1672-1687, 1999. [5] A. A. Ball and H. K. Khalil, “High-gain observer tracking performance in the presence of measurement noise”, American Control Conference, pp 4626-4627, 2009. [6] G. Besançon, Nonlinear observers and applications, Springer, 2007. [7] G. Bornard and H. Hammouri, “A high gain observer for a class of uniformly observable systems”, IEEE, 30th Conference on Decision and Control, December 1991. [8] F. Deza, E. Busvelle, J.P. Gauthier, D. Rakotopara, “High gain estimation for nonlinear systems”, Systems & Control Letters vol. 18, pp 295-299, April 1992. [9] F. Forte, A. Isidori, L. Marconi, “Robust design of internal models by nonlinear regression”, IFAC, 9th Symposium on Nonlinear Control Systems, pp 283-288, September 2013. [10] M. T. J. Gajić, The Lyapunov matrix equation in system stability and control, Academic Press, 1995. [11] J. P. Gauthier, G. Bornard, “Observability for any u(t) of a class of nonlinear systems”, IEEE Transactions on Automatic Control, Vol. AC26, No. 4, August 1981. [12] J.P. Gauthier and I. Kupka, “Deterministic observation theory and applications”, Cambridge University Press, 2004. [13] H Hammouri, G. Bornard and K. Busawon, “High gain observer for structured multi-input multi-output nonlinear systems”, IEEE Transaction on Automatic Control vol. 55, no. 4, pp 987-992, April 2010. [14] H. K. Khalil and L. Praly, “High-gain observers in nonlinear feedback control”, Robust and Nonlinear Control, 24, pp 993-1015, 2014. [15] M. Maggiore, K.M. Passino, “A separation principle for a class of non uniformly completely observable systems”, IEEE Transactions on Automatic Control, vol. 48, no. 7, pp 1122-1133, July 2003. [16] R. G. Sanfelice and L. Praly, “On the performance of high-gain observers with gain adaptation under measurement noise”, Automatica Vol. 47, pp 2165-2176, 2011. [17] A. Teel and L. Praly, “Global stabilizability and observability imply semi-global stabilizability by output feedback”, Systems and Control Letters, 22, pp 313-325, 1994. [18] L. K. Vasiljevic and H. K. Khalil, “Differentiation with high-gain observers the presence of measurement noise”, IEEE 45th Conference on Decision and Control, pp 4717-4722, December 2006.	3cs.SY
arXiv:1711.06855v1 [cs.SY] 18 Nov 2017 A PROBABILISTIC CHARACTERIZATION OF RANDOM AND MALICIOUS COMMUNICATION FAILURES IN MULTI-HOP NETWORKED CONTROL A. CETINKAYA, H. ISHII, AND T. HAYAKAWA Abstract. The control problem of a linear discrete-time dynamical system over a multi-hop network is explored. The network is assumed to be subject to packet drops by malicious and nonmalicious nodes as well as random and malicious data corruption issues. We utilize asymptotic tail-probability bounds of transmission failure ratios to characterize the links and paths of a network as well as the network itself. This probabilistic characterization allows us to take into account multiple failures that depend on each other, and coordinated malicious attacks on the network. We obtain a sufficient condition for the stability of the networked control system by utilizing our probabilistic approach. We then demonstrate the efficacy of our results in different scenarios concerning transmission failures on a multi-hop network. 1. Introduction Network control systems incorporate communication networks to facilitate the exchange of measurement and control data between the plant and the controller [1]. Recently, multi-hop networks have been utilized in networked control operations [2, 3, 4, 5, 6, 7]. A multi-hop network such as a wireless ad hoc network consists of a number of nodes that are connected with a number of communication links. Ensuring orderly operation of a multi-hop networked control system can be challenging, as packets may sometimes fail to be transmitted at different parts of the network due to various reasons. One of the reasons for transmission failures in a multi-hop network is channel noise in the communication links, which may corrupt the contents of state and control input data packets. The occurrence of data corruption in a communication network may be modeled using random processes [8]. In addition to channel noise, network congestion may also cause packet transmission failures. Routers may drop some packets to mitigate congestion [9]. Furthermore, it has become apparent that malicious attacks may also hamper transmissions in a networked control system. For instance, jamming attacks by (A. Cetinkaya, H. Ishii) Department of Computer Science, Tokyo Institute of Technology, Yokohama, 226-8502, Japan. (T. Hayakawa) Department of Systems and Control Engineering, Tokyo Institute of Technology, Tokyo 152-8552, Japan. E-mail addresses: ahmet@sc.dis.titech.ac.jp, ishii@c.titech.ac.jp, hayakawa@sc.e.titech.ac.jp . Key words and phrases. Networked Control, Multi-hop Networks, Malicious Attacks, Random Failures. This work was supported in part by the JST CREST Grant No. JPMJCR15K3 and by JSPS under Grant-in-Aid for Scientific Research Grant No. 15H04020. 1 2 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL an adversary may interfere with the communication on links and effectively prevent transmission of packets. This issue was investigated in several works from the viewpoints of wireless communications [10, 11], as well as control and game theory [12, 13, 14, 15, 16]. Transmissions of state and control input information between the plant and the controller may also fail due to malicious activities of routers. Malicious routers may intentionally drop some of the packets coming from and/or headed to certain nodes of the network [6, 17, 18]. The detection of such routing attacks can be challenging especially when the malicious nodes act normal for certain periods of time (see grayhole attacks in [19]). Furthermore, networks may face both malicious routing and random packet losses due to link errors [20, 21]. Understanding the effects of malicious attacks on networks is important from the viewpoint of cyber security of networked control systems [22, 23]. Our goal in this paper is to explore the effects of random and malicious transmission failures in a general multi-hop communication network and develop a network characterization to be used in the analysis of networked control systems. The key problem here is to characterize the failures for the overall multi-hop network in a nonconservative way while still taking into account mutually-dependent packet failures and coordinated malicious attacks on the network. In the literature, researchers have proposed different characterizations of packet failures in a networked control system. Specifically, [2] explored control over a network with multiple links that introduce random packet drops. The results obtained in [2] utilize packet drop probabilities on the edges that constitute cut-sets of the network graph. This approach is also utilized in network characterization by [24]. Furthermore, [3] discussed almost sure networked control system stability, and [25] studied networked state estimation problem. Recently, [7] investigated mean square stability and robustness under delays and packet losses through a Markov jump linear system framework. In these studies, Markov processes with certain transition probabilities are used for modeling packet losses on a network. In this paper, we consider a network with not only random failures but also malicious activities on nodes or links. As a result, failures may not always be modeled by Markov processes and packet losses may not always have well-defined probabilities. In [6], researchers explored a multi-hop network model with malicious nodes. There, the fault detection and isolation problem is explored for the case where the nodes induce delay in transmissions. Here, we consider a different setup and a different modeling approach. In particular, we do not investigate a detection problem, and we do not model the effect of delays. In this paper, we would like to characterize the overall packet exchange failures on a network between the plant and the controller by using the properties of the paths of the network and the communication links on those paths. We also note that our network characterization provides a high level model and it is tailored to be utilized for stability analysis in networked control as in [2]. Instead of specifying underlying physical channel models and routing protocols, we take a probabilistic approach to characterize transmission failure ratios on the links, paths of a network, as well as the network itself. Our approach for modeling the overall packet failures in a network is built upon tail-probability bounds for the binary-valued processes that describe the occurrences of failures on the network. Specifically, each link on a multi-hop network is described through an asymptotic tail-probability bound of the transmission failure CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 3 ratio of that link. This tail-probability-based approach is different from the typical random packet loss modeling approach of assigning probabilities to failures. Our approach can capture failures that occur due to both malicious and nonmalicious reasons. In fact, we utilized the tail-probability-based approach in [26] to study the combined effects of malicious-attacks following the discrete-time version of the attack model in [13, 16] and random packet losses modeled as Markov chains. Differently from [26], we show in this paper that when tail-probability bounds for the links are available, then we can obtain tail-probability bounds describing the overall failures on individual paths of a network. Then those bounds are used for deriving tail-probability bounds describing the overall failures on the network itself. Using our proposed characterizations, we obtain a probabilistic upper-bound for the average number of packet exchange failures between the plant and the controller, which we use in almost sure stability analysis of a discrete-time linear networked control system. In the multi-hop setting, the location of failures and whether multiple failures depend on each other or not critically affect the quality of communication and hence the stabilization performance of the controller. Especially, the situation becomes more serious when the network is targeted by a number of adversaries that launch coordinated attacks on different locations in the communication network. Our tail-probability bound approach can handle such worst-case situations in a unified manner. In addition to our investigation on those situations, we also explore the case where one or more paths/links are known to be associated with random failures and the corresponding indicator processes are mutually independent. For such cases, we show that tighter results can be obtained. This is done by exploiting certain properties of the hidden Markov models that characterize the random failures. The organization of the rest of the paper is as follows. In Section 2, we explain the networked control problem over a multi-hop network. We then present our multihop network characterization in Sections 3 and 4. Specifically, in Section 3, we give a characterization of the overall transmission failures in a multi-hop network based on the failures on individual paths of that network. Furthermore, in Section 4, we investigate failures on the paths that occur due to data-corruption and packetdropping issues at nodes and communication links. We demonstrate the utility of our results in Section 5, and conclude the paper in Section 6. We note that the conference version of this paper appeared in [27]. In this paper, we provide more detailed discussions throughout the paper. Furthermore, we present new results in Sections 3, 4 and new examples in Section 5. In this paper, we use N0 and N to denote nonnegative and positive integers, respectively. The notation P[·] denotes the probability on a probability space (Ω, F , P) with filtration {Ft }t∈N0 . For binary numbers, the notation ∨ represents the oroperation; moreover, ∧ represents the and-operation. 2. Control over Multi-hop Networks In this section, we investigate the control of a linear plant over a multi-hop network. On this network, the plant and the controller exchange state measurement and control input packets. The transmissions in the network are not subject to delay; however, there may be failures in packet exchange attempts between the plant and the controller. 4 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL We describe the linear dynamics of the plant by (1) x(t + 1) = Ax(t) + Bu(t), x(0) = x0 , t ∈ N0 , where A ∈ Rn×n and B n×m respectively denote the state and input matrices; furthermore, x(t) ∈ Rn and u(t) ∈ Rm are the state and the control input vectors, respectively. The plant and the controller attempt to exchange state measurement and control input packets at each time instant t. We denote success or failure states of packet exchange attempts by using the binary-valued process {l(t) ∈ {0, 1}}t∈N0 . In the case of a successful packet exchange (l(t) = 0), the plant transmits the state measurement to the controller; the controller uses the received state information together with a linear static feedback control law to compute the control input, which is then sent back to the plant. The transmitted control input is applied at the plant side. At certain time instants, either the state packet or the control input packet cannot be delivered due to network issues such as packet drops, jamming attacks, and other communication errors. In such cases packet exchange attempts fail (l(t) = 1), and the control input at the plant side is set to 0, which is one of the common approaches in the literature (see [1] and the references therein). Under this characterization, the control input u(t) applied at the plant side is given by u(t) , (1 − l(t)) Kx(t), (2) t ∈ N0 , m×n where K ∈ R represents the feedback gain. Although we consider a static state-feedback control setup here, the techniques that we develop in this paper can also be used in conjunction with other control approaches. In particular, the predictive control approach of [28] and the event-triggered output-feedback control approach from our earlier work [29] can be studied within the context of multi-hop networked control by using the network characterizations that we develop here. We now introduce a class of processes that is useful in describing packet failure indicators in the paper. Definition 2.1 (Λρ ). Given a scalar ρ ∈ [0, 1] we define the class of binary-valued processes Λρ by ∞ k−1 X X Λρ , l : l(t) ∈ {0, 1}, t ∈ N0 ; P[ l(t) > ρk] < ∞ . k=1 t=0 By this definition, we have Λρ1 ⊆ Λρ2 for ρ1 ≤ ρ2 . Furthermore, any binaryvalued process l satisfies l ∈ Λ1 . In [26], we explored a problem similar to the one that we discuss in this paper. There, we considered a single direct communication channel between the plant and the controller. To describe the packet losses on this channel, we proposed a probabilistic characterization that is based on the following assumption. Assumption 2.2. For the packet exchange failure indicator l(·), we have l ∈ Λρ with ρ ∈ [0, 1]. This assumption allows us to characterize a range of scenarios in a unified manner through the scalar ρ ∈ [0, 1]. For instance, the case where all packet exchange attempts fail (l(t) = 1, t ∈ N0 ) can be described by setting ρ = 1. Moreover, in CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 5 the case where all packet exchange attempts are successful (l(t) = 0, t ∈ N0 ), then l ∈ Λ0 . In addition to these two extreme cases, as we illustrate throughout the paper, Assumption 2.2 can also be used to describe random failures and malicious attacks. When the packet exchange failures in a networked control system satisfy Assumption 2.2, then the scalar ρ ∈ [0, 1] also represents a bound on the asymptotic packet exchange failure ratio. Specifically, it follows from Borel-Cantelli Lemma Pk−1 l(t) ≤ ρ, almost surely. When this inequality holds with that lim supk→∞ k1 t=0 a small ρ, it means that the packet exchanges fail statistically rarely. We showed in [26] that the plant (1) can be stabilized over a network, if ρ is sufficiently small. The stability analysis method developed in [26] allows us to obtain the following result, which presents sufficient stability conditions for the closed-loop networked control system (1), (2). Theorem 2.3. Consider the dynamical system (1), (2). Suppose Assumption 2.2 holds with scalar ρ ∈ [0, 1]. If there exist a positive-definite matrix P ∈ Rn×n and scalars β ∈ (0, 1), ϕ ∈ [1, ∞) such that T (3) (A + BK) P (A + BK) − βP ≤ 0, (4) AT P A − ϕP ≤ 0, (5) (1 − ρ) ln β + ρ ln ϕ < 0, then the zero solution x(t) ≡ 0 of the closed-loop system (1), (2) is asymptotically stable almost surely. In Theorem 2.3, the conditions (3) and (4) characterize stability and instability of the closed-loop and the open-loop dynamics through scalars β and ϕ. These scalars also appear in condition (5). When ρ (and hence the packet exchange failure ratio) is sufficiently small so that (5) is satisfied, then we have almost sure asymptotic stability, which implies P[limt→∞ kx(t)k = 0] = 1. In Sections 3 and 4 below, we will present some key methods for obtaining ρ of a multi-hop networked control system (1), (2) to facilitate its stability analysis with Theorem 2.3. Specifically, we will consider the setting where the state measurement and control input packets are attempted to be transmitted over multi-hop networks instead of a single direct channel, which was considered in [26]. We will use assumptions similar to Assumption 2.2 to characterize packet transmission failures on the paths between the plant and the controller as well as the individual links on those paths. We will then show that the packet exchange failures (represented by l(·)) in the overall networked control system satisfy Assumption 2.2 with a scalar ρ ∈ [0, 1] that depends on the network structure as well as failure models of links. To facilitate the analysis in the following sections, we now define two more classes of binary-valued processes that are distinct from Λρ . Our goal is to characterize more specific models for random and malicious failures. In our earlier work [26], we utilized time-inhomogeneous Markov chains for characterizing random failures in a single communication channel. When we consider a multi-hop network composed of a number of channels that face random transmission failures, we are required to introduce a different characterization. This is because the overall failures in the network depends on the failures on each individual channel. Hence, the overall failures cannot always be described as a Markov 6 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL chain. In order to describe random failures in multi-hop networks, we utilize timeinhomogeneous hidden Markov models (see [30, 31, 32]). The binary-valued process {l(t) ∈ {0, 1}}t∈N0 has a time-inhomogeneous hidden Markov model if l(t) = hl (θl (t)), (6) l t ∈ N0 , l where h : Θ → {0, 1} is a binary valued function on a set Θl of finite number of elements, and moreover, {θl (t) ∈ Θl }t∈N0 is an Ft -adapted, finite-state, and timeinhomogeneous Markov chain with initial distributions ϑlq ∈ [0, 1], q ∈ Θl , and time-varying transition probability functions plq,r : N0 → [0, 1], q, r ∈ Θl , satisfying P[θl (0) = q] = ϑlq , P[θ (t + 1) = r\|θl (t) = q] = plq,r (t), l t ∈ N0 . The process {l(t) ∈ {0, 1}}t∈N0 is also called the output process of a hidden Markov model. Notice that {l(t) ∈ {0, 1}}t∈N0 depends on the Markov chain {θl (t) ∈ Θl }t∈N0 through function hl , but it is not necessarily a Markov chain itself. Specifically, in certain cases, we may have P[l(t + 1) = 1\|Ft ] 6= P[l(t + 1) = 1\|l(t)], which shows that Markov property (see [33]) does not hold. This is for example the case where l(·) represents the failures on a Gilbert-Elliott channel (see [34, 35]). We also note that hidden Markov models naturally arise in the description of multi-hop networks. For instance, l(·) may be the failure indicator of a path with multiple links. Even if the failures on individual links may satisfy the Markov property, the overall failure indicator l(·) does not satisfy it due to dependence on the failure/success states of all individual links. In such cases, l(·) follows a hidden Markov model, where the Markov chain {θl (t) ∈ Θl }t∈N0 represents the combined states of all individual links. For a given binary-valued output process {l(t) ∈ {0, 1}}t∈N0 associated with a time-inhomogeneous hidden Markov model, let Θl0 , Θl1 ⊂ Θl be given by Θl0 , {r ∈ Θl : hl (r) = 0}, Θl1 , {r ∈ Θl : hl (r) = 1}. In the next definition we introduce a class of binary-valued output processes associated with time-inhomogeneous hidden Markov models. Definition 2.4 (Γp0 ,p1 ). Given scalars p0 , p1 ∈ [0, 1] we define the class Γp0 ,p1 of binary-valued output processes of time-inhomogeneous hidden Markov models by X X l plq,r (t) ≤ p1 , q ∈ Θl , t ∈ N0 . pq,r (t) ≤ p0 , q ∈ Θl , t ∈ N0 ; Γp0 ,p1 , l : r∈Θl0 r∈Θl1 The following result establishes the relation between the classes Γp0 ,p1 and Λρ . Proposition 2.5. Consider the binary-valued output process {l(t) ∈ {0, 1}}t∈N0 of a time-inhomogeneous hidden Markov model. If l ∈ Γp0 ,p1 with p1 ∈ (0, 1), then we have l ∈ Λρ for any ρ ∈ (p1 , 1]. Before, we give the proof of Proposition 2.5, we first present a technical result that provides upper bounds on the tail probabilities of sums involving a binaryvalued output process associated with a time-inhomogeneous hidden Markov model. Lemma 2.6. Let {ξ(t) ∈ Ξ}t∈N0 be a finite-state time-inhomogeneous Markov chain with transition probabilities pq,r : N0 → [0, 1], q, r ∈ Ξ, and let Ξ1 ⊂ Ξ be given by Ξ1 , {r ∈ Ξ : h(r) = 1}, where h : Ξ → {0, 1} is a binary-valued function. CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 7 Furthermore, let {χ(t) ∈ {0, 1}}t∈N0 be a binary-valued process that is independent of {ξ(t) ∈ Ξ}t∈N0 . Assume X (7) pq,r (t) ≤ p̃, q ∈ Ξ, t ∈ N0 , r∈Ξ1 ∞ X (8) P[ k−1 X χ(t) > w̃k] < ∞, t=0 k=1 where p̃ ∈ (0, 1), w̃ ∈ (0, 1]. We then have for ρ ∈ (p̃w̃, w̃), (9) P[ k−1 X h(ξ(t))χ(t) > ρk] ≤ ψk , k ∈ N, t=0 w̃k p̃+1) where ψk , σ̃k + φ−ρk+1 ((φ−1) (φ−1)p̃ P∞ N. Moreover, k=1 ψk < ∞. −1 ,φ, ρ w̃ (1−p̃) ρ , p̃(1− w̃ ) σ̃k , P[ Pk−1 t=0 χ(t) > w̃k], k ∈ Lemma 2.6 is an essential tool for dealing with different failure scenarios specific to multi-hop networks, and it generalizes Lemma A.1 of our previous work [26]. In particular, in the case where Ξ = {0, 1} and h(r) = r, r ∈ Ξ, Lemma 2.6 recovers Lemma A.1 of [26]. The proof Lemma 2.6 is given in the Appendix. We are now ready to prove Proposition 2.5. Proof of Proposition 2.5. Notice that l ∈ Λρ for ρ = 1, since l(·) is binary-valued. For the case ρ ∈ (p1 , 1), we show l ∈ Λρ by employing Lemma 2.6. Specifically, let p̃ = p1 , w̃ = 1, and define the processes {ξ(t) ∈ {0, 1}}t∈N0 and {χ(t) ∈ {0, 1}}t∈N0 with ξ(t) = θ1l (t) and χ(t) = 1, t ∈ N0 . Since the conditions in (7) and (8) are satisfied with p̃ = p1 , h = hl , Ξ = Θl , and Ξ1 = Θl1 , it follows from Lemma 2.6 that ∞ X k=1 P[ k−1 X t=0 l(t) > ρk] = ∞ X P[ k=1 which completes the proof. k−1 X hl (θ1l (t)) > ρk] = t=0 ∞ X k=1 P[ k−1 X h(ξ(t))χ(t) > ρk] < ∞, t=0 Next, we introduce a class for binary-valued processes that we employ in characterizing the timing of malicious attacks. Definition 2.7 (Πκ,w ). Given scalars κ ≥ 0, w ∈ [0, 1] we define the class of binary-valued processes Πκ,w by X k−1 Πκ,w , l : l(t) ∈ {0, 1}, t ∈ N0 ; P l(t) ≤ κ + wk = 1, k ∈ N . t=0 The characterization for the class Πκ,w is based on a discrete-time version of the malicious attack model used by [13, 16]. In this model, occurrences of malicious Pk−1 attacks is described by a process l(·) such that P t=0 l(t) ≤ κ + wk = 1, k ∈ N, where κ ≥ 0 represents an upper-bound for a number of initial attacks, and w ∈ (0, 1) represents a bound on the average attack rate. Lemma 2.3 in [26] shows that if l ∈ Πκ,w with w ∈ (0, 1), then l ∈ Λρ for any ρ ∈ (w, 1]. Note that the malicious attack characterization through the class Πκ,w , does not require the process l ∈ Πκ,w to follow a particular distribution at each time. This is the key difference of the class Πκ,w from the class Γp0 ,p1 that represents random failures. 8 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL There are several ways an attacker can strategize when to cause transmission failures. For instance, game-theoretic [14, 15] and optimization-based methods can be used by the attacker to decide the timing of attacks. An important property of the class Πκ,w is that it characterizes attacks by their maximum average attack rate but not by the specific timing strategy they follow. Thus by using Πκ,w , we can capture the uncertainty in the generation of attacks, which may follow a deterministic strategy, or may involve randomness. Interestingly, an attacker can also make use of system dynamics as well as past/present state information to decide the timing of attacks to cause more damage to the system. Such attacks are further illustrated in Section 5.3 and explored through the characterization of Πκ,w . As we discussed above, a process l that belongs to either of the classes Πκ,w and Γp0 ,p1 also belongs to the class Λρ for a suitable value of ρ. This observation suggests us that both random failures and malicious attacks can be characterized by utilizing the class Λρ . As we make it clear in the following sections, the class Λρ has useful properties. For instance, if two processes l1 and l2 belong to classes Λρ1 and Λρ2 , then the processes l∧ and l∨ defined by setting l∧ (t) = l1 (t) ∧ l2 (t) and l∨ = l1 (t) ∨ l2 (t) belong to classes Λρ∧ and Λρ∨ where ρ∧ and ρ∨ depend on ρ1 and ρ2 . Such properties enable us to model the failures on both the links and the paths of a network by using processes that belong to the class Λρ for certain values of ρ. 3. Random and Malicious Packet Failures in Multi-Hop Networks In this section, we present a framework for modeling random and malicious packet transmission failures in the multi-hop networks that are used for exchanging state and control input packets between the plant and the controller. 3.1. Multi-Hop Network Model. We follow the approach of [24] and represent the networks between the plant and the controller by using directed acyclic graphs. To model the network, over which the state packets are transmitted from the plant to the controller, we consider the directed acyclic graph G , (V, E), where V denotes the set of nodes, and E ⊂ {(v1 , v2 ) : v1 , v2 ∈ V } denotes the set of edges. Here the nodes and the edges correspond respectively to communication devices and links. We represent the nodes at the plant and the controller with vP ∈ V and vC ∈ V , respectively. A path P from a node v1 ∈ V to another node vh ∈ V is identified as a sequence of nonrepeating edges P = (v1 , v2 ), (v2 , v3 ), . . . , (vh−1 , vh ) . We write \|P\| to denote the number of edges on the path P. Similarly, the network used for transmission of the control input packets from the controller to the plant is represented by graph G̃ , (Ṽ , Ẽ) with plant and controller nodes ṽP ∈ Ṽ and ṽC ∈ Ṽ . In practice, the same physical network may be used for transmission of both the state and the control input packets. For those cases, the nodes in V and Ṽ would correspond to the same physical devices. We assume that there exists at least one directed path from node vP to node vC in G, and at least one directed path from node ṽC to node ṽP in G̃. Example 3.1. We show an example of G and G̃ in Fig. 1, where the nodes corresponding to the plant and the controller are not directly connected, but state and control input packets can still be transmitted with the help of the intermediate nodes v1 , . . . , v4 in G, and ṽ1 , . . . , ṽ5 in G̃. CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 9 Figure 1. Multi-hop networked control system Intermediate nodes in networks forward data packets that they receive from their incoming edges to their outgoing edges. Depending on the communication protocol, the forwarding method may differ. For instance, in the broadcast method, intermediate nodes forward all data packets that they receive from their incoming edges to all the nodes that they are connected with their outgoing edges. On the other hand, it may also be the case that intermediate nodes follow a specific routing scheme, where a packet coming from a certain incoming edge is forwarded through a certain outgoing edge [36]. We assume that the information packets propagate with no delay in both networks represented with G and G̃, although there may be transmission failures due to: 1) packet drops by malicious nodes to prevent communication and/or by nonmalicious nodes to avoid congestion; 2) data corruption on links because of random channel errors and/or malicious jamming attacks. Hence, a packet exchange between the plant and the controller may fail if the state or the control input packets are dropped or get corrupted. Here note that corrupted data packets are allowed to be transmitted over intermediate nodes, but they are detected and discarded at the plant/controller nodes. Error-detecting codes can be used for this purpose. Note also that if the controller only receives corrupted versions of a state packet, the control input is not computed. In the following sections, we present some key results for the analysis of packet failures on network G, which are directly applicable for analyzing G̃. In particular, we characterize failures on G in terms of the failures on different paths between the plant and the controller. Then we present a set of results that relate data corruption and packet dropping issues of nodes and links to the failures on each individual path of G. These results enable us to obtain ρ ∈ [0, 1] in Assumption 2.2, which is essential for analyzing the system in Fig. 1 with Theorem 2.3. We emphasize that the central problem here is to find ρ for the overall multi-hop network in Fig. 1 in a nonconservative way while still taking into account mutually-dependent packet failures and coordinated attacks on the network. 3.2. Packet Transmission Failures on Networks. We use the binary-valued process {lG (t) ∈ {0, 1}}t∈N0 to indicate transmission failures on G. Specifically, lG (t) = 0 means that the state packet x(t) sent from the plant node vP is successfully 10 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL received at the controller node vC . On the other hand, lG (t) = 1 indicates a failure, that is, the controller does not receive the state x(t). Let c ∈ N denote the number of paths on graph G from the node vP to the node vC , and let Pi , i ∈ {1, . . . , c}, denote these paths. In the example network G in Fig. 1, there are c = 3 paths P1 = (vP , v1 ), (v1 , vC ) , P2 = (vP , v2 ), (v2 , v4 ), (v4 , vC ) , (10) P3 = (vP , v3 ), (v3 , v4 ), (v4 , vC ) . Note that different paths may include the same link. Hence, when packet transmission is attempted on multiple paths, a link that is shared on those paths may be used multiple times. For instance, (v4 , vC ) is on both P2 and P3 . Hence, (v4 , vC ) may be utilized twice to forward the packets coming from v2 and v3 . On the other hand, the framework that we describe below also allows modeling the case where one of the packets is dropped at node v4 and not transmitted further. Furthermore, the packet drop can be random or malicious. We use {lPi (t) ∈ {0, 1}}t∈N0 to indicate whether the state packet x(t) is successfully transmitted to the controller through path Pi or not. Specifically, lPi (t) = 0 represents a successful transmission. On the other hand lPi (t) = 1 may indicate that the path is not utilized for transmission due to the particular routing scheme, or it may indicate a failure. Failures occur if packets get dropped on the path or if they get corrupted. Thus, in network G, the transmission of the state packet x(t) from node vP to node vC results in failure if lPi (t) = 1 for all paths Pi , i ∈ {1, . . . , c}. Therefore, lG (·) is given by lG (t) = lP1 (t) ∧ lP2 (t) ∧ · · · ∧ lPc (t). (11) The following result presents a probabilistic and asymptotic bound for the packet transmission failure ratio of G as a function of the bounds for the individual paths Pi . Proposition 3.2. Assume for each path Pi that we have lPi ∈ ΛρPi , where ρPi ∈ [0, 1]. Then lG ∈ ΛρG with ρG , mini∈{1,...,c} ρPi . Proof. First let i∗ , arg mini∈{1,...,c} ρPi . It follows that ρG = mini∈{1,...,c} ρPi = ρPi∗ . Now, by (11), k−1 X lG (t) ≤ t=0 and hence Pk−1 t=0 lG (t) P[ ≤ k−1 X k−1 X lPi (t), i ∈ {1, . . . , c}, t=0 Pk−1 t=0 lPi∗ (t). Therefore, lG (t) > ρG k] ≤ P[ t=0 The result then follows, since lPi∗ ∈ ΛρPi∗ . k−1 X lPi∗ (t) > ρPi∗ k]. t=0 The scalars ρPi , i ∈ {1, . . . , c}, in Proposition 3.2 represent bounds for asymptotic packet failure ratios on different paths of network G. Proposition 3.2 indicates that the minimum of these scalars is also a bound for the packet failure ratio of the whole network. Observe that if ρPi = 0 for some path Pi , then we have ρG = 0, CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 11 which means that the state can be securely and reliably transmitted to the controller at all time instants. This is because the transmission on path Pi never fails. If, on the other hand, ρPi = 1 for all paths Pi , then ρG = 1, indicating all packet transmission attempts fail, almost surely. Note that in Proposition 3.2, we do not assume that {lPi (t) ∈ {0, 1}}t∈N0 are mutually-independent processes. This allows us to deal with the scenarios where transmission failures on different paths may depend on each other. In particular, we can consider coordinated packet dropout attacks of several malicious routers on different paths. For instance, two malicious routers v2 and v3 in Fig. 1 may skip forwarding packets at the same time. Then transmissions on paths P2 and P3 given in (10) would both fail. Similarly, Proposition 3.2 is also useful when links on different paths are attacked at the same time by coordinated jamming attackers. Another scenario that can be explored through Proposition 3.2 is related to packet drops by nonmalicious routers to prevent congestion [9]. For example, a nonmalicious router v4 in Fig. 1 may choose to forward only one of the packets coming from v2 and v3 . Then, lP2 (·) and lP3 (·) would be dependent processes. In particular, if there are no other failures in the network, then we have lP2 (t) = 1 − lP3 (t). We remark that by utilizing additional properties of the indicator processes lPi (·) for paths, we can obtain a better asymptotic failure bound ρG than the one provided in Proposition 3.2. In particular, if one or more paths are known to be associated with random failures and the corresponding indicator processes are mutually independent, we can obtain tighter results than Proposition 3.2. To this end, we first present the following result on the properties of a process that is obtained by using ∧ operation on the output processes of two mutually-independent hidden Markov models. Theorem 3.3. Consider the binary-valued output processes {l(1) ∈ {0, 1}}t∈N0 and {l(2) ∈ {0, 1}}t∈N0 of hidden Markov models such that l(1) ∈ Γp(1) ,p(1) , l(2) ∈ (1) (1) (2) 0 (2) 1 (1) Γp(2) ,p(2) with p0 , p1 , p0 , p1 ∈ [0, 1]. Suppose that the Markov chains {θl (t) ∈ 0 (1) 1 (2) (2) Θl }t∈N0 and {θl (t) ∈ Θl }t∈N0 associated with the processes l(1) and l(2) are (1) (1) (2) (2) (2) (1) mutually independent. Let p̃0 , min{p0 + p1 p0 , p0 + p1 p0 , 1} and p̃1 , (1) (2) p1 p1 . Then the process {l̃(t) ∈ {0, 1}}t∈N0 defined by ˜ l(t) = l(1) (t) ∧ l(2) (t), (12) t ∈ N0 , is the output process of a time-inhomogeneous hidden Markov model, and moreover, l̃ ∈ Γp̃0 ,p̃1 . (1) (2) Proof. Let Θl̃ , {(q (1) , q (2) ) : q (1) ∈ Θl , q (2) ∈ Θl } and define the bivariate process {θl̃ (t) ∈ Θl̃ }t∈N0 by (1) (2) θl̃ (t) = (θl (t), θl (t)), t ∈ N0 . It follows that {θl̃ (t) ∈ Θl̃ }t∈N0 is a time-inhomogeneous Markov chain with ini(1) (2) (2) (1) tial distribution ϑl̃(q(1) ,q(2) ) = ϑlq(1) ϑlq(2) , q (1) ∈ Θl , q (2) ∈ Θl , and time-varying (1) (2) transition probabilities pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) = plq(1) ,r(1) (t)plq(2) ,r(2) (t), t ∈ N0 . Here, for j ∈ {1, 2}, ϑl (j) (j) and pl (·) respectively denote the initial distribution and the 12 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL (j) (j) transition probability function for the Markov chain {θl (t) ∈ Θl }t∈N0 associated with the output process {l(j) ∈ {0, 1}}t∈N0 . Furthermore, it follows from (12) that (1) (1) (2) (2) ˜ l(t) = l(1) (t) ∧ l(2) (t) = l(1) (t)l(2) (t) = hl (θl (t))hl (θl (t)), t ∈ N0 . Now let hl̃ : Θl̃ → {0, 1} be given by (1) (2) hl̃ ((q, r)) = hl (q)hl (r), (q, r) ∈ Θl̃ . It follows that (6) holds with l replaced with l̃. Thus, {l̃(t) ∈ {0, 1}}t∈N0 is the output process of a time-inhomogeneous hidden Markov model. Our next goal is to prove l̃ ∈ Γp̃0 ,p̃1 by showing X (13) pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) ≤ p̃0 , (r (1) ,r (2) )∈Θl̃0 X (14) pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) ≤ p̃1 , t ∈ N0 . (r (1) ,r (2) )∈Θl̃1 First, we show (13). Observe that Θl̃0 = {(r(1) , r(2) ) ∈ Θl̃ : hl̃ ((r(1) , r(2) )) = 0} (1) (2) = {(r(1) , r(2) ) ∈ Θl̃ : hl (r(1) )hl (r(2) ) = 0}. (15) Now let j1 , j2 ∈ {1, 2} be such that j1 6= j2 . It follows from (15) that Θl̃0 = {(r(1) , r(2) ) ∈ Θl̃ : hl (j1 ) ∪ {(r(1) , r(2) ) ∈ Θl̃ : hl (16) Hence, by (16), we obtain X (r(j1 ) ) = 1, hl (j1 ) (j2 ) (r(j2 ) ) = 0} (r(j1 ) ) = 0}. pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) (r (1) ,r (2) )∈Θl̃0 = X pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) (j ) (j ) r (j1 ) ∈Θ0l 1 ,r (j2 ) ∈Θl 2 + X (j1 ) r (j1 ) ∈Θ1l X = (j1 ) r (j1 ) ∈Θ0l + X X r (j2 ) ∈Θl (17) + (j2 ) ,r (j2 ) ∈Θ0l X r (j2 ) ∈Θl (j1 ) r (j1 ) ∈Θ1l = pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) (j1 ) (j2 ) X (j1 ) (j2 ) r (j2 ) ∈Θ0l (j2 ) (j2 ) (j1 ) (j1 ) X (j1 ) plq(j1 ) ,r(j1 ) (t) (j1 ) r (j1 ) ∈Θ0l plq(j1 ) ,r(j1 ) (t) r (j1 ) ∈Θ1l (j2 ) plq(j1 ) ,r(j1 ) (t)plq(j2 ) ,r(j2 ) (t) plq(j2 ) ,r(j2 ) (t) X (j2 ) plq(j1 ) ,r(j1 ) (t)plq(j2 ) ,r(j2 ) (t) X (j2 ) pql (j2 ) ,r(j2 ) (t). (j2 ) r (j2 ) ∈Θ0l CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 13 P (j1 ) Now, since l(1) ∈ Γp(1) ,p(1) , l(2) ∈ Γp(2) ,p(2) , we have r(j1 ) ∈Θl(j1 ) plq(j1 ) ,r(j1 ) (t) ≤ 0 1 0 1 0 P (j1 ) (j2 ) (j ) P (j ) (j ) p0 1 , r(j1 ) ∈Θl(j1 ) plq(j1 ) ,r(j1 ) (t) ≤ p1 1 , and r(j2 ) ∈Θl(j2 ) plq(j2 ) ,r(j2 ) (t) ≤ p0 2 , t ∈ 1 0 P (j2 ) N0 . Furthermore, we have r(j2 ) ∈Θl(j2 ) plq(j2 ) ,r(j2 ) (t) = 1, since the summation is (j2 ) over all possible states r(j2 ) ∈ Θl . Using these inequalities in (17), we obtain X (j ) (j ) (j ) (18) pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) ≤ p0 1 + p1 1 p0 2 , (r (1) ,r (2) )∈Θl̃0 for t ∈ N0 . Since (18) holds for all j1 , j2 ∈ {1, 2} such that j1 6= j2 , we have (19) X (1) (1) (2) (2) (2) (1) pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) ≤ min{p0 + p1 p0 , p0 + p1 p0 }, t ∈ N0 . (r (1) ,r (2) )∈Θl̃0 P Furthermore, noting that Θl̃0 ⊂ Θl̃ , we obtain (r(1) ,r(2) )∈Θl̃ pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) ≤ 0 P l̃ p (t) = 1, t ∈ N . By using this inequality, it follows l̃ (1) (2) 0 (1) (2) (1) (2) (r ,r )∈Θ (q ,q ),(r ,r ) from (19) that (13) holds. Next, we show (14). Notice that Θl̃1 = {(r(1) , r(2) ) ∈ Θl̃ : hl̃ ((r(1) , r(2) )) = 1} (1) (2) = {(r(1) , r(2) ) ∈ Θl̃ : hl (r(1) )hl (r(2) ) = 1} (1) (2) = {(r(1) , r(2) ) ∈ Θl̃ : hl (r(1) ) = 1, hl (r(2) ) = 1}. (20) Noting that l(1) ∈ Γp(1) ,p(1) , l(2) ∈ Γp(2) ,p(2) , we use (20) to obtain for t ∈ N0 , 1 0 1 0 X X l̃ pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) p(q(1) ,q(2) ),(r(1) ,r(2) ) (t) = (1) (r (1) ,r (2) )∈Θl̃1 = X (1) r (1) ∈Θl1 = X X r (1) ∈Θl1 plq(1) ,r(1) (t)plq(2) ,r(2) (t) (2) r (2) ∈Θl1 X (1) plq(1) ,r(1) (t) (1) r (1) ∈Θl1 (2) ,r (2) ∈Θl1 (2) (1) (2) (1) (2) plq(2) ,r(2) (t) ≤ p1 p1 , (2) r (2) ∈Θl1 which implies (14). Now since (13) and (14) hold, we have l̃ ∈ Γp̃0 ,p̃1 . (1) (2) Theorem 3.3 shows that when two hidden Markov output processes l and l are combined with ∧ operation, the resulting process l̃ is also a hidden Markov output process. Furthermore, Theorem 3.3 provides the values of p̃0 , p̃1 for which l̃ ∈ Γp̃0 ,p̃1 . This result can be applied to obtain ρG . For instance, consider the case c = 2, where lP1 (·) and lP2 (·) are the output processes of hidden Markov models such that lP1 ∈ Γp(1) ,p(1) , lP2 ∈ Γp(2) ,p(2) . It follows from Theorem 3.3 with l(1) = lP1 , 0 1 0 1 (1) (2) l(2) = lP2 , and l̃ = lG that lG ∈ Γp̃0 ,p̃1 with p̃1 , p1 p1 . Now, suppose that (1) (2) p1 p1 < 1. Notice that since lP1 ∈ Γp(1) ,p(1) and lP2 ∈ Γp(2) ,p(2) , we have lP1 ∈ (1) 0 1 (2) 0 1 ΛρP1 and lP2 ∈ ΛρP2 with ρP1 ∈ (p1 , 1) and ρP2 ∈ (p1 , 1). The direct application of Proposition 3.2 gives lG ∈ ΛρG with ρG = min{ρP1 , ρP2 }. However, by applying Proposition 2.5, we can obtain a smaller value for ρG . In fact by Proposition 2.5, 14 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL (1) (2) we obtain lG ∈ ΛρG for any ρG ∈ (p1 p1 , 1). Notice that in the case where c > 2, Theorem 3.3 can be applied repeatedly. For instance, when c = 3, we can use Theorem 3.3 first for lP2 (t) ∧ lP3 (t) and then for lG (t) = lP1 (t) ∧ (lP2 (t) ∧ lP3 (t)). Now consider the case where the graph G possesses some paths with indicator processes that are mutually independent but not all of them are associated with random failures. Even for this case, we can obtain results that are tighter than Proposition 3.2. To this end, we first provide the following result where we derive properties of a process that is obtained by using ∧ operation on a hidden Markov output process l(1) ∈ Γp(1) p(1) and a binary valued process l(2) ∈ Λρ(2) . 0 1 Theorem 3.4. Consider the binary-valued processes {l(1) ∈ {0, 1}}t∈N0 and {l(2) ∈ (1) {0, 1}}t∈N0 that satisfy l(1) ∈ Γp(1) ,p(1) and l(2) ∈ Λρ(2) with p1 ρ(2) < 1. Then the 0 1 process {l̃(t) ∈ {0, 1}}t∈N0 defined by ˜ = l(1) (t) ∧ l(2) (t), (21) l(t) t ∈ N0 , (1) satisfies ˜ l ∈ Λρ̃ for all ρ̃ ∈ (p1 ρ(2) , 1]. (1) (1) Proof. We will first show that l̃ ∈ Λρ̃ for ρ̃ ∈ (p1 ρ(2) , min{p1 , ρ(2) }). Let ǫ , (1) ρ̃ − p1 ρ(2) . Note that ρ(2) + ǫ(1) < 1. It follows that {χ(t) , {0, 1}}t∈N0 defined 2p1 by χ(t) = l (2) (2) (t), t ∈ N0 , satisfies (8) with w̃ = p1 + (1) ǫ (1) . 2p1 Furthermore, {ξ(t) , (1) (1) {0, 1}}t∈N0 defined by ξ(t) = θl (t), t ∈ N0 , satisfies (7) with p̃ = p1 , h = hl , (1) (1) Ξ = Θl , and Ξ1 = Θl1 . It then follows from Lemma 2.6 that for all ρ ∈ (p̃w̃, w̃), we have X X k−1 X k−1 X P[ h(ξ(t))χ(t) > ρk] = P[ l(1) (t) ∧ l(2) (t) > ρk] < ∞. k∈N t=0 k∈N (1) p1 ρ(2) t=0 (1) < p1 ρ(2) and w̃ = ρ(2) + ǫ (1) (1) ρ̃ ∈ (p1 ρ(2) , min{p1 , ρ(2) }). (1) > min{p1 , ρ(2) }, it + Now since p̃w̃ = follows that ˜ l ∈ Λρ̃ holds for This implies ˜l ∈ Λρ̃ , (1) (2) ρ̃ ∈ [min{p1 , ρ }, 1], since Λρ1 ⊆ Λρ2 for any ρ1 , ρ2 ∈ [0, 1] such that ρ1 ≤ ρ2 . (1) (1) (1) Consequently, we have ˜ l ∈ Λρ̃ for ρ̃ ∈ (p1 ρ(2) , 1] = (p1 ρ(2) , min{p1 , ρ(2) }) ∪ (1) (2) [min{p1 , ρ }, 1]. ǫ 2 Theorem 3.4 is concerned with ∧ operation applied to a process l(1) (·) from the hidden Markov model class Γp(1) ,p(1) and another process l(2) (·) from the class Λρ(2) . (1) 0 1 It is shown that if p1 ρ(2) < 1, then this operation results in a process l̃ that satisfies (1) l̃ ∈ Λρ̃ for all ρ̃ ∈ (p1 ρ(2) , 1]. We note that the application of Proposition 3.2 to (1) this situation would allow us to show ˜l ∈ Λρ̃ for all ρ̃ ∈ (min{p1 , ρ(2) }, 1]. Notice (1) (1) that Proposition 3.2 in this case is conservative since min{p1 , ρ(2) } > p1 ρ(2) . On the other hand, we note that Proposition 3.2 allows us to deal with processes that are not mutually independent. 4. Packet Transmission Failures on Paths of a Network So far, in the previous section, we have looked at how packet failures on the paths of a network affect the overall packet transmission rate. In this section, our goal is to explore the effect of the failures at individual nodes and links of a path. CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 15 To this end, we first consider the scenario where packet transmission failures occur due to only data corruption. We then explore the case where data corruption and packet drops may occur on the same path. 4.1. Characterization for Data Corrupting Paths. Let Pi,j denote the jth P edge on path Pi . We use the binary-valued process {lPii,j (t) ∈ {0, 1}}t∈N0 to denote the data corruption indicator for this link. For example, in Fig. 1, consider the (v ,v ) second edge P1,2 = (v1 , vC ) of path P1 in (10). The state lP11 C (t) = 1 indicates that at time t, the packet flowing on path P1 faces data corruption on the link (v1 , vC ). This may be due to a jamming attack on this link, or due to channel noise, and moreover, it may also be the case that the node v1 maliciously corrupts the packet. The notation for the data corruption indicator allows us to distinguish data corruption issues when we consider the same communication link on different paths. For instance, communication link (v4 , vC ) is on both P2 and P3 . It may be the case that node v4 corrupts all packets transmitted along P2 , but none of the packets (v ,v ) transmitted along P3 . This situation can be described by setting lP24 C (t) = 1, (v ,v ) t ∈ N0 , and lP34 C (t) = 0, t ∈ N0 . State packet transmitted through path Pi is subject to data corruption if there is data corruption on one (or more) of the edges in this path. Hence, for each i ∈ {1, . . . , c}, P Pi,\|Pi \| P lPi (t) = lPii,1 (t) ∨ lPii,2 (t) ∨ · · · ∨ lPi (22) (t). The next result shows that an upper-bound for the asymptotic transmission failure rate of a path can be given as the sum of the failure rate bounds of the links on the path. P Proposition 4.1. Consider {lPi (t) ∈ {0, 1}}t∈N0 given by (22). Assume lPii,j ∈ P\|Pi \| Pi,j P ΛρPi,j , j ∈ {1, . . . , \|Pi \|}, where ρPi,j ∈ [0, 1], j ∈ {1, . . . , \|Pi \|}, satisfy j=1 ρPi ≤ i Pi P\|Pi \| Pi,j 1. Then lPi ∈ ΛρPi with ρPi , j=1 ρPi . P\|Pi \| Pk−1 Pi,j Pk−1 Proof. By (22), t=0 lPi (t) ≤ j=1 t=0 lPi (t). Hence, (23) P[ k−1 X lPi (t) > ρPi k] ≤ P[ t=0 \|Pi \| k−1 X X P lPii,j (t) > ρPi k] ≤ j=1 t=0 \|Pi \| X j=1 for k ∈ N, where to obtain the last inequality, we used P[ \|Pi \| X j=1 \|P \| i {γj > 0}] ≤ γj > 0] ≤ P[∪j=1 \|Pi \| X P[ k−1 X P P lPii,j (t) > ρPi,j k], i t=0 P[γj > 0] j=1 Pk−1 P P P k. The result then follows from (23) since lPii,j ∈ with γj , t=0 lPii,j (t) − ρPi,j i Λ Pi,j , j ∈ {1, . . . , \|Pi \|}. ρP i Proposition 4.1 can be used to characterize overall failures on a path. Note that in Proposition 4.1, the indicators for the links are not necessarily mutuallyindependent processes. This allows us to model failures on different links that 16 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL depend on each other. In particular, we can explore the effect of interference between links as well as coordinated jamming attacks targeting several links at the same time. Note that in certain cases the result provided in Proposition 4.1 can be improved in terms of conservativeness. In particular, if one or more links in a path are known to be associated with random failures and the corresponding indicator processes are mutually independent, we can obtain less conservative results in comparison to Proposition 4.1. The following result is the counterpart of Theorem 3.3 for ∨ operation and it is concerned with output processes of two mutually-independent hidden Markov models. Theorem 4.2. Consider the binary-valued output processes {l(1) ∈ {0, 1}}t∈N0 and {l(2) ∈ {0, 1}}t∈N0 of hidden Markov models such that l(1) ∈ Γp(1) ,p(1) , l(2) ∈ (1) (1) (2) 0 (2) 1 (1) Γp(2) ,p(2) with p0 , p1 , p0 , p1 ∈ [0, 1]. Suppose that the Markov chains {θl (t) ∈ 0 (1) 1 (2) (2) Θl }t∈N0 and {θl (t) ∈ Θl }t∈N0 associated with the processes l(1) and l(2) (1) (2) (1) (1) (2) (2) are mutually independent. Let p̃0 , p0 p0 and p̃1 , min{p1 + p0 p1 , p1 + (2) (1) p0 p1 , 1}. Then the process {l̃(t) ∈ {0, 1}}t∈N0 defined by ˜ l(t) = l(1) (t) ∨ l(2) (t), (24) t ∈ N0 , is the output process of a time-inhomogeneous hidden Markov model, and moreover, l̃ ∈ Γp̃0 ,p̃1 . (1) (2) Proof. Let Θl̃ , {(q (1) , q (2) ) : q (1) ∈ Θl , q (2) ∈ Θl } and define the bivariate process {θl̃ (t) ∈ Θl̃ }t∈N0 by (1) (2) θl̃ (t) = (θl (t), θl (t)), t ∈ N0 . It follows that {θl̃ (t) ∈ Θl̃ }t∈N0 is a time-inhomogeneous Markov chain with ini(1) (2) (2) (1) tial distribution ϑl̃(q(1) ,q(2) ) = ϑlq(1) ϑlq(2) , q (1) ∈ Θl , q (2) ∈ Θl , and time-varying (2) (1) transition probabilities pl̃(q(1) ,q(2) ),(r(1) ,r(2) ) (t) = plq(1) ,r(1) (t)plq(2) ,r(2) (t), t ∈ N0 . Here, (j) (j) for j ∈ {1, 2}, ϑl and pl (·) respectively denote the initial distribution and the (j) (j) transition probability function for the Markov chain {θl (t) ∈ Θl }t∈N0 associated with the output process {l(j) ∈ {0, 1}}t∈N0 . Furthermore, it follows from (12) that ˜ l(t) = l(1) (t) ∨ l(2) (t) = l(1) (t) + (1 − l(1) (t))l(2) (t) (1) (1) (1) (1) (2) (2) = hl (θl (t)) + (1 − hl (θl (t)))hl (θl (t)), for t ∈ N0 . Now let hl̃ : Θl̃ → {0, 1} be given by (1) (1) (2) hl̃ ((q, r)) = hl (q) + (1 − hl (q))hl (r), (q, r) ∈ Θl̃ . It follows that (6) holds with l replaced with l̃. Thus, {l̃(t) ∈ {0, 1}}t∈N0 is the output process of a time-inhomogeneous hidden Markov model. To show ˜ l ∈ Γp̃0 ,p̃1 , we can apply Theorem 3.3. To this end, we first note that for an output process l ∈ Γp0 ,p1 , the complementing process {lc (t) ∈ {0, 1}}t∈N0 given by lc (t) = 1 − l(t), t ∈ N0 , is the output process of a time-inhomogeneous hidden Markov model, for which Θlc = Θl , θlc (t) = θl (t), t ∈ N0 . Furthermore, we have Θl0c = Θl1 and Θl1c = Θl0 , since hlc ((q, r)) = 1 − hl ((q, r)), (q, r) ∈ Θlc = Θl . CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 17 As a consequence, we have l ∈ Γp0 ,p1 if and only if lc ∈ Γp1 ,p0 . In the following, we show l̃ ∈ Γp̃0 ,p̃1 by proving that {l̃c (t) ∈ {0, 1}}t∈N0 given by ˜lc (t) = 1 − ˜l(t), t ∈ N0 , satisfies l̃c ∈ Γp̃1 ,p̃0 . lc (t) = 1−˜l(t) = 1−l(1) (t)∨l(2) (t) = (1−l(1) (t))∧(1−l(2) (t)) = First, observe that ˜ (1) (2) (i) lc (t) ∧ lc (t), where lc (t) = 1 − l(i) (t), i ∈ {1, 2}. Since l(1) ∈ Γp(1) ,p(1) , l(2) ∈ (1) Γp(2) ,p(2) , we have lc 0 1 (2) ∈ Γp(1) ,p(1) , lc 1 0 0 1 ∈ Γp(2) ,p(2) . Finally, noting that l̃c (·) is 1 (1) 0 (2) obtained by using ∧ operation on processes lc (·), lc (·), we can use Theorem 3.3, to obtain ˜ lc ∈ Γp̃1 ,p̃0 , which implies ˜l ∈ Γp̃0 ,p̃1 . Theorem 4.2 shows that when two hidden Markov output processes l(1) and l(2) are combined with ∨ operation, the resulting process l̃ is also a hidden Markov output process. Furthermore, it provides the values of p̃0 , p̃1 for which l̃ ∈ Γp̃0 ,p̃1 . Theorem 4.2 can be applied in obtaining ρPi for a given path Pi . Consider for example a path Pi with \|Pi \| = 2 links. Assume that the failure indicator P P processes lPii,1 (·), lPii,2 (·) associated with the links are mutually independent and P P P lPii,1 ∈ Γp(1) ,p(1) , lPii,2 ∈ Γp(2) ,p(2) . It follows from Theorem 4.2 with l(1) (t) = lPii,1 (t) 0 1 0 P 1 (1) (2) and l(2) (t) = lPii,2 (t) that lPi ∈ Γp̃0 ,p̃1 with p̃0 = p0 p0 (1) (2) (2) (2) (1) (1) (1) and p̃1 = min{p1 + (1) (2) (2) (2) (1) p0 p1 , p1 + p0 p1 , 1}. Furthermore, if min{p1 + p0 p1 , p1 + p0 p1 } < 1, (1) (1) (2) (2) then by Proposition 2.5, we have lPi ∈ ΛρPi with ρPi ∈ (min{p1 + p0 p1 , p1 + (2) (1) p0 p1 }, 1]. On the other hand, the direct application of Proposition 4.1 provides P a conservative result. To apply Proposition 4.1, notice first that lPii,1 ∈ ΛρPi,1 , P (2) (1) lPii,2 ∈ ΛρPi,2 with ρPi,1 ∈ (p1 , 1] and ρPi,2 ∈ (p1 , 1]. Hence, by Proposition 4.1, we obtain the value ρPi = ρPi,1 + ρPi,2 , which implies that lPi ∈ ΛρPi with ρPi ∈ (1) (2) (1) (1) (2) (2) (2) (1) (1) (2) (p1 + p1 , 1]. The inequality min{p1 + p0 p1 , p1 + p0 p1 } ≤ p1 + p1 allows us to conclude that Theorem 4.2 provides a less conservative range for ρPi compared to Proposition 4.1. In certain scenarios a path Pi may be composed of communication links with mutually-independent indicator processes some of which are not associated with random failures. In such scenarios, it is again possible to obtain results that are less conservative than those in Proposition 4.1. Specifically, in the following result we obtain properties of a process that is obtained by using ∨ operation on a hidden Markov output process l(1) ∈ Γp(1) p(1) and a binary valued process l(2) ∈ Λρ(2) . 0 1 Theorem 4.3. Consider the binary-valued processes {l(1) ∈ {0, 1}}t∈N0 and {l(2) ∈ (1) (1) {0, 1}}t∈N0 that satisfy l(1) ∈ Γp(1) ,p(1) and l(2) ∈ Λρ(2) with p1 +p0 ρ(2) < 1. Then 0 1 the process {l̃(t) ∈ {0, 1}}t∈N0 defined by (25) ˜ l(t) = l(1) (t) ∨ l(2) (t), t ∈ N0 , (1) (1) satisfies ˜ l ∈ Λρ̃ for all ρ̃ ∈ (p1 + p0 ρ(2) , 1]. Proof. Notice that since ˜ l(·) is a binary-valued process, we have Λρ̃ for ρ̃ = 1. Next, (1) (1) ˜ we show l ∈ Λρ̃ for ρ̃ ∈ (p1 + p0 ρ(2) , 1). First, (25) implies (26) l̃(t) = l(1) (t) + (1 − l(1) (t))l(2) (t), t ∈ N0 . 18 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL Pk−1 ˜ t=0 l(t). It follows from (26) that Let L̃(k) , (27) L̃(k) = k−1 X l(1) (t) + t=0 (1) k−1 X (1 − l(1) (t))l(2) (t), k ∈ N. t=0 (1) (2) ρ(2) −p (1) ρ 0 }, ǫ1 , ǫ − ǫ2 , and define Now, let ǫ , ρ̃ − p1 − p0 ρ(2) , ǫ2 , min{ 2ǫ , 2 Pk−1 (1) (1) (1) (2) (t) and ρ̃1 , p1 + ǫ1 , ρ̃2 , p0 ρ + ǫ2 . Furthermore, let L̃1 (k) , t=0 l Pk−1 (1) (2) L̃2 (k) , t=0 (1 − l (t))l (t). We then have P[L̃(k) > ρ̃k] = P[L̃1 (k) + L̃2 (k) > ρ̃1 k + ρ̃2 k] oi o n hn ≤ P L̃1 (k) > ρ̃1 k ∪ L̃2 (k) > ρ̃2 k ≤ P[L̃1 (k) > ρ̃1 k] + P[L̃2 (k) > ρ̃2 k]. P∞ P In the following we show k=1 P[L̃1 (k) > ρ̃1 k] < ∞ and ∞ k=1 P[L̃2 (k) > ρ̃2 k] < ∞. First, note that (28) (1) ǫ (1) ρ(2) − p0 ρ(2) (1) (1) ρ̃1 = p1 + ǫ − ǫ2 = max{p1 + , p1 + ǫ − } 2 2 (1) (1) (1) ρ(2) − p0 ρ(2) p + ρ̃ − p0 ρ(2) (1) , ρ̃ − p0 ρ(2) − } = max{ 1 2 2 (1) (1) p + ρ̃ − p0 ρ(2) 2ρ̃ − ρ(2) (1 + p0 ) , }. = max{ 1 2 2 (29) (1) p (1) (2) +ρ̃−p (2) ρ (1) 0 As 1 < 1 and 2ρ̃−ρ 2(1+p0 ) < 1, it holds from (29) that ρ̃1 ∈ (p1 , 1). 2 (1) Since l ∈ Γp(1) ,p(1) , we can use Proposition 2.5 with ρ replaced with ρ̃ and l 0 1 P∞ replaced with l(1) to obtain k=1 P[L̃1 (k) > ρ1 k] < ∞. P Next, we use Lemma 2.6 to show that ∞ k=1 P[L̃2 (k) > ρ̃2 k] < ∞. To obtain (1) (2) this result, we first observe that ρ̃2 > p0 ρ , since ǫ2 > 0. Moreover, (1) (1) ǫ ρ(2) − p0 ρ(2) ρ(2) − p0 ρ(2) (1) (1) ρ2 = p0 ρ(2) + min{ , } ≤ p0 ρ(2) + 2 2 2 (1) (2) (1) (2) (2) (2) < p0 ρ + ρ − p0 ρ = ρ , (1) (1) and hence, we have ρ2 ∈ (p0 ρ(2) , ρ(2) ). Let {lc (t) ∈ {0, 1}}t∈N0 be defined by (1) (1) lc (t) = 1 − l(1) (t), i ∈ {1, 2}. Since l(1) ∈ Γp(1) ,p(1) , we have lc ∈ Γp(1) ,p(1) . 0 1 1 0 Furthermore, since l(2) ∈ Λρ(2) , conditions (7), (8) in the Lemma 2.6 hold with (1) p̃ = p0 and w̃ = ρ(2) , together with processes {ξ(t) ∈ {0, 1}}t∈N0 and {χ(t) ∈ {0, 1}}t∈N0 defined by setting ξ(t) = θlc (t), χ(t) = lM (t), t ∈ N0 . Now, we have P∞ Pk−1 L̃2 (k) = t=0 ξ(t)χ(t) and hence, Lemma 2.6 implies k=1 P[L̃2 (k) > ρ̃2 k] < ∞. Finally, by (28), we arrive at ∞ X k=1 P[L̃(k) > ρ̃k] ≤ ∞ X P[L̃1 (k) > ρ̃1 k] + k=1 ∞ X P[L̃2 (k) > ρ̃2 k] < ∞, k=1 (1) (1) which shows that ˜ l ∈ Λρ̃ for all ρ̃ ∈ (p1 + p0 ρ(2) , 1]. (1) Theorem 4.3 is concerned with ∨ operation applied to a process l (·) from the hidden Markov model class Γp(1) ,p(1) and another process l(2) (·) from the class Λρ(2) . 0 1 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 19 It is shown that the ∨ operation results in a process l̃ that satisfies l̃ ∈ Λρ̃ for all (1) (1) ρ̃ ∈ (p1 +p0 ρ(2) , 1]. Notice that the application of Proposition 3.2 to this situation (1) would allow us to show l̃ ∈ Λρ̃ for all ρ̃ ∈ (p1 + ρ(2) , 1]. Proposition 4.1 in this case (1) (1) (1) (1) (1) (1) is conservative since p1 + ρ(2) ≥ p1 + p0 ρ(2) (and p1 + ρ(2) > p1 + p0 ρ(2) if (1) p0 < 1). We remark that the advantage of Proposition 4.1 may be that it allows us to deal with processes that are not mutually independent. 4.2. A General Characterization of Paths with Data Corruption and Packet Drops. In the previous section, we provided a characterization for paths with data-corrupting links. In what follows we extend this characterization to consider effects of both data corruption and packet dropouts. Without loss of generality, we assume that links on a path are either data-corrupting or packet-dropping, but not both. Note that if in the original network, a link (v, w) is subject to both of the issues, we can artificially add a node v ′ and edges (v, v ′ ), (v ′ , w) to the graph, and consider (v, v ′ ) as a packet-dropping link and (v ′ , w) as a data-corrupting link. For a data-corrupting link (v, w), packets available at v are always transmitted to w, but their content may be externally manipulated or damaged during the transmission over this link. On the other hand, if (v, w) is a packet-dropping link, packets available at v may or may not be transmitted to w, but never get corrupted on link (v, w). A packet dropout may occur if v is a malicious router that intentionally skips forwarding packets (see blackhole and grayhole attacks in [19]). A nonmalicious router may also drop packets to avoid congestion. In addition to these two issues, a packet may also be dropped if the header part of the packet, which includes information on the destination of the packet, is corrupted. Furthermore, in scenarios where error-detection is implemented at intermediate nodes, corruption on the data part of a packet can be detected. As a result, a corrupted packet need not be further transmitted. Such a scenario can also be studied within the packet drop framework. Consider a link Pi,j = (v, w) on path Pi . Let t = 0, 1, . . ., denote the indices of packets that node v possesses (or receives from previous nodes on path Pi ). Observe that if there are links before Pi,j that are packet-dropping, then the first packet (with index t = 0) that v receives may be different from x(0), and it may be the state at a later time. This is because x(0) may have been dropped before reaching node v. We use lPi,j (t) ∈ {0, 1} to indicate the status of transmission of the (t + 1)th packet (with index t) that node v possesses to node w. Our goal here is to obtain a relation between the asymptotic packet failure ratio of path Pi and the failure ratios of the links on that path. To this end, we will use a recursive characterization for describing packet failures on paths. Specifically, consider a path (30) P , (v1 , v2 ), (v2 , v3 ), . . . , (vh , vh+1 ) of h ≥ 1 links, and consider the associated process {lP (t) ∈ {0, 1}}t∈N0 . The state lP (t) = 0 indicates that the (t + 1)th packet that the first node v1 possesses can be successfully transmitted to the last node vh+1 , whereas lP (t) = 1 indicates a failure. (v ,v ) If h = 1 in (30), then we have lP (t) = lP 1 2 (t), t ∈ N0 . Now consider the case h ≥ 2, and let f (P) and R(P) respectively denote the first link on P, and the 20 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL subpath composed of the rest of the links, that is, (31) (32) f (P) = (v1 , v2 ), R(P) = (v2 , v3 ), . . . , (vh , vh+1 ) . We illustrate f (P) and R(P) on the left side of Fig. 2. Next, we show that transmission failures on a path P can be characterized f (P) through transmission failures on the link f (P) and the subpath R(P). Let {lP (t) ∈ R(P) {0, 1}}t∈N0 and {lP (t) ∈ {0, 1}}t∈N0 denote indicators of transmission failures on the link f (P) and the subpath R(P). If the link f (P) = (v1 , v2 ) is a dataf (P) R(P) corrupting link, then we would have lP (t) = lP (t) ∨ lP (t). But this relation f (P) does not hold if f (P) is a packet-dropping link, because the index t for lP (t) and R(P) lP (t) represent different packets. R(P) Now, we will introduce a new process {b lP (t) ∈ {0, 1}}t∈N0 for which (33) f (P) lP (t) = lP R(P) (t) ∨ b lP (t), t ∈ N0 . R(P) R(P) If f (P) is a data-corrupting link, then we define b lP (·) by setting b lP (t) = R(P) lP (t), t ∈ N0 . On the other hand, if f (P) is a packet-dropping link, then for t ∈ N0 , ( f (P) 0, if lP (t) = 1, R(P) b lP (t) , R(P) (34) f (P) lP k̂(t + 1) − 1 , if lP (t) = 0, Pt−1 f (P) where k̂(t) , i=0 (1 − lP (i)), t ∈ N. In this definition, k̂(t + 1) denotes the number of packets that are successfully transmitted from node v1 to node v2 among the first t+1 packets that node v1 possesses. Hence, the scalar k̂(t+1)−1 represents R(P) the index of the k̂(t + 1)th packet received by v2 . Moreover, lP k̂(t + 1) − 1 indicates whether this packet is successfully transmitted from v2 over R(P) to vh+1 . R(P) Observe that by (34), b lP (t) is set to 0, if v1 drops the (t + 1)th packet that it R(P) possesses. On the other hand, if v1 transmits this packet to v2 , the state b lP (t) = 1 indicates that further transmission of this packet on subpath R(P) has failed, R(P) f (P) whereas b lP (t) = 0 indicates success. As a result, we have lP (t) = lP (t) + (1 − R(P) f (P) lP (t), and hence, (33) holds. lP (t))b The characterization above is recursive in the sense that it is recursively applied to describe failures on R(P) by means of failures on the first link f (R(P)) and the subpath R(R(P)). Example 4.4. In Fig. 2, we see a path P, its packet-dropping first link f (P) = R(P) f (P) lP (·), (v1 , v2 ), and its subpath R(P), together with sample trajectories of lP (·), b R(P) lP (·) and lP (·). Note that the link f (P) = (v1 , v2 ) drops the first two packets, that R(P) R(P) f (P) lP (1) = 0, and as lP (0) = b is, lP (t) = 1 for t ∈ {0, 1}. By (34), we have b a consequence of (33), lP (0) = lP (1) = 1. Furthermore, the link f (P) successfully f (P) R(P) transmits the 3rd packet, i.e, lP (2) = 0. The value of b lP (2) represents whether this packet is successfully transmitted over the subpath R(P) or not. Since this packet is the first packet that v2 receives, its transmission state is represented by R(P) R(P) R(P) R(P) lR(P) (0), and as a result, b lP (2) = lP (0). We have b lP (2) = lP (0) = 0, CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 21 Figure 2. [Left] Paths P, R(P), and packet-dropping link f (P). f (P) [Right] Trajectories of lP R(P) R(P) (·), b lP (·), lP (·) and lP (·). indicating successful transmission over R(P). Now, by (33), lP (2) = 0, which indicates that the 3rd packet possessed by the first node v1 is successfully transmitted to the last node vh+1 . The following result is essential for characterizing the failures on path P through those on the first link f (P) and the subpath R(P). We obtain the result by first Pk−1 R(P) Pk−1 R(P) establishing the key inequality t=0 b lP (t) ≤ t=0 lP (t), and then applying some of the arguments that we employed for proving Proposition 4.1. f (P) Lemma 4.5. Consider path P given in (30) for h ≥ 1. Assume lP f (P) R(P) ∈ [0, 1], ∈ ΛρR(P) with scalars ρP , ρP P f (P) R(P) follows that lP ∈ ΛρP with ρP = ρP + ρP . R(P) lP such that f (P) ρP + ∈Λ f (P) ρP R(P) ≤ ρP and 1. It R(P) Proof. We will first show that b lP ∈ ΛρR(P) . If f (P) is a data-corrupting link, P R(P) R(P) R(P) ∈ ΛρR(P) . Now consider lP we have b lP (t) = lP (t), t ∈ N0 , and therefore b P the situation where f (P) is a packet-dropping link. For this case, k−1 X t=0 R(P) b lP (t) = ( 0, Pk̂(k)−1 t=0 k̂(k) = 0, R(P) lP (t), k̂(k) 6= 0, k ∈ N. Pk̂(k)−1 R(P) Pk−1 R(P) Since k̂(k) ≤ k and lR(P) (t) ≥ 0, we have t=0 lP (t) ≤ t=0 lP (t), and Pk−1 R(P) Pk−1 bR(P) hence, t=0 lP (t) ≤ t=0 lP (t). It follows that P[ k−1 X t=0 R(P) Now, as lP R(P) R(P) b lP (t) > ρP k] ≤ P[ R(P) ∈ ΛρR(P) , we also have b lP P k−1 X R(P) lP t=0 ∈ ΛρR(P) . P R(P) (t) > ρP k]. 22 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL f (P) Finally, since (33) holds, we have lP (t) ≤ lP P[ k−1 X lP (t) > ρP k] ≤ P[ t=0 = P[ ≤ P[ f (P) lP (t) + f (P) lP (t) + k−1 X k−1 X t=0 f (P) lP (t) > k−1 X t=0 t=0 k−1 X t=0 (35) k−1 X R(P) (t) + b lP (t), t ∈ N0 . Hence, R(P) b lP (t) > ρP k] R(P) f (P) R(P) b lP (t) > (ρP + ρP )k] f (P) ρP k] + P[ k−1 X t=0 t=0 P∞ R(P) R(P) b lP (t) > ρP k]. Pk−1 f (P) f (P) It then follows from (35) together with k=1 P[ t=0 lP (t) > ρP k] < ∞ and P∞ Pk−1 P∞ Pk−1 bR(P) R(P) (t) > ρP k] < ∞ that k=1 P[ t=0 lP (t) > ρP k] < ∞, k=1 P[ t=0 lP which completes the proof. Now, for a given path Pi , repeated application of Lemma 4.5 to Pi , R(Pi ), R(R(Pi )), . . ., allows us to obtain the following result. P P ∈ [0, 1], j ∈ {1, . . . , \|Pi \|}, that Theorem 4.6. Assume lPii,j ∈ ΛρPi,j with ρPi,j i Pi P\|Pi \| Pi,j P\|Pi \| Pi,j ρPi . satisfy j=1 ρPi ≤ 1. Then lPi ∈ ΛρPi with ρPi , j=1 Note that Theorem 4.6 allows us to consider both data-corruption and packetdropouts on links, and hence it generalizes Proposition 4.1. Remark 4.7. By utilizing Theorem 4.6 together with Proposition 3.2, we can obtain ρG , ρG̃ ∈ [0, 1] as upper-bounds for the average number of packet exchange failures on networks G and G̃ such that lG ∈ ΛρG and lG̃ ∈ ΛρG̃ . Note that when lG (t) = 1, then the controller either does not receive the state packet or receives corrupted versions, which are discarded. Hence, when lG (t) = 1, no control input packet is attempted to be transmitted on G̃. This setting is similar to the situation that we discussed above for packet dropping links. Here we can consider the whole networked system as a path P from the node vP to node ṽP , where lG (·) corresponds to the indicator for the first packet dropping link (f (P)) and lG̃ (·) corresponds to packet transmission failure indicator for the rest of the path (R(P)). Hence, using arguments similar to the ones used above, we can show that if ρG + ρG̃ ≤ 1, then l ∈ Λρ with ρ = ρG + ρG̃ . This ρ value is utilized for stability analysis with Theorem 2.3. 5. Illustrative Numerical Examples In this section, we present illustrative examples to demonstrate the utility of our results in the characterization of communication failures on multi-hop networks. We also investigate the effect of those failures on the stability of a multi-hop networked control system. Consider the networked control system (1), (2) with 1 0.1 0.1 A= , B= (36) , K = −2.9012 −0.9411 −0.5 1.1 1.2 together with the networks G and G̃ in Fig. 1. This system was explored previously in [26] with a single channel network model. Here, differently from [26], we consider CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 23 networks G and G̃ that incorporate multiple paths and multiple links for packet transmissions. In what follows, we investigate various scenarios where we demonstrate the utility of our results in Sections 3 and 4 for characterizing overall network failures. For each different scenario, our goal is to find out the level of transmission failures that can be tolerated on the communication links so that the stability of the system (1), (2) is guaranteed. 5.1. Data corruption/packet dropout issues on both networks G and G̃. Consider the scenario where all links on both networks G and G̃ are subject to malicious/nonmalicious data corruption or packet dropout issues. We explore the general situation where the failures may depend on each other. For this general setup, we can use Proposition 3.2 and Theorem 4.6 for the characterization of the overall transmission failures between the plant and the controller. As explained in Remark 4.7, the overall packet exchange failure process l(·) satisfies Assumption 2.2 with ρ = ρG + ρG̃ . Here, ρG and ρG̃ are asymptotic failure ratios for the networks G and G̃, that is, lG ∈ ΛρG , lG̃ ∈ ΛρG̃ . To find the values of ρG and ρG̃ , we can use Proposition 3.2 and Theorem 4.6. In particular, by Proposition 3.2 and Theorem 4.6, we obtain ρG = mini∈{1,...,c=3} ρPi P\|Pi \| Pi,j where ρPi = j=1 ρPi . Similarly, we have ρG̃ = mini∈{1,...,c̃} ρP̃i and ρP̃i = P\|P̃i \| P̃i,j j=1 ρP̃ , where c̃ = 4 is the number of paths (denoted by P̃i ) from the controller i node ṽC to the plant node ṽP on the network G̃. Consequently, Assumption 2.2 holds with ρ= min i∈{1,...,c} \|Pi \| X P + ρPi,j i j=1 \|P̃i \| X min i∈{1,...,c̃} P̃ ρP̃i,j . i j=1 P P̃ The effect of the asymptotic packet failure ratios ρPi,j , ρP̃i,j on the stability of i i the networked system can be analyzed by using Theorem 2.3. First, we identify the values of asymptotic packet exchange failure ratio ρ in Assumption 2.2, for which the stability conditions (3)–(5) hold. For this numerical example, there exist a positive-definite matrix P and scalars β ∈ (0, 1), ϕ ∈ [1, ∞) that satisfy (3)–(5), when ρ is less than 0.411. Hence, Theorem 2.3 guarantees that the zero solution of the closed-loop system (1), (2) is asymptotically stable almost surely if the asymptotic packet transmission failure ratios satisfy (37) min i∈{1,...,c} \|Pi \| X j=1 P ρPi,j + i min i∈{1,...,c̃} \|P̃i \| X P̃ ρP̃i,j ≤ 0.411. i j=1 The system operator can guarantee (37) by ensuring that at least one path in each network is sufficiently secure and reliable. In particular, if there exist a P\|Pi∗ \| Pi∗ ,j path Pi∗ in network G and a path P̃i∗ in network G̃ such that j=1 ρPi∗ + P\|P̃i∗ \| P̃i∗ ,j j=1 ρP̃i∗ ≤ 0.411, then (37) holds and the stability is guaranteed regardless of the security/reliability of all other paths. Another approach to guarantee (37) is to ensure a certain level of security/reliability for all links. For this example, if P P̃ , ρP̃i,j ≤ ρ , 0.411 all links are sufficiently secure and reliable so that ρPi,j 6 , then i i (37) holds and the stability is guaranteed. To see this, first note that the network 24 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL Figure 3. Paths from the plant node vP to the controller node vC in network G Figure 4. Paths from the controller node ṽC to the plant node ṽP in network G̃ G contains 3 paths, and the number of links on these paths are given by 2, 3, 3. Furthermore, the network G̃ contains 4 paths, each of which contains 4 links. It follows that min i∈{1,...,c} \|Pi \| X j=1 P + ρPi,j i min i∈{1,...,c̃} \|P̃i \| X j=1 P̃ ρP̃i,j ≤ i min i∈{1,...,c} \|Pi \| X ρ+ j=1 = min{2ρ, 3ρ, 3ρ} + min{4ρ, 4ρ, 4ρ, 4ρ} = 6ρ ≤ 0.411, min i∈{1,...,c̃} \|P̃i \| X ρ j=1 which implies that (37) holds, and hence stability is guaranteed. Notice that in this example, we have not made any particular assumption on the independence or randomness of the failures on the links. In fact, all links may be subject to failures caused by actions of coordinated adversaries. In such a case, the occurrence of the failures may be nonrandom, and moreover, the binary-valued processes that characterize failures on different links would depend on each other. For example, in the case of data-corruption attacks, the worst-case scenario would be that the failures on the paths are synchronized so that packet transmissions necessarily fail in all parallel paths (such as paths P1 , P2 , and P3 of the graph P P̃ , ρP̃i,j ≤ ρ guarantees that G) at the same time. Notice that the condition ρPi,j i i such failures happen sufficiently rarely in average in the long run. Thus, networked stabilization can be achieved through the successful exchanges of measurement and control data. In the following examples, we will illustrate how our results in Sections 3 and 4 can be used for scenarios where some information on the properties of the communication links are available. CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 25 5.2. Jamming Attacks and Random Transmission Failures on Multiple Links. Consider the network G and its paths shown in Fig. 3. We assume that the paths P1 and P3 are subject to malicious attacks. In particular, the node v3 on P3 is assumed to be controlled by a malicious agent and all packets arriving P at node v3 are dropped. Hence, we have lP33,2 (t) = 1, t ∈ N0 , which implies that lP3 (t) = 1, t ∈ N0 . Moreover, the first and the second links on path P1 are assumed to be subject to jamming attacks that cause data corruption. Here the worst-case scenario happens when the attackers coordinate and jam only one of the links at once. This would maximize the number of packet losses within the total energy constraint of the attackers. Our results take this worst-case into account. P In particular, suppose that the attacked links P1,1 and P1,2 satisfy lP11,1 ∈ Πκ,w , P lP11,2 ∈ Πκ,w with κ ≥ 0 and w ∈ (0, 1). Here, w provides a bound on average failures on each link and it is related to the energy available to each attacker. The P P P P fact that lP11,1 ∈ Πκ,w , lP11,2 ∈ Πκ,w implies lP11,1 ∈ ΛρP1,1 , lP11,2 ∈ ΛρP1,2 for all P P1 P P1 P P ρP11,1 , ρP11,1 ∈ (w, 1]. It then follows from Proposition 4.1 that ρP1 = ρP11,1 + ρP11,1 , and hence lP1 ∈ ΛρP1 for ρP1 ∈ (2w, 1]. Notice that lP1 ∈ ΛρP1 holds even in the worst-case scenario mentioned above. Here ρP1 ∈ (2w, 1] provides an upper-bound on the average number of overall failures on path P1 in the worst-case scenario. We further assume that the link (v4 , vC ) on path P2 is subject to random data P corruption and the associated failure indicator process {lP22,3 (t) ∈ {0, 1}}t∈N0 satP isfies lP22,3 ∈ Γp0 ,p1 with p0 , p1 ∈ (0, 1). We consider ideal communication on the other links on path P2 . Thus, by using (22), we obtain P P P P P lP2 (t) = lP22,1 (t) ∨ lP22,2 (t) ∨ lP22,3 (t) = 0 ∨ 0 ∨ lP22,3 (t) = lP22,3 (t), t ∈ N0 , and hence we also have lP2 ∈ Γp0 ,p1 . To characterize overall failures on network G, we use (11) and obtain lG (t) = lP1 (t) ∧ lP2 (t) ∧ lP3 (t) = lP1 (t) ∧ lP2 (t) ∧ 1 = lP1 (t) ∧ lP2 (t), t ∈ N0 . Now, assuming that the failures on paths P1 and P2 are mutually-independent, it (1) (1) follows from Theorem 3.4 with l(1) (t) = lP2 (t), l(2) (t) = lP1 (t), p0 = p0 , p1 = p1 , and ρ(2) = ρP1 that if p1 ρP1 < 1, then lG ∈ ΛρG with ρG ∈ (p1 ρP1 , 1]. Noting that ρP1 ∈ (2w, 1], we see that if 2p1 w < 1, then lG ∈ ΛρG for all ρG ∈ (2p1 w, 1]. Next, consider the network G̃. Suppose that G̃ is secure against attacks, but it is unreliable and subject to random transmission failures. To characterize the overall transmission failures on the network G̃, we utilize Theorems 3.3 and 4.2. Before we apply these results, we first describe the routing scheme on this network. Specifically, in this network ṽ3 is assumed to be a router that forwards all incoming packets from node ṽ1 only to node ṽ5 , and all incoming packets from node ṽ2 only to node ṽ4 . Among the paths shown in Fig. 4, packets may be transmitted on P̃1 and P̃2 , but are never transmitted on P̃3 and P̃4 due to the routing scheme. Hence lP̃3 (t) = 1, lP̃4 (t) = 1, t ∈ N0 . We assume that the links (ṽ1 , ṽ3 ) and (ṽ3 , ṽ5 ) on path P̃1 and the links (ṽ2 , ṽ3 ) and (ṽ3 , ṽ4 ) on P̃2 face random data corruption issues. Furthermore, the failure (ṽ ,ṽ ) (ṽ ,ṽ ) (ṽ ,ṽ ) indicator processes {lP̃ 1 3 (t) ∈ {0, 1}}t∈N0 , {lP̃ 3 5 (t) ∈ {0, 1}}t∈N0 , {lP̃ 2 3 (t) ∈ (ṽ ,ṽ ) 1 1 2 {0, 1}}t∈N0 , {lP̃ 3 4 (t) ∈ {0, 1}}t∈N0 are assumed to be mutually independent pro2 cesses that belong to the hidden Markov model class Γp0 ,p1 with p0 , p1 ∈ (0, 1) (see 26 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL Definition 2.4). The links that are connected directly to the plant and the controller nodes (ṽP and ṽC ) are considered to be ideal communication links. In other words, P̃ P̃ lP̃i,1 (t) = 0, lP̃i,4 (t) = 0, t ∈ N0 , i ∈ {1, 2}. i i Now, observe that the failure indicators for paths P̃1 and P̃2 satisfy P̃ P̃ P̃ P̃ P̃ P̃ lP̃i (t) = lP̃i,1 (t) ∨ lP̃i,2 (t) ∨ lP̃i,3 (t) ∨ lP̃i,4 (t) = 0 ∨ lP̃i,2 (t) ∨ lP̃i,3 (t) ∨ 0 i = i P̃ lP̃i,2 (t) i ∨ i P̃ lP̃i,3 (t), i i i i t ∈ N0 , i ∈ {1, 2}. P̃ P̃ (1) (2) By applying Theorem 4.2 with l(1) (t) = lP̃i,2 (t), l(2) (t) = lP̃i,3 (t), p0 = p0 = p0 , (1) i (2) i and p1 = p1 = p1 , we obtain lP̃i ∈ Γp̃0 ,p̃1 , i ∈ {1, 2}, where p̃0 = p20 and p̃1 = min{p1 + p0 p1 , p1 + p0 p1 , 1} = min{p1 + p0 p1 , 1}. Next, since lP̃3 (t) = 1, lP̃4 (t) = 1, t ∈ N0 , we have lG̃ (t) = lP̃1 (t) ∧ lP̃2 (t) ∧ lP̃3 (t) ∧ lP̃4 (t) = lP̃1 (t) ∧ lP̃2 (t) ∧ 1 ∧ 1 = lP̃1 (t) ∧ lP̃2 (t), for t ∈ N0 . It then follows from Theorem 3.3 with l(1) (t) = lP̃1 (t), l(2) (t) = lP̃2 (t), (1) (2) p0 = p0 with (1) = p20 , and p1 (1) (2) = p1 (1) (2) (2) = min{p1 + p0 p1 , 1}, that we have lG̃ ∈ ΓpG̃ ,pG̃ 0 (2) (1) (1) 1 (1) (2) pG̃ 0 = min{p0 + p1 p0 , p0 + p1 p0 , 1} = min{p0 + p1 p0 , 1} = min{p20 + min{p1 + p0 p1 , 1}p20 , 1} = min{p20 + min{p20 p1 + p30 p1 , p20 }, 1} = min{min{p20 + p20 p1 + p30 p1 , 2p20 }, 1} = min{p20 + p20 p1 + p30 p1 , 2p20 , 1} (1) (2) 2 2 and pG̃ 1 = p1 p1 = (min{p1 +p0 p1 , 1}) = min{(p1 +p0 p1 ) , 1}. Finally, in the case 2 where (p1 + p0 p1 ) < 1, as a consequence of Proposition 2.5, we obtain lG̃ ∈ Λρ̃G 2 for ρG̃ ∈ (pG̃ 1 , 1] = ((p1 + p0 p1 ) , 1]. Now, we note once again that the overall packet exchange failure indicator {l(t) ∈ {0, 1}}t∈N0 satisfies l ∈ Λρ with ρ = ρG + ρG̃ . Since ρG ∈ (2p1 w, 1] and ρG̃ ∈ ((p1 + p0 p1 )2 , 1], it follows that if 2p1 w + (p1 + p0 p1 )2 < 1, then l ∈ Λρ with ρ ∈ (2p1 w+(p1 +p0 p1 )2 , 1]. As we discussed in Section 5.1, stability of the networked control system can be ensured when ρ ≤ 0.411. It follows that if (38) 2p1 w + (p1 + p0 p1 )2 < 0.411, then l ∈ Λρ holds with ρ = 0.411, implying almost sure asymptotic stability of the networked control system. Observe that by utilizing the results in Sections 3 and 4, we are able to derive the sufficient stability condition (38) in terms of the attack rate w ∈ (0, 1) associated with the links on path P1 that are attacked by jamming attackers as well as random failure parameters p0 , p1 ∈ (0, 1). 5.3. State-dependent Attacks by a Malicious Node. In the scenarios discussed in Sections 5.1 and 5.2, the strategies of the attackers are not specified. However, in certain cases an attacker may have access to the state or control input information and be able to directly cause transmission failures between the plant and the controller. In such scenarios, the goal of the attacker might be to increase state norm with small amount of attacks. In this section, our goal is to illustrate such an attack strategy. In particular, we consider the case where the plant node vP is compromised by an attacker. The attacker is assumed to have access to the state information. CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 27 We consider an attack strategy that is based on an optimization problem. In particular, the attacker at node vP decides whether to transmit the state information on links P1,1 = (vP , v1 ), P2,1 = (vP , v2 ), and P3,1 = (vP , v3 ) after solving an optimization problem for maximizing the norm of the state at a future time. We represent the attackers actions with a binary-valued process {lA (t) ∈ {0, 1}}t∈N0 , where lA (t) = 0 indicates the attacker transmits the state packet on P1,1 = (vP , v1 ), P2,1 = (vP , v2 ), and P3,1 = (vP , v3 ) and lA (t) = 1 indicates no transmission. The attacker decides the values of the binary-valued process lA (t) according to lA (t) = a∗t , (39) where (a∗t , a∗t+1 , · · · , a∗t+N −1 ) ∈ {0, 1}N is a solution to the optimization problem PN 2 maximize i=1 kx̂(t + i)k (a ,··· ,a ) t t+N −1 (40) Pt+N −1 Pt−1 ai ≤ κ + w(t + N ), subject to i=t i=0 lA (t) + where x̂(t+i) = A+(1−at+i−1 )BK A+(1−at+i−2 )BK · · · A+(1−at )BK x(t). In this strategy, the attacker decides to attack (lA (t) = 1) or not (lA (t) = 0) at time t based on the optimization problem where the goal is to maximize the sum of squared state norms over the interval [t + 1, t + N ], while keeping the attack rate below a certain value w ∈ (0, 1). Here N ∈ N is the horizon in the optimization problem, and it can be large if the attacker has sufficient computational resources. Notice that under this attack strategy, we have lA ∈ Πκ,w , and hence, lA ∈ Λρ for any ρ ∈ (w, 1]. Observe that in this scenario, we have P P P lP11,1 (t) =lP22,1 (t) = lP11,1 (t) = lA (t), t ∈ N0 . P Since lA ∈ Πκ,w , we have lA ∈ ΛρA with ρA ∈ (w, 1]. As a consequence, lPii,1 ∈ Λ P ρPi,1 P with ρPii,1 ∈ (wA , 1] for all i ∈ {1, 2, 3}. Notice that lA (t) = 1 implies l(t) = 1, i since the attacker can completely prevent the packet exchange between the plant and the controller. Observe also that if there are other sources of transmission failures on the network, then there may be times when l(t) = 1 even if lA (t) = 0. As a result, if lA (t) = 0 then the attacker may not be able to correctly predict the state x(t + 1) at time t, as there may or may not be a failure in the network that prevents control input to reach the plant. However, as the optimization problem in (40) is solved at each time step, the updated state information is used for decision. Notice that in the case where all links other than P1,1 = (vP , v1 ), P2,1 = (vP , v2 ), and P3,1 = (vP , v3 ) are secure and reliable, we have lG (t) = lA (t), since P P P lG (t) = lP1 (t) ∧ lP2 (t) ∧ lP3 (t) = lP11,1 (t) ∧ lP22,1 (t) ∧ lP33,1 (t) = lA (t) ∧ lA (t) ∧ lA (t) = lA (t), t ∈ N0 . Hence, lG ∈ ΛρG with ρG ∈ (w, 1]. Now suppose that the network G̃ also faces failures. In particular, consider the 2 setup in Section 5.2, where lG̃ ∈ Λρ̃G for ρG̃ ∈ (pG̃ 1 , 1] = ((p1 + p0 p1 ) , 1]. Since 2 l ∈ Λρ with ρ = ρG +ρG̃ , we have l ∈ Λρ for all ρ ∈ (w+(p1 +p0 p1 ) , 1]. Noting that the stability of the networked control system can be ensured when ρ ≤ 0.411, we can impose a sufficient condition on the attack rate w. Specifically, for the scenario PSfrag replacements 28 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL kx(t)k ×101 w = 0.25, κ = 1, N = 2 0.5 0.0 0 kx(t)k ×10 5 10 15 20 1 25 w = 0.25, κ = 1, N = 10 1 0 0 5 10 15 20 25 Time [t] Figure 5. Comparison of state trajectories for different values of N in (40) of this section, if w + (p1 + p0 p1 )2 < 0.411, (41) then l ∈ Λρ holds with ρ = 0.411, and thus the networked control system is almost surely asymptotically stable. To illustrate the effect of different parameters in the attack strategy (40), we conduct simulations. First, we generate 50 sample trajectories of the process lG̃ by P̃ P̃ setting the failure processes lP̃i,2 , lP̃i,3 , i ∈ {1, 2}, as outputs of time-inhomogeneous i i hidden Markov models such that (42) P̃ lP̃i,j (t) = θi,j (t), i t ∈ N0 , i ∈ {1, 2}, j ∈ {2, 3}, where {θi,j (t) ∈ Θl = {0, 1}}t∈N0 is an Ft -adapted finite-state time-inhomogeneous Markov chain with initial distributions and time-varying transition probability functions satisfying P[θi,j (0) = 0] = 0.6 and P[θi,j (t + 1) = 1\|θi,j (t) = 0] = 0.2 + 0.02 cos2 (0.1t), P[θi,j (t + 1) = 1\|θi,j (t) = 1] = 0.2 + 0.02 sin2 (0.2t), P̃ P̃ for t ∈ N0 , i ∈ {1, 2}, j ∈ {2, 3}. Notice that lP̃i,2 , lP̃i,3 ∈ Γp0 ,p1 , i ∈ {1, 2}, with i i p0 = 0.8, p1 = 0.22. Next, for each sample trajectory of lG̃ , we simulate the networked control system (1), (2) under the attack strategy (40). In Fig. 5, we show state trajectories when attack parameters are selected as w = 0.25 and κ = 1. In top part of Fig. 5, the horizon parameter N = 2, and in the bottom part N = 10. Observe that with larger horizon variable, the attacker following the strategy (40) can utilize the resources more efficiently so that the state norm is set to larger values for longer durations even though in both cases N = 2 and N = 10, the attack belongs to the same class Πκ,w . Notice that with p0 = 0.8, p1 = 0.22, and w = 0.25, the inequality (41) holds. Therefore, the zero solution of the closed-loop system is almost surely asymptotically stable by Theorem 2.3. PSfrag replacements t=0 lG̃ (t) 0.0 Pk−1 0.5 1 k CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 1.0 (p1 + p0 p1 )2 = 0.156816 1 k Pk−1 t=0 lG (t) 0 l(t) 100 200 300 400 500 600 700 1.0 w = 0.25 0.5 0.0 0 100 200 300 t=0 1.0 0.5 1 k Pk−1 29 0.0 400 500 600 700 w + (p1 + p0 p1 )2 = 0.406816 0 100 200 300 400 500 600 700 Number of packet transmission attempts [k] Figure 6. Average number of failures on networks G̃ and G together with average number of overall packet exchange failures For w = 0.25, κ = 1, and N = 2, we show in Fig. 6 the average number of failures on the networks G̃ and G as well as the average number of total packet exchange failures between the plant and the controller. These averages are upper bounded in the long run by certain scalars. In particular, for a process l that satisfies l ∈ Λρ with Pk−1 ρ ∈ (ρ, 1], we have lim supk→∞ k1 t=0 l(t) ≤ ρ (see Lemma 3.3 in [26]). As a result, Pk−1 Pk−1 we have lim supk→∞ k1 t=0 lG̃ (t) ≤ (p1 + p0 p1 )2 , lim supk→∞ k1 t=0 lG (t) ≤ w, P k−1 and lim supk→∞ k1 t=0 l(t) ≤ w + (p1 + p0 p1 )2 . Next, we consider the case where κ = 10, which corresponds to the case where the attacker has more initial resources. In this case, we consider two scenarios w = 0.25 and w = 0.75, for which we obtain the sample state trajectories shown in Fig. 7. First, notice that in the case with κ = 10, w = 0.25 (top plot in Fig. 7), the state grows to larger values for longer durations in comparison to the case with κ = 1, w = 0.25 (top plot in Fig. 5). In both cases, the stability conditions hold with w = 0.25, and therefore the state eventually converges to zero as guaranteed by Theorem 2.3. Note that, when w is set to larger values, the attacker can attack at a higher rate. For this example, we set w = 0.75 and observe that for such a high attack rate all sample state trajectories diverge (see bottom plot in Fig. 7). 6. Conclusion In this paper, we explored state feedback control of a linear plant over an unreliable network that may also face malicious attacks. We developed a probabilistic PSfrag replacements 30 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL kx(t)k ×102 w = 0.25, κ = 10, N = 2 0.5 0.0 0 kx(t)k ×10 10 20 30 40 50 4 60 70 w = 0.75, κ = 10, N = 2 2.5 0.0 0 10 20 30 40 50 60 70 Time [t] Figure 7. Comparison of state trajectories for different values of w in (39) approach to characterize the failures on the network in terms of the failures on different paths between the plant and the controller. We showed that the failures on each path can be described as a combination of data-corruption and packet-dropout failures on the communication links of that particular path. We obtained sufficient conditions for almost sure stability of the overall networked control system, which allow us to check stability by using a probabilistic upper-bound obtained for the average number of packet exchange failures between the plant and the controller. Appendix A. Proof of Lemma 2.6 In this section, we prove Lemma 2.6. As in the proof of Lemma A.1 of [26], for proving Lemma 2.6, we use Markov’s inequality and follow the approaches used for obtaining Chernoff-type tail distribution inequalities for sums of random variables (see Appendix B of [37] and Section 1.9 of [33]). In the proof Lemma 2.6, the following result also plays a key role. Lemma A.1. Let {ξ(t) ∈ Ξ}t∈N0 be a finite-state time-inhomogeneous Markov chain with transition probabilities pq,r : N0 → [0, 1], q, r ∈ Ξ. Furthermore, let Ξ1 ⊂ Ξ be given by Ξ1 , {r ∈ Ξ : h(r) = 1}, where h : Ξ → {0, 1} is a binary-valued function. Then for all φ > 1, s ∈ N, and p̃ ∈ [0, 1] such that X (43) pq,r (t) ≤ p̃, q ∈ Ξ, t ∈ N0 , r∈Ξ1 we have (44) Ps E[φ j=1 h(ξ(tj )) ] ≤ φ ((φ − 1)p̃ + 1)s−1 , where t1 , t2 , . . . , ts ∈ N0 denote indices such that 0 ≤ t1 < t2 < . . . < ts . Proof. We use induction to show (44). First, we consider the case where s = 1. In this case, we have (45) Ps E[φ j=1 h(ξ(tj )) ] = E[φh(ξ(t1 )) ] ≤ φ. CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 31 Next, consider the case where s = 2. Observe that t1 ≤ t2 − 1, thus the random variable ξ(t1 ) is Ft2 −1 -measurable. Consequently, Ps E[φ j=1 h(ξ(tj )) ] = E[φh(ξ(t1 )) φh(ξ(t2 )) ] = E[E[φh(ξ(t1 )) φh(ξ(t2 )) \| Ft2 −1 ]] = E[φh(ξ(t1 )) E[φh(ξ(t2 )) \| Ft2 −1 ]], (46) where the last equality follows from the fact that φh(ξ(t1 )) is a measurable function of ξ(t1 ), and hence, it is Ft2 −1 -measurable. Now, since {ξ(t) ∈ Ξ}t∈N0 is a Markov chain, we have E[φh(ξ(t2 )) \| Ft2 −1 ] = E[φh(ξ(t2 )) \| ξ(t2 − 1)]. Therefore, Ps j=1 h(ξ(tj )) ] = E[φh(ξ(t1 )) E[φh(ξ(t2 )) \| ξ(t − 1)]] 2 i h = E φh(ξ(t1 )) φP[h(ξ(t2 )) = 1 \| ξ(t2 − 1)] + P[h(ξ(t2 )) = 0 \| ξ(t2 − 1)] i h = E φh(ξ(t1 )) φP[h(ξ(t2 )) = 1 \| ξ(t2 − 1)] + 1 − P[h(ξ(t2 )) = 1 \| ξ(t2 − 1)] E[φ (47) i h = E φh(ξ(t1 )) (φ − 1)P[h(ξ(t2 )) = 1 \| ξ(t2 − 1)] + 1 . P Here we obtain P[h(ξ(t2 )) = 1 \| ξ(t2 −1)] = P[ξ(t2 ) ∈ Ξ1 \|ξ(t2 −1)] = r∈Ξ1 P[ξ(t2 ) = r\|ξ(t2 − 1)]. Thus, by using (43) and (45), we arrive at i h i Ps E[φ j=1 h(ξ(tj )) ] ≤ E φh(ξ(t1 )) (φ − 1)p̃ + 1 = E[φh(ξ(t1 )) ]((φ − 1)p̃ + 1) ≤ φ((φ − 1)p̃ + 1). (48) Hence, we have that (44) is satisfied for s ∈ {1, 2}. Now, assume that (44) holds for s = s̃ > 2, i.e., Ps̃ (49) E[φ j=1 h(ξ(tj )) s̃−1 ] ≤ φ ((φ − 1)p̃ + 1) . We will to prove that (44) holds for s = s̃ + 1. By employing arguments similar to those used for obtaining (46)–(48), we get Ps̃+1 E[φ j=1 h(ξ(tj )) Ps̃ = E[E[φ j=1 Ps̃ ] = E[φ (50) φ j=1 h(ξ(tj )) Ps̃ = E[φ j=1 h(ξ(tj )) h(ξ(ts̃+1 )) h(ξ(tj )) h(ξ(ts̃+1 )) Ps̃ = E[φ j=1 h(ξ(tj )) h(ξ(ts̃+1 )) E[φ φ ] \| Fts̃+1 −1 ]] \| Fts̃+1 −1 ]] Ps̃ E[φh(ξ(ts̃+1 )) \| ξ(ts̃+1 − 1)]] ≤ E[φ j=1 h(ξ(tj )) Finally, by (49) and (50), we obtain (44) with s = s̃ + 1. ]((φ − 1)p̃ + 1). Next, by utilizing Lemma A.1, we prove Lemma 2.6. Proof of Lemma 2.6. First, we define h(k) , [h(ξ(0)), h(ξ(1)), . . . , h(ξ(k − 1))]T , χ(k) , [χ(0), χ(1), . . . , χ(k − 1)]T , k ∈ N. Next, let Fs,k , {χ ∈ {0, 1}k : χT χ = s}, s ∈ {0, 1, . . . , k}, k ∈ N. Here, we have Fs1 ,k ∩ Fs2 ,k = ∅, s1 6= s2 , and moreover, P[χ(k) ∈ ∪ks=0 Fs,k ] = 1, k ∈ N. By 32 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL utilizing these definitions we obtain for all ρ ∈ (p̃w̃, 1) and k ∈ N, P[ k−1 X T h(ξ(t))χ(t) > ρk] = P[h (k)χ(k) > ρk] t=0 (51) = k X X T P[h (k)χ(k) > ρk \| χ(k) = χ]P[χ(k) = χ]. s=0 χ∈Fs,k Since ξ(·) and χ(·) are mutually-independent, T T P[h (k)χ(k) > ρk \| χ(k) = χ] = P[h (k)χ > ρk]. (52) Therefore, it follows from (51) and (52) that for k ∈ N, P[ k−1 X h(ξ(t))χ(t) > ρk] = t=0 k X X T P[h (k)χ > ρk]P[χ(k) = χ] s=0 χ∈Fs,k ⌊w̃k⌋ = X X T P[h (k)χ > ρk]P[χ(k) = χ] s=0 χ∈Fs,k (53) k X + X T P[h (k)χ > ρk]P[χ(k) = χ]. s=⌊w̃k⌋+1 χ∈Fs,k In what follows, our goal is to find upper bounds for the two summation terms T in (53). We start with the second term. Since P[h (k)χ > ρk] ≤ 1, k ∈ N, we obtain k X X T P[h (k)χ > ρk]P[χ(k) = χ] s=⌊w̃k⌋+1 χ∈Fs,k (54) ≤ k X X P[χ(k) = χ] = P[ k−1 X χ(t) > w̃k] = σ̃k , t=0 s=⌊w̃k⌋+1 χ∈Fs,k for k ∈ N. Next, we obtain an upper bound for the first term in (53). Observe T that P[h (k)χ > ρk] = 0 for χ ∈ F0,k . It then follows that, for all k ∈ N such that ⌊w̃k⌋ = 0, we have ⌊w̃k⌋ X X (55) T P[h (k)χ > ρk]P[χ(k) = χ] = 0. s=0 χ∈Fs,k Moreover, for all k ∈ N such that ⌊w̃k⌋ ≥ 1, we obtain (56) ⌊w̃k⌋ X X s=0 χ∈Fs,k T P[h (k)χ > ρk]P[χ(k) = χ] = ⌊w̃k⌋ X X T P[h (k)χ > ρk]P[χ(k) = χ]. s=1 χ∈Fs,k T To obtain an upper bound for the term P[h (k)χ > ρk], we will utilize Markov’s inequality and Lemma A.1. First, for s ∈ {1, 2, . . . , ⌊w̃k⌋}, let t1 (χ), t2 (χ), . . . , ts (χ) CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 33 denote the indices of the nonzero entries of χ ∈ Fs,k such that t1 (χ) < t2 (χ) < · · · < ts (χ). Consequently, T (57) P[h (k)χ > ρk] = P[ s X htj (χ̄) (k) > ρk] = P[ j=1 s X h(ξ(tj (χ) − 1)) > ρk], j=1 for χ ∈ Fs,k , s ∈ {1, 2, . . . , ⌊w̃k⌋}, and k ∈ N such that ⌊w̃k⌋ ≥ 1. Now observe that φ > 1, since ρ ∈ (p̃w̃, w̃). By using Markov’s inequality, we obtain T P[h (k)χ > ρk] ≤ P[ s X Ps h(ξ(tj (χ) − 1)) ≥ ρk] = P[φ j=1 −ρk ≤φ (58) Ps E[φ j=1 h(ξ(tj (χ)−1)) j=1 h(ξ(tj (χ)−1)) ≥ φρk ] ]. Ps s−1 It follows from Lemma A.1 that E[φ j=1 h(ξ(tj (χ)−1)) ] ≤ φ ((φ − 1)p̃ + 1) . This inequality together with (56) and (58) imply that for all k ∈ N such that ⌊w̃k⌋ ≥ 1, ⌊w̃k⌋ X X T P[h (k)χ > ρk]P[χ(k) = χ] s=0 χ∈Fs,k ⌊w̃k⌋ ≤ X X s−1 φ−ρk φ ((φ − 1)p̃ + 1) P[χk = χ] s=1 χ∈Fs,k ⌊w̃k⌋ = φ−ρk+1 X ((φ − 1)p̃ + 1) X ((φ − 1)p̃ + 1) X s−1 s=1 ⌊w̃k⌋ −ρk+1 =φ P[χk = χ] χ∈Fs,k s−1 P[χk ∈ Fs,k ] s−1 . s=1 ⌊w̃k⌋ −ρk+1 ≤φ (59) X ((φ − 1)p̃ + 1) s=1 Notice that to obtain the last inequality in (59), we employed the fact that P[χk ∈ Fs,k ] ≤ 1. The summation term on the far right-hand side of (59) satisfies ⌊w̃k⌋ (60) X ⌊w̃k⌋ s−1 ((φ − 1)p̃ + 1) s=1 = w̃k ((φ − 1)p̃ + 1) ((φ − 1)p̃ + 1) −1 ≤ ((φ − 1)p̃ + 1) − 1 (φ − 1)p̃ −1 . Therefore, from (59) and (60), we obtain ⌊w̃k⌋ (61) X X s=0 χ∈Fs,k T P[h (k)χ > ρk]P[χ(k) = χ] ≤ φ−ρk+1 ((φ − 1)p̃ + 1)w̃k − 1 , (φ − 1)p̃ for all k ∈ N such that ⌊w̃k⌋ ≥ 1. The right-hand side of this inequality is zero if ⌊w̃k⌋ = 0. Therefore, (61) holds also for all k ∈ N. By using this fact together with (53), (54), we arrive at (9). 34 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL Next, we show ∞ X P∞ k=1 φ−ρk+1 k=1 (62) = ψk < ∞. First of all, we have w̃k ((φ − 1)p̃ + 1) (φ − 1)p̃ −1 ∞ ∞ k=1 k=1 X X φ φ φ−ρk ((φ − 1)p̃ + 1)w̃k − φ−ρk . (φ − 1)p̃ (φ − 1)p̃ We will prove that the summation terms on the right-hand side of (62) are both convergent. First, we have φ−ρ < 1, because φ > 1. Therefore, the geometric series P∞ −ρk P −ρk converges, that is, ∞ < ∞. Next, we show φ−ρ ((φ − 1)p̃ + 1)w̃ < k=1 φ k=1 φ 1. Observe that w̃ ρ w̃ (63) φ−ρ ((φ − 1)p̃ + 1) = φ− w̃ ((φ − 1)p̃ + 1) . Moreover, ρ − w̃ φ − w̃ρ ρ − p̃) w̃ (1 − p̃) ((φ − 1)p̃ + 1) = − 1 p̃ + 1 p̃(1 − w̃ρ ) p̃(1 − w̃ρ ) ρ w̃ 1− w̃ρ p̃w̃ 1 − p̃ = . ρ 1 − w̃ρ ρ w̃ (1 1−p̃ Here, we have p̃ρw̃ , 1− ρ ∈ (0, 1) ∪ (1, ∞). As a result, since ln v < v − 1 for any w̃ v ∈ (0, 1) ∪ (1, ∞), we obtain ρ ρ p̃w̃ 1 − p̃ ρ − w̃ ((φ − 1)p̃ + 1) = ln ln φ + (1 − ) ln w̃ ρ w̃ 1 − w̃ρ ρ p̃w̃ ρ ρ p 1 − p̃ < − 1 + (1 − ) + − p̃ = 0, ρ − 1 = p̃ − w̃ ρ w̃ 1 − w̃ w̃ w̃ ρ w̃ which implies φ− w̃ ((φ − 1)p̃ + 1) < 1. Thus, by (63), φ−ρ ((φ − 1)p̃ + 1) < 1. It then follows that ∞ X w̃k (64) φ−ρk ((φ − 1)p̃ + 1) < ∞. Finally, since P∞ k=1 k=1 −ρk φ < ∞, we obtain P∞ k=1 ψk < ∞ from (62) and (64). References [1] J. P. Hespanha, P. Naghshtabrizi, and Y. Xu, “A survey of recent results in networked control systems,” Proc. IEEE, vol. 95, no. 1, pp. 138–172, 2007. [2] V. Gupta, A. F. Dana, J. P. Hespanha, R. M. Murray, and B. Hassibi, “Data transmission over networks for estimation and control,” IEEE Trans. Autom. Control, vol. 54, no. 8, pp. 1807–1819, 2009. [3] R. Alur, A. D’Innocenzo, K. H. Johansson, G. J. Pappas, and G. Weiss, “Compositional modeling and analysis of multi-hop control networks,” IEEE Trans. Autom. Control, vol. 56, no. 10, pp. 2345–2357, 2011. [4] F. Smarra, A. D’Innocenzo, and M. D. Di Benedetto, “Optimal co-design of control, scheduling and routing in multi-hop control networks,” in Proc. IEEE Conf. Dec. Contr., pp. 1960–1965, 2012. [5] A. D’Innocenzo, M. D. Di Benedetto, and E. Serra, “Fault tolerant control of multi-hop control networks,” IEEE Trans. Autom. Control, vol. 58, no. 6, pp. 1377–1389, 2013. [6] A. D’Innocenzo, F. Smarra, and M. D. Di Benedetto, “Resilient stabilization of multi-hop control networks subject to malicious attacks,” Automatica, vol. 71, pp. 1–9, 2016. CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL 35 [7] F. Smarra, M. D. D. Benedetto, and A. D’Innocenzo, “A sub-optimal method for routing redundancy design over lossy networks,” in Proc. 20th IFAC W. Congress, pp. 2604–2609, 2017. [8] S. A. Khayam and H. Radha, “Markov-based modeling of wireless local area networks,” in Proc. ACM MSWiM, pp. 100–107, 2003. [9] S. Floyd and V. Jacobson, “Random early detection gateways for congestion avoidance,” IEEE/ACM Trans. Networking, vol. 1, no. 4, pp. 397–413, 1993. [10] W. Xu, W. Trappe, Y. Zhang, and T. Wood, “The feasibility of launching and detecting jamming attacks in wireless networks,” in Proc. 6th ACM Int. Symp. Mobile Ad Hoc Network. Comput., pp. 46–57, 2005. [11] K. Pelechrinis, M. Iliofotou, and S. V. Krishnamurty, “Denial of service attacks in wireless networks: The case of jammers,” IEEE Commun. Surveys Tuts., vol. 13, no. 2, pp. 245–257, 2011. [12] S. Amin, A. A. Cárdenas, and S. S. Sastry, “Safe and secure networked control systems under Denial-of-Service attacks,” in Proc. 12th HSCC, pp. 31–45, 2009. [13] C. De Persis and P. Tesi, “Input-to-state stabilizing control under denial-of-service,” IEEE Trans. Autom. Control, vol. 60, no. 11, pp. 2930–2944, 2015. [14] S. Liu, P. X. Liu, and A. El Saddik, “A stochastic game approach to the security issue of networked control systems under jamming attacks,” J. Franklin Inst., vol. 351, no. 9, pp. 4570–4583, 2014. [15] Y. Li, L. Shi, P. Cheng, J. Chen, and D. E. Quevedo, “Jamming attacks on remote state estimation in cyber-physical systems: A game-theoretic approach,” IEEE. Trans. Autom. Control, vol. 60, no. 10, pp. 2831–2836, 2015. [16] C. De Persis and P. Tesi, “Networked control of nonlinear systems under Denial-of-Service,” Syst. Control Lett., vol. 96, pp. 124–131, 2016. [17] B. Awerbuch, R. Curtmola, D. Holmer, C. Nita-Rotaru, and H. Rubens, “ODSBR: An ondemand secure Byzantine resilient routing protocol for wireless ad hoc networks,” ACM Trans. Inf. and System Security, vol. 10, no. 4, p. 6, 2008. [18] M. M. E. A. Mahmoud and X. S. Shen, Security for Multi-hop Wireless Networks. Springer, 2014. [19] R. H. Jhaveri, S. J. Patel, and D. C. Jinwala, “DoS attacks in mobile ad hoc networks: A survey,” in Proc. ACTT, pp. 535–541, 2012. [20] A. T. Mizrak, S. Savage, and K. Marzullo, “Detecting malicious packet losses,” IEEE Trans. Parallel Distrib. Syst., vol. 20, no. 2, pp. 191–206, 2009. [21] T. Shu and M. Krunz, “Privacy-preserving and truthful detection of packet dropping attacks in wireless ad hoc networks,” IEEE Trans. Mobile Computing, vol. 14, no. 4, pp. 813–828, 2015. [22] A. A. Cárdenas, S. Amin, and S. Sastry, “Research challenges for the security of control systems,” in Proc. Conf. Hot Topics in Security, 2008. [23] H. Sandberg, S. Amin, and K. H. Johansson, “Special issue on cyberphysical security in networked control systems,” IEEE Control Syst. Mag., vol. 35, no. 1, 2015. [24] A. F. Dana, R. Gowaikar, R. Palanki, B. Hassibi, and M. Effros, “Capacity of wireless erasure networks,” IEEE Trans. Inf. Theory, vol. 52, no. 3, pp. 789–804, 2006. [25] D. E. Quevedo, A. Ahlén, and K. H. Johansson, “State estimation over sensor networks with correlated wireless fading channels,” IEEE Trans. Autom. Control, vol. 58, no. 3, pp. 581–593, 2013. [26] A. Cetinkaya, H. Ishii, and T. Hayakawa, “Networked control under random and malicious packet losses,” IEEE Trans. Autom. Control, vol. 62, no. 5, pp. 2434–2449, 2017. [27] A. Cetinkaya, H. Ishii, and T. Hayakawa, “Random and malicious packet transmission failures on multi-hop channels in networked control systems,” in Proc. IFAC NecSys, pp. 49–54, 2016. [28] D. E. Quevedo and D. Nesic, “Input-to-state stability of packetized predictive control over unreliable networks affected by packet-dropouts,” IEEE Trans. Autom. Control, vol. 56, no. 2, pp. 370–375, 2011. [29] A. Cetinkaya, H. Ishii, and T. Hayakawa, “Event-triggered output feedback control resilient against jamming attacks and random packet losses,” in Proc. IFAC NecSys, pp. 270–275, 2015. [30] B. D. O. Anderson, “From Wiener to hidden Markov models,” IEEE Control Syst. Mag., vol. 19, no. 3, pp. 41–51, 1999. 36 CHARACTERIZATION OF FAILURES IN MULTI-HOP NETWORKED CONTROL [31] O. Cappé, E. Moulines, and T. Rydén, Inference in Hidden Markov Models. Springer, 2009. [32] M. Vidyasagar, Hidden Markov Processes: Theory and Applications to Biology. Princeton University Press, 2014. [33] P. Billingsley, Probability and Measure. Wiley, 2012. [34] P. Sadeghi, R. A. Kennedy, P. B. Rapajic, and R. Shams, “Finite-state Markov modeling of fading channels,” IEEE Signal Process. Mag., vol. 25, no. 5, pp. 57–80, 2008. [35] M. Ellis, D. P. Pezaros, T. Kypraios, and C. Perkins, “A two-level Markov model for packet loss in UDP/IP-based real-time video applications targeting residential users,” Comput. Netw., vol. 70, no. 9, pp. 384–399, 2014. [36] D. Medhi and K. Ramasamy, Network Routing: Algorithms, Protocols, and Architectures. Morgan Kaufmann, 2007. [37] U. Madhow, Fundamentals of Digital Communication. Cambridge University Press, 2008.	3cs.SY
Two provably consistent divide and conquer clustering algorithms for large networks arXiv:1708.05573v1 [stat.ML] 18 Aug 2017 Soumendu Sundar Mukherjee1 , Purnamrita Sarkar2 , and Peter J. Bickel1 1 2 Department of Statistics, University of California, Berkeley Department of Statistics and Data Sciences, University of Texas, Austin August 21, 2017 Abstract In this article, we advance divide-and-conquer strategies for solving the community detection problem in networks. We propose two algorithms which perform clustering on a number of small subgraphs and finally patches the results into a single clustering. The main advantage of these algorithms is that they bring down significantly the computational cost of traditional algorithms, including spectral clustering, semi-definite programs, modularity based methods, likelihood based methods etc., without losing on accuracy and even improving accuracy at times. These algorithms are also, by nature, parallelizable. Thus, exploiting the facts that most traditional algorithms are accurate and the corresponding optimization problems are much simpler in small problems, our divide-and-conquer methods provide an omnibus recipe for scaling traditional algorithms up to large networks. We prove consistency of these algorithms under various subgraph selection procedures and perform extensive simulations and real-data analysis to understand the advantages of the divide-and-conquer approach in various settings. Contents 0 Glossary of notation 2 1 Introduction 3 2 Two divide and conquer algorithms 2.1 PACE: an averaging algorithm . . . 2.2 GALE: a sequential algorithm . . . 2.3 Remarks on sampling schemes . . . 2.4 Beyond community detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 6 8 10 10 3 Main results 11 3.1 Results on PACE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.2 Results on GALE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 4 Simulations and real-data analysis 17 4.1 Simulations: comparison against traditional algorithms . . . . . . . . . . . . . . . . . . . . . . 17 4.2 Real data analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 1 5 Discussion 20 6 Proofs 21 6.1 Results on PACE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 6.2 Results on GALE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 7 Acknowledgments 30 A Details of DGCluster 33 B Some auxiliary results B.1 Analysis of PACE under stochastic block model . . . B.1.1 Community sizes in random m-subgraphs . . B.1.2 Community sizes in 1-hop ego neighborhoods B.1.3 Analysis of adjacency spectral clustering . . . B.1.4 Analysis of SDP . . . . . . . . . . . . . . . . B.2 Proofs of some results for GALE . . . . . . . . . . . . 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 35 36 36 37 39 39 Glossary of notation For the convenience of the reader, we collect here some of the more frequently used notations used in the paper, and provide a summarizing phrase for each, as well as the page number at which the notation first appears. T S, Sl wl,i,j Nij τ Ĉ Ĉproj M Ẑ` Ẑ (`) nk πk πmin δc (σ, σ 0 ) δ(Z, Z 0 ) δ̃(C, C 0 ) M(Z 0 , Z) (`) yi π (S) nk Yab number of subgraphs used a randomly chosen subgraph weights used in w-PACE the number of times i and j appear in all subgraphs S1 , . . . , ST threshold for Nij , a tuning parameter in PACE estimated clustering matrix returned by PACE randomly projected version of Ĉ, to be used for recovering an estimate of Z confusion matrix between two clusterings, used by Match cluster membership matrix of subgraph S` before alignment cluster membership matrix of subgraph S` after alignment number of nodes from cluster k nk /n, i.e. proportion of nodes from cluster k mink πk , similarly πmax normalized mismatch between two clusterings σ, σ 0 (σ, σ 0 : {1, . . . , n} → {1, . . . , K}) normalized mismatch between two clusterings Z, Z 0 (Z, Z 0 ∈ {0, 1}n×K ) normalized between two clustering matrices C, C 0 nδ(Z, Z 0 ) 1 if i ∈ S` and 0 otherwise the vector (π1 , . . . , πK ) number of nodes from cluster k in a subgraph S \|Sa ∩ Sb \| 2 6 6 6 6 6 6 7 8 9 9 9 9 9 11 11 11 14 14 16 26 28 1 Introduction Community detection, also known as community extraction or network clustering, is a central problem in network inference. A wide variety of real world problems, ranging from finding protein protein complexes in gene networks (Guruharsha et al., 2011) to studying the consequence of social cliques on adolescent behavioral patterns (Hollingshead, 1949) depend on detecting and analyzing communities in networks. In most of these problems, one observes co-occurrences or interactions between pairs of entities, i.e. pairwise edges and possibly additional node attributes. The goal is to infer the hidden community memberships. Many of these real world networks are massive, and hence it is crucial to develop and analyze scalable algorithms for community detection. We will first talk about methodology that uses only the network connections for inference. These can be divided into roughly two types. The first type consists of methods which are derived independently of any model assumption. These typically involve the formulation of global optimization problem; examples include normalized cuts (Shi and Malik, 2000), Spectral Clustering (Ng et al., 2002), etc. On the other end, Statisticians often devise techniques under model assumptions. The simplest statistical model for networks with communities is the Stochastic Blockmodel (SBM) (Holland et al., 1983). The key idea in a SBM is to enforce stochastic equivalence, i.e. two nodes in the same latent community have identical probabilities of connection to all nodes in the network. There are many extensions of SBM. The Degree Corrected Stochastic Blockmodel (Karrer and Newman, 2011) allow one to model varied degrees in the same community, whereas a standard SBM does not. Mixed membership blockmodels (Airoldi et al., 2009) allow a node to belong to multiple communities, whereas in a SBM, a node can belong to exactly one cluster. For an SBM generating a network with n nodes and K communities, one has a hidden community/cluster membership matrix Z ∈ {0, 1}n×K , where Zik = 1 if node i is in community k ∈ [K]. Given these memberships, the link formation probabilities are given as P(Aij = 1 \| Zik = 1, Zjk0 = 1) = Bkk0 , where B is a K × K symmetric parameter matrix of probabilities. The elements of B may decay to zero as n grows to infinity, to model sparse networks. Typically the goal is to estimate the latent memberships consistently. A method outputting an estimate Ẑ is called strongly consistent if P(Ẑ = ZΠ) → 1 for some K × K permutation matrix Π, as n → ∞. A weaker notion of consistency is when the fraction of misclustered nodes goes to zero as n goes to infinity. Typically most of the consistency results are derived where average degree of the network grows faster than the logarithm of n. This is is often called the semi-dense regime. When average degree is bounded, we are in the sparse regime. There are a plethora of algorithms for community detection. These include likelihood-based methods (Amini et al., 2013), modularity based methods (Snijders and Nowicki, 1997; Newman and Girvan, 2004; Bickel and Chen, 2009), spectral methods (Rohe et al., 2011), semi-definite programming (SDP) based approaches (Cai et al., 2015) etc. Among these, spectral methods are scalable since the main bottleneck is computing top K eigenvectors of a large and often sparse matrix. While the theoretical guarantees of Spectral Clustering are typically proven in the semi-dense regime (McSherry, 2001; Rohe et al., 2011; Lei and Rinaldo, 2015), a regularized version of it has been shown to perform better than a random predictor for sparse networks (Le et al., 2015). Profile likelihood methods (Bickel and Chen, 2009) involve greedy search over all possible membership matrices, which makes them computationally expensive. Semidefinite programs are robust to outliers (Cai et al., 2015) and are shown to be strongly consistent in the dense regime (Amini and Levina, 2014) and yield a small but non-vanishing error in the sparse regime (Guédon 3 and Vershynin, 2015). However, semi definite programs are slow and typically only scale to thousands of nodes, not millions of nodes. Methods like spectral clustering on geodesic distances (Bhattacharyya and Bickel, 2014) are provably consistent in the semi-dense case, and can give a small error in sparse cases. However, it requires computing all pairs of shortest paths between all nodes, which can pose serious problems for both computation and storage for very large graphs. Monte Carlo methods (Snijders and Nowicki, 1997; Nowicki and Snijders, 2001), which are popular tools in Bayesian frameworks, are typically not scalable. More scalable alternatives such as variational methods (Gopalan and Blei, 2013) do not have provable guarantees for consistency, and often suffer from bad local minima. So far we have discussed community detection methods which only look at the network connections, and not node attributes, which are often also available and may possess useful information on the community structure (see, e.g., Newman and Clauset (2015)). There are extensions of the methods mentioned earlier which accommodate node attributes, e.g., modularity based (Zhang et al., 2015), spectral (Binkiewicz et al., 2014), SDP based (Yan and Sarkar, 2016), etc. These methods come with theoretical guarantees and have good performance in moderately sized networks. While existing Bayesian methods (McAuley and Leskovec, 2012; Newman and Clauset, 2015; Xu et al., 2012) are more amenable to incorporating covariates in the inference procedure, they often are computationally expensive and lack rigorous theoretical guarantees. While the above mentioned array of algorithms are diverse and each has unique aspects, in order to scale them to very large datasets, one has to apply different computational tools tailored to different algorithmic settings. While stochastic variational updates may be suitable to scale Bayesian methods, pseudo likelihood methods are better optimized using row sums of edges inside different node blocks. In this article, we propose a divide and conquer approach to community detection. The idea is to apply a community detection method on small subgraphs of a large graph, and somehow stitch the results together. If we could achieve this, we would be able to scale up any community detection method (which may involve covariates as well as the network structure) that is computationally feasible on small graphs, but is difficult to execute on very large networks. This would be especially useful for computationally expensive community detection methods (such as SDPs, modularity based methods, Bayesian methods etc.). Another possible advantage concerns variational likelihood methods (such as mean-field) with a large number (depending on n) of local parameters, which typically have an optimization landscape riddled with local minima. For smaller graphs there are less parameters to fit, and the optimization problem often becomes easier. Clearly, the principal difficulty in doing this, is matching the possibly conflicting label assignments from different subgraphs (see Figure 1(a) for an example). This immediately rules out a simple-minded averaging of estimates ẐS of cluster membership matrices ZS , for various subgraphs S, as a viable stitching method. In this regard, we propose two different stitching algorithms. The first is called Piecewise Averaged Community Estimation (PACE); in which we focus on estimating the clustering matrix C = ZZ > , which is labeling invariant, since the (i, j)-th element of this matrix being one simply means that nodes i and j belong to the same cluster, whereas the value zero means i and j belongs to two different clusters. Thus we first compute estimates of ZS ZS> for various subgraphs S and then average over these matrices to obtain an estimate Ĉ of C. Finally we apply some computationally cheap clustering algorithm like approximate K-means, DGCluster† , spectral clustering etc. on Ĉ to recover an estimate of Z. We also propose another algorithm called Global Alignment of Local Estimates (GALE), where we first take a sequence of subgraphs, such that any two consecutive subgraphs on this sequence have a large intersection, and then traverse through this sequence, aligning the clustering based on a subgraph with an † This is a greedy algorithm we propose in this article, see Appendix A. 4 (a) (b) Figure 1: (a) Conflicting label assignments on two subgraphs, on one labels are denoted by 0/1, on the other by +/−, (b) onion neighborhood: the yellow vertices and the solid edges between them constitute the ego network of the root vertex colored red. The green and yellow vertices together with the solid edges between them constitute the 2-hop onion neighborhood of the root averaged clustering of the union of all its predecessor subgraphs in the sequence, which have already been aligned. The alignment is done via an algorithm called Match which identifies the right permutation to align two clusterings on two subgraphs by computing the confusion matrix of these two clusterings restricted to the intersection of the two subgraphs. Whereas a naive approach would entail searching through all K permutations, Match finds the right permutation in K log K time. Once this alignment step is complete, we get an averaged clustering of the union of all subgraphs (which covers all the vertices). By design GALE works with estimates of cluster membership matrices ZS directly to output an estimate of Z, and thus, unlike PACE, avoids the extra overhead of recovering such an estimate from Ĉ. The rest of the paper is organized as follows. In Section 2 we describe our algorithms. In Section 3 we state our main results and some applications. Section 4 contains simulations and real data experiments. In Section 6 we provide proofs of our main results, while relegating some of the details to the Appendix, Section B. Finally, in Section 5 we conclude with some discussions and directions for future work. 2 Two divide and conquer algorithms As we discussed in the introduction, the main issue with divide and conquer algorithms for clustering is that one has to somehow match up various potentially conflicting label assignments. In this vein we propose two algorithms to accomplish this task. Both algorithms first compute the clustering on small patches of a network; these patches can be induced subgraph of a random subsample of nodes, or neighborhoods. However, the stitching procedures are different. 5 2.1 PACE: an averaging algorithm Suppose A is the adjacency matrix of a network with true cluster membership of its nodes being given by the n × K matrix Z where there are K clusters. Set C = ZZ > to be the clustering matrix whose (i, j)-th entry is the indicator of whether nodes i, j belong to the same cluster. Given A we will perform a local clustering algorithm to obtain an estimate of C, from which an estimate Ẑ of the cluster memberships may be reconstructed. Algorithm 1 PACE: Piecewise Averaged Community Estimation 1: Subgraph selection: Fix a positive integer threshold m? for minimum required subgraph size. Fix another positive integer T , that will be the number of subgraphs we will sample. Given A, choose T subsets S1 , . . . , ST of the nodes with \|S` \| ≥ m? by some procedure, e.g., select m ≥ m? nodes at random, or pick h-hop neighborhoods of vertices in G, or ego-neighborhood of vertices. By AS` denote the adjacency matrix of the network induced by S` . 2: Clustering on subgraphs: Perform any of the standard clustering algorithms like Profile Likelihood (PL), Mean Field Likelihood (MFL), Spectral Clustering (SC), Semi-Definite Programming (SDP) etc. on each of these T subgraphs which have size at least m? to obtain estimated clustering matrices Ĉ (S` ) = Ĉ (`) . For the rest of the subgraphs set Ĉ (`) ≡ 0. Extend Ĉ (`) to an n × n matrix by setting (`) Ĉij = 0 if at least one of i, j was not selected in S` . Let’s denote the resulting matrix again by Ĉ (`) . 3: Patching up: Let yij denote the indicator of the event that both i, j were selected in S` . Set Nij = PT (`) `=1 yij . Define the combined estimator Ĉ = Ĉτ by (`) Ĉij = Ĉτ,ij = 1{Nij ≥τ } (`) (`) `=1 yij Ĉij PT Nij = 1{Nij ≥τ } (`) `=1 Ĉij PT Nij . (1) Here 1 ≤ τ ≤ T is an integer tuning parameter. We will call Ĉτ as Piecewise Averaged Community Estimator (also abbreviated as PACE). The τ parameter in PACE seems to reduce variance in estimation quality as it discards information from less credible sources — if a pair of nodes has appeared in only a few subgraphs, we do not trust what the patching has to say about them. Setting τ = θE(Nij ) seems to work well in practice (this choice is also justified by our theory). A slight variant of Algorithm 1 is where we allow subgraph and/or node-pair specific weights w`,i,j in the computation of the final estimate, i.e. P P (`) (`) (`) 1{Nij ≥τ } T`=1 w`,i,j yij Ĉij 1{Nij ≥τ } T`=1 w`,i,j Ĉij Ĉij = Ĉτ,ij = = , (2) Nij Nij P (`) where Nij now equals T`=1 w`,i,j yij . We may call this estimator w-PACE standing for weighted-PACE. If the weights are all equal, w-PACE becomes equivalent to ordinary PACE. There are natural nontrivial choices, including (i) w`,i,j = \|S` \|, which will place more weight to estimates based on large subgraphs, (ii) w`,i,j = degS` (i) + degS` (j), where degS (u) denotes the degree of node u in subgraph S (this will put more weight on pairs which have high degree in S` ). 6 The first prescription above is intimately related to the following sampling scheme for ordinary PACE: pick subgraphs with probability proportional to their sizes. For instance, in Section 4.2 we analyze the political blog data of Adamic and Glance (2005) where neighborhood subgraphs are chosen by selecting their roots with probability proportional to degree. In real world applications, it might make more sense to choose these weights based on domain knowledge (for instance, it may be that certain subnetworks are known to be important). Another (minor) advantage of having weights is that when T = 1 and \|S1 \| = n, we have Nij = w1,i,j and so if w1,i,j ≥ τ , then (1) Ĉij = Ĉij , i.e. w-PACE becomes the estimator based on the full graph. This is for example true with w`,i,j = \|S` \|, because τ is typically much smaller than n. However, ordinary PACE lacks this property unless τ = 1, in fact, with τ > 1, the estimate returned by PACE is identically 0. Anyway, in what follows, we will stick with ordinary PACE because of its simplicity. Before we discuss how to reconstruct an estimate Ẑ of Z from Ĉ, let us note that we may obtain a binary matrix Ĉη by thresholding Ĉ at some level η (for example, η = 1/2): Ĉη := [Ĉ > η]. This thresholding does not change consistency properties (see Lemma B.1). Looking at a plot of this matrix gives a good visual representation of the community structure. In what follows, we work with unthresholded Ĉ. Reconstruction of Ẑ: How do we actually reconstruct Ẑ from Ĉ? The key is to note that members of the same community have identical rows in C and that, thanks to PACE, we have gotten hold of a consistent estimate of C. Thus we may use any clustering algorithm on the rows of Ĉ to recover the community memberships. Another option would be to run spectral clustering on the matrix Ĉ itself. However, as the rows of Ĉ are n-vectors, most clustering algorithms will typically have a running time of O(n3 )† at best. Indeed, the main computational bottleneck of any distance based clustering algorithm in a high dimensional situation like the present one, is computing dij = kĈi? −Ĉj? k which takes O(n) bit operations. However, since we have gotten a good estimate of C, we can project the rows of Ĉ onto lower dimensions, without distorting the distances too much. The famous Johnson-Lindenstrauss Lemma for random projections says that by projecting onto Ω(log n/2 ) dimensions, one can keep, with probability at least 1 − O(1/n), the distances between projected vectors within a factor of (1 ± ) of the true distances. Choosing as inverse polylog(n) we need to project onto polylog(n) dimensions and this would then readily bring the computational cost of any distance based algorithm down from O(n3 ) to O(n2 polylog(n)). Following the discussion of the above paragraph, we first do a random projection of the rows of Ĉ onto s (= polylog(n)) dimensions and then apply a (distance-based) clustering algorithm. Algorithm 2 Recovering Ẑ from Ĉ: random projection followed by distance based clustering 1: Select a dimension s for random projection. Let Cluster(·, K) be a clustering algorithm that operates on the rows of its first argument and outputs K clusters. √ 2: Ĉproj ← ĈR/ s, where R is a standard Gaussian matrix of dimensions n × s. 3: Ẑ ← Cluster(Ĉproj , K). † A word on asymptotic notation: in addition to the standard ‘O’, ‘Ω’, ‘Θ’, ‘o’ notations, and their probabilistic counterparts, we will use the following less standard alternatives to clean up the presentation at places: (i) A B to mean A = Θ(B), and (ii) A B to mean B = o(A). We will also sometimes use a ‘tilde’ over the standard notations to hide polylogarithmic factors. 7 As for Cluster(·, K), we may use approximate K-means or any other distance based clustering algorithm, e.g., DGCluster(Ĉ, Ĉproj , K), a greedy algorithm presented in Appendix A as Algorithm 5. 2.2 GALE: a sequential algorithm First we introduce a simple algorithm for computing the best permutation to align labels of one clustering (Z1 ) to another (Z2 ) of the same set of nodes (with fixed ordering) in a set S. The idea is to first compute the confusion matrix between two clusterings. Note that if the two labelings each have low error with respect to some unknown true labeling, then the confusion matrix will be close to a diagonal matrix up to permutations. The following algorithm below essentially finds a best permutation to align one clustering to another. Algorithm 3 Match: An algorithm for aligning two clusterings of the same set S of nodes. Input Z1 , Z2 ∈ {0, 1}\|S\|×K . Compute K × K confusion matrix M = Z1> Z2 . Set Π ← 0K×K . 2: while there are no rows/columns left do 1: (a) Find i, j, such that Mij = kM k∞ (a tie can be broken arbitrarily). (b) Set Πij = 1. (c) Replace the i-th row and j-th columns in M with −1. 3: Return the permutation matrix Π. Remark 2.1. One can also compute the optimal permutation by searching through all K! permutations of the labels and picking the one which gives smallest mismatch between the two; but Match brings the dependence on K down from exponential to quadratic. Note that if one of the clusterings are poor, then Match may not retrieve the optimal permutation. However, our goal is to cluster many subgraphs using an algorithm which has good accuracy with high probability (but may be computationally intensive, e.g., Profile Likelihood or Semidefinite Programming) and then use the intersections between the subgraphs to align one to another. As we shall show later, as long as there are enough members from each community in S, the simple algorithm sketched above suffices to find an optimal permutation. Now we present our sequential algorithm which aligns labelings across different subgraphs. The idea is to first fix the indexing of the nodes; cluster the subgraphs (possibly with a parallel implementation) using some algorithm, and then align the clusterings along a sequence of subgraphs. To make things precise, we make the following definition. Definition 2.1. Let Sm,T = (V, E) (vertex set V and edge set E) denote a “super-graph of subgraphs” where each subgraph is a node, and two nodes are connectedlif the m corresponding subgraphs have a substantial m2 overlap. For random m-subgraphs, this threshold is m1 := 2n . Define a traversal through a spanning tree T of Sm,T as a sequence x1 , . . . , xJ , xj ∈ [T ], T ≤ J ≤ 2T , covering all the vertices, such that along the sequence xi and xi+1 are adjacent, for 1 ≤ i ≤ J − 1 (i.e. a traversal is just a walk on T of length at most 2T − 1 passing through each vertex at least once). After we construct a traversal, we travel through this traversal such that at any step, we align the current subgraph’s labels using the Match algorithm (Algorithm 3) on its intersection with the union of the previously aligned subgraphs. At the end, all subgraph labellings are aligned to the labeling of the starting subgraph. Now we can simply take an average or a majority vote between these. 8 Algorithm 4 GALE: Global Alignment of Local Estimates. Input adjacency matrix A, parameters m, T, τi , i ∈ [T ], a base algorithm A (e.g., PL, MFL, SC, SDP etc. ) 1: Subgraph selection: Given A, choose T subsets S1 , . . . , ST of the nodes by some procedure, e.g., select them at random, or select T random nodes and then pick their h-hop neighborhoods. By AS` denote the adjacency matrix of the network induced by S` . 2: Clustering on subgraphs: Perform algorithm A on each of these T subgraphs to obtain estimated cluster membership matrices Ẑ` = ẐS` . Extend Ẑ` to a n × K matrix by setting (Ẑ` )jk = 0 for all k ∈ [K] if j ∈ / S` . Traversal of subgraphs: Construct a traversal Sx1 , . . . , SxJ through the T subgraphs. Initial estimate of Z: Ẑ ← Ẑx1 . Also set Ẑ (x1 ) := Ẑx1 . 5: Sequential label aligning: For subgraph Sxi on the traversal (i = 1, . . . , J), if xi has not been visited yet, 3: 4: (a) Compute the overlap between the current subgraph with all subgraphs previously visited, i.e. let S = Sxi ∩ ∪i−1 `=1 Sx` . (b) Compute the best permutation to match the clustering Ẑxi , Ẑ on this set S, i.e. compute Π̂i = Match(Ẑxi S , Ẑ S ). (c) Permute the labels of the nodes of Sxi to get an aligned cluster membership matrix Ẑ (xi ) ← Ẑxi Π̂. P (d) Update Ẑjk ← (x ) `∈{x1 ,...,xi } P Ẑjk ` 1{j∈Sx `∈{x1 ,...,xi } 1{j∈Sx ` } ` } 1{P`∈{x 1 ,...,xi } 1{j∈Sx ` } >τi } , for some threshold τi . (e) Mark Sxi as visited. Implementation details: Constructing a traversal of the subgraphs can be done using a depth first search of the super-graph Sm,T of subgraphs. For our implementation, we start with a large enough subgraph (the parent), pick another subgraph that has a large overlap with it (the child), align it and note that this subgraph has been visited. Now recursively find another unvisited child of the current subgraph, and so on. It is possible that a particular path did not cover all vertices, and hence it is ideal to estimate clusterings with multiple traversals with different starting subgraphs and then align all these clusterings, and take an average. This is what we do for real networks. We also note that at any step, if we find a poorly clustered subgraph, then this can give a bad permutation which may deteriorate the quality of aligning the subsequent subgraphs on the path. In order to avoid this we use a self validation routine. Let S be intersection of current subgraph with union of the previously visited subgraphs. After aligning the current subgraph’s clustering, we compute the classification accuracy of the current labeling of S with the previous labeling of S. If this accuracy is large enough, we use this subgraph, and if not we move to the next subgraph on the path. For implementation, we use a threshold of 0.55. Computational time and storage: The main computational bottleneck in GALE is in building a traversal through the random subgraphs. Let ηm,T be the time for computing the clusterings for T subgraphs in parallel. A naive implementation would require computing intersections between all T2 pairs of mp subsets. As we will show in our theoretical analysis, we take m = ω( n/πmin ), where πmin := mink πk (here πk := nk /n, where nk is the size of the k-th cluster) and T = Ω̃(n/m). Taking πmin = Θ(1), the computation of intersections takes O(T 2 m) = Õ(n3/2 ) time. Further, a naive comparison for computing subsets similar or close to a given one would require T log T time for each subset leading to T 2 log T = Õ(n) computation. However, for building a traversal one only needs access to subsets with large overlap with a given subset, 9 which is a classic example of nearest neighbor search in Computer Science. One such method is the widely used and theoretically analyzed technique of Locality Sensitive Hashing (LSH). A hash function maps any data object of an arbitrary size to another object of a fixed size. In our case, we map the characteristic vector of a subset to a number. The idea of LSH is to compute hash functions for two subsets A and B such that the two functions are the same with high probability if A and B are “close”. In fact, the amount of overlap normalized by m is simply the cosine similarity between the characteristic vectors χ(A) and χ(B) of the two subsets, for which efficient hashing schemes h : {0, 1}n → {0, 1} using random projections exist Charikar (2002), with P(h(A) = h(B)) = 1 − arccos(χ(A)T χ(B)/m)/π. For LSH schemes, one needs to build L := T ρ hash tables, for some ρ < 1, that governs the approximation quality. In each hash table, a “bucket” corresponding to an index stores all subsets which have been hashed to this index. For any query point, one evaluates O(L) hash functions and examines O(L) subsets hashed to those buckets in the respective hash tables. Now for these subsets, the distance is computed exactly. The 1+ρ preprocessing time is O(T 1+ρ ) = Õ(n 2 ), with storage being O(T 1+ρ + T m) = Õ(n), and total query time being O(T 1+ρ m) = Õ(n1+ρ/2 ). This brings down the running time added to the algorithm specific ηm,T from sub-quadratic time Õ(n3/2 ) to nearly linear time, i.e. Õ(n1+ρ/2 ). Thus, for other nearly linear time clustering algorithms GALE may not lead to computational savings. However, for algorithms like Profile Likelihood or SDP which are well known to be computationally intensive, GALE can lead to a significant computational saving without requiring a lot of storage. 2.3 Remarks on sampling schemes With PACE we have mainly used random m-subgraphs, h-hop neighborhoods, and onion neighborhoods, but many other subgraph sampling schemes are possible. For instance, choosing roots of hop neighborhoods with probability proportional to degree, or sampling roots from high degree nodes (we have done this in our analysis of the political blog data, in Section 4.2). As discussed earlier, this weighted sampling scheme is related to w-PACE. A natural question regarding h-hop neighborhoods is how many hops to use. While we do not have a theory for this yet, because of “small world phenomenon” we expect not to need many hops; typically in moderately sparse networks, 2-3 hops should be enough. Although, an adaptive procedure (e.g., cross-validation type) for choosing h would be welcome. Also, since neighborhood size increases exponentially with hop size, an alternative to choosing full hop-neighborhoods is to choose a smaller hopneighborhood and then add some (but not all) randomly chosen neighbors of the already chosen vertices. Other possibilities include sampling a certain proportion of edges at random, and the consider the subgraph induced by the participating nodes. We leave all these possibilities for future work. We have analyzed GALE under the random sampling scheme. For any other scheme, one will have to understand the behavior of the intersection of two samples or neighborhoods. For example, if one takes hhop neighborhoods, for sparse graphs, each neighborhood predominantly has nodes from mainly one cluster. Hence GALE often suffers with this scheme. We show this empirically in Section 4, where GALE’s accuracy is much improved under a random m-subgraph sampling scheme. 2.4 Beyond community detection The ideas behind PACE and GALE are not restricted to community detection and can be modified for application in other interesting problems, including general clustering problems, co-clustering problems (Rohe et al. (2016)), mixed membership models, among others (these will be discussed in an upcoming article). In fact, Mackey et al. (2015) took a similar divide and conquer approach for matrix completion. 10 3 Main results In this section we will state and discuss our main results on PACE and GALE, along with a few applications. 3.1 Results on PACE Let σ and σ 0 be two clusterings (of n objects into K clusters), usually their discrepancy is measured by n δc (σ, σ 0 ) = inf ξ∈SK 1 1X 1{σ(i)6=ξ(σ0 (i))} = inf kξ(σ) − σ 0 k0 , n n ξ∈SK i=1 where SK is the permutation group on [K]. If Z, Z 0 are the corresponding n × K binary matrices, then a related measure of discrepancy between the two clusterings is δ(Z, Z 0 ) = inf Q perm. n1 kZQ − Z 0 k0 . It is easy to see that δ(Z, Z 0 ) = 2δc (σ, σ 0 ). (To elaborate, let Qξ be the permutation matrix corresponding to the 0 , i.e. k(ZQ ) −Z 0 k = permutation ξ, i.e. Qij = 1{ξ(i)=j)} . Then ξ(σ(i) 6= σ 0 (i), if and only if (ZQξ )i? 6= Zi? ξ i? i? 0 2.) For our purposes, however, a more useful measure of discrepancy would be the normalized Frobenius squared distance between the corresponding clustering matrices C = ZZ > and C 0 = Z 0 Z 0> , i.e. δ̃(C, C 0 ) = 1 kC − C 0 k2F . n2 Now we compare these two notions of discrepancies. Proposition 3.1. We have δ̃(C, C 0 ) ≤ 4δ(Z, Z 0 ) = 8δc (σ, σ 0 ) Incidentally, if the cluster sizes are equal, i.e. n/K, then one can show that δ̃(C, C 0 ) ≤ 4 8 δ(Z, Z 0 ) = δc (σ, σ 0 ). K K Although we do not have a lower bound on δ̃(C, C 0 ) in terms of δ(Z, Z 0 ), Lemma A.1 of Tang et al. (2013) gives us (with X = Z, Y = Z 0 ) that there exists an orthogonal matrix O such that p p √ kC − C 0 k( KkCk + KkC 0 k) 2 KnkC − C 0 k 0 kZO − Z kF ≤ ≤ , λmin (C) nmin (Z) where we used the fact that kCk = λmax (C) = nmax (Z) ≤ n. The caveat here is that the matrix O need not be a permutation matrix. To prove consistency of PACE we have to assume that the clustering algorithm A we use has some consistency properties. For example, it will suffice to assume that for a randomly chosen subgraph S (under our subgraph selection procedure), Eδ(ẐS , ZS )† is small. The following is our main result on the expected misclustering rate of PACE. Theorem 3.1 (Expected misclustering rate of PACE). Let S be a randomly chosen subgraph according to our sampling scheme. Let πmax = maxk πk . We have Eδ̃(Ĉ, C) ≤ T × EkĈ (S) − C (S) k2F + πmax × max P(Nij < τ ), i,j τ n2 where EkĈ (S) − C (S) k2F ≤ n2 P(\|S\| < m? ) + 4E\|S\|2 δ(ẐS , ZS )1(\|S\|≥m? ) . † The expectation is taken over both the randomness in the graph and the randomness of the sampling mechanism. 11 (3) The first term in (3) essentially measures the performance of the clustering algorithm we use on a randomly chosen subgraph. The second term measures how well we have covered the full graph by the chosen subgraphs, and only depend on the subgraph selection procedure. The effect of the algorithm we use is felt through the first term only. We can now specialize Theorem 3.1 to various subgraph selection schemes. First, we consider randomly chosen m-subgraphs, which is an easy corollary. Corollary 3.1 (Subgraphs are induced by m ≥ m? randomly chosen nodes). Let p = and τ = θT p. We have m 2 Eδ̃(Ĉ, C) ≤ × Eδ̃(Ĉ (S) , C (S) ) + πmax × e−(1−θ) T p/2 . θ(m − 1) m(m−1) n(n−1) , 0<θ<1 (4) m Notice that the constant θ(m−1) in (4) can be made as close to 1 as one desires, which means that the above bound is essentially optimal. Full h-hop neighborhood subgraphs are much harder to analyze and will not be pursued here. However, ego networks, which are 1-hop neighborhoods minus the root node (see Figure 1(b)), are easy to deal with. One can also extend our analysis to h-hop onion neighborhoods which are recursively defined as follows: O1 (v) = S1 (v) is just the ego network of vertex v; in general, the h-th shell Sh (v) := ]u∈Sh−1 (v) [O1 (u) \ Oh−1 (v) ∪ {v}], and Oh (v) = Oh−1 (v) ] Sh (v), where the operation ] denotes superposition of networks. Here, for ease of exposition, we choose to work with ego networks (1-hop onion neighborhoods). Corollary 3.2 (Ego neighborhoods under a stochastic block model). Let B# = max Bab , B? = min Bab and 2 † ? τ = TB 4 and m? ≤ (n − 1)B? /2. Let θ > 1. Then Eδ̃(Ĉ, C) ≤ 2 16(1 + θ)2 B# B?2 × Eδ(ẐS , ZS )1(\|S\|≥m? ) + ∆, (5) where " ! 2 # θ nB# 4 nB?2 4 nB?4 T B?2 ∆ ≤ 2 × exp − + 2 2 × exp − + πmax × 2 exp − + exp − . 2 B? 16B# n B? 6 16 16B# We will now use existing consistency results on several clustering algorithms A, in conjunction with the above bounds to see what conditions (i.e. conditions on the model parameters, and m, T etc.) are required for PACE to be consistent. We first consider (1 + )-approximate adjacency spectral clustering (ASP) of Lei and Rinaldo (2015) as A. We will use stochastic block model as the generative model and for simplicity will assume that the link probability matrix has the following simple form B = αn [λI + (1 − λ)11> ] with 0 < λ < 1. (6) We now quote a slightly modified version of Corollary 3.2 of Lei and Rinaldo (2015) for this model. Lemma 3.1 (Lei and Rinaldo (2015)). Let c0 , , r > 0. Consider an adjacency matrix A generated from the simple block model (6) where αn ≥ c0 log n/n. If Ẑ is the output of the (1 + )-approximate adjacency spectral clustering algorithm applied to A, then there exists an absolute constant c = c(c0 , r) such that with probability at least 1 − n−r , 1 K πmax −1 δ(Ẑ, Z) ≤ min c (2 + ) 2 (7) 2 n, 1 . 2 λ αn πmin † Actually we can allow τ = θ0 T B?2 , for any 0 < θ0 < 1. As the resulting bound involves complicated constants depending on θ and does not add anything extra as to the nature of the bounds, we have chosen to work with a particular θ0 (= 1/4) to 2 ease our exposition. With a general θ0 , the constant multiplier in the first term in (5) will be 4 (1+θ) , instead of 16(1 + θ)2 . θ0 0 12 Corollary 3.3 ((1 + )-approximate adjacency spectral clustering with random m-subgraphs). Assume the setting of Lemma 3.1. Let r, r0 > 0. We have 2K 8m K πmax 1 2 −1 Cn,m,r0 , 1 + r0 + r + πmax × e−(1−θ) T p/2 . (8) Eδ̃(Ĉ, C) ≤ × min c (2 + ) 2 2 θ(m − 1) λ αn πmin m m n Here the quantity Cn,m,r0 → 1, if 2 m πmin πmax log n → ∞. The proof of Corollary 3.3 follows from Corollary 3.1 and an estimate for Eδ(ẐS , ZS ) given in (33), which max = o(1), is obtained using Lemma 3.1. In order for the first term of (8) to go to zero we need λ2 Kπ mα π 2 n min Kπmax . 2 λ2 αn πmin 1 K) 2 i.e. m Thus for balanced block sizes (i.e. πmax , πmin we need to have m λK 2 α . So, n qualitatively, for large K or small αn or a small separation between the blocks, m has to be large, which is natural to expect. In particular, for fixed K and λ, this shows that we need subgraphs of size m nd−1 n , n2 n2 and T m2 many of them to achieve consistency (here the average degree dn nαn ). Let T = m2 rn and m = dnn sn , where both rn , sn → ∞. Let us see what computational gain we get from this. Spectral clustering on the full graph has complexity O(n3 ), while the complexity of PACE with spectral clustering is 2 3 n n 3 3 2 O(T m ) = O rn m = O n mrn = O rn sn . m2 dn So if dn = Θ(nα ), then the complexity would be O(n3−α rn sn ), which is essentially O(n3−α ). When dn = Θ(log n) the gain is small. Note however that for a parallel implementation, with each source processing M out of the T subgraphs, wemay get a significant boost in running time, at least in terms of constants; the running time would be 3 O Mndn rn sn . Corollary 3.4 ((1 + )-approximate adjacency spectral clustering with ego subgraphs). Assume the setting of Lemma 3.1. Let r, r0 > 0. We have r 32(1 + θ)2 K πmax 16 2K 4 −1 Eδ̃(Ĉ, C) ≤ × min c (2 + ) 4 2 2 Dn,B,r0 , 1 + r0 + + + ∆, (9) λ2 λ αn πmin n nλ2 αn nλαn n where 2 4 nλ2 αn 4 θ nαn nλ4 αn2 T λ2 αn2 + 2 2 2 × exp − + πmax × 2 exp − + exp − . ∆ ≤ 2 2 × exp − λ αn 16 n λ αn 6 16 16 Here the quantity Dn,B,r0 → 1, if 2 λ2 nα πmin n πmax log n → ∞. The proof of Corollary 3.4 follows from Corollary 3.2 and an estimate for Eδ(ẐS , ZS )1(\|S\|≥m? ) given in (34), which is obtained using Lemma 3.1. For the right hand side in (9) to go to zero (assuming K fixed, balanced block sizes), we need min{nαn2 , T αn2 } → ∞. In terms of average degree this means that we need √ 2 dn n, and T dn2 . That with ego neighborhoods we can not go down to dn = Θ(log n) is not surprising, n since these ego networks are rather sparse in this case. One needs to use larger neighborhoods. Anyway, √ 2 writing dn = nrn , T = dn2 sn , where both rn , sn → ∞, the complexity of adjacency spectral clustering, in n 2 this case becomes O(T d3n ) = O(n2 dn rn sn ) and with M processing units gets further down to O( nMdn rn sn ). 13 Although from our analysis, it is not clear why PACE with spectral clustering should work well for sparse settings, in numerical simulations, we have found that in various regimes PACE with (regularized) spectral clustering vastly outperforms ordinary (regularized) spectral clustering (see Table 4). It seems that the reason why PACE works well in sparse settings lies in the weights Nij . With h-hop neighborhoods as the chosen subgraphs, if Puv = 1{ρg (u,v)≤h} , where ρg is the geodesic distance on G, then Nij = (P 2 )ij . It is known that spectral clustering on the matrix of geodesic distances works well in sparse settings (Bhattacharyya and Bickel (2014)). PACE seems to inherit that property through N , although we do not presently have a rigorous proof of this. We conclude this section with an illustration of PACE with random m-subgraphs using SDP as the algorithm A. We shall use the setting of Theorem 1.3 of Guédon and Vershynin (2015) for the illustration, stated here with slightly different notation. Let SDP-GV denote the following SDP (Guédon and Vershynin, 2015, SDP (1.10)) maximize hA, Xi subject to X 0, X ≥ 0, diag(X) = In , 1> X1 = 1> C1. Lemma 3.2 (Guédon and Vershynin (2015), Theorem 1.3). Consider an SBM with P mink Bkk ≥ a/n, 2 maxk6=k0 Bkk0 ≤ b/n, where a > b. Also let the expected variance of all edges n(n−1) 1≤k≤k0 ≤K Bkk0 (1 − Bkk0 )nkk0 = ng , where nkk0 denotes the number of pairs of vertices, one from community k, the other from community k 0 . Fix an accuracy > 0. If g ≥ 9 and (a − b)2 ≥ 484−2 g, then any solution X̂ of SDP-GV satisfies 1 kX̂ − Ck2F ≤ . n2 0 Corollary 3.5 (SDP with random m-subgraphs). Consider the setting 0. of Lemma 3.2. Let c > 1 >0 c > c1 (n−1)g c2 mg n−cm n−c m 2 0 2 and set ḡ1 = m−1 and ḡ2 = n , where c1 = c 1 − cm(nπmin −1) and c2 = (c ) 1 − c0 m(nπmax −1) . Fix an accuracy > 0. Assume that (a − b)2 ≥ 484−2 ḡ1 , and that ḡ2 ≥ 9. Then we have Eδ̃(Ĉ, C) ≤ 4m 1 0 2 2 2 × ( + e3 5−m + Ke−(1−c ) mπmin /4 + Ke−(c−1) mπmin /4 ) + × e−(1−θ) T p/2 . θ(m − 1) 2 For the simple two parameter blockmodel B = n1 ((a − b)I + b11> ) with equal community sizes, we have g a+(K−1)b dn , the average degree of the nodes (note that dn = a+(K−1)b − na ). The assumptions of K K Corollary 3.5 are satisfied when n ndn m = Ω max , . dn 2 (a − b)2 This is exactly similar to what we saw for spectral clustering (take a = nαn , and b = nαn (1 − λ)). In particular, when the average degree dn = Θ(nα ), and a − b = Θ(nα ), we need m = Ω(n1−α /2 ) and T n2α /4 for PACE to succeed. However, in the bounded degree regime, the advantage is negligible, only from a potentially smaller constant, because we need m = Ω(n). Again, from our numerical results, we expect that with h-hop subgraphs, PACE will perform much better. 3.2 Results on GALE We denote the unnormalized miscustering error between estimated labels Ẑ and the true labels Z, (Ẑ, Z ∈ {0, 1}n×K ) of the same set of nodes as M(Z, Ẑ) := nδ(Z, Z 0 ) = minQ perm. kẐ − ZQk0 . Note that since Ẑ, Z 14 are binary, the kẐ − ZQk0 = kẐ − ZQk1 = kẐ − ZQk2F . As discussed earlier, the number of misclustered nodes will be half of this number. The main idea of our algorithm is simple. Every approximately accurate clustering of a set of nodes is only accurate up to a permutation, which can never be recovered without the true labels. However we can align a labeling to a permuted version of the truth, where the permutation is estimated from another labeling of the same set of vertices. This is done by calculating the confusion matrix between two clusterings. We call two clusterings aligned if cluster i from one clustering has a large overlap with cluster i from the other clustering. If the labels are “aligned” and the clusterings agree, this confusion matrix will be a matrix with large diagonal entries. This idea is used in the Match algorithm, where we estimate the permutation matrix to align one clustering to another. Now we present our main result. We prove consistency of a slightly modified and weaker version of Algorithm 4. In Algorithm 4, at every step of a traversal, we apply the Match algorithm on the intersection of the current subgraph and the union of all subgraphs previously aligned to estimate the permutation of the yet unaligned current subgraph. However, in the theorem presented below we use the intersection between the unaligned current subgraph with the last aligned subgraph. Empirically it is better to use the scheme presented in Algorithm 4 since it increases the size of the intersection which requires weaker conditions on the clustering accuracy of any individual subgraph. We leave this for future work. (`) We now formally define our estimator Ẑ GALE . Let yi = 1{i∈S` } . Let Ẑ (`) denote the aligned clustering PT (`) of subgraph S` and let Ni = `=1 yi . Define GALE Ẑik (`) (`) `=1 yi Ẑik PT := Ni 1{Ni ≥τ } . (10) The entries of Ẑ GALE will be fractions, but as we show in Lemma B.2, rounding it to a binary matrix will not change consistency properties. Note that GALE depends on the spanning tree we use and particular the traversal of that spanning tree. Let SpanningTreesG be the set of all spanning trees of a graph G. For T ∈ SpanningTreesG , let TraversalsT be the set of all traversals of T . Let ẐTGALE ,(x1 ,...,xJ ) be the outcome of GALE on the traversal (x1 , . . . , xJ ) of T ∈ SpanningTreesSm,T . Theorem 3.2 (Misclustering rate of GALE). Let 0 < θ < 1 and r, r0 > 0. Let m = Ωr,r0 ,θ q n log n πmin , θT m n . T = Ωr,r0 ,θ (n log n/m), and τ = Consider an algorithm A which labels any random m-subgraph with 0 error ≤ m1 πmin /12 with probability at least 1 − δ. Then we have, with probability at least 1 − T δ − O(1/nr ), that T 1 X 1 GALE max max δ(ẐT ,(x1 ,...,xJ ) , Z) ≤ δ(Ẑ` , Z) + O . (11) T ∈SpanningTreesS θT nr (x1 ,...,xJ )∈TraversalsT m,T `=1 Again, the constant θ can be taken as close to 1 as one desires. Thus the above bound is also essentially optimal. We will now illustrate Theorem 3.2 with several algorithms A. We begin with a result on (1 + )approximate adjacency spectral clustering. Corollary 3.6 ((1 + )-approximate adjacency spectral clustering with q GALE). Assume the setting of n log n 0 00 000 0 Lemma 3.1. Let 0 < θ < 1. Let r, r , r , r > 0. Let m = Ωr,r0 ,θ πmin , T = Ωr,r ,θ (n log n/m), 15 and τ = θT m n . Then we have, with probability at least 1 − T mr00 − 2T K nr000 −O 1 nr 0 , that K πmax 2 1 000 min c−1 (2 + ) 2 , C , 1 +O 2 m n,m,r r T ∈SpanningTreesS θ λ α n π (x ,...,x )∈Traversals 1 n min J T m,T (12) 000 where the constant Cn,m,r is the same as in Corollary 3.1 max max δ(ẐTGALE ,(x1 ,...,xJ ) , Z) ≤ We see that the first term is exactly same as the first term in Corollary 3.1. This, for balanced graphs, 2 again imposes the condition m λK 2 α . In particular, if K = Θ(1) and we are in a dense well separated √ √n regime, with λ = Θ(1), αn = Ω(1/ n), then we need m = Ω( n log n). If K = Θ(1), λ = Θ(1) and αn = Θ(log n/n), then we need m n/ log n. In both cases, we need T = Ω(n log n/m). Thus in the regime where average degree is like log n there is still some computational advantage for very large networks (also factoring in parallelizability); however, for moderately sized networks, GALE may not lead to much computational advantage. Now we present an exact recovery result with SDP as the base algorithm A. We shall use a result† from Yan et al. (2017) on an SDP which they call SDP-λ. Let κ := πmax /πmin . Also let π denote the vector of the cluster proportions (π1 , . . . , πK ). Lemma 3.3 (Theorem 2 of Yan et al. (2017)). Let r > 0. Then Z diag(nπ)−1 Z > is the optimal solution of the SDP-λ, with probability at least 1 − O((nπmin )−r ), if   s max(B , K max B ) kk `6=k k`  min(Bkk − max Bk` ) = Ω̃r κ max . (13) k `6=k k nπk Assuming that any subsequent clustering of the exactly recovered scaled clustering matrix Z diag(nπ)−1 Z > gives the exact clustering Z back (for example, our distance based naive algorithm NaiveCluster‡ will do this), we have the following corollary. Corollary 3.7 (SDP with GALE, exact recovery). Assume the setting of Lemma 3.3. Let 0 < θ < q n log n θT m 0 1. Let r, r0 , r00 , r000 > 0. Let m = Ωr,r0 ,θ πmin , T = Ωr,r ,θ (n log n/m), and τ = n . Then, as long in (13) holds with m replacing n, we have, with probability at least as the separation condition T TK 1 1 − O (mπ )r00 + nr000 + nr0 , that min max T ∈SpanningTreesS max m,T (x1 ,...,xJ )∈TraversalsT δ(ẐTGALE ,(x1 ,...,xJ ) , Z) = O 1 nr , (14) Note that, in the above bound r can taken to be greater than 1. This means that, with high probability, the proportion of misclustered nodes is less than 1/n and hence zero, leading to exact recovery. As for computational complexity, note that the separation condition (13), with n replaced by m, restricts how small m can be. Consider the simple SBM (6) with balanced block sizes for concreteness. In this case, the 2 separation condition essentially dictates, as in the case of spectral clustering, that m λK 2 α . Thus the n remarks made earlier on how large m or T should be chosen apply here as well. † We are not using Lemma 3.2 as it only shows that the solution of SDP-GV has small norm difference from the ideal clustering matrix, but does not relate this directly to misclustering error. ‡ detailed in Appendix A 16 As discussed earlier in Section 2, even a naive implementation of GALE will only result in an O(n3/2 ) running time in addition to the time (ηm,T ) required to cluster the T random m-subgraphs, whereas a more careful implementation will only add a time to ηm,T that is nearly linear in T . Since SDPs are notoriously time intensive to solve, this gives us a big saving. 4 Simulations and real-data analysis In Table 1 we present a qualitative comparison of PACE and GALE with four representative global community detection methods Profile Likelihood (PL), Mean Field Likelihood (MFL), Spectral Clustering (SC) and Semi Definite Programming (SDP). Computationally easy Theoretical complexity Real-world scalability (n) Parallelizability PL 7 NP hard 102 − 103 7 MFL 7 O(nθ1 )† 102 − 103 7 SC 3 O(n3 ) 106 7 SDP 7 O(nθ2 )† 102 − 103 7 PACE 3 O(n2+ )‡ 106 3 GALE 3 O(n2+ )‡ 106 3 Table 1: Qualitative comparison of various methods. 4.1 Simulations: comparison against traditional algorithms For simulations we will use of the following simple block model: B = (p − q)I + qJ = ρn a((1 − r)I + rJ), where I is the K dimensional identity matrix and J is the K × K matrix of all ones. Here ρn will be the degree density, and r will measure the relative separation between the within block and between block connection probabilities, i.e. p and q. If the blocks have prior probabilities πi , i = 1, . . . , K, then the average degree d, under this model is given by X dn = (n − 1)ρn a((1 − r) πi2 + r). i In particular, if the model is balanced, i.e. πi = 1/K for all i, then dn = (n − 1)ρn a(1 + (K − 1)r) . K In order to understand and emphasize the role of PACE and GALE in reducing computational time while maintaining good clustering accuracy, we use different settings of sparsity for different methods. For recovering Ẑ from Ĉ in PACE, we have used random projection plus K-means (abbreviated as RPKmeans below), and spectral clustering (SC). We also want to point out that, for sparse unbalanced networks GALE may return more than K clusters, typically when a very small fraction of nodes has not been visited. However, it is possible that the unvisited nodes have important information about the network structure. For example, all subgraphs may be chosen from the larger clusters, thereby leaving the smallest cluster † ‡ θj ’s (≥ 3) depend on details of implementation and numerical accuracy sought. with O(n3 ) algorithms 17 unvisited. We take care of this by computing the smallest error between the (K + 1)! permutations of GALE’s clustering to the ground truth. This essentially treats the smallest cluster returned by GALE as misclustered. In real and simulated networks we have almost never seen GALE return a large number of unvisited nodes. SDP with ADMM: Interior point methods for SDPs are not very fast in practice. We have solved SDPs using an ADMM based implementation of Yan and Sarkar (2016). From Table 2 we see that PACE and GALE significantly reduces the running time of SDP without losing accuracy too much. In fact, if we use spectral clustering to estimate Ẑ from Ĉ in the last step of PACE, we get zero misclustering error (ME). Mean Field Likelihood: From Table 3 we see that our implementation of mean field on the full graph did not converge to an acceptable solution even after five and half hours, while both PACE and GALE return much better solutions in about two minutes. In fact, with spectral clustering in the last step of PACE, the misclustering error is only 0.14, which is quite good. This begs the question if this improvement is due to spectral clustering only. We show in the next simulation that in certain settings, even when spectral clustering is used as the base algorithm, PACE and GALE lead to significant improvements in terms of accuracy and running time. Algorithm SDP SDP + PACE + SC SDP + PACE + RPKmeans SDP + GALE ME(%) 0 0 9.1 1.2 Algorithm MFL MFL + PACE + SC MFL + PACE + RPKmeans MFL + GALE Time taken 1588s 288s 281s 281s ME(%) 50 1.4 36.5 19.2 Time taken 20439s 131s 125s 126s Table 3: PACE and GALE with Mean Field Likelihood (MFL) as the base method. Simulation settings: n = 5000, average degree = 13, 2-hop neighborhood, 2 equal sized clusters, T = 100, parallel implementation in Matlab with 20 workers. Table 2: PACE and GALE with SDP as the base method. Simulation settings: n = 5000, average degree = 128, m = 500, 4 equal sized clusters, T = 100, parallel implementation in Matlab with 20 workers. Regularized Spectral Clustering: In sparse unbalanced settings, regularized spectral clustering with PACE and GALE performs significantly better than regularized spectral clustering on full graph. In fact, with spectral clustering used in the last step of PACE, we can hit about 5% error or below, which is quite remarkable. See Table 4. In Section 4.2 we will see that PACE and GALE also add stability to spectral clustering (in terms of clustering degree 1 vertices). Profile Likelihood with tabu search: Optimizing profile likelihood (PL) or likelihood modularity (Bickel and Chen (2009)) for community detection is a combinatorial problem, and as such hard to scale, even if we ignore the problem of local minima. In Table 5 we compare running time of profile likelihood (optimized using tabu search) and its divide and conquer versions. We see that the local methods significantly cut down the running time of PL without losing accuracy too much. We also applied profile likelihood on 5000 node graphs with 20 workers. Although PACE and GALE finished in about 22 minutes, the global method did not finish in 3 days. So, here we present results on 1000 node networks instead. Remark 4.1. We have seen from the results presented in this section that, for recovering Ẑ from Ĉ in PACE, spectral clustering outperforms the random projection based algorithms (e.g., RPKmeans). For smaller 18 Random 1500-subgraph 3-hop neighborhood 5-hop onion Algorithm ME(%) Time taken ME(%) Time taken ME(%) Time taken RSC 39.6 87s RSC + PACE + SC 11.1 26s 3.4 21s 5.1 59s RSC + PACE + RPKmeans 34.7 20s 34.2 14s 18 53s RSC + GALE 17.9 23s 33.6 13s 29.7 52s Table 4: PACE and GALE with Regularized Spectral Clustering (RSC) as the base method. Simulation settings: n = 5000, average degree = 7, 2 unequal sized clusters with relative sizes π = (0.2, 0.8), T = 100, parallel implementation in Matlab with 20 workers. Random 310-subgraph Algorithm ME(%) Time taken PL 0 70m PL + PACE + RPKmeans 3.9 PL + GALE 1.2 2-hop neighborhood ME(%) Time taken 30m 3.5 38m 30m 29.5 38m Table 5: PACE and GALE with Profile Likelihood (PL) as the base method. Simulation settings: n = 1000, average degree = 18.47, 2 unequal sized clusters with relative sizes π = (0.4, 0.6), T = 50, parallel implementation in Matlab with 12 workers. We sampled 2-hop neighborhoods by selecting their roots uniformly at random from nodes having degree greater than the 0.1th lower quantile (= 12) of the degree distribution (average neighborhood size was 310). Ordinary PACE with such a scheme may be thought of as w-PACE, as discussed in Section 2. networks, this is not an issue (e.g., spectral clustering on the dense 5000 × 5000 matrix Ĉ in the context of Table 3 took only about 7-8 seconds). However, for networks of much larger scale (say, with several million nodes), that last step would be costly if spectral clustering is used. Designing better algorithms for recovering Ẑ from Ĉ is something we are working on currently. 4.2 Real data analysis Political blog data: This is a directed network (see Figure 2) of hyperlinks between 1490 blogs (2004) that are either liberal or conservative (Adamic and Glance (2005)); we have ground truth labels available for comparison, 758 are liberal, 732 are conservative. We convert it into an undirected network by putting an edge between blogs i and j if there is at least one directed edge between them. The resulting network has lots of isolated nodes and isolated edges. The degree distribution is also quite heterogeneous (so a degree-corrected model would be more appropriate). We focus on the largest connected component. We use Laplacian spectral clustering (row normalized, to correct for degree heterogeneity), with PACE. Tables 6-7 show that PACE and GALE add stability (possibly in eigenvector computation) to spectral clustering. Indeed, with PACE and GALE we are able to cluster “leaf” vertices (i.e. vertices of degree 1), with significantly more accuracy. 19 Figure 2: Network of political blogs, reds are conservative, blues are liberal; picture courtesy: Adamic and Glance (2005). Largest Conn. Comp. With leaves (1222 nodes) Without leaves (1087 nodes) RSC 18.74% 11.87% RSC + PACE 6.79% 4.23% SC 48.12% 3.13% SC + PACE 6.55% 3.86% Table 6: Misclustering rate in the political blog data. PACE was used with T = 10, and h-hop neighborhoods with h = 2, with roots chosen at random from high degree nodes. Largest Conn. Comp. With leaves (1222 nodes) Without leaves (1087 nodes) RSC 18.74% 11.87% RSC + PACE 13.34% 12.8% RSC + GALE 11.62% 10.0% SC 48.12% 4.23% SC + PACE 7.86 % 7.28% SC + GALE 5.81% 6.7% Table 7: Misclustering rate in the political blog data. GALE and PACE were used with T = 50, and m = 300 random subgraphs. 5 Discussion To summarize, we have proposed two divide-and-conquer type algorithms for community detection, PACE and GALE, which can lead to significant computational advantages without sacrificing accuracy. The main idea behind these methods is to compute the clustering for each individual subgraph and then “stitch” them together to produce a global clustering of the entire network. The main challenge of such a stitching procedure comes from the fundamental problem of unidentifiability of label assignments. That is, if two subgraphs overlap, the clustering assignment of a pair of nodes in the overlap may be inconsistent between the two subgraphs. PACE addresses this problem by estimating the clustering matrix for each subgraph and then estimating the global clustering matrix by averaging over the subgraphs. GALE takes a different approach by using overlaps between two subgraphs to calculate the best alignment between the cluster memberships of nodes in 20 the subgraphs. We prove that, in addition to being computationally much more efficient than base methods which typcally run in Ω(n2 ) time, these methods have accuracy at least as good as the base algorithm’s typical accuracy on the type of subgraphs used, with high probability. Experimentally, we show something more interesting — we identify parameter regimes where a local implementation of a base algorithm based on PACE or GALE in fact outperforms the corresponding global algorithm. One example of this is the Meanfield algorithm, which typically suffers from bad local optima for large networks. Empirically, we have seen that on a smaller subgraph, with a reasonable number of restarts, it finds a local optima that is often highly correlated with the ground truth. PACE and GALE take advantage of this phenomenon to improve on accuracy/running time significantly. Another example is Regularized Spectral Clustering on sparse unbalanced networks. We intend to theoretically investigate this further in future work. Finally, working with many subgraphs naturally leads to the question of self consistency of the underlying algorithm. This is often crucial in real world clustering problems with no available ground truth labels. We intend to explore this direction further for estimating model parameters like the number of clusters, algorithmic parameters like the size and number of subgraphs, number of hops to be used for the neighborhood subgraphs, etc. Currently, these are all picked a priori based on the degree considerations. It may also be possible to choose between different models (e.g., standard blockmodels, degree corrected models, dot product models etc.) by examining which model leads to the most self consistent results. We leave this for future work. In conclusion, not only are our algorithms, to the best of our knowledge, the first ever divide-and-conquer type algorithms used for community detection, we believe that the basic principles of our methods will have a broad impact on a range of clustering and estimation algorithms that are computationally intensive. 6 6.1 Proofs Results on PACE Proof of Proposition 3.1. Since both Z, Z 0 are 0, 1-valued, we can safely replace the count by Frobenius norm squared, i.e. 1 δ(Z, Z 0 ) = inf kZQ − Z 0 k2F . Q perm. n Now, note that (ZQ)(ZQ)> = ZZ > for all permutation matrices Q. Thus kC − C 0 kF = k(ZQ)(ZQ)> − Z 0 Z 0> kF = k(ZQ − Z 0 )(ZQ)> + Z 0 ((ZQ)> − Z 0> kF ≤ k(ZQ − Z 0 )(ZQ)> kF + kZ 0 ((ZQ)> − Z 0> kF ≤ kZQ − Z 0 kF kZQ> k2 + kZ 0 k2 k(ZQ)> − Z 0> kF . But kZ 0 k22 is the maximum eigenvalue of Z 0> Z 0 which is diagonal with its maximum diagonal entry being the size of the largest cluster under Z 0 . Thus kZ 0 k22 equals the size of the largest cluster under Z 0 and so is trivially upper bounded by n. Same goes for kZQ> k22 . Therefore we get √ kC − C 0 kF ≤ 2 nkZQ − Z 0 kF . Squaring this, and taking infimum over all permutation matrices Q in the right hand side, we obtain the claimed inequality. 21 Now we will prove Theorem 3.1. The proof will be broken down into two propositions. First we decompose Ĉij − Cij = Ĉij − Cij 1{Nij ≥τ } + Cij 1{Nij ≥τ } − Cij (15) {z } \| {z } \| =:(E1 )ij Note that (E1 )ij = 1{Nij ≥τ } Therefore =:(E2 )ij (`) (`) `=1 yij (Ĉij − Cij )/Nij , PT and (E2 )ij = −1{Nij <τ } Cij , so that (E1 )ij (E2 )ij = 0. 1 1 kĈ − Ck2F = 2 (kE1 k2F + kE2 k2F ). 2 n n We will estimate kE1 kF and kE2 kF separately. (16) Proposition 6.1. We have T EkĈ (S) − C (S) k2F . (17) τ P (`) (`) = Wij T`=1 yij (Ĉij − Cij ). So, by an application of Cauchy- EkE1 k2F ≤ Proof. Let Wij := Schwartz, we have kE1 k2F = 1{Nij ≥τ } . Nij X Wij2 i,j Then (E1 )ij T X !2 (`) (`) yij (Ĉij − Cij ) (C-S) ≤ X T X Wij2 i,j `=1 `=1 \| = T X X ! (`) yij T X ! (`) (`) yij (Ĉij − Cij ) 2 `=1 {z =Nij (`) } (`) Wij yij (Ĉij − Cij )2 `=1 i,j ≤ max Wij ij Note that Wij = 1{Nij ≥τ } Nij (`) (`) yij (Ĉij − Cij )2 `=1 i,j = max Wij ij T X X T X kĈ (`) − C (`) k2F . `=1 ≤ τ1 . On the other hand, since the subgraphs were chosen independently using the same sampling scheme, the kĈ (`) − C (`) kF are identically distributed. Therefore, taking expectations we get T EkE1 k2F 1 X T ≤ × EkĈ (`) − C (`) k2F = EkĈ (S) − C (S) k2F , τ τ `=1 where S is a randomly chosen subgraph under our subgraph selection scheme. Proposition 6.2. Let nmax be the size of the largest block. Then we have EkE2 k2 ≤ nmax n max P(Nij < τ ). (18) i,j P 21 Proof. Since −(E2 )ij = Cij 1{Nij <τ } , we have kE2 k2F = i,j Cij {Nij <τ } , and by taking expectations we get X X 2 2 EkE2 k2F = Ci,j P(Nij < τ ) ≤ max P(Nij < τ ) Ci,j i,j i,j i,j = max P(Nij < τ ) i,j K X a=1 22 n2a ≤ nmax n × max P(Nij < τ ). i,j Proof of Theorem 3.1. Combining Propositions 6.1 and 6.2, we get (3). Finally, note that EkĈ (S) − C (S) k2F = EkĈ (S) − C (S) k2F 1(\|S\|<m? ) + EkĈ (S) − C (S) k2F 1(\|S\|≥m? ) ≤ n2 P(\|S\| < m? ) + 4E\|S\|2 δ(ẐS , ZS )1(\|S\|≥m? ) . Proof of Corollary 3.1. For this sampling scheme \|S\| = m ≥ m? and with p = so that we have, using the Chernoff bound∗ for binomial lower tail, that 2 T p/2 P (Nij < θT p) ≤ e−(1−θ) m(m−1) n(n−1) , Nij ∼ Binomial(T, p) . Finally, we get (4) by plugging in these parameter values and estimates in (3). Proof of Corollary 3.2. The most crucial thing to observe here is that if one removes the root node and its adjacent edges from a 1-hop neighborhood, then the remaining “ego network” has again a blockmodel structure. Indeed, let S be a random ego neighborhood of size ≥ s with root R, i.e. V (S) = {j : ARj = 1}. Then conditional on V (S) being R’s neighbors, and the latent cluster memberships, edges in E(S) are independently generated, i.e. for j, k, `, m ∈ V (S), and s, t ∈ {0, 1}, we have P(Ajk = s, A`m = t\|S, Z) = P(Ajk = s\|Z)P(A`m = t\|Z). This is because the “spoke” edges ARj are independent of Aj,k , j, k ∈ V (S). Therefore, conditional on S, this ego-subgraph is one instantiation of a block model with the same parameters on \|S\| vertices. (`) Now for ego networks, yij ∼ Bernoulli(nij /n), where nij is the total number of ego-subgraphs containing both i and j. Notice that X nij = 1{Ai` =1,Aj` =1} , `6=i,j that is, nij is the sum of (n − 2) independent Bernoulli random variables 1{Ai` =1,Aj` =1} ∼ Bernoulli(Bσ(i)σ(`) Bσ(j)σ(`) ). So 2 (n − 2)B# ≥ Enij = X Bσ(i)σ(`) Bσ(j)σ(`) ≥ (n − 2)B?2 , `6=i,j and we have, by the Chernoff bound, that P(nij ≤ (n − 2)B?2 (Enij − (n − 2)B?2 + ∆)2 − ∆) ≤ exp − 2Enij ! ∆2 ≤ exp − . 2 2(n − 2)B# (19) In order to apply Theorem 3.1 we need the following two ingredients, which we will now work out. ∗ We use a slightly loose but convenient form of the Chernoff bounds: (i) P(X ≤ (1 − δ)µ) ≤ exp(−δ 2 µ/2), and (ii) P(X ≥ (1 + δ)µ) ≤ exp(−δ 2 µ/3), where X = X1 + · · · + Xn and Xi ’s are independent binary random variables, with EX = µ, and 0 < δ < 1. 23 (i) estimate of \|S\|, and (ii) estimate of P(Nij < τ ). (i) Estimate of \|S\|. Note that \|S\| = P k AkR . So P(\|S\| < m? ) = ER P(\|S\| < m? \| R). P Since (n − 1)B? ≤ E(\|S\| \| R) = j6=R Bσ(k)σ(R) ≤ (n − 1)B# , and AkR , 1 ≤ k ≤ n are independent, we have, by Chernoff’s inequality, that (E(\|S\| \| R) − m? )2 P(\|S\| < m? \| R) ≤ exp − 2E(\|S\| \| R) ((n − 1)B? − m? )2 ≤ exp − , 2(n − 1)B# where m? ≤ (n − 1)B? . Therefore, the same upper bound holds for P(\|S\| < m? ). In particular, for m? ≤ (n − 1)B? /2 we have (n − 1)B?2 nB?2 P(\|S\| < m? ) ≤ exp − ≤ exp − . (20) 8B# 16B# Similarly, using Chernoff’s inequality for Binomial upper tail, we can show that, for θ > 0, 2 θ nB# P(\|S\| > (1 + θ)nB# ) ≤ exp − . 6 (ii) Estimate of P(Nij < τ ). Recall that Nij \| nij ∼ Binomial(T, nij /n). Then 2nτ 2nτ P(Nij < τ ) = P Nij < τ, nij < + P Nij < τ, nij ≥ . T T \| {z } \| {z } =:P1 =:P2 Clearly 2nτ nB?2 P1 ≤ P nij < ≤ P nij < . T 2 Now, given nij such that nij ≥ 2nτ T , we can invoke Chernoff’s inequality to get (E(Nij \| nij ) − τ )2 T nij P(Nij < τ \| nij ) ≤ exp − ≤ exp − . 2E(Nij \| nij ) 8n Therefore T nij P2 ≤ E exp − 8n 1{nij ≥ 2nτ } T T nij ≤ E exp − 8n T nij T nij = E exp − 1 − 1 nB 2 + E exp nB 2 {nij < 2 ? } (nij ≥ 2 ? ) 8n 8n nB?2 T B?2 ≤ P nij < + exp − . 2 16 24 (21) Thus T B?2 nB?2 + exp − . P(Nij < τ ) ≤ 2P nij < 2 16 But by (19) nB?2 P nij < 2 2 ? 2 ((n − 2)B?2 − nB 2 ) ≤ exp − 2 2(n − 2)B# ! (n − 2) B?4 ≤ exp − 2 8 B# Thus nB?4 P(Nij < τ ) ≤ 2 exp − 2 16B# ! ! nB?4 ≤ exp − 2 16B# ! . T B?2 + exp − . 16 Now we are ready to use (3). Using our estimates on \|S\| we get that 2 2 2 E\|S\| δ(ẐS , ZS )1(\|S\|≥m? ) ≤ (1 + θ) n 2 Eδ(ẐS , ZS )1(\|S\|≥m? ) B# θ2 nB# + exp − 6 . Next we plug into (3) all the estimates we derived in this subsection to get the desired bound (5). 6.2 Results on GALE For any clustering Ẑi on subgraph Si , let Πi ∈ arg minQ perm. kẐi − Zi Qk1 , where Zi is used a shorthand for ZSi = Z S , the true cluster membership matrix for the members of Si . Define the matrix Fi by the i requirement Ẑi = Zi Πi + Fi , Fi ∈ {0, ±1}n×K . (22) In other words, kFi k1 = M(Ẑi , Zi ). We first analyze Algorithm 3, i.e. Match. Recall that, if two clusterings on some set S agree, then the confusion matrix will be a diagonal matrix up to permutations, with the entries in the diagonal corresponding to the cluster sizes in either of the clusterings. In the following lemma, we consider a noisy version of this, where the two clusterings are not in perfect agreement. This lemma essentially establishes that if supplied with two clusterings whose confusion matrix is a diagonal matrix up to permutations plus noise, then Match will recover the correct aligning permutation, if the noise is not too large. > K×K , kΓk Lemma 6.1. Let d ∈ RK ∞ ≤ mini di /3. Then + . Also let M = Π2 diag(d)Π1 + Γ, where Γ ∈ R > Match returns Π = Π2 Π1 , when applied on the confusion matrix M . Proof of Lemma 6.1. Let D := diag(d). Let ξi be the permutation encoded in Πi , for i ∈ [2], i.e. (Πi )uv = 1{v=ξi (u)} . It is easy to see that Muv = Dξ−2 (u),ξ−1 (v) + Γuv . We have for all u, v such that (Π> 2 Π1 ))uv = 0, 1 2 Muv ≤ kΓk∞ , whereas for all u, v such that (Π> 2 Π1 )uv = 1, Muv ≥ mina Daa − kΓk∞ . Hence min u,v : (Π> 2 Π1 )uv =1 Muv ≥ min Daa − kΓk∞ ≥ 2kΓk∞ ≥ 2 a max u,v : (Π> 2 Π1 )uv =0 Muv . Hence the top K (recall that K is the number of rows in M ) elements are the diagonal elements in D. Thus the elements of Π learned by the Match algorithm will be Πuv = 1{ξ−1 (u)=ξ−1 (v)} . This is equivalent 2 to Π = Π> 2 Π1 . 25 1 Now we will establish that post-alignment misclustering errors on subgraphs equal the original misclustering errors. For this we first need a lemma on what happens when we align two subgraphs based on their intersection. Lemma 6.2. Consider two random m-subgraphs S1 and S2 . Suppose our clustering algorithm A outputs l 2m m clusterings Ẑi , i = 1, 2. Let S = S1 ∩ S2 be of size at least m1 = 2n . Assume that M(Ẑi , Zi ) ≤ (S) m1 πmin /12 and that mink∈[K] nk Match(Ẑ2 S , Ẑ1 S ). ≥ m1 πmin . Then kẐ2 Π − ZΠ1 k = M(Ẑ2 , Z), where Π is the output of Proof of Lemma 6.2. To ease notation, let us write Z̃i := Ẑi S , F̃i := Fi S , i = 1, 2, and Z̃ := Z S . Then by restricting (22) to S, we get Z̃i = Z̃Πi + F̃i , F̃i ∈ {0, ±1}n×K . Now > > > > > Z̃2> Z̃1 = Π> 2 Z̃ Z̃Π1 + Π2 Z̃ F̃1 + F̃2 Z̃Π1 + F̃2 F̃1 . \| {z } =:Γ Note that (a) multiplication by a permutation matrix does not change the k·k1 norm, and (b) for any matrix A, we have kAk1 = kA> k1 . Therefore > kΠ> 2 Z̃ F̃1 k1 (by (a)) = kZ̃ > F̃1 k1 = X X XX X X \| (Z̃ > )ai (F̃1 )ib \| ≤ Z̃ia \|(F̃1 )ib \| = \|(F̃1 )ib \| × Z̃ia = kF̃1 k1 . i a,b a,b i i,b \| a {z } =1 Similarly, kF̃2> Z̃Π1 k1 (by (b)) = > kΠ> 1 Z̃ F̃2 k1 ≤ kF̃2 k1 . Finally, kF̃2> F̃1 k1 ≤ XX a,b \|(F̃2 )ia (F̃1 )ib \| = i X \|(F̃1 )ib \| × X \|(F̃2 )ia )\| = 2kF̃1 k1 , a i,b \| {z =2 } where we have used the fact that each row of F̃i has exactly one 1 and one −1. By (b), kF̃2> F̃1 k1 = kF̃1> F̃2 k1 ≤ 2kF̃2 k1 , and therefore kF̃2> F̃1 k1 ≤ kF̃1 k1 + kF̃2 k1 . Note that, by our assumptions on the individual misclustering errors, kF̃i k1 ≤ kFi k1 = M(Ẑi , Zi ) ≤ m1 πmin /12. Therefore kΓk ≤ 2(kF̃1 k1 + kF̃2 k1 ) ≤ m1 πmin /3. (S) Since by our assumption, (Z̃ > Z̃)kk = nk Match(Z̃2 , Z̃1 ) is Π = Π> 2 Π1 . Hence ≥ m1 πmin , we can apply Lemma 6.1 to see that the output > kẐ2 Π − ZΠ1 k1 = kẐ2 Π> 2 Π1 − ZΠ1 k1 = kẐ2 Π2 − Zk1 = kẐ2 − ZΠ2 k1 = kF2 k1 . 26 Proposition 6.3. Let the T subsets, (Si )Ti=1 be associated with estimated clusterings Ẑi with misclustering error M(Ẑi , Zi ) ≤ m1 πmin /12. Consider a traversal x1 ∼ x2 ∼ · · · ∼ xJ of some spanning tree T of Sm,T as in Algorithm 4, satisfying (Sxi ∩Sxi−1 ) min min nk 2≤j≤J k∈[K] ≥ m1 πmin , (S) where nk is the number of nodes from cluster k in a subgraph S. Applying GALE on this walk: let Ẑ (x1 ) := Ẑx1 , Π̂x1 = I and for 2 ≤ i ≤ J, define recursively Π̂xi ( I = Match(Ẑxi Sxi ∩Sxi−1 , Ẑ (xi−1 ) Sxi ∩Sxi−1 ) if xi was visited before, i.e. xi = xj for some 1 ≤ j < i, otherwise, and set Ẑ (xi ) = Ẑxi Π̂xi . Then, for any 1 ≤ i ≤ J, we have that M(Ẑ (xi ) , Zxi ) = M(Ẑxi , Zxi ). Proof. We claim that, for all 1 ≤ i ≤ J, we have Πx1 ∈ arg min kẐ (xi ) − Zxi Qk, and M(Ẑ (xi ) , Zxi ) = M(Ẑxi , Zxi ). Q (23) We use strong induction to prove this claim. (23) is true by definition, for i = 1. Now assume that it is true for all i ≤ `. Then, for all i ≤ `, we have the representation Ẑ (xi ) = Zxi Πx1 + F̄xi , for some matrix F̄xi ∈ {0, ±1}m×K . If x`+1 has been visited before, i.e. x`+1 = xj for some j ≤ `, then (23) holds by our induction hypothesis. Otherwise, we can apply Lemma 6.2 on the two clusterings Ẑ (x` ) on Sx` and Ẑx`+1 on Sx`+1 , to conclude that kẐ (x`+1 ) − Zx`+1 Πx1 k1 = kẐx`+1 Π̂x`+1 − Zx`+1 Πx1 k1 = M(Ẑx`+1 , Zx`+1 ). But Ẑ (x`+1 ) = Ẑx`+1 Π̂x`+1 , and Π̂x`+1 = Π> x`+1 Πx1 . So, for any permutation matrix Q we have kẐ (x`+1 ) − Zx`+1 Qk1 = kẐx`+1 Π̂x`+1 − Zx`+1 Qk1 = kẐx`+1 − Zx`+1 QΠ̂> x`+1 k1 ≥ M(Ẑx`+1 , Zx`+1 ) = kẐ (x`+1 ) − Zx`+1 Π1 k1 . Thus, indeed Πx1 ∈ arg minQ kẐ (x`+1 ) − Zx` Qk and as a consequence M(Ẑ (x`+1 ) , Zx`+1 ) = M(Ẑx`+1 , Zx`+1 ). The induction is now complete. Now we establish an upper bound on the error of the estimator Ẑ GALE in terms of the errors of the aligned clusterings Ẑ (`) . Proposition 6.4. Let 0 < θ < 1, and set τ = θT m n . Then, for any cluster membership matrix Z̄ ∈ {0, 1}n×K , T kẐ GALE − Z̄k2F 1 X kẐ (`) − Z̄` k2F X ≤ + 1{Ni <τ } . n θT m i `=1 27 (24) Proof. The proof is very similar to the proof of Theorem 3.1. We decompose Z̄ik = Z̄ik 1{Ni ≥τ } + Z̄ik 1{Ni <τ } . Using this we have GALE Ẑik (`) (`) `=1 yi (Ẑik PT − Z̄ik = − Z̄ik ) Ni 1{Ni ≥τ } − Z̄ik 1{Ni <τ } . Therefore kẐ GALE 2 − Z̄k = X GALE (Ẑik 2 − Z̄ik ) = i,k X (`) (`) `=1 yi (Ẑik PT X !2 1{Ni ≥τ } + Ni i,k ≤ − Z̄ik ) (`) (`) `=1 yi (Ẑik − Z̄ik ) !2 1{Ni ≥τ } + Ni \| 2 Z̄ik 1{Ni <τ } i,k PT i,k X X 1{Ni <τ } . i {z } =:E PT P (`) (`) 2 2 We now analyze E. Let Wi := 1(Ni ≥τ ) /Ni . Noting that E = i,k Wi ( `=1 yi (Ẑik − Z̄ik )) , an application of Cauchy-Schwartz inequality gives ! T T T X X X X (`) (`) X (`) (`) (`) E≤ Wi2 yi yi (Ẑik − Z̄ik )2 = Wi yi (Ẑik − Z̄ik )2 i,k `=1 \| {z =Ni } `=1 `=1 i,k ≤ max Wi i ≤ max Wi i Note that Wi = 1(Ni ≥τ ) Ni ≤ 1 τ = n θT m . T X X (`) (`) yi (Ẑik − Z̄ik )2 `=1 i,k T X kẐ (`) − Z̄` k2F . (25) `=1 Therefore E≤ T n X kẐ (`) − Z̄` k2F . θT m (26) `=1 We now mention some auxiliary results, whose proofs are deferred to the Appendix. These will help us control the probability of a “bad” event, defined in the proof of Theorem 3.2. Lemma 6.3. Let τ = θT m n , where 0 < θ < 1. Let r > 0. Then, with probability at least 1 − ! r X 3r log n −(1−θ)2 T m/2n 1{Ni <τ } ≤ ne 1+ . −(1−θ)2 T m/2n ne i 1 nr , we have Recall that we view the T random m-subsets of [n] as nodes of a super-graph Sm,T and put an edge between nodes a(≡ Sa ) and b(≡l Sb )m in Sm,T (we use the shorthand a ∼ b to denote an edge between a and b), if Yab := \|Sa ∩ Sb \| ≥ m1 = graph. m2 2n . The next lemma shows that Sm,T is in fact an Erdös-Rényi random 28 Lemma 6.4. The super-graph Sm,T is an Erdös-Rényi random graph with 2 m2 m ≥ 1 − exp − . P (a ∼ b) = P Yab ≥ 2n 16n (27) Because of our assumptions on m, it follows that Sm,T is well above the connectivity threshold for Erdös-Rényi random graphs. Lemma 6.5. The super-graph Sm,T is connected with probability at least 1 − exp(−O(n)). The next lemma states that intersection of two random m-subgraphs contains enough representatives from each cluster, with high probability. Lemma 6.6. Consider two random m-subgraphs Sa and Sb . Let r > 0 and P (S ∩S ) nk a b m2 πk < n s 1−O 2rn log n m2 πk m 2 πk n log n !!! ≤ ≥ 20r. Then 2 . nr We are now ready to prove our main result on GALE. Proof of Theorem 3.2. First we construct a good set in the sample space. Consider the following “bad” events B1 := {(S1 , . . . , ST ) \| Sm,T is connected}, (Si ∩Sj ) B2 := (S1 , . . . , ST ) \| min min nk < m1 πmin , (i,j) k ( r X −(1−θ)2 T m/2n B3 := (S1 , . . . , ST ) \| 1+ 1{Ni <τ } > e i !) 3r0 log n , ne−(1−θ)2 T m/2n B4 := {(A, S1 , . . . , ST ) \| max M(Ẑi , Zi ) > m1 πmin /12}. 1≤i≤T Let B := ∪4i=1 Bi . Let r00 > 0 By Lemma 6.5, we have P(B1 ) ≤ exp(−O(n)). Lemma 6.6, and a union 2 1 bound gives P(B2 ) ≤ T2 × K × n2r00 ≤ KT 00 . By Lemma 6.3, P(B3 ) ≤ 0 . Finally, by our hypothesis on nr nr individual misclustering errors, we have, by a union bound, that P(B4 ) ≤ T δ. Therefore P(B) ≤ 4 X i=1 KT 2 1 P(Bi ) ≤ exp(−O(n)) + r00 + r0 + T δ ≤ T δ + O n n 1 nr 0 , by choosing r00 suitably large. Now on the good set B c , for any T ∈ SpanningTreesSm,T and for any (x1 , . . . , xJ ) ∈ TraversalsT , the hypothesis of Propositions 6.3 and 6.4 are satisfied. Now note that δ(ẐTGALE ,(x1 ,...,xJ ) , Z) ≤ kẐTGALE ,(x1 ,...,xJ ) − ZΠx1 k1 n = 2 kẐTGALE ,(x1 ,...,xJ ) − ZΠx1 kF n By Proposition 6.3, we also have, for any 1 ≤ ` ≤ T , that kẐ (`) − Z` Πx1 k2F kẐ (`) − Z` Πx1 k1 = = δ(Ẑ (`) , Z) = δ(Ẑ` , Z). m m 29 . Thus, by of Proposition 6.4, taking Z̄ = ZΠx1 and noting then that Z̄` = Z` Πx1 , we have δ(ẐTGALE ,(x1 ,...,xJ ) , Z) ≤ T X 1 X δ(Ẑ` , Z) + 1{Ni <τ } . θT i `=1 Since the bound on the RHS does not depend on the particular spanning tree used, or a particular traversal thereof, we conclude that, on the good event B c , max max T ∈SpanningTreesS m,T (x1 ,...,xJ )∈TraversalsT Now, for our choice of τ , and T ≥ X 1{Ni <τ } ≤ e 2r000 n log n (1−θ)2 m −(1−θ)2 T m/2n δ(ẐTGALE ,(x1 ,...,xJ ) , Z) T X 1 X ≤ δ(Ẑ` , Z) + 1{Ni <τ } . θT i `=1 , we have e r 1+ i −(1−θ)2 T m/2n 3r0 log n ne−(1−θ)2 T m/2n ≤ ! 1 . nr000 r =O Therefore, on the event B c log n n1+r000 ! =O 1 nr , by choosing r000 suitably large. Thus we conclude that with probability at least 1 − T δ − O n1r0 , we have max T ∈SpanningTreesS 7 max m,T (x1 ,...,xJ )∈TraversalsT δ(ẐTGALE ,(x1 ,...,xJ ) , Z) ≤ T 1 X 1 δ(Ẑ` , Z) + O . θT nr `=1 Acknowledgments We thank Arash Amini, David Blei, David Choi for their valuable comments when versions of this work were presented at conferences. SSM thanks Aditya Guntuboyina, Alan Hammond, Luca Trevisan and Bin Yu for their valuable comments during his quals talk at Berkeley, which was based on this paper. References Adamic, L. A. and Glance, N. (2005). The political blogosphere and the 2004 us election: divided they blog. In Proceedings of the 3rd international workshop on Link discovery, pages 36–43. ACM. Airoldi, E. M., Blei, D. M., Fienberg, S. E., and Xing, E. P. (2009). Mixed membership stochastic blockmodels. In Advances in Neural Information Processing Systems, pages 33–40. Amini, A. A., Chen, A., Bickel, P. J., Levina, E., et al. (2013). Pseudo-likelihood methods for community detection in large sparse networks. The Annals of Statistics, 41(4):2097–2122. Amini, A. A. and Levina, E. (2014). On semidefinite relaxations for the block model. arXiv preprint arXiv:1406.5647. Bhattacharyya, S. and Bickel, P. J. (2014). Community detection in networks using graph distance. arXiv preprint arXiv:1401.3915. 30 Bickel, P. J. and Chen, A. (2009). A nonparametric view of network models and newman–girvan and other modularities. Proceedings of the National Academy of Sciences, 106(50):21068–21073. Binkiewicz, N., Vogelstein, J. T., and Rohe, K. (2014). Covariate assisted spectral clustering. arXiv preprint arXiv:1411.2158. Cai, T. T., Li, X., et al. (2015). Robust and computationally feasible community detection in the presence of arbitrary outlier nodes. The Annals of Statistics, 43(3):1027–1059. Charikar, M. S. (2002). Similarity estimation techniques from rounding algorithms. In Proceedings of the Thiry-fourth Annual ACM Symposium on Theory of Computing, STOC ’02, pages 380–388, New York, NY, USA. ACM. Gopalan, P. K. and Blei, D. M. (2013). Efficient discovery of overlapping communities in massive networks. Proceedings of the National Academy of Sciences, 110(36):14534–14539. Greene, E. and Wellner, J. A. (2017). Exponential bounds for the hypergeometric distribution. Bernoulli, 23(3):1911–1950. Guédon, O. and Vershynin, R. (2015). Community detection in sparse networks via grothendieck’s inequality. Probability Theory and Related Fields, pages 1–25. Guruharsha, K., Rual, J.-F., Zhai, B., Mintseris, J., Vaidya, P., Vaidya, N., Beekman, C., Wong, C., Rhee, D. Y., Cenaj, O., et al. (2011). A protein complex network of drosophila melanogaster. Cell, 147(3):690– 703. Holland, P. W., Laskey, K. B., and Leinhardt, S. (1983). Stochastic blockmodels: First steps. Social networks, 5(2):109–137. Hollingshead, A. B. (1949). Elmstown’s youth; the impact of social classes on adolescents. Karrer, B. and Newman, M. E. (2011). Stochastic blockmodels and community structure in networks. Physical Review E, 83(1):016107. Le, C. M., Levina, E., and Vershynin, R. (2015). Sparse random graphs: regularization and concentration of the laplacian. arXiv preprint arXiv:1502.03049. Lei, J. and Rinaldo, A. (2015). Consistency of spectral clustering in stochastic block models. The Annals of Statistics, 43(1):215–237. Mackey, L., Talwalkar, A., and Jordan, M. I. (2015). Distributed matrix completion and robust factorization. Journal of Machine Learning Research, 16:913–960. McAuley, J. J. and Leskovec, J. (2012). Learning to discover social circles in ego networks. In NIPS, volume 2012, pages 548–56. McSherry, F. (2001). Spectral partitioning of random graphs. In Foundations of Computer Science, 2001. Proceedings. 42nd IEEE Symposium on, pages 529–537. IEEE. Newman, M. and Clauset, A. (2015). Structure and inference in annotated networks. arXiv preprint arXiv:1507.04001. 31 Newman, M. E. and Girvan, M. (2004). Finding and evaluating community structure in networks. Physical review E, 69(2):026113. Ng, A. Y., Jordan, M. I., Weiss, Y., et al. (2002). On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 2:849–856. Nowicki, K. and Snijders, T. A. B. (2001). Estimation and prediction for stochastic blockstructures. Journal of the American Statistical Association, 96(455):1077–1087. Rohe, K., Chatterjee, S., and Yu, B. (2011). Spectral clustering and the high-dimensional stochastic blockmodel. Ann. Statist., 39(4):1878–1915. Rohe, K., Qin, T., and Yu, B. (2016). Co-clustering directed graphs to discover asymmetries and directional communities. Proceedings of the National Academy of Sciences, 113(45):12679–12684. Shi, J. and Malik, J. (2000). Normalized cuts and image segmentation. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 22(8):888–905. Snijders, T. A. B. and Nowicki, K. (1997). Estimation and prediction for stochastic blockmodels for graphs with latent block structure. Journal of Classification, 14(1):75–100. Tang, M., Sussman, D. L., Priebe, C. E., et al. (2013). Universally consistent vertex classification for latent positions graphs. The Annals of Statistics, 41(3):1406–1430. Xu, Z., Ke, Y., Wang, Y., Cheng, H., and Cheng, J. (2012). A model-based approach to attributed graph clustering. In Proceedings of the 2012 ACM SIGMOD international conference on management of data, pages 505–516. ACM. Yan, B. and Sarkar, P. (2016). Convex relaxation for community detection with covariates. arXiv preprint arXiv:1606.01869. Yan, B., Sarkar, P., and Cheng, X. (2017). Exact recovery of number of blocks in blockmodels. arXiv preprint arXiv:1705.08580. Zhang, Y., Levina, E., and Zhu, J. (2015). Community detection in networks with node features. arXiv preprint arXiv:1509.01173. 32 A Details of DGCluster Here we detail our distance based greedy algorithm DGCluster. The idea behind DGCluster p is to note that, if ip and j are in the same community, then kCi? − Cj? k = 0, and otherwise kCi? − Cj? k = \|σ(i)\| + \|σ(j)\| = Θ( 2n/K) (when the communities are balanced). Thus, we expect to be able to cluster the vertices using dij = kĈproj,i? − Ĉproj,j? k, namely by starting with a root vertex r1 and for some threshold γ putting all vertices j satisfying dr1 j ≤ γ in the same cluster as r1 , and then picking another root vertex r2 from the remaining set, and putting all vertices j in the remaining set that satisfy dr2 j ≤ γ in the same cluster as r2 , and so on. Here a root vertex ri may be chosen as one of the vertices with the highest degree in the remaining set (or according to a degree-weighted random sampling scheme), to give importance to highly connected vertices. We also note that depending on the threshold γ the number of blocks we get can vary. In practice, we will start with small γ (yielding a large number of communities), and stop at the smallest γ that gives us ≥ K blocks. If we get more than K blocks, we merge, in succession, pairs of blocks having the largest intersectionpin Ĉ (relative to their sizes) until we get exactly K blocks. A rule of thumb would be to start with γ = c 2n/K with c small and then gradually increase c. Algorithm 5 A distance based greedy clustering algorithm: DGCluster 1: Input: C, C 0 , K. Output: σ = DGCluster(C, C 0 , K), a clustering based on distances between the rows of C 0 , with p merging guidance from C, if necessary. 2: Set θ = 2n/K. c = cmin = 0, δ = 0.01. 3: Set Know = n; 4: while flag = TRUE do 5: c ← c + δ. 6: σtemp ← NaiveCluster(C 0 , cθ). 7: Know = maxi σtemp (i). 8: if Know ≥ K then 9: flag ← TRUE. 10: σ ← σtemp . 11: else 12: flag ← FALSE. 13: Know ← maxi σ(i). 14: while Know > K do 15: Let Γi = σ −1 ({i}), i = 1, . . . , Know . 16: Compute the (upper triangle of the) matrix R, where P k∈Γi ,l∈Γj Ckl Rij ← . \|Γi \|\|Γj \| 17: 18: Pick (i? , j? ) ∈ arg maxi,j Rij . σ ← Merge(σ, i? , j? ). Algorithm 5 in turn makes use of the following two algorithms. 33 Algorithm 6 A naive clustering algorithm: NaiveCluster 1: Input: C, γ : a threshold. Output: σ = NaiveCluster(C, γ), a clustering based on distances between the rows of C. 2: Set Unassigned = {1, . . . , n}, b = 1, σ ≡ 0. 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: while Unassigned 6= ∅ do i ← a random (uniform or degree-weighted etc.) index in Unassigned. σ(i) ← b. Unassigned ← Unassigned \ {i}. for j ∈ Unassigned do Compute dij = kĈi? − Ĉj? k if dij ≤ γ then σ(j) ← b Unassigned ← Unassigned \ {j}. b ← b + 1. Algorithm 7 Merge 1: Input: σ, a, b. Output: σ = Merge(σ, a, b), a clustering with blocks a and b merged 2: u ← min{a, b}, v ← max{a, b} 3: for i = 1, . . . , n do 4: if σ(i) = v then 5: σ(i) ← u. 6: else if σ(i) > v then 7: σ(i) ← σ(i) − 1. B Some auxiliary results Lemma B.1 (Thresholding PACE). Let Ĉη := [Ĉ > η]. We have Eδ̃(Ĉη , C) ≤ Eδ̃(Ĉη , C) . min{η 2 , (1 − η)2 } (28) In particular, for η = 1/2, we have Eδ̃(Ĉ1/2 , C) ≤ 4Eδ̃(Ĉ, C). Proof. Note that \|(Ĉη )ij − Cij \| ∼ Ber(qij ), where qij = Cij P(Ĉij ≤ η) + (1 − Cij )P(Ĉij > η). Thus E((Ĉη )ij − Cij )2 = qij . Now E(Ĉij − Cij )2 = E(Ĉij − Cij )2 1{Ĉij ≤η} + E(Ĉij − Cij )2 1{Ĉij ≤η} . Therefore ( (1 − η)2 P(Ĉij ≤ η) = (1 − η)2 qij , E(Ĉij − Cij ) ≥ η 2 P(Ĉij > η) = η 2 qij , 2 34 if Cij = 1 if Cij = 0. This means E((Ĉη )ij − Cij )2 ≤ E(Ĉij − Cij )2 , min{η 2 , (1 − η)2 } which, on summing over i, j, gives us the desired bound. Lemma B.2 (Rounding GALE). Consider an estimated cluster membership matrix Ẑ ∈ [0, 1]n×K and the true cluster membership matrix Z ∈ {0, 1}n×K . Then kround(Ẑ) − Zk1 ≤ 2kẐ − Zk1 . Proof. Let S := {(i, k) ∈ [n]×[K] : \|Ẑik −Zik \| ≥ 1/2}. Then by Markov’s inequality, we have \|S\| ≤ 2kẐ−Zk1 . Note that for (i, k) ∈ [n] × [K] \ S, round(Ẑik ) = Zik . Thus, X \|round(Ẑik ) − Zik \| ≤ \|S\| ≤ 2kẐ − Zk1 kround(Ẑ) − Zk1 = (i,k)∈S We shall need the following Bernstein-type concentration result for hypergeometric random variables. Lemma B.3 (Corollary 1 of Greene and Wellner (2017) (restated here using slightly different notation)). Consider a random variable H ∼ Hypergeometric(k, `, L). Let µ = `/L, σ 2 = µ(1 − µ), f = (k − 1)/(L − 1). Then for λ > 0, ! √ H λ2 /2 P k − µ > λ ≤ exp − . λ k σ 2 (1 − f ) + 3√ k Let us deduce from this two convenient Chernoff type bounds on the upper and lower tail of a hypergeometric variable. Lemma B.4. Consider a random variable H ∼ Hypergeometric(k, `, L). Then for 0 < < 1, we have 2 2 EH EH max{P (H < (1 − )EH) , P (H > (1 + )EH)} ≤ exp − ≤ exp − . 2(1 + /3) 4 Proof. Note that k − H ∼ Hypergeometric(k, L − `, L). Using Lemma B.3 we get P Taking √λ k = ` L , H ` λ − < −√ k L k ! √ k−H L−` λ2 /2 k − > λ ≤ exp − ` =P . λ √ k L L + 3 k we get k` P H < (1 − ) L `3 k2 /(2L2 ) 2 k` ≤ exp − = exp − . (1 + /3)`/(L) 2(1 + /3)L This gives us the desired bound for the lower tail, the upper tail can be handled similarly. B.1 Analysis of PACE under stochastic block model Recall that we need to know how an algorithm performs on a randomly selected subgraph under our subgraph selection procedure. Here we will discuss how one can obtain such guarantees under the stochastic blockmodel. This will require us to understand the behavior of sizes of different communities in a (randomly) chosen subgraph. 35 B.1.1 Community sizes in random m-subgraphs (The results of this subsection do not depend on any modeling assumption.) Let S be a random m(S) (S) (S) subgraph, and set m/n = q. Then if (n1 , . . . , nK ) is the cluster size vector for ZS , then clearly nk ∼ Hypergeometric(m, nk , n) and therefore, by Lemma B.4, ((nk − nmin )q + ∆)2 (S) P(nk ≤ nmin q − ∆) ≤ exp − 4nk q 2 ∆ . ≤ exp − 4nmax q Therefore, by union bound, (S) P(nmin ≤ nmin q − ∆) ≤ X (S) P(nk ≤ nmin q − ∆) k ∆2 ≤ K exp − 4nmax q Choosing ∆ = √ . 4r0 nmax q log n, where r0 > 0, we see that with probability at least 1 − (S) nmin ≥ nmin q − p 4r0 nmax q log n K nr 0 we have (29) Similarly, we can show that P(n(S) max and then, taking ∆ = ≥ nmax q + ∆) ≤ K exp − , √ 0 4r nmax q log n, conclude that with probability at least 1 − n(S) max ≤ nmax q + B.1.2 ∆2 4nmax q K nr0 p 4r0 nmax q log n. (30) Community sizes in 1-hop ego neighborhoods Let now S be a randomly chosen 1-hop ego neighborhood. Note that the size of the k-th block in this neighborhood satisfies X (S) nk = Xj , j : σ(j)=k where Xj = 1{j∈S} . Now it is not hard to see that the Xj ’s are independent conditional R, the root of S. We also have that E(Xj \| R) = Bσ(j)σ(R) 1(j 6= R). It follows that (S) E(nk \| R) = X X E(Xj \| R) = j : σ(j)=k Bσ(j)σ(R) 1{j6=R} , j : σ(j)=k which means that (S) (nmin − 1)B∗ ≤ (nk − 1)B? ≤ E(nk \| R) ≤ nk B# ≤ nmax B# . 36 Therefore, by Chernoff’s inequality, (S) (S) P(nk ≤ (nmin − 1)B? − ∆ \| R) ≤ exp − ≤ exp − (E(nk \| R) − nmin B? + ∆)2 ! (S) 2E(nk \| S) . ∆2 2nmax B# Since the right hand side does not depend on R, we can take expectations of both sides with respect to R to get ∆2 (S) P(nk ≤ (nmin − 1)B? − ∆) ≤ exp − . 2nmax B# This implies, by an application of the union bound, that (S) P(nmin Choosing, ∆ = ∆2 ≤ (nmin − 1)B? − ∆) ≤ K exp − 2nmax B# p 2r0 nmax B# log n we conclude, with probability at least 1 − (S) nmin ≥ (nmin − 1)B? − p . K , nr0 that 2r0 nmax B# log n. (31) One can prove similarly that P(n(S) max and then, taking ∆ = ≥ nmax B# + ∆) ≤ K exp − , p 3r0 nmax B# log n, conclude, with probability at least 1 − n(S) max ≤ nmax B# + B.1.3 ∆2 3nmax B# K , nr0 that p 3r0 nmax B# log n. (32) Analysis of adjacency spectral clustering We shall use the community size estimates from the previous subsection along with Lemma 3.1. Random m-subgraphs: From (29) and (30) we have, with probability at least 1 − n2K r 0 , that √ nmax q + 4r0 nmax q log n nmax √ 0 ≤ ≤ 2 Cn,m,r0 , 2 (S) 2 nmin q (nmin q − 4r nmax q log n) (nmin ) (S) nmax where s Cn,m,r0 = 1+ 4r0 log n nmax q n2 s ! 1− q 4r0 nmax log n n2min q !−2 . n2 q Note that Cn,m,r0 stays bounded, e.g., by 6, if nmaxminlog n ≥ 16r0 . In fact, it approaches 1 as nmaxminlog n → ∞. 1 Thus, from (7), with probability at least 1 − n2K r 0 over randomness in S and with probability at least 1 − mr over randomness in AS , we have 1 K nmax −1 δ(ẐS , ZS ) ≤ min c (2 + ) 2 Cn,m,r0 , 1 . 2 λ αn n2min q 37 We conclude that 2K 1 K nmax 1 −1 Cn,m,r0 , 1 + r0 + r . Eδ(ẐS , ZS ) ≤ min c (2 + ) 2 2 2 λ αn nmin q m n 1-hop ego neighborhoods: By (7), given S (non-empty), with probability at least 1 − randomness in AS , we have ( ) (S) 1 n K max δ(ẐS , ZS ) ≤ min c−1 (2 + ) 2 , 1 =: JS . 2 λ αn (n(S) )2 (33) 1 \|S\|r over the min So 1 1 E[δ(ẐS , ZS ) \| S] ≤ JS + . 2 \|S\|r Therefore 1 Eδ(ẐS , ZS )1(\|S\|≥m? ) ≤ EJS 1(\|S\|≥m? ) + E 2 1(\|S\|≥m? ) \|S\|r . From (31) and (32) we have, with probability at least 1 − n2K r 0 , that p (S) nmax B# + 3r0 nmax B# log n nmax B# nmax nmax p ≤ Dn,B,r0 = 2 Dn,B,r0 , ≤ 2 2 (S) 2 0 2 nmin B? nmin λ2 αn ((nmin − 1)B? − 2r nmax B# log n) (nmin ) where !−2 2r0 nmax B# log n Dn,B,r0 = 1 + 1− − nmin n2min B?2 s ! !−2 r 3r0 log n 2r0 nmax log n 1 = 1+ . 1− − nmax αn nmin n2min λ2 αn q n2min λ2 αn 0 stays bounded, e.g., by 16(1 + 38 ), if nmax log n ≥ 8r . In fact, it approaches 1 as s Note that Dn,B,r0 n2min λ2 αn nmax log n 3r0 log n nmax B# ! 1 s → ∞. Thus −1 EJS 1(\|S\|≥m? ) ≤ min c K nmax 2K (2 + ) 4 2 2 Dn,B,r0 , 1 + r0 . λ αn nmin n On the other hand, E 1(\|S\|≥m? ) \|S\|r ≤E =E 1(\|S\|≥1) \|S\|r ! 1( (n−1)B? >\|S\|≥1) 2 \|S\|r +E 1(\|S\|≥ (n−1)B? ) ! 2 \|S\|r r (n − 1)B? 2 ≤ P \|S\| < + 2 (n − 1)B? r nB?2 2 ≤ exp − + (by (20)) 16B# (n − 1)B? r 16B# 4 r 16 4 ≤ + = + . nB?2 nB? nλ2 αn nλαn 38 So, finally, we have r 4 16 2K 1 −1 K nmax + Eδ(ẐS , ZS )1(\|S\|≥m? ) ≤ min c . Dn,B,r0 , 1 + r0 + 2 λ4 αn2 n2min nλ2 αn nλαn n B.1.4 Analysis of SDP a n In the setting of Corollary 3.5, for a random m-subgraph S we have mink Bkk ≥ b n (34) b̃ m am n , (S) nkk0 bm n . = ã 0 m , maxk6=k Bkk0 ≤ = where ã = b̃ = Let denote the number of pairs of vertices in the subgraph S such that one of them is from community k and the other is from community k 0 . Now by Lemma B.4 (S) P(nk 2 mπ ≥ cnk m/n) ≤ e−(c−1) min /4 . 2 So with probability at least 1 − Ke−(c−1) mπmin /4 we have, for all 1 ≤ k ≤ k 0 ≤ K, that ( (S) (S) for k < k 0 , n n 0 ≤ c2 nk n0k m2 /n2 ≤ c1 nkk0 m2 /n2 , (S) nkk0 = 1 k (S) k (S) 1 2 2 for k = k 0 , 2 nk (nk − 1) ≤ 2 cnk m/n(cnk m/n − 1) ≤ c1 nkk m /n , where c1 = c2 1 + cm −1 n cm (nπ min −1) n 0 2 mπ . Similarly, we can show that, with probability at least 1−Ke−(1−c ) (S) nkk0 ( c2 nkk0 m2 /n2 , ≥ c2 nkk m2 /n2 , min /4 for k < k 0 , for k = k 0 , c0 m −1 n (S) where c2 = 1 + c0 m . Let G = {S \| c2 nkk0 m2 /n2 ≤ nkk0 ≤ c1 nkk0 m2 /n2 , for all 1 ≤ k ≤ (nπmax −1) n P (S) g̃ 2 k 0 ≤ K}. Given S, the expected variance of edges in S is m(m−1) 1≤k≤k0 ≤K Bkk0 (1 − Bkk0 )nkk0 =: m , say. Then, if S ∈ G, we have (c0 )2 c2 g c2 m(n − 1) g g̃ c1 m(n − 1) g m ≤ × ≤ ≤ × = 2 × ḡ. n (m − 1)n n m (m − 1)n n n 2 and so, using our assumptions on a, b, g, we get that g̃ ≥ 9 and ã, b̃ satisfy (ã − b̃)2 ≥ 484−2 m ḡ ≥ 484−2 g̃. n2 Therefore, using Lemma 3.2 on the subgraph S, we conclude that, given S ∈ G, we have 1 kĈ (S) − Ĉ (S) k2F ≤ , m2 with probability at least 1 − e3 5−m , where Ĉ (S) is a solution of SDP-GV on the subgraph S. As P(S ∈ / G) ≤ 2 0 2 Ke−(c−1) mπmin /4 + Ke−(1−c ) mπmin /4 , we conclude that Eδ̃(Ĉ (S) , C (S) ) ≤ + e3 5−m + Ke−(c−1) B.2 2 mπ min /4 0 2 mπ + Ke−(1−c ) min /4 . Proofs of some results for GALE In this section we collect the proofs of the auxiliary results we presented in Section 6.2. 39 Proof of Lemma 6.3. Note that the Ni ’s are i.i.d. Binomial(T ; m/n) random variables and hence by Chernoff’s inequality we have, 2 P(Ni < τ ) ≤ e−(1−θ) T m/2n . (35) Another use of Chernoff gives, for δ > 0, ! P X 1{Ni <τ } > nP(N1 < τ )(1 + δ) ≤ exp(−δ 2 nP(N1 < τ )/3). i Setting δ = q 3r log n nP(N1 <τ ) , we conclude that with probability at least 1 − X 1{Ni <τ } ≤ nP(N1 < τ ) + 1 nr , p 3rnP(N1 < τ ) log n i ≤ ne −(1−θ)2 T m/2n r 1+ 3r log n ne−(1−θ)2 T m/2n ! . Proof of Lemma 6.4. We will first show that for three different nodes a, b, c in Sm,T the overlap variables Yab and Yac are independent. This is an immediate consequence of the following two facts: (i) Yab and y (a) are independent which follows from the observation that for any m-subset S of [n], Yab \| y (a) = 1S ∼ Hypergeometric(m, m, n). (ii) Given y (a) , the overlaps Yab and Yac are independent, which follows from the fact that y (b) and y (c) are independent. From the discussion above, Yab ∼ Hypergeometric(m, m, n). Using Lemma B.4 we get, with = 1/2, we get that 2 m m2 P(a ∼ b) = P Yab ≥ ≥ 1 − exp − . 2n 16n (S) Proof of Lemma 6.6. Notice that nk \| Yab ∼ Hypergeometric(Yab , nπk , n), which follows from the fact that given Yab , Sa ∩ Sb is distributed as a uniform Yab -subset of [n]† . So, Lemma B.4 gives us s ! ! 4r log n 1 (S) P nk < Yab πk 1 − Yab ≤ r . Yab πk n Therefore the same bound holds for the unconditional probability. Another application of Lemma B.4 for Yab ∼ Hypergeometric(m, m, n) gives us !! r m2 4rn log n 1 P Yab < 1− ≤ r. n m2 n † Thanks to Satyaki Mukherjee for making this observation, which makes the proof considerably shorter than our original approach based on concentration inequalities for sums of dependent Bernoulli random variables. 40 Using these two bounds, we conclude that with probability at least 1 − (S) nk as long as m2 πk n log n ≥ m2 π n s k 1−O 2 n log n m2 πk 2 nr , we have that !! , is large enough ( nmlogπkn ≥ 20r suffices to make the RHS above ≥ 0). 41	10math.ST
arXiv:1704.07811v1 [math.GR] 25 Apr 2017 R-triviality of some exceptional groups Maneesh Thakur Indian Statistical Institute, 7-S.J.S. Sansanwal Marg New Delhi 110016, India e-mail: maneesh.thakur@ gmail.com Abstract The main aim of this paper is to prove R-triviality for simple, simply connected 78 or E 78 , defined over a field k of arbitrary algebraic groups with Tits index E8,2 7,1 characteristic. Let G be such a group. We prove that there exists a quadratic extension K of k such that G is R-trivial over K, i.e., for any extension F of K, G(F )/R = {1}, where G(F )/R denotes the group of R-equivalence classes in G(F ), in the sense of Manin (see [23]). As a consequence, it follows that the variety G is retract K-rational and that the Kneser-Tits conjecture holds for these groups over K. Moreover, G(L) is projectively simple as an abstract group for any field extension L of K. In their monograph ([51]) J. Tits and Richard Weiss conjectured that for an Albert division algebra A over a field k, its structure group Str(A) is generated by scalar homotheties and its U -operators. This is known to 78 . We be equivalent to the Kneser-Tits conjecture for groups with Tits index E8,2 settle this conjucture for Albert division algebras which are first constructions, in affirmative. These results are obtained as corollaries to the main result, which shows that if A is an Albert division algebra which is a first construction and Γ its structure group, i.e., the algebraic group of the norm similarities of A, then Γ(F )/R = {1} for any field extension F of k, i.e., Γ is R-trivial. Keywords: Exceptional groups, Algeraic groups, Albert algebras, Structure group, Kneser-Tits conjecture 1 Introduction The main aim of this paper is to prove R-triviality of certain exceptional algebraic groups. More precisely, let G be a simple, simply connected algebraic group defined over a field k of 78 or E 78 (see [48] for the index notation). Groups arbitrary characteristic, with Tits index E8,2 7,1 with these indices exist over a field k if k admits a central division algebra of degree 3 with non-surjective reduced norm, or if k admits degree 3 central division algebras with unitary involutions over a quadratic extension of k, with non-surjective reduced norm map (see [46], 78 , E 78 do not exist over finite 3.2, 3.3.1). Note also that the groups with Tits indices E8,2 7,1 fields, number fields, algebraically closed fields, local fields, field of reals. However, these exist, for example, over Q(t), the function field over Q in one variable. 1 2 Maneesh Thakur 78 are classified by isotopy classes of Albert division algebras over k Groups with index E8,2 and their associated buildings are Moufang hexagons defined by a hexagonal system of type 27/K where K is either a quadratic extension of k or K = k (see [51], [42]). Groups with 78 are classified by isotopy classes of Albert division algebras and they are associated index E7,1 to Moufang sets M(A), for A an Albert division algebra (e.g. see [43] or [2]). Let G be an algebraic group over k, with Tits index as one of the above mentioned indices. We prove that there exists a quadratic field extension K of k such that G is R-trivial over K, i.e., for any field extension L of K, G(L)/R = {1} in the sense of Manin ([23]). Since E8 has 78 , there is a quadratic extension K of k trivial center, it follows that when G has index E8,2 such that G(L) is a simple (abstract) group for any field extension L of K. It is well known (see [46]) that the anisotropic kernels of these groups correspond (upto a central torus) to the structure groups of Albert division algebras over k. Albert division algebras over a field k are obtainable from two (rational) constructions due to Tits, called the first construction and the second construction. We prove that if the anisotropic kernel of G (for G as above) comes from a first construction Albert division algebra, then G is R-trivial over k. In this situation G splits over a cubic extension of k. The Kneser-Tits conjecture predicts that if G is a connected, simple, simply connected algebraic group, defined and isotropic over a field k, then G(k)/G(k)† = {1}, where G(k)† is the subgroup of G(k) generated by the k-rational points of the unipotent radicals of kparabolics of G (see [33], §7.2, [13]). By a theorem of Gille, this is equivalent to proving G(k)/R = {1} ([13], Thm. 7.2). A conjecture of J. Tits and Richard M. Weiss asserts that if A is an Albert division algebra over a field k, then its structure group Str(A) is generated by scalar homotheties and U -operators (defined in §2) of A (see [51], 37.41, 37.42 and page 418). This conjecture 78 (see [51]). is equivalent to the Kneser-Tits conjecture for groups with Tits index E8,2 Our results prove the Tits-Weiss conjecture when A is a first Tits construction. For an extensive survey and new results on the Kneser-Tits problem, we refer the reader to ([13]) along with the classics ([45], [47]) and ([33]) and the paper ([38]) for reduction to the relative rank one case. We mention some recent results: The case of simply connected group with 29 was settled in ([12]), where a modern version (in terms of W -triviality) of F. D. index 2 E6,1 Veldkamp’s proof ([49], [50]) is presented. The case of the trialitarian forms of D4 over global fields was settled in ([37]). The paper [37]) also presents a description of the Whitehead group (over an infinite field k) of an absolutely simple, simply connected algebraic group of type 3,6 D4 of k-rank 1. The result for such groups over a general field is covered in ([13], 66 over an arbitrary field was §6.1). The Kneser-Tits conjecture for groups with Tits index E8,2 settled in ([36]). This paper is the second in a series of papers dedicated to this problem. In the first paper ([44]) we proved, among other results, that the conjecture is true for pure first constructions. We had also proved that for such an Albert division algebra A over k, the group Aut(A) is R-trivial. The problem of R-equivalence for semisimple algebraic groups has been studied by several mathematicians, see for example ([4], [5], [7], [8], [3]) and the references therein. This connects with some algebro-geometric properties of algebraic groups; e.g. if G is a simple, simply connected algebraic group, isotropic and defined over a field k, then G is R-trivial if and only if the variety of G is retract k-rational (see [13], Thm. 5.9). 78 or E 78 , the Remark : For a simple, simply connected algebraic group G with Tits index E8,2 7,1 (semisimple) anisotropic kernel H of G is isomorphic to Isom(A), the full group of isometries of the norm form of an Albert division algebra A over k. The anisotropic kernel H of G is R-triviality 3 split by a cubic extension of k if and only if A is a first Tits construction (see Proposition 2.2 and Corollary 2.1). Retract rationality : By ([13]), our results also prove that if G is a simple, simply connected 78 or E 78 , then there exists a quadratic algebraic group defined over k with Tits index E8,2 7,1 extension K of k such that the underlying variety of G is retract K-rational. If the anisotropic kernel of G is split over a cubic extension of k, then G is retract k-rational. Plan of the paper : In this paper, k will always denote a base field, infinite of arbitrary characteristic. In §2, we give a quick introduction to Albert algebras, Tits processes and the structure group of an Albert algebra. A description of strongly inner groups of type E6 and those 78 and E 78 is given in detail. This is followed by a brief introduction to with Tits index E8,2 7,1 R-equivalence on algebraic groups. At the end of §2, we state our main results and some of their consequences. §3 and §4 contain some of the key results of the paper. §3 is mainly concerned with extension of automorphisms of a 9-dimensional subalgebra of an Albert algebra A to an automorphism of A and also discusses R-triviality of some subgroups of Aut(A). §4 contains results on extending norm similarities of 9-dimensional subalgebras to those of the Albert algebra (see Theorem 4.5). Some of the most important results on R-triviality are proved §5. Here we develop the R-triviality results on various subgroups of Aut(A) and Str(A) for Albert algebras in general as well as for the special case of first constructions. These results culminate in the main theorem (Theorem 5.4) , where we prove Str(A) is R-trivial for first Tits construction Albert algebras. §6 discusses a proof of the R-triviality of groups of type 78 , with anisotropic kernel split over a cubic extension. A proof of the Tits-Weiss conjecture E8,2 for Albert division algebras, arising from first Tits construction is presented (Corollary 6.1). 78 , having anisotropic kernel split over a cubic In §7, we discuss groups with Tits index E7,1 extension. We prove that such groups are R-trivial. We end §with some R-triviality results for reduced Albert algebras. In §8 some retract rationality results are derived. We prove, as a consequence of our results 78 or E 78 and the anisotropic kernel is in §6 and §7, that any group G whose Tits index is E7,1 8,2 split over a cubic extension of k, is retract k-rational. 2 Preliminaries and results For much of the preliminary material on Albert algebras, we refer the reader to ([27]), ([39]) and ([21]). We recall below, very briefly, some notions we will need. All base fields considered in this paper will be assumed to be infinite and of arbitrary characteristic. To define the notion of an Albert algebra over a field of arbitrary characteristics, we take the approach via cubic norm structure, as defined in ([32] or [30]) . Cubic norm structures : Let J be a finite dimensional vector space over a field k. A cubic norm structure on J consists of a tuple (N, #, c), where c ∈ J is a base point, called the identity element of the norm structure and 1. N : J → k is a cubic form on J, 2. N (c) = 1, 3. the trace form T : J × J → k, defined by T (x, y) := −∆xc ∆y logN , is nondegenerate, 4 Maneesh Thakur 4. the adjoint map # : J → J defined by T (x# , y) = ∆yx N , is a quadratic map such that 5. x## = N (x)x, 6. c# = c, 7. c×x = T (x)c−x, where T (x) := T (x, c) is the trace of x and x×y := (x+y)# −x# −y # , and these hold in all scalar extensions of J. In the above, ∆yx f denotes the directional derivative of a polynomial function f on J, in the direction y, evaluated at x. For differential calculus of rational maps and related notation, we refer to ([16], Chap. VI, Sect. 2). Let, for x ∈ J, Ux (y) := T (x, y)x − x# × y, y ∈ J. Then with 1J := c and Ux as above, one gets a unital quadratic Jordan algebra structure on J (see [24]), sometimes denoted by J(N, c). The linear operators Ux as defined above, are called the U -operators of the Jordan algebra J. An element x ∈ J is invertible if and only if N (x) 6= 0 and in that case one has x−1 = N (x)−1 x# . Recall that J(N, c) is called a division algebra if Ux is surjective for all x 6= 0, which is equivalent to N (x) 6= 0 for all x 6= 0. We list some examples and notation that we will use in the paper: Example 1. Let D denote a separable associative algebra over k, of degree 3. Let ND denote its norm and TD the trace. Then with 1D the unit element of D and # : D → D the (classical) adjoint map, we get a quadratic Jordan algebra structure on D, which we will denote by D+ . Example 2. Let (B, σ) be a separable associative algebra over k with an involution σ of the second kind (over its center). With the unit element 1 of B and the restriction of the norm NB of B to (B, σ)+ := {b ∈ B\|σ(b) = b}, we get a cubic norm structure on (B, σ)+ which is a substructure of B+ . Example 3. Let C be an octonion algebra over a field k with n as its norm map and t its trace. Let x 7→ x denote its canonical involution. Let γ = diag(γ1 , γ2 , γ3 ) ∈ GL3 (k) and H3 (C, γ) := {x = (xij ) ∈ M3 (C)\|γ −1 xt γ = x and xii ∈ k, 1 ≤ i ≤ 3}, where M3 (C) is the algebra of 3 × 3 matrices with It follows that any x ∈ H3 (C, γ) has the form  α1 γ2 c  x = γ1 c α2 γ1 b γ2 a entries in C and xt is the transpose of x.  γ3 b γ3 a , α3 where αi ∈ k, a, b, c ∈ C. Define N (x) := α1 α2 α3 − γ2 γ3 α1 n(a) − γ3 γ1 α2 n(b) − γ1 γ2 α3 n(c) + γ1 γ2 γ3 t(abc), 1 := e11 + e22 + e33 , and x# := α2 α3 − γ2 γ3 n(a) + α3 α1 − γ3 γ1 n(b) + α1 α2 − γ1 γ2 n(c) +(γ1 bc − α1 a)[23] + (γ2 ca − α2 b)[31] + (γ3 ab − α3 c)[12], 5 R-triviality where a[ij] := γj aeij + γi aeji in terms of the matrix units. This defines a cubic norm structure (see [30]), hence a Jordan algebra, called a reduced Albert algebra over k. We call a reduced Albert algebra split if the coordinate octonion algebra is split, in which case H3 (C, γ) ∼ = H3 (C) := H3 (C, 1). Tits process : 1. Let D be a finite dimensional associative k-algebra of degree 3 with norm ND and trace TD . Let λ ∈ k∗ . We define a cubic norm structure on the k-vector space D ⊕ D ⊕ D by 1 := (1, 0, 0), N ((x, y, z)) := ND (x) + λND (y) + λ−1 ND (z) − TD (xyz), (x, y, z)# := (x# − yz, λ−1 z # − xy, λy # − zx). The (quadratic) Jordan algebra associated to this norm structure is denoted by J(D, λ). We regard D+ as a subalgebra of J(D, λ) through the first summand. Recall (see [32], 5.2) that J(D, λ) is a division algebra if and only if λ is not a norm from D. This construction is called the first Tits process arising from the parameters D and λ. 2. Let K be a quadratic étale extension of k and B be a separable associative algebra of degree 3 over K with a K/k involution σ. Let u ∈ (B, σ)+ be a unit such that NB (u) = µµ for some µ ∈ K ∗ , where a 7→ a denotes the nontrivial k-automorphism of K. We define a cubic norm structure on the k-vector space (B, σ)+ ⊕ B by the formulae: N ((b, x)) := NB (b) + TK (µNB (x)) − TB (bxuσ(x)), (b, x)# := (b# − xuσ(x), µσ(x)# u−1 − bx), 1 := (1B , 0). The (quadratic) Jordan algebra corresponding to this cubic norm structure is denoted by J(B, σ, u, µ). We note that (B, σ)+ can be identified with a subalgebra of J(B, σ, u, µ) through the first summand. Recall (see [32], 5.2) that J(B, σ, u, µ) is a division algebra if and only if µ is not a norm from B. This construction is called the second Tits process arising from the parameters B, u, µ. Note that when K = k × k, then B = D × D ◦ and σ is the switch involution, where D ◦ is the opposite algebra of D. In this case, the second construction J(B, σ, u, µ) can be identified with a first construction J(D, λ). Albert algebras : In the Tits process (1) described above, if we start with a central simple algebra D and λ ∈ k∗ , we get a first Tits construction Albert algebra A = J(D, λ) over k. Similarly, in the Tits process (2) above, if we start with a central simple algebra (B, σ) with center a quadratic étale algebra K over k and an involution σ of the second kind, u, µ as described above, we get a second Tits construction Albert algebra A = J(B, σ, u, µ) over k. One knows that all Albert algebras can be obtained via Tits constructions. An Albert algebra is a division algebra if and only if its (cubic) norm N is anisotropic over k (see [21], §39). We also recall here that a first construction Albert algebra A is either a division algebra, or it is split, i.e., is isomorphic to H3 (C) with C the split octonion algebra k. Moreover, any Albert division algebra arising from a first construction is split over any of its cubic subfields. If A = J(B, σ, u, µ) as above, then A ⊗k K ∼ = J(B, µ) as K-algebras, where K is the center of B (see [21], 39.4). Hexagonal systems : A hexagonal system is a cubic norm structure whose norm is anisotropic. We say a hexagonal system is of type 27/k if it arises from a first Tits construction Albert division algebra over k, and of type 27/K, for K a quadratic field extension of k, 6 Maneesh Thakur if it arises from a Tits second construction Albert division algebra, corresponding to (B, σ)+ and K is the center of B. Pure first and second constructions : Let A be an Albert algebra over k. If A is a first construction, but not a second construction with respect to a degree 3 central simple algebra whose center is a quadratic field extension of k, then we call A a pure first construction Albert algebra. Similarly pure second construction Albert algebras are those which are not obtainable through the first construction. We recall at this stage that if A is an Albert division algebra, then any subalgebra of A is either k or a cubic subfield of A or of the form (B, σ)+ for a degree 3 central simple algebra B with an involution σ of the second kind over its center K, a quadratic étale extension of k (see [16], Chap. IX, §12, Lemma 2, [30]). Norm similarities of Albert algebras : Let A be an Albert algebra over k and N denote its norm map. A norm similarity of A is a bijective linear map f : A → A such that N (f (x)) = ν(f )N (x) for all x ∈ A and some ν(f ) ∈ k ∗ . Since we are working over k infinite, norm similarities are synonimous with isotopies of Albert algebras (see [16], Chap. VI, Thm. 6, Thm. 7). Adjoints and U -operators : Let A be an Albert algebra over k and N be its norm map. Recall, for a ∈ A the U -operator Ua is given by Ua (y) := T (a, y)a − a# × y, y ∈ A. When a ∈ A is invertible, it can be shown that Ua is a norm similarity of A, in fact, for any x ∈ A, N (Ua (x)) = N (a)2 N (x) when a is invertible. For a central simple algebra D of degree 3 over k, the adjoint map a 7→ a# satisfies ND (a) = aa# = a# a, ∀ a ∈ D. It can be shown that (xy)# = y # x# for all x, y ∈ D. It also follows that N (x# ) = N (x)2 . Algebraic groups from Albert algebras : In this paper, for a k-algebra X and a field extension L of k, We will denote by XL the L-algebra X ⊗k L. Let A be an Albert algebra over k with norm N and k be an algebraic closure of k. The full group of automorphisms Aut(A) := Aut(Ak ) is a simple algebraic group of type F4 defined over k and all simple groups of type F4 defined over k arise this way. When A is a division algebra, Aut(A) is anisotropic over k. We will denote the group of k-rational points of Aut(A) by Aut(A). Note that A is a division algebra precisely when the norm form N of A is anisotropic (see [40], Thm. 17.6.5). The full group Str(A) of norm similarities of N , called the structure group of A, is a connected reductive group over k, of type E6 . We denote by Str(A) the group of krational points Str(A)(k). The U -operators Ua , for a ∈ A invertible, are norm similarities and generate a normal subgroup of Str(A), called the inner structure group of A, which we will denote by Instr(A). The commutator subgroup Isom(A) of Str(A) is the full group of isometries of N and is a simple, simply connected group of type E6 , anisotropic over k if and only if N is anisotropic, if and only if A is a division algebra (see [39], [40]). Note that Aut(A) is the stabilizer of 1 ∈ A in Str(A). On occasions, we will need to treat an algebra X as an algebraic variety (particularly when dealing with R-equivalence etc.). In such a situation, if no confusion is likely, we shall continue to denote the underlying (affine) space by X. Morphisms X → Y or rational maps would carry similar meanings. Base change of an object X defined over a base field k to an extension L of k will be denoted by XL . We record the following observations with proofs : Proposition 2.1. Let A be an Albert division algebra over k. Then Aut(A), Str(A) and Isom(A) are split by a degree 6 extension of k. 7 R-triviality 1. A is a first construction if and only if A ( hence Aut(A) ) has a degree 3 splitting field and, A is a pure first construction if and only if every minimal splitting field of A is cubic cyclic. 2. A is a pure second construction if and only if A ( hence Aut(A) ) has no splitting field of degree 3 over k. Proof. Write A as a second construction A = J(B, σ, u, µ) for suitable parameters as decribed in the preliminaries. If A is a first construction, any cubic subfield of A splits A (see [28], Thm. 4.7), hence splits Aut(A) and the statement follows. Let K be the center of B, which must be a quadratic field extension of k if A is not a first construction. Base changing to K we have A ⊗ K ∼ = J(B, µ). By the argument presented, there is a cubic extension L of K that splits A ⊗ K. Hence A is split by the degree 6 extension LK of k. To prove (1), note that if A is a first construction, A has a cubic splitting field. Conversely, if A has a degree 3 splitting field L over k, then the invariant f3 (A) ∈ H 3 (k, Z/2) vanishes over L and hence over k by Springer’s theorem. Hence f3 (A) = 0 and A is a first construction ([21], 40.5). The assertion about pure first construction follows from ([44], Thm. 2.1). Proof of (2) follows from (1). Proposition 2.2. Let A be an Albert division algebra over k. Then A is a first Tits construction if and only if Str(A) (equivalently Isom(A)) is split by a cubic extension of k. Proof. First assume that A is an Albert division algebra over k arising from a first Tits construction. Then A is split by any cubic extension L ⊂ A of k (see [28], Thm. 4.7) and hence, by ([40], Thm. 17.6.3), L also splits the structure group Str(A) of A. Conversely, if Str(A) splits over a cubic extension L of k, then, by ([40], Thm. 17.6.5), the Albert algebra A is reduced over L. By ([40], Thm. 17.6.3), the stabilizer of 1 ∈ A ⊗k L is the split group of type F4 over L and equals Aut(A ⊗k L). Hence, by ([21], 40.5), Aut(A ⊗k L) has its mod-2 invariant f3 (A ⊗k L) = f3 (A) ⊗k L = 0. It follows that the coordinatizing octonion algebra of A ⊗k L is split over L. Hence, by Springer’s theorem, f3 (A) = 0. Therefore A must be a first Tits construction over k ([21], 40.5). Strongly inner group of type E6 : Let H be a simple, simply connected algebraic group defined and strongly inner of type E6 over k (see [46] for definition). Then H is isomorphic to Isom(A), the (algebraic) group of all norm isometries of an Albert algebra A over k. Moreover, H is anisotropic if and only if A is a division algebra (see [40], Thm. 17.6.5). 78 : Let G be a simple, simply connected algebraic group with Groups with Tits index E8,2 78 over k. The index has associated diagram as below: Tits index E8,2 • • • • • • • • The anisotropic kernel of such a group is a simple, simply connected, strongly inner group of type E6 and by ([51], [46]), can be realized as the group Isom(A) of norm isometries of an Albert division algebra A over k. Let Γ denote the Tits building associated to G. Then Γ is a Moufang hexagon associated to a hexagonal system of type 27/F , where F is a quadratic field extension K of k (this is when A is written as a second construction J(B, σ, u, µ) with Z(B) = K) or F = k (when A is a first construction). The group G(k) is the group of “linear” 78 as automorphisms of a automorphisms of Γ. The explicit realization of groups of type E8,2 Moufang hexagon as above is due to Tits and Weiss and is detailed in their monograph ([51]). 8 Maneesh Thakur 78 : Let H be a simple, simply connected algebraic group over Groups with Tits index E7,1 78 k with Tits index E7,1 over k. Then the anisotropic kernel of H is a simple, simply connected strongly inner group of type E6 and is isomorphic to Isom(A) for an Albert division algebra A over k (see [46]). The Tits index of H has associated diagram as below: • • • • • • • Following Max Koecher ([20]), one can realize H as follows: fix an Albert divsion algebra A over k. Let Ξ(A) be the (algebraic) group of birational transformations of A generated by the structure group Str(A) and the maps x 7→ ta (x) = x + a ( a ∈ Ak ), x 7→ j(x) = −x−1 (x ∈ Ak ). 78 and any algebraic group over k with this Tits index arises this way Then Ξ(A) has type E7,1 (see [46]). Any element of f ∈ Ξ(A) is of the form f = w ◦ ta ◦ j ◦ tb ◦ j ◦ tc , w ∈ Str(A), a, b, c ∈ Ak . This was proved by Koecher when k is infinite and has characteristic 6= 2 and by O. Loos in arbitrary chatacteristics ([22]), in the context of Jordan pairs. Remark : In light of Proposition 2.1 and Proposition 2.2, the anisotropic kernel of a simple, 78 or E 78 , is split by a degree 6 simply connected group G defined over k, with Tits index E8,2 7,1 extension of k. We have 78 Corollary 2.1. Let G be a simple, simply connected algebraic group over k with index E8,2 78 or E7,1 , such that the anisotropic kernel H of G is split by a cubic extension of k. Then H∼ = Isom(A) for an Albert division algebra A over k arising from the first Tits construction. Proof. This is immediate from Proposition 2.2 and the above discussion. R-equivalence in algebraic groups : Let X be an irreducible variety over a field k with X(k) nonempty. Following Manin ([23]) we define points x, y ∈ X(k) to be R-equivalent if there exists a sequence x0 = x, x1 , · · · , xn = y of points in X(k) and rational mape fi : A1k → X, 1 ≤ i ≤ n, defined over k and regular at 0 and 1, such that fi (0) = xi−1 , fi (1) = xi . When G is a connected algebraic group defined over k, the set of elements of G(k) which are R-equivalent to 1 ∈ G(k) is a normal subgroup RG(k) of G(k). The set G(k)/R of Requivalence classes in G(k) is in natural bijection with the quotient G(k)/RG(k) and therefore carries a natural group structure, and we identify G(k)/R with the group G(k)/RG(k). This group measures the obstruction to rational parameterizing of points of G(k). We say G is R-trivial if G(L)/R = {1} for all field extensions L of k. Recall that a variety X defined over k is said to be k-rational if X is birationally isomorphic over k to an affine space. It is well known (see [52], Chap. 6, Prop. 2) that if G is k-rational then G is R-trivial. Let G be a connected reductive group defined over k and assume G is k-isotropic, i.e., admits a k-embedding Gm ֒→ G such that the image is not contained in Z(G), the center of G. Let G(k)† denote the subgroup of G(k) generated by the k-rational points of the unipotent radicals of the k-parabolic subgroups of G. Then G(k)† is normal in G(k) and the quotient W (k, G) = G(k)/G(k)† is the Whitehead group of G. R-triviality 9 By a theorem of Gille ([13], Thm. 7.2) if G is a semisimple, simply connected and absolutely almost simple group, defined and isotropic over k, then W (k, G) ∼ = G(k)/R. The Kneser-Tits problem asks if for such a group G, its Whitehead group is trivial, or more generally, to compute the Whitehead group. In this paper, we prove 78 or Theorem 2.1. Let G be a simple, simply connected algebraic group over k of type E7,1 78 or strongly inner type E , whose anisotropic kernel is split by a cubic extension of k. E8,2 6 Then G is R-trivial. Corollary 2.2. Let G be a simple, simply connected algebraic group, defined and isotropic 78 or E 78 , and anisotropic kernel split by a cubic extension over a field k, with Tits index E8,2 7,1 of k. then the Kneser-Tits conjecture holds for G. Moreover, G(L) is projectively simple, for any field extension L or k. Corollary 2.3. Let G be a simple, simply connected algebraic group, defined and isotropic 78 or E 78 . Then there exists a quadratic field extension K over a field k, with Tits index E8,2 7,1 of k such that G is R-trivial over K, i.e., for any extension F of K, G(F )/R = {1}. We will prove this result in steps. The key result in proving the above theorem is our Main theorem: Theorem 2.2. Let A be an Albert division algebra over k arising from the first Tits construction. Then Str(A) is R-trivial. Corollary 2.4. Tits-Weiss conjecture : Let A be an Albert division algebra arising from the first Tits construction. Then Str(A) = C.Instr(A), where C is the subgroup of Str(A) consisting of scalar homotheties Ra : A → A, a ∈ k∗ . 3 Automorphisms and R-triviality Let A be an Albert algebra over k and S a subalgebra of A. We will denote by Aut(A/S) the (closed) subgroup of Aut(A) consisting of automorphisms of A which fix S pointwise and Aut(A, S) will denote the closed subgroup of automorphisms of A stabilizing S. The group of k-rational points of these groups will be denoted by ordinary fonts, for example Aut(A) = Aut(A)(k) and so on. When A is a division algebra, any proper subalgebra S of A is either k or a cubic subfield or a 9-dimensional (degree 3) Jordan subalgebra, which is of the form D+ for a degree 3 central division algebra D over k or of the form (B, σ)+ for a central division algebra B of degree 3 over a quadratic extension K of k, with an involution σ of the second kind over K/k (cf. Thm. 1.1, [31]). We record below a description of some subgroups of Aut(A) which we shall use in the sequel, we refer to ([21], 39.B) and ([11]) for details. Proposition 3.1. Let A be an Albert algebra over k. 1. Suppose S = D+ ⊂ A for D a degree 3 central simple k-algebra. Write A = J(D, µ) for some µ ∈ k∗ . Then Aut(A/S) ∼ = SL(1, D), the algebraic group of norm 1 elements in D. 2. Let S = (B, σ)+ ⊂ A for B a degree 3 central simple algebra of degree 3 over a quadratic extension K of k with an involution σ of the second kind over K/k. Write A = J(B, σ, u, µ) for suitable parameters. Then Aut(A/S) ∼ = SU(B, σu ), where σu := Int(u) ◦ σ. 10 Maneesh Thakur In particular, the subgroups described in (1) and (2) are simply connected, simple of type A2 , defined over k and hence are rational, therefore are R-trivial. The last assertion is due to the fact that connected reductive groups of absolute rank at most 2 are rational (see [6]). For an algebraic group G defined over k, we denote the (normal) subgroup of elements in G(k) that are R-equivalent to 1 ∈ G(k) by R(G(k)). Extending automorphisms : Let A be an Albert (division) algebra and S ⊂ A a subalgebra. Then, under certain conditions, we can extend automorphisms of S to automorphisms of A (e.g. when S is nine dimensional, see [34], Thm. 3.1, [26], Thm. 5.2). We need certain explicit extensions for our results, we proceed to describe these (cf. [21], 39.B). We will need the following important result in the sequel frequently, so we record it below ([25] or [21], 40.13 and Lemma 5.4 of [26] for arbitrary characteristics): Lemma 3.1. Let (B, σ) be a degree 3 central simple algebra with center K, a quadratic étale algebra and σ an involution of the second kind. Let w ∈ B ∗ be such that λ := NB (w) satisfies λλ = 1. Then there exists q ∈ U (B, σ) such that λ = NB (q). First, let S be a 9-dimensional subalgebra of A. Then we may assume S = (B, σ)+ for a degree 3 central simple algebra with center K, a quadratic étale extension of k (see [31]). We then have A ∼ = J(B, σ, u, µ) for suitable parameters. Recall that any automorphism of S is of the form x 7→ pxp−1 with p ∈ Sim(B, σ), where Sim(B, σ) = {g ∈ B ∗ \|gσ(g) ∈ k∗ }. This is immediate from the fact that any norm similarity of S is of the form x 7→ γgxσ(g) for some γ ∈ k∗ and g ∈ B ∗ (see Thm. 5.12.1, [17]) and that among these, automorphisms are precisely the norm similarities that fix the identity element of S (see [18], Thm. 4). Recall the notation σu := Int(u) ◦ σ. We have, Proposition 3.2. Let A = J(B, σ, u, µ) be an Albert algebra over k, S := (B, σ)+ and K = Z(B) be as above. Let φ ∈ Aut(S) be given by φ(x) = gxg −1 for g ∈ Sim(B, σ) with gσ(g) = λ ∈ k∗ and ν := NB (g) ∈ K ∗ . Let q ∈ U (B, σu ) be arbitrary with NB (q) = ν̄ −1 ν. Then the map φe : A → A, given by (a, b) 7→ (gag −1 , λ−1 σ(g)# bq), is an automorphism of A extending φ. Proof. To show φe is an automorphism, we first note that φe is a k-linear bijection. Hence, e by ([24], Page 507), it suffices to check φ(1) = 1 and φe preserves norms. The first condition is obvious. Let us verify the second. We first note some formulae which will be used in the computation below: NB (σ(g)) = ν, NB (g# ) = NB (g)2 = ν 2 , NB (σ(g)# ) = ν 2 , NB (gσ(g)) = νν = λ3 , NB (g) = gg # = ν, λ2 = (gσ(g))# = σ(g)# g # , NB (q) = ν −1 ν. Also, since q ∈ U (B, σu ), we have 1 = qσu (q) = quσ(q)u−1 , qu = uσ(q)−1 . Now, for (a, b) ∈ A, we compute using the above formulae, e b)) = N ((gag −1 , λ−1 σ(g)# bq)) N (φ(a, 11 R-triviality = NB (gag −1 )+TK (µNB (λ−1 σ(g)# bq))−TB (gag −1 (λ−1 σ(g)# bq)uσ(λ−1 σ(g)# bq)) = NB (a)+TK (µλ−3 ν̄ 2 NB (b)NB (q))−TB (gag −1 (λ−1 σ(g)# bq)u(σ(q)σ(b)g# λ−1 )) = NB (a)+TK (µ(ν ν̄)−1 ν̄ 2 NB (b)ν̄ −1 ν)−TB (gag −1 λ−1 λ2 (g # )−1 bquσ(q)σ(b)νg−1 λ−1 ) = NB (a)+TK (µNB (b))−TB (gag −1 gν −1 bquσ(q)σ(b)νg−1 ) = NB (a)+TK (µNB (b))−TB (abuσ(b)) = N ((a, b)). Hence φe is an automorphism of A extending φ. Corollary 3.1. Let A = J(D, µ) be an Albert algebra, which is a first construction. Let φ ∈ Aut(D+ ) be given by φ(x) = gxg −1 for g ∈ D ∗ and x ∈ D+ . Then the map (x, y, z) 7→ (gxg −1 , gyh−1 , hzg−1 ), for any h ∈ D ∗ with ND (g) = ND (h), is an automorphism of A extending φ. Proof. We identify A = J(D, µ) with the second construction J(D × D ◦ , ǫ, 1, (µ, µ−1 )), where ǫ is the switch involution on B = D × D ◦ and apply the proposition. Note that, for g ∈ Sim(B, σ), the extension φe of φ as constructed above, can be simplified as follows: let p = λν −1 g, where gσ(g) = λ ∈ k∗ and ν = NB (g) ∈ K ∗ . Then λ3 = ν ν̄ and p ∈ U (B, σ), since pσ(p) = λν −1 gλν̄ −1 σ(g) = λ2 (ν ν̄)−1 (gσ(g)) = λ3 λ−3 = 1. e using gσ(g) = λ, σ(g# )g# = λ2 , gg# = ν, becomes, Hence the expression for φ, e b) = (gag −1 , λ−1 σ(g# )bq) φ(a, = (λν −1 gaλ−1 νg−1 , λ−1 ν̄σ(g)−1 bq) = (pap−1 , σ(λ−1 νg −1 )bq) = (pap−1 , σ(p)−1 bq) = (pap−1 , pbq). We have, Proposition 3.3. Let A = J(B, σ, u, µ) be an Albert algebra and S := (B, σ)+ ⊂ A and K = Z(B), a quadratic étale extension of k. Then any automorphism of A stabilizing S is of the form (a, b) 7→ (pap−1 , pbq) for p ∈ U (B, σ) and q ∈ U (B, σu ) with NB (p)NB (q) = 1. We have Aut(A, (B, σ)+ ) ∼ = [U (B, σ) × U (B, σu )]det /K (1) , where [U (B, σ) × U (B, σu )]det := {(p, q) ∈ U (B, σ) × U (B, σu )\|NB (p) = NB (q)}, and K (1) denotes the group of norm 1 elements in K, embedded diagonally. Proof. Let φ ∈ Aut(A, S) and let η := φ\|S ∈ Aut(S). Then η(a) = gag −1 for some g ∈ Sim(B, σ). Let λ := gσ(g) ∈ k∗ and ν := NB (g) ∈ K ∗ . We extend η to an automorphism ηe of A by the formula ηe(a, b) = (pap−1 , pbq ′ ), 12 Maneesh Thakur where p = λν −1 g and q ′ ∈ U (B, σu ) is arbitrary with NB (q ′ )NB (p) = 1. Then p ∈ U (B, σ) as discussed above and the automorphism ηe−1 φ of A fixes S = (B, σ)+ pointwise. Hence, by ([21], 39.16), ηe−1 φ(a, b) = (a, bq ′′ ) for some q ′′ ∈ SU (B, σu ). Therefore we get φ(a, b) = ηe(a, bq ′′ ) = (pap−1 , pbq ′′ q ′ ). We set q = q ′′ q ′ ∈ U (B, σu ) and the proof of the first assertion is complete. Now, if q ∈ U (B, σu ) is arbitrary, with NB (q) = ν ∈ K ∗ . Then νν = 1. Hence, by Lemma 3.1, there exists p ∈ U (B, σ) with NB (p) = ν −1 . Then, by the computations done before, it follows that (b, x) 7→ (pbp−1 , pxq) is an automorphism of A stabilizing (B, σ)+ . Hence any automorphism φ of A which stabilizes (B, σ)+ is precisely of the form (b, x) 7→ (pbp−1 , pxq) for some p ∈ U (B, σ), q ∈ U (B, σu ) with NB (p)NB (q) = 1. The map f : [U (B, σ) × U (B, σu )]det → Aut(A, (B, σ)+ ) given by (p, q) 7→ [(b, x) 7→ (pbp−1 , pxq −1 )] is a surjective homomorphism with ker(f ) = {(p, p)\|p ∈ K (1) } ∼ = K (1) , inducing the required isomorphism between the two groups. Corollary 3.2. Let A = J(D, µ) be a first Tits construction Albert algebra. Then any automorphism of A stabilizing D+ ⊂ A is of the form (x, y, z) 7→ (gxg −1 , gyh−1 , hzg−1 ), where g, h ∈ D ∗ and ND (h) = ND (g). Proof. As before, we identify A = J(D, µ) with a second Tits construction J(D×D ◦ , ǫ, 1, (µ, µ−1 )) and apply the proposition. 4 Norm similarities and R-triviality In the last section, we extended automorphisms of a 9-dimensional subalgebra of an Albert algebra to an automorphism of the Albert algebra. In this section, we analyse norm similarities of an Albert algebra A in the same spirit, study rationality and R-triviality properties of some subgroups of Str(A). First we make some simple observations. Since any norm similarity of A fixing the identity element 1 of A is necessarily an automorphism of A, it follows that for any k-subalgebra S ⊂ A the subgroup Str(A/S) of all norm similarities that fix S pointwise, is equal to the subgroup Aut(A/S) of Aut(A), i.e., Str(A/S) =Aut(A/S). Recall that we denote the (normal) subgroup of Str(A) generated by the U -operators of A, by Instr(A). We will now prove some results on norm similarities that will be needed in the paper. We begin with Theorem 4.1. Let A be any Albert algebra. Then Instr(A) ⊂ R(Str(A)(k)). Proof. Let ψ ∈ Instr(A). Then ψ = Ua Ub Uc · · · with a, b, c, · · · ∈ A. Let, for t ∈ A1k , at := (1 − t)a + t ∈ Ak , similarly bt , ct , · · · be defined. Let θ : A1k →Str(A) be defined by θ(t) = Uat Ubt Uct · · · Then θ is a rational map defined over k, with domain the (open) subset of A1k consisting of t such that at , bt , ct , · · · are invertible elements of Ak̄ . Clearly θ is regular and defined at 0 and 1 and θ(t) ∈ Str(A) with θ(0) = ψ, θ(1) = 1. Hence the assertion is proved. R-triviality 13 The theorem below is a generalization of Corollary 4.2 of ([44]) over a field k of arbitrary characteristic : Theorem 4.2. Let A be an Albert division algebra over k and φ ∈ Aut(A) an automorphism of A. Then φ fixes a cubic subfield of A pointwise. Proof. Let Aφ denote the subalgebra of fixed points of φ. In light of the fact that proper subalgebras of an Albert division algebras are of dimension 1, 3 or 9, and that Dimk (Aφ ) is invariant under base change, we may work with Ak . Since the eigenvalues of φ are the same as the eigenvalues of its semisimple part, we may further assume φ is semisimple. Let T be a maximal torus containing φ. By (Theorem 3, [9]), the group of all automorphisms of Ak fixing a primitive idempotent is isomorphic to Spin(9). We can conjugate T to a maximal torus in Spin(9) (over k) to conclude that T fixes a vector w, w ∈ / k. Hence φ fixes some v ∈ A, v ∈ / k. Then v generates a cubic subfield L of A and φ fixes L pointwise. The following theorem is a generalization of Theorem 5.3 of ([44]) to arbitrary characteristics : Theorem 4.3. Let A be a first Tits construction Albert division algebra over k. Let φ ∈ Aut(A) be such that φ fixes a cubic cyclic subfield of A pointwise. Then φ is a product of two automorphisms, each stabilizing a 9-dimensional subalgebra of A. Proof. The proof goes along the exact lines of the proof given in ([44]), except that we use Theorem 4.2 above in place of (Corollary 4.2, [44]). We have the following corollary to the above, with proof exactly as in ([44], Cor. 5.1), we will use this in the paper for k of arbitrary characteristics: Corollary 4.1. Let A be a first Tits construction Albert division algebra over k and φ ∈ Aut(A) be such that φ(L) = L for a cubic cyclic subfield L ⊂ A. Then φ is a product of three automorphisms of A, each stabilizing a 9-dimensional subalgebra of A. The following result is a generalization of Theorem 5.10 of ([44]), for arbitrary characteristics : Theorem 4.4. Let A be an Albert division algebra arising from the first Tits construction. Let L ⊂ A be a cubic cyclic subfield. Then Str(A, L) ⊂ C.Instr(A).H, where H denotes the subgroup of Aut(A) generated by automorphisms stabilizing 9-dimensional subalgebras of A and C is the group of all scalar homotheties in Str(A) . Proof. The proof goes along the exact lines of the proof of Thm 5.10 in ([44]), except that we use Corollary 4.1 above instead of (Cor.5.1, [44]). Extending norm similarities : Let A be an Albert (division) algebra and S a 9-dimensional subalgebra of A. Given an element ψ ∈ Str(S), we wish to extend it to an element of Str(A). Though this can be achieved using a recent result of Garibaldi-Petersson ([15], Prop. 7.2.4), we need certain explicit extension for the purpose of proving R-triviality of the algebraic groups described in the introduction, which we now proceed to describe. We have S ∼ = (B, σ)+ for a suitable degree 3 central division algebra B over a quadratic extension K of k with an involution σ of the second kind. Therefore we have A ∼ = J(B, σ, u, µ) for suitable parameters. For our purpose, by conjugating with a suitable automprphism of 14 Maneesh Thakur A, we may assume S = (B, σ)+ , A = J(B, σ, u, µ) and ψ ∈ Str((B, σ)+ ). Let b0 = ψ(1) and λ = NB (b0 ). By ([17], Chap.V, Thm. 5.12.10), Str((B, σ)+ ) consists of the maps x 7→ λaxσ(a), x ∈ (B, σ)+ , a ∈ B ∗ , λ ∈ k∗ . Hence there exists γ ∈ k∗ and g ∈ B ∗ such that ψ(x) = γgxσ(g), ∀x ∈ S. Let ν := NB (g) ∈ K = Z(B) and let δ = NB (σu (g)−1 g) = NB (g) −1 NB (g) = ν̄ −1 ν, where σu = Int(u) ◦ σ. Then δδ̄ = 1. By Lemma 3.1, there exists q ∈ U (B, σu ) such that δ = NB (q). Define ψe : A → A by setting e ψ((b, x)) = γ(gbσ(g), σ(g)# xq). (∗) e Clearly ψ((b, 0)) = (ψ(b), 0). Hence ψe is an extension of ψ to A and is clearly a k-linear bijective endomorphism of A. We now compute e N (ψ((b, x))) = γ 3 N ((gbσ(g), σ(g)# xq)) 2 = γ 3 [NB (g)NB (g)NB (b) + TK (µNB (g) NB (x)NB (q)) −TB (gbσ(b)σ(g)# xquσ(q)σ(x)g # )]. −1 Now substituting N (q) = NB (g) NB (g), ν = NB (g) = gg # and qu = uσ(q)−1 using qσu (q) = 1 = quσ(q)u−1 in the RHS of the above equation and simplifying, we get e N (ψ((b, x))) = γ 3 ν ν̄N ((b, x)). Therefore ψe ∈ Str(A, S). We have proved Theorem 4.5. Let A be an Albert division algebra over k and S ⊂ A be a 9-dimensional subalgebra. Then every element ψ ∈ Str(S) admits an extension ψe ∈ Str(A). Let S = (B, σ)+ and A = J(B, σ, u, µ) for suitable parameters. Let ψ ∈ Str(S) be given by b 7→ γgbσ(g) for b ∈ (B, σ)+ , g ∈ B ∗ and γ ∈ k∗ . Then for any q ∈ U (B, σu ) with NB (q) = NB (σ(g)−1 g), the map ψe given by e ψ((b, x)) = γ(gbσ(g), σ(g)# xq), is a norm similarity of A extending ψ. Corollary 4.2. Let A = J(D, µ) be a first construction Albert algebra. Let ψ ∈ Str(D+ ) be given by x 7→ γaxb for γ ∈ k∗ and a, b ∈ D ∗ . Then the map ψe : A → A given by (x, y, z) 7→ γ(axb, b# yc, c−1 za# ), where b, c ∈ D ∗ are such that ND (a) = ND (b)ND (c), is a norm similarity of A extending ψ. 15 R-triviality Now, let η ∈ Str(A, S) be arbitrary and let φ := η\|S. We extend φ to an element φe ∈ Str(A) according to (∗). Then we have φe−1 ◦ η ∈ Str(A/S) = Aut(A/S) ∼ = SU (B, σu ). Hence there exists q ′ ∈ SU (B, σu ) such that φe−1 ◦ η((b, x)) = (b, xq ′ ) ∀b ∈ (B, σ)+ , x ∈ B. Write θq′ for the automorphism (b, x) 7→ (b, xq ′ ). Then η = φe ◦ θq′ . Also, since η((B, σ)+ ) = (B, σ)+ , there exists γ ∈ k∗ and g ∈ B ∗ , such that η((b, 0)) = γ(gbσ(g), 0) ∀b ∈ (B, σ)+ . Therefore, we have η((b, x)) = γ(gbσ(g), σ(g)# xq0 ) for all (b, x) ∈ A and some q0 ∈ U (B, σu ) with N (q0 ) = NB (σ(g)−1 g). We have therefore associated a triple (γ, (g, q0 )) ∈ k∗ × B ∗ × U (B, σu ) to η ∈ Str(A, S) with γ ∈ k∗ and g ∈ B ∗ arbtrary and q0 ∈ U (B, σu ) with NB (q0 ) = NB (σ(g)−1 g). Conversely, for any triple (γ, (g, q0 )) ∈ k∗ × B ∗ × U (B, σu ) with γ ∈ k∗ , g ∈ B ∗ arbitrary and q0 ∈ U (B, σu ) such that NB (q0 ) = NB (σ(g)−1 g), we define η : A → A by the formula (∗) : η((b, x)) = γ(gbσ(g), σ(g)# xq0 ). By a straight forward calculation exactly as above, it follows that N (η((b, x))) = γ 3 ν ν̄N ((b, x)). Therefore η ∈ Str(A, S). Hence we have proved Theorem 4.6. Let A = J(B, σ, u, µ) be an Albert division algebra, written as a second Tits construction and S = (B, σ)+ . The group Str(A, S) consists of the maps (b, x) 7→ γ(gbσ(b), σ(g)# xq) where γ ∈ k∗ , g ∈ B ∗ are arbitrary and q ∈ U (B, σu ) satisfies NB (q) = NB (σ(g)−1 g). We have, ∗ ∼ k × H0 , Str(A, (B, σ)+ ) = K∗ where H0 = {(g, q) ∈ B ∗ × U (B, σu )\|NB (q)NB (σ(g)−1 g) = 1} and K ∗ ֒→ k∗ × H0 via α 7→ (NK (α)−1 , α, α−1 α). Proof. Only the last assertion needs to be proved. Let H0 = {(g, q) ∈ B ∗ × U (B, σu )\|NB (q)NB (σ(g)−1 g) = 1}. We define f : k∗ × H0 → Str(A, (B, σ)+ ) by (γ, (g, q)) 7→ [(b, x) 7→ γ.(gbσ(g), σ(g)# xq −1 )]. Then, by our calculations before, f is a surjective homomorphism. We compute ker(f ). Let (γ, (g, q)) ∈ Ker(f ). Then γ(gbσ(g), σ(g)# xq −1 ) = (b, x) ∀b ∈ (B, σ)+ , x ∈ B. (∗∗) 16 Maneesh Thakur Taking b = 1, x = 1 in the above relation gives γgσ(g) = 1, γσ(g)# q −1 = 1. (∗ ∗ ∗) Now gσ(g) = γ −1 gives NB (g)NB (g) = γ −3 and σ(g)# = γ −2 (g# )−1 . Also gg # = NB (g) gives (g# )−1 = NB (g)−1 g. Therefore, we get from the above equation γσ(g)# = γ(γ −2 (g# )−1 ) = γ −1 NB (g)−1 g = q, and thus g = γNB (g)q and σ(g)# = γ −1 q. Substituting this in (∗∗) we get γ(γNB (g)qbσ(q)NB (g)γ, γ −1 qxq −1 ) = (b, x) ∀b ∈ (B, σ)+ , x ∈ B. This gives, using NB (g)NB (g) = γ −3 , (qbq −1 , qxq −1 ) = (b, x) ∀b ∈ (B, σ)+ , x ∈ B. Hence q ∈ K ∗ ∩ U (B, σu ) = K (1) . Thus g = γNB (g)q ∈ K ∗ and, by (∗ ∗ ∗), NK (g) = gσ(g) = γ −1 . Therefore, q = γ −1 NB (g)−1 g = γ −1 g −3 g = γ −1 g −2 = σ(g)gg−2 = σ(g)g −1 . Hence (γ, g, q) = (NK (g)−1 , g, σ(g)g−1 ). This shows ker(f ) = {(NK (α)−1 , α, α−1 α)\|α ∈ K ∗ }, and the proof is complete. Corollary 4.3. Let A = J(D, µ) be a first construction Albert algebra. Then the group Str(A, D+ ) consists of the maps (x, y, z) 7→ γ(axb, b# yc, c−1 za# ), where a, b, c ∈ D ∗ , γ ∈ k∗ and ND (a) = ND (b)ND (c). 5 R-triviality results In this section, we proceed to develop our results on R-triviality of various groups and also prove our main theorem. We need a version of a theorem of Yanchevskii ([21], 17.25) for degree 3 division algebras in arbitrary characteristics, which we prove now: Proposition 5.1. Let (B, σ) be a degree 3 central division algebra over its center K, a quadratic étale extension of k, with σ an involution of the second kind. Then every nonzero element g ∈ B admits a factorization g = zs with s ∈ (B, σ)+ and z ∈ B ∗ is such that zσ(z) = σ(z)z. Proof. We may assume characteristic of k is 2, the proof when characteristic is not 2 is as in ([21], 17.25). Then (B, σ)+ = Sym(B, σ) = Skew(B, σ) and Dimk (Sym(B, σ)) = 9. Since σ restricts to the nontrivial k-automorphism of K, K 6⊂ Sym(B, σ). Hence Dimk (K + Sym(B, σ)) = 10. Also g ∈ B ∗ , hence Dimk (g.Sym(B, σ)) = Dimk (Sym(B, σ)) = 9. 17 R-triviality Hence g.Sym(B, σ) ∩ (K + Sym(B, σ)) 6= 0. Thus we can find s0 ∈ Sym(B, σ) − {0}, λ ∈ K and z0 ∈ Sym(B, σ) such that gs0 = λ + z0 . Let z := λ + z0 . Then z ∈ B ∗ and zσ(z) = (λ + z0 )(σ(λ) + z0 ) = (σ(λ) + z0 )(λ + z0 ) = σ(z)z. Moreover, g.s0 = z implies g = z.(s0 )−1 . Hence we have the required factorization g = zs for s = s−1 0 . We now have Theorem 5.1. Let A be an Albert (division) algebra. Let S ⊂ A be a 9-dimensional subalgebra. Then, with the notations as above, Str(A, S) is R-trivial. Proof. Let ψ ∈ Str(A, S). Since Dim(S) = 9, S ∼ = (B, σ)+ for a suitable degree 3 central simple algebra over a quadratic étale extension K of k, with an involution σ of the second kind. Conjugating by a suitable automorphism of A, we may assume S = (B, σ)+ and A = J(B, σ, u, µ) for suitable parameters. Let H0 = {(g, q) ∈ B ∗ × U (B, σu )\|NB (σ(g)−1 g) = NB (q)}. Clearly H0 is a subgroup of B ∗ ×U (B, σu ). Then, by Theorem 4.6, ψ((b, x)) = γ(gbσ(g), σ(g)# xq) for γ ∈ k∗ and (g, q) ∈ H0 . By Proposition 5.1, we can write g = zs with s, z ∈ B ∗ satisfying −1 σu (z)z = zσu (z), σu (s) = s. Then, since NB (σ(s)−1 s) = NB (s) NB (s) = 1, it follows that (s, 1) ∈ H0 . Since the scalar homotheties in Str(A, S) constitute a 1 dimensional k-split torus, by ([52], §16, Prop. 2, 16.2, Prop. 3), scalar multiplications over k are in R(Str(A, S)(k)). Hence we may assume ψ((b, x)) = (gbσ(g), σ(g)# xq). Now, using g = zs as above (with σu (s) = s, σu (z)z = zσu (z)) and that (s, 1) ∈ H0 , writing θ for the norm similarity (b, x) 7→ (sbσ(s), σ(s)# x), we have ψ = ψ ′ ◦ θ, where ψ ′ ∈ Str(A, S) with ψ ′ ((b, x)) := (zbσ(z), σ(z)# xq), σu (z)z = zσu (z) and (z, q) ∈ H0 . Claim : θ ∈ R(Str(A, S)(k)). To see this, define st := (1 − t)s + t ∈ Bk and let λ : A1k → Str(A, S) be defined by λ(t)((b, x)) = (st bσ(st ), σ(st )# x), (b, x) ∈ Ak = J(Bk , σ, u, µ) = (Bk , σ)+ ⊕ Bk . Then λ is a rational map, defined over k, regular at 0 and 1 and λ(0) = θ, λ(1) = 1 in Str(A, S). This settles the claim. Hence it only remains to show that ψ ′ ∈ R(Str(A, S)(k)). For this, observe that p := σu (z)−1 z ∈ U (B, σu ) since zσu (z) = σu (z)z. We have, NB (p) = NB (z) −1 −1 NB (z) = NB (gs−1 ) NB (gs−1 ) = NB (σ(g)−1 g) = NB (q). Hence q = pq ′ for some q ′ ∈ SU (B, σu ). We therefore have (z, q) = (z, pq ′ ) ∈ H0 for some q ′ ∈ SU (B, σu ). Hence (z, q) = (z, p)(1, q ′ ) and (1, q ′ ) ∈ H0 . Also SU(B, σu ) ⊂ Str(A, S) via η 7→ [(b, x) 7→ (b, xη)], η ∈ SU(B, σu ). By ([6]), SU(B, σu ) is rational, hence R-trivial. Hence the automorphism ψ0′ : (b, x) 7→ (b, xq ′ ) belongs to R(Str(A, S)(k)). Let ψ0 ∈ Str(A, S) be defined by ψ0 (b, x) := (zbσ(z), σ(z)# xp). Then ψ ′ = ψ0′ ◦ ψ0 . Claim : ψ0 ∈ R(Str(A, S)(k)). To see this, define zt = (1 − t)z + t ∈ Bk . Then σu (zt )zt = zt σu (zt ). Let qt = σu (zt )−1 zt whenever zt ∈ Bk∗ . Then qt ∈ U(B, σu ) and the map f : A1k → Str(A, S), f (t)((b, x)) = (zt bσ(zt ), σ(zt )# xqt ) 18 Maneesh Thakur is defined on an open subset of A1k , is a rational map and f (t) ∈ Str(A, S) corresponds to (zt , qt ) ∈ H0 . In particular, f (0) corresponds to the pair (z0 , q0 ) = (z, p), hence f (0) = ψ0 . Now f (1) corresponds to (z1 , q1 ) = (1, q1 ) ∈ H0 . But q1 = σu (z1 )−1 z1 = 1 since z1 = 1. Hence f (1) = 1 ∈ Str(A, S). This settles the claim and the proof of the theorem is complete. Corollary 5.1. Let A be an Albert (division) algebra and S a 9-dimensional subalgebra of A. Then Aut(A, S) ⊂ R(Str(A)(k)). We now can prove Theorem 5.2. Let A be a first Tits construction Albert division algebra and L ⊂ A a cubic cyclic subfield. Then Str(A, L) ⊂ R(Str(A)(k)). Proof. By Theorem 4.4, Str(A, L) ⊂ C.Instr(A).H. By Theorem 4.1 it follows that C.Instr(A) ⊂ R(Str(A)(k)), where C is the subgroup of Str(A) consisting of scalar homotheties and H is the subgroup of Aut(A) generated by all automorphisms stabilizing 9-dimensional subalgebras of A. So it suffices to prove H ⊂ R(Str(A)(k)). Any element of H is a product of automorphisms, each stabilizing some 9-dimensional subalgebra of A. Let h ∈ H and write h = h1 h2 · · · hr with hi ∈ Aut(A, Si ), Si a 9-dimensional subalgebra of A. But by the above theorem Aut(A, Si ) ⊂ Str(A, Si ) ⊂ R(Str(A)(k)). Hence h ∈ R(Str(A)(k)) and the proof is complete. Corollary 5.2. Assume that A is a first Tits construction and L ⊂ A a cyclic cubic subfield. Then Aut(A, L) ⊂ R(Str(A)(k)). Proof. We have Aut(A, L) ⊂ Str(A, L), hence the result follows from Theorem 5.2. Theorem 5.3. Let A be an Albert division algebra, arising from the first Tits construction. Then Aut(A) ⊂ R(Str(A)(k)). Proof. Fix a cubic cyclic extension E ⊂ A, which is possible by ([29]). Let φ ∈ Aut(A). If φ(E) = E, then φ ∈ Aut(A, E) and by Theorem 5.2, φ ∈ R(Str(A)(k)). So we may assume F := φ(E) 6= E. Let S be the subalgebra of A generated by E and F . Then (cf. [31], Thm. 1.1), S is 9-dimensional. Hence S = (B, σ)+ for a degree 3 central simple algebra B over a quadratic étale extension K of k, with an involution of the second kind over K. Hence we may write A = J(B, σ, u, µ) for suitable parameters. If Z(B) = K = k ×k, then we have (B, σ)+ = D+ for a degree 3 central division algebra D over k, and A = J(D, µ0 ) for some µ0 ∈ k∗ . Then, by the classical Skolem-Noether theorem, φ\|E : E → F extends to an automorphism of D+ and hence, by Corollary 3.1, to an automorphism θ of A, θ(D+ ) = D+ and θ\|E = φ\|E. Hence ψ := φ−1 ◦ θ fixes E pointwise. By Corollary 5.2, ψ ∈ Aut(A, E) ⊂ R(Str(A)(k)). Also, θ ∈ Aut(A, D+ ). Hence by Corollary 5.1, θ ∈ Aut(A, D+ ) ⊂ R(Str(A)(k)). It follows that φ ∈ R(Str(A)(k)). So now we assume K is a field. First extend φ\|E to φ′ : EK → F K ⊂ B to a K-linear isomorphism. By Skolem-Noether theorem, there exists g ∈ B ∗ such that φ′ = Int(g) : EK → F K. Since φ(E) = F , it follows that Int(g)(E) = F = φ(E) ⊂ (B, σ)+ . Therefore σ(gxg −1 ) = gxg −1 ∀x ∈ E. This yields σ(g)−1 xσ(g) = gxg −1 ∀x ∈ E. 19 R-triviality Hence σ(g)g centralizes E and hence EK in B. Since EK is maximal commutative in B, it follows that σ(g)g ∈ EK. Also σ(g)g ∈ (B, σ)+ . Hence σ(g)g ∈ E. Similarly, it follows that gσ(g) ∈ F . Let e := σ(g)g ∈ E and f := gσ(g) ∈ F . Define ψ ∈ Str(S) by ψ(x) = gxσ(g), x ∈ S. Then for any x ∈ E we have xe ∈ E. Hence, ψ(x) = gxσ(g) = gxeg −1 = φ(xe) ∈ φ(E) = F. Therefore ψ(E) = F and ψ(1) = f ∈ F . Since A is a first Tits construction and F is cyclic, there is a degree 3 central division algebra D over k, such that F ⊂ D+ ⊂ A (see [28], Cor. 4.5). By ([44], Lemma 5.1), there exists χ ∈ C.Instr(A) such that χ(f ) = 1, χ(D+ ) = D+ and χ(F ) = F . Hence χ ◦ ψ(1) = 1. We extend ψ to an element ψe ∈ Str(A, S) by the formula in Theorem 4.5. Then χ ◦ ψe ∈ e e Str(A) and χ ◦ ψ(1) = 1. Hence χ ◦ ψe ∈ Aut(A) with χ ◦ ψ(E) = F . Therefore φ−1 ◦ χ ◦ ψe ∈ −1 −1 e Aut(A) and φ ◦ χ ◦ ψ(E) = E. Hence, by Theorem 5.2, φ ◦ χ ◦ ψe ∈ R(Str(A)(k)). Also, χ ∈ Str(A, D+ ), and ψe ∈ Str(A, S). We have already shown in Theorem 5.1 that these are subgroups of R(Str(A)(k)). It now follows that φ ∈ R(Str(A)(k)). Lemma 5.1. Let L be a cubic extension of k and λ ∈ k∗. Let S denote the first Tits process J(L, λ) over k. Let a ∈ L∗. Define χ ∈ Str(S) by χ = RNL (a) U(0,0,1) U(0,NL (a)−1 ,0) . Then χ(a) = 1. Proof. The proof follows from a calculation exactly same as in the proof of (Lemma 5.1, [44]). We are now in a position to prove our main result: Theorem 5.4. Let A be an Albert division algebra arising from the first Tits construction. Then Str(A) is R-trivial. Proof. Let ψ ∈ Str(A) and a := ψ(1). If a ∈ k ∗ , we have Ra ∈ R(Str(A)(k)) and Ra−1 ◦ ψ(1) = 1. Hence, by Theorem 5.3, Ra−1 ◦ ψ ∈ Aut(A) ⊂ R(Str(A)(k)). This implies ψ ∈ R(Str(A)(k)). So we may assume a ∈ / k∗ . Then L := k(a) ⊂ A is a cubic separable subfield and, by ([28], Cor. 4.5), there is λ ∈ k∗ such that the Tits process S = J(L, λ) ⊂ A. Define χ ∈ Str(S) by χ = RN (a) U(0,0,1) U(0,N (a)−1 a,0) . By the above lemma χ(a) = 1. We now extend χ to χ e ∈ Str(A) using (∗). Then χ e(S) = S and hence by Theorem 5.1, χ e ∈ R(Str(A)(k)). Also, χ e ◦ ψ(1) = χ e(a) = 1. Hence χ e ◦ ψ ∈ Aut(A) ⊂ R(Str(A)(k)) and it follows that ψ ∈ R(Str(A)(k)). Theorem 5.5. Let A be an Albert division algebra arising from the first Tits construction over a field k. Then Isom(A) is R-trivial. 20 Maneesh Thakur Proof. Let ψ ∈ Isom(A) = Isom(A)(k) ⊂ Str(A) = Isom(A)(k). By Theorem 5.4, Str(A) is R-trivial. Let θ : A1k → Str(A) be a rational map connecting ψ to 1 in Str(A), θ(0) = ψ, θ(1) = 1. Let ψt := θ(t), at := ψt (1), λt := N (at ). Let χt := Rλ−1 Ua−1 Uat ψt . Then 0 t χt (1) = −1 λ−1 t Ua0 Uat ψt (1) = 3 λ−1 t Ua0 (at ). Since N (a0 ) = N (ψ0 (1)) = N (ψ(1)) = 1, we have 2 3 N (χt (1)) = λ−3 t N (a0 ) N (at ) = 1, It follows that χt ∈ Isom(A). Also, t 7→ χt is a rational map A1k → Isom(A). Now λ0 = N (ψ0 (1)) = N (ψ(1)) = 1 as ψ is a norm isometry. Hence χ0 = Ua−1 Uψ(1) ψ0 = ψ, 0 and χ1 = 1. This proves the assertion. 6 78 Tits-Weiss conjecture and E8,2 In this section, we will prove the Tits-Weiss conjecture for Albert division algebras that arise from the first Tits construction. Recall that in this paper, we denote by Str(A) the group Str(A)(k) of k-rational points of the algebraic group Str(A), the full group of norm similarities of A. We denote by Isom(A) the full group of norm isometries of A and Isom(A) := Isom(A)(k). By ([39], Thm. 7.3.2) Isom(A) is a connected, simple, simply connected algebraic group of type E6 defined over k. It follows that Isom(A) is the commutator subgroup of Str(A). Let C denote the subgroup of Str(A) consisting of all scalar homotheties Ra , a ∈ k∗ . Tits-Weiss conjecture : Let A be as above, then Str(A) = C.Instr(A). Str(A) = {1}. This is equivalent to the KneserEquivalently, the conjecture predicts that C.Instr(A) 78 Tits conjecture for groups of type E8,2 . 78 and G(k)† Let G be a simple, simply connected group defined over k, with Tits index E8,2 be the subgroup of G(k) generated by the k-rational points of the unipotent radicals of G(k) parabolic k-subgroups of G. The Kneser-Tits conjecture for G predicts that G(k) † = {1}. The (reductive) anisotropic kernel of G is the structure group Str(A) of a uniquely (upto isotopy) determined Albert division algebra A defined over k (see [46], 3.3.1). It is proved G(k) ∼ Str(A) in ([51], 37.41, 42.3.6) that G(k) † = C.Instr(A) . Hence, to prove the Tits-Weiss conjecture, we need to show the first quotient is trivial. By (Thm. 7.2, [13]), it suffices to prove G is R-trivial. In fact, it suffices to prove W (k, G) = G(k)/G(k)† ∼ = G(k)/R = {1}. Let S be a maximal k-split torus in G. By the Bruhat-Tits decomposition for G, it follows that there is a birational k-isomorphism G ∼ = U × ZG (S) × U, where U is the unipotent radical of a minimal parabolic k-subgroup containing S. It is well known that the underlying variety of U is rational over k. Hence, by ([52], Prop. 1, §16), it follows that G(k) ∼ ZG (S)(k) . = R R Again, by ([14], 1.2), ZG (S)(k) ∼ ZG (S) (k)/R. = R S 21 R-triviality Hence, to prove the Tits-Weiss conjecture, it suffices to prove that the second quotient above is trivial. We now proceed to prove this. Note that H := ZG (S) is a connected reductive k-subgroup of G, is the reductive anisotropic kernel of G and its commutator subgroup H′ is simple k-anisotropic, strongly inner of type E6 . Hence, by ([46], 3.3), H′ = Isom(A). We first prove Lemma 6.1. The connected center of H equals S. Proof. We know that H is connected reductive and its connected center contains S. We have H = H′ .Z(H)◦ , an almost direct product of groups. Since H′ is simple of type E6 , Rank(H′ ) = 6. Let T be any maximal k-torus in H′ . The connected center Z(H)◦ of H is a torus S′ ⊃ S and S′ ∩ T is finite. Hence T.S′ is a torus in H of dimension 6 + rank(S′ ) and rank(S′ ) ≥ 2. Therefore Rank(G) = 8 ≥ 6 + rank(S′ ). It follows that S′ = S. In light of this lemma, we have ZG (S) H′ S H′ (k)/R = (k)/R ∼ (k)/R. = ′ S S H ∩S It is well known (see Thm. 7.3.2, [39]. Thm. 14.27, [41]) that for an Albert algebra A over k, Str(A) is a connected reductive group defined over k. We need information on its center: Lemma 6.2. Let A be an Albert algebra over k. Then the group Str(A) is connected reductive with center Gm . The commutator [Str(A), Str(A)] = Isom(A) and the sequence {1} → Isom(A) → Str(A) → Gm → {1} is exact, where Str(A) → Gm maps a norm similarity to its factor of similitude. Proof. That Str(A) is connected reductive is proved in (3.3.1, [46], [39]). Let ψ ∈ Z(Str(A)). We choose two 9-dimensional degree 3 separable k-subalgebras of A, say S1 and S2 such that S1 ∩ S2 = k (see [35], Lemma 5.2). Let θi ∈ Aut(A/Si ) be such that Aθi = Si . Then since ψ commutes with θi , i = 1, 2, ψ must stabilize both S1 as well as S2 . Hence ψ stabilizes S1 ∩ S2 = k. Thus a := ψ(1) ∈ k∗ and φ = Ra−1 ◦ ψ ∈ Aut(A). Hence φ ∈ Z(Aut(A)). But Aut(A) is a group of type F4 , hence has trivial center. It follows that ψ = Ra . Hence Z(Str(A)) = Gm , the k-subgroup of Str(A) consisting of all scalar homotheties Ra , a ∈ Gm . The remaining assertions follow from ([39], Thm. 7.3.2) and ([41]. Thm. 14.27). Now we observe that maximal k-split tori in G are conjugate by G(k) and G ⊃ Str(A) over k (see [51], 42.6). Hence we may assume that S ⊃ Z(Str(A)). Denote by Z the center of Str(A). We have Str(A) = [Str(A), Str(A)].Z = Isom(A).Z. Since both S and Z are split over k, by ([1], 8.5. Corollary) we have a decomposition S = Z.S′ of k-tori, with Z ∩ S′ = {1}. Hence, since H′ = [Str(A), Str(A)] = Isom(A), we have H′ S = H′ Z.S′ = Isom(A)Z.S′ = Str(A).S′ . Also, since Z ∩ S′ = {1} and S′ ⊂ S = Z(H)◦ , it follows that Str(A) ∩ S′ = {1}. We therefore have ZG (S) H′ S ∼ Str(A) × S′ ∼ Str(A) = . = = S S Z × S′ Z 22 Maneesh Thakur Finally, by Hilbert Theorem-90, H 1 (k, Z) = {1}. Hence, it follows that Str(A) Str(A)(k) (k) ∼ . = Z Z(k) We can now state and prove Theorem 6.1. Let G be a simple, simply connected algebraic group over k with Tits index 78 , whose anisotropic kernel is split by a cubic extension of k. Then G is R-trivial. E8,2 Proof. Let H denote the reductive anisotropic kernel of G as in the above discussion. Then, by Corollary 2.1, [H, H] = Isom(A) for an Albert division algebra A over k, arising from a first Tits construction. We have already shown G(k)/R ∼ = Str(A) Str(A)(k) (k)/R ∼ /R. = Z Z(k) Also, by Theorem 5.4, when A is a first Tits construction, Str(A) is R-trivial. Let L be any field extension of k. If A ⊗k L is split, then G splits over L and hence W (L, G) = {1}. If A⊗k L is division, then the Tits index of G over L does not change and the anisotropic kernel of G over L corresponds to Str(A ⊗k L). Hence W (L, G) = G(L)/R = Str(A)(L) /R = {1}, Z(L) the last equality holds because Str(A) is R-trivial : since Str(A) Z (L) = Str(A)(L) Z(L) , any element Str(A) Z (L) of is represented by an element of Str(A)(L) and any element of Str(A)(L) can be joined to the identity element by an L-rational image of A1L , as Str(A) is R-trivial. The result now follows. Remark : Note that we have proved above also that Str(A) Z , the adjoint form of E6 corresponding to a first Tits construction Albert algebra A over a field k is R-trivial. Corollary 6.1. Tits-Weiss Conjecture : Let A be a first Tits construction Albert division algebra. Then Str(A) = C.Instr(A). Proof. This follows immediately from the above theorem and that G is as in the above theorem. 7 Str(A) C.Instr(A) ∼ = G(k) R , where 78 Groups with index E7,1 78 . The In this brief section, we wish to explore R-triviality issues for simple groups of type E7,1 reason for considering such groups is that these groups also have their anisotropic kernels as the structure groups of Albert division algebras and all groups of this type arise this way (see [46], 3.3.1). We will consider the groups Ξ with this index whose anisotropic kernel is the structure group of an Albert division algebra, that is a Tits first construction. These groups were studied by Max Koecher in ([20]). We use the explicit description of Ξ given by Max Koecher. Fix an Albert division algebra A over k. Let Ξ(A) be the full group of birational transformations generated by j : Ak → Ak , j(x) := (−x)−1 and the translations ta : x 7→ a+x, a, x ∈ Ak . Then every element of Ξ(A) has an expression f = w ◦ ta ◦ j ◦ tb ◦ j ◦ tc , a, b, c ∈ Ak . 23 R-triviality In particular Str(A) is contained in Ξ(A) and Isom(A) is the anisotropic kernel (up to a central torus) of Ξ(A). More precisely, let Z denote, as before, the center of Str(A), which is a Gm defined over k. Let S ⊂ Ξ(A) be a maximal k-split torus containing Z. Then H = ZΞ (S) is a connected reductive k-subgroup of G, is the reductive anisotropic kernel of Ξ = Ξ(A). Since k-rank of Ξ is 1, we have S = Z. It follows from this that Z(H)◦ = Z. We have, as in the previous section ZΞ (S) Ξ(k)/R ∼ (k)/R. = ZΞ (S)(k)/R ∼ = S Now, H = ZΞ (S) = H′ .Z(H)◦ = Isom(A).Z = Str(A). Hence ZΞ (S) Str(A) = . S Z We have shown in the last section that Str(A) Z (k)/R = {1}. Hence we have, by arguments 78 , quite analougous to the proof of R-triviality of groups with index E8,2 Theorem 7.1. Let Ξ be a simple, simply connected algebraic group over k with Tits index 78 , whose semisimple anistropic kernel is split over a cubic extension of k. Then Ξ is E7,1 R-trivial. Reduced Albert algebras : In this mini-section, we wish to explore the group Str(A) when A is a reduced Albert algebra, mainly the R-triviality of such groups. We recall at this stage our result from ([44]), which. in fact, holds over fields of arbitrary characteristics : Theorem 7.2. Let A be a reduced Albert algebra over any field k. Then Str(A) = C.Instr(A), where C denotes the subgroup of Str(A) of scalar homotheties of A and Instr(A) is the inner structure group. Proof. This follows from a result of J. R. Faulkner ([9], Thm. 1, [10]), that for any reduced Albert algebra A, the group Isom(A) is simple modulo its center. Given ψ ∈ Str(A), let a := ψ(1) and α := N (a). Then α 6= 0 and Ua is invertible. We now consider χ := Rα−1 Ua ψ ∈ Str(A). Then χ ∈ Isom(A). Since Instr(A) is a normal subgroup of Str(A) and norm similarities of the form Ua1 Ua2 ...Uar ∈ Isom(A) whenever ΠN (ai ) = 1, it follows that (Instr(A) ∩ Isom(A)).Z 6= {1} and hence (Instr(A) ∩ Isom(A)).Z = Isom(A), where Z is the center of Isom(A) and consists of scalar homotheties Rc with c ∈ k∗ , c3 = 1. Therefore χ ∈ C.Instr(A) and hence ψ ∈ C.Instr(A). This completes the proof. We have already shown in §4 that Instr(A) ⊂ R(Str(A)(k)). Hence it follows from the above theorem that Str(A) ⊂ R(Str(A)(k)). This inclusion is independent of the base field, hence Str(A) is R-trivial. We record this as Theorem 7.3. Let A be a reduced Albert algebra over a field k. Then Str(A) is R-trivial. Theorem 7.4. Let A be an Albert algebra arising from the first Tits construction. Then any norm similarity of A is a product of norm similarities, each stabilizing a 9-dimensional subalgebra of A. 24 Maneesh Thakur Proof. First assume A is split. Let ψ ∈ Str(A) and ψ(1) = a. Then, χ := RN (a)−1 ◦ Ua ◦ ψ ∈ Isom(A) and RN (a)−1 ◦ Ua ∈ Str(A) stabilizes any subalgebra of A containing a. By a result of Faulkner cited above, Isom(A) is a simple group modulo its center, which contains scalar homotheties. Hence Isom(A) is generated by norm isometries which stabilize 9-dimensional subalgebras and the result follows. Now assume A is a division algebra. As above, we may assume ψ ∈ Isom(A). As in the proof of Theorem 5.4, we reduce to the case when ψ ∈ Aut(A). Again, using the arguments in the proof of Theorem 5.3, we may assume that ψ stabilizes a cubic cyclic extension E ⊂ A, i.e. ψ ∈ Str(A, E). By Theorem 4.4, Str(A, E) ⊂ C.Instr(A).H, where C is the subgroup of scalar homoteties on A, Instr(A) is the subgroup of Str(A) generated by the U -operators of A and H is the subgroup of Aut(A) generated by automorphisms that stabilize 9-dimensional subalgebras of A. Hence ψ admits the required factorization. Theorem 7.5. Let A be a reduced Albert algebra over a field k. Then Isom(A) is R-trivial. Proof. Let ψ ∈ Isom(A) = Isom(A)(k) ⊂ Str(A). By Theorem 7.3 Str(A) is R-trivial. Now, by an argument exactly as in the proof of Theorem 5.5, the result follows. 8 Retract rationality Recall that a an irreducible variety X defined over a field k is retract k-rational if there exists a non-empty open subset U ⊂ X, defined over k such that the identity map of U factors through an open subset V of an affine space Am k , i.e. there exists morphisms f : U → V and r : V → U such that r ◦ f = idU (see 2.2, [13]). The following result is proved in ([13], Prop. 5.4, Thm. 5.9): Theorem 8.1. Let G be a semisimple, simply connected, absolutely almost simple group, defined and isotropic over an infinite field k, then the following are equivalent 1. G is W -trivial, i.e. W (F, G) = {1} for all field extensions F/k. 2. G is a retract k-rational variety. Combining this with our results, we have, Corollary 8.1. Let G be a simple, simply connected algebraic group defined over a field k 78 or E 78 whose anisotropic kernel splits over a cubic extension of k. Then with Tits index E8,2 7,1 G is retract k-rational. Proof. By Corollary 2.1, the anisotropic kernel of G is isomorphic to Isom(A) for a first Tits construction Albert division algebra A over k. Since if A is a first Tits construction, for any field extension F of k, A ⊗k F is either split or a division algebra (see [28], Cor. 4.2). Let AF := A ⊗k F . By Theorem 7.2, Str(AF ) = C.Instr(AF ) when AF is split and by our result on Tits-Weiss conjecture, Str(AF ) = C.Instr(AF ) also when AF is a division algebra. Hence, by Theorem 6.1 and Theorem 7.1 we have W (F, G) = {1}. This proves the assertion on G. Theorem 8.2. Let A be a reduced Albert algebra over a field k. Then Isom(A) is retract rational. R-triviality 25 Proof. Note that when A is a reduced Albert algebra, the algebraic group Isom(A) is simple, simply connected and k-isotropic. The assertion now follows from the R-triviality of this group, settled in Theorem 7.5. Acknowledgement Part of this work was done when I visited Linus Kramer of the Mathematics Institute, University of Muenster, in March 2016. The stay was supported by the Deutsche Forschungsgemeinschaft through SFB 878. His help is gratefully acknowledged. I thank Richard Weiss and Tom De Medts for some very useful discussions and suggestions. I have immensely benefitted from discussions with Bernhard Muhlherr, Gopal Prasad and Dipendra Prasad. I thank Holger Petersson and Otmar Loos for their encouragement and interest in this work. References [1] A. Borel, Linear Algebraic Groups, GTM 126, Springer Verlag, New York, 1991. [2] Lien Boelaert, Tom De Medts, Anastasia Stavrova, Moufang sets and structurable division algebras, Preprint, available on the arXiv: http://arxiv.org/abs/1603.00780. [3] Chetan Balwe and Anand Sawant, R-equivalence and A1 -connectedness in anisotropic groups, to appear in IMRN. [4] V. Chernousov, A. Merkurjev, R-equivalence and Special Unitary Group, J. Algebra, 209 (1998), 175-198. [5] V. Chernousov, A. Merkurjev, R-equivalence in spinor groups, J. Amer. Math. Soc., 14 (2001), no. 3, 509-534. [6] Vladimir I. Chernousov and Vladimir P. Platonov, The rationality problem for semisimple group varieties, J. Reine Angew. Math. 504 (1998), 1-28. [7] J.-L. Colliot-Thélene, J.-J. Sansuc, La R-équivalence sur les tores, Ann. Scient. Ec. Norm. Sup., 4e serie, 10 (1977), 175-230. [8] V. Chernousov, L. Timoshenko, On the group of R-equivalence classes of semisimple groups over arithmetic fields, (Russian) Algebra i Analiz 11 (1999), no. 6, 191-221; translation in St. Petersburg Math. J. 11 (2000), no. 6, 1097-1121. [9] John R. Faulkner, Octonion planes in characteristic 2, Bull. AMS. 75 (1969), 980-983. [10] John R. Faulkner, Octonion planes defined by quadratic Jordan algebras, Mem. AMS., 104, 1970. [11] Joseph C. Ferrar and Holger P. Petersson, Exceptional simple Jordan algebras and Galois cohomology, Arch. Math., Vol. 61 (1993), 517-520. [12] Skip Garibaldi, Kneser-Tits for a rank 1 form of E6 (After Veldkamp), Compositio Math. 143 (2007), 191-200. [13] Philippe Gille, Le Probléme de Kneser-Tits, Séminaire Bourbaki 982 (2007), Astérisque 326 (2009), 39-81. 26 Maneesh Thakur [14] Philippe Gille, Spécialisation de la R-équivalence pour les groupes réductifs, Trans. AMS. 356 (2004), 4465-4474. [15] Skip Garibaldi and Holger P. Petersson, Outer automorphisms of algebraic groups and a Skolem-Noether theorem for Albert algebras, to appear in Documenta Mathematica. [16] N. Jacobson, Structure and representations of Jordan algebras, AMS. Providence, R. I., 1968, AMS. Colloquium Publications, Vol. XXXIX. [17] N. Jacobson, Finite-dimensional division algebras over fields, Springer Verlag, Berlin, 1996. [18] N. Jacobson, Some groups of transformations defined by Jordan algebras-I, J. reine angew. Math. 201 (1959), No. 3/4, 178-195. [19] N. Jacobson, Some groups of transformations defined by Jordan Algebras. III. Groups of type E6 , J. reine angew. Math. 207 (1961), 61-85. [20] Max Koecher, Uber eine Gruppe von rationalen Abbildungen, Invent. Math. 3, (1967) 136-171. [21] M. A. Knus, A. Merkurjev, M. Rost, J. P. Tignol, The Book of Involutions, AMS. Colloquium Publications, Vol. 44, 1998. [22] Ottmar Loos, On algebraic groups defined by Jordan pairs, Nagoya Math. J., Vol. 74 (1979), 23-66. [23] Yu. I. Manin, Cubic forms, North-Holland, Amsterdam, 1974. [24] K. McCrimmon, The Freudenthal-Springer-Tits constructions of exceptional Jordan algebras, Trans. AMS. 139 (1969), 495-510. [25] A. S. Merkurjev, The norm principle for algebraic groups, Algebra i Analiz 7 (1996), no. 2, 243-264. [26] H. P. Petersson, Structure theorems for Jordan algebras of degree three over fields of arbitrary characteristic, Comm. Algebra 32 (2004), no. 3, 1019–1049. [27] H. P. Petersson, Albert algebras, mini-course, Fields Institute workshop on exeptional algebras and groups, University of Ottawa, 2012. (available on the URL http://www.fernuni-hagen.de/petersson/vortraege). [28] H. P. Petersson, M. Racine, Springer forms and the first Tits construction of exceptional Jordan division algebras, Manuscripta Math. 45 (1984), no. 3, 249-272. [29] H. P. Petersson, M. Racine, An elementary approach to the Serre-Rost invariant of Albert algebras, Indag. Mathem., N.S. 7 (1996), no. 3, 343-365. [30] H. P. Petersson, M. Racine, Albert algebras, Jordan algebras (Oberwolfach, 1992)(W. Kaup and McCrimmon, eds), de Gruyter, Berlin, 1994, no. 3, 197-207. [31] H. P. Petersson, M. Racine, Classification of algebras arising from the Tits process, J. Algebra 98 (1986), no. 1, 244–279. R-triviality 27 [32] H. P. Petersson, M. Racine, Jordan algebras of degree 3 and the Tits process, J. Algebra 98 (1986), no.1, 211-243. [33] V. P. Platonov and A. Rapinchuk, Algebraic groups and number theory, Pure and Applied Mathematics, Vol. 139, Academic Press Inc., Boston, MA, 1994. [34] R. Parimala, R. Sridharan, M. L. Thakur, A classification theorem for Albert algebras, Trans. AMS. 350 (1998), no. 3, 1277-1284. [35] R. Parimala, R. Sridharan, Maneesh L. Thakur, Tits’ constructions of Jordan algebras and F4 bundles on the plane, Compositio Math. 119 (1999), 13-40. [36] R. Parimala, J. P. Tignol, Richard M. Weiss, The Kneser-Tits conjecture for groups with 66 over an arbitrary field, Transform. Groups. 17 (2012), no. 1, 209-231. Tits-index E8,2 [37] Gopal Prasad, On the Kneser-Tits problem for triality forms, Comment. Math. Helv. 83 (2008), no. 4, 913-925. [38] Gopal Prasad, M. S. Raghunathan, On the Kneser-Tits problem, Comment. Math. Helv. 60 (1985), no. 1, 107-121. [39] T. A. Springer and F. D. Veldkamp, Octonions, Jordan Algebras and Exceptional Groups, Springer Monographs in Mathematics, Springer-Verlag, Berlin, (2000). [40] T. A. Springer, Linear Algebraic Groups, Second Edition, Progress in Mathematics, Birkhauser, Boston, 1998. [41] T. A. Springer, Jordan algebras and algebraic groups, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 75, Springer-Verlag, 1973. [42] Tom De Medts and Hendrik Van Maldeghem, Moufang Generalized Polygons, Topics in Diagram Geometry, Quaderni di Matematica, Volume 12 (edited by Antonio Pasini), 59-126, Dept. Math., Seconda Univ. Napoli, Caserta, 2003. [43] Tom De Medts and Yoav Segev, A course on Moufang sets, Innovations in Incidence Geometry, Vol. 9 (2009), pp. 79-122. [44] Maneesh Thakur. Automorphisms of Albert algebras and a conjecture of Tits and Weiss, Trans. AMS. 365 (2013), no. 6., 3041-3068. [45] J. Tits, Groupes de Whitehead de groupes algébriques simples sur un corps, in Séminaire Bourbaki, Vol. 1976/77, Exp. 505, Lecture Notes in Math., No. 677, Springer, BerlinHeidelberg-New York, 1978, pp. 218-236. [46] J. Tits, Strongly inner anisotropic forms of simple algebraic groups, Journal of Alg. 131 (1990), 648-677. [47] J. Tits, Classification of buildings of spherical type and Moufang polygons: A Survey, Colloquio Internazionale sulle Teorie Combinatorie (Roma, 1973), Tomo I, pp. 229–246. Atti dei Convegni Lincei, No. 17, Accad. Naz. Lincei, Rome, 1976. [48] J. Tits, Classification of algebraic semisimple groups, Algebraic Groups and Discontinuous Subgroups (Proc. Sympos. Pure Math., Boulder, Colo., 1965) (A. Borel and G. D. Mostow, eds.), Vol. 9, AMS., Providence, R.I., 1966, 33-62. 28 Maneesh Thakur [49] F. D. Veldkamp, Unitary groups in projective octave planes, Compositio Math. 19 (1968), 213-258. [50] F. D. Veldkamp, Unitary groups in Hjelmslev-Moufang planes, Math. Z. 108 (1969), 288-312. [51] J. Tits, R. M. Weiss, Moufang Polygons, Springer Monographs in Mathematics, Springer Verlag, 2002. [52] V. E. Voskresenskii, Algebraic Groups and Their Birational Invariants, Translation of Math. Monographs, Vol 179, AMS, Providence, RI, 1998. [53] R. M. Weiss, The Structure of Spherical Buildings, Princeton University Press, 2003, Princeton. [54] R. M. Weiss, Quadrangular Algebras, Princeton University Press, Mathematical Notes 46, Princeton University Press, 2006, Princeton.	4math.GR
arXiv:1708.02740v1 [cs.LG] 9 Aug 2017 A Data Prism: Semi-Verified Learning in the Small-α Regime Michela Meister mmeister@stanford.edu Gregory Valiant valiant@stanford.edu August 10, 2017 Abstract We consider a simple model of unreliable or crowdsourced data where there is an underlying set of n binary variables, each “evaluator” contributes a (possibly unreliable or adversarial) estimate of the values of some subset of r of the variables, and the learner is given the true value of a constant number of variables. We show that, provided an α-fraction of the evaluators are “good” (either correct, or with independent noise rate p < 1/2), then the true values of a (1 − ǫ) fraction of the n underlying variables can be deduced as long as α > 1/(2 − 2p)r . For example, if each “good” worker evaluates a random set of 10 items and there is no noise in their responses, then accurate recovery is possible provided the fraction of good evaluators is larger than 1/1024. This result is optimal in that if α ≤ 1/(2 − 2p)r , the large dataset can contain no information. This setting can be viewed as an instance of the semi-verified learning model introduced in [3], which explores the tradeoff between the number of items evaluated by each worker and the fraction of “good” evaluators. Our results require the number of evaluators to be extremely large, > nr , although our algorithm runs in linear time, Or,ǫ (n), given query access to the large dataset of evaluations. This setting and results can also be viewed as examining a general class of semi-adversarial CSPs with a planted assignment. This extreme parameter regime, where the fraction of reliable data is small (inverse exponential in the amount of data provided by each source), is relevant to a number of practical settings. For example, settings where one has a large dataset of customer preferences, with each customer specifying preferences for a small (constant) number of items, and the goal is to ascertain the preferences of a specific demographic of interest. Our results show that this large dataset (which lacks demographic information) can be leveraged together with the preferences of the demographic of interest for a constant number of randomly selected items, to recover an accurate estimate of the entire set of preferences, even if the fraction of the original dataset contributed by the demographic of interest is inverse exponential in the number of preferences supplied by each customer. In this sense, our results can be viewed as a “data prism” allowing one to extract the behavior of specific cohorts from a large, mixed, dataset. 1 Introduction Imagine that you have access to a large dataset of market research. Specifically, the dataset consists of customer evaluations of products. While the total set of products is large, of size n, each customer is only asked to evaluate a small (perhaps randomly selected) subset of r = 2, 3, etc. of those products. Long after the dataset is collected, suppose you wish to identify the preferences of some special demographic of customers—perhaps the customers who are full-time students. Let α denote a lower bound on the fraction of the surveyed customers that were full-time students, but assume that we do not have this demographic information in our dataset–all we have is the set of evaluations of each customer. How can we leverage this dataset to learn anything about the student-demographic? 1 If α ≪ 1/2, this problem seems hopeless because the amount of data contributed by non-students might swamp the portion of the dataset contributed by the demographic of interest. Nevertheless, the main result of this paper shows that one could hire some students to evaluate a constant, k, number of (random) products in the set of size n, and then leverage that constant amount of information together with the large dataset to return accurate evaluations of the student-demographic preferences on all n items. This claim will hold provided the number of items evaluated by each of the customers in the dataset, r > log2 (1/α). The guarantees of the algorithm will ensure that, with high probability, at most an ǫ-fraction of the returned evaluations are incorrect (where k—the number of products evaluated by the hired students, is a function of ǫ that is independent of the total number of items, n). In particular, this strong success guarantee holds irrespective of the behavior of the non-student demographics in the original dataset–in particular, they could even be adversarial, provided by a single malicious entity who is trying to disguise the feedback provided by the student-demographic. The above setting, where one has a large dataset reflecting a number of demographics, and wishes to leverage the large dataset in conjunction with a very small set of “verified” datapoints from one demographic of interest, seems widely applicable beyond the market research domain. Indeed, there are many biological or health-related datasets where the “demographic of interest” might be a trait that is expensive to evaluate. For example, perhaps one has a large database of medical records, and wishes to investigate the propensity of certain medical conditions for the subset of people with a specific genetic mutation. The large dataset of medical records will likely not contain information about whether individuals have the mutation in question. Nevertheless, our results imply that accurate inferences about this subset of people can likely be made as long as 1) the fraction of people with the mutation in the large dataset is not minuscule, and 2) one can obtain a small (i.e. constant) amount of data from individuals with the genetic mutation in question, for example studying a constant number of individuals who are known to have the mutation. 1.1 Formal Model We formally model this problem as an instance of the semi-verified learning model proposed by Charikar, Steinhardt, and Valiant [3]. Suppose there is a set of n Boolean variables, V = {v1 , . . . , vn }, and m “workers” who each provide an evaluation of the values of a randomly selected subset of r of the variables. Suppose that an α-fraction of the workers are “reliable” and submit evaluations with the property that each of their r reported values is incorrect independently with probability ≤ prel . We make no assumptions on the evaluations submitted by the (1 − α)m unreliable workers—these evaluations could be biased, arbitrary, or even adversarially chosen with the goal of confounding the learning algorithm. In addition to this large dataset, we also receive k ≪ n, m “verified” data points that consist of the values of a random subset of the variables of size k. The goal of the learner will be to return assignments to the n variables, such that with probability at least 1 − δ, at most ǫn of these returned assignments differ from their true values. Previous work [12, 3] focussed on the regime where the number of workers, m = Θ(n). In contrast, we will allow m ≫ n, and focus on the interplay between the number of variables evaluated by each individual, r, and the fraction of reliable workers, α. Throughout, our positive results hold when the number of verified data points, k, is a constant that is independent of n, but dependent on ǫ, δ, and α. 1.2 Summary of Results and Connections to Random CSPs Our main result is the following: Theorem 1. Fix a failure probability δ > 0 and accuracy parameter ǫ > 0. Consider a set of n items that each have a Boolean value, and m reviewers who each evaluate a uniformly random subset of r out of the n items. Suppose that αm of the reviewers are “good” in that each of their r reviews is correct (independently) with probability at least 1 − p ≥ 1/2. Given sufficiently many reviewers, accurate reviews 2 of at least (1 − ǫ)n items can be inferred given the true values of a constant (independent of n) sized random 1 subset of the variables, provided the fraction of good reviewers satisfies α > (2−2p) r. Specifically, given the values of a random, constant-sized subset of the items of size k = Õ 1ǫ · 22r log(1/δ) , with probability at least 1−δ one can recover accurate evaluations of at least (1−ǫ)n of the items, provided 1 r α > (2−2p) r and the number of reviewers m = Θ̃α,δ,ǫ (n ). Additionally, the algorithm runs in time linear in the number of items, n, given the ability to query the dataset for reviewers who have evaluated a given set of items in constant time. Specifically, the runtime of the algorithm is Oδ,ǫ,r (n), where the hidden constant hides an exponential dependence on r, and polynomial dependence on 1/ǫ and log(1/δ). The following straightforward observation demonstrates that the above theorem is optimal in the relationship between the fraction of good reviewers, α, and the number of items reviewed by each individual, r, and the error rate of each good reviewer, p: Observation 1. If each good reviewer incorrectly reviews each item independently with probability p, and 1 the fraction of good reviewers satisfies α = (2−2p) r where r denotes the number of items evaluated by each reviewer, then the remaining (1 − α) fraction of reviewers can behave such that for every set of r items, for a randomly selected reviewer, the distribution of reviews for those items will be uniform over the 2r possible review vectors. Hence the dataset contains no useful information. One reason why Theorem 1 is surprising is that this inverse exponential dependence between the number of reviews per reviewers, r, and the fraction of “good” reviewers, can not be attained via the usual approach of low-rank matrix approximation that is often applied to this problem of recommendation systems (e.g. [2, 9]). To see why these approaches cannot be applied, note that for any matrix in which all rows have at most r entries, there is a rank r matrix that exactly agrees with all entries. Intuitively, each of these r factors is capable of representing a different subset of the reviewers. Still, at best this would result in an algorithm that is capable of capturing r different groups of reviewers; in other words, it seems extremely unlikely that such approaches could yield positive results in the setting where the fraction of “good” reviewers was less than 1/r, in contrast to our results that allow this fraction to be 1/exp(r). The setting of Theorem 1 can be easily mapped into the language of a constraint satisfaction problem. Given the evaluations of the reviewers, we build the constraint satisfaction problem by associating a Boolean variable to each of the n items, and for every set of r variables, we define the set of allowable assignments to those variables to include any of the 2r review vectors that constitutes more than a 1/2r fraction of the review vectors for the associated items. (In other words, if at most a 1/2r fraction of the reviewers who evaluated a given set of r items submitted a vector of reviews σ = (σ1 , . . . , σr ), then σ is not an allowable 1 assignment for those variables.) The requirement that α > (2−2p) r guarantees that, for every set of r items, irrespective of the behaviors of the (1 − α) fraction of bad reviewers, for a randomly selected reviewer, the probability that the r reviews are all correct is strictly larger than 1/2r . Additionally, our requirement on the number of reviewers, m, ensures that with high probability (by elementary concentration bounds) for every set of r items, there are sufficiently many reviewers assigned to that set of r items, so as to ensure that the number of accurate ratings (provided by the good reviewers) exceeds a 1/2r fraction of the overall reviews for that set of r items. Hence, with high probability, we obtain a constraint satisfaction problem such that for every set of r variables 1) the correct assignment is in the set of allowable assignments, and 2) at least one of the 2r possible assignments is disallowed. Given this mapping from the review/evaluation setting to constraint satisfaction problems, Theorem 1 will follow immediately from the following result concerning a class of adversarial constraint satisfaction problems: 3 Theorem 2. Consider a set of n Boolean variables, and a planted assignment σ ∈ {0, 1}n . Suppose that for each subset of r variables, t = {v1 , . . . , vr }, there is a subset Ct ⊂ {0, 1}r of assignments such that \|Ct \| ≤ 2r − 1 and the planted assignment σ (restricted to the variables in t) is in set Ct . Given the ability to query the planted assignment values for a constant number of variables chosen uniformly at random, the planted assignment can be recovered with up to ǫn errors, for any constant ǫ > 0. Specifically, for any ǫ, δ > 0, after querying the values of 1 2r · 2 log(1/δ) k = Õ ǫ variables, with probability at least 1 − δ we can output an assignment σ ′ ∈ {0, 1}n that differs from the planted assignment, σ, in at most ǫn values. Additionally, the algorithm will run in time Or,ǫ,δ (n). There is a simple V C-dimension argument together with a sphere-packing result of Haussler [7] that yields a tighter information theoretic recovery result, yielding an analog of the above theorem with polynomial (rather than super-exponential) dependence on r.1 Specifically, the number of verified assignments must be k = O( 1ǫ (r log(1/ǫ) + log(1/δ)). This V C-dimension approach, however, seems to yield an algorithm with runtime at least nr , as opposed to the linear time algorithms of Theorems 1 and 2. For practical settings, having a linear-time algorithm seems quite important; that said, exploring this problem from an information theoretic perspective is also worthwhile. One natural question is whether one can achieve a best-of-both-worlds: a near-linear time algorithm with a polynomial dependence on r. We discuss this problem more in Section 4. Proposition 1. As in Theorem 2, consider a set of n Boolean variables, and a planted assignment σ ∈ {0, 1}n . Suppose that for each subset of r variables, t = {v1 , . . . , vr }, there is a subset Ct ⊂ {0, 1}r of assignments such that \|Ct \| ≤ 2r −1 and the planted assignment σ (restricted to the variables in t) is consistent with Ct . Given the ability to query the planted assignment values for k = O 1ǫ (r log(1/ǫ) + log(1/δ)) random entries, with probability at least 1 − ǫ one can recover an assignment that disagrees with σ on at most ǫn values. Proof. Let S ⊂ {0, 1}n be the set of assignments that are consistent with all of the sets of partial assignments to the r-tuples specified by the sets Ct . The Vapnik-Chervonenkis (VC) dimension of the set S is at most r, since, by assumption, for every r-tuple of variables, t = {v1 , . . . , vr }, there are at most \|Ct \| ≤ 2r − 1 possible assignments to those variables. As was shown by Haussler (Theorem 1 in [7]), for any subset S of n the Boolean hypercube with VC dimension at most r, for every ǫ > 0 there exists a set T ⊂ {0, 1} of size 2e r at most e(r + 1) ǫ such that for every point x ∈ S, there exists a point tx ∈ T that agrees with x on at least (1 − ǫ)n coordinates. Let Tǫ/2 denote such a covering set corresponding to the set S, such that every x ∈ S is distance at most nǫ/2 from an element of Tǫ/2 . We can use our k = O 1ǫ (r log(1/ǫ) + log(1/δ)) random coordinates of the vector σ ∈ S to find, with probability at least 1 − δ, a point in Tǫ/2 of distance at most nǫ from σ by simply choosing the element of Tǫ/2 that agrees with the largest fraction of the k random samples. This follows from 1) leveraging a Chernoff bound to show that out of the k samples, at most a (2/3)ǫ fraction will disagree with the element of Tǫ/2 that has distance ǫn/2, and 2) a union bound over \|Tǫ/2 \| Chernoff bounds to argue that none of the elements of Tǫ/2 that have distance at least ǫn will disagree in fewer than a (2/3)ǫ fraction of indices. Together, this yields that the probability that the element of Tǫ/2 that agrees with the largest fraction of the k random samples has distance greater than ǫn from the true assignment, is at most \|Tǫ/2 \|exp(O(−kǫ)) = \|Tǫ/2 \|(1/ǫ)O(r) δ, which is at most δ for a suitable choice of the constant in the “O” term of k = O 1ǫ (r log(1/ǫ) + log(1/δ)) . 1 We thank an anonymous reviewer of an early version of this paper for drawing our attention to this. 4 One implication of the above result is that for any Boolean constraint satisfaction problem for which 1) there exists a satisfying assignment, and 2) for every subset of r variables the constraints forbid at least one of the 2r possible assignments, it must be the case that there are only a constant number of “ǫ-similar solution clusters,” where an ǫ-similar solution cluster is a set of assignments that differ from each other in at most ǫn locations. Indeed, the number of such clusters will be at most 2k , where k = Õ(r/ǫ) is as specified in Theorem 2 and Proposition 1, is a bound on the number of variables whose assigned value must be queried to achieve a constant probability of failure δ < 1. Note that this number of solution clusters is independent of n. This structure of the satisfying assignments is slightly surprising given the following two simple examples: the first example illustrates that it is possible for such CSPs to have at least two extremely different satisfying assignments, and the second illustrates that it is possible for such CSPs to have super-constant sized solution clusters—clusters of size Ω(n)—although all the assignments in such a cluster are quite similar. Example 1. Consider the setting where the underlying assignment to all n variables is T , and for every pair of variables, the set of allowable assignments is {(F, F ), (T, T )}. Based on these constraints, there are two possible satisfying assignments—either all T or all F . A single “verified” data point is sufficient to distinguish between these two sets of assignments. The following example illustrates that, in general, it is impossible to guarantee that the learner will correctly output the exact assignment, unless the number of verified datapoints k = Θ(n). Example 2. Consider the setting where each set of r values has the constraint that precludes the (F, F, . . . , F ) r-tuple. In this case, there is a single solution cluster consisting of all assignments to the n variables such that at most r − 1 of the variables are F and the remaining n − r + 1 are T . In this case, it is impossible to distinguish between these assignments with any significant probability using fewer than Θ(n) verified evaluations. Despite the above examples, it is still unclear whether the information theoretic bound of Proposition 1 is tight; particularly for small constant ǫ, it is not clear the extent to which the number of ǫ-separated solution clusters can grow as ǫ decreases. 1.3 Related Work Motivated by the increasing practical importance of robust estimation—and more generally, robust learning and optimization—there has been recent interest in these problems from both an information theoretic and computational perspective. Recent works tackled this general problem in several basic settings, including robust linear regression [1], and robustly estimating the mean and covariance of natural classes of distribution, including multivariate Gaussians [4, 10]. The focus of these works was largely on establishing computationally efficiency algorithms for these tasks that approach the information theoretic (minimax) guarantees achieved by more naive or brute-force algorithms. All three works focussed on the regime in which a majority of the data is assumed to be “good”—drawn from the distribution or cohort of interest. In the case of [1], the recovery guarantees require that this fraction of good data satisfies α ≥ 64 65 . The recent works [12] and [3] consider the setting where a minority of the data is “good” (i.e. α < 1/2), with the latter paper formally proposing the “semi-verified” learning model where one may obtain a small amount of “verified” data that has been drawn from the distribution/cohort in question. The former paper, [12] considers a similar item evaluation setting to the setting we consider, but focusses on the regime where the number of evaluators is on the same order as the number of items being evaluated. In this regime, they show that ǫ-accurate recovery is possible provided that the number of items reviewed by each evaluator is O( ǫ41α3 ) . 5 In contrast, we consider the regime in which the number of evaluators might be significantly larger than the number of items, but establish an optimal tradeoff between the fraction of good reviewers and the number of items evaluated by each reviewer, demonstrating the surprising ability to tolerate a fraction of good evaluators that is inverse exponential in the number of items evaluated by each evaluator. For the context of leveraging these techniques as a “prism” to extract information about specific demographics from a large, mixed dataset, this small-α regime seems especially significant. The techniques of this paper, via local algorithms and the constraint-satisfaction perspective, also differ significantly from the previous approaches to robust estimation which rely on more geometric or spectral structure. The general challenge of developing algorithms and estimators that are robust to corruptions in the input data dates back to the early work of [13], and has led to a significant body of work on “Robust Statistics”, which explores a number of different models for the data corruptions, and largely focusses on the regime in which a majority of the data is “good.” Much of this work is orthogonal to the objectives of this paper, and we refer the reader to the surveys of [8] and [6]. 2 The Algorithm In this section we describe a simplified algorithm that obtains the claimed result of Theorem 2 with the exception of two key properties: as-described, the runtime of this algorithm is O(nr ) rather than Or (n), and this algorithm will require a number of “verified” samples that is inverse polynomial in the error parameter ǫ, as opposed to the nearly inverse linear dependence specified in the theorem. The algorithm to which Theorem 2 applies is an extension of this algorithm, and we describe it in Section 3. The overall structure of the algorithm is to reduce an instance of the problem with non-trivial constraints on all sets of r variables, to an instance of the problem that has non-trivial constraints on all sets of r − 1 variables. In general, the true assignment might not satisfy the constraints that we derive on the sets of r − 1 variables, though we will be able to leverage any such derived constraints that are discovered to be false. We begin by providing the intuition for the algorithm in the case that r = 2, and then in Section 2.2 describe the intuition for the reduction from constraints on sets of r variables to constraints on r − 1-tuples. We formally describe the general algorithm in Section 3. 2.1 Intuition: Restricting to Pessimistic Constraints Our algorithm will proceed iteratively, with the goal of each iteration being to inspect at most a constant number of randomly sampled “verified” variable values, and return accurate guesses for at least a constant fraction of the variables. The algorithm will then recursively iterate this procedure on the remaining variables until all but 2ǫ n variables have been assigned guesses; assignments to these last ≤ ǫn/2 variables can be chosen arbitrarily. To begin, consider the setting where r = 2, and for every pair of variables (x, y) we have a set of allowable assignments, C(x,y) ⊂ {T, F }2 , with \|C(x,y) \| ≤ 3. Each such set provides at least two implications, one of the form x = X =⇒ y = Y and one of the form y = Y ′ → x = X ′ for some choice of X, X ′ , Y, Y ′ ∈ {T, F }. For example, if the assignment (T, F ) 6∈ C(x,y) , then we have the implications x = T =⇒ Y = T and y = F =⇒ x = F . In other words, there is at least one value of variable x that would imply the value of variable y, and similarly for y. Hence, if we fix variable x, and consider the implications derived from the sets C(x,y) as y ranges over all n − 1 other variables, there must be an assignment to variable x that would imply the values of at least n/2 variables. We will refer to this assignment as the “optimistic” value of x, as this assignment to x would immediately yield the values of at least half the remaining variables, and we would be done with the current iteration of the algorithm, and would then recurse on the remaining variables that have not been assigned 6 values. The first key idea of our algorithm is that we will assume that all variables take their “pessimistic” values. We will then “check” this assumption by revealing the true values of a random sample of O(log(1/δ)/ǫ) of these variables. If all of these values are consistent with the “pessimistic” values, we can conclude that with probability at least 1 − δ, at least (1 − ǫ)n of the variables actually take their “pessimistic” values, and hence we can simply output this assignment. If, however, any of our O(log(1/δ)/ǫ) random checks fails, that means that we have found a variable that takes its “optimistic” value, and hence that one variable, x, together with the n − 1 constraint sets C(x,·) that involve it, imply the values of at least n/2 variables. In either case, our constant (dependent on ǫ, δ) number of checks has yielded an accurate assignment to at least half the variables. This simple algorithm in the r = 2 case is summarized in the following pseudo-code: FindAssigments, r = 2: Input: Set of n variables, and for every pair (x, y), a set of allowable assignments to those variables C(x,y) ⊂ {T, F }2 , with \|Ct \| ≤ 3. Error parameter ǫ > 0 and failure parameter δ > 0. Output: Assignments to each of the n variables. • While there exists ≥ ǫn/2 variables without assignments – Let n′ denote the number of remaining variables, and for each of these, determine an “optimistic” assignment that would imply the values of at least n′ /2 other variables, and define a variable’s “pessimistic” value to be the opposite assignment. randomly chosen variables and their “verified” assignments. (If – Consider a set of 10 log(1/δ) ǫ2 of these variables lie in the set of n′ > ǫn/2 variables in consideration, fewer than log(1/δ) ǫ output FAIL) – If all the verified assignments for variables in the set of n′ agree with their pessimistic assignments, then assign these n′ variables their pessimistic assignments. – Otherwise, we must have found a variable whose verified assignment is its optimistic assignment, and we can assign the values of at least n′ /2 variables accordingly. 2.2 From r-tuples to r − 1-tuples: Pessimism All The Way Down Given the algorithm for the r = 2 case, which is successful provided every pair of variables has at least one forbidden assignment, the question is how to reduce the setting with constraints on sets of r ≥ 3 variables, to the setting of constraints on sets of r − 1 variables. The following trivial lemma is the key to this reduction: Lemma 1. Given an r-tuple and set of at most 2r − 1 allowable assignments to those r variables, then for any subset of r − 1 of those variables, there exists an assignment to those r − 1 variables that would imply the value of the rth variable. Proof. Consider a r − 1 tuple, t, and an additional variable v, and the set of ≤ 2r − 1 allowable assignments to the r-tuple (t ∪ v). If the restriction of these assignments to the r − 1 variables in t contains all 2r−1 possible assignments, it must be the case that for at least one of these assignments, there is a unique value that v must assume, otherwise this r-tuple would have all 2r possible assignments. If the restriction of the 2r − 1 assignments to the r − 1 tuple do not contain all 2r−1 assignments, then any such assignment would (vacuously) imply the value of the rth variable. The utility of this lemma is that if we have an r − 1-tuple of variables, t, then by considering all possible additional variables v 6∈ t, there exists an assignment to t that determines the value of at least a 1/2r−1 7 fraction of the variables not in t. Hence we can designate an “optimistic” assignment with the property that if that assignment holds, then it will imply assignments to at least a 1/2r−1 fraction of the remaining variables. We will then assume that this “optimistic” assignment is not allowed, thereby reducing the set of allowable assignments of variables in t to size 2r−1 − 1, and proceed inductively. In this sense, at some intermediate step of this algorithm where we are considering sets of r ′ < r variables, the allowable sets of assignments that we are considering may not be completely accurate, as we are not verifying whether the sets actually do take their “optimistic” assignments or not. However, if a r ′ -tuple of variables actually takes the values of a forbidden/optimistic assignment, then either it will immediately imply the values of a constant (i.e. at least 1/2r ) fraction of variables, or it must be a subset of a larger tuple that takes its “optimistic” assignment. Which of these two cases holds can be easily decided via querying the values of a (constant) number of random variables. We describe the full algorithm in the following section. 2.3 The Basic Algorithm The high-level structure of the algorithm described in the previous part takes the form of a “descending” pass followed by an “ascending” pass. In the descending pass, we iteratively turn constraints on r0 tuples into constraints on r0 − 1 tuples, then r0 − 2 tuples, etc; all the while, we forbid “optimistic” assignments to ensure that in the rth level, each r tuple has at most 2r − 1 allowable assignments. This descending phase terminates with r = 1, where we have our “pessimistic” conjectured assignments to all variables. We then randomly check a few of these values; if we do not discover any inconsistencies with the conjectured values, then we can safely conclude that most of the conjectured values are correct. If we have discovered any inconsistencies, then we begin the ascending phase that investigates and checks any discovered “optimistic” assignments. One minor wrinkle is that we should not trust the > 1/2r fraction of values that appear to be implied by an optimistic assignment to a set of r < r0 −1 variables. These implications might be the result of forbidding an optimistic assignment for some larger tuple. Nevertheless, if we randomly check some of the implications, then we will either verify the accuracy of these implications, or have found an optimistic assignment to a r + 1 tuple. In this sense, the ascending phase will either terminate upon satisfactorily verifying a significant (constant sized) subset of the set of output assignments, or we will have found an “optimistic” assignment to a r0 − 1 tuple, and the implications of r0 − 1 tuples are based directly on the given set of constraints to r0 -tuples, which are valid by assumption. Hence each phase of the algorithm will return assignments to a constant (at least 1/2r0 ) fraction of the variables. FindAssigments: Input: Set of n variables, integer r0 , and for every tuple t ⊂ [n]r0 of r0 distinct variables, a set of allowable assignments to those variables Ct ⊂ {T, F }r0 , with \|Ct \| ≤ 2r0 − 1. Error parameter ǫ > 0 and failure parameter δ > 0. Output: Assignments to at least (1 − 2ǫ )n variables. • While there exists ≥ ǫn/2 variables without assignments, run DESCEND on the set of unassigned variables and their corresponding sets of allowable assignments. 8 DESCEND: Input: For each r-tuple, a set C of assignments to those variables, with \|C\| ≤ 2r − 1. 1. If r = 1, AscendAndVerify(set of assigned values to each variable, r = 1). 2. Else, for every r − 1 tuple, t, we will create a set Ct of ≤ 2r−1 − 1 assignments: • Find an “optimistic” assignment σt that would determine at least a 1/2r fraction of variables not in t. (The existence of such an assignment is guaranteed by Lemma 1.) • Set Ct = {T, F }r−1 \ σt . 3. Run DESCEND on the set of r − 1-tuples and their corresponding sets of assignments, each of size 2r−1 − 1. ASCEND AND VERIFY: Input: Proposed assignments σv for each variable v ∈ V for some set V of variables. Integer r indicating the size of the tuples whose constraints generated the proposed assignments, and assignment σt to a r − 1-tuple t, such that Ct∪v provided the implication σt → σv . Access to sets of allowable assignments corresponding to all tuples of size r ′ ∈ {r, . . . , r0 }. Constant A = 2r0 log(1/δ) log(1/ǫ))/ǫ2 . 1. Randomly sample A verified variable assignments. 2. If all verified variable assignments agree with the proposed assignments, σv , then permanently assign v ∈ V with their proposed assignments, σv . 3. Otherwise, let v denote a variable whose true/verified assignment av 6= σv , disagrees with the proposed assignment to v. Hence (σt , av ) 6∈ Ct∪v so assignment (σt , av ) together with the constraints on the r + 1 tuples must imply at least a 1/2r fraction of variable assignments. Denote these assignments by σ new . • Run AscendAndVerify(σ new , r + 1, {t ∪ v}, (σt , av )) 3 An Efficient Algorithm The linear-time variant of the basic algorithm described in the previous section hinges on two observations. The first is that for a given r-tuple t, rather than consulting all Θ(n) constraints Ct∪xi for all xi 6∈ t to determine the “optimistic” assignment to t, one can determine an assignment that implies at least a 21 21r fraction of the variable values, with high probability, via sampling a constant (independent of n but dependent on r, ǫ, δ) number of such constraints. Note that this sampling does not look at any of the “verified” variable assignments—it just samples which of the constraints to consider. We formalize this ability to efficiently determine an “optimistic” assignment via the following subroutine, and the following lemma characterizing its performance. 9 FIND OPTIMISTIC ASSIGNMENT: Input: Set of n variables X, r-tuple t, the ability to query constraints Ct′ for \|t′ \| = r0 ≥ r (i.e. the ability to find optimistic assignments to tuples t′ with \|t′ \| = r0 ) and probability of failure γ > 0. Output: An optimistic assignment σt to t that would, with probability at least 1 − γ, imply the assignments to at least a 1/2r+1 fraction of other variables via the constraints Ct∪x . We define Ct := {T, F }r \ {σt }. 1. If r = r0 then return constraint Ct . 2. Else • Select s = 3 · 2\|t\| log(1/γ) variables x1 , . . . , xs uniformly at random from X \ t. • For each of these s variables, xi , compute Ct∪xi via a (recursive) call to F indOptimisticAssignment(X, t ∪ xi , P robF ailure = γ/(2s)). • Define assignment σt ∈ {T, F }r to be the lexicographically first assignment that, via the constraints {Ct∪xi }, imply at least a 1/2r fraction of variables {x1 , . . . , xs }. [Note that such an assignment exists, since for each xi , \|Ct∪xi \| ≤ 2r+1 − 1 has at least one out of the 2r possible assignment to t that would imply that value of xi .] • Call σt the “optimistic” assignment to tuple t, and store Ct = {T, F }r \ σt . The following two lemmas quantify the performance of the above algorithm. The first lemma characterizes the probability of failure, and the proof follows immediately from standard Chernoff tail bounds. Lemma 2. With probability at least 1 − γ the optimistic assignment σt returned by algorithm FindOptimisticAssignment on input X and t has the property that for at least a 1/2\|t\|+1 fraction of variables x ∈ X, the assignment σt together with the constraint set Ct∪x that would be computed by the algorithm on input tuple t ∪ x, implies the value of variable x. Proof. Letting p denote the true fraction of variables, x, whose assignments are implied by σt and Ct∪x . Recall that σt was chosen based on s independent samples, yielding an empirical estimate p̂ ≥ 1/2\|t\| , and − s standard tail bounds yield that Pr[p̂ > 2p] ≤ e 3·2\|t\| , yielding the lemma, since s = 3 · 2\|t\| log(1/γ). Lemma 3. Given constant-time query access to the constraint sets Ct′ for tuples satisfying \|t′ \| = r0 , for any tuple t, algorithm FindOptimisticAssignment on input t and probability of failure γ > 0 returns Ct and 2 runs in time/queries (2 log(1/γ))O(r0 ) , which is independent of the size of the variable set, \|X\|. Proof. Note that computing Ct calls O(2\|t\| log(1/γ)) computations of Ct′ for \|t′ \| = \|t\| + 1, each called with error parameter 2\|t\| smaller. When \|t\| = r0 , Ct is obtained via a single constant-time query. Expanding this recursion yields the above lemma. The second observation that underpins the efficient algorithm is that we do not need to determine the optimistic assignments and form constraints Ct for all nr r-tuples t. For each phase of the algorithm, which returns assignments to a constant fraction of the unassigned variables—at least 1/2r0 —it suffices to find a single tuple t that takes its “optimistic” assignment. Indeed, such a tuple, by definition, takes values that imply assignments to a constant fraction of the remaining variables. And for each of these variables, x, whose assignment is implied by the assignment to the tuple t, the value of variable x can be determined in constant time by consulting the constraint Ct∪x . This observation is clarified in the following algorithm, which is an adaptation of the Descend/AscendAndVerify algorithm described in the previous section. Finally, we highlight the fact that the algorithm proceeds iteratively. Given an initial set of variables, Y , at some intermediate step in the algorithm, we let X denote the set of variables for which we have not 10 yet output an assignment. The algorithm will terminate when \|X\| ≤ ǫ\|Y \|/2. The goal of the current step of the algorithm will be to output assignments to at least a 1/2r0 fraction of variables in X, such that the \|Y \| ǫ fraction of such assignments that are incorrect is bounded by 2 log(2/ǫ) \|X\| . Given this bound on the fraction of incorrect returned at this phase of the algorithm, the total fraction of errors is bounded by R 1 assignments ǫ · 1t dt = ǫ where the first ǫ/2 is a bound on the error due to the arbitrary assignments to ǫ/2 + t=ǫ/2 2 log(2/ǫ) the last ≤ ǫ\|Y \|/2 variables. The benefit of having the target accuracy increase as \|X\| decreases is because we are given verified samples, drawn uniformly at random from Y . To “check” a proposed assignment to set X to a target accuracy of γ, we need at least 1/γ verified samples from the set X (ignoring the logarithmic dependence on the probability of failure). To guarantee that this number of verified samples is obtained from \|Y \| verified samples from Y . Using the above trick of having the desired set X, we will need to draw ≈ γ\|X\| accuracy degrade as \|X\| decreases, for each phase of the algorithm, a set of \|Y \|/\|X\| \|Y \| ǫ 2 log(2/ǫ) \|X\| = Õ(1/ǫ) verified samples is required—as opposed to the Θ(1/ǫ2 ) samples that would have been required if we had fixed the target error rate to be ǫ for all rounds of the algorithm. EFFICIENT FIND ASSIGNMENTS: Input: Set of n variables Y , integer r0 and for every r0 -tuple t ⊂ Y , a set of allowable assignments Ct with \|Ct \| ≤ 2r0 − 1. Error parameter ǫ > 0 and probability of failure δ. Output: Set of T /F assignments to each x ∈ Y . • Set T = r0 · 2r0 +1 log(2/ǫ). • While there are at least ǫn/2 unassigned variables: 1. Let X ⊂ Y denote the set of unassigned variables. 2. Let ǫX = \|Y \| ǫ 2 log(2/ǫ) \|X\| \| denote the target accuracy of this round, and set s = 10 ǫX\|Y\|X\| log(10T /δ). 3. Take s verified samples, revealing the planted assignment values for each of these variables. Let X1 ⊂ X denote the subset of these variables that are in set X, and for each x ∈ X1 let ax denote \|X\| the verified assignment to variable x. If \|X1 \| < s 2\|Y \| output FAIL. 4. For each x ∈ X1 , determine Cx via FindOptimisticAssignments with failure parameter γ = δ/T . 5. If, for all x ∈ X1 , ax = Cx , then for every variable x ∈ X, compute and output assignment Cx . 6. Otherwise, let x1 ∈ X1 denote a variable for which ax EfficientAscend(X, i, x1 , ax1 , s). 11 6= Cx , and run EFFICIENT ASCEND: Input: Set of variables X, integer i ∈ {1, . . . , r0 − 1}, tuple t with \|t\| = i, verified assignments at ∈ {T, F }\|t\| to tuple t, and parameter s. Output: Output to a subset of variables in set X. 1. If i ≥ r0 output FAIL. 2. Take si = s · 2i verified samples, and let Xi+1 denote the intersection of X with this set of variables with verified assignments, with ax denoting the verified assignment to variable x ∈ Xi+1 . 3. For each x ∈ Xi+1 , determine Ct∪x via a call to FindOptimisticAssignment(X, t∪x, F ailureP rob = ′ δ/(10T · si )), and let Xi+1 ⊂ Xi+1 denote the subset of variables x ∈ Xi+1 for which the constraint ′ \| ≤ s \|X\| output FAIL. Ct∪x together with at implies a value σx for x. If \|Xi+1 i 4·2i \|Y \| ′ , it holds that σ = a , then for every variable x ∈ X, compute C 4. If, for all x ∈ Xi+1 x x t∪x and output assignment σx if σx is implied by Ct∪x and at . ′ 5. Otherwise, let xi+1 ∈ Xi+1 denote a variable for which Ct∪xi+1 and at implies assignment σxi+1 6= axi+1 . Run EfficientAscend(X, i + 1, t ∪ xi+1 , (at , axi+1 ), s). Proposition 2. Algorithm EfficientFindAssignments, when run with error parameter ǫ and probability of failure δ, has the following properties: • The algorithm will require at most Õ 1ǫ · 22r0 log(1/δ) verified samples drawn uniformly at random from the set of variables, Y . • With probability at least 1 − δ, the algorithm will output assignments to each variable x ∈ Y , such that at most an ǫ fraction of the assignments disagree with the planted assignment. • The algorithm runs in time Or0 ,ǫ,δ (n), where the hidden constant is a function of r0 , ǫ, δ. Proof. The high level outline of the execution of algorithm EfficientFindAssignments is that in each step of the outer WHILE loop, an assignment to at least a 1/2r0 +1 fraction of the remaining unassigned variables, X, will be output. This continues until \|X\| ≤ ǫn/2, at which point these remaining variables can be assigned arbitrary labels and the algorithm terminates. Hence there will be at most O (2r0 log(1/ǫ)) iterations of the while loop. In the iteration conducted on unassigned variable set X ⊂ Y , the goal will be to return assignn ǫ ments such that the fraction of returned assignments that are incorrect is at most 2 log(2/ǫ) \|X\| , where \|Y \| = n is the total number of initial variables. Provided these accuracy goals are met at each step of the algorithm, R1 ǫ the overall fraction of errors will be bounded by ǫ/2 + f =ǫ/2 f1 · 2 log(2/eps) df = ǫ, where the first term is the errors due to the arbitrary assignment to the remaining ≤ ǫn/2 variables. Additionally, the number of verified samples required in each iteration is at most O(r0 s · s2r0 ) = Õ 2r0 1ǫ log(1/δ) , hence the total number of verified samples across the O (2r0 log(1/ǫ)) iterations will be bounded by Õ(22r0 log(1/δ)/ǫ), as claimed. We now analyze each run of the WHILE loop in EfficientFindAssignments, and the recursive calls to EfficientAscend. At a high level, in each recursive call to EfficientAscend, either an assignment to at least a 1/2r0 +1 fraction of the remaining unassigned variables is returned via the implications from some (verified) optimistic assignment to a tuple, t; or, we have found a tuple t∪xi+1 for which we have verified assignments to all \|t\| + 1 variables, and for which that assignment, (at , axi+1 ) 6∈ Ct∪xi+1 is the optimistic assignment, in which case the subsequent call to EfficientAscend considers this strictly larger tuple t′ = t ∪ xi+1 . 12 To bound the runtime of the algorithm, note that each run of the algorithm requires constant time (dependent on r0 , ǫ, δ but independent of the number of variables, \|X\|, up until the point in the algorithm when an assignment will be output (Step 4 in EfficientAscend). At this point in the algorithm, at a computational expense of Or0 ,ǫ,δ (\|X\|), an assignment to a constant fraction, at least 1/2r0 +1 of the remaining variables will be output, and the algorithm will then be repeated on the remaining unassigned variables. Hence, the overall runtime of the algorithm will be linear in the number of variables. To bound the probability that a given run of the WHILE loop fails to successfully output an assignY ǫ ment to at least \|X\|/2r0 +1 variables that meets the target accuracy of 2 log(2/ǫ) \|X\| , we will leverage a union bound over a number of standard Chernoff tail bounds. First, note that the probability that EfficientFindAssignments outputs ’FAIL’ in Step 3 in a given round of the algorithm is bounded by the probability that \|X1 \| ≤ E[\|X1 \|/2], where X1 is a sum of i.i.d 0/1 random variables, hence this probability is bounded by δ , where T, as specified in EfficientFindAssignments is a bound on the number of exp(−E[\|X1 \|]/8) ≤ 10T calls to EfficientAscend which bounds the number of runs of the WHILE loop. Given that \|X1 \| ≥ E[\|X1 \|/2], the probability that the assignment output in Step 5 of EfficientFindAssignments does not meet the target Y ǫ \|X1 \| ≤ δ . accuracy, ǫX = 2 log(2/ǫ) 10T \|X\| , is bounded by (1 − ǫX ) The remaining probability of failure stems from the execution of EfficientAscend. In this algorithm, failure can stem from three different issues: 1) the constant number of constraints C· computed via FindOptimisticAssignment prior to Step 4 of EfficientAscend can be erroneous and fail to imply the desired fraction of assignments. The probability of this is bounded by δ/(10T si ), which is sufficient to guarantee that every optimistic assignment/constraint set C· that is computed during the execution of the algorithm is accurate and implies the desired fraction of assignments, aside from the O(\|Y \|) constraints computed during the assignment output steps—Step 4 of EfficientAscend and Step 5 of EfficientFindAssignments. EfficientAscend will never output FAIL during Step 1, as the constraints corresponding to to i = r0 are the constraints on r0 -tuples, which are satisfied by assumption. The final potential failure mode of the algorithm is Step 3 of EfficientAscend, if the random set of verified assignments is insufficiently large to verify (to the target accuracy) a given potential set of assignments implied by an optimistic assignment via Ct∪x . Given that the assignment at to tuple t is optimistic, as guaranteed by the validity of FindOptimisticAssignments described above, this probability of failure is also a trivial application of standard Chernoff bounds, guaranteeing that ′ \| in Step 3 of EfficientAscend deviates from a lower bound on its expectation by the random variable \|Xi+1 at most a factor of 1/2. A union bound over these probabilities of failure for each of the ≤ T runs of the EfficientAscend algorithm yields the desired proposition. 4 Future Work This work shows that it is possible to tolerate a fraction of “good” data, α, that is inverse exponential in r0 , the sparsity of each datapoint (i.e. the number of evaluations submitted per reviewer), provided the number of datapoints/reviewers is sufficiently large to ensure that each set of r items has been evaluated by a significant number of good reviewers. Our algorithm runs in time linear in the number of items to review (provided the ability to query summary statistics of the set of reviewers who have evaluated a given sets of items), and uses a constant number of “verified” reviews, which is independent of the total number of items to review, and depends inverse linearly on the desired error (to logarithmic factors). One natural question is prompted by the results of [12], which provide efficient algorithms for the regime where r = poly(1/α), but where the number of reviewers is linear in the number of items being reviewed (and uses a constant, dependent on α, ǫ, δ verified reviews). Is it possible to achieve the best-of-both-worlds: r = polylog(1/α), and a number of reviewers that is linear, or grows significantly more slowly than the nr that we require, while leveraging a constant number of verified reviews? 13 To this end, our algorithm only ever considers “single-hop” implications of proposed assignments: an assignment to a set of r variables is considered “optimistic” if it directly implies values for a significant fraction of the other variables. It is easy to imagine extending this definition to also consider longer chains of implication. Perhaps a specific assignment to r variables would imply values to c1 additional variables, which in turn would imply values to c2 variables, etc. Indeed, in the basic setting of r = 2, this approach can be realized to yield an algorithm that only requires constraints on a random subset of size O(n3/2 ), as opposed to the O(n2 ) constraints assumed in this work. From a computational perspective, it seems unlikely that such an approach could be pushed to yield an efficient algorithm for the regime in which fewer than nr/2 sets of r variables have nontrivial constraints. Indeed, even for random instances of r − SAT with a planted solution, efficient algorithms below this threshold have been elusive (see, for example, the recent related work on random CSPs with planted assignments [5, 11]). From a purely information theoretic perspective—the picture is not entirely clear either. In contrast to random CSPs, our setting is complicated by the adversarial nature of the constraints that are placed on the r-tuples. Even for a semi-adversarial setting where tuples are chosen at random, and an adversary chooses the constraints to place on those random tuples, it is not immediately clear how to analyze the extent to which implications “propagate”. A second difficulty is that the goal of our setting is not just to find a satisfying assignment, but to find something close to a specific planted assignment. Our results imply, for the setting we consider, that there are at most a constant number of solution clusters. It seems interesting to investigate the extent to which this holds for semi-adversarial CSPs with fewer constraints, perhaps with nr/2 constraints being the threshold between a constant and super-constant number of such clusters. References [1] Kush Bhatia, Prateek Jain, and Purushottam Kar. Robust regression via hard thresholding. In Advances in Neural Information Processing Systems, pages 721–729, 2015. [2] Emmanuel J Candes and Yaniv Plan. Matrix completion with noise. Proceedings of the IEEE, 98(6):925–936, 2010. [3] M. Charikar, J. Steinhardt, and G. Valiant. Learning from untrusted data. In Symposium on Theory of Computing (to appear), 2017. [4] Ilias Diakonikolas, Gautam Kamath, Daniel M Kane, Jerry Li, Ankur Moitra, and Alistair Stewart. Robust estimators in high dimensions without the computational intractability. In Foundations of Computer Science (FOCS), 2016 IEEE 57th Annual Symposium on, pages 655–664. IEEE, 2016. [5] Vitaly Feldman, Will Perkins, and Santosh Vempala. On the complexity of random satisfiability problems with planted solutions. In Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, pages 77–86. ACM, 2015. [6] Frank R Hampel, Elvezio M Ronchetti, Peter J Rousseeuw, and Werner A Stahel. Robust statistics: the approach based on influence functions, volume 114. John Wiley & Sons, 2011. [7] David Haussler. Sphere packing numbers for subsets of the boolean n-cube with bounded vapnikchervonenkis dimension. Journal of Combinatorial Theory, Series A, 69(2):217–232, 1995. [8] Peter J Huber. Robust statistics. Springer, 2011. [9] Raghunandan H Keshavan, Andrea Montanari, and Sewoong Oh. Matrix completion from a few entries. IEEE Transactions on Information Theory, 56(6):2980–2998, 2010. 14 [10] Kevin A Lai, Anup B Rao, and Santosh Vempala. Agnostic estimation of mean and covariance. In Foundations of Computer Science (FOCS), 2016 IEEE 57th Annual Symposium on, pages 665–674. IEEE, 2016. [11] Prasad Raghavendra, Satish Rao, and Tselil Schramm. Strongly refuting random csps below the spectral threshold. arXiv preprint arXiv:1605.00058, 2016. [12] Jacob Steinhardt, Gregory Valiant, and Moses Charikar. Avoiding imposters and delinquents: Adversarial crowdsourcing and peer prediction. In Advances in Neural Information Processing Systems, pages 4439–4447, 2016. [13] John W Tukey. A survey of sampling from contaminated distributions. Contributions to probability and statistics, 2:448–485, 1960. 15	7cs.IT
SCUT-FBP5500‡: A Diverse Benchmark Dataset for Multi-Paradigm Facial Beauty Prediction Abstract—Facial beauty prediction (FBP) is a significant visual recognition problem to make assessment of facial attractiveness that is consistent to human perception. To tackle this problem, various data-driven models, especially state-of-the-art deep learning techniques, were introduced, and benchmark dataset become one of the essential elements to achieve FBP. Previous works have formulated the recognition of facial beauty as a specific supervised learning problem of classification, regression or ranking, which indicates that FBP is intrinsically a computation problem with multiple paradigms. However, most of FBP benchmark datasets were built under specific computation constrains, which limits the performance and flexibility of the computational model trained on the dataset. In this paper, we argue that FBP is a multi-paradigm computation problem, and propose a new diverse benchmark dataset, called SCUT-FBP5500, to achieve multi-paradigm facial beauty prediction. The SCUTFBP5500 dataset has totally 5500 frontal faces with diverse properties (male/female, Asian/Caucasian, ages) and diverse labels (face landmarks, beauty scores within [1, 5], beauty score distribution), which allows different computational models with different FBP paradigms, such as appearance-based/shape-based facial beauty classification/regression model for male/female of Asian/Caucasian. We evaluated the SCUT-FBP5500 dataset for FBP using different combinations of feature and predictor, and various deep learning methods. The results indicates the improvement of FBP and the potential applications based on the SCUT-FBP5500. I. I NTRODUCTION Assessing facial beauty seems natural for human being, but an absolute definition of facial beauty remains elusive. Recently, facial beauty prediction (FBP) have attracted evergrowing interest in the pattern recognition and machining learning communities [8]–[10], which aims to achieve automatic human-consistent facial attractiveness assessment by a computational model. It has application potential in facial makeup synthesis/recommendation [10]–[12], content-based image retrieval [16], aesthetic surgery [10], or face beautification [4]–[7]. From the computational perspective, FBP is still a challenging problem. It is involved with the formulation of visual This work was supported in part by the National Natural Science Foundation of China (NSFC) under Grant 61472144 and Grant 61502176; in part by GDSTP under Grant 2015B010101004, Grant 2015B010130003, Grant 2015B010131004; in part by the National Key Research & Development Plan of China under Grant 2016YFB1001405; and in part by Fundamental Research Funds for the Central Universities (No.2017BQ058). Corresponding author: Lianwen Jin. Beau utyscore e arXiv:1801.06345v1 [cs.CV] 19 Jan 2018 Lingyu Liang, Luojun Lin, Lianwen Jin, Duorui Xie and Mengru Li South China University of Technology, Guangzhou 510641, China E-mail:{lianwen.jin, lianglysky}@gmail.com ‡ Dataset Download URL: https://github.com/HCIILAB/SCUT-FBP5500-Database-Release 5 4 3 2 1 0 Asia Female 1 2 3 4 5 Asia Male Caucasian Female Caucasian Male Fig. 1. The images with different facial properties and beauty scores from the proposed SCUT-FBP5500 benchmark dataset. The dataset download URL is shown below the title. representation and predictor for the abstract concept of facial beauty. To tackle this problem, various data-driven models, were introduced into FBP. One line of the works follows the classic pattern recognition process, which constructs the FBP system using the combination of the hand-crafted features and the shallow predictors. The related hand-crafted feature derived from visual recognition includes the geometric features, like the geometric ratios and landmark distances [8], [15], [19]–[21], [23], and the texture features, like the Gabor/SIFT-like features [24]–[26]. Then, a shallow FBP predictoris trained by the extracted feature in a statistical manner. Another line of works is advanced by the reviving of neural networks, especially the stat-of-the-art deep learning [14]. The hierarchial structure of the deep learning model allows to build an end-to-end FBP system that automatically learns both the representation and the predictor of facial beauty simultaneously from the data. Many works indicate that FBP TABLE I R EPRESENTATIVE DATABASES FOR FACIAL B EAUTY P REDICTION Database Image Num. Labelers/Image Beauty Class Face Property Face Landmarks Publicly Available × √ × Y. Eisenthal et al. [15] 184 28 or 18 7 Caucasian Female F. Chen et al. [22] 23412 unknown 2 Asian Male/Female H. Gunes et al. [35] 215 46 10 Female J. Fan et al. [23] 432 30 7 Generated Female M. Redi et al. [34] 10141 78-549 10 Multiple (Sampled from AVA [16]) SCUT-FBP [1] 500 70 5 Asian Female SCUT-FBP5500 5500 60 5 Asian/Caucasian; Male/Female based on deep learning is superior to the shallow predictors with hand-crafted facial feature [1]–[3], [29], [30], [33]. Most of current FBP models are data-driven, which makes benchmark dataset become one of the essential elements for FBP. There have been many works of the benchmark datasets [16], [22]–[24], [30], [33]–[35] involved with FBP, but most of these datasets focus on a specific problem with specific computation constrains, as shown in Table I. Yan et al. [24] regarded FBP as a ranking problem, and proposed a dataset with low-resolution images gathered from social networks. Fan et al. [23] focused on the geometry analysis of FBP and proposed a dataset containing computergenerated faces with different facial proportions. The Northeast China database [22], Shanghai database [21], Hot-Or-Not database [32], [33], AVA database [16] and re-sampled face subset of AVA database [34] are large-scale databases involved with FBP, where the Northeast China [22] and Shanghai database [21] are limited for geometric facial beauty analysis without attractiveness ratings; Hot-Or-Not database [32], [33] only focuses on the appearance-based FBP; and the AVA database [16], [34] is originally designed for aesthetic analysis of entire images but not the facial attractiveness. In our previous work, Xie et al. [1] published a SCUT-FBP benchmark dataset, which has led to many FBP models [1]–[3], [31], [37], especially the hierarchial CNN-based FBP models with the state-of-the-art deep learning [1]–[3], [37]. Despite the prevalent usage of the SCUT-FBP, it only contains 500 Asian Female faces, which may limit the performance of the datademanded model for FBP. We find that FBP have been formulated the recognition of facial beauty as a specific supervised learning problem of classification [15], [20], [27], [30], regression [13], [19], [23], [29], [38] or ranking [24], [25], [37]. It indicates that FBP is intrinsically a computation problem with multiple paradigms. Previous databases built under specific computation constrains would limit the performance and flexibility of the computational model trained on the dataset, and it is difficult to compare different models derived from the dataset with specific computation paradigm. Therefore, this paper argues that FBP is a multi-paradigm computation problem, and proposes a new diverse benchmark dataset, called SCUTFBP5500, to achieve multi-paradigm facial beauty prediction. √ √ × √ √ × × × √ √ √ The SCUT-FBP5500 dataset has totally 5500 frontal faces with diverse properties (male/female, Asian/Caucasian, ages) and diverse labels (face landmark, beauty score, beauty score distribution), which allows different computational model with different FBP paradigms, such as appearance-based/shapebased facial beauty classification/regression/ranking model for male/female with Asian/Caucasian. Furthermore, the diverse faces with beauty scores gathered from 60 different labelers can lead to many interesting research, such as cross-culture facial beauty analysis, personalized FBP [36], or automatic face beautification [4]–[7]. Both shallow prediction model with hand-crafted feature and the state-of-the-art deep learning models were evaluated on the dataset, and the results indicates the improvement of FBP and the potential applications by the SCUT-FBP5500. The main contributions of this paper can be summarized as following: 1) Dataset. We propose a new large-scale SCUT-FBP5500 benchmark dataset that has totally 5500 frontal faces with diverse properties and diverse labels, which allows construction of FBP models with different paradigms. 2) Benchmark Analysis. We analyze the samples, score labels, labelers and facial landmarks of the SCUTFBP5500 statistically, and the visualization of the data illustrates the properties of the SCUT-FBP5500. 3) Facial Beauty Prediction Evaluation. Both shallow prediction model with hand-crafted feature and deep learning models are trained on the SCUT-FBP5500 for evaluation, and the results indicates the improvement of FBP based on the proposed dataset with better diversity. II. C ONSTRUCTION OF SCUT-FBP5500 DATASET A. Face Images Collection The SCUT-FBP5500 Dataset contains 5500 frontal, unoccluded faces aged from 15 to 60 with neutral expression. It can be divided into four subsets with different races and gender, including 2000 Asian females, 2000 Asian males, 750 Caucasian females and 750 Caucasian males. Most of the images of the SCUT-FBP5500 were collected from Internet, where some portions of Asian faces were from the DataTang [39] and some Caucasian faces were from the 10k US Adult database [40]. TABLE II T HE O UTLIER N UMBER AND P ORTION OF B EAUTY S CORES OF C AUCASIAN FEMALE (CF), C AUCASIAN MALE (CM), A SIAN FEMALE (AF) AND A SIAN MALE (AM). Subset Total Score Num. Outlier Num. Outlier Portion CF 45000 143 0.3% CM 45000 181 0.4% AF 120000 356 0.3% AM 120000 497 0.4% (a) Score distribution of CF (b) Score distribution of CM (c) Score distribution of AF (d) Score distribution of AM Fig. 2. Gaussian fitting (yellow curve) and piecewise fitting (red and blue curve) for the visualization of the beauty score distribution of Caucasian female (CF), Caucasian male (CM), Asian female (AF) and Asian male (AM). Fig. 3. Distribution of standard deviations of Caucasian female, Asian female, Caucasian male and Asian male, respectively. (a) Box figure of CF (b) Box figure of AF (c) Box figure of CM (d) Box figure of AM B. Facial Beauty Scores and Facial Landmarks All the images are labeled with beauty scores ranging from [1, 5] by totally 60 volunteers aged from 18-27 (average 21.6), where the beauty score 5 means most attractive and so on. We developed a web-based GUI system to obtain the facial beauty scores. The labeling system was deployed on the Ali Cloud, and the labeling tasks are distributed to each volunteer in crowdsourcing manners. The four subset, Asian male/female and Caucasian male/female, are labeled separately, where each face of the subset are randomly shown to the volunteer. Then, the volunteer are asked to select the beauty scores within [1, 5] for the face. To reduce the variance in the labeling process, about 10% faces recurred randomly. If the correlation coefficient of the two beauty score of the same faces is less than 0.7, the volunteer would be asked to rate this face once more to decide the final score. To allow geometric analysis of facial beauty, 86 facial landmarks are located to the significant facial components of each images, such as eyes, eyebrows, nose and mouth. A GUI landmarks location system is developed, where the original location of the landmarks are initialized by the active shape model (ASM) trained by the SCUT-FBP dataset. Then, the detected landmarks by ASM are modified manually by volunteers to ensure the accuracy. Fig. 4. Box figures of standard deviations of Caucasian female (CF), Asian female (AF), Caucasian male (CM) and Asian male (AM), respectively. III. B ENCHMARK A NALYSIS OF THE SCUT-FBP5500 We made benchmark analysis of the beauty scores, labelers and face landmarks of the SCUT-FBP5500 with different gender and races, including Asian female (AF), Asian male (AM), Caucasian female (CF) and Caucasian male (CM). A. Distribution of Beauty Scores We visualize the distribution of the beauty scores of AF, AM, CF and CM, respectively. To obtain better visualization, we preprocess the data and filter the outliers of the beauty scores. We regard the average score of all the 60 labelers as the ground-truth. If the score of specific labeler for the same face differs from the ground-truth over 2, the score is treated as outlier and is removed from the distribution visualization. TABLE III C ORRELATION C OEFFICIENTS BETWEEN M ALE AND F EMALE L ABELERS FOR B EAUTY S CORE OF C AUCASIAN FEMALE (CF), C AUCASIAN MALE (CM), A SIAN FEMALE (AF) AND A SIAN MALE (AM). Female Labelers Male Labelers All Labelers CF 0.785 0.791 0.788 AF 0.800 0.795 0.785 CM 0.747 0.763 0.743 AM 0.793 0.797 0.782 All Faces 0.785 0.781 0.770 distributions of the original data and the preprocessed data is mostly similar. Therefore, we visualized the score distribution of the SCUT-FBP5500 using the preprocessed data for all the four subset, respectively. Two distribution fitting schemes are used: one is Gaussian fitting (yellow curve), the other is piecewise fitting (red and blue curve), as shown in Fig. 2. The results indicates that the beauty scores of all the four subset can be approximately fitted by a mixed distribution model with two Gaussian components. B. Standard Deviation of Beauty Scores Female Labeler Male Labeler All Labeler CF AF CM AM All Faces We calculate the standard deviation of the scores gathered from different labelers to the ground-truth, and illustrate the results as histogram in Fig. 3 and as box figure in Fig. 4. We observe that the distribution of standard deviations is similar to Gaussian distribution, and most of the standard deviations are within a reasonable range of [0.6, 0.7]. C. Correlation of Male/Female Labelers 2nd PC 1st PC Fig. 5. Correlation Coefficient of male and female labelers for beauty score of CF, AF, CM, AM and all the faces in SCUT-FBP5500. In this subsection, we investigate the correlation between the male and female Asian labelers for the beauty scores, as shown in Table III and Fig. 5. We observe that the correlation of Asian faces is persistently larger than Caucasian. It is consistent to the psychological research that human have better facial beauty perception for the faces from the same race. 5th PC 4th PC 3rd PC D. PCA Analysis of Facial Geometry mean-4sigma mean mean+4sigma (a) PCA Analysis of AF 1st PC ⬋ᗻϢཇᗻҎ㜌ᔶ⢊ⱘЏ៤ߚব࣪ We visualize the 86-points face landmarks of the dataset using principle component analysis (PCA). Fig. 6 illustrate the mean and the five first principle component of the facial geometry of Asian female (AF) and Asian male (AM), where the landmarks data of Caucasian share similar distribution to Asian faces. We observe that the face shape is one of the main component influence the face geometry of beauty, which is consistent to related psychological research and previous works in [21], [22] 2nd PC IV. FBP E VALUATION VIA H AND -C RAFTED F EATURE AND S HALLOW P REDICTOR 4th PC 3rd PC This section, we evaluate the SCUT-FBP5500 using the hand-crafted feature with shallow predictor, while the next section introduce some state-of-the-art deep learning model to achieve FBP. 5th PC A. Geometric Feature with Shallow Predictor We extract a 18-dimensional ratio feature vector from the faces and formulate FBP based on different regression mod(b) PCA Analysis of AM els, such as the linear regression (LR), Gaussian regression ೒ ⬋ᗻϢཇᗻҎ㜌ᔶ⢊ⱘЏ៤ߚব࣪ Fig. 6. PCA analysis of face landmarks of Asian female (AF) and Asian (GR), and support vector regression (SVR). Comparison were performed for Caucasian female/male and Asian female/male male (AM). subsets, and the performance of different model are measured using pearson correlation coefficient (PC) [15], maximum abThe number and portion of the outliers in the four subset are solute error (MAE) and root mean square error (RMSE) after listed in Table II, and the small portion of outlier indicate the 10 folds cross validation. The results are listed in Table IV and reliability of the labeling process for beauty score. Table V, which can be regarded as a baseline for the geometric Since the outlier portion of the beauty score is tiny, the analysis of FBP. mean-4sigma mean mean+4sigma TABLE IV FACIAL BEAUTY PREDICTION USING GEOMETRIC FEATURE WITH SHALLOW MODELS FOR SUBSETS OF DIFFERENT RACES AND GENDER PC MAE RMSE PC MAE RMSE Asian Female LR GR SVR 0.6771 0.7057 0.7008 0.402 0.387 0.3876 0.5246 0.5057 0.5089 Caucasian Female LR GR SVR 0.6809 0.7263 0.7093 0.3986 0.3862 0.4001 0.5239 0.4908 0.5087 Asian Male LR GR SVR 0.6348 0.6923 0.6816 0.3894 0.3572 0.356 0.5085 0.4752 0.4823 Caucasian Male LR GR SVR 0.6063 0.63 0.6397 0.3871 0.3689 0.3617 0.4899 0.4784 0.4739 TABLE V FACIAL BEAUTY PREDICTION USING G EOMETRIC F EATURE WITH TABLE VII C OMPARISON OF A LEX N ET [41], R ES N ET-18 [42] AND R ES N E X T-50 [43] IN MEASUREMENT OF PC, MAE AND RMSE BY 5- FOLD CROSS VALIDATION PC AlexNet ResNet-18 ResNeXt-50 MAE AlexNet ResNet-18 ResNeXt-50 RMSE AlexNet ResNet-18 ResNeXt-50 1 0.8667 0.8847 0.8985 1 0.2633 0.248 0.2306 1 0.3408 0.3258 0.3025 2 0.8645 0.8792 0.8932 2 0.2605 0.2459 0.2285 2 0.3449 0.3286 0.3084 3 0.8615 0.8929 0.9016 3 0.2681 0.243 0.226 3 0.3538 0.3184 0.3016 4 0.8678 0.8932 0.899 4 0.2609 0.2383 0.2349 4 0.3438 0.3107 0.3044 5 0.8566 0.9004 0.9064 5 0.2728 0.2383 0.2258 5 0.3576 0.2994 0.2918 Average 0.8634 0.89 0.8997 Average 0.2651 0.2419 0.2291 Average 0.3481 0.3166 0.3017 SHALLOW MODELS FOR THE WHOLE DATASET Linear Regression Gaussian Regression SVR 0.5948 0.4289 0.5531 0.6738 0.3914 0.5085 0.6668 0.3898 0.5132 PC MAE RMSE TABLE VI FACIAL BEAUTY PREDICTION USING G ABOR FEATURE WITH TWO SAMPLING SCHEME ON WHOLE DATASET PC MAE RMSE 86-keypoints GR SVR 0.7472 0.6691 0.3554 0.3891 0.4599 0.5065 64UniSample GR SVR 0.6764 0.8065 0.4014 0.3976 0.5177 0.5126 Fig. 7. Two sampling schemes to extract the appearance feature of FBP, where the left one is the 86-keypoints method and the right one is the UniSamplePoint method. B. Appearance Feature with Shallow Predictor We extract 40 Gabor feature maps from every original image in five directions and eight angles. Then, we obtain the appearance feature of FBP using two different sampling schemes that extracts some component of the Gabor feature maps as following: • Sample 86-keypoints from each 40 Gabor feature maps to obtain a 3340-dimensional feature vector, as shown in the right sub-figure of Fig. 7; • Use 64UniSample to obtain a 2560-dimensional feature vector, as shown in the left sub-figure of Fig. 7. Finally, we use PCA to reduce the extracted feature dimension before we train the predictor. The results of the appearancebased shallow predictors for all the data are shown in Table VI. TABLE VIII C OMPARISON OF A LEX N ET [41], R ES N ET-18 [42] AND R ES N E X T-50 [43] IN MEASUREMENT OF PC, MAE AND RMSE BY 60% TRAINING AND 40% TESTING PC MAE RMSE AlexNet ResNet-18 ResNeXt-50 0.8298 0.2938 0.3819 0.8513 0.2818 0.3703 0.8777 0.2518 0.3325 V. FBP E VALUATION VIA D EEP P REDICTOR We evaluate three recently proposed CNN models with different structures for FBP, including AlexNet [41], ResNet18 [42] and ResNeXt-50 [43]. All these CNN models are trained by initializing weights using networks pre-trained on the ImageNet dataset. The evaluation of were performed under two different experiment settings as following: 1) The models were trained and tested using 5-fold cross validation, which means each fold containing 20% samples (1100 images). The accuracy of each fold and the average of all the 5 fold are shown in Table VII. 2) The models were trained using 60% samples (3300 images), and tested with the rest 40% (2200 images). The results are shown in Table VIII. The results illustrates that the deepest CNN-based ResNeXt50 model [43] obtains the best performance comparing to the ResNet-18 and AlexNet in both the experiment setting. It can be observed that all the deep CNN model are superior to the shallow predictor with hand-crafted geometric feature in Table V or appearance feature in Table VI. It indicates the effectiveness and powerfulness of the end-to-end feature learning deep model for FBP. Comparing the results of Table VII and Table VIII, we also find that the accuracy of all the 5-fold cross validation is slightly higher than the results of the split of 60% training and 40% testing. One of the reasons may be due to the amounts and diversity of the training samples, since the 5-fold cross validation use 80% samples to train the models. This observation indicates that the data augmentation techniques may further improve the performance of the deep FBP model, which merits exploring in the future. VI. C ONCLUSION In this paper, we introduce a new diverse benchmark dataset, called SCUT-FBP5500, to achieve multi-paradigm facial beauty prediction. The SCUT-FBP5500 dataset has totally 5500 frontal faces with diverse properties (male/female, Asian/Caucasian, ages) and diverse labels (face landmarks, beauty scores within [1, 5], beauty score distribution). Benchmark analysis have been made for the beauty scores and landmarks in SCUT-FBP5500, and the visualization of the data shows the statistical properties of the dataset. Since the SCUTFBP5500 is designed for multi-paradigm, it can be adapted to different FBP models for different tasks, like appearancebased or shape-based model for facial beauty classification/regression/ranking. We evaluated the SCUT-FBP5500 using different combinations of feature and predictor, and deep learning models, where the results indicates the reliability of the dataset. R EFERENCES [1] D. Xie, L. Liang, L. Jin, J. Xu and M. Li, “SCUT-FBP: A Benchmark Dataset for Facial Beauty Perception,” in Proc. of IEEE SMC, pp. 18211826, 2015. [2] L. Liang, D. Xie, L. Jin, J. Xu, M. Li, and L. Lin, “Region-aware scattering convolution networks for facial beauty prediction,” in Proc. of IEEE ICIP, pp. 2861-2865, 2017. [3] J. Xu, L. Jin, L. Liang, Z. Feng, D. Xie, and H. Mao, “Facial attractiveness prediction using psychologically inspired convolutional neural network (PI-CNN),” in Proc. of IEEE ICASSP, pp. 1657-1661, 2017. [4] L. Liang, L. Jin, and D. Liu, “Edge-aware label propagation for mobile facial enhancement on the cloud,” IEEE Trans. on Circuits and Systems for Video Technology, vol. 27, no. 1, pp. 125-138, 2017. [5] L. Liang, L. Jin, and X. Li, “Facial skin beautification using adaptive region-aware mask,” IEEE Trans. on Cybernetics, vol. 44, no. 12, pp. 2600-2612, 2014. [6] J. Li, X. Chao, L. Liu, X. Shu, and S. Yan, “Deep face beautification,” Proc. ACM International Conference on Multimedia, pp. 793-794, 2015. [7] T. Leyvand, D. Cohen-Or, G. Dror, and D. Lischinski, “Data-driven enhancement of facial attractiveness,” ACM Trans. Graph., pp. 28:1-10, 2008. [8] H. Gunes, “A survey of perception and computation of human beauty,” in Proc. of J-HGBU, pp. 19-24, 2011. [9] A. Laurentini and A. Bottino, “Computer analysis of face beauty: A survey,” Comput. Vision and Image Underst., vol. 125, pp. 184–199, 2014. [10] D. Zhang, F. Chen, and Y. Xu, Computer models for facial beauty analysis. Springer International Publishing Switzerland, 2016. [11] L. Liu, J. Xing, S. Liu, H. Xu, X. Zhou, and S. Yan, “Wow! you are so beautiful today!,” ACM Transactions on Multimedia Computing, Communications, and Applications, vol. 11, no. 1s, p. 20, 2014. [12] K. Scherbaum, T. Ritschel, M. Hullin, T. Thormählen, V. Blanz and H. Seidel, “Computer-suggested facial makeup,” Comput. Graph. Forum, vol. 30, no. 2, pp. 485-492, 2011. [13] Y. Mu, “Computational facial attractiveness prediction by aestheticsaware features,” Neurocomputing, vol. 99, pp. 59-64, 2013. [14] Y. LeCun, Y. Bengio and G. Hinton, “Deep learning,” Nature, vol. 251, pp. 436-444, 2015. [15] Y. Eisenthal, G. Dror and E. Ruppin, “Facial Attractiveness: Beauty and the Machine,” Neural Computation, vol. 18, pp. 119-142, 2006. [16] N. Murray, L. Marchesotti L, and F. Perronnin, “AVA: A large-scale database for aesthetic visual analysis,” in Proc. of CVPR, pp. 2408-2415, 2012. [17] A. Laurentini and A. Bottino, “Computer analysis of face beauty: A survey,” Computer Vision and Image Understanding, vol. 125, pp. 184199, 2014. [18] I. Stephen, M. Law Smith, M. Stirrat, and D. Perrett, “Facial skin coloration affects perceived health of human faces,” International Journal of Primatology, vol. 30, no. 6, pp. 845-857, 2009. [19] A. Kagian, G. Dror, T. Leyvand, D. Cohen-Or, and E. Ruppin, “A humanlike predictor of facial attractiveness,” in Proc. of NIPS, pp. 649656, 2006. [20] H. Mao, L. Jin and M. Du, “Automatic classification of Chinese female facial beauty using Support Vector Machine,” in Proc. of IEEE SMC, pp. 4842-4846, 2009. [21] D. Zhang, Q. Zhao and F. Chen, “Quantitative analysis of human facial beauty using geometric features,” Pattern Recognition, vol. 44, no. 4, pp. 940-950, 2011. [22] F. Chen and D. Zhang, “A benchmark for geometric facial beauty study,” Medical Biometrics, pp. 21-32, 2010. [23] J. Fan, K. P. Chau, X. Wan, L. Zhai and E. Lau, “Prediction of facial attractiveness from facial proportions,” Pattern Recognition, vol. 45, pp. 2326-2334, 2012. [24] H. Yan, “Cost-sensitive ordinal regression for fully automatic facial beauty assessment,” Neurocomputing, no. 129, pp. 334-342, 2014. [25] H. Altwaijry and S. Belongie, “Relative ranking of facial attractiveness,” in IEEE Workshop on WACV, pp. 117-124, 2013. [26] Y. Chen, H. Mao, L. Jin, “A novel method for evaluating facial attractiveness”, in IEEE Proc. ICALIP, pp. 1382-1386, 2010. [27] W. Chiang, H. Lin, C. Huang, L. Lo, and S. Wan, “The cluster assessment of facial attractiveness using fuzzy neural network classifier based on 3D Moir features,” Pattern Recognition, vol. 47, no. 3, pp. 1249-1260, 2014. [28] S. Kalayci, H. K. Ekenel, and H. Gunes, “Automatic analysis of facial attractiveness from video,” in Proc. of ICIP, pp. 4191-4195, 2014. [29] J. Gan, L. Li, Y. Zhai and Y. Liu, “Deep self-taught learning for facial beauty prediction,” Neurocomputing, no. 144, pp. 295-303, 2014. [30] S. Wang, M. Shao and Y. Fu, “Attractive or not? Beauty prediction with attractiveness-aware encoders and robust late fusion,” ACM Multimedia, pp. 805-808, 2014. [31] Y. Ren and X. Geng, “Sense beauty by label distribution learning,” in Proc, IJCAI, pp. 2648–2654, 2017. [32] R. White, A. Eden, and M. Maire, “Automatic prediction of human attractiveness,” UC Berkeley CS280A Proj., vol. 1, p. 2, 2004. [33] D. Gray, K. Yu, W. Xu and Y. Gong, “Predicting facial beauty without landmarks,” in Proc. ECCV, 2010. [34] M. Redi, N. Rasiwasia, G. Aggarwal and A. Jaimes, “The Beauty of Capturing Faces: Rating the Quality of Digital Portraits,” in IEEE Workshops on AFGR, pp. 1-8, 2015. [35] H. Gunes and M. Piccardi, “Assessing facial beauty through proportion analysis by image processing and supervised learning,” International Journal of Human-Computer Studies, vol. 64, no. 12, pp. 1184-1199, 2006. [36] J. Whitehill and J.R. Movellan, “Personalized facial attractiveness prediction,” in Proc. AFGR, 2008. [37] Y. Fan, S. Liu, B. Li, Y. Fan, Z. Guo, A. Samal, J. Wan and Stan Z. Li “Label Distribution based facial attractiveness computation by deep residual learning”, IEEE Transactions on Multimedia, 2017 (Online). [38] B. Davis and S. Lazebnik, “Analysis of human attractiveness using manifold kernel regression,” in Proc. ICIP, 2008. [39] “DataTang,” url: http://datatang.com/. [40] W.A. Bainbridge, P. Isola and A. Oliva, “The intrinsic memorability of face images.” Journal of Experimental Psychology: General, vol. 142, no. 4, pp. 1323-1334, 2013. [41] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Proc, NIPS, 2012, pp. 1097– 1105. [42] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc, CVPR, 2016, pp. 770–778. [43] S. Xie, R. Girshick, P. Dollár, Z. Tu, and K. He, “Aggregated residual transformations for deep neural networks,” arXiv preprint arXiv:1611.05431, 2016.	1cs.CV
1 Degrees of Freedom of the Broadcast Channel with Hybrid CSI at Transmitter and Receivers arXiv:1709.02884v1 [cs.IT] 9 Sep 2017 Mohamed Fadel, Student Member, IEEE, and Aria Nosratinia, Fellow, IEEE Abstract In general the different links of a broadcast channel may experience different fading dynamics and, potentially, unequal or hybrid channel state information (CSI) conditions. The faster the fading and the shorter the fading block length, the more often the link needs to be trained and estimated at the receiver, and the more likely that CSI is stale or unavailable at the transmitter. Disparity of link fading dynamics in the presence of CSI limitations can be modeled by a multi-user broadcast channel with both non-identical link fading block lengths as well as dissimilar link CSIR/CSIT conditions. This paper investigates a MISO broadcast channel where some receivers experience longer coherence intervals (static receivers) and have CSIR, while some other receivers experience shorter coherence intervals (dynamic receivers) and do not enjoy free CSIR. We consider a variety of CSIT conditions for the above mentioned model, including no CSIT, delayed CSIT, or hybrid CSIT. To investigate the degrees of freedom region, we employ interference alignment and beamforming along with a product superposition that allows simultaneous but non-contaminating transmission of pilots and data to different receivers. Outer bounds employ the extremal entropy inequality as well as a bounding of the performance of a discrete memoryless multiuser multilevel broadcast channel. For several cases, inner and outer bounds are established that either partially meet, or the gap diminishes with increasing coherence times. Index Terms Broadcast channel, Channel state information, Coherence time, Coherence diversity, Degrees of freedom, Fading channel, Multilevel broadcast channel, Product superposition. The authors are with the Department of Electrical Engineering, University of Texas at Dallas, Richardson, TX 750830688 USA, E-mail: mohamed.fadel@utdallas.edu;aria@utdallas.edu. This work was presented in part at the IEEE International Symposium on Information Theory (ISIT), Germany, June 2017 [1]. March 26, 2018 DRAFT 2 I. I NTRODUCTION The performance of a broadcast channel depends on both the channel dynamics as well as the availability and the quality of channel state information (CSI) on the two ends of each link [2]–[5]. The two issues of CSI and the channel dynamics are practically related. The faster the fading, the more often the channel needs training, thus consuming more channel resources, while a very slow fading link requires infrequent training, therefore slow fading models often assume that CSIR is available due to the cost of training being small when amortized over time. In practice, in a broadcast channel some links may fade faster or slower than others. Recently, it has been shown [6], [7], that the degrees of freedom of the broadcast channel are affected by the disparity of link fading speeds, but existing studies have focused on a few simple and uniform CSI conditions, e.g., neither CSIT nor CSIR were available in [6], [7] for any user. This paper studies a broadcast channel where the links experience both disparate fading conditions as well as non-uniform or hybrid CSI conditions. A review of the relevant literature is as follows. Under perfect instantaneous CSI, the degrees of freedom of a broadcast channel increase with the minimum of the transmit antennas and the total number of receivers antennas [8], [9]. However, due to the time-varying nature of the channel and feedback impairments, perfect instantaneous transmit-side CSI (CSIT) may not be available, and also receive-side CSI (CSIR) can be assumed for slow-fading channels only. Broadcast channel with perfect CSIR has been investigated under a variety of CSIT conditions, including imperfect, delayed, or no CSIT [2]–[4], [10]–[12]. In the absence of CSIT, Huang et al. [2] and Vaze and Varanasi [3] showed that the degrees of freedom collapse to that of the single-receiver, since the receivers are stochastically equivalent with respect to the transmitter. For a MISO broadcast channel Lapidoth et al. [4] conjectured that as long as the precision of CSIT is finite, the degrees of freedom collapse to unity. This conjecture was recently settled in the positive by Davoodi and Jafar in [10]. Moreover, for a MISO broadcast channel under perfect delayed CSIT Maddah-Ali and Tse in [11] showed using retrospective interference alignment that the degrees of freedom are 1 1 1+ 12 +...+ K > 1, where K is the number of the transmit antennas and also the number of receivers. A scenario of mixed CSIT was investigated in [12], where the transmitter has partial knowledge on the current channel in addition to delayed CSI. The potential variation between the quality of feedback links has led to the model of hybrid March 26, 2018 DRAFT 3 CSIT, where the CSIT with respect to different links may not be identical [10], [13]–[15]. A MISO broadcast channel with perfect CSIT for some receivers and delayed for the others was studied by Tandon et al. in [13] and Amuru et al. in [14]. Davoodi and Jafar in [10] showed that for a MISO two-receiver broadcast channel under perfect CSIT for one user and no CSIT for the other, the degrees of freedom collapse to unity. Tandon et al. in [15] considered a MISO broadcast channel with alternating hybrid CSIT to be perfect, delayed, or no CSIT with respect to different receivers. As mentioned earlier, investigation of broadcast channels under unequal link fading dynamics is fairly recent. An achievable degrees of freedom region for one static and one dynamic receiver was given in [16]–[19] via product superposition, producing a gain that is now known as coherence diversity. Coherence diversity gain was further investigated in [6], [7] for a K-receiver broadcast channel with neither CSIT nor CSIR. Also, a broadcast channel was investigated in [20], where the receivers MIMO fading links experience nonidentical spacial correlation. In this paper, we consider a multiuser model under a hybrid CSIR scenario where a group of receivers, denoted static receivers, are assumed to have CSIR, and another group with shorter link coherence time, denoted dynamic receivers, do not have free CSIR. We consider this model under a variety of CSIT conditions, including no CSIT, delayed CSIT, and two hybrid CSIT scenarios. In each of these conditions, we analyze the degrees of freedom region. A few new tools are introduced, and inner and outer bounds are derived that partially meet in some cases. The results of this paper are cataloged as follows. In the absence of CSIT, an outer bound on the degrees of freedom region is produced via bounding the rates of a discrete memoryless multilevel broadcast channel [21], [22] and then applying the extremal entropy inequality [23], [24]. Our achievable degrees of freedom region meets the outer bound in the limiting case where the coherence times of the static and dynamic receivers are the same. For delayed CSIT, we use the outdated CSI model that was used by Maddah-Ali and Tse [11] under i.i.d. fading and assuming global CSIR at all nodes. Noting that our model does not have uniform CSIR, we produced a technique with alignment over super-symbols to utilize outdated CSIT but merge it together with product superposition to reuse the pilots of the dynamic receivers for the purpose of transmission to static receivers. Moreover, we develop an outer bound that is suitable for block-fading channels with different coherence times, by appropriately enhancing March 26, 2018 DRAFT 4 static Tx dynamic Fig. 1. A broadcast channel with multiple static and multiple dynamic users the channel to a physically-degraded broadcast channel and then applying the extremal entropy inequality [23], [24]. For one static and one dynamic receiver, our achievable degrees of freedom partially meet our outer bound, and furthermore the gap decreases with the dynamic receiver coherence time T . Under hybrid CSIT, we analyze two conditions: First, we consider perfect CSIT for the static receivers and no CSIT with respect to the dynamic receivers. The achievable degrees of freedom in this case are obtained using product superposition with the dynamic receiver’s pilots reused and beamforming for the static receivers to avoid interference. Second, we consider perfect CSIT with respect to the static receivers and delayed CSIT with respect to the dynamic receivers. An achievable transmission scheme is proposed via a combination of beamforming, interference alignment, and product superposition methodologies. The outer bounds for the two hybrid-CSIT cases were based on constructing an enhanced physically degraded channel and then applying the extremal entropy inequality. For one static receiver with perfect CSIT and one dynamic receiver with delayed CSIT, the gap between the achievable and the outer sum degrees of freedom is 1 . T II. S YSTEM M ODEL Consider a broadcast channel with multiple single-antenna receivers and the transmitter is equipped with Nt antennas. The expressions “receiver” and “user” are employed without distinction throughout the paper, indicating the receiving terminals in the broadcast channel. The channels of the users are modeled as Rayleigh block fading where the channel coefficients remain constant over each block and change independently across blocks [25], [26]. As shown in Fig. 1, the users are partitioned into two sets based on channel availability and the length of the coherence interval: one set of dynamic users and another set of static users. The former contains March 26, 2018 DRAFT 5 TABLE I N OTATION Static Users Dynamic Users m′ m MISO channel gains g1 , . . . , gm′ h1 , . . . , hm received signals (continuous) ′ y1′ , . . . , ym ′ y1 , . . . , ym DMC receive variables ′ Y1′ , . . . , Ym ′ Y1 , . . . , Ym transmission rates ′ R1′ , . . . , Rm ′ R1 . . . , Rm ′ M1′ , . . . , Mm ′ M1 , . . . , Mm d′1 , . . . , d′m′ d1 , . . . , dm number of users messages degrees of freedom coherence time T ′ T General Variables X transmit signal ρ signal-to-noise ratio Ui , Vj , W auxiliary random variables H set of all channel gains Dx vertex of degrees of freedom region ei canonical coordinate vector m dynamic users having coherence time T and no free CSIR1 , and the latter contains m′ static users having coherence time T ′ and perfect instantaneous CSIR. We consider the transmitter is equipped with more antennas than the number of dynamic and static users, i.e., Nt ≥ m′ + m. The received signals yj′ (t), yi (t) at the static user j, and the dynamic user i, respectively, at time instant t are yj′ (t) = gj† (t)x(t) + zj′ (t), j = 1, . . . , m′ , yi (t) = h†i (t)x(t) + zi (t), i = 1, . . . , m, (1) where x(t) ∈ CNt is the transmitted signal, zj′ (t), zi (t) denote the corresponding additive i.i.d. Gaussian noise of the users, and gj (t) ∈ CNt , hi (t) ∈ CNt denote the channels of the static user j and the dynamic user i whose coefficients stay the same over T ′ and T time instances, 1 This means that the cost of knowing CSI at the receiver, e.g., by channel estimation, is not ignored. March 26, 2018 DRAFT 6 respectively. The distributions of gj and hi are globally known at the transmitter and at the users.2 Having CSIR, the value of gj (t) is available instantaneously and perfectly at the static user j. Furthermore, the static user j obtains an outdated version of the dynamic users channels hi , and also the dynamic user i obtains an outdated version of the static users channel gi (completely stale) [11]. CSIT for each user can take one of the following: • Perfect CSIT: the channel vectors, gj (t), hi (t), are available at the transmitter instantaneously and perfectly. • Delayed CSIT: the channel vectors, gj (t), hi (t), are available at the transmitter after they change independently in the following block (completely stale [11]). • No CSIT: the channel vectors, gj (t), hi (t), cannot not be known at the transmitter. We consider the broadcast channel with private messages for all users and no common ′ messages. More specifically, we assume that the independent messages Mj′ ∈ [1 : 2nRi (ρ) ], Mi ∈ [1 : 2nRi (ρ) ] associated with rates Rj′ (ρ), Ri (ρ) are communicated from the transmitter to the static user j and dynamic user i, respectively, at ρ signal-to-noise ratio. The degrees of freedom of the static and dynamic users achieving rates Rj′ (ρ), Ri (ρ) can be defined as Rj′ (ρ) , ρ→∞ log(ρ) Ri (ρ) di = lim , ρ→∞ log(ρ) j = 1, . . . , m′ , d′j = lim i = 1, . . . , m. (2) The degrees of freedom region is defined as ′ +m ′ ∃(R1′ (ρ), . . . , Rm D = (d′1 , . . . , d′m′ , d1 , . . . , dm ) ∈ Rm ′ (ρ), R1 (ρ), . . . , Rm (ρ)) ∈ C(ρ), + Rj′ (ρ) Ri (ρ) , di = lim , j = 1, . . . , m′ , i = 1, . . . , m , ρ→∞ log(ρ) ρ→∞ log(ρ) d′j = lim (3) where C(ρ) is the capacity region at ρ signal-to-noise ratio. The sum degrees of freedom is defined as Csum (ρ) , ρ→∞ log(ρ) dsum = lim where ′ Csum (ρ) = max m X j=1 2 Rj′ (ρ) + m X (4) Ri (ρ). (5) i=1 Also, the coherence times of all channels are globally known at the transmitter and at the users. March 26, 2018 DRAFT 7 Y1’ X Y2’ p(y3’\|y2’) p(ym’’\|ym’-1’) p(y3\|y2) p(ym\|ym-1) Ym’’ p(y1’,y1\|x) Y1 Fig. 2. p(y2’\|y1’) p(y2\|y1) Y2 Ym Discrete memoryless multiuser multilevel broadcast channel In the sequel, we study the degrees of freedom of the above MISO broadcast channel under different CSIT scenarios that could be perfect, delayed or no CSIT. III. N O CSIT FOR A LL U SERS In this section, we study the broadcast channel defined in Section II when there is no CSIT for all users. In particular, we give outer and achievable degrees of freedom regions in Section III-B and Section III-C, respectively. The outer degrees of freedom region is based on the construction of an outer bound on the rates of a multiuser multilevel discrete memoryless channel that is given in Section III-A. A. Multiuser Multilevel Broadcast Channel The multilevel broadcast channel was introduced by Borade et al. [21] as a three-user broadcast discrete memoryless broadcast channel where two of the users are degraded with respect to each other. The capacity of this channel under degraded message sets was established by Nair and El Gamal [22]. Here, we study a multiuser multilevel broadcast channel with two sets of degraded users (see Fig. 2). One set contains m′ users with Yj′ received signal at user j, and the other set contains m users with Yi received signal at user i. Therefore, X →Y1′ → Y2′ → · · · → Ym′ ′ X →Y1 → Y2 → · · · → Ym March 26, 2018 (6) DRAFT 8 form two Markov chains. We consider a broadcast channel with (m′ + m) private messages and no common message. An outer bound for the above multilevel broadcast channel is given in the following theorem. Theorem 1: The rate region of the multilevel broadcast channel with two sets of degraded users (Eq. (6)) is outer bounded by the intersection of R1 ≤I(Um′ , W ; Y1\|V1 ) − I(W ; Ym′ ′ \|Um′ ), (7) Ri ≤I(Vi−1 ; Yi \|Vi ), (8) Rj′ ≤I(Uj−1 ; Yj′ \|Uj ), i = 2, . . . , m, j = 1, . . . , m′ − 1, ′ ′ ′ ′ Rm ′ ≤I(W ; Ym′ \|Um′ ) + I(X; Ym′ \|Um′ , W ) − I(X; Ym′ \|Um′ −1 ), (9) (10) and Ri ≤I(Ũi−1 ; Yi \|Ũi ), i = 1, . . . , m − 1, (11) Rm ≤I(W̃ ; Ym \|Ũm ) + I(X; Ym\|Ũm , W̃ ) − I(X; Ym\|Ũm−1 ), (12) R1′ ≤I(Ũm , W̃ ; Y1′ \|Ṽ1 ) − I(W̃ ; Ym \|Ũm ), (13) Rj′ ≤I(Ṽj−1; Yj′ \|Ṽj ), (14) j = 2, . . . , m′ , for some pmf p(u1 , . . . , um′ , ũ1 , . . . , ũm , v1 , . . . , vm , ṽ1 , . . . , ṽm′ , w, w̃, x), (15) where Um′ → · · · → U1 →X → (Y1 , . . . , Ym, Y1′ , . . . , Ym′ ′ ) Vm → · · · → V1 → (W, Um′ ) →X → (Y1 , . . . , Ym, Y1′ , . . . Ym′ ′ ) Ũm → · · · → Ũ1 →X → (Y1 , . . . Ym , Y1′ , . . . , Ym′ ′ ) Ṽm′ → · · · → Ṽ1 → (W̃ , Ũm ) →X → (Y1 , . . . Ym , Y1′ , . . . , Ym′ ′ ) (16) forms Markov chains and U0 = Ũ0 , X. Proof: See Appendix I. Remark 1: Theorem 1 is an extension of the Körner-Marton outer bound [27, Theorem 5] to more than two users, and it recovers the Körner-Marton bound when m = m′ = 1. March 26, 2018 DRAFT 9 Remark 2: For the multiuser multilevel broadcast channel characterized by (6), we establish the capacity for degraded message sets in Appendix II, where one common message is communicated to all receivers and one further private message is communicated to one receiver. B. Outer Degrees of Freedom Region In the sequel, we give an outer bound on the degrees of freedom of the broadcast channel defined in Section II when there is no CSIT for all users. The outer bound development depends on the results of Theorem 1 in Section III-A. Theorem 2: An outer bound on the degrees of freedom region of the fading broadcast channel characterized by Eq. (1) without CSIT is, ′ m X d′j ≤ 1, m X di ≤ 1 − (17) j=1 i=1 m′ X j=1 d′j + m X i=1 di ≤   1  4 3 1 , T T = T ′ , ∆T = 0 (18) (19) otherwise, where ∆T is the offset between the two coherence intervals. Proof: Equations (17) and (18) are outer bounds for a broadcast channel whose users are either all homogeneously static or all homogeneously dynamic [7], [18]. The remainder of the proof is dedicated to establishing (19). We enhance the channel by giving all users global CSIR. When T ′ = T and ∆T = 0, (19) follows directly from [7], [18]. When T ′ 6= T or ∆T 6= 0, having no CSIT, the channel belongs to the class of multiuser multilevel broadcast channels in Section III-A. We then use the two outer bounds developed for the multilevel broadcast channels to generate two degrees of freedom bounds, and merge them to get the desired result. We begin with the outer bound described in (7)-(10); we combine these equations to obtain P P partial sum-rate bounds on the static ( Rj′ ) and dynamic ( Ri ) receivers: ′ m X Rj′ j=1 ≤ ′ −1 m X ′ ′ I(Uj−1 ; yj′ \|Uj , H) + I(W ; ym ′ \|Um′ , H) + I(x; ym′ \|Um′ , W, H) j=1 ′ − I(x; ym ′ \|Um′ −1 , H) March 26, 2018 DRAFT 10 = ′ −1 m X ′ ′ h(yj′ \|Uj , H) − h(yj′ \|Uj−1 , H) + I(W ; ym ′ \|Um′ , H) + h(ym′ \|Um′ , W, H) j=1 ′ − h(ym ′ \|Um′ −1 , H) + o(log(ρ)) (20) ′ ′ =I(W ; ym ′ \|Um′ , H) + h(ym′ \|Um′ , W, H) + o(log(ρ)), (21) where H is the set of all channel vectors, (20) follows from the chain rule, h(yj′ \|x, H) = o(log(ρ)), and (21) follows since the received signals of all static users, yj′ , have the same statistics [7], [18]. Also, using Theorem 1, m X Rj ≤I(Um′ , W ; y1\|V1 , H) − ′ I(W ; ym ′ \|Um′ , H) + m X I(Vj−1; yj \|Vj , H) j=2 j=1 =h(y1 \|V1 , H) − h(y1 \|Um′ , W, H) − ′ I(W ; ym ′ \|Um′ , H) + m X h(yj \|Vj , H) j=2 − h(yj \|Vj−1, H) (22) ′ = − h(y1 \|Um′ , W, H) − I(W ; ym ′ \|Um′ , H) + h(ym \|Vm , H) + o(log(ρ)) (23) ′ ≤ − h(y1 \|Um′ , W, H) − I(W ; ym ′ \|Um′ , H) + log(ρ) + o(log(ρ)), (24) where (22) follows from the chain rule, (23) follows since yj have the same statistics, and (24) ′ follows since h(ym \|Vm , H) ≤ n log(ρ) + o(log(ρ)). Define Yj,k to be the received signal of user j at time instance k. From (21) and (24), we can obtain the bound (27) on the rates. m′ m 1X ′ X 1 1 ′ ′ Rj + Rj ≤ I(W ; ym h(ym ′ \|Um′ , H) + ′ \|Um′ , W, H) − h(y1 \|Um′ , W, H) 2 j=1 2 2 j=1 ′ − I(W ; ym ′ \|Um′ , H) + log(ρ) + o(log(ρ)) 1 ′ = h(ym ′ \|Um′ , W, H) − h(y1 \|Um′ , W, H) + log(ρ) + o(log(ρ)) 2 1 ′ ≤ h(ym ′ , y1 \|Um′ , W, H) − h(y1 \|Um′ , W, H) + log(ρ) + o(log(ρ)) 2 n X 1 ′ ′ ′ ≤ h(ym ′ ,k , y1,k \|Um′ , W, H, ym′ ,1 , . . . , ym′ ,k−1 , y1,1 , . . . , y1,k−1 ) 2 k=1 (25) ′ ′ − h(y1,k \|Um′ , W, H, ym ′ ,1 , . . . , ym′ ,k−1 , y1,1 , . . . , y1,k−1 ) + log(ρ) + o(log(ρ)) ≤ March 26, 2018 max Tr{Σx }≤ρ,Σx <0 (26) EH 1 log \|I + HΣx H† \| − log(1 + h†1 Σx h1 ) + log(ρ) 2 DRAFT 11 + o(log(ρ)), (27) where (25) and (26) follow from the chain rule and that conditioning does not increase differential entropy, and (27) follows from extremal entropy inequality [23], [24], [28]. In order to bound (27), we use a specialization of [29, Lemma 3] as follows. Lemma 1: Consider two random matrices H1 ∈ CN1 ×Nt and H2 ∈ CN2 ×Nt , where N1 ≥ N2 . For a covariance matrix Σx , where Tr{Σx } ≤ ρ, we have 1 1 log \|I + H1 Σx H†1 \| − log \|I + H2 Σx H†2 \| ≤ o(log(ρ)). (28) Σx min{Nt , N1 } min{Nt , N2 } The proof of Lemma 1 is omitted as it directly follows from [29, Lemma 3]. Lemma 1 yields max the following outer bound on the degrees of freedom: m′ m 1X ′ X d + di ≤ 1. 2 j=1 j i=1 (29) We now repeat the exercise of bounding the sum rates and deriving degrees of freedom, this time starting from (11)-(14). By following bounding steps parallel to (21), (24), (27), ′ m X m d′j j=1 1X di ≤ 1. + 2 i=1 (30) Adding (29) and (30) yields the outer bound (19), completing the proof of Theorem 2. C. Achievable Degrees of Freedom Region Theorem 3: The fading broadcast channel described by Eq. (1) can achieve the following degrees of freedom without CSIT: m X di ≤ 1 − m X di ≤ 1. i=1 m′ X j=1 d′j + 1 , T (31) (32) i=1 Proof: The achievable scheme uses product superposition [17], [18], where the transmitter uses one antenna to send the super symbol to two users: one dynamic and one static, x† = xs x†d , (33) where xs ∈ C is a symbol intended for the static user, x†d = [xτ , x†δ ] March 26, 2018 (34) DRAFT 12 where xτ ∈ C is a pilot and xδ ∈ CT −1 is a super symbol intended for the dynamic user. Since degrees of freedom analysis is insensitive to the additive noise, we omit the noise component in the following. y† = hxs [xτ , x†δ ] = [hxτ , hx†δ ], (35) where h = hxs . The dynamic user estimates the equivalent channel h during the first time instance and then decodes xδ coherently based on the channel estimate. The static receiver only utilizes the received signal during the first time instance: y1′ = gxs . (36) Knowing its channel gain g, the static receiver can decode xs . The achievable degrees of freedom of the two users are, (d′ , d) = 1 1 . ,1 − T T (37) We now proceed to prove that the degrees of freedom region characterized by (31) and (32) can be achieved via a combination of two-user product superposition strategies that were outlined above, and single-user strategies. For clarity of exposition we refer to (31), which describes the degrees of freedom constraints of the dynamic receivers, as the non-coherent bound, and to (32) as the coherent bound. The non-negativity of degrees of freedom restricts them to the ′ m+m . The intersection of the coherent bound and the non-negative orthant non-negative orthant R+ is a (m′ + m)–simplex that has m + m′ + 1 vertices. The non-coherent bound is a hyperplane that partitions the simplex with m′ + 1 vertices on one side of the non-coherent bound and m on the other. Therefore the intersection of the simplex with the non-coherent bound produces a polytope with (m′ + 1)(m + 1) vertices.3 For illustration, see Fig. 3 showing the three-user degrees of freedom with two static users and Fig. 4 with one static user. We now verify that each of the (m′ + 1)(m + 1) vertices can be achieved with either a single-user strategy, or via a two-user product superposition strategy: • m′ vertices corresponding to single-user transmission to each static user j achieving one degree of freedom. 3 This can be verified with a simple counting exercise involving the number of edges of the simplex that cross the non-coherent bound. March 26, 2018 DRAFT 13 d d1 Non-coherent bound Non-coherent bound d2 Coherent bound d’2 d’ d’1 Fig. 3. Achievable degrees of freedom region of one dynamic and two static users • Coherent bound Fig. 4. Achievable degrees of freedom region of one static and two dynamic users m vertices corresponding to single-user transmission to each dynamic user i achieving (1 − T1 ) degrees of freedom. • m′ m vertices corresponding to product superposition applied to all possible pairs of static and dynamic users, achieving 1 T degrees of freedom for one static user and (1 − T1 ) degrees of freedom for one dynamic user. • One trivial vertex at the origin, corresponding to no transmission achieving zero degrees of freedom for all users. Hence, the number of the vertices is m′ + m + m′ m + 1 = (m + 1)(m′ + 1). This completes the achievability proof of Theorem 3. Remark 3: When the static and dynamic users have the same coherence time, the inner and outer bounds on degrees of freedom coincide. In this case it is degrees of freedom optimal to serve two-users at a time (one dynamic and one static). IV. D ELAYED CSIT FOR ALL USERS Under delayed CSIT, the transmitter knows each channel gain only after it is no longer valid. This condition is also known as outdated CSIT. We begin by proving inner and outer bounds March 26, 2018 DRAFT 14 when transmitting only to static users, only to dynamic users, and to one static and one dynamic user. We then synthesize this collection of bounds into an overall degrees of freedom region. A. Transmission to Static Users Theorem 4: The degrees of freedom region of the fading broadcast channel characterized by Eq. (1) with delayed CSIT and having m′ static users and no dynamic users is 1 j = 1, . . . , m′ . (38) 1 , 1 + + . . . + m′ Proof: The special case of fast fading (T ′ = 1) was discussed by Maddah-Ali and Tse d′j ≤ 1 2 in [11], where the achievability was established by retrospective interference alignment that aligns the interference using the outdated CSIT, and the converse was proved by generating an improved channel without CSIT having a tight degrees of freedom region against TDMA according to the results in [2], [3]. For T ′ ≥ 1, the achievability is established by employing retrospective interference alignment presented in [11] over super symbols each of length T ′ . The converse is proved by following the same procedures in [11] to generate a block-fading improved channel without CSIT and with identical coherence intervals of length T ′ . According to the results of [7], [18], TDMA is tight against the degrees of freedom region of the improved channel. B. Transmission to Dynamic Users Theorem 5: The fading broadcast channel characterized by Eq. (1), with delayed CSIT and having m dynamic users and no static users, can achieve the degrees of freedom di ≤ 1 1 + + ...+ 1 2 1 (1 m − m ), T i = 1, . . . , m. (39) An outer bound on the degrees of freedom region is di ≤1 − m X 1 , T (40) m . (41) 1 + + . . . + m1 i=1 Proof: The achievability part can be proved as follows. At the beginning of each super di ≤ 1 2 symbol, m pilots are sent for channel estimation. Then retrospective interference alignment in [11] over super symbols is employed during the remaining (T − m) instances, to achieve (39). March 26, 2018 DRAFT 15 For the converse part, (41) is proved by giving the users global CSIR, and then applying Theorem 4. Moreover, (40) is the single-user bound for each dynamic user that can be proved as follows. For a single user with delayed CSIT, feedback does not increase the capacity [30], and consequently the assumption of delayed CSIT can be removed. Hence, the single-user bound for each dynamic user with delayed CSIT is the same as the single-user bound without CSIT [26]. C. Transmission to One Static and One Dynamic User Theorem 6: The fading broadcast channel characterized by Eq. (1), with delayed CSIT and having one static and one dynamic user, can achieve the following degrees of freedom 2 1 2 2 (1 + ), (1 − ) , 3 T 3 T 1 1 D2 : (d′ , d) = ( , 1 − ). T T D1 : (d′ , d) = (42) (43) Furthermore, the achievable degrees of freedom region is the convex hull of the above degrees of freedom pairs. Proof: From Section III-C, product superposition achieves the pair (43) that does not require CSIT for any of the two users. The remainder of the proof is dedicated to the achievability of the pair (42). We provide a transmission scheme based on retrospective interference alignment [11] along with product superposition. 1) The transmitter first emits a super-symbol intended for the static user: X1 = [X1,1 , · · · , X1,ℓ ], where ℓ = T′ , T (44) and each X1,n ∈ C2×T occupies T time instances and has the following structure: X1,n = [Ūn , Ūn Un ], n = 1, . . . , ℓ, (45) both the diagonal matrix Ūn ∈ C2×2 and Un ∈ C2×(T −2) contain symbols intended for the ′ ′ ′ † † static user. The components of y1† = [y1,1 , · · · , y1,ℓ ] are: ′ † y1,n =[g1† Ūn , g1† Ūn Un ], † † =[g̃1,n , g̃1,n Un ], March 26, 2018 n = 1, . . . , ℓ (46) DRAFT 16 † = g1† Ūn . The static user by definition knows g1 so it can decode Ūn which where g̃1,n ′ yields 2 TT degrees of freedom. The remaining † Un involve (T − 2) observations in g̃1,n T′ T ′ 2 TT (T − 2) unknowns, so they require a further T′ T (T − 2) independent observations for reliable decoding. † † The components of y1† = [y1,1 , · · · , y1,ℓ ] are † =[h†1,n Ūn , h†1,n Ūn Un ], y1,n n = 1, . . . , ℓ =[h̃†1,n , h̃†1,n Un ], (47) where h̃†1,n = h†1,n Ūn is the equivalent channel estimated by the dynamic user. The dynamic user saves h̃†1,n Un for interference cancellation in the upcoming steps. 2) The transmitter sends a second super symbol intended for the dynamic user: X2 = [X2,1 , · · · , X2,ℓ ], (48) where X2,n = [Ũn , Ũn Vn ], n = 1, . . . , ℓ, (49) Ũn ∈ C2×2 is diagonal and includes 2 independent symbols intended for the static user, and Vn ∈ C2×(T −2) contains independent symbols intended for the dynamic user. The † † components of y2† = [y2,1 , · · · , y2,ℓ ] are † y2,n =[h†2,n Ũn , h†2,n Ũn Vn ], n = 1, . . . , ℓ =[h̃†2,n , h̃†2,n Vn ], (50) where h̃†2,n = h†2,n Ũn is the equivalent channel estimated by the dynamic user. The dynamic user saves h̃†2,n Vn which includes T′ T T′ (T − T ′† y2,ℓ ] are unknowns, and hence an additional ′ ′ † components of y2† = [y2,1 , ··· , ′ (T − 2) independent observations about 2 TT (T − 2) 2) observations are needed to decode Vn . The ′ y2,n =[g2† Ũn , g2† Ũn Vn ], n = 1, . . . , ℓ † † =[g̃2,n , g̃2,n Vn ], (51) † where g̃2,n = g2† Ũn is the equivalent channel estimated by the static user; the static user † saves g̃2,n Vn for the upcoming steps. Knowing g2 , the static user achieves 2 TT further ′ degrees of freedom from decoding Ũn . March 26, 2018 DRAFT 17 3) The transmitter emits a third super symbol consisting of a linear combination of the signals generated from the first and the second super symbols. X3 = [X3,1 , · · · , X3,ℓ ], (52) where † X3,n = [Ûn , Ûn (h̃†1,n Un + g̃2,n Vn )], n = 1, . . . , ℓ, (53) Ûn ∈ C2×2 is diagonal and contains 2 independent symbols intended for the static user, ′ and hence the static user achieves further 2 TT degrees of freedom. † The static user cancels g̃2,n Vn saved during the second super symbol and obtains h̃†1,n Un that includes the additional independent T′ T (T − 2) observations needed for decoding Un . T′ Therefore, the static user achieves 2 T (T − 2) further degrees of freedom. The dynamic user estimates the equivalent channel h̃†3,n = h†3,n Ûn , cancels h̃†1,n Un saved during the † Vn that contains the additional observations needed first super symbol, and obtains g̃2,n ′ for decoding Vn . Hence, the dynamic user achieves 2 TT (T − 2) degrees of freedom. In aggregate, over 3T ′ time instants, the static and dynamic user achieve the degrees of freedom d′ = 6 T′ T′ + 2 (T − 2), T T d=2 T′ (T − 2). T (54) This completes the proof of Theorem 6. Theorem 7: An outer bound on the degrees of freedom region of the fading broadcast channel characterized by Eq. (1), with one static and one dynamic user having delayed CSIT, is d′ + d ≤ 1, 2 d d′ + ≤ 1, 2 (55) (56) 1 . (57) T Proof: The inequality (57) represents the single-user outer bound [26]. We prove the d≤1− bound (55) as follows. We enhance the original channel by giving both users global CSIR. In addition, the channel output of the dynamic user, y(t), is given to the static user. Therefore, the channel outputs at time instant t are (y ′(t), y(t), H) at the static user, and (y(t), H) at the dynamic user. The enhanced channel is physically degraded [31], [32], hence, removing the delayed CSIT does not reduce the capacity [33]. Also, R′ ≤I(x(t); y ′(t), y(t)\|U, H) = h(y ′(t), y(t)\|U, H) − h(y ′ (t), y(t)\|U, x(t), H) March 26, 2018 DRAFT 18 R ≤I(U; y(t)\|H) = h(y(t)\|H) − h(y(t)\|U, H), (58) where U is an auxiliary random variable, and U → x → (y ′(t), y(t)) forms a Markov chain. Therefore, 1 R′ + R ≤h(y(t)\|H) + h(y ′(t), y(t)\|U, H) − h(y(t)\|U, H) + o(log(ρ)) 2 2 1 (59) ≤ log(ρ) + h(y ′(t), y(t)\|U, H) − h(y(t)\|U, H) + o(log(ρ)) 2 1 ≤ log(ρ) + max EH log \|I + HΣx H† \| − log(1 + h† (t)Σx h(t)) + o(log(ρ)) Tr{Σx }≤ρ,Σx <0 2 (60) ≤ log(ρ) + o(log(ρ)), (61) where (59) follows since h(y(t)\|H) ≤ log(ρ)+o(log(ρ)) [34], (60) follows from extremal entropy inequality [23], [24], [29], and (61) follows from Lemma 1. Hence, the bound (55) is proved. A similar argument, with the role of the two users reversed, leads to the bound (56). Remark 4: The inner and outer bounds obtained for the two-user case partially meet, with the gap diminishing with the coherence time of the dynamic user as shown in Fig. 5 and Fig. 6 for T = 15 and T = 30, respectively. D. Transmission to arbitrary number of static and dynamic users Theorem 8: The fading broadcast channel characterized by Eq. (1), with delayed CSIT, can achieve the multiuser degrees of freedom characterized by vectors Di , 1 D1 : 1 1 + 2 + ...+ ′ m X 1 m′ i=1 e†i , 1 2 2 2 D2 , . . . , Dmm′ +1 : (1 + )e†j + (1 − )e†m′ +i , j = 1, . . . , m′ , i = 1, . . . , m, 3 T 3 T m m † 1 m X † Dmm′ +2 , . . . , Dmm′ +m′ +2 : ej + e , j = 1, . . . , m′ , (1 − ) T T i=1 i 1 + 21 + . . . + m1 (62) (63) (64) where ej is the canonical coordinate vector. Their convex hull characterized an achievable degrees of freedom region. Proof: The achievability of (62) was proved in Section IV-A via multiuser transmission to static users. The achievability of (63) was proved in Section IV-C, via a two-user transmission to a dynamic-static pair. March 26, 2018 DRAFT 19 1.5 Achievable region Outer region 1 0.5 0 0 Fig. 5. 0.2 0.4 0.6 0.8 1 1.2 One static and one dynamic with delayed CSIT and T = 15 1.5 Achievable region Outer region 1 0.5 0 0 Fig. 6. 0.2 0.4 0.6 0.8 1 1.2 One static and one dynamic with delayed CSIT and T = 30 March 26, 2018 DRAFT 20 We now show the achievability of (64) via retrospective interference alignment [11] along with product superposition. Over a super symbol of length T , consider the following transmission: X = [U, UV], (65) where U ∈ Cm×m is diagonal and includes m independent symbols intended for the static user j, and V ∈ Cm×(T −m) is a super symbol containing independent symbols intended for the dynamic users according to retrospective interference alignment [11]. Therefore, the static user decodes U. Thus, over T time instants, the static user achieves m degrees of freedom and the dynamic users achieve 1 1 (T 1+ 12 +...+ m − m), hence (64) is achieved. Theorem 9: An outer bound on the degrees of freedom of the fading broadcast channel characterized by Eq. (1), with delayed CSIT, is ′ m X d′j di + ≤ 1, ′ m + m i=1 m (66) m m X d′j X di + ≤ 1, ′ ′ m m + m j=1 i=1 (67) m X j=1 ′ d′j ≤ 1, ∀j = 1, . . . , m′ , (68) 1 , ∀i = 1, . . . , m. (69) T Proof: The inequalities (68), (69) represent the single-user bounds on the static and the di ≤ 1 − dynamic users, respectively [26], [34]. The remainder of the proof is dedicated to establishing the bounds (66) and (67). We enhance the channel by providing global CSIR as well as allowing full cooperation among static users and full cooperation among dynamic users. The enhanced channel is equivalent to a broadcast channel with two users: one static equipped with m′ antennas and one dynamic ′ equipped with m antennas. Define Y ′ ∈ Cm and Y ∈ Cm to be the received signals of the static and the dynamic super-user, respectively, in the enhanced channel. We further enhance the channel by giving Y to the static user, generating a physically degraded channel since X → (Y ′ , Y) → Y forms a Markov chain. Feedback including delayed CSIT has no effect on capacity [33], therefore we remove it from consideration. Subsequently, we can utilize the Körner-Marton outer bound [27], ′ m X Rj′ ≤I(X; Y ′, Y\|U, H) j=1 March 26, 2018 DRAFT 21 m X Ri ≤I(U; Y\|H). (70) i=1 Therefore, from applying extremal entropy inequality [23], [29], [35] and Lemma 1, ′ m X j=1 m X Rj′ 1 1 Ri ′ + ≤ I(X; Y , Y\|U, H) + I(U; Y\|H) m′ + m i=1 m m′ + m m = m′ 1 1 1 h(Y ′, Y\|U, H) + o(log(ρ)) + h(Y\|H) − h(Y\|U, H) +m m m ≤ log(ρ) + o(log(ρ)). (71) Therefore, the bound (66) is proved. Similarly, we can prove the bound (67) using the same steps after switching the roles of the two users in the enhanced channel. V. H YBRID CSIT: P ERFECT CSIT FOR THE S TATIC U SERS AND N O CSIT FOR THE DYNAMIC U SERS Theorem 10: The fading broadcast channel characterized by Eq. (1), with perfect CSIT for the static users and no CSIT for the dynamic users, can achieve the following multiuser degrees of freedom, ′ D1 : m X e†j , (72) j=1 ′ D2 , . . . , Dm+1 m 1 1X † ej + (1 − )e†i , : T j=1 T i = 1, . . . , m. (73) Therefore, their convex hull is also achievable. Proof: D1 is achieved by inverting the channels of the static users at the transmitter and then every static user achieves one degree of freedom. D2 , . . . , Dm+1 in (73) are achieved using product superposition along with channel inversion as follows. The transmitted signal over T instants is, X = [u, uv† ], where u = Pm′ j=1 bj uj , (74) uj is a symbol intended for the static user j, gj† bj = 0, and v ∈ CT −1 contain independent symbols intended for the dynamic user i. Each of the static users receive an interference-free signal during the first time instant achieving one degrees of freedom. The dynamic user estimates its equivalent channel during the first time instant and decodes v during the remaining (T − 1) time instants. March 26, 2018 DRAFT 22 Theorem 11: An outer bound on the degrees of freedom of the fading broadcast channel characterized by Eq. (1), with perfect CSIT for the static users and no CSIT for the dynamic users, is ′ m X j=1 m X d′j + di ≤ 1, m′ + 1 i=1 d′j ≤ 1, m X (75) ∀j = 1, . . . , m′ , (76) 1 . (77) T i=1 Proof: The inequalities (76) represent single-user bounds for the static users [34], and (77) di ≤ 1 − is a time-sharing outer bound for the dynamic users that was established in [7], [18]. It remains to prove (75), as follows. We enhance the channel by giving global CSIR to all users and allowing full cooperation between the static users. This gives rise to an equivalent static user with m′ antennas receiving Y ′ over an equivalent channel G and noise Z′ . At this point, we have a multi-user system where CSIT is available with respect to one user, but not others. We then bound the performance of this system with that of another (similar) system that has no CSIT. To do so, we use the local statistical equivalence property developed and used in [13], [15], [36]. First, we draw G̃, Z̃ according to the distribution of G, Z′ and independent of them. We enhance the channel by providing Ỹ = G̃X + Z̃ to the static receiver and G̃ to all receivers. Because we do not provide G̃ to the transmitter, there is no CSIT with respect to Ỹ. According to [36], we have h(Ỹ, Y ′\|H) = h(Y ′ \|H) + o(log(ρ)), where H = (G, G̃, h1 , . . . , hm ), therefore we can remove Y ′ from the enhanced channel without reducing its degrees of freedom. This new equivalent channel has one user with m′ antennas receiving (Ỹ, H), m single-antenna users receiving (yi , H), and no CSIT.4 Having no CSIT, the enhanced channel is in the form of a multilevel broadcast channel studied in Section III-A, and hence using Theorem 1, ′ m X Rj′ ≤I(W ; Ỹ\|U, H) + I(X; Ỹ\|U, W, H) j=1 4 In the enhanced channel after removal of Y ′ , the transmitter and receivers still share information about G, but this random variable is now independent of all (remaining) transmit and receive variables. March 26, 2018 DRAFT 23 R1 ≤I(U, W ; y1\|V1 , H) − I(W ; Ỹ\|U, H) Ri ≤I(Vi−1 ; yi \|Vi , H), i = 2, . . . , m. (78) The dynamic receiver received signals have the same distribution. By following bounding steps parallel to (22), (23), (24), m X Ri ≤ log(ρ) + o(log(ρ)) − I(W ; Ỹ\|U, H) − h(y1 \|U, W, H). (79) j=1 Therefore, ′ m X j=1 m X Rj′ 1 h(Ỹ\|U, W, H) + R ≤ log(ρ) + o(log(ρ)) + ( − 1)I(W ; Ỹ\|U, H) + i m′ + 1 j=1 m′ + 1 m′ + 1 − h(y1 \|U, W, H), ≤ log(ρ) + o(log(ρ)) + (80) h(Ỹ, y1 \|U, W, H) − h(y1 \|U, W, H) m′ + 1 ≤ log(ρ) + o(log(ρ)), (81) (82) where the last inequality follows from applying the extremal entropy inequality [23], [29], [35] and Lemma 1. This concludes the proof of the bound (75). VI. H YBRID CSIT: P ERFECT CSIT FOR THE S TATIC U SERS AND D ELAYED CSIT FOR THE DYNAMIC U SERS We begin with inner and outer bounds for one static and one dynamic user, then extend the result to multiple users. The transmitter knows the channel of the static users perfectly and instantaneously, and an outdated version of the channel of the dynamic users. A. Transmitting to One Static and One Dynamic User Theorem 12: For the fading broadcast channel characterized by Eq. (1) with one static and one dynamic user, with perfect CSIT for the static user and delayed CSIT for the dynamic user, the achievable degrees of freedom region is the convex hull of the vectors, 1 1 1 , − ), 2T 2 2T 1 1 D2 :(d′, d) = ( , 1 − ). T T D1 :(d′, d) = (1 − March 26, 2018 (83) (84) DRAFT 24 Proof: The degrees of freedom (84) can be achieved by product superposition as discussed in Section III, without CSIT. We proceed to prove the achievability of (83). 1) Consider [u1 , ··· , uT −1 ] to be a complex 2 × (T − 1) matrix containing symbols intended for the static user, [v1 , · · · , vT −1 ] intended for the dynamic user, and b ∈ C is a beamforming vector so that g† b = 0. In addition we define u0 = 0, v0 = 1. Using these components, the transmitter constructs and transmits a super-symbol of length T , whose value at time t is: x†1 (t) = ut + b vt . (85) Note that x1 (0) = b does not carry any information for either user, and serves as a pilot. The received super symbol at the static user is: ′ y1† = [0, g† u1 , · · · , g† uT −1 ]. (86) The received super symbol at the dynamic user y1† =[h†1 b, (h†1 u1 + h†1 bv1 ), · · · , (h†1 uT −1 + h†1 bvT −1 )]. (87) The dynamic user estimates its equivalent channel h†1 b from the received value in the first time instant. The remaining terms include symbols intended for the dynamic user plus some interference, whose cancellation is the subject of the next step. 2) The transmitter next sends a second super symbol of length T , x2 = [ū, ū(h†1 u1 ), · · · , ū(h†1 uT −1 )], (88) where ū ∈ C is a symbol intended for the static user. Hence, y2† = [h2 ū, h2 ū(h†1 u1 ), · · · , h2 ū(h†1 uT −1 )]. (89) The dynamic user estimates the equivalent channel h2 ū during the first time instant and then acquires h†1 ut , the interference in (87). Therefore, using y1 , y2 , the dynamic user solves for vt achieving (T − 1) degrees of freedom. Furthermore, ′ y2† = [g1 ū, g1 ū(h†1 u1 ), · · · , g1 ū(h†1 uT −1 )]. (90) The static user solves for ū achieving one degree of freedom and also uses h†1 ut to solve for ut achieving further 2 (T − 1) degrees of freedom. March 26, 2018 DRAFT 25 In summary, during 2T instants, the static user achieves (2T − 1) degrees of freedom and the dynamic user achieves (T − 1) degrees of freedom. This shows the achievability of (83). Theorem 13: For the fading broadcast channel characterized by Eq. (1) with one static and one dynamic user, where there is perfect CSIT for the static user and delayed CSIT for the dynamic user, an outer bound on the degrees of freedom region is, d′ + d ≤1, 2 d′ ≤1, (91) (92) 1 . (93) T Proof: The inequalities (92) and (93) represent the single-user outer bounds [26], [34]. It d ≤1 − only remains to prove the outer bound (91), as follows. 1) We enhance the channel by giving global CSIR to both users and also give y to the static user. The enhanced channel is physically degraded having (Y ′ , G) at the static user and (y, G) at the dynamic user, where Y ′ , (y ′, y) and G , (h, g). In a physically degraded channel, causal feedback (including delayed CSIT) does not affect capacity [33], so we can remove the delayed CSIT with respect to the dynamic user. 2) We now use another enhancement with the motivation to remove the remaining CSIT (noncausal, with respect to the static user). This is accomplished, similar to Theorem 11, via local statistical equivalence property [13], [15], [36] in the following manner. We create a channel G̃, and noise Z̃ with the same distribution but independently of the true channel and noise, and a signal Ỹ = G̃X + Z̃. A genie will give Ỹ to the static receiver and G̃ to both receivers. It has been shown [36] that h(Ỹ, Y ′ \|H) = h(Y ′ \|H) + o(log ρ), where H = (G, G̃), therefore we can remove Y ′ from the enhanced channel without reducing its degrees of freedom. 3) The enhanced channel is still physically degraded, therefore [31], [32] R′ ≤I(x; Ỹ\|U, H) = h(Ỹ\|U, H) + o(log(ρ)) R ≤I(U; y\|H) = h(y\|H) − h(y\|U, H), (94) where U is an auxiliary random variable, and U → x → (y ′ , y) forms a Markov chain. Therefore, 1 1 ′ R + R ≤h(y\|H) + h(Ỹ\|U, H) − h(y\|U, H) + o(log(ρ)) 2 2 March 26, 2018 DRAFT 26 1.5 Achievable region Outer region 1 0.5 0 0 Fig. 7. 0.2 0.4 0.6 0.8 1 1.2 One static and one dynamic user with hybrid CSIT and T = 15 ≤ log(ρ) + o(log(ρ)), (95) where the last inequality follows from extremal entropy inequality and Lemma 1 [23], [29], [35]. This concludes the proof of the bound (91). Remark 5: For the above broadcast channel with hybrid CSIT, the achievable sum degrees of freedom is dsum = 3 2 − T1 , and the outer bound on the sum degrees of freedom is dsum ≤ 32 . The gap decreases with the dynamic user coherence time (see Fig. 7 and 8). B. Multiple Static and Dynamic Users Theorem 14: The fading broadcast channel characterized by Eq. (1), with perfect CSIT for the static users and delayed CSIT for the dynamic users, can achieve the following degrees of freedom, ′ D1 : m X e†j , (96) j=1 D2 , . . . , Dmm′ +1 : (1 − March 26, 2018 1 1 † 1 † )ej + ( − )e , 2T 2 2T i j = 1, . . . , m′ , i = 1, . . . , m, (97) DRAFT 27 1.5 Achievable region Outer region 1 0.5 0 0 Fig. 8. 0.2 0.4 0.6 0.8 1 1.2 One static and one dynamic user with hybrid CSIT and T = 30 ′ Dmm′ +2 , . . . , Dmm′ +m+2 m′ Dmm′ +m+3 m 1 1X † ej + (1 − )e†i , : T j=1 T mX † 1 : ej + ( 1 T j=1 1 + 2 + ...+ i = 1, . . . , m, (98) m m X † (1 − )) e. 1 T i=1 i m (99) The achievable region consists of the convex hull of the above vectors. Proof: D1 is achieved by inverting the channel of the static users at the transmitter, providing one degree of freedom per static user. The achievability of D2 , . . . , Dmm′ +1 was established in Section VI-A, and that of Dmm′ +2 , . . . , Dmm′ +m+2 was proved in Section V without CSIT for the dynamic user, so it remains achievable with delayed CSIT. Dmm′ +m+3 is achieved by retrospective interference alignment [11] along with product superposition, as follows. The transmitted signal over T instants is X = [Ū, ŪV], (100) where Ū ∈ Cm×m contains independent symbols intended for the static users sent by inverting the channels of the static users. Therefore, during the first m time instants, each static user receives an interference-free signal and achieves m degree of freedom, and furthermore the dynamic users estimate their equivalent channels. During the remaining time instants, each dynamic receiver March 26, 2018 DRAFT 28 obtains coherent observations of (T − m) transmit symbols, which are pre-processed, combined and interference-aligned into super-symbols V according to retrospective interference alignment techniques of [11]. Accordingly, each dynamic receiver achieves 1 1 (1 1+ 12 +...+ m − m ) T degrees of freedom. Theorem 15: An outer bound on the degrees of freedom region of the fading broadcast channel characterized by Eq. (1), with perfect CSIT for the static users and delayed CSIT for the dynamic users, is ′ m X j=1 m X d′j di + ≤ 1, ′ m + m i=1 m m X i=1 di ≤ 1+ d′j ≤ 1, (101) 1 2 m + ...+ 1 m , j = 1, . . . , m′ , (102) (103) 1 , i = 1, . . . , m. (104) T Proof: The inequalities (103) and (104) represent the single-user outer bounds for the static di ≤ 1 − and dynamic users, respectively [26], [34]. According to Theorem 5, (102) represents an outer bound for the dynamic users. It only remains to prove (101) as follows. 1) The original channel is enhanced by giving the users global CSIR. Furthermore, we assume full cooperation between the static users and between the dynamic users. The resulting enhanced channel is a broadcast channel with two users: one static user equipped with m′ antennas, received signal Y ′ , channel G, and noise noise Z′ , and one dynamic user equipped with m antennas, received signal Y, channel H, and noise Z. 2) We further enhance the channel by giving Y to the static user, constructing a physically degraded channel. For the enhanced channel, the static receiver is equipped with m′ + m ′ antennas and has received signal Ŷ = [Y † , Y † ]† , channel Ĝ = [G† , H† ]† , and noise ′ Ẑ = [Z† , Z † ]† . Since any causal feedback (including delayed CSIT) does not affect the capacity of a physically degraded channel [33], the delayed CSIT for the dynamic receiver can be removed. 3) We now use another enhancement with the motivation to remove the remaining CSIT (non-causal, with respect to the static user). We create an artificial channel and noise, G̃, Z̃, with the same distribution but independent of Ĝ, Ẑ, and a signal Ỹ = G̃X + Z̃. A March 26, 2018 DRAFT 29 genie will give Ỹ to the static receiver and G̃ to both receivers. It has been shown [36] that h(Ỹ, Ŷ\|H) = h(Ŷ\|H) + o(log ρ), where H = (Ĝ, G̃), therefore we can remove Ŷ from the enhanced channel without reducing its degrees of freedom. 4) The enhanced channel is physically degraded without CSIT, therefore [31], [32], ′ m X Rj′ ≤ I(X; Ỹ\|U, H) m X Ri ≤ I(U; Y\|H). j=1 (105) i=1 Hence, ′ m X j=1 m X Rj′ 1 1 1 Ri + ≤ ′ h(Ỹ\|U, H) + h(Y\|H) − h(Y\|U, H) + o(log(ρ)) ′ m + m j=1 m m + m m m ≤ log(ρ) + o(log(ρ)), (106) where the last inequality follows from the extremal entropy inequality [23], [29], [35] and Lemma 1 and since h(Y\|H) ≤ m log(ρ) + o(log(ρ)) [34]. This concludes the proof of the bound (101). VII. C ONCLUSION A multiuser broadcast channel was studied where some receivers experience longer coherence intervals and have CSIR while other receivers experience a shorter coherence interval and do not enjoy free CSIR. The degrees of freedom were studied under delayed CSIT, hybrid CSIT, and no CSIT. Among the techniques employed were interference alignment and beamforming along with product superposition for the inner bounds. The outer bounds involved a bounding of the rate region of the multiuser (discrete memoryless) multilevel broadcast channel. Some highlights of the results are: for one static and one dynamic user with delayed CSIT, the achievable degrees of freedom region partially meets the outer bound. For one static user with perfect CSIT and one dynamic user with delayed CSIT, the gap between the achievable and the outer sum degrees of freedom is inversely proportional to the dynamic user coherence time. For each of the considered CSI conditions, inner and outer bounds were also found for arbitrary number of users. From these results we conclude that in the broadcast channel, coherence diversity delivers gains that are distinct from, and augment, the gains from beamforming and interference alignment. March 26, 2018 DRAFT 30 The authors anticipate that the tools and results of this paper can be helpful for future studies of hybrid CSIT/CSIR in other multi-terminal networks. A PPENDIX I P ROOF OF T HEOREM 1 Recall, Mj′ , Mi are the messages of users j = 1, . . . , m′ and i = 1, . . . , m, respectively. We ′ ′ enhance the channel by assuming that user j = 1, . . . , m′ knows the messages Mj+1 , . . . , Mm ′ and M1 , . . . , Mm and user i = 1, . . . , m knows the messages Mi+1 , . . . , Mm . Using Fano’s inequality, chain rule, and data processing inequality we can bound the rates of the static user j = 1, . . . , m′ , ′ ′ ′ ′ nRj′ ≤ I(Mj′ ; Yj,1 , . . . , Yj,n \|Mj+1 , . . . , Mm ′ , M1 , . . . , Mm ) (107) = n X ′ I(Mj′ ; Yj,k \|Uj,k ) (108) ≤ n X ′ ′ ′ I(Mj′ , Uj,k , Yj−1,1 , . . . , Yj−1,k−1 ; Yj,k \|Uj,k ) (109) = n X ′ I(Uj−1,k ; Yj,k \|Uj,k ) (110) k=1 k=1 k=1 where ′ ′ ′ ′ , . . . , Mm Uj,k = Mj+1 ′ , M1 , . . . , Mm , Yj,1 , . . . , Yj,k−1 , ′ denotes the received signal of user j at time instant k, Yj,k Um′ → · · · → U1 → X → (Y1′ , . . . , Ym′ ′ , Y1 , . . . , Ym) forms a Markov chain, and U0 = X. The rate of static user m′ can be bounded as n X ′ I(Um′ −1,k ; Ym′ ′ ,k \|Um′ ,k ) nRm′ ≤ = ≤ = k=1 n X k=1 n X k=1 n X k=1 March 26, 2018 I(Xk ; Ym′ ′ ,k \|Um′ ,k ) − n X I(Xk ; Ym′ ′ ,k \|Um′ −1,k ) (111) (112) k=1 I(Xk , Y1,k+1, . . . , Y1,n ; Ym′ ′ ,k \|Um′ ,k ) − n X I(Xk ; Ym′ ′ ,k \|Um′ −1,k ) (113) k=1 I(Y1,k+1, . . . , Y1,n ; Ym′ ′ ,k \|Um′ ,k ) + n X I(Xk ; Ym′ ′ ,k \|Um′ ,k , Y1,k+1, . . . , Y1,n ) k=1 DRAFT 31 − n X I(Xk ; Ym′ ′ ,k \|Um′ −1,k ) (114) k=1 = n X I(Wk ; Ym′ ′ ,k \|Um′ ,k ) + n X I(Xk ; Ym′ ′ ,k \|Um′ ,k , Wk ) I(Xk ; Ym′ ′ ,k \|Um′ −1,k ), k=1 k=1 k=1 − n X (115) n where Wk = Y1,k+1 . Similarly, nRi ≤I(Mi ; Yi,1, . . . , Yi,n \|Mi+1 , . . . , Mm ) = = = n X k=1 n X k=1 n X (116) I(Mi ; Yi,k \|Vi,k ) (117) I(Mi , Vi,k , Yi−1,k+1, . . . , Yi−1,n ; Yi,k \|Vi,k ) (118) I(Vi−1,k ; Yi,k \|Vi,k ), (119) k=1 where we define Vi,k , (Mi+1 , . . . , Mm , Yi,k+1, . . . , Yi,n ), which leads to the Markov chain Vm → · · · → V1 → (Um′ , W ) → X → (Y1′ , . . . , Ym′ ′ , Y1 , . . . , Ym ) . Using the chain rule and Csiszár sum identity [37], we obtain the bound (125). R1 ≤ n X I(M1 , . . . , Mm ; Y1,k \|V1,k ) (120) ≤ n X I(M1 , . . . , Mm , Y1,k+1, . . . , Y1,n ; Y1,k \|V1,k ) (121) = n X I(M1 , . . . , Mm , Y1,k+1, . . . , Y1,n , Ym′ ′ ,1 , . . . , Ym′ ′ ,k−1 ; Y1,k \|V1,k ) (122) k=1 k=1 k=1 − n X I(Ym′ ′ ,1 , . . . , Ym′ ′ ,k−1; Y1,k \|M1 , . . . , Mm , Y1,k+1, . . . , Y1,n ) (123) k=1 = n X = n X I(Um′ ,k , Wk ; Y1,k \|V1,k ) − n X I(Y1,k+1, . . . , Y1,n ; Ym′ ′ ,k \|Um′ ,k ) (124) I(Um′ ,k , Wk ; Y1,k \|V1,k ) − n X I(Wk ; Ym′ ′ ,k \|Um′ ,k ). (125) k=1 k=1 k=1 k=1 By introducing a time-sharing auxiliary random variable, Q, [38] and defining X ,(XQ , Q), March 26, 2018 ′ Yj′ , (Yj,Q , Q) DRAFT 32 Yi ,(Yi,Q , Q), Ui , (Ui,Q , Q) Vj ,(Vj,Q , Q), W , (WQ , Q), (126) we establish (7)-(10). Similarly, we can follow the same steps to prove (11)-(14) after switching the role of the two sets of variables Y1′ , . . . , Ym′ ′ and Y1 , . . . , Ym . This completes the proof of Theorem 1. A PPENDIX II M ULTILEVEL B ROADCAST C HANNEL WITH D EGRADED M ESSAGE S ETS Here, we study the capacity of the multiuser multilevel broadcast channel that is characterized by (6) with degraded message sets. In particular, M0 ∈ 1 : 2nR0 is to be communicated to all receivers, and furthermore M1 ∈ 1 : 2nR1 is to be communicated to receiver Y1 .5 A threereceiver special case was studied by Nair and El Gamal [22] where the idea of indirect decoding was introduced, and the capacity is the set of rate pairs (R1 , R0 ) such that R0 ≤ min I(U; Y2 ), I(V ; Y1′ ) , R1 ≤I(X; Y1\|U), R0 + R1 ≤I(V ; Y1′ ) + I(X; Y1\|V ), (127) for some pmf p(u, v)p(x\|v). In the sequel, we give a generalization of Nair and El Gamal for multiuser multilevel broadcast channel. Theorem 16: The capacity of multiuser multilevel broadcast channel characterized by (6), with degraded message sets, is the set of rate pairs (R1 , R0 ) such that R0 ≤ min I(U; Ym ), I(V ; Ym′ ′ ) , R1 ≤I(X; Y1\|U), R0 + R1 ≤I(V ; Ym′ ′ ) + I(X; Y1 \|V ), (128) for some pmf p(u, v)p(x\|v). Proof: The converse parallels the proof of the converse of the three-receiver case studied by Nair and El Gamal in [22] after replacing Y2 , Y1′ with Ym , Ym′ ′ , respectively. In particular, U 5 For compactness of expression, here we refer to each receiver by the variable denoting its received signal. March 26, 2018 DRAFT 33 and V are defined as follows. Uk , (M0 , Y1,1, . . . , Y1,k−1, Ym,k+1, . . . , Ym,n ), Vk , (M0 , Y1,1, . . . , Y1,k−1, Ym′ ′ ,k+1, . . . , Ym′ ′ ,n ), k = 1, . . . , n, and let Q be a time-sharing random variable uniformly distributed over the set {1, . . . , n} and independent of X n , Y1n , Ym,1 , . . . , Ym,n , Ym′ ′ ,1 , . . . , Ym′ ′ ,n . We then set U = (UQ , Q), V = (VQ , Q), X = XQ , Y1 = Y1,Q , Ym = Ym,Q , and Ym′ ′ = Ym′ ′ ,Q . This completes the converse part of the proof. The achievability part uses superposition coding and indirect decoding as follows. • Rate splitting: divide the private message M1 into two independent messages M10 at rate R10 and M11 at rate R11 , where R1 = R10 + R11 . • Codebook generation: fix a pmf p(u, v)p(x\|v) and randomly and independently generate Q 2nR0 sequences un (m0 ), m0 ∈ 1 : 2nR0 , each according to nk=1 pU (uk ). For each m0 , randomly and conditionally independently generate 2nR10 sequences v n (m0 , m10 ), m10 ∈ [1 : Q 2nR10 ], each according to nk=1 pV \|U (vk \|uk (m0 )). For each pair (m0 , m10 ), randomly and conditionally independently generate 2nR11 sequences xn (m0 , m10 , m11 ), m11 ∈ [1 : 2nR11 ], Q each according to nk=1 pX\|V (xk \|vk (m0 , m10 )). • Encoding: to send the message pair (m0 , m1 ) = (m0 , m10 , m11 ), the encoder transmits xn (m0 , m10 , m11 ). • Decoding at the users Y2 , . . . , Ym : decoder i declares that m̂0i ∈ [1 : 2nR0 ] is sent if it is (n) the unique message such that (un (m̂0i ), yin) ∈ Tǫ . Hence, by law of large numbers and the packing lemma [38], the probability of error tends to zero as n → ∞ if R0 < min {I(U; Yi ) − δ(ǫ)}, 2≤i≤m = I(U; Ym ) − δ(ǫ), (129) where the last equality follows from applying data processing inequality on the Markov chain U → X → Y1 → Y2 → · · · → Ym . • Decoding at Y1 : decoder 1 declares that (m̂01 , m̂10 , m̂11 ) is sent if it is the unique message triple such that un (m̂01 ), v n (m̂01 , m̂10 ), xn (m̂01 , m̂10 , m̂11 ), y1n ∈ [1 : 2nR0 ]. Hence, by law of large numbers and the packing lemma [38], the probability of error tends to zero as March 26, 2018 DRAFT 34 n → ∞ if R11 < I(X; Y1\|V ) − δ(ǫ), R10 + R11 < I(X; Y1\|U) − δ(ǫ), R0 + R10 + R11 < I(X; Y1) − δ(ǫ). • (130) Decoding at users Y1′ , . . . , Ym′ ′ : decoder j decodes m0 indirectly by declaring m̃0j is sent (n) if it is the unique message such that (un (m̃0j ), v n (m̃0j , m10 ), zjn ) ∈ Tǫ for some m10 ∈ [1 : 2nR0 ]. Hence, by law of large numbers and packing lemma, the probability of error tends to zero as n → ∞ if R0 + R10 < min {I(U, V ; Yj′ ) − δ(ǫ)}, min {I(V ; Yj′ ) − δ(ǫ)}, 1≤j≤m′ = 1≤j≤m′ = I(V ; Ym′ ′ ) − δ(ǫ), (131) where the last two equalities follow from applying the chain rule and data processing inequality on the Markov chain U → V → X → Y1′ → Y2′ → · · · → Ym′ ′ . By combining the bounds in (129), (130), (131), substituting R10 + R11 = R1 , and eliminating R10 and R11 by the Fourier-Motzkin procedure [22], the proof of the achievability is completed. R EFERENCES [1] M. Fadel and A. Nosratinia, “Block-fading broadcast channel with hybrid CSIT and CSIR,” in IEEE International Symposium on Information Theory (ISIT), June 2017, pp. 1873–1877. [2] C. Huang, S. Jafar, S. Shamai, and S. Vishwanath, “On degrees of freedom region of MIMO networks without channel state information at transmitters,” IEEE Trans. Inf. Theory, vol. 58, no. 2, pp. 849–857, Feb. 2012. [3] C. Vaze and M. Varanasi, “The degree-of-freedom regions of MIMO broadcast, interference, and cognitive radio channels with no CSIT,” IEEE Trans. Inf. Theory, vol. 58, no. 8, pp. 5354–5374, Aug. 2012. [4] A. Lapidoth, S. Shamai, and M. Wigger, “On the capacity of fading MIMO broadcast channels with imperfect transmitter side-information,” arXiv preprint cs/0605079, 2006. [5] S. Jafar, “Blind interference alignment,” IEEE J. Sel. Topics Signal Process., vol. 6, no. 3, pp. 216–227, June 2012. [6] M. Fadel and A. Nosratinia, “Broadcast channel under unequal coherence intervals,” in IEEE International Symposium on Information Theory (ISIT), July 2016, pp. 275–279. [7] ——, “Coherence disparity in broadcast and multiple access channels,” IEEE Trans. Inf. Theory, vol. 62, no. 12, pp. 7383–7401, Dec. 2016. March 26, 2018 DRAFT 35 [8] G. Caire and S. Shamai, “On the achievable throughput of a multiantenna Gaussian broadcast channel,” IEEE Trans. Inf. Theory, vol. 49, no. 7, pp. 1691–1706, July 2003. [9] H. Weingarten, Y. Steinberg, and S. Shamai, “The capacity region of the Gaussian multiple-input multiple-output broadcast channel,” IEEE Trans. Inf. Theory, vol. 52, no. 9, pp. 3936–3964, Sept. 2006. [10] A. Davoodi and S. Jafar, “Aligned image sets under channel uncertainty: Settling a conjecture by Lapidoth, Shamai and Wigger on the collapse of degrees of freedom under finite precision CSIT,” arXiv preprint arXiv:1403.1541, 2014. [11] M. Maddah-Ali and D. Tse, “Completely stale transmitter channel state information is still very useful,” IEEE Trans. Inf. Theory, vol. 58, no. 7, pp. 4418–4431, July 2012. [12] T. Gou and S. Jafar, “Optimal use of current and outdated channel state information: Degrees of freedom of the MISO BC with mixed CSIT,” IEEE Commun. Lett., vol. 16, no. 7, pp. 1084–1087, July 2012. [13] R. Tandon, M. A. Maddah-Ali, A. Tulino, H. V. Poor, and S. Shamai, “On fading broadcast channels with partial channel state information at the transmitter,” in International Symposium on Wireless Communication Systems (ISWCS), Aug. 2012, pp. 1004–1008. [14] S. Amuru, R. Tandon, and S. Shamai, “On the degrees-of-freedom of the 3-user MISO broadcast channel with hybrid CSIT,” in IEEE International Symposium on Information Theory (ISIT), 2014, pp. 2137–2141. [15] R. Tandon, S. Jafar, S. Shamai, and V. Poor, “On the synergistic benefits of alternating CSIT for the MISO broadcast channel,” IEEE Trans. Inf. Theory, vol. 59, no. 7, pp. 4106–4128, July 2013. [16] Y. Li and A. Nosratinia, “Product superposition for MIMO broadcast channels,” IEEE Trans. Inf. Theory, vol. 58, no. 11, pp. 6839–6852, Nov. 2012. [17] ——, “Coherent product superposition for downlink multiuser MIMO,” IEEE Trans. Wireless Commun., vol. PP, no. 99, pp. 1–9, 2014. [18] M. Fadel and A. Nosratinia, “Coherent, non-coherent, and mixed–CSIR broadcast channels: Multiuser degrees of freedom,” in IEEE International Symposium on Information Theory (ISIT), June 2014, pp. 2574–2578. [19] ——, “Coherence disparity in time and frequency,” in Proc. IEEE Global Telecommunication Conference (GLOBECOM’16), Dec. 2016, pp. 1–6. [20] F. Zhang, M. Fadel, and A. Nosratinia, “Spatially correlated MIMO broadcast channel: Analysis of overlapping correlation eigenspaces,” in IEEE International Symposium on Information Theory (ISIT), June 2017, pp. 1097–1101. [21] S. Borade, L. Zheng, and M. Trott, “Multilevel broadcast networks,” in IEEE International Symposium on Information Theory (ISIT), June 2007, pp. 1151–1155. [22] C. Nair and A. Gamal, “The capacity region of a class of three-receiver broadcast channels with degraded message sets,” IEEE Trans. Inf. Theory, vol. 55, no. 10, pp. 4479–4493, Oct. 2009. [23] T. Liu and P. Viswanath, “An extremal inequality motivated by multiterminal information-theoretic problems,” IEEE Trans. Inf. Theory, vol. 53, no. 5, pp. 1839–1851, May 2007. [24] R. Liu, T. Liu, V. Poor, and S. Shamai, “A vector generalization of Costa’s entropy-power inequality with applications,” IEEE Trans. Inf. Theory, vol. 56, no. 4, pp. 1865–1879, Apr. 2010. [25] T. Marzetta and B. Hochwald, “Capacity of a mobile multiple-antenna communication link in Rayleigh flat fading,” IEEE Trans. Inf. Theory, vol. 45, no. 1, pp. 139–157, Jan. 1999. [26] L. Zheng and D. Tse, “Communication on the Grassmann manifold: a geometric approach to the noncoherent multipleantenna channel,” IEEE Trans. Inf. Theory, vol. 48, no. 2, pp. 359–383, Feb. 2002. March 26, 2018 DRAFT 36 [27] K. Marton, “A coding theorem for the discrete memoryless broadcast channel,” IEEE Trans. Inf. Theory, vol. 25, no. 3, pp. 306–311, May 1979. [28] S. Yang, M. Kobayashi, D. Gesbert, and X. Yi, “Degrees of freedom of time correlated MISO broadcast channel with delayed CSIT,” IEEE Trans. Inf. Theory, vol. 59, no. 1, pp. 315–328, Jan. 2013. [29] X. Yi, S. Yang, D. Gesbert, and M. Kobayashi, “The degrees of freedom region of temporally correlated MIMO networks with delayed CSIT,” IEEE Trans. Inf. Theory, vol. 60, no. 1, pp. 494–514, Jan. 2014. [30] C. Shannon, “The zero error capacity of a noisy channel,” IEEE Trans. Inf. Theory, vol. 2, no. 3, pp. 8–19, Sept. 1956. [31] P. Bergmans, “Random coding theorem for broadcast channels with degraded components,” IEEE Trans. Inf. Theory, vol. 19, no. 2, pp. 197–207, Mar. 1973. [32] ——, “A simple converse for broadcast channels with additive white Gaussian noise,” IEEE Trans. Inf. Theory, vol. 20, no. 2, pp. 279–280, Mar. 1974. [33] A. E. Gamal, “The feedback capacity of degraded broadcast channels (corresp.),” IEEE Trans. Inf. Theory, vol. 24, no. 3, pp. 379–381, May 1978. [34] E. Telatar, “Capacity of multi-antenna Gaussian channels,” European transactions on telecommunications, vol. 10, no. 6, pp. 585–595, 1999. [35] H. Weingarten, T. Liu, S. Shamai, Y. Steinberg, and P. Viswanath, “The capacity region of the degraded multiple-input multiple-output compound broadcast channel,” IEEE Trans. Inf. Theory, vol. 55, no. 11, pp. 5011–5023, Oct. 2009. [36] P. Mukherjee, R. Tandon, and S. Ulukus, “Secure degrees of freedom region of the two-user MISO broadcast channel with alternating CSIT,” IEEE Trans. Inf. Theory, vol. PP, no. 99, pp. 1–1, Apr. 2017. [37] I. Csiszár and J. Körner, “Information theory : Coding theorems for discrete memoryless channels,” Budapest: Akadémiai Kiadó, 1981. [38] A. E. Gamal and Y. Kim, Network information theory. March 26, 2018 Cambridge University Press, 2011. DRAFT	7cs.IT
arXiv:1208.3882v1 [cs.DC] 19 Aug 2012 Coordination Level Modeling and Analysis of Parallel Programs using Petri Nets Francisco Heron de Carvalho-Junior Departamento de Computação Universidade Federal do Ceará (UFC) Fortaleza, Brazil Rafael Dueire Lins Centro de Informática Universidade Federal de Pernambuco Recife, Brazil March 2005 Abstract In the last fifteen years, the high performance computing (HPC) community has claimed for parallel programming environments that reconciles generality, higher level of abstraction, portability, and efficiency for distributed-memory parallel computing platforms. The Hash component model appears as an alternative for addressing HPC community claims for fitting these requirements. This paper presents foundations that will enable a parallel programming environment based on the Hash model to address the problems of “debugging”, performance evaluation and verification of formal properties of parallel program by means of a powerful, simple, and widely adopted formalism: Petri nets. 1 Introduction Haskell# is a parallel extension to Haskell [6], the most widely used non-strict pure functional programming language [25]. It makes possible the coordination of a set of functional processes written in Haskell through a configuration language, called HCL (Haskell# Configuration Language). Thus, Haskell# separate the programming task in two levels: the computation level, where functional processes are written in Haskell, and the coordination level, where functional processes are coordinated. The coordination media of Haskell# is also called the Hash component model. In recent works, we have generalized the Hash component model in order to support other programming languages than Haskell at the computation level. Functional processes are now units, possibly written 1 in any programming language supported by a programming environment that complies to the Hash component model. In this paper, we still suppose that units are the functional processes of Haskell#. The coordination level of Haskell# , i. e. the Hash component model, was designed in order to make possible to translate the coordination media of Haskell# programs onto Petri nets. This paper addresses the issue of presenting a translation schema for that, demonstrating its use to formal analysis of them, involving verification of formal properties. The use of Petri nets allows for reusing existing automatic tools based on this formalisms for reasoning about Hash programs, such as PEP [4] and INA [21]. Further extensions could make possible performance evaluation using timed or stochastic Petri Nets variants [14, 26]. In the rest of the paper, we use the Hash component model to refer to the coordination level of Haskell# , i.e. it is assumed that functional processes are written in Haskell (functional modules) and they communicate through either stream-based or singleton communication channels that link their input and output ports. Such concepts are not present in the general definition of the Hash component model. In additional to this introduction, this paper comprises the following sections. Section 2 presents additional details about the Hash component model. Section 3 presents a translation schema of Hash programs onto Petri nets. Section 4 demonstrates how Petri net models of Hash programs may be used for verification of formal concurrency properties of Hash programs. 2 The Hash Component Model Distributed parallel programs may be viewed as collections of processes that interact by exchanging messages during execution. Current programming models provide the ability to describe computation of processes by augmenting common languages with notations for explicit message passing. However, they do not provide ability to modularize concerns that appear in the design of parallel applications, including the concern of parallelism itself, which are scattered across the implementation of processes. We advocate that this is the key feature for integrating advanced software engineering techniques in the development environment of HPC parallel applications. In sequential programming, the focus is on modularization of concerns, since there are a unique conceptual “process” and efficiency requirements are less restrictive. This is the essential difference that makes sequential programming actually more suitable for current software engineering techniques for large scale applications than current parallel programming. The Hash component model may be viewed as a new paradigm for developing message passing programs. Now, they may be viewed from two orthogonal perspective dimensions: the dimension of processes and the dimension of components. A process correspond to the related notion derived from conventional message passing programming. Thus, they are agents that perform computational 2 tasks, communicating through communication channels. Conceptually, Hash channels, like in OCCAM [17], are point-to-point, synchronous, typed and unidirectional. Bounded buffers are also supported. The disciplined use of channels is the feature that makes possible formal analysis of parallel programs by using Petri nets, the main topic of this paper. A component is an abstract entity that address a functional or non-functional concern of the application or its execution environment in the parallel program. A component describe the role of a set of processes with respect to a given concern. The sets of components that respectively implement a set of concerns may overlap, allowing for modular separation of concerns that are interlaced across implementation of processes (cross-cutting concerns). The separation of cross-cutting concerns is an active research area in object-oriented programming of large scale applications [19]. A Hash program is defined by a main component that address the overall application concern. Some common examples of cross-cutting non-functional concerns that appears in HPC applications are: placement of processes onto processors, secure policies for accessing computing resources on grids, fault-tolerance schemes for long-running applications, parallel debugging, execution timing, and so on. Hash programming is performed from the perspective of components, instead of processes, but resulting in the specification of the topology of a network of parallel processes. Components may be composed or simple. Composed components are programmed using the Hash configuration language (HCL), being built by hierarchical composition of other components, called inner components. HCL may be viewed as a language for gluing and orchestrating components, i.e. a connector language. It is distinguished from other compositional languages because it supports composition of parallel components by overlapping the concerns they address. Conventional compositional languages only allows nested composition of sequential components. Simple components addresses functional concerns, implemented using a host language, supposed to be sequential. Simple components are the atoms of functionality in Hash programs, constituting the leaves in their component hierarchy. The Hash component model supports parallel programming with skeletons [11] without any additional language support. Partial topological skeletons may expose topological patterns of interaction between processes in a Hash program, which may be used to produce more efficient code for specific architectures and execution environments [8]. They are implemented as composed components parameterized by their addressed concerns. The Hash component model has origins in Haskell# [6], where the host language used to program simple components is Haskell. Haskell enables separation between coordination and computation code, by attaching lazy streams to communication channels at coordination level, avoiding the use of communication primitives in communication code [5]. This paper on focused in Haskell# . In the next section, it is described how composed components, and skeletons, are programmed using Hash configurations, while programming of simple components, in Haskell, is described in Section 2.2. 3 # Programming Message Passing Programming (Efficient, Portable, Expressive) (Structured, High Level of Abstraction) # Program recv! recv! send! # compiler − code generation − # compiler − unfolding − send! computable computable computable send! send! send! send! recv! recv! recv! send! send! non−computable Unit send! Process # Process Process View Component View Process View concerns are scattered among processes,but it is possible to factorize them into components concerns are encapsulated into components send! recv! send! Process Slice Port Channel recv! recv! requires programmer intervention Component recv! recv! recv! concerns are scattered among processes and it is not possible to factorize them into components Figure 1: Component Perspective versus Process Perspective 2.1 Programming Composed Components Composed components define coordination media of Haskell# programs, where all parallelism concerns are addresses without mention to entities at computation levels, where computations are specified. Composed components are written in HCL (Hash Configuration Language). Also, they define the core of the Hash component model, supported by Haskell#. The configuration of a composed component specifies how a collection units, agents that perform specific tasks, interact by means of point-to-point, typed, and unidirectional communication channels, for addressing a given parallel programming concern. For that, a unit is instantiated from an interface and associated to a component (Figure 2). The latter specifies the task performed by the unit, since functional modules describe addressed concerns, while the former specifies how the unit interacts with the coordination medium. A unit associated with a simple component is called process, while a unit associated with a composed component is called a cluster. A interface of a unit is defined by a set of typed input and output ports and a protocol. The protocol of an interface specifies the order in which ports may be activated during execution of the units instantiated from that interface, by means of an embedded language whose constructors have semantic equivalence to regular expressions controlled by semaphores. This formalism is equivalent to place/transition Petri nets, allowing for formal property analysis, simulation and performance evaluation of programs using available Petri net tools [10], such as PEP [16] and INA [21]. A port is activated between the time in which it becomes ready to perform a Assignmennt Instantiation interface declaration Unit (non−virtual) Virtual Unit Interface or Interface Class unit declaration assign declaration Figure 2: Instantiating and Configuring a Unit 4 A C e b F d F unification A CDE B D a e a e b G d B b G c factorization E c Figure 3: Unification/Factorization communication operation and the time it completes this operation, according to the communication mode of the channel where the port is connected. Units interact through communication channels, which connect an output port from an unit (transmitter ) to an input port of another one (receiver ). The types of the connected ports must be the same. The supported communication modes, inspired on MPI, are : synchronous, buffered and ready. In a unit specification, interface ports can be replicated to form groups. Groups may be of two kinds: any and all, according to the semantics of activation. The activation of a group of ports of kind all implies activation of all its port members. The activation of a group of ports of kind any implies that one of the port members will be put ready for communication but only one will complete communication. The chosen port is one of the activated ports whose communication pairs are also activated at that instant. From the internal perspective of the unit, groups are treated as indivisible entities, while from the perspective of the coordination medium, port members are referred directly in order to forming channels. Input and output ports (groups individually) of the unit interface must be mapped to arguments and return points of the component assigned to it, respectively. Wire functions are useful when it is necessary to transform values at the boundary between ports and arguments/exit points. A particularly useful use of wire functions is to aggregate data received from input ports belonging to a group of ports of kind all to a unique value, passed to the associated argument. Similarly, wire functions allow that a value produced in an exit point to be mapped onto a collection of values in order to be sent by the port members of the associated group of output ports of kind all. Wire functions A A[1] B B[1] A[2] replication C C D D Figure 4: An Illustrative Example of Replication 5 B[2] TORUS<3> SQMATMULT X9 PIPELINE<3> overlapping FARM<9> topological reorganization O V E R L A P P I N G n w MMShift ab e c s Figure 5: Topology of Matrix Multiplication Using a Torus increases the changes for reusing a component, resolving possible conflicts. Two operations are defined on units: unification and factorization. Unification allows to unify a collection of units, forming a single unit. Factorization is the inverse of unification, allowing units to be divided into many virtual units. Replication, a third operation applied to units, allows that the network induced by a collection of units to be replicated. All operations assume that units are fully connected. Behavioral and connectivity preserving restrictions are applied, but not formalized here. Connectivity restrictions imply the possibility of to replicate ports whenever it is necessary to adjust topological connectivity after an operation. Figures 3 and 4 present illustrative examples of these operations. 2.1.1 Virtual Units and Skeletons To allow overlapping of components and the support for skeletons, the notion of virtual unit has been introduced. A unit is virtual whenever there is no component associated with it. In other terms, the task performed by a virtual unit is not defined. Components are partially parameterized in its addressed concern by means of placing virtual units in its constitution. A component that comprises at least one virtual unit is called an abstract component, or a partial topological skeleton. These terms are used as synonyms. When a programmer re-use an abstract component for specification of a Hash program, it must to assign components to the virtual units comprising it. Abstract components may not instantiate applications. It is necessary to describe the computation performed by their constituent virtual units. The assignment operation is used allows to associate a component to a virtual unit, making it a non-virtual unit. Also, there is a superseding operation, which allows to take a non-virtual unit for replacing a virtual unit of the topology. The behavioural compatibility restrictions from the non-virtual unit to the replaced virtual unit guarantees that any sequence of communication actions that is 6 ... X9= ... ... ... ... ... P i p e − L i n e ... Figure 6: Composing Pipe-Lines to Form a Systolic Mesh valid in the non-virtual unit remains valid in the virtual unit. The superseding operation is a “syntactic sugar” of HCL, since it may be implemented using unification and assignment. Figure 6 shows how a skeleton describing a systolic mesh of processes, implemented by overlapping a collection of pipe-line skeleton instances. The abstract components PipeLine is used to describe interaction of processes placed at the mesh lines and columns. Each unit in SystolicMesh abstract component is formed by two slices: one described by the unit that comes from the vertical PipeLine component and the other described by the unit that comes from the horizontal one. 2.2 Programming Simple Components Simple components, also called functional modules, are atoms of functionalities in Hash programming. The collection of simple components in a Hash program describes its computation media. Simple components might be programmed virtually in any general purpose language, called host language. For that, it is needed to define which host language constructions correspond to the arguments and return points of the underlying functional module. It is preferred that no extensions to the host language be necessary for this purpose, keeping transparency between coordination and computation media. Simple components may be overlapped when configuring composed components. The goal is to implement a really multi-lingual approach for parallel programming. For that, it has been proposed to use CCA (Common Component Architecture) [2], a recent standard proposed for integrating components written in different languages in a parallel environment. Another possibility is to use heterogeneous implementations of MPI [24, 23], recently proposed, facilitating this task since, in cluster environments, Hash programs are compiled to MPI. Since the translation schema onto Petri nets is defined on top of coordination 7 component CPipeLine <N> with component SqMatMult<N> where iterator i range [1,N] interface ICPipe where ports: i* → o* protocol: repeat seq{o!; i?} until <o & i> [/ unit pipe[i] where ports: ICPipe /] iterator i, j range [1,N] use Skeletons.Common.{Torus, Farm} use MMShift interface ISqMatMult where ports: j → r # ITorus protocol: seq {j?; repeat seq {s!;e!;n?;w?} counter N } connect pipe[i]→o to pipe[i+1]←i, buffered component Torus <N> with unit mm torus; assign Torus<N> to mm torus unit mm farm; assign Farm<N> to mm farm use Skeletons.Common.CPipeLine iterator i, j range [1,N] interface ITorus where ports: ICPipe @ n → s # ICPipe @ e → w protocol: repeat seq {par {s!; w!}; par {n?; e?}} until <n & e & s & w> [/ unit vpipe[i]; assign CPipeLine<N> to vpipe[i] /] [/ unit hpipe[j]; assign CPipeLine<N> to hpipe[j] /] [/ unify farm.worker[i + j × N ], torus.node[i][j] to sqmm[i][j] /] where ports: ISqMatMult unify farm.distributor, farm.collector, sqmm[0][0] to sqmm root where ports: () → ab # c → () # ISqMatMult @ mm protocol: seq {ab!; c?; do mm } [/ assign MMShift to sqmm[i][j] /] [/ unify vpipe[i].pipe[j], hpile[j].pipe[i] to node[i][j] where ports: ITorus /] module MMShift(main) where component Farm<N> with main :: Num t ⇒ t → t → [t] → [t] → ([t],[t],t) main a b as i bs i = (as o,bs o,c) where c = matmult as o bs o (as o, bs o) = (a:as i, b:bs i) unit distributor where ports: () → job unit worker where ports: job → result protocol: seq {job?; result!} unit collector where ports: result → () connect distributor.job to worker.job, synchronous connect worker.result to collector.result, synchronous matmult :: Num t ⇒ t → [t] → [t] → t matmult [] [] = 0 matmult (a:as) (b:bs) = ab + matmult as bs replicate N: worker Figure 7: Configuration Code of Matrix Multiplication on a Torus module Tracking(main) where import Track import Tallies import Mcp types main :: User spec info → [(Particle,Seed)] → ([[Event]],[Int]) main user info particle list = let events’s = map f particle list in (events’s, tally bal event lists) where f (particle@( , , , e, ), sd) = (Create source e):(track user info particle [] sd) Figure 8: A Functional Module from MCP-Haskell# 8 <Component> → component ({u1 :<Unit>, . . ., ui :<unit>}, {c1 :<Channel>, . . .,cj :<Channel>}) <Unit> → unit (id, {p1 :<Port>, . . ., pk :<Port>}, <Typeu >) <Typeu > → repetitive \| non-repetitive <Behavior> → protocol ({idsem , . . . , idsem }, <Action>) n 1 <Action> → \| \| \| \| \| \| \| \| \| skip seq {a1 :<Action>; . . .; ak :<Action>} par {a1 :<Action>; . . .; ak :<Action>} alt {g1 :<Action>; . . .; gk :<Action>} repeat until (<Action>, C) repeat counter (<Action>, N) repeat forever <Action> signal id wait id activate id <Port> → port (id,<Direction>,<Multiplicity>,<Typep >,nesting f actor) <Multiplicity> → single \| group (<Typeg >, {p1 :<Port>, . . ., pn :<Port>}) <Direction> → input \| output <Typep > → stream \| non-stream <Typeg > → any \| all <Channel> → connect (idport , idport , <ChanMode>) <ChanMode> → synchronous \| buffered \| ready Figure 9: Abstract Hash Configuration Language Syntax media, abstracting from computation media concerns, specific details about programming simple components is not provided in this paper. For illustration, Figure 8 presents an example of functional module from MCP-Haskell# program [9], written in Haskell. 2.3 An Abstract Representation for Hash Components In Figure 2.2, it is defined a simplified syntax for an abstract representation of Hash configurations, named Abstract Hash. The abstract Hash configuration language (AHCL) only captures information strictly relevant to the translation schema onto Petri nets further presented. For example, interface declarations and operations over units, such as unifications, factorizations, replications, and assign operation are not represented in AHCL. It is supposed that these operations are all resolved before translation process. AHCL is not a simplification. Indeed, in the compilation process of Hash configurations, AHCL correspond to the intermediate code generated by the front-end compiler module, which serves as input to all back-end modules. A back-end was developed for generating PNML (Petri Net Markup Language) code from a Hash configuration. The following paragraph describes the structure of AHCL. A component is composed by a set of units (u1 , . . . , ui ) and a set of channels (c1 , . . . , cj ). Units are described by an identifier, a collection of ports (< Port >). A unit can be repetitive or non-repetitive (< Typeu >). The interface of a unit is defined by a collection of ports and a protocol, described by means of an embedded language that specifies valid orders for activation of ports. This language have constructors equivalent to combinators of regular 9 <COMPONENT> unidades repetitivas <UNIT> P2 ... P4 T2 ... P6 P3 T4 <UNIT> P2 ... P4 T2 P5 P3 (P1) PROGRAM_BEGIN <UNIT> T1 P2 ... (P2) PROCESS_STARTED[proc_i] (P3) PROCESS_RESTART_ENABLED[proc_i] P4 (P4) PROCESS_FINISHED[proc_i] (P5) PROGRAM_END_READY ... ~ unidades nao−repetitivas P1 (P6) PROGRAM_END T3 <UNIT> P2 ... (T1) T.PROCESSES_FORK (T2) T.PROCESS_RESTART[proc_i] P4 (T3) T.PROCESSES_NR_JOIN (T4) T.PROCESSES_ALL_JOIN Figure 10: Translating a Component expressions controlled by balanced semaphores. A port is described by an identifier (id), a direction (< Direction >), a type (stream or non-stream) and a nesting factor. Additionally, port multiplicity specifies if a port is a single port or a group of ports. Notice that a group of port can be of two types: any or all. All port identifiers are assumed to be distinct in an abstract Hash programs. A channel connect two ports and is associated to a mode (synchronous, buffered and ready). Notice that, abstract Hash syntax does not force that these two ports be from opposite directions. This restriction is implicitly assumed. 3 Translating Hash programs into Petri Nets In this section, a schema for translating Hash programs into Petri nets is introduced. In order to make translation schema easier to understand, it is informally described using diagrams. The possibility of making intuitive visual descriptions is an interesting feature of Petri nets. The translation schema is specified inductively in the hierarchy of components at the coordination medium of the Hash program. Thus, simple components are ignored. The overall steps in the translation procedure of a Hash program into an interlaced Petri net are: 1. Translating units: For each unit comprising the component (unit declarations), its interface is used for yielding an interlaced Petri net describing activation order of their interface ports. In Hash configuration language, it is defined by an embedded language, in interface declarations, whose combinators have correspondence to operators of regular expression controlled by semaphores. This formalism was proved to have expressiveness 10 Interface or Interface Class Assignmennt Instantiation interface declaration Unit (non−virtual) Virtual Unit unit declaration assign declaration Figure 11: Unit = Interface (by instantiation) + Component (by assignment) equivalence to Petri nets according to formal language theory. If a unit is a cluster (a composed component is assigned to the unit), it is necessary to generate the Petri net that corresponds to the assigned component. Using information about mapping of argument/return points to input/output ports of the unit, it is possible to synchronize behavior of the unit with behavior of the component, in such way that they are compatible; 2. Synchronize units. Now that a Petri net exists for describing communication behavior (traces) of each unit, communication channels (connect declarations) may be used for coordinating synchronized behavior of these Petri nets (units); 3. Synchronize streams. For each port carrying a stream, a Petri net describing a protocol for stream synchronization is overlapped with the Petri net produced in the last step. The semantics of stream communication is described separately because it increases complexity of the generated interlaced Petri net, making computationally hard its analysis. Thus, in Hash programming environment, the programmer may decide not to include stream synchronization protocol. Obviously, information may be lost, but it may not be necessary for some useful analysis. The next sections provide details about the above translation steps. Also, it is discussed how higher-level information encompassed in skeletons may be used to simplify the generated network. 3.1 Modelling Components (ΥC ) In Figure 10, the Petri net resulted from application of translation function ΥC to the configuration of a component is illustrated. The translation function ΥU is applied to each unit comprising the component, generating a Petri net that describes its communication behavior. The resulting Petri nets are connected in order to model parallel execution of units. The places process started[j] and process finished[j], 1 ≤ j ≤ n, where n is the number of units, correspond to start places and stop places of the Petri nets modelling units, respectively. When a token is placed on process started[j], the unit j is ready to initiate execution, and when a token is placed on process finished[j], the unit j had 11 (T1) GROUP_PREPARE[g] (P2) PORT_COMPLETE[p] (T2) GROUP_COMPLETE[g] ... (P4) GROUP_COMPLETE[g] (T3) PORT_SEND[p] / PORT_RECV[p] ] (P3) GROUP_PREPARED[g] a[n (P1) PORT_PREPARED[p] ] a[2 ] a[1 a ANY ALL Group of Ports P1 T3 P2 T3 P2 T1 T2 P3 P2 P1 T2 P4 ... P4 P2 P1 T3 ... T3 P1 ... P3 <PORT> group any P2 ... T1 P2 P1 port ... T3 P1 T3 ... Individual Port <PORT> group all T3 <PORT> single P1 P2 Figure 12: Translating Unit’s Ports finished. The transitions process restart[k], 1 ≤ k ≤ r, where r is the number of repetitive units, allows that repetitive units return back to its initial state after finalization. They are introduced in the Petri net generated for each unit. The place program end ready receives a mark when all non-repetitive units terminates. In this case, the tokens in places process restart enabled[k], 1 ≤ k ≤ r, are removed, preventing repetitive processes execute again. At this state, the program terminates after all repetitive processes also terminate, which causes transition processes all join to be fired and a token to be deposited in program end. 3.2 Modelling Units (ΥU ) This section intends to describe how individual units are translated into Petri nets. Firstly, Figure 12 presents interlaced Petri nets that model the activation of the ports of the interface of a unit. A mark in place port prepared[p] indicates that the port p is prepared for communication. The firing of transition port send[p]/port recv[p] models communication, causing the deposit of a mark in place port complete[p] for indicating that communication has been completed on port p. It is possible that two ports be active at the same time. In groups of ports, the places group prepare[g] and group complete[g] are connected to places port prepared[p] and port complete[p] of each port p belonging to the group g that model local preparation and completion of communication in ports belonging to the group, according to their semantics: any or all. In groups of ports of kind all, all individual ports are activated in consequence of activation of the group. In groups of ports of kind any, only one port is chosen among the ports that are ready for communication completion. For obeying this semantic restriction, the firing of a transition port send[p]/port recv[p] of a port p in group g of kind any causes removal of the marks in places port prepared[p′ ], for all places p′ belonging to g, in 12 <ACTION> skip p Figure 13: Translating “Null Action” such a way that all ports p, such that p 6= p, cannot complete communication (firing of transition port send[p]/port recv[p]). The order in which marks are placed in places port prepared[p], for any p, is controlled by an interlaced Petri net that models the protocol of the unit. The following sections discusses how primitive actions and action combinators of behavior expressions are translated into Petri nets. The primitive actions are: skip (null action), wait (increment semaphore primitive), signal (decrement semaphore primitive), ? (activation of input port), and ! (activation of output ports). The combinators of actions are: seq (sequential actions), par (concurrent actions), alt (non-deterministic choice among n actions), and if (conditional choice between two actions). 3.2.1 Null Action (skip) The skip combinator have no communication effect. Because that, it is known as the “null action”. Only one place is needed, where it is both start and stop place (Figure 13) of the interlaced Petri net generated. 3.2.2 Sequencing (seq) The seq combinator describes a total ordering for execution of a set of actions (sequential execution), represented by a1 , a2 , . . . , an . It may be modelled by sequential composition of the Petri nets induced for each action (Figure 14). 3.2.3 Concurrency (par) The par combinator describes a concurrent (interleaving) execution of a set of actions, represented by a1 , a2 , . . . , an . It may be modelled by parallel composi<ACTION> seq <ACTION> T1 <ACTION> T1 T1 (T1) SEQ_NEXT_ACTION[seq_label, i] Figure 14: Sequence of Actions 13 <ACTION> <ACTION> par (P1) PAR_BEGIN[par_label] <ACTION> (P2) PAR_END[par_label] T1 T2 <ACTION> P1 P2 <ACTION> (T1) PAR_FORK_ACTIONS[par_label] (T2) PAR_JOIN_ACTIONS[par_label] Figure 15: Interleaving (Concurrency) Among Actions tion of the Petri nets induced by each action (Figure 15). 3.2.4 Non-Deterministic Choice (alt) The alt combinator describes a conceptually non-deterministic choice among a set of actions, represented by a1 , a2 , . . . , an . It may be modelled by composing the Petri nets induced for each action using a conflict in place alt begin[alt label], its start place (Figure 16). The firing of transitions alt select branch[alt label, i], 1 ≤ i ≤ n, models the choice. 3.2.5 Streams and Conditions for Checking Stream Termination The two combinators modelled in the next sections, repeat and if, requires testing a condition in order to choose the next action to be performed. The condition is defined by a logical predicate in its disjunctive normal form (DNF). <ACTION> alt <ACTION> T2 T1 (P1) ALT_BEGIN[alt_label] <ACTION> P2 P1 T2 ... T1 (P2) ALT_END[alt_label] (T1) ALT_SELECT_BRANCH[alt_label,i] (T2) ALT_END_ACTION[alt_label,i] <ACTION> T1 T2 Figure 16: Choice Among Actions 14 Logical variables are references to ports that carry streams. This section attempts to formalize the notion of streams in the Hash component model and valuation of logical variables of termination-test conditions of streams. A stream is defined as a sequence of semantically related data items, terminated with a special mark, that is transmitted through a channel. By making an analogy with conventional message passing programming using MPI, a trivial example of stream is a sequence of data items transmitted in calls to a specific occurrence of MPI Send primitive in the context of an iteration. At each iteration, an item of the stream is transmitted. The termination of the iteration is modelled, in the Hash program, by the end mark carried by the stream. Communication on channels that carry streams may be implemented using persistent communication objects in the underlying messaging passing library, which may reduce communication overhead. Hash streams may be nested. Streams of streams, at any nesting depth may (1,d) be defined. A stream Sx , where d is the nesting factor of the stream and x is a positive integer that indicates the order of a nested stream in a stream, may be defined as following: (i,d) Sx (i+1,d) = hS1 VALUE (i+1,d) , S2 (i+1,d) , . . . , Sn , EOS[i]i, n ≥ 0 , if i < d , if i = d (1) where VALUE is a data item and EOS[i] is a termination value at nesting level i. Notice that termination values should carry an integer indicating the nesting level of the stream being terminated. The feature of nested streams appeared in consequence of design of Haskell# . Streams at coordination level must be associated to lazy lists in computation media 1 . The experience with Haskell# programming have shown that laziness of Haskell nested lists may be useful in some applications. This feature is analogous to communication operations that occurs in context of nested iterations in MPI parallel programming. In Hash configuration language, streams are declared by placing “” symbols after the identifier of a port in the declaration of interfaces. The number of “”’s indicates the nesting factor of the stream carried by the port. Only ports carrying streams of the same nesting levels may be connected through a channel. It is defined that a port that transmit a single value (non-streamed) have nesting level zero. Now that the notion of streams is defined, it is possible to define the syntax and semantics of predicates for testing synchronized termination of streams, a necessary feature for combinators if and repeat. This kind of predicate will be referred as stream predicate. Syntactically, a stream predicate is a logical predicate in its disjunctive normal form. The logical operators supported are “&”, the logical and, and “\|”, the logical or. Disjunctions may be enclosed by “h” and “i” delimiters. Logical variables are references to interface ports of a unit. The formal syntax of stream predicates is shown below: 1 In Haskell# , simple components are functional modules written in Haskell 15 stream predicate → sync conjunction 1 ‘\|’ . . . ‘\|’ sync conjunction n (n ≥ 1) sync conjunction → ‘h’ simple conjunction ‘i’ \| simple conjunction simple conjunction → port id \| ( port id1 ‘&’ . . . ‘&’ port idn ) (n ≥ 1) Let d be the depth of nesting of an occurrence R of a repeat or if combinator in relation to an outermost occurrence, if it exists (d = 0, if it does not exists). Only port carrying streams with nesting factor equal or less than d can appear in the termination condition of R. It is now possible to define semantics of stream predicates, by defining how values of logical variables may be inferred in execution of units. For instance, let p be a port carrying stream S (1,n) . Its value is false whenever a data value (VALUE) or an ending value at nesting level i (EOS[i]), such that i < d, was transmitted (sent or received) in its last activation. Otherwise, it is true. The value of the stream predicate may evaluate to true, false or fail. A value true is obtained by evaluating the stream predicate ignoring semantics of angle brackets delimiters. A fail is obtained if negation of some conjunction enclosed in angle brackets evaluates to true, assuming the following identity that defines angle brackets semantics: ¬ha(k,r) ∧ a(k,r+1) ∧ . . . ∧ a(k,s) i = ¬a(k,r) ∧ ¬a(k,r+1) ∧ . . . ∧ ¬a(k,s) (2) If stream predicate evaluates neither to true or error, the value of the stream predicate is false. Angle brackets delimiters are used for ensuring synchronization of the nature of values transmitted by streams, whenever necessary. In order to make possible to model test of stream predicates with Petri nets, it is firstly necessary to model stream communication by using this formalism. Particularly, it is necessary to introduce, in the Petri net of the Hash program, places that can remember the kind of value transmitted in the last activation of ports that carry streams. In a Hash program, for each port s that carries a stream with nesting factor n, there are two sets of n + 1 places, here referred as Stream Flags(s) and Stream Flags(s): Stream Flags(s) = {stream port flag[s, k] \| 0 ≤ k ≤ n} Stream Flags(s) = {stream port flag dual[s, k] \| 0 ≤ k ≤ n} For some stream port s, the places in set Stream Flags(s) form a split binary semaphore: X M (p) = 1 p∈Stream Flags(s) Also, they are mutually exclusive with its corresponding places in Stream Flags(s): ∀k : 0 ≤ k ≤ n : M (stream port flag[s, k]) + M (stream port flag dual[s, k]) = 1 For a port s carrying a stream with nesting factor n, the places Stream Flags(s) and Stream Flags(s) are used to remember which kind of value was transmitted in the last activation of s. There are n + 1 possibilities: 16 P10 ... T5 ... P5 P10 P10 T7 P10 P10 T7 ... ... T5 P10 ... P10 ... P10 (P1) RU_CHECKING_CONDITIONS[r] (P2) RU_CHECKING_CONDITION_L[r] (P3) RU_CHECKING_CONDITION_T[r] (P4) RU_CHECKING_CONDITION_F[r] (P5) RU_SATISFIED_CONDITION_L[r] (P6) RU_SATISFIED_CONDITION_T[r] (P7) RU_SATISFIED_CONDITION_F[r] (P8) RU_TERMINATED (P9) RU_FAILED (P10) STREAM_PORT_FLAG[p,i] P2 P7 P4 T4 P9 <ACTION> repeat ... until T3 <ACTION> P8 P1 T1 T2 P3 P10 T6 P10 ... T6 ... P10 P6 ... (T1) RU_CHECK_CONDITION[r] (T2) RU_TERMINATE[r] (T3) RU_FAIL[r] (T4) RU_LOOP[r] (T5) RU_SATISFY_CONJUNCTION_L[r,i] (T6) RU_SATISFY_CONJUNCTION_T[r,i] (T7) RU_SATISFY_CONJUNCTION_F[r,i] P10 Figure 17: Stream Controlled Repetition • An ending marker at nesting level i (EOS[i]), for 0 ≤ i ≤ n − 1; • A data value. The places stream port flag[s,k], such that 0 ≤ k ≤ n − 1, are respectively associated to ending marks EOS[k] of a port s carrying a stream with nesting factor n. The place stream port flag[s,n] is associated to a data value. Assuming the restrictions above, if there is a mark on the stream port flag[s,k] place, then a value of its corresponding kind was transmitted in the last activation of the port. All the above restrictions are guaranteed by the Petri net protocol for synchronization of streams introduced further, in Section 3.4. The next two sections present respectively how to model repeat and if combinators, assuming the existence of sets of places Stream Flags(s) and Stream Flags(s). 3.2.6 Repetition Controlled by Stream Predicates The repeat combinator is used to model repeated execution of an action. The termination condition may be provided in a until or a counter clause. The later will be introduced in the next section. The former uses a stream predicate for testing termination of a repetition after each iteration. In the Petri net modelling of repeat combinator with until clause, it is assumed existence of the sub-network modelling stream communication introduced in Section 3.2.5. In Figure 17, it is illustrated the Petri net resulted from translation of repeat combinator with an until clause to check termination. The conflict in place ru checking conditions models the decision on to terminate or not the 17 <ACTION> repeat ... counter T1 P2 (P1) RC_ENTERING[r] (P2) RC_REMAINING[r] (P3) RC_PREPARED[r] (P4) RC_PERFORMED[r] (P5) RC_EXITING[r] T2 n P1 P3 <ACTION> (T1) RC_ENTER[r] (T2) RC_START_ACTION[r] (T3) RC_FINISH_ACTION[r] (T4) RC_EXIT[r] P5 T4 n T3 P4 Figure 18: Repeated Action by a Fixed Number of Times <ACTION> repeat ... forever <ACTION> RI_REPEAT[r] Figure 19: Infinite Repetition (repeat) 18 P11 ... T7 ... P5 P11 P11 T6 P11 P11 P8 <ACTION> if ... then ... else ... T6 ... ... T7 P11 ... P10 P11 ... P11 P7 P4 T3 <ACTION> then T2 P9 P1 <ACTION> else P2 T1 T4 P6 P11 T5 P11 ... T5 ... P11 P3 ... (P1) IF_STARTED[r] (P2) IF_CHECKING_CONDITIONS[r] (P3) IF_CHECKING_CONDITION_E[r] (P4) IF_CHECKING_CONDITION_T[r] (P5) IF_CHECKING_CONDITION_F[r] (P6) IF_SATISFIED_CONDITION_E[r] (P7) IF_SATISFIED_CONDITION_T[r] (P8) IF_SATISFIED_CONDITION_F[r] (P9) IF_TERMINATED (P10) IF_FAILED (P11) STREAM_PORT_FLAG[p,i] P11 (T1) IF_CHECK_CONDITION[r] (T2) IF_THEN[r] (T3) IF_FAIL[r] (T4) IF_ELSE[r] (T5) IF_SATISFY_CONJUNCTION_E[r,i] (T6) IF_SATISFY_CONJUNCTION_T[r,i] (T7) IF_SATISFY_CONJUNCTION_F[r,i] Figure 20: Conditional iteration. Values of logical variables corresponding to ports in the stream predicate are tested using their respective Stream Flags(s) and Stream Flags(s) places. The arrangement of places and transitions in Figure 17 allows for testing the value of stream predicates at each iteration. The mutual exclusive firing of transitions ru terminate, ru fail and ru loop correspond, respectively, to values true (execute one more iteration), false (termination) and error (abort the program) for the stream predicate. 3.2.7 Bounded Repetition The combinator repeat with termination condition defined by a counter clause models repeated execution of an action by a fixed number of times. Its translation into Petri nets is illustrated in Figure 18. For a certain fixed bounded repetition r, the weight of arcs (rc enter[r], rc remaining[r]) and (rc performed[r], rc exit[r]) (n) define the number of repetitions of < ACT ION >. 3.2.8 Infinite Repetition Whenever no termination condition is defined for a given occurrence of the repeat combinator, the given action is repeated infinitely. The Petri net that models this kind of repeat combinator is illustrated in Figure 19. 3.2.9 Conditional Choice (if ) The if combinator describes a conditional choice between to two actions. Its translation into Petri nets is illustrated in Figure 20. The reader may notice 19 (P4) SEM_WAIT_BEGIN (P1) SEM_SIGNAL_BEGIN <ACTION> signal (P2) SEM_SIGNAL_END (P5) SEM_WAIT_END P3 P3 (T2) SEM_WAIT (T1) SEM_SIGNAL <ACTION> wait (P3) SEM_COUNTER P1 P2 P5 P4 T1 T2 Figure 21: signal (left) and wait (right) Semaphore Primitives the analogy between the construction of the Petri net in this diagram with the Petri net generated for the repeat ... until combinator, mainly concerning the test of the stream predicate. 3.2.10 Semaphore Primitives The wait and signal (balanced counter) semaphore primitives and the par combinator make behavior expressions comparable to labelled Petri nets in descriptive power [18]. In absence of semaphore primitives, only regular patterns of unit behavior could be described. The semantic of semaphore primitives is now defined using the notation introduced in [1] for concurrent synchronization: wait(s): <await s > 0 → s := s − 1> signal(s): < s := s + 1 > The angle brackets model mutual exclusion (atomic actions in concurrent processes), while the await statement models condition synchronization. The wait primitive causes a process to delay until the value of the semaphore s is greater than zero in order to decrement its value. Thus, the value of a semaphore is greater than zero in any instant of execution of a unit. In concurrent systems where synchronization is controlled by balanced counter semaphores, like in behavior expressions here defined, the value of a semaphore must be the same at the initial and final states of the system. The signal and wait primitives are modelled using Petri nets as described in Figure 21. Given a semaphore s, the number of marks in place sem counter[s] models the semaphore value at a given state. 3.2.11 Port Activation The Hash primitive actions ! and ? models respectively send and receive primitives in message passing programming. In Hash programs, they cause activation of groups of ports and individual ports that are not member of any group. In Section 3.2, the Petri net slice that model individual ports and groups of ports is illustrated in Figure 12. An individual port p is prepared for communication whenever a mark is deposited in place port prepared[p]. A group of ports 20 of kind all is prepared whenever all of its ports are prepared, while a group of kind any is prepared whenever any of its ports are prepared. The communication is completed whenever a place is deposited in place port complete[p] or group complete[p]. The activation a of a port p is defined as the time between preparation of a port and completion of communication. The format of Petri net slices induced by translation of occurrences of primitives ! and ? is illustrated in Figure 22. The firing of transition activate start[a] prepares port p for communication and firing of transition activate stop[a] occurs whenever the port completes communication. Notice that whenever a mark is deposited in place activate on[p], the port p is active. 3.3 Communication between Units In the translation of a component, the Petri net slices resulted from translation of their units form an interlaced Petri net that models their asynchronous execution. Information regarding synchronization of units, by means of communication channels, is not yet included. In Figure 23, Petri net slices that model, respectively, the three kinds of channels that may occur in a Hash program (synchronous, buffered and ready) are presented. The translation function ΥL is applied to each communication channel in a component, generating Petri net slices according to the translation schema illustrated in Figure 23. These Petri net slices are overlapped with the Petri net slices that models behavior of units in order to model synchronous execution of the units. For synchronous channels, communication pairs must be active at the same time for communication to complete. For implementing that, it is only necessary to unify the respective transitions modelling completion of communication from the respective pairs. For buffered channels (bounded buffers), the sender does not need to wait completion of communication operation for resuming execution. For that, whenever the sender port s in a channel c is activated, transition port send[s] must be activated if there is a mark in <ACTION> activate PORT_PREPARED[port] PORT_COMPLETE[port] T2 T1 P3 P2 P1 (P1) ACTIVATE_ENTERING[a] (P2) ACTIVATE_ON[a] (P3)ACTIVATE_EXITING[a] from translation of ports (T1) ACTIVATE_START[a] (T2) ACTIVATE_FINISH[a] Figure 22: Activation of Ports 21 S <CHAN NEL> P1 P2 P1 <CHAN NEL> P1 P2 T3 T1 #n P1 P2 P1 P5 #1 P4 P2 (P1) PORT_PREPARED[p] (P2) PORT_COMPLETE[p] T3/T6 T5 P1 P2 <mode> = ready <mode> = buffered n <mode> = synchronous (P5) CHAN_READY_IS_OPEN[c] (P3) CHAN_BUFFER_FREE[c] (T3) CHAN_SYNCHRONIZE[c] (T1) PORT_SEND[p] (T2) PORT_RECV[p] P3 P2 T4 T2 T2 P2 P1 T1 connect a.o to b.i <mode> P1 <CHAN NEL> P2 (P4) CHAN_BUFFER_USED[c] T3 = fusion of T1 and T2 (T4) CHAN_READY_SENDER_CLOSE[c] (T5) CHAN_READY_RECEIVER_TRY_ENTER[c] (T6) CHAN_READY_SENDER_OPEN[c] from translation of ports Figure 23: Modelling Communication Channels place chan buffer free[c], which models the number of empty slots in the buffer. The place chan buffer used[c], which receives a mark after activation of port send[s], models the number of used slots in the buffer. Notice that the sender blocks whenever there is no empty slots in the buffer. Channels supporting ready mode require a more complex protocol. The place chan ready is open[c] ensures that communication proceeds only if activation of the sender port precedes activation of the receiver. Notice that whenever receiver is activated before sender, the sender cannot proceed, causing a deadlock that may be detected by a Petri net verification tool. Using this approach, it is possible to verify, for example, if a certain parallel program using ready channels may fail in some program state during execution. Ready communication mode may improve communication performance of MPI programs, but unfortunately it is hard to ensure that communication semantics is safe in arbitrary parallel programs. The modelling of communication semantics with Petri nets may overcome these difficulties for debugging. 3.4 Synchronization of Streams Protocol In Section 3.2.5, it was introduced two sets of places that must exist for each stream port s: Stream Flags(s) and Stream Flags(s). Additionally, restrictions were introduced for their markings. They allow to check the kind of the transmitted value in the last activation of the port, making possible to check stream termination conditions that occurs in repeat and if combinators. This section presents a protocol for updating marking of places Stream Flags(s) and Stream Flags(s) in such a way that restrictions introduced in Section 3.2.5 are obeyed. In Figure 24, it is illustrated how the network presented in Figure 22 (activation of ports) may be enriched in order to introduce a protocol for updating places in Stream Flags(s) and Stream Flags(s), for an arbitrary stream port s with nesting factor n. The Petri net slice introduced have the transitions sp clear flag[s,i] e sp set flag[s,i], for 0 ≤ i ≤ n as its main components. They are arranged in such a way that transitions inside the sets are mutually ex- 22 P6’s Output Port/Group of Ports P6’s Input Port/Group of Ports #1 ... ... #1 T2’s T4’s #1 T2’s T5 T3 P2 P3 P4 P5 ... ... P3 P2 T4’s #1 T5 T3 P4 P5 ... ... ... ... #1 #1 ... ... P7’s P7’s The receiver may decide which value has been received only after completion of communication Before the activation, the sender may decide which kind of value is going to be transmitted (P2) SP_CLEARING_FLAGS[s] (P3) SP_CLEARED_FLAGS[s] (P4) SP_SETTING_FLAGS[s] (P5) SP_EXITING[s] (P6) STREAM_PORT_FLAG[s,i] (P7) STREAM_PORT_FLAG_DUAL[s,i] (T2) SP_CLEAR_FLAG[s,i] (T3) SP_START_SET_FLAGS[s] (T4) SP_SET_FLAG[s,i] (T5) SP_PREPARE_PORT[s] Figure 24: Activation of Stream Ports clusive. The firing of a transition in sp clear flag[s,i] clears the set of places Stream Flags(s), by moving the mark in the corresponding cleared place to the corresponding place in set Stream Flags(s). After that, all places in Stream Flags(s) have exactly one mark, while all places in Stream Flags(s) have zero mark. In sequence, one transition of set sp set flag[s,i] is fired causing the moving of a mark from one of the places in Stream Flags(s), chosen non-deterministically, to the corresponding place in Stream Flags(s). This sequence of actions models the test of the kind of the value transmitted in the current activation of the port. Notice that the moment of the choice is different for input and output ports. An input port only updates its Stream Flags(s) after communication has been completed. This is in accordance with implementation semantics, once only after receiving the value, the receiver may check the kind of the transmitted value. 3.4.1 Ensuring Consistency of Communication Channel communication semantics imposes that the kind of value transmitted by the sender in a given activation of a stream port is the same as the kind of the value that the receiver receives in the corresponding activation. In Figure 25, it is shown how this restriction is ensured for channels with synchronous and ready mode of communication. For the following description, consider that individual ports are groups containing only one port. Thus, consider a channel c connecting a sender stream port s and a receiver stream port r, both with nesting factor n. The groups where s and r are contained are respectively gs and gr . For each i, 0 ≤ i ≤ n, it is necessary to create an arc that links the place stream port flag[gs ,i] to the transition sp set flag[gr ,i], in such way that consistency of Stream Flags(s) and Stream Flags(r) is forced after com- 23 #1 ... #1 ... ... ... ... #1 Sender Side ... #1 ... Receiver Side #1 ... ... ... ... Synchronous/Ready Channel #1 ... The kind of the received value must be the same kind of the sent value Figure 25: Ensuring the Consistency of Streams Connected through Synchronous Channels munication completion. If two different senders, connected to ports belonging to a group gr of kind all, decide to transmit values of different nesting levels in a give activation of gr , a deadlock occurs. But it was possible to introduce a Petri net slice for detecting this event. Bounded buffered communication imposes a more complicated approach, illustrated in Figure 26. Consider that channel c have a buffer of size b. For each buffer slot k, 0 ≤ k ≤ b − 1, in a channel connecting ports with nesting factor n, there is a set of places buf slot flag[c,k,i], for 0 ≤ i ≤ n. They remember the kind of value stored in a buffer slot k, after communication in a channel c. Essentially, after firing transition sp set flag[gs ,i], the marking of stream port flag[gs ,i] at the moment of the activation of the group gs is saved in buf slot flag[c,k,i], where k is the number of the next available buffer slot. If all slots are filled, the sender blocks until a slot is freed by the receiver. If gr is a group of kind all (notice that all groups with one port is of kind all), an arc from each place buf slot flag[c,0,i] to transition sp set flag[gr ,i] ensures that the marking of Stream Flags(r) reflects the kind of the oldest value placed in the buffer by the sender, as semantics of buffered communication imposes. If gr is a group of kind any, it is necessary to introduce the Petri net slice shown in Figure 27. It ensures to copy the marking of the places buf slot flag[c,0,i] of the chosen receiver port r in gr . For that, the mutually exclusive places any group port activated[r], for each r belonging to gr , remember which port of gr was chosen. They enable the appropriate set of transitions any group copy flag[r], which are connected to places buf slot flag[c,0,i] of the channel where the chosen port is connected. The marking of this set of places is copied to places any group copied flag[r,i], which are connected to transitions sp set flag[gr , i]. 24 #1 ... #1 ... ... ... ... #1 Sender Side P1 ... T1 P3 #1 T3’s 2 1 ... 2 1 n−1 n ... P6 P5 P5 P6 P6 ... 2 ... 2 n−1 P5 ... ... 2 P5 2 P5 ... ... 2 P5 P5 T3’s T3’s 2 P5 ... 2 ... P5 Buffered Channel #n P2 P4 (T1) BUF_SLOTS_UNLOCK[c] (T2) BUF_SLOT_SELECT[c,s] (T3) BUF_SLOT_FLAG_SET[c,s,i] (T4) BUF_SLOT_SHIFT_FLAG[c,s,i] (T5) BUF_SLOTS_START_SHIFT[c] (T6) BUF_SLOTS_FINISH_SHIFT[c] T2’s 2 Buffer Management (P1) BUF_SLOTS_LOCKED[c] (P2) BUF_SLOTS_UNLOCKED[c] (P3) BUF_SLOT_READY_UPDATE[c] (P4) BUF_SLOT_UPDATED[c] (P5) BUF_SLOT_FLAG[c,s,i] (P6) BUF_SLOT_EMPTY[c,s] (P7) BUF_SLOT_READY_SHIFT[c,s] (P8) BUF_SLOTS_SHIFTED[c] (P9) BUF_SLOTS_SHIFTING[c,s] Receiver Side T4’s T4’s P7 ... #1 P7 P7 ... P8 #1 ... ... ... ... P9 T6 T5 #1 ... The kind of the received value must be the same kind of the oldest sent value Figure 26: Ensuring Consistency Streams Connected through Buffered Channels P3’s Activation of Groups of Ports of Kind Any ... T3 ... P1 T3 P3’s P2 T3 ... P1 ... T3 P3’s T1 (P1) ANY_GROUP_PORT_ACTIVATED[p] (P2) ANY_GROUP_COPIED_FLAG[g,i] (P3) BUF_SLOT_FLAG[p,0,i] ... ... T2 P2 T2 T2 P2 T3 T3 P1 ... (T1) PORT_SEND[p] / PORT_RECV[p] (T2) SP_SET_FLAG[g,i] (T3) ANY_GROUP_COPY_FLAG[p,i] ... T1 ... ... T1 T3 ... ... ... ... ... ... T3 T3 Figure 27: Copying Protocol for Groups of Input Ports of Kind any 25 The arrangement of the places buf slots locked[c] and buf slots unlocked[c] and the transition buf slots unlock[c] avoids accesses to the buffer while it is being updated after a transmission. The firing of a transition buf slot select[c,k] means the selection of the next empty slot. Whenever the buffer is full (n marks are deposited in place chan buffer used[c]), the sender must block for waiting that the receiver consume the contents of the first slot entry. After copying the kind of the value in the first slot of the buffer, it is discarded. For that, a shift operation, that occurs while a mark is deposited in place buf slots shifting[c], allows to save the marking of places buf slot flag[c,k,i] to buf slot flag[c,k − 1,i], for 1 ≤ k ≤ n. 3.4.2 Ensuring Consistency of Order of Kind of Transmitted Values The following example illustrates the needs for imposing one more restriction in the Petri net slice of the stream control protocol. For instance, consider a nested Haskell list of Int’s with nesting factor 4. [[[[1],[5,6]],[[2,3]]],[],[[[[4,5,7],[8,9]]],[[6],[7,9]]]] The correspondent stream in the Hash component model will transmit the following values at each activation of the corresponding port: {1,Eos 3, 5, 6,Eos 3,Eos 1, 2, 3,Eos 3,Eos 2,Eos 1,Eos 1, 4, 5, 7, Eos 3, 8, 9,Eos 3,Eos 2,Eos 1, 6,Eos 3, 7, 9,Eos 3,Eos 2,Eos 1,Eos 0} Notice that after transmitting a value Eos 2, it is not possible to transmit a value Eos 0, since the enclosing stream at nesting level 1 was not finalized yet. In this case, only values Eos 1, Eos 2, Eos 3 and a data item may be transmitted. Now, consider the general case for the transmission of given end marker at nesting level k. If k ≥ 1, the next value to be transmitted may be any of the end markers at nesting level greater than k − 1 or a data item. If k = 0, the stream is finalized. Any attempt to read a value in a finalized stream is considered an error. In Figure 28, it is shown how the Petri net slice presented in Figure 24 (output port) may be enriched in order to support the restriction stated in the last paragraph. A mark is placed in place sp order fail[s] whenever an attempt to activate a finalized stream port s (the kind of last transmitted value is Eos 0) occurs. For each nesting level i, four places and two transitions controls the consistency of the order of the transmitted value. The place sp flag open[s,i] receives a value whenever a value Eos i may be sent. Its mutually exclusive dual place, named sp flag open dual[s,i] allows for resetting the marking of sp flag open[s,i] after activation. Resetting procedure is implemented using the next elements described here. The place sp cleaning flag[s,i] has a mark whenever the resetting procedure is enabled for the port s in nesting level i. Its corresponding place sp cleaned flag[s,i] has a mark after resetting procedure finishes. The transition sp clean[s,i] resets the place sp flag open[s,i] 26 P5 T3’s (P1) SP_FLAG_OPEN[s,i] (P2) SP_FLAG_OPEN_DUAL[s,i] (P3) SP_CLEANING_FLAG[s,i] (P4) SP_CLEANED_FLAG[s,i] (P5) SP_FAIL T1 P3 P1 P2 P4 ... T2 T1 P3 P1 (T1) SP_CLEAN[s,i] (T2) SP_KEEP_CLEANED[s,i] P2 T2 P4 ... ... ... (T3) SP_CLEAR_FLAG[s,i] (T4) SP_SET_FLAG[s,i] (T5) SP_PREPARE_PORT[s] T1 P3 ... P1 P2 From Stream Output Port T2 P4 T4’s Figure 28: Consistency of the Order of Kind of Transmitted Stream Values to its original state (zero marks), while transition sp keep cleaned[s,i] fires whenever the place sp flag open[s,i] is already cleaned. Notice that the transitions sp keep cleaned[s,i] and sp clean[s,i] are mutual exclusive, since sp flag open[s,i] and sp flag open dual[s,i] are too. 3.5 The Complexity of the Generated Petri Net After overlapping the Petri net slice that models stream communication semantics, allowing to make precise analysis about behavior of Hash programs at coordination level, the Petri net of simple Hash programs may become very large. Large Petri nets may turn impossible for programmers the analysis without help of some automatic or higher level means. Also, it makes hard and memory consuming the computations performed by the underlying Petri net tools, such as computation of reachability and coverability graphs, place and transition invariants, etc. These difficulties comes from transitory technological limitations, since processing power and memory amount of machines have increased rapidly in the recent years and it is expected that they will continue to increase in the next decades. Also, it is possible to use parallel techniques to perform high computing demanding analysis, but this approach has been exploited by few designers of Petri net tools. It is a reasonable assumption that parallel programmers have access to some parallel computer. Despite these facts, it is desirable to provide ways for helping programmers to work with large Petri nets or simplifying the generated Petri net. The following techniques have been proposed: • During the process of analysis, the programmer may decide not to augment the Petri net of the Hash program with the protocol for modelling stream communication. This approach makes sense whenever the infor27 T5 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. component IS<problem class, num procs, max key log2, num buckets log2, total keys log2, max iterations, max procs, test array size> with #define PARAMETERS (IS Params problem class num procs max key log2 num buckets log2 total keys log2 max iterations max procs test array size) iterator i range [1, num procs] use Skeletons.{Misc.RShift, Collective.{AllReduce, AllToAllv}} use IS FM −− IS Functional Module unit bs comm assign AllReduce<num procs, MPI SUM, MPI INTEGER> to bs comm unit kb comm assign AllToAllv<num procs> to kb comm unit k shift assign RShift<num procs> 0 → to k shift interface IIS (bs, kb, k) → (bs, kb, k) where: bs@IAllReduce (UArray Int Int) # kb@IAllToAllv (Int, Ptr Int) # k@RShift Int behaviour: seq {repeat seq {do bs; do kb} until <bs & kb>; do k} [/ unify bs comm.p[i] # bs, kb comm.p[i] # kb, k comm.p[i] # k to is peer[i] # IIS assign IS FM (PARAMETERS, bs, kb, k) → (bs, kb, k) to is peer[i] # bs # kb # k /] Figure 29: The Configuration Code of IS Program mation provided by the protocol is not always necessary for the analysis being conducted. This is the reason for the separation of the stream communication protocol from the rest of the Petri net in the Hash program; • Another approach that have been proposed, but for further works, is to build higher level environments for analysis of Hash programs on top of Petri net tools. Instead of programmers to manipulate Petri net components, they manipulate Hash program elements of abstraction. Then, the analysis are transparently and automatically translated into proving sequences using INA tool; • Specific translation schemas for skeletons might be specified in order to simplify the generated Petri net. This approach is illustrated in the next section. For instance, it is defined how higher-level information provided by the use of collective communication skeletons might be used in the translation process of Hash programs into Petri nets. 3.6 Modelling Collective Communication Skeletons Message passing libraries, such as MPI, support special primitives for collective communication. In Hash programming environment, it is defined a library of skeletons that implement the pattern of communication involved in collective communication operations supported by MPI. They are: Bcast, Scatter, Scatterv, Reduce Scatter, Scan, Gather, Gatherv, Reduce, AllGather, AllGatherv, AllToAll, AllToAllv, and AllReduce. In the next paragraphs, the use of collective communication skeletons is illustrated by means of an example. In Figure 29, the code for the Hash version of IS program, from NPB (NAS Parallel Benchmarks) [3] is presented. IS is a parallel implementation of the bucket sort algorithm, originally written in C/MPI. In this section, IS is used to 28 collective port (a) T1 P1 (b) P1 join transitions P2 P2 T1 P2 P1 T1 ... ... T1 P1 ... P2 ... P1 P1 P2 P2 protocol for modelling stream communication semantics (P3) SP_CLEARING_FLAGS[d] (P4) SP_CLEARED_FLAGS[d] (P5) SP_SETTING_FLAGS[d] (P6) SP_EXITING[d] (P7) STREAM_PORT_FLAG[d,i] (P8) STREAM_PORT_FLAG_DUAL[d,i] P8’s (c) (P1) COLLECTIVE_PREPARED[d] (P2) COLLECTIVE_COMPLETE[d] #1 ... (T1) COLLECTIVE_JOIN[d] (T2) COLLECTIVE_UNJOIN[d] #1 P1 T1 P1 P5 P2 P6 T2 ... ... ... ... P1 T4 P4 P3 P2 T5’s T3’s P2 #1 ... P7’s shared among all collective ports (T3) SP_CLEAR_FLAG[d,i] (T4) SP_START_SET_FLAG[d,i] (T5) SP_SET_FLAG[d,i] Figure 30: Modelling Collective Communication Skeleton Semantics illustrate the use of collective communication skeletons in a Hash program, motivating definition of a specific strategy for translating collective communication patterns of interaction among units in a Hash program. In the line 9 of Figure 29, it is declared that skeletons2 AllReduce and AllToAllv will be used in the configuration. In lines 12, 13, and 14, three units are declared, named bs comm, kb comm and k shift. The first two have collective communication skeletons assigned to them, respectively AllReduce and AllToAllv. Since skeletons are composed components, these units are clusters of units that interact using the collective communication patterns described by the skeleton. In line 20, the unification of correspondent units that comprise clusters bs comm, kb comm and k shift forms the units that comprise the IS topology. The unification of virtual units from distinct cluster allows to overlap skeletons. The behavior of the virtual units that result from unification, 2 Partial topological skeletons are composed components where at least one unit is virtual. In the case of collective communication skeletons, all units are virtual. 29 named is peer[i], 1 ≤ i ≤ num procs, is specified by interface IIS. The simple component IS FM is assigned to them for defining the computation of each process in the IS program. Notice that IS is a SPMD program, where the task performed by all processes is defined by the same simple component [7]. The interface IIS is declared from composition of IAllReduce, IAllToAllv and IRShift interfaces, the interface slices of IIS. The interface slices are respectively identified by bs, kb, and k. The use of combinator do is an abbreviation that avoids to rewrite the behavior of interface slices. Thus, “do bs” relates to the sequence of actions encapsulated in the specification of interface IAllReduce. Using the above conventions for overlapping collective communication skeletons in order to form more complex topologies, it is simple to define a specific translation rule for patterns of collective communication interactions. The identifiers of interface slices from collective communication skeletons might be viewed as special kinds of ports (collective ports) and the operator do as its activation operator. Notice that interface slices may be used in termination conditions of streams, like in line 18. Collective ports have no direction, since all processes participate in communication. All communication operations are synchronous. The Figure 30(b) illustrates a Petri net slice that models a collective communication operation. The involved ports are collective ports that correspond to interface slices of units that participates in a certain collective communication operation, defined by the cluster of units that defines it. In Figure 30(c), it is illustrated how to augment the Petri net slice shown in Figure 30(b) with a protocol for modelling stream communication semantics. It is important to notice that the set of places Stream Flags(s) and Stream Flags(s) are shared by all collective ports involved in a collective operation. 4 Petri Net Analysis of Formal Properties In this section, solutions for two well known synchronization problems, implemented in Hash approach, illustrates the use of Petri nets for analyzing Hash programs by verifying their formal properties. 4.1 Dining Philosophers The dining philosophers problem is one of the most relevant synchronization problems in concurrency theory. Since it was originally proposed by Dijkstra in 1968 [12], it has been widely used for exercising the ability of concurrent languages and models for providing elegant solutions for avoiding deadlocks in concurrent programs. The dining philosophers problem is stated in the following way: Five philosophers are sited around a table for dinner. The philosophers spend their times eating and thinking. When a philosopher wants to eat, he takes two forks from the table. When a philosopher 30 + + phil[0] + + phil[0] + + phil[1] + + + + phil[4] + + + + phil[3] + + phil[4] + + phil[2] + + phil[1] phil[3] (a) phil[2] (b) Figure 31: Hash Topologies for the Dining Philosophers Problem wants to think, he keeps the two forks available on the table. However, there are only five forks available, requiring that each philosopher share their two forks with their neighbors. Thus, whenever a philosopher is eating, its neighbors are thinking. A solution to the dining philosopher problem establishes a protocol for ordering the activity of the philosophers. In Figure 31, Hash topologies for solutions for the dining philosophers problem are presented. The first one, whose code is presented in Figure 32(a), is an “anarchical” solution, where philosophers are free to decide when to think or to eat. In this solution, the reader may observe the use of buffered channels and groups of ports of kind any composing the network topology. The one-slot buffered channel allows to model the fact that a fork may be on the table waiting for a philosopher to acquire them. This occurs whenever there is one message pending on the buffer. At the beginning of the interaction, all philosophers have a fork and release them. When a philosopher releases a fork, the semantics of group of ports of kind any ensures that if there is a philosopher waiting for the fork, he will obtain the fork immediately. Otherwise, it is possible that the philosopher that released the fork have a chance to acquire the fork again. This solution does not satisfy some of the enunciated requisites for a good solution to an instance of the critical section problem. For example, if all philosophers acquire their right (left) forks, they will be in deadlock. It is also possible that a philosopher never get a chance to obtain the forks (eventual entry). The second solution, whose code is presented in Figure 32(b), ensure all requisites. Additionally, it ensures maximal parallelism. In any state of execution, there are two philosophers eating. Figure 33 presents the Petri nets that model the respective behaviors of individual philosophers in the first and in the second solutions. Figure 34 presents the Petri nets modelling the interaction among the five philosophers, after modelling communication channels. The figures are only illustrative, since the networks have a number of components intractable by simple visual inspection. 31 component DiningPhilosophers<N> with index i range [0,N-1] interface IPhil where ports: (rf get, lf get) → (rf put, rf put) protocol: seq {rf put!; repeat seq {par {lf get?; rf get?}; par {lf put!; rf put!}} until < lf get & rf get & lf put & rf put >} [/ unit phil[i] where ports: IPhil grouping: rf get {neighbor,self} any, lf get {neighbor,self} any, rf put {neighbor,self} any, lf put {neighbor,self} any /] [/ connect connect connect connect phil[i]→rf put[neighbor] phil[i]→rf put[self] phil[i]→lf put[neighbor] phil[i]→lf put[self] to to to to phil[i-1 mod N] ←lf get[neighbor] , phil[i] ←rf get[self] , phil[i+1 mod N] ←rf get[neighbor] , phil[i] ←lf get[self] , buffered buffered buffered buffered 1 1 1 1 /] (a) component DiningPhilosophers<N> with index i range [0,N-1] interface IPhil [0] where ports: (rf get, lf get) → (rf put, rf put) protocol: repeat seq {lf put!;rf get?; lf get?; rf put!} interface IPhil [1] where ports: (rf get, lf get) → (rf put, rf put) protocol: repeat seq {rf get?; lf put!; rf put!; lf get?} [/ unit phil[i] where ports: IPhil [i mod 2] /] [/ connect phil[i]→rf put to phil[i-1 mod N]←lf get, buffered connect phil[i]→lf put to phil[i+1 mod N]←rf get, buffered /] (b) Figure 32: Hash Code for the First Solution of the Dining Philosophers Problem 32 (a) (b) Figure 33: Petri Net Modelling Behavior of One Philosopher 4.1.1 Proving Properties on the Dining Philosophers Solution In this section, INA (Integrated Network Analyzer) is used as an underlying engine for verifying formal properties about the above Hash solutions for the dining philosophers problem. INA allows to perform several structural and dynamic analysis on the Petri net induced from the two solutions. Among other possible analysis approaches, INA provides model checking facilities that allows to check validity of CTL (Constructive Tree Logic) formulae, describing properties about the Hash program, on the reachability graph of its corresponding Petri net. CTL is a suitable formalism, from branching-time temporal logics, for expressing and verifying safety (invariance) and liveness properties of dynamic systems. It allows temporal operators to quantify over paths that are possible from a given state. There exist a superset of CTL, named CTL∗ , that augments expressive power of CTL by allowing to express fairness constrains that are not allowed to be expressed in CTL. However, INA restricts to CTL because model checking algorithms for CTL are more efficient (linear in the formula size) that in CTL∗ (exponential in the formula size). Now, let us to introduce relevant properties that may be proved about solutions to the dining philosophers problems and to model these properties using CTL. The dining philosophers problem may be thought as an instance of the critical section problem. In fact, forks are critical sections, since it cannot be taken by more than one philosopher. A good solution to an instance of the critical section problem must ensure three properties [1]: • Mutual exclusion: Two adjacent philosophers cannot obtain the same 33 (a) (b) Figure 34: Petri nets for Dining Philosophers Hash Solutions 34 (a) Macros About Channels sender prepared[c] receiver prepared[c] rendezvous[c] buffer full[c] buffer empty[c] sender blocked[c] sender blocked[c] receiver blocked[c] receiver blocked[c] (b) Macros About Ports port pair prepared[p] port pair prepared[p] = = = = = = = = = = = port prepared[∂0 (c)] port prepared[∂1 (c)] sender ready[c] ∧ receiver ready[c] ¬chan buffer free[c] ¬chan buffer used[c] sender ready[c] ∧ ¬receiver ready[c] sender ready[c] ∧ buffer full[c] receiver ready[c] ∧ ¬sender ready[c] receiver ready[c] ∧ buffer empty[c] port prepared[∂1 ◦ ∂0− 1(p)] port prepared[∂0 ◦ ∂1− 1(p)] (c) Macros About Groups of Ports group prepared[G] group prepared[G] = = i • s • • • b • • • • • r • • • • • • • o • (∀p)∈G : port prepared[p] (∃p)∈G : port prepared[p] all • any • Table 1: Some Useful Formula Macros for Hash programming fork (enter critical section); • Absence of deadlock: A deadlock occurs whenever there is at least one active philosopher and all active (not terminated) philosophers are blocked. The classical deadlock situation in dining philosophers problem occurs when all philosophers acquire its right, or left, forks. In this state, all philosophers may not proceed and they are not finished. Thus, they are in deadlock; • Absence of unnecessary delay: If a philosopher demands its right (left) fork and their right (left) neighbor is thinking, the philosopher is not prevented from obtaining the fork; • Eventual entry: Each philosopher that demands for a fork eventually will obtain it. In the following paragraphs, the above properties are characterized using CTL formulae. But before, intending to facilitate concise and modular specification of complex CTL formulas, we define the notion of CTL-formula macro. A CTL-formula macro is new kind of CTL-formula of the form hmacro namei[q1 , q2 , . . . , qn ], where hmacro namei is a macro name, qi , 1 ≤ i ≤ n, are qualifiers. CTLformulae macros may be expanded in flat CTL-formulae, by applying recursively their definitions. For instance, A CTL-formula macro is defined using the following syntax: hmacro namei[v1 , v2 , . . . , vn ] :: f where hmacro namei is a name for the macro, vi , 1 ≤ i ≤ n, are qualifier variables, and f is a CTL-formula (possibly making reference to CTL-formula 35 (a) Macros About Philosophers (p = 0 . . . N − 1) phil demands lf[p] = phil demands rf[p] = phil posseses lf[p] = phil posseses rf[p] = phil is eating[p] = phil posseses lf[p] ∧ phil posseses rf[p] phil is thinking[p] = ¬phil is eating[p] phil waiting[p] = phil demands rf[p] ∨ phil demands lf[p] phil finished[p] = process finished[phil[p]] all phil finished = (∀p)<N : phil finished[p] (b) Macros About Forks (f = 0 . . . N − 1) fork is free[f ] = phil is thinking[f ] ∧ phil is thinking[f ⊕N 1] fork in use by right[f ] = phil posseses lf[f ] fork in use by left[f ] = phil posseses rf[f ⊕N 1] fork in use[f ] = fork in use by right[f ] ∨ fork in use by left[f ] Table 2: Some Useful Formula Macros for Dining Philosophers macros). In a flat CTL-formula that appears in the right-hand-side of a CTLformula macro, qualifier variables are used as qualifiers for place and transition identifiers and for references to enclosed CTL-formulae macros. In Table 1, some useful CTL-formulae macros are defined for simplifying specification of CTL formulae on the restrict domain of Hash programs. In Table 2, other CTL-formulae macros are defined, but now on the restrict domain of the dining philosophers problem (application oriented). Notice that CTL-formulae macros of Table 2 are defined on top of that defined in Table 1. Unlike the later macros, the implementation of the former ones is not sensitive to modifications in the underlying translation schema. This illustrates the transparency provided by the use of CTL-formulae macros in an environment for proof and analysis of formal properties. We have used INA for proving the three properties enunciated above for dining philosophers. The first three ones are safety properties. It may be proved by negating a predicate describing a state that cannot be reached in execution (bad state). The last one is a liveness property, for which validity of a predicate must be checked in all possible states. Proof of Mutual Exclusion. Safety property. One valid formulation for the corresponding bad state is: bad = EF (∃f )0≤f <N : fork in use by right[f ] ∧ fork in use by left[f ] Proof of Absence of deadlock. the corresponding bad state is: bad = EF Safety property. One valid formulation for (∀p)0≤p<N : phil waiting[p] ∨ phil finished[p] ∧ ¬all phil finished Proof of Absence of unnecessary delay. mulation for the corresponding bad state is: 36 Safety property. One valid for- out is [t] out transmitter as [(t,Bit)] as await asl [(t,Bit)] as corrupt_send bs2 [Err (t,Bit)] [Err Bit] bs ds corrupt_ack as [Bit] cs in [Err (t,Bit)] bs ds sender side bs1 ack receiver side os [t] in receiver Figure 35: Hash Topology for of ABP (Alternating Bit Protocol) EF bad = (∀p)0≤f <N : fork is free[f ] ∧ (phil demands lf[f ] ∨ phil demands rf[f ⊕N 1]) Proof of Eventual entry. Liveness property. One valid formulation for the corresponding good state is: : good = AG [phil waiting[p] ⇒ (AF [phil eating[p]])] 4.2 The Alternating Bit Protocol The alternating bit protocol (ABP) is a simple and effective technique for managing retransmission of lost messages in fault-tolerant low level implementations of message passing libraries. Given a transmitter process A and a receiver process B, connected by a point-to-point stream channel, ABP ensures that whenever a message sent from A to B is lost, it is retransmitted. The Hash implementation described here is based on a functional implementation described in [13]. The Figure 35 illustrates the topology of the component ABP, which might be used for implementing ABP protocol. The virtual units transmitter and receiver model the processes involved in the communication. The other units implement the protocol. The units transmitter, out, await, and corrupt ack implement the sender side of the ABP protocol, while the units receiver, in, ack, and corrupt send implement the receiver side. The await process may retransmit a message repetitively until the message is received by process ack. Retransmissions are modelled using streams with nesting factor 2 (streams of streams). The elements of the nested stream correspond to the retransmission attempts of a given value. The correct arrive of the message is performed by inspecting the value received through the port ds. The processes corrupt ack and corrupt send verify the occurrence of errors in the messages that arrive at the sender and receiver, respectively, modelling the unreliable nature of the communication channel. The Hash configuration code presented in Figure 36 implements the ABP component. The Petri net induced by translating ABP component is presented in Figure 37. 37 component ABP with use Out, Await, Corrupt, Ack, In interface ABP Transmitter where ports: () → (out::t) protocol: repeat out! until out interface ABP Receiver t where ports: (in::t) → () protocol: repeat in? until in interface Out where ports: (is::t) → (as::(t,Bit)) protocol: repeat seq {is?; as!} until <is & as> interface Await where ports: (as::(t,Bit),ds::Err Bit) → (as’::(t,Bit)) protocol: repeat seq { as?; repeat seq {as’ !; ds?} until <as’ & ds>} until <as & as’ & ds> interface Corrupt where ports: (as::(t,Bit)) → (bs::Err (t,Bit)) protocol: repeat seq {as?; bs!} until <as & bs> interface Ack where ports: (bs::Err (t,Bit)) → (cs::Bit) protocol: repeat seq {bs?; cs!} until <bs & cs> interface In where ports: (bs*::Err (t,Bit)) → (os::t) protocol: repeat seq { repeat bs? until bs; os!} until <bs & os> unit transmitter where ports: ABP Transmitter () → out unit receiver where ports: ABP Receiver in → () unit unit unit unit unit unit out await corrupt ack in ack corrupt send connect connect connect connect connect connect connect connect * * * * * * * * where where where where where where ports: ports: ports: ports: ports: ports: transmitter→out in→os out→as await→as corrupt ack→ds corrupt send→bs[0] corrupt send→bs[1] ack→cs IOut IAwait ICorrupt IIn IAck ICorrupt grouping: bs2 all to to to to to to to to out←is receiver←in await←as corrupt send←as ack←ds in←bs ack←bs corrupt ack←cs assign assign assign assign assign assign Out to Await to Corrupt to In to Ack to Corrupt to , buffered , buffered Figure 36: ABP Component 38 out await corrupt ack in ack corrupt send transmitter await receiver corrupt_send out corrupt_ack ack in Figure 37: Petri Net Induced by ABP Component 39 References [1] G. Andrews. Concurrent Programming: Principles and Practice. Addison Wesley, 1991. [2] R. Armstrong, D. Gannon, A. Geist, K. Keahey, S. Kohn, L. McInnes, S. Parker, and B. Smolinski. Towards a Common Component Architecture for High-Performance Scientific Computing. In The 8th IEEE International Symposium on High Performance Distributed Computing. IEEE, 1999. [3] D. H. Bailey, T. Harris, W. Shapir, R. van der Wijngaart, A. Woo, and M. Yarrow. The NAS Parallel Benchmarks 2.0. Technical Report NAS-95-020, NASA Ames Research Center, December 1995. http://www.nas.nasa.org/NAS/NPB. [4] E. Best, J. Esparza, B. Grahlmann, S. Melzer, S. Rmer, and F. Wallner. The PEP Verification System. In Workshop on Formal Design of Safety Critical Embedded Systems (FEmSys’97), 1997. [5] F. H. Carvalho Junior, R. M. F. Lima, and R. D. Lins. Coordinating Functional Processes with Haskell# . In ACM Press, editor, ACM Symposium on Applied Computing, Track on Coordination Languages, Models and Applications, pages 393–400, March 2002. [6] F. H. Carvalho Junior and R. D. Lins. Haskell# : Parallel Programming Made Simple and Efficient. Journal of Universal Computer Science, 9(8):776–794, August 2003. [7] F. H. Carvalho Junior and R. D. Lins. On the Implementation of SPMD Applications using Haskell#. In 15th Brazilian Symposium on Computer Architecture and High Performance Computing (SBAC-PAD 2003). IEEE Press, November 2003. [8] F. H. Carvalho Junior and R. D. Lins. Topological Skeletons in Haskell# . In International Parallel and Distributed Processing Symposium (IPDPS). IEEE Press, April 2003. 8 pages. [9] F. H. Carvalho Junior, R. D. Lins, and R. M. F. Lima. Parallelising MCP-Haskell# for Evaluating Haskell# Parallel Programming Environment. In UnB, editor, 13th Brazilian Symposium on Computer Architecture and High-Performance Computing (SBAC-PAD 2001), September 2001. [10] F. H. Carvalho Junior, R. D. Lins, and R. M. F. Lima. Translating Haskell# Programs into Petri Nets. Lecture Notes in Computer Science (VECPAR’2002), 2565:635–649, 2002. [11] M. Cole. Algorithm Skeletons: Structured Management of Paralell Computation. Pitman, 1989. 40 [12] E. C. Djikstra. The Structure of THE Multiprogramming System. Communications of the ACM, 11:341–346, November 1968. [13] P. Dybjer and H. P. Sander. A Functional Programming Approach to the Specification and Verification of Concurrent Systems. Formal Aspects of Computing, 1:303–319, 1989. [14] R. German. SPNL: Processes as Language-Oriented Building Blocks of Stochastic Petri Nets. In 9th Conference on Computer Performance Evaluation, Modelling Techniques and Tools, pages 123–134. Springer Verlag, 1997. [15] J. Gischer. Shuffle Languages, Petri Nets, and Context-Sensitive Grammars. Communications of the ACM, 24(9):597–605, September 1981. [16] B. Grahlmann and E. Best. PEP - More than a Petri Net Tool. In Lecture Notes in Computer Science (Tools and Algorithms for the Construction and Analysis of Systems, Second Int. Workshop, TACAS’96, Passau, Germany), volume 1055, pages 397–401. Springer Verlag, March 1996. [17] Inmos. Occam Programming Manual. Prentice-Hall, C.A.R. Hoare Series Editor, 1984. [18] T. Ito and Y. Nishitani. On Universality of Concurrent Expressions with Synchronization Primitives. Theoretical Computer Science, 19:105–115, 1982. [19] G. Kiczales, J. Lamping, Menhdhekar A., Maeda C., C. Lopes, J. Loingtier, and J. Irwin. Aspect-Oriented Programming. In Lecture Notes in Computer Science (Object-Oriented Programming 11th European Conference – ECOOP ’97), volume 1241, pages 220–242. Springer-Verlag, November 1997. [20] S. L. Peyton Jones and J. (editors) Hughes. Report on the Programming Language Haskell 98, A Non-strict, Purely Functional Language, February 1999. [21] S. Roch and P. Starke. Manual: Integrated Net Analyzer Version 2.2, 1999. [22] A. C. Shaw. Software Descriptions with Flow Expressions. IEEE Transactions on Software Enginnering, SE-4(3):299–325, May 1978. [23] J. M. Squyres and A. Lumsdaine. A Component Architecture for LAM/MPI. In Proceedings, 10th European PVM/MPI Users’ Group Meeting, number 2840 in Lecture Notes in Computer Science, Venice, Italy, September 2003. Springer-Verlag. [24] J. M. Squyres, A. Lumsdaine, W. L. George, J. G. Hagedorn, and J. E. Devaney. The interoperable message passing interface (IMPI) extensions to LAM/MPI. In Proceedings, MPIDC’2000, March 2000. 41 [25] S. Thompson. Haskell, The Craft of Functional Programming. AddisonWesley Publishers Ltd., 1996. [26] A. Zimmermann, J. Freiheit, R. German, and G. Hommel. Petri Net Modelling and Performability Evaluation with TimeNET 3.0. In 11th Int. Conf. on Modelling Techniques and Tools for Computer Performance Evaluation (TOOLS’2000), pages 188–202. Lecture Notes in Computer Sciente, 2000. A The Formal Syntax of HCL In what follows, it is described a context-free grammar for HCL, the Haskell# Configuration Language, whose syntax and programming abstractions were informally presented in Section ??. Examples of HCL configurations and their meanings were presented in Sections ?? and ??. The notation employed here is similar to that used for describing syntax of Haskell 98 [20]. Indeed, some non-terminals from that grammar are reused here, once some Haskell code appears in HCL configurations. They are faced italic and bold. A minor difference on notation resides on the use of (. . .)? , instead of [. . .], for describing optional terms. For simplicity, notation for indexed notation is ignored from the description of formal syntax of HCL. It may be resolver by a pre-processor, before parsing. A.1 Top-Level Definitions configuration header static parameter list component interface declaration A.2 → → → → → \| \| \| header declaration 1 . . . declaration n (n ≥ 0) component ID static parameter list ? component interface ? < ID1 . . . IDn > (n ≥ 0) ports naming \| iterator decl \| interface decl import decl \| use decl unit decl \| assign decl \| replace decl \| channel decl unify decl \| factorize decl \| replicate decl \| bind decl haskell code Use Declaration use decl → use use spec use spec → id \| id.use spec \| id.{ use spec 1 , . . . , use spec n } A.3 (n ≥ 1) Import Declaration import decl → impdecl A.4 Iterator Declaration iterator decl → iterator id1 , . . ., idn range [ numeric exp , numeric exp ] (n ≥ 1) 42 A.5 Interface Declaration interface decl → interface (context =>)? ID tyvar 1 . . . tyvar k interface spec interface spec → interface ports spec (where : interface inheritance)? (behavior : behavior expression)? A.5.1 Interface Ports Description interface ports spec → port spec list -> port spec list port spec list → port spec \| ( port spec 1 , . . . , port spec n ) (n ≥ 2) port spec → id ()? (:: atype)? \| id A.5.2 Interface Composition → interface slice 1 # . . . # interface slice k (k ≥ 1) interface inheritance interface slice → id @ ID \| ID ports naming composition ports naming composition → ports naming \| ( ports naming 1 # . . . # ports naming n ) (n ≥ 1) ports naming → port naming list -> port naming list port naming list → id \| ( id1 , . . . , idn ) (n ≥ 1) A.5.3 Interface Behavior behavior expression → action → \| \| \| condition → disjunction → sync conjunction → simple conjunction → A.6 Unit Declaration unit decl unit spec unit interface wf setup group spec group type wire function A.7 (sem id1 , . . . , idn )? : action (n ≥ 1) par { action 1 ; . . . ; action n } \| seq { action 1 ; . . . ; action n } alt { action 1 ; . . . ; action n } \| repeat action condition ? if condition then action else action id ! \| id ? \| signal id \| wait id (n ≥ 2) until disjunction \| counter numeric exp sync conjunction 1 ‘\|’ . . . ‘\|’ sync conjunction n (n ≥ 1) h simple conjunction i \| simple conjunction id \| ( id1 & . . . & idn ) (n ≥ 1) → → → → → → → unit unit spec ()? id (# unit interface)? (wire wf setup 1 , . . . , wf setup n )? ID ports naming composition ? \| interface spec id (group type group spec)? (: wire function)? { id1 , . . ., idn } \| numeric exp any \| all ? \| exp Assignment Declaration assign decl → assign assigned component to assigned unit 43 assigned component → ID actual parameter list ? ports naming composition ? actual parameter list → < numeric exp 1 , . . . , numeric exp n > (n ≥ 1) assigned unit → qid ports naming composition ? A.8 Replace Declaration replace decl → replace qid ports naming composition ? by operand unit A.9 Channel Declaration channel decl → connect qid -> qid to qid <- qid , comm mode comm mode → synchronous \| buffered numeric exp \| ready A.10 Unification Declaration → unify operand unit 1 , . . . , operand unit n to unit spec adjust wire wf setup 1 , . . . , wf setup k (n ≥ 2, k ≥ 1) operand unit → qid # interface pattern 1 . . . # interface pattern n (n ≥ 1) interface pattern → port pattern list -> port pattern list \| id port pattern list → pattern \| ( pattern 1 , . . . , pattern n ) pattern → id \| @ qid \| \| unify decl A.11 Factorization Declaration factorize decl → factorize operand unit to unit spec 1 . . . unit spec n adjust wire wf setup 1 , . . . , wf setup k (n ≥ 2, k ≥ 1) A.12 Replication Declaration replicate decl → replicate operand unit 1 , . . . , operand unit n into numeric exp adjust wire wf setup 1 , . . . , wf setup k (n ≥ 2, k ≥ 1) A.13 Bind Declaration bind declaration → bind qid -> qid to -> id \| bind qid <- qid to <- id A.14 Miscelaneous haskell code → topdecls qid → id1 ‘.’ . . . ‘.’ idn (n ≤ 2) qID → ID1 ‘.’ . . . ‘.’ IDn (n ≤ 2) 44 B Foundations and Notations In this section, it is discussed the formalisms that comprise the formal framework for the development of this work, concerning modelling of communication behavior of processes, according to the Hash component model design principles. B.1 Formal Languages The theory of formal languages will be employed as a framework for the study of patterns of communication interaction of Hash process in a parallel program. The main interest is to investigate relations between descriptive power of concurrent expressions and Petri nets, in order to define a language for expressing communication behavior of processes, embedded in the Hash language. Definition 1 (Alphabet) An alphabet is a finite set of indivisible symbols, denoted by Σ. Definition 2 (Word) A word is a finite sequence of symbols of some alphabet Σ. The symbol ǫ denotes the empty word, whose length is zero. Definition 3 (Kleene’s Closure of an Alphabet ) A Kleene’s closure of an alphabet Σ, denoted by Σ∗ , is defined as below: Σ∗ = {w \| w is a word in Σ} Thus, any sequence of symbols in Σ, including ǫ, belongs to Σ∗ . It is common to define Σ+ as: Σ+ = Σ∗ − {ǫ} Definition 4 Given an alphabet Σ, a formal language L is defined as follows: L ⊂ Σ∗ B.2 Labelled Petri Nets and Formal Languages Now, notations and definitions concerning Petri nets are presented. Definition 5 (Place/Transition Petri Net) A place/transition Petri net is a directed bipartite graph that can be formalized as a quadruple (P, T, A, M0 ), where: 1. P is a finite set of places, which can store an unlimited number of marks; 2. T is a finite set of transitions. 3. P ∩ T = ∅. 45 4. A defines a set of arcs, in such way that A ⊆ ((P × T ) ∪ (T × P )) × N aturais. Thus, an arc can go from a transition to a place or from a place to a transition. A number is associated to the arc, indicating its weight. For simplicity, if the weight is omitted it is one ((p, t) ≡ (p, t, 1)). 5. The relation M0 ⊂ P × N defines the initial marking, or the number of marks that are stored in each place at the initial state of the Petri net; From the initial marking, a Petri net defines a set reachable markings. A marking is reachable if it can be obtained from the initial marking by firing a sequence of transitions, according to firing rules, formalized in what follows. Definition 6 (Enabled Transition) Be Π, Π = (P, T, A, MO ), a Petri net and t, t ∈ T , a transition of Π: t is enabled ⇔ (∀p ∈ P, ∃m, n ∈ N : (p, t, m) ∈ A ∧ (p, n) ∈ M : n ≥ m) Thus, a transition t is enabled if the number of marks in each one of its input places is greater than or equal to the weight of the arc that links it to the transition. Definition 7 (Firing Rule and Reachable Markings) Be Π = (P, T, A, M0 ) a Petri net and a marking M of Π. The transition from the marking M to a new marking M’ is a side effect of firing an enabled transition t, and represented by t the relation M → M ′ , where: M ′ = M̄ ∪ MI ∪ MO MI = {(p, n − m) \| ∃m, n ∈ N : p ∈ P, (p, t, m) ∈ A, (p, n) ∈ M } MO = {(p, n + m) \| ∃m, n ∈ N : p ∈ P, (t, p, m) ∈ A, (p, n) ∈ M } The sub-marking MI indicates the new marking for the input places of the transition t, while MO indicates the new marking for output places of t. The transition can be fire if it is enabled and the effect of firing is to remove marks from input places and add marks to output places, according to the weights of the arcs that link these places to t. It is necessary to generalize this definition to cover the concept of (transitively) reachable marking. Thus, a marking Mn is reachable from a mark M if there is a sequence of firing of transitions from M to Mn : t t t t n 3 2 1 Mn ··· → M2 → M1 → M→ This notation can be abbreviated to: σ M → Mn , where σ = t1 t2 · · · tn The symbol σ denotes a sequence of firings of transitions. 46 Definition 8 (Labelled Petri nets) Be Σ an alphabet. A labelled Petri net is a 7-uple (P, T, A, M0 , ρ), where (P, T, A, M0 ) is a place/transition Petri net and ρ : T → Σ ∪ {λ} is a function that associate transitions to a symbol in Σ. Transitions labelled with λ are called silent transitions. Labelled Petri nets are an extension of Petri nets for generating formal languages. There are two classes of Petri net languages. Definition 9 (Petri Net Language) Given a labelled Petri net Π = (P, T, A, M0 , ρ), we define the formal language generated by Π as: σ L(Π, M0 ) = {ρ(σ) \| ∃M : M0 → M }. Definition 10 (Petri Net Terminal Language) Given a Petri net Π(P, T, A, M0 , ρ) and a final marking Mf , we can define the terminal language generated by Π, with respect to Mf , by: σ T (N, M0 , Mf ) = {ρ(σ) \| M0 → Mf } Given a labelled Petri net Π, the language generated by Π defines the possible sequence of firing traces from its initial marking. The terminal language generated by Π, with respect to final marking Mf , differs from Π language because only that traces from M0 to Mf are considered. Definition 11 (Classes of Petri Nets Languages) We denote the class of all Petri net languages as ℓλ 0 , and the class of all Petri net terminal languages as ℓλ 1 . It is simple to demonstrable that ℓλ 0 ⊂ ℓλ 1 . B.3 Interlaced Petri Nets Interlaced Petri nets are an alternative extension to labelled Petri nets, introduced in this article for simplifying specification of the translation schema of Hash programs into labelled Petri nets. Like hierarchic Petri nets [], interlaced Petri nets allows that complex and large scale Petri nets be implemented in a simpler and modular way. It does not modify descriptive power of labelled Petri nets. However, while hierarchic Petri nets allow only nesting composition of Petri nets, interlaced Petri nets also allow overlapping of them. Given a set of Petri net slices, each one addressing different concerns, they may be overlapped to form an interlaced Petri net. Interlaced Petri nets may also be viewed as Petri net slices for composing higher level interlaced Petri nets. Definition 12 (Interlaced Petri Nets) An interlaced Petri net Π can be defined inductively as: 1. (Base) A tuple (P, T, A, M0 , ρ, δ), where (P, T, A, M0 , ρ) represents a labelled Petri net and δ : (P ∪ T ) → 2∆ is a function that maps Petri net nodes (places and transitions) onto a list of qualifiers, represents a simple interlaced Petri net; 47 2. (Induction) Let Π1 , Π2 , . . . , Πn be interlaced Petri nets, called Petri net slices in this context ( hypothesis). Π, such that Π = hΠ1 , Π2 , . . . , Πn i, is a composed interlaced Petri net ( induction step). 3. Anything that may not be formed from application of inductive rules 1 and 2 is not an interlaced Petri net. Definition 13 (Unfolding Composite Interlaced Petri Nets) An unfolded interlaced Petri net Π̆ is obtained from a interlaced petri net Π by applying the transformation function µ, defined below: 1. µ [(P, T, A, M0 , ρ, δ)] = (P, T, A, M0 , ρ, δ) ˘ µ [Π2 ] ∪ ˘ ...∪ ˘ µ [Πn ] (composite) 2. µ [< Π1 , Π2 , . . . , Πn >] = µ [Π1 ] ∪ ˘ correspond to the Petri net union operator, defined The binary operator ∪ as below: ˘ (P1 , T1 , A1 , M0 1 , ρ1 , δ1 )∪(P 2 , T2 , A2 , M02 , ρ2 , δ2 ) = (P1 ∪ P2 , T1 ∪ T2 , A1 ∪ A2 , M01 ∪ M0 2 , ρ1 ∪ ρ2 , δ1 ∪ δ2 ) Qualifiers are used to identify components of interlaced Petri nets (places or transitions) that must be treated as the same component. These components are said to be equivalent components. They may belong to distinct slices. Next, qualifiers and identification rules are formalized. Definition 14 (Qualifier) Let Θ be a finite set of symbols. The set ∆ is informally defined as all qualifiers that can be induced from set Θ. A qualifier δ, δ ∈ ∆, is defined from Θ as following: • a, a ∈ Θ, is a qualifier ( primitive qualifier); • A tuple (a1 , a2 , . . . , an ) is a qualifier ( composed qualifier), assuming that each ai , 1 ≤ i ≤ n, is a primitive qualifier; The following rule teaches how to identify equivalent vertices in an interlaced Petri net, using their qualifiers. Definition 15 (Identification of Vertices (Places or Transitions)) Let Π be an interlaced Petri net. Consider its unfolded variant µ [Π] = (P, T, A, M0 , ρ, δ). Let V be the collection of components of µ [Π] (P ∪T ). The following equivalence relation is defined between two vertices v1 and v2 ({v1 , v2 } ⊂ V ): v1 ≡ v2 ⇔ ({v1 , v2 } ⊂ T ∨ {v1 , v2 } ⊂ P ) ∧ (δ(v1 ) ∩ δ(v2 ) 6= ∅) 48 B.4 Regular Expressions Controlled by Balanced Semaphores (SCRE) The following definitions are presented in order to introduce the class of regular expressions controlled by balanced semaphores (SCRE). This generalization to regular expressions, with descriptive power comparable to labelled Petri nets, will be used to model a language for specification of communication behavior of Hash processes. Definition 16 (Regular Expressions) A regular expression E over an alphabet Σ is inductively defined as follows: 1. a ∈ Σ is a regular expression; 2. λ is a regular expression; 3. ∅ is a regular expression; 4. Be S1 and S2 regular expression. Thus, (S1 ), S1 · S2 , S1 + S2 e, S1 ∗ are also regular expressions. Definition 17 (Regular Language) Be E a regular expression over Σ. The formal language generated by E, LRE (E), is defined in the following way: 1. LRE (∅) = ∅ 2. LRE (λ) = {λ} 3. LRE (a) = {a}, para a ∈ Σ, 4. LRE ((E)) = LRE (E) 5. LRE (E1 .E2 ) = {xy \| x ∈ LRE (E1 ) ∧ y ∈ LRE(E2 )} 6. LRE (E1 + E2 ) = {x \| x ∈ LRE (E1 ) ∨ x ∈ LRE (S2 )} S 7. LRE (E ∗ ) = LRE (Ei ) where: 1. E 0 = λ 2. E i = E i−1 E, i > 0 The class of regular languages are denoted by RE. A previous version of the Hash language used regular expressions to model communication traces of processes [5]. However, Petri nets are far more expressive than simple regular expressions for describing communication traces. This fact motivated us to generalize the adopted approach, using some class of synchronized concurrent expressions shown to be equivalent to Petri nets. Concurrent expressions are an extension to regular expressions defined to model concurrency. Some authors refer to these languages as shuffle languages. 49 Definition 18 (Concurrent Expressions) Be Σ an alphabet. A concurrent expression S over Σ is defined by the following rules: 1. If S is a regular expression over Σ, then S is a concurrent expression; 2. If S1 e S2 are concurrent expressions, then S1 ⊙ S2 , e S1 ⊗ are concurrent expressions; The operators ⊙ and ⊗ distinguish, at least syntactically, non-regular concurrent expressions. The language generated by concurrent expressions are defined below. Definition 19 (Concurrent Languages) Be S a concurrent expression over Σ. The concurrent language LCE (S) is defined according to the following rules: 1. LCE (S) = LRE(S), if S is a well-formed regular expression. 2. LCE (S1 ⊙ S2 ) = {x1 y1 x2 y2 · · · xk yk \| x1 x2 · · · xk ∈ LCE (S1 ) ∧ y1 y2 · · · yk ∈ LCE (S2 )}, S 3. LCE (S ⊗ ) = LCE (S⊙i ). where: 1. S ⊙0 = λ 2. S ⊙i = S ⊙i−1 ⊙ S, i > 0. The class of concurrent languages are denoted by CE. It is clear that CE ⊃ RE, because every regular expression is a concurrent expression. But there is a strongest and important result that relates concurrent and regular languages. Theorem 1 ([22], Relating Concurrent and Regular Expressions) If S is a concurrent expression that does not makes use of ⊗ operator, then LCE (S) is regular. Another important result gives bounds to the expressiveness of concurrent expressions. Theorem 2 ([15], Bounds to Concurrent Expressions Expressiveness) If S is a concurrent expression, than LCE (S) is a context-sensitive language. In order to increase expressivity of concurrent expressions, allowing than to express recursively enumerable languages, synchronized concurrent expressions[18] were proposed. They extend concurrent expressions with synchronization mechanisms and has been used extensively in 80’s to analyse expressiveness of synchronization mechanisms of concurrency, mainly that based on semaphores. 50 Definition 20 (Synchronized Concurrent Expressions) Be Σ an alphabet and Ω a set of symbols that denote synchronization primitives, where Σ and Ω are disjoint. A concurrent expression E over Σ ∪ Ω is called synchronized concurrent expression and the language (LS (Ω)) is said to be a synchronization mechanism over E. A synchronized concurrent expression over Σ ∪ Ω, adopting the synchronization mechanism K, where K= LS (Ω), will be denoted by (E, Σ, Ω, K). The class of synchronized concurrent expression is denoted by SCE. The next definition defines the language of a synchronized concurrent expression, which gives a meaning for the synchronization mechanism. Definition 21 (Synchronized Concurrent Languages) Be (S, Σ, Ω, K) a synchronized concurrent expression, where K = LS (Ω). The language of S, LSCE (S) is defined as below: LSCE (S) = {h(x) \| x ∈ LCE (x), h(x) ∈ K}, where the homomorphisms h and h are defined as follwing: a a∈Σ h(a) = λ a∈Ω h(a) = λ a a∈Σ a∈Ω The class of synchronized concurrent languages that uses a synchronization mechanism K is denoted by SCEK . This definition parameterizes the adopted synchronization mechanism. The most common are that based on semaphores, whose most significative examples are presented in the following paragraphs Be Ωn a set of synchronization primitives com n sı́mbolos, where: Ωn = {ωi , σi \| i = 1, · · · , n} The following semaphore-based synchronization mechanisms can be defined over Ωn [18]. 1. Counter semaphore: LS (Ωn ) = C(n), where: C(n) = LCE ((σ1 ·ω1 + σ1 ))⊗ ⊙ (σ2 ·ω2 + sigma2 ))⊗ ⊙ · · ·⊙ (σn ·ωn + σn ))⊗ ) C = {C(n) \| n ≥ 0, whereC(0) = λ} 2. [0]-counter semaphore: LS (Ωn ) = C0 (n), where: C0 (n) = LCE ((σ1 · ω1 ))⊗ ⊙ (σ2 · ω2 ))⊗ ⊙ · · · ⊙ (σn · ωn ))⊗ ) C0 = {C0 (n) \| n ≥ 0, whereC0 (0) = λ} 51 3. Binary semaphore: LS (Ωn ) = B(n), where: B(n) = LCE ((σ1 · ω1 + σ1 ))∗ ⊙ (σ2 · ω2 + sigma2))∗ ⊙ · · · ⊙ (σn · ωn + σn ))∗ ) B = {B(n) \| n ≥ 0, whereB(0) = λ} 4. [0]-binary semaphore: LS (Ωn ) = B0 (n), where: B0 (n) = LCE ((σ1 · ω1 ))∗ ⊙ (σ2 · ω2 ))∗ ⊙ · · · ⊙ (σn · ωn ))∗ ) B0 = {B0 (n) \| n ≥ 0, whereB0 (0) = λ} C, C0 , B, B0 denote the sets of synchronization primitives with any number of primitives. This notation can be generalized to any synchronization mechanism other than semaphores. Let CE(n) beSthe family of synchronization mechanisms with n primitives. Then CE = CE(n), n = 0 . . . n. Thus, C(n), C0 (n), B(n), B0 (n) are special cases of CE(2n). A concurrent expression controlled by a semaphore system will be called as a semaphore controlled concurrent expression. Many important results about the expressive power of synchronized concurrent expressions were presented[]. But, in this work, one deserve special attention. It establishes the equivalence of synchronized regular expressions that uses [0]-counter semaphores as synchronization protocol to Petri nets. Synchronized regular expressions are defined below. Definition 22 (Regular Expressions Controlled by Balanced Semaphores) Regular expressions controlled by balanced semaphores (RECBS) are defined as [0]-counter synchronized concurrent expressions that does not make use of ⊗ operator. The class of RECBS’s is denoted by RECBS. Remember that theorem 1 guarantees that concurrent expressions without ⊗ operator are equivalent to regular expressions. However the presence of ⊙ operator and [0]-counter semaphores guarantees that this kind of expression may generate a richer class of formal languages than simple regular expressions, a fact enunciated by the following theorem. Theorem 3 (Equivalence of RECBS to Petri nets [] ) The class of languages generated by a RECBS is ℓλ 1 . The result in theorem 3 are convenient for our purpose to make descriptive power of the Hash language equivalent to descriptive power of Petri nets. The use of RECBS avoids the use of ⊗ operator. Shaw [22], when introducing flow expressions to make software descriptions gave two interpretations for ⊗: a parallel loop and as a sequential loop that creates a process (fork ) in each iteration. Neither interpretation is practical in the Hash language, because it assumes static parallelism. 52	6cs.PL
P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 arXiv:1704.07234v2 [cs.PL] 8 May 2017 Scaling Reliably: Improving the Scalability of the Erlang Distributed Actor Platform Phil Trinder, Natalia Chechina, Nikolaos Papaspyrou, Konstantinos Sagonas, Simon Thompson, Stephen Adams, Stavros Aronis, Robert Baker, Eva Bihari, Olivier Boudeville, Francesco Cesarini, Maurizio Di Stefano, Sverker Eriksson, Viktoria Fordos, Amir Ghaffari, Aggelos Giantsios, Rockard Green, Csaba Hoch, David Klaftenegger, Huiqing Li, Kenneth Lundin, Kenneth MacKenzie, Katerina Roukounaki, Yiannis Tsiouris, Kjell Winblad RELEASE EU FP7 STREP (287510) project www.release-project.eu/ April 25, 2017 Abstract Distributed actor languages are an effective means of constructing scalable reliable systems, and the Erlang programming language has a well-established and influential model. While the Erlang model conceptually provides reliable scalability, it has some inherent scalability limits and these force developers to depart from the model at scale. This article establishes the scalability limits of Erlang systems, and reports the work of the EU RELEASE project to improve the scalability and understandability of the Erlang reliable distributed actor model. We systematically study the scalability limits of Erlang, and then address the issues at the virtual machine, language and tool levels. More specifically: (1) We have evolved the Erlang virtual machine so that it can work effectively in large scale single-host multicore and NUMA architectures. We have made important changes and architectural improvements to the widely used Erlang/OTP release. (2) We have designed and implemented Scalable Distributed (SD) Erlang libraries to address language-level scalability issues, and provided and validated a set of semantics for the new language constructs. (3) To make large Erlang systems easier to deploy, monitor, and debug we have developed and made open source releases of five complementary tools, some specific to SD Erlang. Throughout the article we use two case studies to investigate the capabilities of our new technologies and tools: a distributed hash table based Orbit calculation and Ant Colony Optimisation (ACO). Chaos Monkey experiments show that two versions of ACO survive random process failure and hence that SD Erlang preserves the Erlang reliability model. While we report measurements on a range of NUMA and cluster architectures, the key scalability experiments are conducted on the Athos cluster with 256 hosts (6144 cores). Even for programs with no global recovery data to maintain, SD Erlang partitions the network to reduce network traffic and hence improves performance of the Orbit and ACO benchmarks above 80 hosts. ACO measurements show that maintaining global recovery data dramatically limits scalability; however scalability is recovered by partitioning the recovery data. We exceed the established scalability limits of distributed Erlang, and do not reach the limits of SD Erlang for these benchmarks at this scale (256 hosts, 6144 cores). 1 I. Introduction stant messaging server [77]. In Erlang, the actors are termed processes and are managed by a sophisticated Virtual Machine on a single multicore or NUMA host, while distributed Erlang provides relatively transparent distribution over networks of VMs on multiple hosts. Erlang is supported by the Open Telecom Platform (OTP) libraries that capture common patterns of reliable distributed computation, such as the client-server pattern and process supervision. Any largescale system needs scalable persistent storage and, following the CAP theorem [39], Erlang uses and indeed implements Dynamostyle NoSQL DBMS like Riak [49] and Cassandra [50]. While the Erlang distributed actor model conceptually provides reliable scalability, it has some inherent scalability limits, and indeed large-scale distributed Erlang systems must depart from the distributed Erlang paradigm in order to scale, e.g. not maintaining a fully connected graph of hosts. The EU FP7 RELEASE project set out to establish and address the scalability limits of the Erlang reliable distributed actor model [67]. After outlining related work (Section II) and the benchmarks used throughout the article (Section III) we investigate the scalability limits of Erlang/OTP, seeking to identify specific issues at the virtual machine, language and persistent storage levels (Section IV). We then report the RELEASE project work to address these issues, working at the following three levels. Distributed programming languages and frameworks are central to engineering large scale systems, where key properties include scalability and reliability. By scalability we mean that performance increases as hosts and cores are added, and by large scale we mean architectures with hundreds of hosts and tens of thousands of cores. Experience with high performance and data centre computing shows that reliability is critical at these scales, e.g. host failures alone account for around one failure per hour on commodity servers with approximately 105 cores [11]. To be usable, programming languages employed on them must be supported by a suite of deployment, monitoring, refactoring and testing tools that work at scale. Controlling shared state is the only way to build reliable scalable systems. State shared by multiple units of computation limits scalability due to high synchronisation and communication costs. Moreover shared state is a threat for reliability as failures corrupting or permanently locking shared state may poison the entire system. Actor languages avoid shared state: actors or processes have entirely local state, and only interact with each other by sending messages [2]. Recovery is facilitated in this model, since actors, like operating system processes, can fail independently without affecting the state of other actors. Moreover an actor can supervise other actors, detecting failures and taking remedial action, e.g. restarting the failed actor. Erlang [6, 19] is a beacon language for reliable scalable computing with a widely emulated distributed actor model. It has influenced the design of numerous programming languages like Clojure [43] and F# [74], and many languages have Erlang-inspired actor frameworks, e.g. Kilim for Java [73], Cloud Haskell [31], and Akka for C#, F# and Scala [64]. Erlang is widely used for building reliable scalable servers, e.g. Ericsson’s AXD301 telephone exchange (switch) [79], the Facebook chat server, and the Whatsapp in- 1. We have designed and implemented a set of Scalable Distributed (SD) Erlang libraries to address language-level reliability and scalability issues. An operational semantics is provided for the key new s_group construct, and the implementation is validated against the semantics (Section V). 2. We have evolved the Erlang virtual machine so that it can work effectively in large-scale single-host multicore and NUMA architectures. We have improved the shared 2 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 ETS tables, time management, and load balancing between schedulers. Most of these improvements are now included in the Erlang/OTP release, currently downloaded approximately 50K times each month (Section VI). 3. To facilitate the development of scalable Erlang systems, and to make them maintainable, we have developed three new tools: Devo, SDMon and WombatOAM, and enhanced two others: the visualisation tool Percept, and the refactorer Wrangler. The tools support refactoring programs to make them more scalable, easier to deploy at large scale (hundreds of hosts), easier to monitor and visualise their behaviour. Most of these tools are freely available under open source licences; the WombatOAM deployment and monitoring tool is a commercial product (Section VII). Throughout the article we use two benchmarks to investigate the capabilities of our new technologies and tools. These are a computation in symbolic algebra, more specifically an algebraic ‘orbit’ calculation that exploits a non-replicated distributed hash table, and an Ant Colony Optimisation (ACO) parallel search program (Section III). We report on the reliability and scalability implications of our new technologies using Orbit, ACO, and other benchmarks. We use a Chaos Monkey instance [14] that randomly kills processes in the running system to demonstrate the reliability of the benchmarks and to show that SD Erlang preserves the Erlang language-level reliability model. While we report measurements on a range of NUMA and cluster architectures as specified in Appendix A, the key scalability experiments are conducted on the Athos cluster with 256 hosts and 6144 cores. Having established scientifically the folklore limitations of around 60 connected hosts/nodes for distributed Erlang systems in Section 4, a key result is to show that the SD Erlang benchmarks exceed this limit and do not reach their limits on the Athos clus- ter (Section VIII). Contributions. This article is the first systematic presentation of the coherent set of technologies for engineering scalable reliable Erlang systems developed in the RELEASE project. Section IV presents the first scalability study covering Erlang VM, language, and storage scalability. Indeed we believe it is the first comprehensive study of any distributed actor language at this scale (100s of hosts, and around 10K cores). Individual scalability studies, e.g. into Erlang VM scaling [8], or language and storage scaling have appeared before [38, 37]. At the language level the design, implementation and validation of the new libraries (Section V) have been reported piecemeal [21, 60], and are included here for completeness. While some of the improvements made to the Erlang Virtual Machine (Section i) have been thoroughly reported in conference publications [66, 46, 47, 69, 70, 71], others are reported here for the first time (Sections iii, ii). In Section VII, the WombatOAM and SDMon tools are described for the first time, as is the revised Devo system and visualisation. The other tools for profiling, debugging and refactoring developed in the project have previously been published piecemeal [52, 53, 55, 10], but this is their first unified presentation. All of the performance results in Section VIII are entirely new, although a comprehensive study of SD Erlang performance is now available in a recent article by [22]. II. Context i. Scalable Reliable Programming Models There is a plethora of shared memory concurrent programming models like PThreads or Java threads, and some models, like OpenMP [20], are simple and high level. However synchronisation costs mean that these models generally do not scale well, often struggling to exploit even 100 cores. Moreover, re3 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 liability mechanisms are greatly hampered by the shared state: for example, a lock becomes permanently unavailable if the thread holding it fails. The High Performance Computing (HPC) community build large-scale (106 core) distributed memory systems using the de facto standard MPI communication libraries [72]. Increasingly these are hybrid applications that combine MPI with OpenMP. Unfortunately, MPI is not suitable for producing general purpose concurrent software as it is too low level with explicit message passing. Moreover, the most widely used MPI implementations offer no fault recovery:1 if any part of the computation fails, the entire computation fails. Currently the issue is addressed by using what is hoped to be highly reliable computational and networking hardware, but there is intense research interest in introducing reliability into HPC applications [33]. Server farms use commodity computational and networking hardware, and often scale to around 105 cores, where host failures are routine. They typically perform rather constrained computations, e.g. Big Data Analytics, using reliable frameworks like Google MapReduce [26] or Hadoop [78]. The idempotent nature of the analytical queries makes it relatively easy for the frameworks to provide implicit reliability: queries are monitored and failed queries are simply re-run. In contrast, actor languages like Erlang are used to engineer reliable general purpose computation, often recovering failed stateful computations. ii. Actor Languages The actor model of concurrency consists of independent processes communicating by means of messages sent asynchronously between processes. A process can send a message to any other process for which it has the address (in Erlang the “process identifier” or pid), and the remote process may reside on a different host. While the notion of actors originated 1 Some fault tolerance is provided in less widely used MPI implementations like [27]. 4 in AI [42], it has been used widely as a general metaphor for concurrency, as well as being incorporated into a number of niche programming languages in the 1970s and 80s. More recently it has come back to prominence through the rise of not only multicore chips but also larger-scale distributed programming in data centres and the cloud. With built-in concurrency and data isolation, actors are a natural paradigm for engineering reliable scalable general-purpose systems [1, 41]. The model has two main concepts: actors that are the unit of computation, and messages that are the unit of communication. Each actor has an address-book that contains the addresses of all the other actors it is aware of. These addresses can be either locations in memory, direct physical attachments, or network addresses. In a pure actor language, messages are the only way for actors to communicate. After receiving a message an actor can do the following: (i) send messages to another actor from its address-book, (ii) create new actors, or (iii) designate a behaviour to handle the next message it receives. The model does not impose any restrictions in the order in which these actions must be taken. Similarly, two messages sent concurrently can be received in any order. These features enable actor based systems to support indeterminacy and quasicommutativity, while providing locality, modularity, reliability and scalability [41]. Actors are just one message-based paradigm, and other languages and libraries have related message passing paradigms. Recent example languages include Go [28] and Rust [62] that provide explicit channels, similar to actor mailboxes. Probably the most famous message passing library is MPI [72], with APIs for many languages and widely used on clusters and High Performance Computers. It is, however, arguable that the most important contribution of the actor model is the one-way asynchronous communication [41]. Messages are not coupled with the sender, and neither they are transferred synchronously to a temporary container where transmission takes place, e.g. a buffer, a queue, or a mailbox. Once a message is sent, P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 the receiver is the only entity responsible for that message. Erlang [6, 19] is the pre-eminent programming language based on the actor model, having a history of use in production systems, initially with its developer Ericsson and then more widely through open source adoption. There are now actor frameworks for many other languages; these include Akka for C#, F# and Scala [64], CAF2 for C++, Pykka3 , Cloud Haskell [31], PARLEY [51] for Python, and Termite Scheme [35], and each of these is currently under active use and development. Moreover, the recently defined Rust language [62] has a version of the actor model built in, albeit in an imperative context. iii. Erlang’s Support for Concurrency In Erlang, actors are termed processes, and virtual machines are termed nodes. The key elements of the actor model are: fast process creation and destruction; lightweight processes, e.g. enabling 106 concurrent processes on a single host with 8GB RAM; fast asynchronous message passing with copying semantics; process monitoring; strong dynamic typing, and selective message reception. By default Erlang processes are addressed by their process identifier (pid), e.g. Pong_PID = spawn(fun some_module:pong/0) spawns a process to execute the anonymous function given as argument to the spawn primitive, and binds Pong_PID to the new process identifier. Here the new process will execute the pong/0 function which is defined in some_module. A subsequent call Pong_PID ! finish sends the messaged finish to the process identified by Pong_PID. Alternatively, processes can be given names using a call of the form: register(my_funky_name, Pong_PID) 2 http://actor-framework.org 3 http://pykka.readthedocs.org/en/latest/ which registers this process name in the node’s process name table if not already present. Subsequently, these names can be used to refer to or communicate with the corresponding processes (e.g. send them a message): my_funky_process ! hello. A distributed Erlang system executes on multiple nodes, and the nodes can be freely deployed across hosts, e.g. they can be located on the same or different hosts. To help make distribution transparent to the programmer, when any two nodes connect they do so transitively, sharing their sets of connections. Without considerable care this quickly leads to a fully connected graph of nodes. A process may be spawned on an explicitly identified node, e.g. Remote_Pong_PID = spawn(some_node, fun some_module:pong/0). After this, the remote process can be addressed just as if it were local. It is a significant burden on the programmer to identify the remote nodes in large systems, and we will return to this in Sections ii, i.2. iv. Scalability and Erlang Systems Reliability in Erlang was designed to solve a particular set of problems, namely those in building telecommunications’ infrastructure, where systems need to be scalable to accommodate hundreds of thousands of calls concurrently, in soft real-time. These systems need to be highlyavailable and reliable: i.e. to be robust in the case of failure, which can come from software or hardware faults. Given the inevitability of the latter, Erlang adopts the “let it fail” philosophy for error handling. That is, encourage programmers to embrace the fact that a process may fail at any point, and have them rely on the supervision mechanism, discussed shortly, to handle the failures. Figure 1 illustrates Erlang’s support for concurrency, multicores and distribution. Each Erlang node is represented by a yellow shape, 5 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Figure 1: Conceptual view of Erlang’s concurrency, multicore support and distribution. and each rectangle represents a host with an IP address. Each red arc represents a connection between Erlang nodes. Each node can run on multiple cores, and exploit the inherent concurrency provided. This is done automatically by the VM, with no user intervention needed. Typically each core has an associated scheduler that schedules processes; a new process will be spawned on the same core as the process that spawns it, but work can be moved to a different scheduler through a work-stealing allocation algorithm. Each scheduler allows a process that is ready to compute at most a fixed number of computation steps before switching to another. Erlang built-in functions or BIFs are implemented in C, and at the start of the project were run to completion once scheduled, causing performance and responsiveness problems if the BIF had a long execution time. Scaling in Erlang is provided in two different ways. It is possible to scale within a single node by means of the multicore virtual machine exploiting the concurrency provided by the multiple cores or NUMA nodes. It is also possible to scale across multiple hosts using multiple distributed Erlang nodes. Reliability in Erlang is multi-faceted. As in all actor languages each process has private state, preventing a failed or failing process from corrupting the state of other processes. Messages enable stateful interaction, and contain a deep copy of the value to be shared, with no references (e.g. pointers) to the senders’ internal state. Moreover Erlang avoids type errors by enforcing strong typing, albeit dynamically [7]. Connected nodes check liveness 6 with heartbeats, and can be monitored from outside Erlang, e.g. by an operating system process. However, the most important way to achieve reliability is supervision, which allows a process to monitor the status of a child process and react to any failure, for example by spawning a substitute process to replace a failed process. Supervised processes can in turn supervise other processes, leading to a supervision tree. The supervising and supervised processes may be in different nodes, and on different hosts, and hence the supervision tree may span multiple hosts or nodes. To provide reliable distributed service registration, a global namespace is maintained on every node, which maps process names to pids. It is this that we mean when we talk about a ‘reliable’ system: it is one in which a named process in a distributed system can be restarted without requiring the client processes also to be restarted (because the name can still be used for communication). To see global registration in action, consider a pong server process global:register_name(pong_server, Remote_Pong_PID). Clients of the server can send messages to the registered name, e.g. global:whereis_name(pong_server)!finish. If the server fails the supervisor can spawn a replacement server process with a new pid and register it with the same name (pong_server). Thereafter client messages to the pong_server will be delivered to the new server process. We return to discuss the scalability limitations of maintaining a global namespace in Section V. v. ETS: Erlang Term Storage Erlang is a pragmatic language and the actor model it supports is not pure. Erlang processes, besides communicating via asynchronous message passing, can also share data in public memory areas called ETS tables. P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 The Erlang Term Storage (ETS) mechanism is a central component of Erlang’s implementation. It is used internally by many libraries and underlies the in-memory databases. ETS tables are key-value stores: they store Erlang tuples where one of the positions in the tuple serves as the lookup key. An ETS table has a type that may be either set, bag or duplicate_bag, implemented as a hash table, or ordered_set which is currently implemented as an AVL tree. The main operations that ETS supports are table creation, insertion of individual entries and atomic bulk insertion of multiple entries in a table, deletion and lookup of an entry based on some key, and destructive update. The operations are implemented as C built-in functions in the Erlang VM. The code snippet below creates a set ETS table keyed by the first element of the entry; atomically inserts two elements with keys some_key and 42; updates the value associated with the table entry with key 42; and then looks up this entry. Table = ets:new(my_table, [set, public, {keypos, 1}]), ets:insert(Table, [ {some_key, an_atom_value}, {42, {a,tuple,value}}]), ets:update_element(Table, 42, [{2, {another,tuple,value}}]), [{Key, Value}] = ets:lookup(Table, 42). ETS tables are heavily used in many Erlang applications. This is partly due to the convenience of sharing data for some programming tasks, but also partly due to their fast implementation. As a shared resource, however, ETS tables induce contention and become a scalability bottleneck, as we shall see in Section i. III. Benchmarks for scalability and reliability The two benchmarks that we use throughout this article are Orbit, that measures scalability without looking at reliability, and Ant Colony Optimisation (ACO) that allows us to measure the impact on scalability of adding global namespaces to ensure reliability. The source code for the benchmarks, together with more detailed documentation, is available at https://github. com/release-project/benchmarks/. The RELEASE project team also worked to improve the reliability and scalability of other Erlang programs including a substantial (approximately 150K lines of Erlang code) SimDiasca simulator [16] and an Instant Messenger that is more typical of Erlang applications [23] but we do not cover these systematically here. i. Orbit Orbit is a computation in symbolic algebra, which generalises a transitive closure computation [57]. To compute the orbit for a given space [0..X ], a list of generators g1 , g2 , ..., gn are applied on an initial vertex x0 ∈ [0..X ]. This creates new values ( x1 ...xn ) ∈ [0..X ], where xi = gi ( x0 ). The generator functions are applied on the new values until no new value is generated. Orbit is a suitable benchmark because it has a number of aspects that characterise a class of real applications. The core data structure it maintains is a set and, in distributed environments is implemented as a distributed hash table (DHT), similar to the DHTs used in replicated form in NoSQL database management systems. Also, in distributed mode, it uses standard peer-to-peer (P2P) techniques like a credit-based termination algorithm [61]. By choosing the orbit size, the benchmark can be parameterised to specify smaller or larger computations that are suitable to run on a single machine (Section i) or on many nodes (Sectioni). Moreover it is only a few hundred lines of code. As shown in Figure 2, the computation is initiated by a master which creates a number of workers. In the single node scenario of the benchmark, workers correspond to processes but these workers can also spawn other processes to apply the generator functions on a subset of their input values, thus creating intraworker parallelism. In the distributed version 7 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Figure 2: Distributed Erlang Orbit (D-Orbit) architecture: workers are mapped to nodes. of the benchmark, processes are spawned by the master node to worker nodes, each maintaining a DHT fragment. A newly spawned process gets a share of the parent’s credit, and returns this on completion. The computation is finished when the master node collects all credit, i.e. all workers have completed. ii. Ant Colony Optimisation (ACO) Ant Colony Optimisation [29] is a metaheuristic which has been applied to a large number of combinatorial optimisation problems. For the purpose of this article, we have applied it to an NP-hard scheduling problem known as the Single Machine Total Weighted Tardiness Problem (SMTWTP) [63], where a number of jobs of given lengths have to be arranged in a single linear schedule. The goal is to minimise the cost of the schedule, as determined by certain constraints. The ACO method is attractive from the point of view of distributed computing because it can benefit from having multiple cooperating colonies, each running on a separate compute node and consisting of multiple “ants”. Ants are simple computational agents which concur- Figure 3: Distributed Erlang Two-level Ant Colony Optimisation (TL-ACO) architecture. 8 rently compute possible solutions to the input problem guided by shared information about good paths through the search space; there is also a certain amount of stochastic variation which allows the ants to explore new directions. Having multiple colonies increases the number of ants, thus increasing the probability of finding a good solution. We implement four distributed coordination patterns for the same multi-colony ACO computation as follows. In each implementation, the individual colonies perform some number of local iterations (i.e. generations of ants) and then report their best solutions; the globallybest solution is then selected and is reported to the colonies, which use it to update their pheromone matrices. This process is repeated for some number of global iterations. Two-level ACO (TL-ACO) has a single master node that collects the colonies’ best solutions and distributes the overall best solution back to the colonies. Figure 3 depicts the process and node placements of the TL-ACO in a cluster with NC nodes. The master process spawns NC colony processes on available nodes. In the next step, each colony process spawns NA ant processes on the local node. Each ant iterates I A times, returning its result to the colony master. Each colony iterates I M times, reporting their best solution to, and receiving the globally-best solution from, the master process. We validated the implementation by applying TL-ACO to a number of standard SMTWTP instances [25, 13, 34], obtaining good results in all cases, and confirmed that the number of perfect solutions increases as we increase the number of colonies. Multi-level ACO (ML-ACO). In TL-ACO the master node receives messages from all of the colonies, and thus could become a bottleneck. ML-ACO addresses this by having a tree of submasters (Figure 4), with each node in the bottom level collecting results from a small number of colonies. These are then fed up through the tree, with nodes at higher levels selecting the best solutions from their children. Globally Reliable ACO (GR-ACO). In ML-ACO if a single colony fails to report back the sys- P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Figure 4: Distributed Erlang Multi-level Ant Colony Optimisation (ML-ACO) architecture. tem will wait indefinitely. GR-ACO adds fault tolerance, supervising colonies so that a faulty colony can be detected and restarted, allowing the system to continue execution. Scalable Reliable ACO (SR-ACO) also adds fault-tolerance, but using supervision within our new s_groups from Section i.1, and the architecture of SR-ACO is discussed in detail there. i. Scaling Erlang on a Single Host To investigate Erlang scalability we built BenchErl, an extensible open source benchmark suite with a web interface.4 BenchErl shows how an application’s performance changes when resources, like cores or schedulers, are added; or when options that control these resources change: • the number of nodes, i.e. the number of Erlang VMs used, typically on multiple hosts; IV. Erlang Scalability Limits This section investigates the scalability of Erlang at VM, language, and persistent storage levels. An aspect we choose not to explore is the security of large scale systems where, for example, one might imagine providing enhanced security for systems with multiple clusters or cloud instances connected by a Wide Area Network. We assume that existing security mechanisms are used, e.g. a Virtual Private Network. • the number of cores per node; • the number of schedulers, i.e. the OS threads that execute Erlang processes in parallel, and their binding to the topology of the cores of the underlying computer node; • the Erlang/OTP release and flavor; and • the command-line arguments used to start the Erlang nodes. 4 Information about BenchErl is available at http:// release.softlab.ntua.gr/bencherl/. 9 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 20000 with intra-worker parallelism, 10048, 128 without intra-worker parallelism, 10048, 128 18000 16000 time (ms) 14000 12000 10000 8000 6000 4000 2000 0 0 10 20 30 40 50 60 70 60 70 # Schedulers 40 with intra-worker parallelism, 10048, 128 without intra-worker parallelism, 10048, 128 35 30 Speedup 25 20 15 10 5 0 0 10 20 30 40 50 # Schedulers Figure 5: Runtime and speedup of two configurations of the Orbit benchmark using Erlang/OTP R15B01. Using BenchErl, we investigated the scalability of an initial set of twelve benchmarks and two substantial Erlang applications using a single Erlang node (VM) on machines with up to 64 cores, including the Bulldozer machine specified in Appendix A. This set of experiments, reported by [8], confirmed that some programs scaled well in the most recent Erlang/OTP release of the time (R15B01) but also revealed VM and language level scalability bottlenecks. Figure 5 shows runtime and speedup curves for the Orbit benchmark where master and workers run on a single Erlang node in configurations with and without intra-worker parallelism. In both configurations the program scales. Runtime continuously decreases as we add more schedulers to exploit more cores. The speedup of the benchmark without intraworker parallelism, i.e. without spawning additional processes for the computation (green 10 curve), is almost linear up to 32 cores but increases less rapidly from that point on; we see a similar but more clearly visible pattern for the other configuration (red curve) where there is no performance improvement beyond 32 schedulers. This is due to the asymmetric characteristics of the machine’s micro-architecture, which consists of modules that couple two conventional x86 out-of-order cores that share the early pipeline stages, the floating point unit, and the L2 cache with the rest of the module [3]. Some other benchmarks, however, did not scale well or experienced significant slowdowns when run in many VM schedulers (threads). For example the ets_test benchmark has multiple processes accessing a shared ETS table. Figure 6 shows runtime and speedup curves for ets_test on a 16-core (eight cores with hyperthreading) Intel Xeon-based machine. It shows that runtime increases beyond two schedulers, and that the program exhibits a slowdown instead of a speedup. For many benchmarks there are obvious reasons for poor scaling like limited parallelism in the application, or contention for shared resources. The reasons for poor scaling are less obvious for other benchmarks, and it is exactly these we have chosen to study in detail in subsequent work [8, 46]. A simple example is the parallel BenchErl benchmark, that spawns some n processes, each of which creates a list of m timestamps and, after it checks that each timestamp in the list is strictly greater than the previous one, sends the result to its parent. Figure 7 shows that up to eight cores each additional core leads to a slowdown, thereafter a small speedup is obtained up to 32 cores, and then again a slowdown. A small aspect of the benchmark, easily overlooked, explains the poor scalability. The benchmark creates timestamps using the erlang:now/0 built-in function, whose implementation acquires a global lock in order to return a unique timestamp. That is, two calls to erlang:now/0, even from different processes are guaranteed to produce monotonically increasing values. This lock is precisely the bot- P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 10000 160000 [2512,1264,640] 9000 8000 120000 7000 100000 time (ms) time (ms) [70016,640] 140000 6000 5000 80000 60000 4000 40000 3000 20000 2000 1000 0 0 2 4 6 8 10 12 14 16 0 10 20 30 # Schedulers 40 50 60 70 # Schedulers 1.6 1 [2512,1264,640] [70016,640] 0.9 1.4 0.8 1.2 Speedup Speedup 0.7 1 0.8 0.6 0.5 0.4 0.6 0.3 0.4 0.2 0.2 0.1 0 2 4 6 8 10 12 14 16 # Schedulers 0 10 20 30 40 50 60 70 # Schedulers Figure 6: Runtime and speedup of the ets_test BenchErl benchmark using Erlang/OTP R15B01. Figure 7: Runtime and speedup of the BenchErl benchmark called parallel using Erlang/OTP R15B01. tleneck in the VM that limits the scalability of this benchmark. We describe our work to address VM timing issues in Figure iii. tion as there is no need to discover nodes, and the design works well for small numbers of nodes. However at emergent server architecture scales, i.e. hundreds of nodes, this design becomes very expensive and system architects must switch from the default Erlang model, e.g. they need to start using hidden nodes that do not share connection sets. We have investigated the scalability limits imposed by network connectivity costs using several Orbit calculations on two large clusters: Kalkyl and Athos as specified in Appendix A. The Kalkyl results are discussed by [21], and Figure 29 in Section i shows representative results for distributed Erlang computing orbits with 2M and 5M elements on Athos. In all cases performance degrades beyond some scale (40 nodes for the 5M orbit, and 140 nodes for the 2M orbit). Figure 32 illustrates the additional network traffic induced by the fully connected network. It allows a comparison be- Discussion Our investigations identified contention for shared ETS tables, and for commonly-used shared resources like timers, as the key VM-level scalability issues. Section VI outlines how we addressed these issues in recent Erlang/OTP releases. ii. Distributed Erlang Scalability Network Connectivity Costs When any normal distributed Erlang nodes communicate, they share their connection sets and this typically leads to a fully connected graph of nodes. So a system with n nodes will maintain O(n2 ) connections, and these are relatively expensive TCP connections with continual maintenance traffic. This design aids transparent distribu- 11 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 tween the number of packets sent in a fully connected network (ML-ACO) with those sent in a network partitioned using our new s_groups (SR-ACO). paper [37] investigates the impact of data size, and computation time in P2P calls both independently and in combination, and the scaling properties of the common Erlang/OTP generic server processes gen_server and gen_fsm. Global Information Costs Maintaining global information is known to limit the scalability of distributed systems, and crucially the process namespaces used for reliability are global. To investigate the scalability limits imposed on distributed Erlang by such global information we have designed and implemented DE-Bench, an open source, parameterisable and scalable peer-to-peer benchmarking framework [37, 36]. DE-Bench measures the throughput and latency of distributed Erlang commands on a cluster of Erlang nodes, and the design is influenced by the Basho Bench benchmarking tool for Riak [12]. Each DE-Bench instance acts as a peer, providing scalability and reliability by eliminating central coordination and any single point of failure. To evaluate the scalability of distributed Erlang, we measure how adding more hosts increases the throughput, i.e. the total number of successfully executed distributed Erlang commands per experiment. Figure 8 shows the parameterisable internal workflow of DEBench. There are three classes of commands in DE-Bench:(i) Point-to-Point (P2P) commands, where a function with tunable argument size and computation time is run on a remote node, include spawn, rpc, and synchronous calls to server processes, i.e. gen_server or gen_fsm. (ii) Global commands, which entail synchronisation across all connected nodes, such as global:register_name and global:unregister_name. (iii) Local commands, which are executed independently by a single node, e.g. whereis_name, a look up in the local name table. The benchmarking is conducted on 10 to 100 host configurations of the Kalkyl cluster (in steps of 10) and measures the throughput of successful commands per second over 5 minutes. There is one Erlang VM on each host and one DE-Bench instance on each VM. The full Here we focus on the impact of different proportions of global commands, mixing global with P2P and local commands. Figure 9 shows that even a low proportion of global commands limits the scalability of distributed Erlang, e.g. just 0.01% global commands limits scalability to around 60 nodes. Figure 10 reports the latency of all commands and shows that, while the latencies for P2P and local commands are stable at scale, the latency of the global commands increases dramatically with scale. Both results illustrate that the impact of global operations on throughput and latency in a distributed Erlang system is severe. 12 Explicit Placement While network connectivity and global information impact performance at scale, our investigations also identified explicit process placement as a programming issue at scale. Recall from Section iii that distributed Erlang requires the programmer to identify an explicit Erlang node (VM) when spawning a process. Identifying an appropriate node becomes a significant burden for large and dynamic systems. The problem is exacerbated in large distributed systems where (1) the hosts may not be identical, having different hardware capabilities or different software installed; and (2) communication times may be non-uniform: it may be fast to send a message between VMs on the same host, and slow if the VMs are on different hosts in a large distributed system. These factors make it difficult to deploy applications, especially in a scalable and portable manner. Moreover while the programmer may be able to use platform-specific knowledge to decide where to spawn processes to enable an application to run efficiently, if the application is then deployed on a different platform, or if the platform changes as hosts fail or are added, this becomes outdated. P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Figure 8: DE-Bench’s internal workflow. Figure 9: Scalability vs. percentage of global commands in Distributed Erlang. Discussion Our investigations confirm three language-level scalability limitations of Erlang from developer folklore.(1) Maintaining a fully connected network of Erlang nodes limits scalability, for example Orbit is typically limited to just 40 nodes. (2) Global operations, and crucially the global operations required for reliability, i.e. to maintain a global namespace, seriously limit the scalability of distributed Erlang systems. (3) Explicit process placement makes it hard to built performance portable applications for large architectures. These issues cause designers of reliable large scale systems in Erlang to depart from the standard Erlang model, e.g. using techniques like hidden nodes and stor- ing pids in data structures. In Section V we develop language technologies to address these issues. iii. Persistent Storage Any large scale system needs reliable scalable persistent storage, and we have studied the scalability limits of Erlang persistent storage alternatives [38]. We envisage a typical large server having around 105 cores on around 100 hosts. We have reviewed the requirements for scalable and available persistent storage and evaluated four popular Erlang DBMS against these requirements. For a target scale 13 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Figure 10: Latency of commands as the number of Erlang nodes increases. of around 100 hosts, Mnesia and CouchDB are, unsurprisingly, not suitable. However, Dynamo-style NoSQL DBMS like Cassandra and Riak have the potential to be. We have investigated the current scalability limits of the Riak NoSQL DBMS using the Basho Bench benchmarking framework on a cluster with up to 100 nodes and independent disks. We found that that the scalability limit of Riak version 1.1.1 is 60 nodes on the Kalkyl cluster. The study placed into the public scientific domain what was previously well-evidenced, but anecdotal, developer experience. We have also shown that resources like memory, disk, and network do not limit the scalability of Riak. By instrumenting the global and gen_server OTP libraries we identified a specific Riak remote procedure call that fails to scale. We outline how later releases of Riak are refactored to eliminate the scalability bottlenecks. Discussion We conclude that Dynamo-like NoSQL DBMSs have the potential to deliver reliable persistent storage for Erlang at our target scale of approximately 100 hosts. Specifically an Erlang Cassandra interface is available and Riak 1.1.1 already provides scalable and available persistent storage on 60 nodes. Moreover 14 the scalability of Riak is much improved in subsequent versions. V. Improving Language Scalability This section outlines the Scalable Distributed (SD) Erlang libraries [21] we have designed and implemented to address the distributed Erlang scalability issues identified in Section ii. SD Erlang introduces two concepts to improve scalability. S_groups partition the set of nodes in an Erlang system to reduce network connectivity and partition global data (Section i.1). Semi-explicit placement alleviates the issues of explicit process placement in large heterogeneous networks (Section i.2). The two features are independent and can be used separately or in combination. We overview SD Erlang in Section i, and outline s_group semantics and validation in Sections ii, iii respectively. i. SD Erlang Design i.1 S_groups reduce both the number of connections a node maintains, and the size of name spaces, i.e. they minimise global information. Specifically names are registered on, and synchronised between, only the nodes within the s_group. An s_group has the following parameters: a name, P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 a list of nodes, and a list of registered names. A node can belong to many s_groups or to none. If a node belongs to no s_group it behaves as a usual distributed Erlang node. The s_group library defines the functions shown in Table 1. Some of these functions manipulate s_groups and provide information about them, such as creating s_groups and providing a list of nodes from a given s_group. The remaining functions manipulate names registered in s_groups and provide information about these names. For example, to register a process, Pid, with name Name in s_group SGroupName we use the following function. The name will only be registered if the process is being executed on a node that belongs to the given s_group, and neither Name nor Pid are already registered in that group. s_group:register_name(SGroupName, Name, Pid) -> yes \| no To illustrate the impact of s_groups on scalability we repeat the global operations experiment from Section ii (Figure 9). In the SD Erlang experiment we partition the set of nodes into s_groups each containing ten nodes, and hence the names are replicated and synchronised on just ten nodes, and not on all nodes as in distributed Erlang. The results in Figure 11 show that with 0.01% of global operations throughput of distributed Erlang stops growing at 40 nodes while throughput of SD Erlang continues to grow linearly. The connection topology of s_groups is extremely flexible: they may be organised into a hierarchy of arbitrary depth or branching, e.g. there could be multiple levels in the tree of s_groups; see Figure 12. Moreover it is not necessary to create an hierarchy of s_groups, for example, we have constructed an Orbit implementation using a ring of s_groups. Given such a flexible way of organising distributed systems, key questions in the design of an SD Erlang system are the following. How should s_groups be structured? Depending on the reason the nodes are grouped – reducing the number of connections, or reducing the namespace, or both – s_groups can be freely structured as a tree, ring, or some other topology. How large should the s_groups be? Smaller s_groups mean more inter-group communication, but the synchronisation of the s_group state between the s_group nodes constrains the maximum size of s_groups. We have not found this constraint to be a serious restriction. For example many s_groups are either relatively small, e.g. 10-node, internal or terminal elements in some topology, e.g. the leaves and nodes of a tree. How do nodes from different s_groups communicate? While any two nodes can communicate in an SD Erlang system, to minimise the number of connections communication between nodes from different s_groups is typically routed via gateway nodes that belong to both s_groups. How do we avoid single points of failure? For reliability, and to minimise communication load, multiple gateway nodes and processes may be required. Information to make these design choices is provided by the tools in Section VII and by benchmarking. A further challenge is how to systematically refactor a distributed Erlang application into SD Erlang, and this is outlined in Section i. A detailed discussion of distributed system design and refactoring in SD Erlang provided in a recent article [22]. We illustrate typical SD Erlang system designs by showing refactorings of the Orbit and ACO benchmarks from Section III. In both distributed and SD Erlang the computation starts on the Master node and the actual computation is done on the Worker nodes. In the distributed Erlang version all nodes are interconnected, and messages are transferred directly from the sending node to the receiving node (Figure 2). In contrast, in the SD Erlang version nodes are grouped into s_groups, and messages are transferred between different s_groups via Submaster nodes (Figure 13). A fragment of code that creates an s_group on a Sub-master node is as follows: create_s_group(Master, GroupName, Nodes0) -> case s_group:new_s_group(GroupName, Nodes0) of {ok, GroupName, Nodes} -> 15 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Table 1: Summary of s_group functions. Function Description new_s_group(SGroupName, Nodes) Creates a new s_group consisting of some nodes. delete_s_group(SGroupName) Deletes an s_group. add_nodes(SGroupName, Nodes) Adds a list of nodes to an s_group. remove_nodes(SGroupName, Nodes) Removes a list of nodes from an s_group. s_groups() Returns a list of all s_groups known to the node. own_s_groups() Returns a list of s_group tuples the node belongs to. own_nodes() Returns a list of nodes from all s_groups the node belongs to. own_nodes(SGroupName) Returns a list of nodes from the given s_group. info() Returns s_group state information. register_name(SGroupName, Name, Pid) Registers a name in the given s_group. re_register_name(SGroupName, Name, Pid) Re-registers a name (changes a registration) in a given s_group. unregister_name(SGroupName, Name) Unregisters a name in the given s_group. registered_names({node,Node}) Returns a list of all registered names on the given node. registered_names({s_group,SGroupName}) Returns a list of all registered names in the given s_group. whereis_name(SGroupName, Name) Return the pid of a name registered in the given whereis_name(Node, SGroupName, Name) s_group. send(SGroupName, Name, Msg) Send a message to a name registered in the given send(Node, SGroupName, Name, Msg) s_group. Master ! {GroupName, Nodes}; _ -> io:format("exception: message") end. distributed Erlang and SD Erlang Orbit and ACO is presented in Section VIII. Similarly, we introduce s_groups in the GR-ACO benchmark from Section ii to create Scalable Reliable ACO (SR-ACO); see Figure 12. Here, apart from reducing the number of connections, s_groups also reduce the global namespace information. That is, instead of registering the name of a pid globally, i.e. with all nodes, the names is registered only on all nodes in the s_group with s_group:register_name/3. A comparative performance evaluation of i.2 16 Semi-Explicit Placement Recall from Section iii that distributed Erlang spawns a process onto an explicitly named Erlang node, e.g. spawn(some_node,fun some_module:pong/0). and also recall the portability and programming effort issues associated with such explicit placement in large scale systems discussed in Section ii. P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 1e+09 9e+08 Throughput 7e+08 5e+08 3e+08 1e+08 10 20 30 40 50 60 70 80 90 100 Number of nodes Distributed Erlang SD Erlang Figure 11: Global operations in Distributed Erlang vs. SD Erlang. represents a sub-master process represents a colony process master node Level 0 master process represents an ant process P1 represents an s_group P2 Level 1 P16 represents a node node1 node2 node16 node1 node1 node16 node1 node1 node16 colony nodes in this level Level 2 Figure 12: SD Erlang ACO (SR-ACO) architecture. To address these issues we have developed a semi-explicit placement library that enables the programmer to select nodes on which to spawn processes based on run-time information about the properties of the nodes. For example, if a process performs a lot of computation one would like to spawn it on a node with considerable computation power, or if two processes are likely to communicate frequently then it would be desirable to spawn them on the same node, or nodes with a fast interconnect. ated hosts, such as total and currently available RAM, installed software, hardware configuration, etc. The second deals with a notion of communication distances which models the communication times between nodes in a distributed system. Therefore, instead of specifying a node we can use the attr:choose_node/1 function to define the target node, i.e. We have implemented two Erlang libraries to support semi-explicit placement [60]. The first deals with node attributes, and describes properties of individual Erlang VMs and associ- [60] report an investigation into the communication latencies on a range of NUMA and cluster architectures, and demonstrate the effectiveness of the placement libraries using the spawn(attr:choose_node(Params), fun some_module:pong/0). 17 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Figure 13: SD Erlang (SD-Orbit) architecture. ML-ACO benchmark on the Athos cluster. ii. S_group Semantics For precise specification, and as a basis for validation, we provide a small-step operational semantics of the s_group operations [21]. Figure 14 defines the state of an SD Erlang system and associated abstract syntax variables. The abstract syntax variables on the left are defined as members of sets, denoted {}, and these in turn may contain tuples, denoted (), or further sets. In particular nm is a process name, p a pid, ni a node_id, and nis a set of node_ids. The state of a system is modelled as a four tuple comprising a set of s_groups, a set of f ree_groups, a set of f ree_hidden_groups, and a set of nodes. Each type of group is associated with nodes and has a namespace. An s_group additionally has a name, whereas a f ree_hidden_group consists of only one node, i.e. a hidden node simultaneously acts as a node and as a group, because as a group it has a namespace but does not share it with any other node. Free normal and hidden groups have no names, and are uniquely defined by the nodes associated with them. Therefore, group names, gr_names, are either NoGroup or a set of s_group_names. A namespace is a set of name and process id, pid, pairs and is 18 replicated on all nodes of the associated group. A node has the following four parameters: node_id identifier, node_type that can be either hidden or normal, connections, and group_names, i.e. names of groups the node belongs to. The node can belong to either a list of s_groups or one of the free groups. The type of the free group is defined by the node type. Connections are a set of node_ids. Transitions in the semantics have the form (state, command, ni) −→ (state0 , value) meaning that executing command on node ni in state returns value and transitions to state0 . The semantics is presented in more detail by [21], but we illustrate it here with the s_group:registered_names/1 function from Section i.1. The function returns a list of names registered in s_group s if node ni belongs to the s_group, an empty list otherwise (Figure 15). Here ⊕ denotes disjoint set union; IsSGroupNode returns true if node ni is a member of some s_group s, false otherwise; and OutputNms returns a set of process names registered in the ns namespace of s_group s. iii. Semantics Validation As the semantics is concrete it can readily be made executable in Erlang, with lists replacing sets throughout. Having an executable seman- P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 (grs, fgs, fhs, nds) ∈ {state} ≡ {s_group}, {free_group}, {free_hidden_group}, {node} gr ∈ grs ≡ {s_group} ≡ s_group_name, {node_id}, namespace fg ∈ fgs ≡ {free_group} ≡ {node_id}, namespace fh ∈ fhs ≡ {free_hidden_group} ≡ node_id, namespace nd ∈ nds ≡ {node} ≡ node_id, node_type, connections, gr_names gs ∈ { gr_names} ≡ NoGroup, {s_group_name} ns ∈ {namespace} ≡ {(name, pid)} cs ∈ {connections} ≡ {node_id} nt ∈ {node_type} ≡ { Normal, Hidden} s ∈ { NoGroup, s_group_name} Figure 14: SD Erlang state [21]. (grs,fgs, fhs, nds), s_group : registered_names(s), ni −→ (grs, fgs, fhs, nds), nms) −→ (grs, fgs, fhs, nds), {} where (s, {ni} ⊕ nis, ns) ⊕ grs0 ≡ grs nms ≡ OutputNms(s, ns) if IsSGroupNode(ni, s, grs otherwise IsSGroupNode(ni, s, grs) = ∃nis, ns, grs0 . (s, {ni} ⊕ nis, ns) ⊕ grs0 ≡ grs OutputNms(s, ns) = (s, nm) \| (nm, p) ∈ ns Figure 15: SD Erlang Semantics of s_group:registered_names/1 Function tics allows users to engage with it, and to understand how the semantics behaves vis à vis the library, giving them an opportunity to assess the correctness of the library against the semantics. Better still, we can automatically assess how the system behaves in comparison with the (executable) semantics by executing them in lockstep, guided by the constraints of which operations are possible at each point. We do that by building an abstract state machine model of the library. We can then generate random sequences (or traces) through the model, with appropriate library data gener- ated too. This random generation is supported by the QuickCheck property-based testing system [24, 9]. The architecture of the testing framework is shown in Figure 16. First an abstract state machine embedded as an “eqc_statem” module is derived from the executable semantic specification. The state machine defines the abstract state representation and the transition from one state to another when an operation is applied. Test case and data generators are then defined to control the test case generation; this includes the automatic generation of eligible s_group operations and the input data to 19 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 those operations. Test oracles are encoded as the postcondition for s_group operations. During testing, each test command is applied to both the abstract model and the s_group library. The application of the test command to the abstract model takes the abstract model from its current state to a new state as described by the transition functions; whereas the application of the test command to the library leads the system to a new actual state. The actual state information is collected from each node in the distributed system, then merged and normalised to the same format as the abstract state representation. For a test to be successful, after the execution of a test command, the test oracles specified for this command should be satisfied. Various test oracles can be defined for s_group operations; for instance one of the generic constraints that applies to all the s_group operations is that after each s_group operation, the normalised system state should be equivalent to the abstract state. Thousands of tests were run, and three kinds of errors — which have subsequently been corrected — were found. Some errors in the library implementation were found, including one error due to the synchronisation between nodes, and the other related to the remove_nodes operation, which erroneously raised an exception. We also found a couple of trivial errors in the semantic specification itself, which had been missed by manual examination. Finally, we found some situations Figure 16: Testing s_groups using QuickCheck. 20 where there were inconsistencies between the semantics and the library implementation, despite their states being equivalent: an example of this was in the particular values returned by functions on certain errors. Overall, the automation of testing boosted our confidence in the correctness of the library implementation and the semantic specification. This work is reported in more detail by [54]. VI. Improving VM Scalability This section reports the primary VM and library improvements we have designed and implemented to address the scalability and reliability issues identified in Section i. i. Improvements to Erlang Term Storage Because ETS tables are so heavily used in Erlang systems, they are a focus for scalability improvements. We start by describing their redesign, including some improvements that pre-date our RELEASE project work, i.e. those prior to Erlang/OTP R15B03. These historical improvements are very relevant for a scalability study and form the basis for our subsequent changes and improvements. At the point when Erlang/OTP got support for multiple cores (in release R11B), there was a single reader-writer lock for each ETS table. Optional fine grained locking of hash-based ETS tables (i.e. set, bag or duplicate_bag tables) was introduced in Erlang/OTP R13B02-1, adding 16 reader-writer locks for the hash buckets. Reader groups to minimise read synchronisation overheads were introduced in Erlang/OTP R14B. The key observation is that a single count of the multiple readers must be synchronised across many cache lines, potentially far away in a NUMA system. Maintaining reader counts in multiple (local) caches makes reads fast, although writes must now check every reader count. In Erlang/OTP R16B the number of bucket locks, and the default number of reader groups, were both upgraded from 16 to 64. We illustrate the scaling properties of the P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 50 12 R11B-5 R13B02-1 R14B R16B 45 40 1 RG 2 RG 4 RG 8 RG 16 RG 32 RG 64 RG 10 35 8 30 6 25 20 4 15 10 2 5 0 0 60 10 20 30 40 50 60 70 R11B-5 R13B02-1 R14B R16B 50 0 0 10 20 30 40 50 14 12 10 40 60 70 1 RG 2 RG 4 RG 8 RG 16 RG 32 RG 64 RG 8 30 6 20 4 10 0 2 0 10 20 30 40 50 60 70 0 0 10 20 30 40 50 60 70 Figure 17: Runtime of 17M ETS operations in Erlang/OTP releases: 10% updates (left) and 1% updates (right). The y-axis shows time in seconds (lower is better) and the x-axis is number of OS threads (VM schedulers). Figure 18: Runtime of 17M ETS operations with varying numbers of reader groups: 10% updates (left) and 1% updates (right). The y-axis shows runtime in seconds (lower is better) and the x-axis is number of schedulers. ETS concurrency options using the ets_bench BenchErl benchmark on the Intel NUMA machine with 32 hyperthreaded cores specified in Appendix A. The ets_bench benchmark inserts 1M items into the table, then records the time to perform 17M operations, where an operation is either a lookup, an insert, or a delete. The experiments are conducted on a hash-based (set) ETS table with different percentages of update operations, i.e. insertions or deletions. Figure 17 shows the runtimes in seconds of 17M operations in different Erlang/OTP versions, varying the number of schedulers (xaxis), reflecting how the scalability of ETS tables has improved in more recent Erlang/OTP releases. Figure 18 shows the runtimes in seconds of 17M operations on an ETS table with different numbers of reader groups, again varying the number of schedulers. We see that one reader group is not sufficient with 10% updates, nor are two with 1% updates. Beyond that, different numbers of reader groups have little impact on the benchmark performance except that using 64 groups with 10% updates slightly degrades performance. We have explored four other extensions or redesigns in the ETS implementation for better scalability. 1. Allowing more programmer control over the number of bucket locks in hashbased tables, so the programmer can reflect the number of schedulers and the expected access pattern. 2. Using contention-adapting trees to get better scalability for ordered_set ETS tables as described by [69]. 3. Using queue delegation locking to improve scalability [47]. 4. Adopting schemes for completely eliminating the locks in the meta table. A more complete discussion of our work on ETS can be found in the papers by [69] and [47]. Here we outline only our work on contention-adapting (CA) trees. A CA tree monitors contention in different parts of a 21 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 ii. Operations / Microsecond 20 15 10 set ordset AVL-CA tree Treap-CA tree 5 0 0 10 20 30 40 Number of Threads 50 60 70 30 40 Number of Threads 50 60 70 Operations / Microsecond 25 20 set ordset AVL-CA tree Treap-CA tree 15 10 5 0 0 10 20 Figure 19: Throughput of CA tree variants: 10% updates (left) and 1% updates (right). tree-shaped data structure, introducing routing nodes with locks in response to high contention, and removing them in response to low contention. For experimental purposes two variants of the CA tree have been implemented to represent ordered_sets in the virtual machine of Erlang/OTP 17.0. One extends the existing AVL trees in the Erlang VM, and the other uses a Treap data structure [5]. Figure 19 compares the throughput of the CA tree variants with that of ordered_set and set as the number of schedulers increases. It is unsurprising that the CA trees scale so much better than an ordered_set which is protected by a single readers-writer lock. It is more surprising that they also scale better than set. This is due to hash tables using fine-grained locking at a fixed granularity, while CA trees can adapt the number of locks to the current contention level, and also to the parts of the key range where contention is occurring. 22 Improvements to Schedulers In the Erlang VM, a scheduler is responsible for executing multiple processes concurrently, in a timely and fair fashion, making optimal use of hardware resources. The VM implements preemptive multitasking with soft realtime guarantees. Erlang processes are normally scheduled on a reduction count basis where one reduction is roughly equivalent to a function call. Each process is allowed to execute until it either blocks waiting for input, typically a message from some other process, or until it has executed its quota of reductions. The Erlang VM is usually started with one scheduler per logical core (SMT-thread) available on the host machine, and schedulers are implemented as OS threads. When an Erlang process is spawned it is placed in the run queue of the scheduler of its parent, and it waits on that queue until the scheduler allocates it a slice of core time. Work stealing is used to balance load between cores, that is an idle scheduler may migrate a process from another run queue. Scheduler run queues are visualised in Figure iii.2. The default load management mechanism is load compaction that aims to keep as many scheduler threads as possible fully loaded with work, i.e. it attempts to ensure that scheduler threads do not run out of work. We have developed a new optional scheduler utilisation balancing mechanism that is available from Erlang/OTP 17.0. The new mechanism aims to balance scheduler utilisation between schedulers; that is, it will strive for equal scheduler utilisation on all schedulers. The scheduler utilisation balancing mechanism has no performance impact on the system when not enabled. On the other hand, when enabled, it results in changed timing in the system; normally there is a small overhead due to measuring of utilisation and calculating balancing information, which depends on the underlying primitives provided by the operating system. The new balancing mechanism results in a better distribution of processes to schedulers, P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 reducing the probability of core contention. Together with other VM improvements, such as interruptable BIFs and garbage collection, it results in lower latency and improved responsiveness, and hence reliability, for soft real-time applications. iii. Improvements to Time Management Soon after the start of the RELEASE project, time management in the Erlang VM became a scalability bottleneck for many applications, as illustrated by the parallel benchmark in Figure i. The issue came to prominence as other, more severe, bottlenecks were eliminated. This subsection motivates and outlines the improvements to time management that we made; these were incorporated into Erlang/OTP 18.x as a new API for time and time warping. The old API is still supported at the time of writing, but its use is deprecated. The original time API provides the erlang:now/0 built-in that returns “Erlang system time” or time since Epoch with micro second resolution. This time is the basis for all time internally in the Erlang VM. Many of the scalability problems of erlang:now/0 stem from its specification, written at a time when the Erlang VM was not multi-threaded, i.e. SMT-enabled. The documentation promises that values returned by it are strictly increasing and many applications ended up relying on this. For example applications often employ erlang:now/0 to generate unique integers. Erlang system time should align with the operating system’s view of time since Epoch or “OS system time”. However, while OS system time can be freely changed both forwards and backwards, Erlang system time cannot, without invalidating the strictly increasing value guarantee. The Erlang VM therefore contains a mechanism that slowly adjusts Erlang system time towards OS system time if they do not align. One problem with time adjustment is that the VM deliberately presents time with an in- accurate frequency; this is required to align Erlang system time with OS system time smoothly when these two have deviated, e.g. in the case of clock shifts when leap seconds are inserted or deleted. Another problem is that Erlang system time and OS system time can differ for very long periods of time. In the new API, we resolve this using a common OS technique [59], i.e. a monotonic time that has its zero point at some unspecified point in time. Monotonic time is not allowed to make leaps forwards and backwards while system time is allowed to do this. Erlang system time is thus just a dynamically varying offset from Erlang monotonic time. Time Retrieval Retrieval of Erlang system time was previously protected by a global mutex, which made the operation thread safe, but scaled poorly. Erlang system time and Erlang monotonic time need to run at the same frequency, otherwise the time offset between them would not be constant. In the common case, monotonic time delivered by the operating system is solely based on the machine’s local clock and cannot be changed, while the system time is adjusted using the Network Time Protocol (NTP). That is, they will run with different frequencies. Linux is an exception with a monotonic clock that is NTP adjusted and runs with the same frequency as system time [76]. To align the frequencies of Erlang monotonic time and Erlang system time, we adjust the frequency of the Erlang monotonic clock. This is done by comparing monotonic time and system time delivered by the OS, and calculating an adjustment. To achieve this scalably, one VM thread calculates the time adjustment to use at least once a minute. If the adjustment needs to be changed, new adjustment information is published and used to calculate Erlang monotonic time in the future. When a thread needs to retrieve time, it reads the monotonic time delivered by the OS and the time adjustment information previously published and calculates Erlang monotonic time. To preserve monotonicity it is important that all threads that read the same OS 23 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 180000 15B01 (now()) [70016,640] 17.4 (now()) [70016,640] 18.1 (now()) [70016,640] 18.1 (monotonic_time()) [70016,640] 160000 140000 time (ms) 120000 100000 80000 60000 40000 20000 0 0 10 20 30 40 50 60 70 60 70 # Schedulers 30 18.1 (monotonic_time()) [70016,640] 25 20 Speedup monotonic time map this to exactly the same Erlang monotonic time. This requires synchronisation on updates to the adjustment information using a readers-writer (RW) lock. This RW lock is write-locked only when the adjustment information is changed. This means that in the vast majority of cases the RW lock will be read-locked, which allows multiple readers to run concurrently. To prevent bouncing the lock cache-line we use a bespoke reader optimised RW lock implementation where reader threads notify about their presence in counters on separate cache-lines. The concept is similar to the reader indicator algorithm described by [48, Fig. 11] and alternatives include the ingress-egress counter used by [18] and the SNZI algorithm of [30]. 15 10 Timer Wheel and BIF Timer The timer wheel contains all timers set by Erlang processes. The original implementation was protected by a global mutex and scaled poorly. To increase concurrency, each scheduler thread has been assigned its own timer wheel that is used by processes executing on the scheduler. The implementation of timers in Erlang/OTP uses a built in function (BIF), as most low-level operations do. Until Erlang/OTP 17.4, this BIF was also protected by a global mutex. Besides inserting timers into the timer wheel, the BIF timer implementation also maps timer references to a timer in the timer wheel. To improve concurrency, from Erlang/OTP 18 we provide scheduler-specific BIF timer servers as Erlang processes. These keep information about timers in private ETS tables and only insert one timer at the time into the timer wheel. Benchmarks We have measured several benchmarks on a 16-core Bulldozer machine with eight dual CPU AMD Opteron 4376 HEs.5 We present three of them here. The first micro benchmark compares the execution time of an Erlang receive with that of a receive after that specifies a timeout and 5 See §2.5.4 of the RELEASE project Deliverable 2.4 (http://release-project.eu/documents/D2.4.pdf). 24 5 0 0 10 20 30 40 50 # Schedulers Figure 20: BenchErl parallel benchmark using erlang:monotonic_time/0 or erlang:now/0 in different Erlang/OTP releases: runtimes (left) and speedup obtained using erlang:monotonic_time/0 (right). provides a default value. The receive after sets a timer when the process blocks in the receive, and cancels it when a message arrives. In Erlang/OTP 17.4 the total execution time with standard timers is 62% longer than without timers. Using the improved implementation in Erlang/OTP 18.0, total execution time with the optimised timers is only 5% longer than without timers. The second micro benchmark repeatedly checks the system time, calling the built-in erlang:now/0 in Erlang/OTP 17.4, and calling both erlang:monotonic_time/0 and erlang:time_offset/0 and adding the results in Erlang/OTP 18.0. In this machine, where the VM uses 16 schedulers by default, the 18.0 release is more than 69 times faster than the 17.4 release. The third benchmark is the parallel BenchErl P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 benchmark from Section i. Figure 20 shows the results of executing the original version of this benchmark, which uses erlang:now/0 to create monotonically increasing unique values, using three Erlang/OTP releases: R15B01, 17.4, and 18.1. We also measure a version of the benchmark in Erlang/OTP 18.1 where the call to erlang:now/0 has been substituted with a call to erlang:monotonic_time/0. The graph on its left shows that: 1. the performance of time management has remained roughly unchanged between Erlang/OTP releases prior to 18.0; 2. the improved time management in Erlang/OTP 18.x make time management less likely to be a scalability bottleneck even when using erlang:now/0, and 3. the new time API (using erlang:monotonic_time/0 and friends) provides a scalable solution. The graph on the right side of Figure 20 shows the speedup that the modified version of the parallel benchmark achieves in Erlang/OTP 18.1. VII. Scalable Tools This section outlines five tools developed in the RELEASE project to support scalable Erlang systems. Some tools were developed from scratch, like Devo, SDMon and WombatOAM, while others extend existing tools, like Percept and Wrangler. These include tooling to transform programs to make them more scalable, to deploy them for scalability, to monitor and visualise them. Most of the tools are freely available under open source licences (Devo, Percept2, SD-Mon, Wrangler); while WombatOAM is a commercial product. The tools have been used for profiling and refactoring the ACO and Orbit benchmarks from Section III. The Erlang tool “ecosystem” consists of small stand-alone tools for tracing, profiling and debugging Erlang systems that can be used separately or together as appropriate for solving the problem at hand, rather than as a single, monolithic, super-tool. The tools presented here have been designed to be used as part of that ecosystem, and to complement already available functionality rather than to duplicate it. The Erlang runtime system has built-in support for tracing many types of events, and this infrastructure forms the basis of a number of tools for tracing and profiling. Typically the tools build on or specialise the services offered by the Erlang virtual machine, through a number of built-in functions. Most recently, and since the RELEASE project was planned, the Observer6 application gives a comprehensive overview of many of these data on a node-bynode basis. As actor frameworks and languages (see Section ii) have only recently become widely adopted commercially, their tool support remains relatively immature and generic in nature. That is, the tools support the language itself, rather than its distinctively concurrent aspects. Given the widespread use of Erlang, tools developed for it point the way for tools in other actor languages and frameworks. For example, just as many Erlang tools use tracing support provided by the Erlang VM, so can other actor frameworks, e.g. Akka can use the Kamon7 JVM monitoring system. Similarly, tools for other actor languages or frameworks could use data derived through OS-level tracing frameworks DTrace8 and SystemTap9 probes as we show in this section for Erlang, provided that the host language has tracing hooks into the appropriate infrastructure. i. Refactoring for Scalability Refactoring [65, 32, 75] is the process of changing how a program works without changing what it does. This can be done for readability, for testability, to prepare it for modification or extension, or — as is the case here — in order to improve its scalability. Because refactoring involves the transformation of source code, it is typically performed using machine support in a refactoring tool. There are a number of tools that support refactoring in Erlang: in the RELEASE project we have chosen to extend Wrangler10 [56]; other tools include Tidier [68] 6 http://www.erlang.org/doc/apps/observer/ 7 http://kamon.io 8 http://dtrace.org/blogs/about/ 9 https://sourceware.org/systemtap/wiki 10 http://www.cs.kent.ac.uk/projects/wrangler/ 25 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 and RefactorErl [44]. Supporting API Migration The SD Erlang libraries modify Erlang’s global_group library, becoming the new s_group library; as a result, Erlang programs using global_group will have to be refactored to use s_group. This kind of API migration problem is not uncommon, as software evolves and this often changes the API of a library. Rather than simply extend Wrangler with a refactoring to perform this particular operation, we instead added a framework for the automatic generation of API migration refactorings from a user-defined adaptor module. Our approach to automatic API migration works in this way: when an API function’s interface is changed, the author of this API function implements an adaptor function, defining calls to the old API in terms of the new. From this definition we automatically generate the refactoring that transforms the client code to use the new API; this refactoring can also be supplied by the API writer to clients on library upgrade, allowing users to upgrade their code automatically. The refactoring works by generating a set of rules that “fold in” the adaptation to the client code, so that the resulting code works directly with the new API. More details of the design choices underlying the work and the technicalities of the implementation can be found in a paper by [52]. Support for Introducing Parallelism We have introduced support for parallelising explicit list operations (map and foreach), for process introduction to complete a computationally intensive task in parallel, for introducing a worker process to deal with call handling in an Erlang “generic server” and to parallelise a tail recursive function. We discuss these in turn now; more details and practical examples of the refactorings appear in a conference paper describing that work [55]. Uses of map and foreach in list processing are among of the most obvious places where parallelism can be introduced. We have added a small library to Wrangler, called 26 para_lib, which provides parallel implementations of map and foreach. The transformation from an explicit use of sequential map/foreach to the use of their parallel counterparts is very straightforward, even manual refactoring would not be a problem. However a map/foreach operation could also be implemented differently using recursive functions, list comprehensions, etc.; identifying this kind of implicit map/foreach usage can be done using Wrangler’s code inspection facility, and a refactoring that turns an implicit map/foreach to an explicit map/foreach can also be specified using Wrangler’s rule-based transformation API. If the computations of two non-trivial tasks do not depend on each other, then they can be executed in parallel. The Introduce a New Process refactoring implemented in Wrangler can be used to spawn a new process to execute a task in parallel with its parent process. The result of the new process is sent back to the parent process, which will then consume it when needed. In order not to block other computations that do not depend on the result returned by the new process, the receive expression is placed immediately before the point where the result is needed. While some tail-recursive list processing functions can be refactored to an explicit map operation, many cannot due to data dependencies. For instance, an example might perform a recursion over a list while accumulating results in an accumulator variable. In such a situation it is possible to “float out” some of the computations into parallel computations. This can only be done when certain dependency constraints are satisfied, and these are done by program slicing, which is discussed below. Support for Program Slicing Program slicing is a general technique of program analysis for extracting the part of a program, also called the slice, that influences or is influenced by a given point of interest, i.e. the slicing criterion. Static program slicing is generally based on program dependency including both control dependency and data dependency. Backward P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 intra-function slicing is used by some of the refactorings described above; it is also useful in general, and made available to end-users under Wrangler’s Inspector menu [55]. Our work can be compared with that in PaRTE11 [17], a tool developed in another EU project that also re-uses the Wrangler front end. This work concentrates on skeleton introduction, as does some of our work, but we go further in using static analysis and slicing in transforming programs to make them suitable for introduction of parallel structures. ii. Scalable Deployment We have developed the WombatOAM tool12 to provide a deployment, operations and maintenance framework for large-scale Erlang distributed systems. These systems typically consist of a number of Erlang nodes executing on different hosts. These hosts may have different hardware or operating systems, be physical or virtual, or run different versions of Erlang/OTP. Prior to the development of WombatOAM, deployment of systems would use scripting in Erlang and the shell, and this is the state of the art for other actor frameworks; it would be possible to adapt the WombatOAM approach to these frameworks in a straightforward way. Architecture The architecture of WombatOAM is summarised in Figure 21. Originally the system had problems addressing full scalability because of the role played by the central Master node; in its current version an additional layer of Middle Managers was introduced to allow the system to scale easily to thousands of deployed nodes. As the diagram shows, the “northbound” interfaces to the web dashboard and the command-line are provided through RESTful connections to the Master. The operations of the Master are delegated to the Middle Managers that 11 http://paraphrase-enlarged.elte.hu/ downloads/D4-3_user_manual.pdf 12 WombatOAM (https://www.erlang-solutions. com/products/wombat-oam.html) is a commercial tool available from Erlang Solutions Ltd. engage directly with the managed nodes. Each managed node runs a collection of services that collect metrics, raise alarms and so forth; we describe those now. Services WombatOAM is designed to collect, store and display various kinds of information and event from running Erlang systems, and these data are accessed and managed through an AJAX-based Web Dashboard; these include the following. Metrics WombatOAM supports the collection of some hundred metrics — including, for instance, numbers of processes on a node and the message traffic in and out of the node — on a regular basis from the Erlang VMs running on each of the hosts. It can also collect metrics defined by users within other metrics collection frameworks, such as Folsom13 , and interface with other tools such as graphite14 which can log and display such information. The metrics can be displayed as histograms covering different windows, such as the last fifteen minutes, hour, day, week or month. Notifications As well as metrics, it can support event-based monitoring through the collection of notifications from running nodes. Notifications, which are one time events can be generated using the Erlang System Architecture Support Libraries, SASL, which is part of the standard distribution, or the lager logging framework15 , and will be displayed and logged as they occur. Alarms Alarms are more complex entities. Alarms have a state: they can be raised, and once dealt with they can be cleared; they also have identities, so that the same alarm may be raised on the same node multiple times, and each instance will need to be 13 Folsom collects and publishes metrics through an Erlang API: https://github.com/boundary/folsom 14 https://graphiteapp.org 15 https://github.com/basho/lager 27 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Wombat Web Dashboard Wombat CLI REST REST Northbound Interface Metrics API Notif. API Alarms API Orchestration API Node manager Metrics Notif. Alarms Orchestration Node manager Metrics Notif. Alarms Orchestration Topology API Erlang Master Core Core DB (including Topology and Orchestration data) Topology Erlang Middle Manager Core Erlang elibcloud/ Libcloud Erlang/RPC Provision VMs (REST) Provision software (SSH) Infrastructure Provider Agents Metrics agent Virtual machine error_logger agent lager agent Alarms agent Managed Erlang Node Figure 21: The architecture of WombatOAM. dealt with separately. Alarms are generated by SASL or lager, just as for notifications. Topology The Topology service handles adding, deleting and discovering nodes. It also monitors whether they are accessible, and if not, it notifies the other services, and periodically tries to reconnect. When the nodes are available again, it also notifies the other services. It doesn’t have a middle manager part, because it doesn’t talk to the nodes directly: instead it asks the Node manager service to do so. Node manager This service maintains the connection to all managed nodes via the 28 Erlang distribution protocol. If it loses the connection towards a node, it periodically tries to reconnect. It also maintains the states of the nodes in the database (e.g. if the connection towards a node is lost, the Node manager changes the node state to DOWN and raises an alarm). The Node manager doesn’t have a REST API, since the node states are provided via the Topology service’s REST API. Orchestration This service can deploy new Erlang nodes on already running machines. It can also provision new virtual machine instances using several cloud providers, and deploy Erlang nodes on those instances. For communicating with P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 the cloud providers, the Orchestration service uses an external library called Libcloud,16 for which Erlang Solutions has written an open source Erlang wrapper called elibcloud, to make Libcloud easier to use from WombatOAM. Note that WombatOAM Orchestration doesn’t provide a platform for writing Erlang applications: it provides infrastructure for deploying them. Deployment The mechanism consists of the following five steps. 1. Registering a provider. WombatOAM provides the same interface for different cloud providers which support the OpenStack standard or the Amazon EC2 API. WombatOAM also provides the same interface for using a fixed set of machines. In WombatOAM’s backend, this has been implemented as two driver modules: the elibcloud driver module which uses the elibcloud and Libcloud libraries to communicate with the cloud providers, and the SSH driver module that keeps track of a fixed set of machines. 2. Uploading a release. The release can be either a proper Erlang release archive or a set of Erlang modules. The only important aspect from WombatOAM’s point of view is that start and stop commands should be explicitly specified. This will enable WombatOAM start and stop nodes when needed. 3. Defining a node family. The next step is creating the node family, which is the entity that refers to a certain release, contains deployment domains that refer to certain providers, and contains other information necessary to deploy a node. 4. Defining a deployment domain. At this step a deployment domain is created that specifies(i) which providers should be used for 16 The unified cloud API [4]: apache.org https://libcloud. provisioning machines; (ii) the username that will be used when WombatOAM connects to the hosts using SSH. 5. Node deployment. To deploy nodes a WombatOAM user needs only to specify the number of nodes and the node family these nodes belong to. Nodes can be dynamically added to, or removed from, the system depending on the needs of the application. The nodes are started, and WombatOAM is ready to initiate and run the application. iii. Monitoring and Visualisation A key aim in designing new monitoring and visualisation tools and adapting existing ones was to provide support for systems running on parallel and distributed hardware. Specifically, in order to run modern Erlang systems, and in particular SD Erlang systems, it is necessary to understand both their single host (“multicore”) and multi-host (“distributed”) nature. That is, we need to be able to understand how systems run on the Erlang multicore virtual machine, where the scheduler associated with a core manages its own run queue, and processes migrate between the queues through a work stealing algorithm; at the same time, we have to understand the dynamics of a distributed Erlang program, where the user explicitly spawns processes to nodes.17 iii.1 Percept2 Percept218 builds on the existing Percept tool to provide post hoc offline analysis and visualisation of Erlang systems. Percept2 is designed to allow users to visualise and tune the parallel performance of Erlang systems on a single node on a single manycore host. It visualises Erlang application level concurrency and identifies concurrency bottlenecks. Percept2 uses 17 This is in contrast with the multicore VM, where programmers have no control over where processes are spawned; however, they still need to gain insight into the behaviour of their programs to tune performance. 18 https://github.com/RefactoringTools/percept2 29 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Erlang built in tracing and profiling to monitor process states, i.e. waiting, running, runnable, free and exiting. A waiting or suspended process is considered an inactive and a running or runnable process is considered active. As a program runs with Percept, events are collected and stored to a file. The file is then analysed, with the results stored in a RAM database, and this data is viewed through a web-based interface. The process of offline profiling for a distributed Erlang application using Percept2 is shown in Figure 22. Percept generates an application-level zoomable concurrency graph, showing the number of active processes at each point during profiling; dips in the graph represent low concurrency. A lifetime bar for each process is also included, showing the points during its lifetime when the process was active, as well as other per-process information. Percept2 extends Percept in a number of ways — as detailed by [53] — including most importantly: • Distinguishing between running and runnable time for each process: this is apparent in the process runnability comparison as shown in Figure 23, where orange represents runnable and green represents running. This shows very clearly where potential concurrency is not being exploited. • Showing scheduler activity: the number of active schedulers at any time during the profiling. • Recording more information during execution, including the migration history of a process between run queues; statistics about message passing between processes: the number of messages and the average message size sent/received by a process; the accumulated runtime per-process: the accumulated time when a process is in a running state. • Presenting the process tree: the hierarchy structure indicating the parent-child relationships between processes. 30 • Recording dynamic function call graph/count/time: the hierarchy structure showing the calling relationships between functions during the program run, and the amount of time spent in a function. • Tracing of s_group activities in a distributed system. Unlike global group, s_group allows dynamic changes to the s_group structure of a distributed Erlang system. In order to support SD Erlang, we have also extended Percept2 to allow the profiling of s_group related activities, so that the dynamic changes to the s_group structure of a distributed Erlang system can be captured. We have also improved on Percept as follows. • Enabling finer-grained control of what is profiled. The profiling of port activities, schedulers activities, message passing, process migration, garbage collection and s_group activities can be enabled/disabled, while the profiling of process runnability (indicated by the “proc” flag) is always enabled. • Selective function profiling of processes. In Percept2, we have built a version of fprof, which does not measure a function’s own execution time, but measures everything else that fprof measures. Eliminating measurement of a function’s own execution time gives users the freedom of not profiling all the function calls invoked during the program execution. For example, they can choose to profile only functions defined in their own applications’ code, and not those in libraries. • Improved dynamic function callgraph. With the dynamic function callgraph, a user is able to understand the causes of certain events, such as heavy calls of a particular function, by examining the region around the node for the function, including the path to the root of the graph. Each edge in the callgraph is annotated with P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Figure 22: Offline profiling of a distributed Erlang application using Percept2. Figure 23: Percept2: showing processes running and runnable (with 4 schedulers/run queues). the number of times the target function is called by the source function as well as further information. Finally, we have also improved the scalability of Percept in three ways. First we have parallelised the processing of trace files so that multiple data files can be processed at the same time. We have also compressed the representation of call graphs, and cached the history of generated web pages. Together these make the system more responsive and more scalable. iii.2 DTrace/SystemTap tracing DTrace provides dynamic tracing support for various flavours of Unix, including BSD and Mac OS X, and SystemTap does the same for Linux; both allow the monitoring of live, running systems with minimal overhead. They can be used by administrators, language developers and application developers alike to examine the behaviour of applications, language implementations and the operating system during development or even on live production systems. In comparison to other similar tools and instrumentation frameworks, they are relatively lightweight, do not require special recompiled versions of the software to be examined, nor special post-processing tools to create meaningful information from the data gathered. Using these probs, it is possible to identify bottlenecks both in the VM itself and 31 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Run Queues 4500 RunQueue15 RunQueue14 RunQueue13 RunQueue12 RunQueue11 RunQueue10 RunQueue9 RunQueue8 RunQueue7 RunQueue6 RunQueue5 RunQueue4 RunQueue3 RunQueue2 RunQueue1 RunQueue0 4000 3500 3000 2500 2000 1500 1000 500 0 Figure 24: DTrace: runqueue visualisation of bang benchmark with 16 schedulers (time on Xaxis). in applications. VM bottlenecks are identified using a large number of probes inserted into Erlang/OTP’s VM, for example to explore scheduler runqueue lengths. These probes can be used to measure the number of processes per scheduler, the number of processes moved during work stealing, the number of attempts to gain a run-queue lock, how many of these succeed immediately, etc. Figure 24 visualises the results of such monitoring; it shows how the size of run queues vary during execution of the bang BenchErl benchmark on a VM with 16 schedulers. Application bottlenecks are identified by an alternative back-end for Percept2, based on DTrace/SystemTap instead of the Erlang builtin tracing mechanism. The implementation reuses the existing Percept2 infrastructure as far as possible. It uses a different mechanism for collecting information about Erlang programs, a different format for the trace files, but the same storage infrastructure and presentation facilities. iii.3 Devo Devo19 [10] is designed to provide real-time online visualisation of both the low-level (single node, multiple cores) and high-level (multi19 https://github.com/RefactoringTools/devo 32 ple nodes, grouped into s_groups) aspects of Erlang and SD Erlang systems. Visualisation is within a browser, with web sockets providing the connections between the JavaScript visualisations and the running Erlang systems, instrumented through the trace tool builder (ttb). Figure 25 shows visualisations from devo in both modes. On the left-hand side a single compute node is shown. This consists of two physical chips (the upper and lower halves of the diagram) with six cores each; with hyperthreading this gives twelve virtual cores, and hence 24 run queues in total. The size of these run queues is shown by both the colour and height of each column, and process migrations are illustrated by (fading) arc between the queues within the circle. A green arc shows migration on the same physical core, a grey one on the same chip, and blue shows migrations between the two chips. On the right-hand side of Figure 25 is a visualisation of SD-Orbit in action. Each node in the graph represents an Erlang node, and colours (red, green, blue and orange) are used to represent the s_groups to which a node belongs. As is evident, the three nodes in the central triangle belong to multiple groups, and act as routing nodes between the other nodes. The colour of the arc joining two nodes represents the current intensity of communication between the nodes (green quiescent; red busiest). iii.4 SD-Mon SD-Mon is a tool specifically designed for monitoring SD-Erlang systems. This purpose is accomplished by means of a shadow network of agents, that collect data from a running system. An example deployment is shown in Figure 26, where blue dots represent nodes in the target system and the other nodes make up the SD-Mon infrastructure. The network is deployed on the basis of a configuration file describing the network architecture in terms of hosts, Erlang nodes, global group and s_group partitions. Tracing to be performed on moni- P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 p0 9152 0 p0 10890 0 p0 10891 0 p0 9501 0 Figure 25: Devo: low and high-level visualisations of SD-Orbit. tored nodes is also specified within the configuration file. An agent is started by a master SD-Mon node for each s_group and for each free node. Configured tracing is applied on every monitored node, and traces are stored in binary format in the agent file system. The shadow network follows system changes so that agents are started and stopped at runtime as required, as shown in Figure 27. Such changes are persistently stored so that the last configuration can be reproduced after a restart. Of course, the shadow network can be always updated via the User Interface. Each agent takes care of an s_group or of a free node. At start-up it tries to get in contact with its nodes and apply the tracing to them as stated by the master. Binary files are stored in the host file system. Tracing is internally used in order to track s_group operations happening at runtime. An asynchronous message is sent to the master whenever one of these changes occurs. Since each process can only be traced by a single process at a time, each node (included those belonging to more than one s_group) is controlled by only one agent. When a node is removed from a group or when the group is deleted, another agent takes over, as shown in Figure 27. When an agent is stopped, all traces on the controlled nodes are switched off. The monitoring network is also supervised, in order to take account of network fragility, and when an agent node goes down another node is deployed to play its role; there are also periodic consistency checks for the system as a whole, and when an inconsistency is detected then that part of the system can be restarted. SD-Mon does more than monitor activities one node at a time. In particular inter-node and inter-group messages are displayed at runtime. As soon as an agent is stopped, the related tracing files are fetched across the network by the master and they are made available in a readable format in the master file system. SD-Mon provides facilities for online visualisation of this data, as well as post hoc offline analysis. Figure 28 shows, in real time, messages that are sent between s_groups. This data 33 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Figure 26: SD-Mon architecture. can also be used as input to the animated Devo visualisation, as illustrated in the right-hand side of Figure 25. VIII. Figure 27: SD-Mon evolution. Before (top) and after eliminating s_group 2 (bottom). Systemic Evaluation Preceding sections have investigated the improvements of individual aspects of an Erlang system, e.g. ETS tables in Section i. This section analyses the impact of the new tools and technologies from Sections V, VI, VII in concert. We do so by investigating the deployment, reliability, and scalability of the Orbit and ACO benchmarks from Section III. The experiments reported here are representative. Similar experiments show consistent results for a range of micro-benchmarks, several benchmarks, and the very substantial (approximately 150K lines of Erlang) Sim-Diasca case study [16] on several state of the art NUMA architectures, and the four clusters specified in Appendix A. A coherent presentation of many of these results is available in an article by [22] and in a RELEASE project deliverable20 . The bulk of the experiments reported here are conducted on the Athos cluster using Erlang/OTP 17.4 and the 20 See Deliverable 6.2, available online. http://www. release-project.eu/documents/D6.2.pdf 34 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Figure 28: SD-Mon: online monitoring. associated SD Erlang libraries. The experiments cover two measures of scalability. As Orbit does a fixed size computation, the scaling measure is relative speedup (or strong scaling), i.e. speedup relative to execution time on a single core. As the work in ACO increases with compute resources, weak scaling is the appropriate measure. The benchmarks also evaluate different aspects of s_groups: Orbit evaluates the scalability impacts of network connections, while ACO evaluates the impact of both network connections and the global namespace required for reliability. D-Orbit performs better on a small number of nodes as communication is direct, rather than via a gateway node. As the number of nodes grows, however, SD-Orbit delivers better speedups, i.e. beyond 80 nodes in case of 2M orbit elements, and beyond 100 nodes in case of 5M orbit elements. When we increase the size of Orbit beyond 5M, the D-Orbit version fails due to the fact that some VMs exceed the available RAM of 64GB. In contrast SD-Orbit experiments run successfully even for an orbit with 60M elements. ii. i. Orbit Figure 29 shows the speedup of the D-Orbit and SD-Orbit benchmarks from Section i.1. The measurements are repeated seven times, and we plot standard deviation. Results show that Ant Colony Optimisation (ACO) Deployment The deployment and monitoring of ACO, and of the large (150K lines of Erlang) Sim-Diasca simulation using WombatOAM (Section ii) on the Athos cluster is detailed in [22]. 35 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 450 Relative Speedup 400 300 200 100 0 0 50(1200) 100(2400) 150(3600) Number of nodes (cores) D-Orbit W=2M D-Orbit W=5M 200(4800) 257(6168) SD-Orbit W=2M SD-Orbit W=5M Figure 29: Speedup of distributed Erlang vs. SD Erlang Orbit (2M and 5M elements) [Strong Scaling]. An example experiment deploys 10,000 Erlang nodes without enabling monitoring, and hence allocates three nodes per core (i.e. 72 nodes on each of 139 24-core Athos hosts). Figure 30 shows that WombatOAM deploys the nodes in 212s, which is approximately 47 nodes per second. It is more common to have at least one core per Erlang node, and in a related experiment 5000 nodes are deployed, one per core (i.e. 24 nodes on each of 209 24-core Athos hosts) in 101s, or 50 nodes per second. Crucially in both cases the deployment time is linear in the number of nodes. The deployment time could be reduced to be logarithmic in the number of nodes using standard divideand-conquer techniques. However there hasn’t been the demand to do so as most Erlang systems are long running servers. The measurement data shows two important facts: it shows that WombatOAM scales well (up to a deployment base of 10,000 Erlang nodes), and that WombatOAM is non-intrusive because its overhead on a monitored node is typically less than 1.5% of effort on the node. We conclude that WombatOAM is capable of deploying and monitoring substantial distributed Erlang and SD Erlang programs. The experiments in the remainder of this section 36 use standard distributed Erlang configuration file deployment. Reliability SD Erlang changes the organisation of processes and their recovery data at the language level, so we seek to show that these changes have not disrupted Erlang’s world-class reliability mechanisms at this level. As we haven’t changed them we don’t exercise Erlang’s other reliability mechanisms, e.g. those for managing node failures, network congestion, etc. A more detailed study of SD Erlang reliability, including the use of replicated databases for recovering Instant Messenger chat sessions, finds similar results [23]. We evaluate the reliability of two ACO versions using a Chaos Monkey service that kills processes in the running system at random [14]. Recall that GR-ACO provides reliability by registering the names of critical processes globally, and SR-ACO registers them only within an s_group (Section ii). For both GR-ACO and SR-ACO a Chaos Monkey runs on every Erlang node, i.e. master, submasters, and colony nodes, killing a random Erlang process every second. Both ACO versions run to completion. Recovery, at this failure frequency, has no measurable P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Time (sec) 200 160 120 80 0 2000 4000 6000 Number of nodes (N) Experimental data 8000 10000 Best fit: 0.0125N+83 Figure 30: Wombat Deployment Time Execution time (s) 16 12 8 4 0 0 50(1200) GR-ACO 100(2400) 150(3600) Number of nodes (cores) SR-ACO 200(4800) 250(6000) ML-ACO Figure 31: ACO execution times, Erlang/OTP 17.4 (RELEASE) [Weak Scaling]. 37 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 impact on runtime. This is because processes are recovered within the Virtual machine using (globally synchronised) local recovery information. For example, on a common X86/Ubuntu platform typical Erlang process recovery times are around 0.3ms, so around 1000x less than the Unix process recovery time on the same platform [58]. We have conducted more detailed experiments on an Instant Messenger benchmark, and obtained similar results [23]. We conclude that both GR-ACO and SRACO are reliable, and that SD Erlang preserves distributed Erlang reliability model. The remainder of the section outlines the impact of maintaining the recovery information required for reliability on scalability. Scalability Figure 31 compares the runtimes of the ML, GR, and SR versions of ACO (Section ii) on Erlang/OTP 17.4(RELEASE). As outlined in Section i.1, GR-ACO not only maintains a fully connected graph of nodes, it registers process names for reliability, and hence scales significantly worse than the unreliable ML-ACO. We conclude that providing reliability with standard distributed Erlang process registration dramatically limits scalability. While ML-ACO does not provide reliability, and hence doesn’t register process names, it maintains a fully connected graph of nodes which limits scalability. SR-ACO, that maintains connections and registers process names only within s_groups scales best of all. Figure 31 illustrates how maintaining the process namespace, and fully connected network, impacts performance. This reinforces the evidence from the Orbit benchmarks, and others, that partitioning the network of Erlang nodes significantly improves performance at large scale. To investigate the impact of SD Erlang on network traffic, we measure the number of sent and received packets on the GPG cluster for three versions of ACO: ML-ACO, GR-ACO, and SR-ACO. Figure 32 shows the total number of sent packets. The highest traffic (the red line) belongs to the GR-ACO and the lowest traffic belongs to the SR-ACO (dark blue line). This 38 shows that SD Erlang significantly reduces the network traffic between Erlang nodes. Even with the s_group name registration SR-ACO has less network traffic than ML-ACO that has no global name registration. iii. Evaluation Summary We have shown that WombatOAM is capable of deploying and monitoring substantial distributed Erlang and SD Erlang programs like ACO and Sim-Diasca. The Chaos Monkey experiments with GR-ACO and SR-ACO show that both are reliable, and hence that SD Erlang preserves the distributed Erlang language-level reliability model. As SD Orbit scales better than D-Orbit, SRACO scales better than ML-ACO, and SRACO has significantly less network traffic, we conclude that, even when global recovery data is not maintained, partitioning the fullyconnected network into s_groups reduces network traffic and improves performance. While the distributed Orbit instances (W=2M) and (W=5M) reach scalability limits at around 40 and 60 nodes, Orbit scales to 150 nodes on SD Erlang (limited by input size), and SR-ACO is still scaling well on 256 nodes (6144 cores). Hence not only have we exceeded the 60 node scaling limits of distributed Erlang identified in Section ii, we have not reached the scaling limits of SD Erlang on this architecture. Comparing GR-ACO and ML-ACO scalability curves shows that maintaining global recovery data, i.e. a process name space, dramatically limits scalability. Comparing GRACO and SR-ACO scalability curves shows that scalability can be recovered by partitioning the nodes into appropriately-sized s_groups, and hence maintaining the recovery data only within a relatively small group of nodes. These results are consistent with other experiments. IX. Discussion Distributed actor platforms like Erlang, or Scala with Akka, are a common choice for internet-scale system architects as the model, P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Figure 32: Number of sent packets in ML-ACO, GR-ACO, and SR-ACO. with its automatic, and VM-supported reliability mechanisms makes it extremely easy to engineer scalable reliable systems. Targeting emergent server architectures with hundreds of hosts and tens of thousands of cores, we report a systematic effort to improve the scalability of a leading distributed actor language, while preserving reliability. The work is a vade mecum for addressing scalability of reliable actor languages and frameworks. It is also high impact, with downloads of our improved Erlang/OTP running at 50K a month. We have undertaken the first systematic study of scalability in a distributed actor language, covering VM, language and persistent storage levels. We have developed the BenchErl and DE-Bench tools for this purpose. Key VM-level scalability issues we identify include contention for shared ETS tables and for commonly-used shared resources like timers. Key language scalability issues are the costs of maintaining a fully-connected network, maintaining global recovery information, and explicit process placement. Unsurprisingly the scaling issues for this distributed actor language are common to other distributed or parallel languages and frameworks with other paradigms like CHARM++ [45], Cilk [15], or Legion [40]. We establish scientifically the folklore limitations of around 60 connected nodes for distributed Erlang (Section IV). The actor model is no panacea, and there can still be scalability problems in the algorithms that we write, either within a single actor or in the way that we structure communicating actors. A range of pragmatic issues also impact the performance and scalability of actor systems, including memory occupied by processes (even when quiescent), mailboxes filling up, etc. Identifying and resolving these problems is where tools like Percept2 and WombatOAM are needed. However, many of the scalability issues arise where Erlang departs from the private state principle of the actor model, e.g. in maintaining shared state in ETS tables, or a shared global process namespace for recovery. We have designed and implemented a set of Scalable Distributed (SD) Erlang libraries to address language-level scalability issues. The key constructs are s_groups for partitioning the network and global process namespace, and semi-explicit process placement for deploying distributed Erlang applications on large heterogeneous architectures in a portable way. We have provided a state transition operational semantics for the new s_groups, and validated 39 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 the library implementation against the semantics using QuickCheck (Section V). To improve the scalability of the Erlang VM and libraries we have improved the implementation of shared ETS tables, time management and load balancing between schedulers. Following a systematic analysis of ETS tables, the number of fine-grained (bucket) locks and of reader groups have been increased. We have developed and evaluated four new techniques for improving ETS scalability:(i) programmer control of number of bucket locks; (ii) a contention-adapting tree data structure for ordered_sets; (iii) queue delegation locking; and (iv) eliminating the locks in the meta table. We have introduced a new scheduler utilisation balancing mechanism to spread work to multiple schedulers (and hence cores), and new synchronisation mechanisms to reduce contention on the widely-used time management mechanisms. By June 2015, with Erlang/OTP 18.0, the majority of these changes had been included in the primary releases. In any scalable actor language implementation such thoughtful design and engineering will be required to schedule large numbers of actors on hosts with many cores, and to minimise contention on shared VM resources (Section VI). To facilitate the development of large Erlang systems, and to make them understandable we have developed a range of tools. The proprietary WombatOAM tool deploys and monitors large distributed Erlang systems over multiple, and possibly heterogeneous, clusters or clouds. We have made open source releases of four concurrency tools: Percept2 now detects concurrency bad smells; Wrangler provides enhanced concurrency refactoring; the Devo tool is enhanced to provide interactive visualisation of SD Erlang systems; and the new SD-Mon tool monitors SD Erlang systems. We anticipate that these tools will guide the design of tools for other large scale distributed actor languages and frameworks (Section VII). We report on the reliability and scalability implications of our new technologies using a range of benchmarks, and consistently use the Orbit and ACO benchmarks throughout the 40 article. While we report measurements on a range of NUMA and cluster architectures, the key scalability experiments are conducted on the Athos cluster with 256 hosts (6144 cores). Even when global recovery data is not maintained, partitioning the network into s_groups reduces network traffic and improves the performance of the Orbit and ACO benchmarks above 80 hosts. Crucially we exceed the 60 node limit for distributed Erlang and do not reach the scalability limits of SD Erlang with 256 nodes/VMs and 6144 cores. Chaos Monkey experiments show that two versions of ACO are reliable, and hence that SD Erlang preserves the Erlang reliability model. However the ACO results show that maintaining global recovery data, i.e. a global process name space, dramatically limits scalability in distributed Erlang. Scalability can, however, be recovered by maintaining recovery data only within appropriately sized s_groups. These results are consistent with experiments with other benchmarks and on other architectures (Section VIII). In future work we plan to incorporate RELEASE technologies, along with other technologies in a generic framework for building performant large scale servers. In addition, preliminary investigations suggest that some SD Erlang ideas could improve the scalability of other actor languages.21 For example the Akka framework for Scala could benefit from semiexplicit placement, and Cloud Haskell from partitioning the network. Appendix A: Architecture Specifications The specifications of the clusters used for measurement are summarised in Table 2. We also use the following NUMA machines. (1) An AMD Bulldozer with 16M L2/16M L3 cache, 128GB RAM, four AMD Opteron 6276s at 2.3 GHz, 16 “Bulldozer” cores each, giving a total of 64 cores. (2) An Intel NUMA with 128GB RAM, four Intel Xeon E5-4650s at 2.70GHz, 21 See Deliverable 6.7 (http://www.release-project. eu/documents/D6.7.pdf), available online. P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 Table 2: Cluster Specifications (RAM per host in GB). Cores per Name max Hosts host total avail. 20 16 320 320 Kalkyl 348 8 2,784 1,408 TinTin 160 16 2,560 2,240 Athos 776 24 18,624 6,144 GPG Processor RAM Intel Xeon E5-2640v2 8C, 2GHz Intel Xeon 2.26GHz 5520v2 4C, 64 24–72 Interconnection 10GB Ethernet InfiniBand Gb/s 20 AMD Opteron 6220v2 Bull2:1 oversub64–128 scribed dozer 8C, 3.0GHz QDR Infiniband Intel Xeon E5-2697v2 12C, 2.7GHz 64 Infiniband FDR14 each with eight hyperthreaded cores, giving a total of 64 cores. [3] Bulldozer (microarchitecture), 2015. https://en.wikipedia.org/wiki/ Bulldozer_(microarchitecture). Acknowledgements [4] Apache SF. Liblcoud, 2016. https:// libcloud.apache.org/. We would like to thank the entire RELEASE project team for technical insights and administrative support. Roberto Aloi and Enrique Fernandez Casado contributed to the development and measurement of WombatOAM. This work has been supported by the European Union grant RII3-CT-2005-026133 “SCIEnce: Symbolic Computing Infrastructure in Europe”, IST-2011287510 “RELEASE: A High-Level Paradigm for Reliable Large-scale Server Software”, and by the UK’s Engineering and Physical Sciences Research Council grant EP/G055181/1 “HPCGAP: High Performance Computational Algebra and Discrete Mathematics”. [5] C. R. Aragon and R. G. Seidel. Randomized search trees. In Proceedings of the 30th Annual Symposium on Foundations of Computer Science, pages 540–545, October 1989. [6] Joe Armstrong. Programming Erlang: Software for a Concurrent World. Pragmatic Bookshelf, 2007. [7] Joe Armstrong. Erlang. Commun. ACM, 53:68–75, 2010. [1] Gul Agha. ACTORS: A Model of Concurrent Computation in Distributed Systems. PhD thesis, MIT, 1985. [8] Stavros Aronis, Nikolaos Papaspyrou, Katerina Roukounaki, Konstantinos Sagonas, Yiannis Tsiouris, and Ioannis E. Venetis. A scalability benchmark suite for Erlang/OTP. In Torben Hoffman and John Hughes, editors, Proceedings of the Eleventh ACM SIGPLAN Workshop on Erlang, pages 33–42, New York, NY, USA, September 2012. ACM. [2] Gul Agha. An overview of Actor languages. SIGPLAN Not., 21(10):58–67, 1986. [9] Thomas Arts, John Hughes, Joakim Johansson, and Ulf Wiger. Testing telecoms References 41 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 software with Quviq QuickCheck. In Proceedings of the 2006 ACM SIGPLAN workshop on Erlang, pages 2–10, New York, NY, USA, 2006. ACM. [10] Robert Baker, Peter Rodgers, Simon Thompson, and Huiqing Li. Multi-level Visualization of Concurrent and Distributed Computation in Erlang. In Visual Languages and Computing (VLC): The 19th International Conference on Distributed Multimedia Systems (DMS 2013), 2013. [11] Luiz André Barroso, Jimmy Clidaras, and Urs Hölzle. The Datacenter as a Computer. Morgan and Claypool, 2nd edition, 2013. Symposium, TFP 2014. Revised Selected Papers, volume 8843 of LNCS, pages 104–121. Springer, 2015. [18] Irina Calciu, Dave Dice, Yossi Lev, Victor Luchangco, Virendra J. Marathe, and Nir Shavit. NUMA-aware reader-writer locks. In Proceedings of the 18th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, pages 157–166, New York, NY, USA, 2013. ACM. [19] Francesco Cesarini and Simon Thompson. Erlang Programming: A Concurrent Approach to Software Development. O’Reilly Media, 1st edition, 2009. [12] Basho Technologies. Riakdocs. Basho Bench, 2014. http://docs.basho. com/riak/latest/ops/building/ benchmarking/. [20] Rohit Chandra, Leonardo Dagum, Dave Kohr, Dror Maydan, Jeff McDonald, and Ramesh Menon. Parallel Programming in OpenMP. Morgan Kaufmann, San Francisco, CA, USA, 2001. [13] J. E. Beasley. OR-Library: Distributing test problems by electronic mail. The Journal of the Operational Research Society, 41(11):1069–1072, 1990. Datasets available at http://people.brunel.ac. uk/~mastjjb/jeb/orlib/wtinfo.html. [21] Natalia Chechina, Huiqing Li, Amir Ghaffari, Simon Thompson, and Phil Trinder. Improving the network scalability of Erlang. J. Parallel Distrib. Comput., 90(C):22–34, 2016. [14] Cory Bennett and Ariel Tseitlin. Chaos monkey released into the wild. Netflix Blog, 2012. [15] Robert D Blumofe, Christopher F Joerg, Bradley C Kuszmaul, Charles E Leiserson, Keith H Randall, and Yuli Zhou. Cilk: An efficient multithreaded runtime system. In Proceedings of the fifth ACM SIGPLAN symposium on Principles and practice of parallel programming, pages 207–216. ACM, 1995. [16] Olivier Boudeville. Technical manual of the sim-diasca simulation engine. EDF R&D, 2012. [17] István Bozó, Viktória Fördős, Dániel Horpácsi, Zoltán Horváth, Tamás Kozsik, Judit Kőszegi, and Melinda Tóth. Refactorings to enable parallelization. In Jurriaan Hage and Jay McCarthy, editors, Trends in Functional Programming: 15th International 42 [22] Natalia Chechina, Kenneth MacKenzie, Simon Thompson, Phil Trinder, Olivier Boudeville, Viktória Fördős, Csaba Hoch, Amir Ghaffari, and Mario Moro Hernandez. Evaluating scalable distributed Erlang for scalability and reliability. IEEE Transactions on Parallel and Distributed Systems, 2017. [23] Natalia Chechina, Mario Moro Hernandez, and Phil Trinder. A scalable reliable instant messenger using the SD Erlang libraries. In Erlang’16, pages 33–41, New York, NY, USA, 2016. ACM. [24] Koen Claessen and John Hughes. QuickCheck: a lightweight tool for random testing of Haskell programs. In Proceedings of the 5th ACM SIGPLAN International Conference on Functional Programming, pages 268–279, New York, NY, USA, 2000. ACM. P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 [25] H. A. J. Crauwels, C. N. Potts, and L. N. van Wassenhove. Local search heuristics for the single machine total weighted tardiness scheduling problem. INFORMS Journal on Computing, 10(3):341–350, 1998. The datasets from this paper are included in Beasley’s ORLIB. [26] Jeffrey Dean and Sanjay Ghemawat. Mapreduce: Simplified data processing on large clusters. Commun. ACM, 51(1):107– 113, 2008. [27] David Dewolfs, Jan Broeckhove, Vaidy Sunderam, and Graham E. Fagg. FTMPI, fault-tolerant metacomputing and generic name services: A case study. In EuroPVM/MPI’06, pages 133–140, Berlin, Heidelberg, 2006. Springer-Verlag. [28] Alan AA Donovan and Brian W Kernighan. The Go programming language. Addison-Wesley Professional, 2015. [29] Marco Dorigo and Thomas Stützle. Ant Colony Optimization. Bradford Company, Scituate, MA, USA, 2004. [30] Faith Ellen, Yossi Lev, Victor Luchangco, and Mark Moir. SNZI: scalable NonZero indicators. In Proceedings of the 26th Annual ACM Symposium on Principles of Distributed Computing, pages 13–22, New York, NY, USA, 2007. ACM. [31] Jeff Epstein, Andrew P. Black, and Simon Peyton-Jones. Towards Haskell in the Cloud. In Haskell ’11, pages 118–129. ACM, 2011. [32] Martin Fowler. Refactoring: Improving the Design of Existing Code. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1999. [33] Ana Gainaru and Franck Cappello. Errors and faults. In Fault-Tolerance Techniques for High-Performance Computing, pages 89–144. Springer International Publishing, 2015. [34] Martin Josef Geiger. New instances for the Single Machine Total Weighted Tardiness Problem. Technical Report Research Report 10-03-01, March 2010. Datasets available at http://logistik. hsu-hh.de/SMTWTP. [35] Guillaume Germain. Concurrency oriented programming in termite scheme. In Proceedings of the 2006 ACM SIGPLAN workshop on Erlang, pages 20–20, New York, NY, USA, 2006. ACM. [36] Amir Ghaffari. De-bench, a benchmark tool for distributed erlang, 2014. https: //github.com/amirghaffari/DEbench. [37] Amir Ghaffari. Investigating the scalability limits of distributed Erlang. In Proceedings of the 13th ACM SIGPLAN Workshop on Erlang, pages 43–49. ACM, 2014. [38] Amir Ghaffari, Natalia Chechina, Phip Trinder, and Jon Meredith. Scalable persistent storage for Erlang: Theory and practice. In Proceedings of the 12th ACM SIGPLAN Workshop on Erlang, pages 73–74, New York, NY, USA, 2013. ACM. [39] Seth Gilbert and Nancy Lynch. Brewer’s conjecture and the feasibility of consistent, available, partition-tolerant web services. ACM SIGACT News, 33:51–59, 2002. [40] Andrew S Grimshaw, Wm A Wulf, and the Legion team. The Legion vision of a worldwide virtual computer. Communications of the ACM, 40(1):39–45, 1997. [41] Carl Hewitt. Actor model for discretionary, adaptive concurrency. CoRR, abs/1008.1459, 2010. [42] Carl Hewitt, Peter Bishop, and Richard Steiger. A universal modular ACTOR formalism for artificial intelligence. In IJCAI’73, pages 235–245, San Francisco, CA, USA, 1973. Morgan Kaufmann. [43] Rich Hickey. The Clojure programming language. In DLS’08, pages 1:1–1:1, New York, NY, USA, 2008. ACM. 43 P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 [44] Zoltán Horváth, László Lövei, Tamás Kozsik, Róbert Kitlei, Anikó Nagyné Víg, Tamás Nagy, Melinda Tóth, and Roland Király. Building a refactoring tool for Erlang. In Workshop on Advanced Software Development Tools and Techniques, WASDETT 2008, 2008. [45] Laxmikant V Kale and Sanjeev Krishnan. Charm++: a portable concurrent object oriented system based on c++. In ACM Sigplan Notices, volume 28, pages 91–108. ACM, 1993. [46] David Klaftenegger, Konstantinos Sagonas, and Kjell Winblad. On the scalability of the Erlang term storage. In Proceedings of the Twelfth ACM SIGPLAN Workshop on Erlang, pages 15–26, New York, NY, USA, September 2013. ACM. [47] David Klaftenegger, Konstantinos Sagonas, and Kjell Winblad. Delegation locking libraries for improved performance of multithreaded programs. In Euro-Par 2014 Parallel Processing., volume 8632 of LNCS, pages 572–583. Springer, 2014. [48] David Klaftenegger, Konstantinos Sagonas, and Kjell Winblad. Queue delegation locking. IEEE Transactions on Parallel and Distributed Systems, 2017. To appear. [49] Rusty Klophaus. Riak core: Building distributed applications without shared state. In ACM SIGPLAN Commercial Users of Functional Programming, CUFP ’10, pages 14:1–14:1, New York, NY, USA, 2010. ACM. [50] Avinash Lakshman and Prashant Malik. Cassandra: A decentralized structured storage system. SIGOPS Oper. Syst. Rev., 44(2):35–40, April 2010. [51] J. Lee et al. Python actor runtime library, 2010. osl.cs.uiui.edu/parley/. [52] Huiqing Li and Simon Thompson. Automated API Migration in a User-Extensible Refactoring Tool for Erlang Programs. In 44 Tim Menzies and Motoshi Saeki, editors, Automated Software Engineering, ASE’12. IEEE Computer Society, 2012. [53] Huiqing Li and Simon Thompson. Multicore profiling for Erlang programs using Percept2. In Proceedings of the Twelfth ACM SIGPLAN workshop on Erlang, 2013. [54] Huiqing Li and Simon Thompson. Improved Semantics and Implementation Through Property-Based Testing with QuickCheck. In 9th International Workshop on Automation of Software Test, 2014. [55] Huiqing Li and Simon Thompson. Safe Concurrency Introduction through Slicing. In PEPM ’15. ACM SIGPLAN, January 2015. [56] Huiqing Li, Simon Thompson, György Orosz, and Melinda Töth. Refactoring with Wrangler, updated. In ACM SIGPLAN Erlang Workshop, volume 2008, 2008. [57] Frank Lubeck and Max Neunhoffer. Enumerating large Orbits and direct condensation. Experimental Mathematics, 10(2):197– 205, 2001. [58] Andreea Lutac, Natalia Chechina, Gerardo Aragon-Camarasa, and Phil Trinder. Towards reliable and scalable robot communication. In Proceedings of the 15th International Workshop on Erlang, Erlang 2016, pages 12–23, New York, NY, USA, 2016. ACM. [59] LWN.net. The high-resolution timer API, January 2006. https://lwn.net/ Articles/167897/. [60] Kenneth MacKenzie, Natalia Chechina, and Phil Trinder. Performance portability through semi-explicit placement in distributed Erlang. In Proceedings of the 14th ACM SIGPLAN Workshop on Erlang, pages 27–38. ACM, 2015. [61] Jeff Matocha and Tracy Camp. A taxonomy of distributed termination detection P. Trinder, N. Chechina, N. Papaspyrou, K. Sagonas, S. Thompson, et al. • Scaling Reliably • 2017 algorithms. Journal of Systems and Software, 43(221):207–221, 1998. [62] Nicholas D Matsakis and Felix S Klock II. The Rust language. In ACM SIGAda Ada Letters, volume 34, pages 103–104. ACM, 2014. [63] Robert McNaughton. Scheduling with deadlines and loss functions. Management Science, 6(1):1–12, 1959. [64] Martin Odersky et al. The Scala programming language, 2012. www.scala-lang.org. [65] William F. Opdyke. Refactoring ObjectOriented Frameworks. PhD thesis, University of Illinois at Urbana-Champaign, 1992. [66] Nikolaos Papaspyrou and Konstantinos Sagonas. On preserving term sharing in the Erlang virtual machine. In Torben Hoffman and John Hughes, editors, Proceedings of the 11th ACM SIGPLAN Erlang Workshop, pages 11–20. ACM, 2012. [67] RELEASE Project Team. EU Framework 7 Project 287510 (2011–2015), 2015. http: //www.release-project.eu. [68] Konstantinos Sagonas and Thanassis Avgerinos. Automatic refactoring of Erlang programs. In Principles and Practice of Declarative Programming (PPDP’09), pages 13–24. ACM, 2009. range updates using contention adapting search trees. In Languages and Compilers for Parallel Computing - 28th International Workshop, volume 9519 of LNCS, pages 37–53. Springer, 2016. [72] Marc Snir, Steve W. Otto, D. W. Walker, Jack Dongarra, and Steven HussLederman. MPI: The Complete Reference. MIT Press, Cambridge, MA, USA, 1995. [73] Sriram Srinivasan and Alan Mycroft. Kilim: Isolation-typed actors for Java. In ECOOP’08, pages 104–128, Berlin, Heidelberg, 2008. Springer-Verlag. [74] Don Syme, Adam Granicz, and Antonio Cisternino. Expert F# 4.0. Springer, 2015. [75] Simon Thompson and Huiqing Li. Refactoring tools for functional languages. Journal of Functional Programming, 23(3), 2013. [76] Marcus Völker. Linux timers, May 2014. [77] WhatsApp, 2015. https://www.whatsapp. com/. [78] Tom White. Hadoop - The Definitive Guide: Storage and Analysis at Internet Scale (3. ed., revised and updated). O’Reilly, 2012. [79] Ulf Wiger. Industrial-Strength Functional programming: Experiences with the Ericsson AXD301 Project. In Implementing Functional Languages (IFL’00), Aachen, Germany, September 2000. Springer-Verlag. [69] Konstantinos Sagonas and Kjell Winblad. More scalable ordered set for ETS using adaptation. In Proc. ACM SIGPLAN Workshop on Erlang, pages 3–11. ACM, September 2014. [70] Konstantinos Sagonas and Kjell Winblad. Contention adapting search trees. In Proceedings of the 14th International Symposium on Parallel and Distributed Computing, pages 215–224. IEEE Computing Society, 2015. [71] Konstantinos Sagonas and Kjell Winblad. Efficient support for range queries and 45	6cs.PL
Top Tips to Make Your Research Irreproducible Neil P. Chue Hong1 , Tom Crick2 , Ian P. Gent3 , Lars Kotthoff4 and Kenji Takeda5 1 arXiv:1504.00062v2 [cs.CE] 8 Apr 2015 2 Software Sustainability Institute, University of Edinburgh, UK Department of Computing & Information Systems, Cardiff Metropolitan University, UK 3 School of Computer Science, University of St Andrews, UK 4 Insight Centre for Data Analytics, University College Cork, Ireland 5 Microsoft Research, Cambridge, UK 1 n.chuehong@software.ac.uk http://www2.epcc.ed.ac.uk/~ neilc/ 2 tcrick@cardiffmet.ac.uk http://drtomcrick.com/ 3 ian.gent@st-andrews.ac.uk http://ian.gent 4 lars.kotthoff@insight-centre.org http://4c.ucc.ie/~ larsko/ 5 kenji.takeda@microsoft.com http://research.microsoft.com/en-us/people/kenjitak/ 1 April 2015 We have noticed (and contributed to) a number of manifestos, guides and top tips on how to make research reproducible [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]; however, we have seen very little published on how to make research irreproducible. Irreproducibility is the default setting for all of science, and irreproducible research is particularly common across the computational sciences. The study of making your work irreproducible without reviewers complaining is a much neglected area; we feel therefore that by encapsulating some of our top tips1 on irreproducibility, we will be filling a much-needed gap in the domain literature. By following our starter tips, you can ensure that if your work is wrong, nobody will be able to check it; if it is correct, you will make everyone else do disproportionately more work than you to build upon it. In either case you are the beneficiary. It is an unfortunate convention of science that research should pretend to be reproducible; our top tips will help you salve the conscience of certain reviewers still bound by this fussy conventionality, enabling them to enthusiastically recommend acceptance of your irreproducible work. 1. Think “Big Picture”. People are interested in the science, not the dull experimental setup, so don’t describe it. If necessary, camouflage this absence with brief, high-level details of insignificant aspects of your methodology. 2. Be abstract. Pseudo-code is a great way of communicating ideas quickly and clearly while giving readers no chance to understand the subtle implementation details (particularly the custom toolchains and manual interventions) that actually make it work. 3. Short and sweet. Any limitations of your methods or proofs will be obvious to the careful reader, so there is no need to waste space on making them explicit2 . However much work it takes colleagues to fill in the gaps, you will still get the credit if you just say you have amazing experiments or proofs (with a hat-tip to Pierre de Fermat: “Cuius rei demonstrationem mirabilem sane detexi hanc marginis exiguitas non caperet.”). 4. The deficit model. You’re the expert in the domain, only you can define what algorithms and data to run experiments with. In the unhappy circumstance that your methods do not do well on 1 N.B. We are by no means claiming this is an exhaustive list for making your research irreproducible... Space saved in this way can be used to cite the critical papers in the field, i.e. those papers that will inflate your own (as well as potential reviewers’) h-index. 2 1 community curated benchmarks, you should create your own bespoke benchmarks and use those (and preferably not make them available to others). 5. Don’t share. Doing so only makes it easier for other people to scoop your research ideas, understand how your code actually works3 instead of why you say it does, or worst of all to understand that your code doesn’t actually work at all. However, our most important tip is deceptively but beautifully simple: to ensure your work is irreproducible, make sure that you cannot reproduce it yourself . If you were able to reproduce it, there would always be the danger of somebody else being able to do exactly the same as you. Much else follows from this; for example, complete confidence in your own inability to reproduce work saves tedious time revising your work on advice from reviewers: if you are unable to browbeat the editor into accepting it as is, you can always resubmit elsewhere. A major advantage of this key insight is that no strict discipline is required to ensure self-irreproducibility: in our experience, irreproducibility can happily occur after only the tiniest amount of carelessness at one of any number of stages. We make a simple conjecture: an experiment that is irreproducible is exactly equivalent to an experiment that was never carried out at all. The happy consequences of this conjecture for experts in irreproducibility will be published elsewhere, with extremely impressive experimental support. We close with a mantra for scientists interested in irreproducibility: After Publishing Research, Irreproducibility Lets False Observations Obtain Longevity! References [1] Andreas Prlić and James B. Procter. Ten Simple Rules for the Open Development of Scientific Software. PLoS Computational Biology, 12(8), 2012. [2] Geir Kjetil Sandve, Anton Nekrutenko, James Taylor, and Eivind Hovig. Ten Simple Rules for Reproducible Computational Research. PLoS Computational Biology, 9(10), 2013. [3] Ian P. Gent. The Recomputation Manifesto. Available from: http://arxiv.org/abs/1304.3674, April 2013. [4] Lucas N. Joppa, Greg McInerny, Richard Harper, Lara Salido, Kenji Takeda, Kenton O’Hara, David Gavaghan, and Stephen Emmott. Troubling Trends in Scientific Software Use. Science, 340(6134):814–815, 2013. [5] Ian P. Gent and Lars Kotthoff. Recomputation.org: Experience of its first year and lessons learned. In Recomputability 2014, December 2014. [6] Greg Wilson, D. A. Aruliah, C. Titus Brown, Neil P. Chue Hong, Matt Davis, Richard T. Guy, Steven H. D. Haddock, Kathryn D. Huff, Ian M. Mitchell, Mark D. Plumbley, Ben Waugh, Ethan P. White, and Paul Wilson. Best Practices for Scientific Computing. PLoS Biology, 12(1), 2014. [7] Carole Goble. Better Software, Better Research. IEEE Internet Computing, 18(5):4–8, 2014. [8] Tom Crick, Benjamin A. Hall, and Samin Ishtiaq. “Can I Implement Your Algorithm?”: A Model for Reproducible Research Software. In Proceedings of 2nd International Workshop on Sustainable Software for Science: Practice and Experiences (WSSSPE2), 2014. [9] Tom Crick, Benjamin A. Hall, Samin Ishtiaq, and Kenji Takeda. “Share and Enjoy”: Publishing Useful (and Usable) Scientific Models. In Proceedings of 1st International Workshop on Recomputability, 2014. [10] Victoria Stodden and Sheila Miguez. Best Practices for Computational Science: Software Infrastructure and Environments for Reproducible and Extensible Research. Journal of Open Research Software, 2(1):1–6, 2014. 3 An exemplary example: http://www.phdcomics.com/comics.php?f=1689 2	5cs.CE
RESTRICTION ESTIMATES FOR THE FREE TWO STEP NILPOTENT GROUP ON THREE GENERATORS arXiv:1704.07876v1 [math.GR] 25 Apr 2017 VALENTINA CASARINO AND PAOLO CIATTI Abstract. Let G be the free two step nilpotent Lie group on three generators and let L be a subLaplacian on it. We compute the spectral resolution of L and prove that the operators arising from this decomposition enjoy a Tomas-Stein type estimate. 1. Introduction Starting from the observation of E. Stein that, when the Lebesgue exponent p is sufficiently close to 1, the Fourier transform of an Lp function restricts in a sense, that may be made precise, to a compact hypersurface of nonvanishing curvature, various forms of restriction theorems for the Fourier transform became one of the main theme of analysis. The most satisfactory result obtained so far in this theory is the Stein-Tomas theorem, which concerns the restriction in the L2 - sense of the Fourier transform of a function n`1 in Lp pRn q, with 1 ď p ď 2 n`3 , to the sphere S n´1 . This result was proved in the mid seventies of the last century and, as any very important theorem, has now a great deal of extensions and applications in different branches of mathematics. In particular, it was observed by Stein himself and by R. Strichartz in [Str] that the Stein-Tomas estimates can be interpreted as bounds concerning the mapping properties between Lebesgue spaces of the operators arising in the spectral decomposition of the Euclidean Laplacian. This point of view is emphasised in the works of C. Sogge, who studied the boundedness of the spectral projections of the Laplace-Beltrami operator on the spheres in [So1] and more generally on compact Riemannian manifolds in [So2]. Few years later D. Müller proved for the first time a result of this sort for a subelliptic operator, the subLaplacian on the Heisenberg group. In the series of papers [C1], [C2], [CCi2], [CCi4] the authors of this article studied a sort of combination of the previous results considering estimates for the joint spectral projections of a Laplace-Beltrami operator and a subLaplacian. Some years ago we started in [CCi2] and [CCi3] an investigation devoted to extend the result of Müller to subLaplacians on more general two step nilpotent Lie groups. In those papers we considered groups enjoying a special nondegeneracy condition, according to which the quotient with respect to a codimension one subspace of the center is isomorphic to a Heisenberg group. In this paper we are instead concerned with the mapping properties of the operators arising in the spectral decomposition of an invariant subLaplacian on the free two step 2000 Mathematics Subject Classification. 22E25; 22E30, 43A80, 47B40. Key words and phrases. Free nilpotent Lie groups. Sub-Laplacians. Restriction theorems. Research supported by the Italian Ministero dell’Istruzione, dell’Università e della Ricerca through PRIN “Real and Complex Manifolds: Geometry, Topology and Harmonic Analysis” and the GNAMPA project 2016 “Calcolo funzionale per operatori subellittici su varietà”. 1 2 VALENTINA CASARINO AND PAOLO CIATTI nilpotent Lie group on three generators. In this group the quotient with respect to a codimension one subspace of the center is isomorphic to the direct product of the three dimensional Heisenberg group and the real line. Therefore, the nondegeneracy property we exploited in the previous works here is absent. However, the analysis on this group is made easier by the fact that the abelian component in the quotients with respect to the planes in the center is always one dimensional and by the fact that the combined action of the rotation group on the two layers of the Lie algebra gives rise to a family of automorphisms. We conclude the introduction with a short description of the next sections. In Section 2 we describe some features of the group G we are concerned with. Decomposing its Lie algebra g into the direct sum z ‘ v, where z is the center, we see that once a nontrivial linear form µ on z has been fixed, the quotient of g with respect to the null space of µ is the direct product of a three dimensional Heisenberg algebra times the real line. Moreover, these two components, the Heisenberg and the abelian one, depend only on the line in z˚ spanned by µ. Then in Section 3 we derive the spectral decomposition of an L2 function f with respect to the subLaplacian. This decomposition is expressed first of all in terms of the Fourier transform f µ of f on z, which for every µ in the dual of z is a function living on v. This transform is followed by the section of the Fourier transform of f µ in the direction of the radical of µ, which is finally decomposed into eigenfunctions of the two dimensional twisted Laplacian. Finally, in Section 4 we prove that the operators arising from the spectral decomposition of the subLaplacian are bounded for 1 ď s ď 6{5 and 1 ď p ď 2 from the nonisotropic Lebesgue space Lsz Lpv , in which the integrations in z and v are weighted with exponents s 1 and p, to Lsz L2v . The basic ingredients in these estimates are provided by the Stein-Tomas theorem, which in fact dictates the range of the exponents concerning the integration on z, and by the estimates proved some years ago by H. Koch and F. Ricci in [KR] for the twisted Laplacian. The range of exponents for which our estimates hold is sharp as examples analogue to those provided in [CCi2] show. 2. Generalities Let G be a free two step nilpotent Lie group on three generators. We assume that G is connected and simply connected. The Lie algebra, g, of G splits as a vector space into the direct sum g “ v ‘ z, where z is the centre. Both v and z are three dimensional vector spaces. To proceed it is convenient to introduce on g an inner product x¨, ¨y with respect to which v and z are orthogonal subspaces. The inner product induces a norm on g and a norm on g˚ , the space of linear forms on g, which we will both denote by \| ¨ \|. We shall always identify G with its Lie algebra g by means of the exponential mapping and use coordinates px, zq with x “ px1 , x2 , x3 q P R3 labeling the points of v and z “ pz1 , z2 , z3 q P R3 the points of z. The vector fields X1 “ B x 1 ` x3 x2 B z2 ´ B z3 , 2 2 X2 “ B x 2 ` x1 x3 B z3 ´ B z1 , 2 2 X3 “ B x 3 ` x2 x1 B z1 ´ B z2 2 2 RESTRICTION ESTIMATES FOR THE FREE GROUP 3 are left invariant and satisfy rX1 , X2 s “ Bz3 “ Z3 , rX2 , X3 s “ Bz1 “ Z1 , rX3 , X1 s “ Bz2 “ Z2 . (2.1) Fix a point ω in the unit sphere S “ tν P z˚ : \|ν\| “ 1u; here z˚ denotes the space of linear forms on z. We call rω the radical of the skew-symmetric form v ˆ v Q pX, Y q ÞÑ ωprX, Y sq, that is the space rω “ tX P v : ωprX, Y sq “ 0 for all Y P vu, which as follows from (2.1) is one dimensional. If Zω P z satisfies ωpZω q “ 1, then \|Zω \| “ 1 and rX, Y s “ ωprX, Y sqZω , X, Y P v. Thus if we fix two unit vectors Xω and Yω in vω satisfying ωprXω , Yω sq “ 1, then the subspace hω spanned by tXω , Yω , Zω u is a Lie algebra isomorphic to the three dimensional Heisenberg algebra. The subspace kω of z orthogonal to Zω coincides with the kernel of ω. Therefore the quotient of g with respect to kω is isomorphic as a Lie algebra to the direct product of hω and rω . To equip this algebra with a system of coordinates, we decompose a vector into the sum vω Vω ` xω Xω ` yω Yω ` zω Zω , where Vω is a unit vector in rω and pvω , xω , yω , zω q P R4 . We shall identify, with a slight abuse of notation, ω with the linear form in v˚ canonically associated to Vω , writing vω “ ωpXq. Fixing two orthonormal vectors tWω,1 , Wω,2 u in kω , we obtain an orthogonal basis tZω , Wω,1 , Wω,2 u of z. Then any vector in g can be uniquely written as vω Vω ` xω Xω ` yω Yω ` zω Zω ` wω,1 Wω,1 ` wω,2 Wω,2 , where pvω , xω , yω , zω , wω,1 , wω,2q P R6 are coordinates adapted to ω. The transformation mapping X “ px1 , x2 , x3 q to pxω , yω , vω q is a rotation Rω P SOp3q fixing Vω and hence pxω , yω q “ xω Xω ` yω Yω “ Rω pXq ´ ωpXqVω . Since by the universal property rotations of v extend to automorphisms of g, Rω preserves Zω and maps pz1 , z2 , z3 q to pwω,1 , wω,2, zω q. 3. The subLaplacian The vector fields X1 , X2 , X3 generate by (2.1) the entire Lie algebra g. Therefore by Hörmander’s theorem the subLaplacian L “ ´X12 ´ X22 ´ X32 4 VALENTINA CASARINO AND PAOLO CIATTI is a hypoelliptic differential operator. It is easy to see that 1 1 1 L “ ´Bx21 ´ Bx22 ´ Bx23 ` px2 Bx3 ´ x3 Bx1 qBz1 ` px3 Bx1 ´ x2 Bx3 qBz2 ` px1 Bx2 ´ x2 Bx1 qBz3 2 2 2 1 2 1 2 1 2 2 2 2 2 2 2 ´ px2 ` x3 qBz1 ´ px1 ` x3 qBz2 ´ px1 ` x2 qBz3 4 4 4 and hence that L is homogeneous with respect to the dilations defined by δǫ pX, Zq “ pǫx, ǫ2 zq, ǫ ą 0, that is Lpf ˝ δǫ q “ ǫ2 pLf q ˝ δǫ . The subLaplacian is a symmetric operator on the Schwartz class and extends to a selfadjoint operator on L2 pGq with positive spectrum. Our plan in the section consists in deriving its spectral decomposition. Our construction is largely inspired by the analogous derivations in [ACDS] and [MMü]. To accomplish the task, given a Schwartz function f on G, we take its partial Fourier transform in the central variables ż µ f pXq “ e´iµpZq f pX, ZqdZ, µ P z˚ . z˚ We then introduce spherical coordinates in z˚ , writing µ “ ρω, where ρ “ \|µ\| and ω belongs to S. In these coordinates the Fourier inversion formula in z˚ reads ż ż8 f pX, Zq “ eiρωpZq f ρω pXqρ2 dρdω. (3.1) S 0 In particular, we have Xa f pX, Zq “ ż ż8 S eiρωpZq Xaρω f ρω pXqρ2 dρdω, a “ 1, 2, 3, (3.2) 0 where the vector fields Xaρω are the differential operators on v defined by ` ˘ Xa e´iρωpZq gpXq “ e´iρωpZq Xaρω gpXq, a “ 1, 2, 3, for any Schwartz function g. They are more explicitly given by i i i X1ρω “ Bx1 ` ρpω3 x2 ´ω2 x3 q, X2ρω “ Bx2 ` ρpω1 x3 ´ω3 x1 q, X3ρω “ Bx3 ` ρpω2 x1 ´ω1 x2 q, 2 2 2 ρω where ωa “ ω pZa q, a “ 1, 2, 3. Defininig the differential operator L on v by ` ˘ L e´iρωpZq gpXq “ e´iρωpZq Lρω gpXq, (3.3) we find i i i Lρω “ ´Bx21 ´ Bx22 ´ Bx23 ` ρω1 px3 Bx1 ´ x2 Bx3 q ` ρω2 px2 Bx3 ´ x3 Bx1 q ` ρω3 px2 Bx1 ´ x1 Bx2 q 2 2 2 1 1 1 2 2 2 ` ρ ω1 px2 ` x23 q ` ρ2 ω22px21 ` x23 q ` ρ2 ω32ρ2 px21 ` x22 q. 4 4 4 In the coordinates pvω , xω , yω , zω , wω,1 , wω,2q the operator Lρω becomes i 1 2 2 2 2 2 2 Lρω xω ,yω ,vω “ ´Bvω ´ Bxω ´ Byω ` ρpyω Bxω ´ xω Byω q ` ρ pxω ` yω q 2 4 “ ´Bv2ω ` ∆ρxω ,yω , (3.4) RESTRICTION ESTIMATES FOR THE FREE GROUP 5 where ∆λs,t ˙ ˆ ˘ B2 λ2 ` 2 B2 i B B “´ 2 ´ 2 ` λ t ´s ` s ` t2 Bs Bt 2 Bs Bt 4 is the two dimensional λ-twisted Laplacian. The spectrum of ∆λs,t consists of eigenvalues, which are given by λp2k ` 1q with k “ 0, 1, 2, ¨ ¨ ¨ . We write the spectral decomposition of g with respect to ∆λs,t as gps, tq “ 8 ÿ Λλk gps, tq, (3.5) k“0 where Λλk maps L2 pR2 q onto the eigenspace associated to λp2k ` 1q. These are integral operators (for more details about them see, for instance, [Th, p.19 ff]) satisfying the estimates 1 1 1 }Λλk f }L2 pR2 q ď Cλ p ´ 2 p2k ` 1qγp p q }f }Lp pR2 q , 1 ď p ď 2 , (3.6) proved by H. Koch and F. Ricci in [KR]; here γ is the piecewise affine function on r 21 , 1s defined by # ˆ ˙ 1 ´1 if 1 ď p ď 56 , 1 γ (3.7) :“ p1 1 1 p p ´ p q if 56 ď p ď 2. 2 2 In order to decompose f ρω into eigenfunctions of the twisted Laplacian, we define the function gρ,ω “ f ρω ˝ Rω´1 , where Rω´1 is the inverse of Rω . Since pxω , yω , vω q “ Rω px1 , x2 , x3 q “ Rω X, recalling that vω “ ωpXq and pxω , yω q “ Rω pXq ´ ωpXqVω , we have f ρω pXq “ gρ,ω pxω , yω , vω q “ gρ,ω pRω pXq ´ ωpXqVω , ωpXqq. Then it makes sense to decompose the function gρ,ω pxω , yω , vω q in eigenfuctions of ∆ρxω ,yω , obtaining f ρω pXq “ gρ,ω pxω , yω , vω q 8 ÿ “ Λρk gρ,ω pxω , yω , vω q “ k“0 8 ÿ k“0 ` ˘ Λρk pf ρω ˝ Rω´1 q Rω pXq ´ ωpXqVω , ωpXq . Taking the partial Fourier transform of gρ,ω in rω , it follows that ż8 ` ˘ Fω gρ,ω pxω , yω ; ξq “ e´isξ gρ,ω Rω pXq ´ ωpXqVω ` sVω ds “ ´8 8 ÿż8 k“0 ´8 and also Fω gρ,ω pxω , yω ; ξq “ ` ˘ e´isξ Λρk gρ,ω Rω pXq ´ ωpXqVω ` sVω ds, 8 ÿ k“0 ` ˘ Λρk Fω gρ,ω Rω pXq ´ ωpXqVω ; ξ , 6 VALENTINA CASARINO AND PAOLO CIATTI since the operators Λρk and Fω , acting on different spaces, namely vω and rω , commute. Thus, inverting the Fourier transform we obtain 8 ż8 ÿ gρ,ω pxω , yω , vω q “ eivω ξ Λρk Fω gρ,ω pxω , yω ; ξqdξ. k“0 ´8 Hence, we have 8 ż8 ÿ ρω f pXq “ k“0 ´8 ` ˘ eiωpXqξ Λρk Fω pf ρω ˝ Rω´1 q Rω pXq ´ ωpXqVω ; ξ dξ, (3.8) which plugged in (3.1) yields ż8 8 ż ż8 ÿ ` ˘ iρωpZq e eiωpXqξ Λρk Fω pf ρω ˝ Rω´1 q Rω pXq ´ ωpXqVω ; ξ dξρ2 dρdω. f pX, Zq “ k“0 S ´8 0 (3.9) We may now deduce from (3.9) the spectral decomposition of f with respect to L by replacing the arguments of the sum and the integrals with generalised eigenfunctions. To do that we notice that by (3.2) we have ż ż8 Lf pX, Zq “ eiρωpZq Lρω f ρω pXqρ2 dρdω, S 0 which by (3.8) and (3.4) implies ¸ ˜ ż ż ż8 8 8 ÿ ` ˘ eiωpXqξ Λρk Fω pf ρω ˝ Rω´1 q Rω pXq ´ ωpXqVω ; ξ dξ ρ2 dρdω eiρωpZq Lρω S “ k“0 ´8 0 8 ÿ ż ż8 k“0 S eiρωpZq 0 ˆ ˙ˇ ż8 ` ˘ ˇ 2 ρ ivω ξ ρ ρω ´1 ˆ p´Bvω ` ∆xω ,yω q e Λk Fω pf ˝ Rω q xω , yω ; ξ dξ ˇ “ 8 ż ż8 ÿ k“0 S ˆ ˆż 8 ´8 0 ivω ξ e 8 ż ż8 ÿ k“0 S ˆ ˆż 8 2 pξ ` ivω ξ e k“0 S ˆ ´8 ˝ Rω´1 q ` ˙ˇ ˘ ˇ xω , yω ; ξ dξ ˇ Rω pXq ρ2 dρdω 0 8 ż ż8 ÿ ˆż 8 ∆ρxω ,yω qΛρk Fω pf ρω eiρωpZq 2 pξ ` ρp2k ` ´8 “ ρ2 dρdω eiρωpZq ´8 “ Rω pXq 1qqΛρk Fω pf ρω ˝ Rω´1 q eiρωpZq ` ˙ˇ ˘ ˇ xω , yω ; ξ dξ ˇ Rω pXq ρ2 dρdω 0 iωpXqξ e 2 pξ ` ρp2k ` 1qqΛρk Fω pf ρω ˝ Rω´1 q ` ˙ ˘ Rω pXq ´ ωpXqVω ; ξ dξ ρ2 dρdω, RESTRICTION ESTIMATES FOR THE FREE GROUP 7 ˇ where ˇRω pXq means that the integral must be computed in xω Xω `yω Yω “ Rω pXq´ωpXqVω and vω “ ωpXq. Replacing in the last integral ρp2k ` 1q with ρ we obtain ż8 ż8 ż ÿ 8 ´3 p2k ` 1q eiρωpZq{p2k`1q eiωpXqξ pξ 2 ` ρq Lf pX, Zq “ ˆ ´8 0 S k“0 ` ρ{p2k`1q Λk Fω pf ρω{p2k`1q ˝ Rω´1 q Rω pXq ˘ ´ ωpXqVω ; ξ ρ2 dρdξdω, from which, setting µ “ ξ 2 ` ρ and using Fubini’s theorem, we deduce ż8 ż8 ż ÿ 8 2 ´3 p2k ` 1q eipµ´ξ qωpZq{p2k`1q eiωpXqξ pµ ´ ξ 2 q2 Lf pX, Zq “ ´8 S k“0 ξ2 ` ˘ 2 pµ´ξ 2 q{p2k`1q ˆ µΛk Fω pf pµ´ξ qω{p2k`1q ˝ Rω´1 q Rω pXq ´ ωpXqVω ; ξ dµdξdω ż8 ż ÿ ż ?µ 8 2 “ µ p2k ` 1q´3 ? eipµ´ξ qωpZq{p2k`1q eiωpXqξ pµ ´ ξ 2 q2 ˆ ´ µ S k“0 0 2 pµ´ξ 2 q{p2k`1q Λk Fω pf pµ´ξ qω{p2k`1q ` ˘ ˝ Rω´1 q Rω pXq ´ ωpXqVω ; ξ dξdωdµ. The expression for Lf just derived shows that if we write ż8 f pX, Zq “ Pµ f pX, Zqdµ, (3.10) 0 where Pµ f pX, Zq “ ż ÿ 8 p2k ` 1q ´3 ? eipµ´ξ 2 qωpZq{p2k`1q eiωpXqξ pµ ´ ξ 2 q2 (3.11) ´ µ S k“0 pµ´ξ 2 q{p2k`1q ˆ Λk ż ?µ Fω pf pµ´ξ 2 qω{p2k`1q then ` ˘ ˝ Rω´1 q Rω pXq ´ ωpXqVω ; ξ dξdω, LPµ f “ µPµ f. Thus, (3.10) provides the spectral resolution of f with respect to L. 4. The restriction theorem In this section we show that the operators Pµ satisfy some restriction estimates. To state and prove this result we introduce nonisotropic norms on G defined, for 1 ď p, s ă 8, by ˜ż ˆż ¸ p1 ˙p s }f }Lsz Lpv “ \|f pX, Zq\|s dZ dV , v z with the obvious modifications when s or p equal 8. We shall prove the following theorem. Theorem 4.1. Let f be a Schwartz function on G and let 1 ď p ď 2 and 1 ď s ď 65 , then }Pµ f }Lsz 1 L2v ď Cµ3p s ´ s1 q`3p p ´ 2 q }f }Lsz Lpv . 1 These estimates are false for s ą 65 . 1 1 1 (4.1) 8 VALENTINA CASARINO AND PAOLO CIATTI Proof. The sharpness of the range of s where the estimates hold may be proved with a suitable and easy modification of the example provided in [CCi2]. So we prove only the bound (4.1). The dependence on µ in the right hand side of (4.1) is dictated by the homogeneity of L. Therefore, it suffices to discuss the case µ “ 1. To reduce the complexity of the notation and make the formulas more readable we consider a tensor function f pX, Zq “ αpZqβpXq, with α and β Schwartz functions on z and v. Then (3.11) reduces to ˆ ˙ ż ż1 ÿ 8 p1 ´ ξ 2 q2 i p1´ξ2 qωpZq iωpXqξ p1 ´ ξ 2 qω P1 f pX, Zq “ α̂ e 2k`1 e 3 2k ` 1 S ´1 k“0 p2k ` 1q 1´ξ2 2k`1 ˆ Λk ` ˘ Fω pβ ˝ Rω´1 q Rω pXq ´ ωpXqVω ; ξ dξdω, where α̂ denotes the Fourier transform of α. Consider another tensor function gpX, Zq “ γpZqδpXq, with }γ}Lsz “ 1 and }δ}Lqv “ 1, and compute ż ż xP1 f, gy “ P1 f pX, ZqgpX, ZqdZdX. v z Then we have ˆ ˙ ż ż ˜ż ż 1 ÿ 8 p1 ´ ξ 2 q2 i p1´ξ2 qωpZq iωpXqξ p1 ´ ξ 2 qω xP1 f, gy “ α̂ e 2k`1 e 3 2k ` 1 v z S ´1 k“0 p2k ` 1q ¸ 1´ξ2 ` ˘ ˆ Λk2k`1 Fω pβ ˝ Rω´1 q Rω pXq ´ ωpXqVω ; ξ dξdω γpZqδpXqdZdX ˙ ˆ ˙ ˆ ż ż1 ÿ 8 p1 ´ ξ 2 q2 p1 ´ ξ 2 qω p1 ´ ξ 2 qω “ α̂ γ̂ 3 2k ` 1 2k ` 1 S ´1 k“0 p2k ` 1q ˜ż ¸ 1´ξ2 ` ˘ ˆ eiωpXqξ Λk2k`1 Fω pβ ˝ Rω´1 q Rω pXq ´ ωpXqVω ; ξ δpXqdX dξdω. v Changing the variables in the integral over v we obtain (since \| det Rω \| “ 1) ˙ ˆ ˙ ˆ ż ż1 ÿ 8 p1 ´ ξ 2 qω p1 ´ ξ 2 q2 p1 ´ ξ 2 qω γ̂ xP1 f, gy “ α̂ 3 2k ` 1 2k ` 1 S ´1 k“0 p2k ` 1q ¸ ˜ż 1´ξ2 ` ˘ ˆ eivω ξ Λk2k`1 Fω pβ ˝ Rω´1 q xω , yω ; ξ pδ ˝ Rω´1 qpxω , yω , vω qdxω dyω dvω dξdω v ˙ ˆ ˙ ˆ ż ż1 ÿ 8 p1 ´ ξ 2 q2 p1 ´ ξ 2 qω p1 ´ ξ 2 qω “ α̂ γ̂ 3 2k ` 1 2k ` 1 S ´1 k“0 p2k ` 1q ˜ż ¸ 1´ξ2 ` ˘ Λk2k`1 Fω pβ ˝ Rω´1 q xω , yω ; ξ Fω pδ ˝ Rω´1 qpxω , yω ; ξqdxω dyω dξdω. ˆ vω RESTRICTION ESTIMATES FOR THE FREE GROUP 9 Setting Ψpω; ξ, kq “ ż vω we write 1´ξ2 ` ˘ Λk2k`1 Fω pβ ˝ Rω´1 q xω , yω ; ξ Fω pδ ˝ Rω´1 qpxω , yω ; ξqdxω dyω , (4.2) ¸ ˙ ˆ ˙ ż ˜ż ˆ 8 ÿ p1 ´ ξ 2 q2 1 p1 ´ ξ 2 qω p1 ´ ξ 2 qω Ψpω; ξ, kqdω dξ. α̂ xP1 f, gy “ γ̂ p2k ` 1q3 ´1 S 2k ` 1 2k ` 1 k“0 Then the Cauchy-Schwarz inequality yields ż1 8 ÿ 1 \|xP1 f, gy\| ď p1 ´ ξ 2 q2 Φpξ, kq 3 p2k ` 1q ´1 k“0 ˜ż ˇ ˆ ˙ˇ2 ¸ 21 ˜ż ˇ ˆ ˙ˇ2 ¸ 21 2 2 ˇ ˇ ˇ ˇ ˇα̂ p1 ´ ξ qω ˇ dω ˇγ̂ p1 ´ ξ qω ˇ dω ˆ dξ, ˇ ˇ 2k ` 1 ˇ 2k ` 1 ˇ S S where we set Φpξ, kq “ sup \|Ψpω; ξ, kq\| (4.3) ωPS to simplify the notation. We remind that according to the Tomas-Stein theorem in R3 the inequality ˙ 21 ˆż 1 2 \|η̂ prωq\| dω ď Cs r ´3{s }η}Lsz S holds for 1 ď s ď 4 3 and r ą 0, where \|xP1 f, gy\| ď }α}Lsz 8 ÿ 1 s ` 1 s1 “ 1. Thus, we obtain (being }γ}Lsz “ 1) 1 k“0 p2k ` 1q 3´ s61 ż1 6 p1 ´ ξ 2 q2´ s1 Φpξ, kqdξ. (4.4) ´1 To estimate the integral in the last formula, we remind the definition (4.3) of Φ (see also (4.2)) and use the Cauchy-Schwarz inequality in the integral on vω to obtain ż1 ż1 6 2 2´ s61 Φpξ, kqdξ ď p1 ´ ξ 2 q2´ s1 p1 ´ ξ q ´1 ´1 ˇż ˇ 2 1´ξ ˇ ˇ ` ˘ 2k`1 ´1 ´1 ˆ ˇˇ Λk Fω pβ ˝ Rω q xω , yω ; ξ Fω pδ ˝ Rω qpxω , yω ; ξqdxω dyω ˇˇ dξ vω ď ż1 6 p1 ´ ξ 2 q2´ s1 ´1 ¸ 12 ˜ż ˇ ˇ2 1´ξ2 ˇ ˇ 2k`1 ` ˘ ´1 ˇ ˇΛ ˇ k Fω pβ ˝ Rω q xω , yω ; ξ ˇ dxω dyω vω ˆ ˆż vω ˇ ˇ ˇFω pδ ˝ Rω´1 qpxω , yω ; ξqˇ2 dxω dyω ˙1 2 dξ. 10 VALENTINA CASARINO AND PAOLO CIATTI A further application of the Cauchy-Schwarz inequality then implies ż1 6 p1 ´ ξ 2 q2´ s1 Φpξ, kqdξ ´1 ď ˜ż 1 12 p1 ´ ξ 2q4´ s1 ´1 (4.5) ¸1 ˇ2 ż ˇ 1´ξ2 2 ˇ 2k`1 ˇ ` ˘ ´1 ˇΛ ˇ ˇ k Fω pβ ˝ Rω q xω , yω ; ξ ˇ dxω dyω dξ vω ˆ ˆż 8 ż ´8 vω ˇ ˇ ˇFω pδ ˝ R´1 qpxω , yω ; ξqˇ2 dxω dyω dξ ω ˙ 12 . Since the Plancherel theorem applied to Fω yields ż8 ż ż ż8 ˇ ˇ ˇ ˇ2 ´1 ˇpδ ˝ Rω´1 qpxω , yω , vω qˇ2 dxω dyω dvω ˇFω pδ ˝ Rω qpxω , yω ; ξqˇ dξdxω dyω “ vω ´8 ´8 vω “ }δ}L2v “ 1, (being \| det Rω \| “ 1), (4.5) reduces to ż1 6 p1 ´ ξ 2 q2´ s1 Φpξ, kqdξ ´1 ď ˜ż 1 p1 ´ ξ 2 q 4´ 12 s1 ´1 ż vω (4.6) ¸1 ˇ 1´ξ2 ˇ2 2 ˇ 2k`1 ˇ ` ˘ ´1 ˇΛ ˇ . ˇ k Fω pβ ˝ Rω q xω , yω ; ξ ˇ dxω dyω dξ Then in force of the Koch-Ricci estimates (3.6) we deduce from (4.6) that ż1 1 1 1 6 p1 ´ ξ 2 q2´ s1 Φpξ, kqdξ ď C p2k ` 1qγ p p q´p p ´ 2 q ´1 ˆ ˜ż 1 p1 ´ ξ 2 q p 4´ 12 `2 p1 ´ 12 s1 ´1 q ˆż vω ˇ ` ˘ˇ ˇFω pβ ˝ R´1 q xω , yω ; ξ ˇp dxω dyω ω ˙2 p dξ (4.7) ¸ 21 . Applying the Hölder inequality in (4.7) to the integration in ξ with exponents p1 {2 and p {pp1 ´ 2q “ p{p2 ´ pq, we deduce that ż1 6 p1 ´ ξ 2 q2´ s1 Φpξ, kqdξ 1 ´1 ď C p2k ` 1q γ p p1 q´p p1 ´ 12 q ˆż 1 4p 3 p1 ´ ξ qp1´ s1 q 2´p `1 dξ 2 ´1 ˙ p1 ´ 21 ˛ 11 ¨ p ˙ pp1 ż 8 ˆż ˇ ` ˘ˇp ´1 ˇFω pβ ˝ R q xω , yω ; ξ ˇ dxω dyω dξ ‚ . ˆ˝ ω ´8 vω The first integral is finite since s1 ě 4, which implies that ˆ ˙ 3 2 4p 1´ 1 `1ě ą 0. s 2´p 2´p RESTRICTION ESTIMATES FOR THE FREE GROUP 11 Therefore, we have ż1 6 p1 ´ ξ 2 q2´ s1 Φpξ, kqdξ ´1 ¨ ż 8 ˆż γ p p1 q´p 1p ´ 21 q ˝ ď C p2k ` 1q ´8 vω ˇ ` ˘ˇ ˇFω pβ ˝ Rω´1 q xω , yω ; ξ ˇp dxω dyω ˙ pp1 ˛ 11 p dξ ‚ . Now, being p1 {p ě 1, we may use the Minkowski integral inequality, which gives ż1 6 p1 ´ ξ 2 q2´ s1 Φpξ, kqdξ ´1 ď C p2k ` 1q γ p p1 q´p p1 ´ 12 q ˜ż ˆż 8 ´8 vω ˇ ` ˘ˇ 1 ˇFω pβ ˝ R´1 q xω , yω ; ξ ˇp dξ ω ˙ p1 p dxω dyω ¸ p1 . Then we apply the Hausdorff-Young inequality in the inner integral to Fω and replace the coordinates xω , yω , vω with x, y, v, ż1 6 p1 ´ ξ 2 q2´ s1 Φpξ, kqdξ ´1 ď C p2k ` 1q γ p p1 q´p p1 ´ 12 q ˆż ˆż 8 ´8 vω ď C p2k ` 1q γ p p1 q´p p1 ´ 12 q deducing that ż1 ´1 ˆż v ˙ ˙ 1p ˇ ˇ ` ˇpβ ˝ R´1 q xω , yω , vω qˇp dvω dxω dyω ω ˇ ` ˇ ˇβ x, y, vqˇp dxdydv ˙ p1 , p1 ´ ξ 2q2´ s1 Φpξ, kqdξ ď C p2k ` 1qγ p p q´p p ´ 2 q }β}Lpv . 1 6 1 1 (4.8) Finally, plugging (4.8) in (4.4) we obtain \|xP1 f, gy\| ď C}α} }β} Lsz Lpv 8 ÿ p2k ` 1q s1 ´3`γ p p q´p p ´ 2 q 6 1 1 1 (4.9) k“0 ď C}α}Lsz }β}Lpv “ C}f }Lsz Lpv . The series in fact converges for 1 ď s ď 4{3 and 1 ď p ď 2 since by (3.7), when 1 ď p ď 56 , we have ˙ ˆ ˙ ˆ 1 1 6 1 ´ ď ´2 ´3`γ ´ s1 p p 2 and, when 6 5 ď p ď 2, 6 ´3`γ s1 ˙ ˆ ˙ ˆ 3 1 1 1 ´ ď´ . ´ p p 2 2 From (4.9) the estimate asserted in the statement follows by duality, proving the theorem. 12 VALENTINA CASARINO AND PAOLO CIATTI References [ACDS] [C1] [C2] [CCi1] [CCi2] [CCi3] [CCi4] [KR] [MMü] [Mü] [So1] [So2] [Str] [Th] F. Astengo, M. Cowling, B. Di Blasio, M. Sundari, Hardy’s uncertainty principle on certain Lie groups, J. London Math. Soc. (2), 62 (2000), no. 2, 461–472. V. Casarino, Norms of complex harmonic projection operators, Canad. J. Math., 55 (2003), no. 6, 1134–1154. V. Casarino, Two-parameter estimates for joint spectral projections on complex spheres, Math. Z., 261 (2009), no. 2, 245–259. V. Casarino and P. Ciatti, Transferring Lp eigenfunction bounds from S 2n`1 to hn , Studia Math., 194 (2009), no. 1, 23–42. V. Casarino and P. Ciatti, A Restriction Theorem for Métivier Groups, Advances in Mathematics, 245 (2013), 52–77. V. Casarino and P. Ciatti, Restriction estimates for the full Laplacian on Métivier Groups, Rend. Lincei Mat. Appl. 24 (2013), 165–179. V. Casarino and P. Ciatti, Lp Joint Eigenfunction Bounds on Quaternionic Spheres, J. Fourier Anal. Appl., electronically published on 14 October 2016, DOI: http://dx.doi.org/10.1007/s00041-016-9506-6 (to appear in print). H. Koch and F. Ricci, Spectral projections for the twisted Laplacian, Studia Math. 180 (2007), no. 2, 10–110. A. Martini and D. Müller, Lp spectral multipliers on the free group N3,2 , Studia Math. 217 (2013), no. 1, 41–55. D. Müller, A restriction theorem for the Heisenberg group, Ann. of Math. (2), 131 (1990), no. 3, 567–587. C. Sogge, Oscillatory integrals and spherical harmonics, Duke Math. J. (2), 53 (1986), no. 1, 43–65. C. Sogge, Concerning the Lp norm of spectral clusters for second-order elliptic operators on compact manifolds, J. Funct. Anal., 77 (1988), no. 1, 123–138. R. Strichartz, Harmonic analysis as spectral theory of Laplacians, J. Funct. Anal., 87 (1989), no.1, 51–148. S. Thangavelu, Harmonic Analysis on the Heisenberg Group, Progress in Mathematics, Vol. 159, Birkhäuser Boston Inc., Boston, MA, 1998. Università degli Studi di Padova, Stradella san Nicola 3, I-36100 Vicenza, Italy E-mail address: valentina.casarino@unipd.it Università degli Studi di Padova, Via Marzolo 9, I-35100 Padova, Italy E-mail address: paolo.ciatti@unipd.it	4math.GR
arXiv:1606.03809v3 [cs.PL] 29 Dec 2017 Why Nominal-Typing Matters in Object-Oriented Programming Moez A. AbdelGawad College of Mathematics and Econometrics, Hunan University Changsha 410082, Hunan, P.R. China Informatics Research Institute, SRTA-City New Borg ElArab, Alexandria, Egypt moez@cs.rice.edu January 1, 2018 Abstract type systems. Recently, a domain-theoretic model of nominally-typed OOP was compared to well-known models of structurally-typed OOP. Combined, these comparisons provide both a clear technical account and a deep mathematical account of the relation between nominal and structural OO type systems that have not been presented before, and they help demonstrate the key value of nominal typing and nominal subtyping to OO developers and language designers. We believe a clearer understanding of the key semantic advantage of pure nominal OO typing over pure structural OO typing can help remedy the existing schism. We believe future foundational OO PL research, to further its relevance to mainstream OOP, should be based less on structural models of OOP and more on nominal ones instead. The statements ‘inheritance is not subtyping’ and ‘mainstream OO languages unnecessarily place restrictions over inheritance’ have rippled as mantras through the PL research community for years. Many mainstream OO developers and OO language designers, however, do not accept these statements. In nominally-typed OO languages that these developers and language designers are dearly familiar with, inheritance simply is subtyping; and they believe OO type inheritance is an inherently nominal notion, not a structural one. Nominally-typed OO languages, such as Java, C#, C++, and Scala, are among the most used programming languages today. However, the value of nominal typing to mainstream OO developers, as a means for designing robust OO software, seems to be in wait for full appreciation among PL researchers—thereby perpetuating an unnecessary schism between many OO developers and language designers and many OO PL researchers, with each side discounting, if not even disregarding, the views of the other. In this essay we strengthen and complement earlier efforts to demonstrate the semantic value of nominal typing by presenting a technical comparison between nominal OO type systems and structural OO 1 Introduction In 1990, Cook et al. shocked the programming languages (PL) research community by declaring that, in object-oriented programming, ‘inheritance is not subtyping,’ meaning there is no one-to-one correspondence between type inheritance and subtyping in OO programming languages, further adding that mainstream OO languages unnecessarily ‘place re1 strictions over inheritance.’ Over the years, these statements rippled, as mantras, through the PL research community. To this day, however, many mainstream OO developers and OO language designers cannot digest or accept not identifying type inheritance with subtyping. Simply, the statement of Cook et al. is not true in nominally-typed OO languages that these developers and language designers are dearly familiar with. Further, they see that the so-called “restriction” of type inheritance in nominally-typed OO languages is strongly justified, or, more so, that a structural view of inheritance is in fact an unjustified redefinition of type inheritance, which they view as an inherently nominal notion. Nominally-typed OO languages are among the top most-used programming languages today. Examples of nominally-typed OO languages include many industrial-strength mainstream OO programming languages such as Java [44], C# [1], C++ [2], and Scala [61]. Nominally-typed OO languages have remained among the top most-used programming languages for over a decade [14, 5]. And, even by the most conservative measures, these languages are expected to remain among the top most-used OO programming languages in the near future, if not the far one, too. In spite of this, the value of nominal typing and nominal subtyping to mainstream OO developers, as a means for designing robust OO software that can be readily understood and maintained, as well as the value of properties of OO type systems that depend on nominality to them (such as the identification of type inheritance with subtyping), seem to be not yet fully appreciated among OO PL researchers. This has led to a continuing tension and schism between two large and significant communities: many mainstream OO developers and OO language designers, on one side, and many OO PL researchers, on the other side, with each of both sides discounting, and even disregarding, the views and opinions of the other. In nominally-typed (a.k.a., nominatively-typed) OO languages, objects and their types are nominal, meaning that objects and their types carry class names information as part of the meaning of objects and of the meaning of their types. Class names— and interface names and trait names in OO languages that support these notions—are used as type names in nominally-typed OO languages. Class—and interface and trait—behavioral contracts, typically written informally in code documentation comments, are specifications of the behavioral design intentions of OO software developers. In nominally-typed OOP, a reference to a class (or interface or trait) name is invariably considered a reference to the associated contract too. Given this association of type names to corresponding behavioral contracts, nominal typing allows associating types of objects with (formal or informal) behavioral contracts. By using type names in their code, OO developers using nominally-typed OO languages have a simple way to refer to the corresponding contracts— referring to them as richer specifications of object state and behavior that can be checked statically and that can be used during runtime. This readily access to richer object specifications—which cannot be expressed in a natural way using non-nominal (a.k.a., structural) record types that, by definition, include no class names information—makes nominally typed OO languages closer to being semantically typed languages than structurally typed OO languages are. The first mathematical models of OOP to gain widespread recognition among programming languages (PL) researchers were developed while OOP was making its first steps into mainstream computer programming. (See Section 2.) These early models were structural models of OOP. As the developers of these models themselves (e.g., Cardelli) explained, this was due to influence from functional programming research extant at that time. These models of OOP, thus, reflected a view of OOP that does not include class names information. Being structural, objects were viewed in these models simply as being records (of functions). Object types, in accordance, were viewed as record types, where the type of an object only specifies the structure of the object, meaning that object types carry information only on the names of the members of objects (i.e., fields and methods) and, inductively, on the (structural) types of those members. Examples of structurally-typed OO languages include lesserknown languages such as O’Caml [48], Modula-3 [27], 2 Moby [39], PolyTOIL [21], and Strongtalk [19].1 In pure structurally-typed OO languages, class names information (also called nominal information) is not used as part of the identity of objects and their types, neither during static type checking nor at runtime. Accordingly, nominal information is missing in structural mathematical models of OOP. The main practical advantage of structural typing over nominal typing in OO languages seems to be their “flexibility,” i.e., the ability in a structurallytyped OO language to have supertypes get defined “after the fact” (i.e., after their subtypes are already defined). In light of mainstream OO developers of statically-typed OO languages not adopting structural typing, the “inflexibility” of nominally-typed OO languages seems not to be enough justification for wider use of structural typing, particularly in light of the advantages of nominal typing we discuss in this essay. We attempt thus in this essay to further close the gap that exists between programming language researchers who maintain a structurally-typed view of OOP (and who believe in conclusions based on this view, such as inheritance and subtyping not being in one-to-one correspondence) and mainstream OO software developers and OO language designers who maintain a nominally-typed view of OO software (and who, accordingly, reject conclusions based on the structural view) by giving a precise technical account of the relation between nominal and structural OO type systems. The essay complements the recent mathematical comparison of a nominal domain-theoretic model of OOP to structural domain-theoretic models of OOP. This essay is structured as follows. First, in Section 2 we give some details on the history of modeling OOP, particularly details relevant to realizing differences between nominal typing and structural typing and to the development of nominal and structural models of OOP. Given that structural typing, more or less, is understood well among PL researchers, in Section 3 we directly demonstrate the value of behavioral contracts and nominal typing in mainstream nominally-typed OO languages using a comparison, followed by a discussion of the comparison. In Section 3.1 we first discuss, in some detail, the value of contracts and the value of identifying inheritance with subtyping in mainstream OO software design. Then, using code examples, in Section 3.2 and Section 3.3 we compare nominally-typed OO type systems and structurallytyped ones to more vividly illustrate the main technical differences between them. We then conclude in Section 3.4 by discussing the nominal and structural views of type names and of recursive-types, and the importance of recursive types in mainstream OOP. We conclude the essay by summarizing our findings, and making some final remarks, in Section 4. 2 Related Work Even though object-oriented programming emerged in the 1960s, and got mature and well-established in mainstream software development in the 1980s, the differences between nominally-typed and structurally-typed OO programming languages started getting discussed by PL researchers only in the 1990s [53, 65, 68]. In spite of an early hint by Cardelli (see below), the value of investigating nominal typing and nominal subtyping and their value to OO developers was not appreciated much however—that is, until about a decade later, around the year 2000. In the eighties, while mainstream OOP was in its early days, Cardelli built the first denotational model of OOP [24, 25]. Cardelli’s work was pioneering, and naturally, given the research on modeling functional programming extant at that time (which Cardelli heavily referred to and relied on), the model Cardelli constructed was a structural denotational model of OOP.2 1 A discussion of statically-typed versus dynamically-typed OO languages (including the non-well-defined so-called “ducktyping”), and the merits and demerits of each, is beyond the scope of this essay. The interested reader should check [56]. In this essay, we focus on nominal and structural statically-typed OO languages. 2 Quite significantly, Cardelli in fact also hinted at looking for investigating nominal typing [26, p.2]. Sadly, Cardelli’s hint went largely ignored for years, and structural typing was rather assumed superior to nominal typing instead, particularly after 3 In the late eighties/early nineties, Cook and his colleagues worked to improve on Cardelli’s model. Unlike Cardelli, Cook et al. emphasized in their work the importance of self-references in OOP, at the value level and at the type level. Their research led them to break the identification of type/interface inheritance with subtyping [32, 35, 34].3 In 1994, Bruce and others presented a discussion of the problem of binary methods in OOP [22] (a “binary method” is a method that takes a parameter or more of the same type as the class the method is declared in4 ). Later, Bruce and Simons promoted the structural view of OOP and conclusions based on it in a number of publications (e.g., [23] and [66]), in spite of the disagreement between these conclusions and the fundamental intuitions of a significant portion of mainstream OO developers and language designers [28]. Under the pressure of this disagreement, some PL researchers then started in the late nineties/early 2000s stressing the significance of the differences between nominally-typed OOP and structurally-typed OOP, and they started acknowledging the practical value of nominal typing and nominal subtyping [63]. Accordingly some attempts were made to develop OO languages with complex type systems that are both nominally- and structurally-typed [38, 62, 42, 54, 55, 61, 4]. However, in the eyes of mainstream OO developers, these “hybrid” languages have more complex type systems than those of languages that are either simply purely nominally-typed or purely structurallytyped. This more complexity typically results in lesser productivity for developers who attempt to use both of the typing approaches in their software (see also discussion at the end of Section 3.4). As to operational models of OOP, Abadi and Cardelli were the first to present such a model [6, 7]. Again, their model had a structural view of OOP. However, operational models of nominallytyped OOP got later developed. In their seminal work, Igarashi, Pierce, and Wadler presented Featherweight Java (FJ) [45] as an operational model of a nominally-typed OO language. Even though FJ is not the first operational model of nominallytyped OOP (see [36], [60] and [40, 41], for example), FJ is the most widely known operational model of (a tiny core subset of) a nominally-typed mainstream OO language, namely Java. The development of FJ and other operational models of nominallytyped OOP motivated the construction of NOOP as the first domain-theoretic model of nominally-typed OOP [8, 10].5 Given the different basis for deriving data structuring in functional programming (based on standard branches of mathematics) and in object-oriented programming (based on biology and taxonomy6), 5 Featherweight Java [45] offers a very clear operational semantics for a tiny nominally-typed OO language. It is worth mentioning that domain-theoretic models of nominally-typed OO languages, as more foundational models that have fewer assumptions than operational ones, provide a denotational justification for the inclusion of nominal information in FJ. The inclusion of nominal information in NOOP was crucial for proving the identification of inheritance with subtyping in nominally-typed OOP [8, 10]. In FJ [45], the identification of inheritance with subtyping was taken as an assumption rather than being proven as a consequence of nominality. Also, domain-theoretic models such as NOOP allows discussing issues of OOP such as type names, ‘self-types’ and binary methods on a more foundational level than provided by operational models of OOP such as FJ. The more abstract description of denotational models results in a conceptually clearer understanding of the programming notions described as well as the relations between them. (It is also worth mentioning that NOOP was developed, partially, in response to the technical challenge Pierce presented in his LICS’03 lecture [64] in which Pierce looked for the precise relation between structural and nominal OO type systems, notably after the development of FJ was concluded, implying that the question about the relation remained an open question after the development of FJ. As to their purpose, it is customary that denotational models and operational ones play complementary roles, where denotational models are usually of more interest to programming language designers, while operational ones are usually of more interest to programming language implementers.) 6 Which is a fact that seemingly is nowadays forgotten by some PL researchers but that Cardelli explicitly mentions in [24, 25]. the publication of Cook et al.’s and Bruce et al.’s work as we later discuss. 3 A discussion of Cardelli’s model and of Cook’s model and a comparison of NOOP, a nominal domain-theoretic model of OOP [8, 10], versus the structural models of Cardelli and Cook is presented in [11, 13]. 4 See our later discussion of ‘false/spurious binary methods’—methods that are mistakenly identified by structural OO type systems as being binary methods but their semantics are not those of true binary methods. 4 we uncover a third so-far-unrecognized problem in pure structurally-typed OO languages that we call the problem of ‘spurious binary methods.’ We then conclude our demonstration of the value of nominal typing by discussing, in some depth, the nominal and structural views of type names and of recursive-types, and the importance of recursive types in mainstream OOP in Section 3.4. The discussions and comparisons in this section demonstrate that nominal typing in nominally-typed OO programming languages causes typing and subtyping in these languages to be closer to semantic typing and subtyping, respectively, because of the association of nominal information with class contracts. This closeness to semantic typing, the simplicity of the resulting software design mental model, and the importance of recursive types to mainstream OO developers and OO language designers help explain the practical value of nominal typing to mainstream OO developers and OO language designers. some PL researchers have also expressed dissatisfaction with assuming that the views of programming based on researching functional programming (including a view that assumes structural typing) may apply, without qualifications, to object-oriented programming. In addition to Pierce and other earlier and later researchers pointing out the importance of distinguishing between nominal typing and structural typing, MacQueen [51], for example, also noted many mismatches between Standard ML (a popular functional programming language [59]) and classbased OO languages such as Java and C++. Later, Cook [33] also pointed out differences between objects of OOP and abstract data types (ADTs) that are common in functional programming. These research results run in a similar vein as ours since they somewhat also point to some mismatches between the theory and practice of programming languages—theory being more mathematics-based, functional, and structurally-typed, and practice being more biology/taxonomy-based, object-oriented, and nominally-typed. 3 3.1 Nominally-Typed OOP versus Structurally-Typed OOP Contracts, Nominality and The Liskov Substitution Principle (LSP) Contracts are widely-used notions in mainstream OO software development. A contract in an OO program is similar to a contract in the real world: It specifies what an object expects of client objects and what client objects can expect of it. Members of an object—i.e., its fields and methods—and properties of these members form the object’s interface with the outside world; the buttons on the front of a television set, for example, are the interface between us and the electrical wiring on the other side of the TV’s plastic casing. One presses the “power” button and he or she are promised this will turn the television on and off. In its most common form, an interface is a group of related methods together with a contract giving promises on the behavior of these methods. Similarly, a class contract is an agreement that instances of the class will expose (present as their public interface, or API) certain methods, certain properties, and certain behaviors. In this section we first informally and less-technically, discuss the importance of contracts, and of nominal typing and nominal subtyping to mainstream OO developers in Section 3.1—briefly discussing DbC (Design by Contract) in the process, then we discuss how LSP (Liskov’s Substitution Principle) expresses the importance of preserving contracts upon inheritance. In Section 3.2 and Section 3.3, we then present code examples that illustrate how nominally-typed OOP and structurally-typed OOP compare to each other from a technical point of view, and illustrate how structural typing and structural subtyping sometimes force the breaking of contracts. In the comparison we discuss two key problems with structural OO type systems, namely, ‘spurious subsumption,’ and its converse, ‘missing subsumption.’7 In Section 3.3.1, 7 A technical overview of main OO typing notions that explains some of the technical jargon used in this essay is pre- sented in [9]. 5 Provider Client Contract Examples Examples of contracts in OO software are plenty. Examples familiar to most Java developers, for example, include the contract of the Comparable interface promising its clients a total ordering on its elements and requiring that classes that implement the interface adhere to this promise, and the contract of class Object promising that the equals() method and the hashCode() method are in agreement and requiring subclasses that override one of the two methods to override the other method accordingly. Also in Java, class JComponent contains a default implementation of all of the methods in the Accessible interface, but JComponent is not actually declared to implement the interface, because the contract associated with Accessible is not satisfied by the default implementation provided in JComponent. This example stresses the association of inherited contracts with superclass names. In mainstream OO software, if a class extends another class or implements an interface it is declaring that it inherits the contract associated with the superclass or superinterface and will maintain it. Likewise, if a class does not maintain the contract associated with another class or interface it does not declare itself as extending or implementing the class or interface. We discuss this point further in Section 3.1.1. Other examples of contracts may also include a class that implements tree layout algorithms. The contract of such a class may require the input graph to be a tree, and may promise as result, if the input is a tree, to produce a layout that has no overlapping nodes, edges or labels. In general, contracts, whether written formally or informally, usually contain the following pieces of information: side effects, preconditions, postconditions, invariants, and, sometimes, even performance guarantees. In Java, class contracts, as a set of requirements and promises, are usually stated in Javadoc comments. Requirements of a contract are simply any conditions on the use of the class, for example: conditions on argument values, conditions on the order of execution of methods, conditions on execution in a multi-threaded or parallel environment. Two further, rather artificial, examples for contracts, that we will expound on below to show the dif- Benefit (4) Output guaranteed to comply to postconditions (no need to check output) (2) Input guaranteed to comply to preconditions (no need to check input) Obligation (1) Satisfy preconditions (3) Satisfy postconditions Table 1: Design By Contract (DbC): Benefits and Obligations [Source: [46]] ferences between nominal typing and structural typing, are the promise that an animal can play with any another animal, and the promise that a (mathematical) set contains no repeated elements. In particular, we use these two examples to show how structural subtyping can lead to breaking contracts associated with classes/interfaces/traits. From the presented examples, it is easy to see that a contract is made of two parts: requirements upon the caller (“the client”) made by the class (“the provider”) and promises made by the class to the caller. If the caller fulfills the requirements, then the class promises to deliver some well-defined service. Requirements may be enforced by throwing checked or unchecked exceptions when the stated conditions are violated. Promises can be enforced by assertions at the end of a method. Further, according to proponents of ‘Design By Contract’ (DbC), classes of a software system communicate with one another on the basis of precisely defined benefits and obligations [57, 58]. If preconditions are not obeyed by the client of the class method, the service provider will deny its service. If any postcondition or invariant is violated, it uncovers a problem on the service provider side [46]. As such, the benefits and obligations of clients and providers, along with their relative chronological order, can be summarized as in Table 1. Nominally-Typed OOP Moving on from DbC to the design of mainstream OO type systems, OO languages should ideally include behavioral contracts in object types. Motivated by DbC, contracts are used 6 Provider Client by mainstream OO developers for constructing robust, reliable, reusable and maintainable OO software, since contracts promise specified properties of objects. For example, in his widely-known book titled ‘Effective Java’ [16, 17], Joshua Bloch reflects on the use of contracts in mainstream OOP and asserts the value of contracts to OO software design by writing that, ‘No class is an island. Instances of one class are frequently passed to another. Many classes depend on the objects passed to them obeying the contracts associated with their superclasses ... once you’ve violated the contract, you simply don’t know how other objects will behave when confronted with your object.’ Obligation (1) Satisfy contract requirements (3) Satisfy contract promises Table 2: Contracts in Nominally-Typed OOP: Benefits and Obligations world. In a statically-typed language the tractable component of the contract is enforced at build time by the compiler. In Java, for example, if a class claims to implement an interface, all methods defined by that interface must appear in its source code before the class will successfully compile, and during run time it is assumed the promises given by the interface are maintained by the class. The same happens if a class claims to extend another class, or if an interface claims to extend another interface. This inheritance of requirements and promises is sometimes referred to as interface inheritance, contract inheritance, or type inheritance.8 While using inheritance, it may be necessary to make changes to a superclass contract. Some changes to a specification/contract will break the caller, and some will not. For determining if a change will break a caller, professional OO developers use the memorable phrase “require no more, promise no less”: if the new specification does not require more from the caller than before, and if it does not promise to de- In practice however, the inclusion of behavioral contracts in object types is too much to ask of a type checker (because of the general problem of not being able to statically check contracts, since behavioral contracts are remarkably expressive). The solution OO language designers choose is to go with an approximation. The association of class names with contracts, and OO type systems respecting nominal information in typing and subtyping decisions, allows a nominally-typed OO type system to be a tractable approximation of DbC; hence, OO language designers of many mainstream OO languages use nominal typing. Nominally-typed OO languages typically do not require the enforcement of requirements and promises of contracts; requirements and promises are rather assumed to hold, thereby encouraging but not requiring developers to enforce the contracts. To accurately reflect how contracts are used in nominally-typed OOP, Table 1 can as such be modified into Table 2. 3.1.1 Benefit (4) Output assumed to comply to contract promises (no need to check output) (2) Input assumed to comply to contract requirements (no need to check input) 8 According to [67], ‘Programmers employ inheritance for a number of different purposes: to provide subtyping, to reuse code, to allow subclasses to customise superclasses’ behaviour, or just to categorise objects’. Inheritance as only being ‘a method by which classes share implementations’ (i.e., it being a ‘code sharing/code reuse’ technique), is a very limited notion of inheritance that, unfortunately, is still entertained by some OO PL researchers. Code sharing is only a part of the fuller picture of OO inheritance, and is only a means towards the higher goal of classes sharing their contracts (and even their architectures [37, 14]), not just their code. Inheritance as ‘contract sharing’ is the notion of inheritance that we are generally interested in and discuss in this essay. Inheritance, Subsumption and Contract Preservation As to the inheritance of contracts in mainstream OOP, implementing an interface, for example, allows a class in an OO program to become more formal about the behavior it promises to provide. Interfaces form a contract between a class and the outside 7 liver less than before, then the new specification is compatible with the old, and will not break the caller. Bloch [16, 17], hinting at the conventional wisdom among mainstream OO developers that identifies type inheritance with subtyping, proceeds to conclude, based on his earlier observations about contracts, that programming. Whenever these principles are violated, a program becomes more difficult to understand and to maintain. (For example, after his discussion of contracts Bloch then gives examples demonstrating problems, in the Java libraries which Bloch himself coauthored, that resulted from violating these principles [16, 17].) Based on the discussion of contracts and inheritance, two clear OO design principles among professional mainstream OO developers are: ‘inheritance [of contracts] is appropriate only in circumstances where the subclass really is a subtype of the superclass.’ 1. Whenever the contract of a class/interface/trait is obeyed by an inheriting class (i.e., whenever we have type inheritance) we should have subsumption between corresponding class types (i.e., we should have subtyping), and, conversely, As such, Bloch concludes that ‘it is the responsibility of any subclass overriding the methods of a superclass to obey their general contracts; failure to do so will prevent other classes that depend on the contracts from functioning properly in conjunction with the subclass.’ 2. Whenever we have subsumption between two class types (i.e., whenever we have subtyping) the contracts (of the superclass type) should be obeyed by the subclass type (i.e., we should have type inheritance). The requirement that subclasses maintain the contracts of their superclasses is expressed among professional OO developers by stating that well-designed OO software should obey the Liskov substitution principle (LSP) [49, 50]. According to Bloch [16, 17], Given the importance of the LSP (as expressing the preservation of contracts upon inheritance) to mainstream OO developers, and given that contracts are typically specifications of object behavior, it is easy to conclude that basing typing on contracts so as to make typing and subtyping closer to behavioral typing and behavioral subtyping (sometimes also called ‘semantic typing’ and ‘semantic subtyping’) is a desirable property of an OO language. To further illustrate the importance of identifying inheritance with subtyping to OO software developers, in the following two sections we present code examples that point out two problems with structural subtyping that do not exist in nominally-typed OOP. The first problem is what is sometimes called the problem of ‘spurious subsumption’ ([63, p.253]), and the second is the problem of inheritance not implying subtyping (which we call ‘missing subsumption’), i.e., that inheritance between two classes does not imply subsumption between the two corresponding class types (the converse of the spurious subsumption problem.)10 ‘The LSP says that any important property of a type should also hold for its subtypes, so that any method written for the type should work equally well on its subtypes [49, 50]’.9 As such, as demonstrated by Bloch, it is common knowledge among professional mainstream OO developers that subsumption (as expressed by the LSP) and the identification of inheritance with subsumption (i.e., subtyping between object types) are an integral part of the mental model of object-oriented 9 The LSP is the third of the five OO design principles (the L in ‘SOLID’) mainstream OO developers follow to design robust OO software. In the jargon of OO developers, OO code “smells” (in particular, it has a ‘refused bequest’) if some class in the code does not obey LSP, i.e., if a derived class—that is, a subclass—in the code breaks the contract of one of its base classes-–that is, of one of its superclasses. The LSP thus expresses that class contracts are preserved by inheriting classes. 10 The code examples can be skipped by a reader familiar with these two problems. 8 In the spurious subsumption problem, we have two classes whose instances do not maintain the same contract but are considered subtypes according to structural subtyping rules, demonstrating an example of structural subtyping breaking LSP. In the missing subsumption problem, we have two classes whose instances do maintain the same behavioral contract but that are not considered subtypes in a structural type system, due to structural type systems rebinding self-types upon inheritance, demonstrating an example of structural subtyping thus breaking the identification between inheritance and subtyping. Further, due to pure structural OO type systems always requiring the rebinding of self-types upon inheritance, we also point out a so-far-unknown problem with pure structural subtyping that we call the problem of ‘spurious binary methods.’ We note in the code for class Set and class MultiSet that the two classes support precisely the same set of four operations on their instances having the same signatures for these four operations. The different contracts associated with the two classes, specifying that the semantics and run-time behavior of their instances should agree with the corresponding mathematical notions, are reflected only in the different class names of the two classes but not in the structure of the two classes. Given that nominally-typed OOP respects class names (and thus the associated class contracts) while structurally-typed OOP ignores them, we have the final assignment in 3.2 correctly disallowed in nominally-typed OOP, but is wrongly allowed in structurally-typed OOP. The assignment is not allowed in nominally-typed OOP because class MultiSet does not inherit from class Set, while it is allowed in structurally-typed OOP because of matching signatures of all operations supported by the two classes. The assignment should not be allowed because, if allowed (as demonstrated in the code above), it will allow repetition of a set’s elements in the value bound to variable s by the assignment (given that it is an instance of MultiSet). Variable s, by its declaration, is assumed to be a Set (with no repeated elements), and the assignment, if allowed, will thus break the contract of class Set associated with variable s. The problem of spurious subsumption is similar to the problem of accidentally mistaking values of a datatype for those of another datatype (think of using floats for modeling euros and dollars then mistaking euros for dollars or mistaking floats for either [63].) Similarly, OO software developers think of an object in the context of its class hierarchy and of the contracts associated with its class members. A key prescript of nominally-typed OOP is that class contracts are inherited along with class members and should be maintained by instances of inheriting subclasses. Spurious subsumption in structurally-typed OOP al- MultiSet m = new MultiSet (); m . insert (2); m . insert (2); Set s = m ; // Allow assignment ? Spurious Subsumption In structurally-typed OOP, subtyping does not imply inheritance, that is, we may have subtyping between types corresponding to two classes (instances of one can be used as that of the other) but not have an inheritance relation between these two classes. To illustrate, let us assume the following definitions for class Set and class MultiSet (where the contract of class Set disallows repetition of elements of a Set, in agreement with the mathematical definition of sets, whereas the contract of class MultiSet allows repetition of elements of a MultiSet, in agreement with the mathematical definition of multisets, sometimes also called bags), class Set { Boolean equals ( Object s ) { ... } Void insert ( Object o ) { ... } Void remove ( Object o ) { ... } Boolean isMember ( Object o ) { ... } } class MultiSet { Boolean equals ( Object ms ) { ... } Void insert ( Object o ) { ... } Void remove ( Object o ) { ... } Boolean isMember ( Object o ) { ... } } 9 lows for unintended breaking of this rule, since a structural type checker fails to reject a program that uses an object of one type where a behaviorally different, but structurally compatible, type is expected.11 . 3.3 Animal mate ( Animal a ) { ... } } class Cat extends Animal { ... // behavior specific to cats Inheritance Is Not Subtyping Another problem with structural subtyping is the converse of the spurious subsumption problem. In structurally-typed OO languages, structural subtyping does not require subtyping between types of classes that are in the inheritance relation (and thus have inherited contracts), i.e., inheritance does not imply subtyping. In combination with subtyping not implying inheritance (i.e., spurious subsumption), structural subtyping thus totally separates the notions of inheritance and subtyping, based on its nonnominal view of inheritance which ignores the inheritance of class contracts associated with class names. To illustrate inheritance not implying subtyping in structurally-typed OOP, assume the following definitions for class Animal and class Cat class Animal { Void move ( Point to ) { ... } Void eat ( Food some ) { ... } Void breathe () { ... } Void sleep ( Time period ) { ... } ... // more generic animal behavior , // and ... 11 For example, as mentioned earlier, interface Comparable in Java, consisting of the single abstract method int compareTo(Object o), has a public contract asserting that compareTo defines a total ordering on instances of any class inheriting from Comparable, and that clients of Comparable can thus depend on this property. An arbitrary class with a method int compareTo(Object o), however, generallyspeaking does not necessarily obey this contract. In structurally-subtyped OO languages, instances of such a class can be bound, by spurious subsumption, to variables of type Comparable (similar to allowing the binding of the instance first bound to variable m to variable s in the Set/MultiSet example above). This is in spite of the fact that when a developer asserts that a class C implements Comparable, he or she is asserting that compareTo() defines a total ordering on class C, and the author of Comparable conversely asserting that if a class C does not implement the Comparable interface, instances of C cannot be bound to variables of type Comparable, since the compareTo method of C may unintentionally or intentionally not define a total ordering on instances of C. // generic animal behavior inherited // from class Animal // but , do Cats ... Cat mate ( Cat c ) { ... } // OR do they ... Animal mate ( Animal a ) { ... } // ??? } Structurally-typed and nominally-typed OOP disagree on the signature of method mate() in class Cat. Structurally-typed OOP assumes that mate() is a binary method (See [22], where binary methods were recognized as problematic and multiple approaches were suggested for dealing with them. We discuss them more technically in Section 3.4), and requires the method to have the “more natural” signature Cat mate(Cat c), at the expense of making Cats (i.e., instances of class Cat) not be Animals (i.e., instances of class Animal)12, when they (quite naturally) should be ones13 . Nominally-typed OOP, on the other hand, does not assume the mate() method in class Animal to be a binary method. It thus keeps using the same signature for the method upon its inheritance by class Cat. In nominally-typed OOP, thus, Cats are indeed Animals, meaning that nominally-typed mainstream OOP does identify inheritance with subtyping. Given how nominally-typed OOP and structurally-typed OOP differ on whether inheritance implies subtyping, 12 That is, in structurally-typed OO languages the structural type corresponding to class Cat is not a subtype of the structural type corresponding to class Animal (because of the contravariance of types of method arguments), unless some unintuitive structural notion, like ‘matching’ [23] (which expresses ‘the similarity of recursive structure’ between class Cat and class Animal, but which did not gain traction or support in mainstream OOP), is added to the language to be used as a pseudo-replacement for subtyping. (See also discussions in Section 3.4 and in [11, 13]) 13 One may here recall Cardelli’s noting, in [24, 25], of the ‘biological origin’ of OOP. This biological origin is the reason OO inheritance is called ‘inheritance’ in the first place. 10 the assignment Animal a = new Cat() is correctly allowed in nominally-typed OOP, but is wrongly disallowed in structurally-typed OOP. To summarize, the code examples presented above demonstrate a fundamental difference between structurally-typed OOP and nominally-typed OOP from the perspective of OO developers. In structurally-typed OOP, where only class structure is inherited but not class contracts, MultiSets are Sets (contrary to their mathematical definition) and Cats are not Animals (contrary to their “biological definition”). In nominally-typed OOP, where class contracts are inherited (via class names) in addition to class structure, MultiSets are not Sets (in agreement with their mathematical definition) and Cats are Animals (in agreement with their “biological definition”.) In nominally-typed OOP whether subtyping is needed or not is indicated by the presence or absence of explicit inheritance declarations. Accordingly, the code examples above make it clear, more generally, that: Structurally-typed OOP sometimes forces subtyping when it is unneeded, and sometimes bars it when it is needed, while nominally-typed OOP only forces subtyping when it is explicitly needed, and bars it when, by omission, it is explicitly unneeded. This conclusion demonstrates a fundamental semantic and practical value of nominal information to OO developers of nominally-typed OO programming languages. method any method inside a class that takes an argument that has the same type as that of the class the method is declared in. This causes methods like playWith in class Animal, whose semantics is not that of true binary methods, to be mistaken as ones, and thus, in subclass Cat, for example, the playWith method will have the restrictive signature Void playWith(Cat a) (allowing Cats to play only with other Cats but not with other Animals.) We call this so-far-unrecognized problem with structural typing as the problem of ‘spurious binary methods’ (or, ‘false binary methods’), since a method is inadvertently considered as being a binary method when it should not be. Nominallytyped OO languages do not suffer from this problem, because nominally-typed OO languages treat any such method as a regular (i.e., non-binary) method, and thus the signature of the method does not change upon inheritance in a nominally-typed OO language.15 3.4 Nominal Typing, Type Names, Recursive Types and Binary Methods Based on the discussion in the previous sections, a fundamental technical difference between nominallytyped OO languages and structurally-typed OO languages clearly lies in how the two approaches of typing OO languages differently view and treat type names. 3.3.1 Spurious Binary Methods In structurally-typed OO languages, type names are viewed as being names for type variables that abA lesser-recognized problem with structurally-typed breviate type expressions (i.e., are “shortcuts”). As OOP, which is also related to binary methods, is the such type names in structurally-typed OO languages fact that, in a pure structurally-typed OO language (i.e., one with no nominal typing features), a class 15 In nominally-typed OOP, ‘F-bounded Generics’ (as used, like Animal cannot have a method like, say, for example, to define the generic class Enum in Java [43, 44]) offers a somewhat better alternative—if also not a fully satisfacVoid playWith(Animal a) that keeps having the same signature in its subclass tory one—to support true binary methods, while keeping the identification between inheritance and subtyping. Based on Cat14 and any other subclasses of class Animal. Pure some preliminary research we made, we expect future research structurally-typed OOP always treats as a binary to offer a more satisfactory alternative for supporting true bi14 So as to allow Cats to play with Dogs and Mouses (i.e., mice!), for example. nary methods in nominally-typed OO languages—hopefully a fully satisfactory alternative. (Also see related discussion close to the end of Section 3.4.) 11 are useful, and are even necessary for defining recursive type expressions. As variable names, however, recursive type names in structurally-typed OO languages (such as the name of a class when used inside the definition of the class, which gets interpreted as ‘self-type’) get rebound to different types upon inheritance, and they get rebound to types that, if they were subtypes, could break the contravariant subtyping rule of method parameter types and thus break the type safety of structurally-typed OO languages. Structurally-typed OO languages resolve this situation by breaking the one-to-one correspondence between inheritance and subtyping as we demonstrated in the earlier code examples in Section 3.2 and Section 3.3. In nominally-typed OO languages, however, nominality of types means type names are viewed as part of the identity and meaning of type expressions given the association of type names with public class contracts. This means that class names cannot be treated as variable names. Accordingly, in a nominally-typed OO program type names have fixed meanings that do not change upon inheritance. Further, the fixed type a type name is bound to in a nominally-typed OO program does not break the contravariant subtyping of method parameters when the method and its type get inherited by subclasses, thus not necessitating breaking the identification between inheritance and subtyping as we demonstrated by the earlier code examples in Section 3.2 and Section 3.3. In class-based OOP, a class can directly refer to itself (using class names) in the signature of a field, or that of a method parameter or return value. This kind of reference is called a self-reference, a recursive reference, or, sometimes, a circular reference. Also, mutually-dependent classes, where a class refers to itself indirectly (i.e., via other classes), are allowed in class-based OOP. This kind of reference inside a class, indirectly referencing itself, is called an indirect self-reference, a mutually-recursive reference, or, a mutually-circular reference. As Pierce [63] noted, nominally-typed OO languages allow readily expressing circular (i.e., mutually-dependent) class definitions. Since objects in mainstream OOP are characterized as being selfreferential values (are ‘self-aware,’ or ‘autognostic’ ac- cording to Cook [33]), and since self-referential values can be typed using recursive types [52], there is a strong and wide need for circular class definitions in OOP. As such, direct and indirect circular type references are quite common in mainstream OOP [33]. The ease by which recursive typing can be expressed in nominally-typed OO languages is one of the main advantages of nominally-typed OOP. According to Pierce [63, p.253], ‘The fact that recursive types come essentially for free in nominal systems is a decided benefit [of nominally-typed OO languages].’ As a demonstration of the influence of views of selfreferential classes on properties of OO type systems, when nominal and structural domain-theoretic models of OOP are compared (as done in brief, e.g., in [8, 10], and in detail, e.g., in [11]), it is easy to see that self-referential classes are viewed differently by nominal models of OOP than by structural models and that these different views of self-referential classes, in particular, make nominal domain-theoretic models of OOP lead to a simple mathematical proof of the identification between type inheritance and subtyping—a different conclusion than the one reached based on structural models. (In particular, the inclusion of class signature constructs in NOOP led to the simplicity of the mathematical proof of this identification. See [8, 10, 13] for more details.) Aside from theory, the difference between the nominal and the structural views of type names in OOP demonstrates itself most prominently, in practice, in the different support and the different treatment provided by nominally-typed OO languages and by structurally-typed OO languages to “binary methods”. As mentioned in Section 3.3, a “binary method” is a method that takes a parameter or more of the same type as the class the method is declared in [22]. The problem of binary methods and requiring them to be supported in OO languages was a main factor behind structural models of OOP leading to not identifying type inheritance with subtyping. Structurally-typed OO languages, given their view of type names as type variable names that can get rebound, require the type of the argument of a method, 12 when identified as a binary method and upon inheritance of the method, to be that of the type corresponding to the subclass. Nominally-typed OO languages, on the other hand, with their fixed interpretation of type names, offer a somewhat middle-ground solution between totally avoiding binary methods and overly embracing them as pure structurally-typed OO languages do. Nominally-typed OO languages treat a method taking in an argument of the same class as that in which the method is declared like any other method, needing no special treatment. Nominally-typed OOP thus does not quite support binary methods, but, for good reasons (i.e., so as to not break the identification of inheritance of contracts with subtyping nor lose other advantages of nominal typing), offers only a good approximation to binary methods. Given that the meaning of types names in nominally-typed OO languages does not change upon inheritance, these languages provide methods whose type, upon inheritance, only approximates that of true binary methods: The type of the input parameter of a method that approximates a binary method is guaranteed to be a supertype of its type if it were a true binary method. Given that the type of the parameter does not change in subclasses, the degree of approximation gets lesser the deeper in the inheritance hierarchy the method gets inherited. In light of the ‘spurious binary methods’ problem we uncovered in structural OO type systems (see Section 3.3), we believe providing approximations to binary methods is a smart design choice by nominal OO type systems, even if it is likely that avoiding spurious binary methods may have not been consciously intended. It should also be noted that the problem of spurious binary methods provides justification for nominally-typed OO languages being cautious about fully embracing binary methods by treating a method that looks like a binary method as indeed being one. Also, as we hinted to in earlier sections, in our opinion structural typing having arguably better support for binary methods does not justify using structural OO typing, since structural type systems have their own problems in their support of binary methods (i.e., the problem of spurious binary methods). We conjecture that F-bounded generics, or some other notion16 may provide a better solution for binary methods in nominally-typed (and possibly also in structurally-typed OOP) that does not require breaking the identification of inheritance with subtyping (and thus does not require sacrificing the closeness of nominal typing/subtyping to semantic and behavioral typing/subtyping and other advantages of nominal typing). In light of the ‘spurious binary methods’ problem, and requiring no explicit use of generics, in our opinion a better approach towards supporting binary methods in mainstream OO languages might be by allowing developers to explicitly mark or flag binary methods as being such, or, even more precisely, to allow developers to mark specific parameters of methods as being parameters that need to be treated as those of binary methods. As to “hybrid” OO languages, which add (some) structural typing features to nominally-typed OO languages17 , or vice versa, we conjecture that the useful part of the claimed “flexibility” of structural typing may be possible to achieve in nominally-typed OO languages by supporting a separate notion of ‘contract names,’ thereby splitting class names from contract names, then allowing classes to additionally define themselves as satisfying supercontracts of other already-defined (sub)classes. We have not explored this suggestion further however, since we believe it may complicate nominal OO type systems. But we believe that, if flexibility is an absolute necessity, then splitting class names from contract names may be a suggestion worthy of investigation. We believe this suggestion to be a more viable option—simpler to reason about and more in agreement with the nominal spirit of nominally-typed OO languages—than using multiple dispatch ([30, 18, 31]), which was discussed in [22] as a possible solution to the problem of binary methods, and also more viable than creating hybrid languages. We believe that having an OO type system be both nominally and structurally typed, as 16 For example, “implicit self-type-variables”, e.g., This/Self, which are implicit type variables because they do not get included in class signatures, similar to “implicit self-variables”, e.g., this/self, which are not included in method signatures. 17 Due to problems with supporting recursive types mentioned above, we believe most of these “hybrid” languages do not support recursive structural types, for example. 13 in [38, 62, 42, 54, 55, 61, 4], makes the type system does not necessarily imply the similarity of behavvery complex (and probably even lends its “hybrid” ior, in nominally-typed OOP type inheritance implies features unusable.) refined contracts, and refined contracts imply subsumption between class types, and vice versa (i.e., in nominally-typed OOP, subsumption between class 4 Concluding Remarks types implies refined contracts, implying type inheritance.) In this essay we added to earlier efforts that aimed Putting these facts together, it is clear that in to demonstrate the semantic value of nominal typing, nominally-typed OOP different class names informaparticularly the association of class names with be- tion implies different contracts/types and different havioral class contracts, by making a technical com- contracts/types imply different class names informaparison between nominal OO type systems and struc- tion. This identification of types with contracts, and tural OO type systems. Recently, a domain-theoretic of subtyping with inheritance of contracts, makes model of nominally-typed OOP, namely NOOP, was nominal typing and nominal subtyping closer to sealso compared to models of structurally-typed OOP. mantic typing and semantic subtyping. These comparisons provide a clear and deep account In the essay we thus stressed the practical value of for the relation between nominal and structural OO nominal typing in mainstream OOP by particularly type systems that, due to earlier lack of a domain- showing theoretic model of nominally-typed OOP, has not 1. The value of nominal subtyping, at compile time been presented before, and they should help furand at runtime, in respecting behavioral conther demonstrate, to OO PL researchers, the value tracts and thus respecting design intents, of nominal typing and nominal subtyping to mainstream OO developers and language designers, and 2. The value of the resulting identification between instill in them a deeper appreciation of it. inheritance and subtyping in providing a simpler In the essay we particularly noted that nominal conceptual model of OO software and of OO softtyping prevents types that structurally look the same ware components, leading to a simpler design from being confused as being the same type. Since process of OO software, and some objects having the same structure does not necessarily imply these objects have the same behav3. The value of making recursive types readily ior, nominal typing identifies types only if they have expressible, this being necessary for the static the same class names information (nominal informatyping of “autognostic” objects. tion) and thus only if they assert maintaining the same contract and not just assert having the same Our comparison also revealed the problem of ‘spuristructural interface. Thus, in nominally-typed OOP ous binary methods,’ so far an unrecognized problem objects having the same class type implies them as- in structurally-typed OO languages. serting they maintain the same contracts, and them Further, the recent comparison of nominal and asserting they maintain the same contracts implies structural denotational models of OOP demonstrates them having the same type. them having different views of fundamental notions of Similarly, nominal subtyping allows subtyping re- mainstream OOP, namely of objects, of type names, lations to be decided based on the refinement of con- of object/class types, of subtyping and of the relation tracts maintained by the objects, not just based on between subtyping and type inheritance. In particthe refinements of their structure. By inclusion of ular, this comparison demonstrates that an object contracts in deciding the subtyping relation, nominal in mainstream nominally-typed OOP is a record tosubtyping thus also prevents types that are superfi- gether with nominal information, that class types are cially (i.e., structurally) similar from being confused record types whose elements (i.e., objects/class inas being subtypes. Since the similarity of structure stances) additionally respect statements of class con14 tracts, and that type inheritance is correctly identi- mainstream OOP, should be based less on structural models of OOP and more on nominal ones instead. fied with nominal subtyping. Table 3 on the following page summarizes the main differences between nominal typing and structural References typing we pointed out in this essay. [1] C# language specification, version 3.0. http://msdn.microsoft.com/vcsharp, 2007. We hope the development of mathematical models of nominally-typed OOP and the comparisons presented in this essay and elsewhere are significant steps in providing a full account of the relation between nominal and structural OO type systems. We further hope this essay clearly explains the rationale behind our belief that the significant practical value and the significant semantic value of nominal typing are the reasons for industrial-strength mainstream OO software developers correctly choosing to use nominally-typed OO languages. We believe that having a clear view of the rationale behind many OO developers’ preference of nominally-typed OO languages, and having a more accurate technical and mathematical view of nominally-typed OO software, present programming languages researchers with better chances for progressing mainstream OO languages and for making PL research relevant to more OO language designers and more mainstream OO software developers.18 Finally, we believe a clearer understanding and a deeper appreciation of a key semantic advantage of nominal OO typing over structural OO typing can help remedy the existing schism between OO PL researchers on one hand and OO developers and OO language designers on the other hand, offering thereby better chances for progressing mainstream OO languages. In particular, we believe future foundational OO PL research, to further its relevance to 18 Generics, for example, add to the expressiveness of type systems of nominally-typed OO programming languages [15, 20, 29, 43, 44, 1, 17, 3, 47, 69]. As hinted to earlier, generics— particularly so-called ‘F-bounded generics’—seem to improve the support for binary methods in nominally-typed OO languages while maintaining the identification between inheritance and subtyping. We believe building a domain-theoretic model of generic nominally-typed OOP (e.g., along the lines of [12]) may offer better chances for having a deeper understanding of features of generic mainstream OO languages, such as Java erasure, variance annotations (including the notorious Java wildcards), polymorphic methods, generic type inference and so on. [2] ISO/IEC 14882:2011: Programming Languages: C++. 2011. [3] The Scala programming language. www.scalalang.org, 2014. [4] The Go programming www.golang.org, 2015. language. [5] The TIOBE index. http://www.tiobe.com/index.php/ content/paperinfo/ tpci/index.html, 2016. [6] Martin Abadi and Luca Cardelli. A semantics of object types. In Proc. LICS’94, 1994. [7] Martin Abadi and Luca Cardelli. A Theory of Objects. Springer-Verlag, 1996. [8] Moez A. AbdelGawad. NOOP: A Mathematical Model of Object-Oriented Programming. PhD thesis, Rice University, 2012. [9] Moez A. AbdelGawad. An overview of nominaltyping versus structural-typing in objectoriented programming (with code examples). Technical report, arXiv.org:1309.2348 [cs.PL], 2013. [10] Moez A. AbdelGawad. A domain-theoretic model of nominally-typed object-oriented programming. Journal of Electronic Notes in Theoretical Computer Science (ENTCS), DOI: 10.1016/j.entcs.2014.01.002., 301:3–19, 2014. [11] Moez A. AbdelGawad. A comparison of NOOP to structural domain-theoretic models of objectoriented programming. Preprint available at http://arXiv.org/abs/1603.08648, 2016. 15 Nominally-Typed OOP Object Interfaces Nominal; Include class names Contracts (as richer specifications of object behavior) Types of Objects Object Type Expressions Included (via class names) in the meaning of objects, of their interfaces and of their types Class types Structurally-Typed OOP Structural; Do not include class names Ignored in the meaning of objects, of object interfaces, and of object types Record types Class signatures Record type expressions Class signatures, as object types, are included in object meanings Class names, being associated with public class contracts, are used as type names Fixed. Cannot be rebound upon inheritance Includes inheritance of class contracts Respects contracts; respects (type) inheritance Objects do not carry their types as part of object meanings Type names are abbreviations/ synonyms for type expressions (with no inherent fixed meaning) Can be rebound upon inheritance Type Reification/Nominality Type Names Meaning of Type Names (Type) Inheritance Subtyping (Type) Inheritance versus Subtyping OO Software Design Mental Model Binary Methods Spurious and Missing Subsumption Spurious Bin. Methods Ignores behavioral class contracts Ignores behavioral class contracts One-to-one correspondence Two relations independent Simple; Inheritance hierarchy = Subtyping hierarchy Not supported. Approximations provided Complex; An inheritance hierarchy and a separate, independ’t subtyping hierarchy Fully supported (including false ones) Neither can exist Both can exist Cannot exist Can exist Special constructs needed for explicit expression Further from semantic/behavioral typing and subtyping Recursive Types Readily and naturally expressed Typing and Subtyping Closer to semantic/behavioral typing and subtyping Table 3: OO Nominal-Typing vs. OO Structural-Typing 16 [12] Moez A. AbdelGawad. Towards understand- [22] Kim Bruce, Luca Cardelli, Giuseppe Castagna, ing generics. Technical report, arXiv:1605.01480 The Hopkins Objects Group, Gary Leavens, and [cs.PL], 2016. Benjamin C. Pierce. On binary methods. Theory and Practice of Object Systems, 1994. [13] Moez A. AbdelGawad and Robert Cartwright. In nominally-typed OOP, objects are not mere [23] Kim B. Bruce. Foundations of Object-Oriented records and inheritance Is subtyping. Submitted Languages: Types and Semantics. MIT Press, for journal publication, 2016. 2002. [14] Jonathan Aldrich. The power of interoperabil[24] Luca Cardelli. A semantics of multiple inherity: Why objects are inevitable. In Proceedings itance. In Proc. of the internat. symp. on seof the 2013 ACM International Symposium on mantics of data types, volume 173, pages 51–67. New Ideas, New Paradigms, and Reflections on Springer-Verlag, 1984. Programming & Software, Onward! 2013, pages 101–116, New York, NY, USA, 2013. ACM. [25] Luca Cardelli. A semantics of multiple inheritance. Inform. and Comput., 76:138–164, 1988. [15] Joseph A. Bank, Barbara Liskov, and Andrew C. Myers. Parameterized types and Java. Technical [26] Luca Cardelli. Structural subtyping and the report, 1996. notion of power type. In ACM Proceedings of POPL, 1988. [16] Joshua Bloch. Effective java programming language guide, sun microsystems. Inc., Mountain [27] Luca Cardelli, James Donahue, Lucille GlassView, CA, 2001. man, Mick Jordan, Bill Kalsow, and Greg Nel[17] Joshua Bloch. Effective Java. Prentice Hall son. Modula-3 Report (Revised), volume 52. DigPTR, 2008. ital Systems Research Center, 1989. [18] John Boyland and Giuseppe Castagna. Parasitic [28] Robert Cartwright and Moez A. AbdelGawad. methods: An implementation of multi-methods Inheritance Is subtyping (extended abstract). for Java. In OOPSLA, 1997. In The 25th Nordic Workshop on Programming Theory (NWPT), Tallinn, Estonia, 2013. [19] G. Bracha and D. Griswold. Strongtalk: typechecking Smalltalk in a production environment. [29] Robert Cartwright and Jr. Steele, Guy L. In OOPSLA’93, pages 215–230, 1993. Compatible genericity with run-time types for [20] Gilad Bracha, Martin Odersky, David the Java programming language. In Craig Stoutamire, and Philip Wadler. Making Chambers, editor, ACM Symposium on Objectthe future safe for the past: Adding genericity Oriented Programming: Systems, Languages to the Java programming language. In Craig and Applications (OOPSLA), volume 33, pages Chambers, editor, ACM Symposium on Object201–215, Vancouver, BC, October 1998. ACM, Oriented Programming: Systems, Languages ACM SIGPLAN. and Applications (OOPSLA), volume 33, pages 183–200, Vancouver, BC, October 1998. ACM, [30] C. Chambers. Object-oriented multi-methods in Cecil. In ECOOP, 1992. ACM SIGPLAN. [21] K. Bruce, A. Schuett, R. van Gent, and A. Fiech. [31] C. Clifton, T. Millstein, G. Leavens, and PolyTOIL: A type-safe polymorphic objectC. Chambers. MultiJava: Design rationale, comoriented language. ACM Transactions on Propiler implementation and applications. ACM gramming Languages and Systems, 25(2):225– Transactions on Programming Languages and 290, 2003. Systems, 28(3):517–575, 2006. 17 [32] William R. Cook. A Denotational Semantics of Inheritance. PhD thesis, Brown Univ., 1989. [43] James Gosling, Bill Joy, Guy Steele, and Gilad Bracha. The Java Language Specification. Addison-Wesley, 2005. [33] William R. Cook. On understanding data abstraction, revisited. volume 44, pages 557–572. [44] James Gosling, Bill Joy, Guy Steele, Gilad Bracha, and Alex Buckley. The Java Language ACM, 2009. Specification. Addison-Wesley, 2014. [34] William R. Cook, Walter L. Hill, and Peter S. [45] Atsushi Igarashi, Benjamin C. Pierce, and Philip Canning. Inheritance is not subtyping. In Wadler. Featherweight Java: A minimal core POPL’90 Proceedings, 1990. calculus for Java and GJ. ACM Transactions on Programming Languages and Systems, [35] William R. Cook and Jens Palsberg. A denota23(3):396–450, May 2001. tional semantics of inheritance and its correctness. In ACM Symposium on Object-Oriented [46] Reto Kramer. Examples of Design by Contract Programming, Systems, Languages and Applicain Java. In Design and Components - Object tions (OOPSLA), pages 433–444, 1989. World - Berlin, 1999. [36] Sophia Drossopoulou, Susan Eisenbach, and Sar- [47] Angelika Langer. The Java Generics FAQ. www.angelikalanger.com/GenericsFAQ/Javafraz Khurshid. Is the java type system sound? GenericsFAQ.html, 2015. TAPOS, 5(1):3–24, 1999. [37] Mohamed Fayad and Douglas C. Schmidt. [48] X. Leroy, D. Doligez, J. Garrigue, D. Rémy, and J. Vouillon. The Objective Caml system. AvailObject-oriented application frameworks. Comable at http://caml.inria.fr/. mun. ACM, 40(10):32–38, October 1997. [49] Barbara Liskov. Keynote address-data abstrac[38] Robert Bruce Findler, Matthew Flatt, and tion and hierarchy. In ACM Sigplan Notices, volMatthias Felleisen. Semantic casts: Contracts ume 23, pages 17–34. ACM, 1987. and structural subtyping in a nominal world. In ECOOP 2004–Object-Oriented Programming, [50] Barbara H Liskov and Jeannette M Wing. A pages 365–389. Springer, 2004. behavioral notion of subtyping. ACM Transactions on Programming Languages and Systems [39] K. Fisher and J. Reppy. The design of a class (TOPLAS), 16(6):1811–1841, 1994. mechanism for Moby. In PLDI, 1999. [51] David B. MacQueen. Should ML be objectoriented? Formal Aspects of Computing, 13:214– [40] Matthew Flatt, Shriram Krishnamurthi, and 232, 2002. Matthias Felleisen. Classes and mixins. In Proceedings of the 25th ACM SIGPLAN-SIGACT [52] David B. MacQueen, Gordon D. Plotkin, and symposium on Principles of programming lanR. Sethi. An ideal model for recursive polymorguages, pages 171–183. ACM, 1998. phic types. Information and Control, 71:95–130, 1986. [41] Matthew Flatt, Shriram Krishnamurthi, and Matthias Felleisen. A programmer’s reduction [53] Boris Magnusson. Code reuse considered harmsemantics for classes and mixins. In Formal ful, 1991. syntax and semantics of Java, pages 241–269. [54] Donna Malayeri and Jonathan Aldrich. InSpringer, 1999. tegrating nominal and structural subtyping. In ECOOP 2008–Object-Oriented Programming, [42] J. Gil and I. Maman. Whiteoak: Introducing pages 260–284. Springer, 2008. structural subtyping in Java. In OOPSLA, 2008. 18 [55] Donna Malayeri and Jonathan Aldrich. Is struc- [67] Ewan Tempero, Hong Yul Yang, and James Noble. What programmers do with inheritural subtyping useful? an empirical study. In tance in java. In Proceedings of the 27th EuESOP, 2009. ropean Conference on Object-Oriented Program[56] Erik Meijer and Peter Drayton. Static typing ming, ECOOP’13, pages 577–601, Berlin, Heiwhere possible, dynamic typing when needed: delberg, 2013. Springer-Verlag. The end of the cold war between programming [68] Kresten Krab Thorup and Mads Torgersen. Unilanguages. In OOPSLA, 2004. fying genericity. In ECOOP 99–Object-Oriented Programming, pages 186–204. Springer, 1999. [57] Bertrand Meyer. Applying ‘design by contract’. Computer, 25(10):40–51, 1992. [69] Yizhou Zhang, Matthew C. Loring, Guido Salvaneschi, Barbara Liskov, and Andrew C. Myers. [58] Bertrand Meyer. Object-Oriented Software ConLightweight, flexible object-oriented generics. In struction. Prentice Hall, 1995. Proceedings of the 36th ACM SIGPLAN Conference on Programming Language Design and Im[59] R. Milner, M. Tofte, R. Harper, and D. Macplementation, PLDI 2015, pages 436–445, New Queen. The Definition of Standard ML (ReYork, NY, USA, 2015. ACM. vised). MIT Press, 1997. [60] Tobias Nipkow and David Von Oheimb. Javalight is type-safe–definitely. In Proceedings of the 25th ACM SIGPLAN-SIGACT symposium on Principles of programming languages, pages 161–170. ACM, 1998. [61] Martin Odersky. The scala language specification, v. 2.7. http://www.scala-lang.org, 2009. [62] Klaus Ostermann. Nominal and structural subtyping in component-based programming. Journal of Object Technology, 7(1):121–145, 2008. [63] Benjamin C. Pierce. Types and Programming Languages. MIT Press, 2002. [64] Benjamin C. Pierce. Types and programming languages: The next generation. LICS’03, 2003. [65] Harry H Porter III. Separating the subtype hierarchy from the inheritance of implementation. Journal of Object-Oriented Programming, 4(6):20–29, 1992. [66] Anthony J. H. Simons. The theory of classification, part 1: Perspectives on type compatibility. Journal of Object Technology, 1(1):55–61, MayJune 2002. 19	6cs.PL
arXiv:1307.8162v2 [math.AC] 24 Jan 2018 STILLMAN’S QUESTION FOR EXTERIOR ALGEBRAS AND HERZOG’S CONJECTURE ON BETTI NUMBERS OF SYZYGY MODULES JASON MCCULLOUGH Abstract. Let K be a field of characteristic 0 and consider exterior algebras of finite dimensional K-vector spaces. In this short paper we exhibit principal quadric ideals in a family whose Castelnuovo-Mumford regularity is unbounded. This negatively answers the analogue of Stillman’s Question for exterior algebras posed by I. Peeva. We show that these examples are dual to modules over polynomial rings that yield counterexamples to a conjecture of J. Herzog on the Betti numbers in the linear strand of syzygy modules. 1. Introduction Let K be a field and let S be the symmetric algebra of some finite dimensional K-vector space. Stillman [12, Problem 3.14] posed the following question: Can the projective dimension pdS (S/I) of a homogeneous ideal I be bounded purely in terms of the number and degrees of the minimal generators of I? Caviglia showed that this question was equivalent to the parallel question where one replaces projective dimension by regularity. (cf. [10, Theorem 2.4]) Ananyan-Hochster [1] recently gave a positive answer to this question in full generality. Now let E be the positively graded exterior algebra of a finite dimensional Kvector space. While resolutions over E need not be finite, the regularity regE (M ) of a finitely generated E-module is finite since E is a Koszul algebra. Here regE (M ) is defined as regE (M ) := sup{j − i\| TorE i (M, K)j 6= 0}. Irena Peeva posed the following variant of Stillman’s Question at the Joint Introductory Workshop at MSRI in the fall of 2012: Question 1.1 (Peeva). Can the regularity regE (E/I) of a homogeneous ideal I of E be bounded purely in terms of the number and degrees of the generators? Surprisingly, and contrary to the symmetric algebra case, the answer to this question is no. In Section 4 we present a family of principal quadric ideals in exterior algebras over an arbitrary field whose regularity is unbounded. Now let M be a finitely generated, graded S-module where S = K[x1 , . . . , xn ]. Let d denote the minimal degree of a generator of M . We consider the Betti numbers in the linear strand of M , that is: βilin (M ) := βi,d+i (M ), 2010 Mathematics Subject Classification. Primary: 13D02 ; Secondary: 15A75. 1 2 JASON MCCULLOUGH where βi,j (M ) = dimK TorSi (M, K)j . The length of the linear strand of M is max{i \| βilin (M ) 6= 0}. Herzog proposed the following lower bound on the Betti numbers in the linear strand of kth syzygy modules. Conjecture 1.2 (Herzog [8]). If M is a graded kth syzygy module over S with linear strand of length p, then p+k lin βi (M ) ≥ . i+k The conjecture has been proved in the following cases: (1) Herzog proved the case k = 0 in [8]. (2) Herzog was motivated by a result of Green [7] which contained the case k = 1, i = 0. See also similar results by Eisenbud-Koh [5]. (3) Reiner and Welker [13] proved the case when M is a monomial ideal and k = 1. (4) Römer [14] proved the following cases: (a) When k = 1 and 0 ≤ i ≤ p. p+k lin (b) When p > 0 and i = p − 1; i.e. βp−1 (M ) ≥ p−1+k = p + k. (c) When M is Zn -graded, Römer proved the conjecture in full generality. By modifying a recent construction of Iyengar-Walker [9], we use the BernsteinGel’fand-Gel’fand (BGG) correspondence to produce counterexamples for virtually all other cases of Herzog’s Conjecture. More precisely, in Section 3 we construct for each n ≥ 1 a finitely generated, graded S = K[x1 , . . . , x2n ]-module M that is an nth sygyzy module such that M has linear resolution of length n and has graded Betti numbers 2n 2n lin βi (M ) = − , for 0 ≤ i ≤ n. n+i n+i+2 We note that for 0 ≤ i ≤ n − 2 that 2n 2n n+n lin βi (M ) = − < , n+i n+i+2 n+i contradicting the conjecture. These examples are BGG dual to the principal quadric E-ideals mentioned above. We note that a different construction of Conca-Herbig-Iyengar [2] also gives a family of counterexamples to Conjecture 1.2, although that was not their aim. Specifically, they show [2, Theorem 5.1] that if S = K[x1 , . . . , xn , y1 , . . . , yn ] and I is the ideal I = ({xi yj , xi yj − xj yi \| 1 ≤ i, j ≤ n with i 6= j}) , then the cyclic module B = S/I has graded Betti numbers  1 for i, j = 0     2n − 2n − 2 n for 1 ≤ i ≤ n − 1 and j = i + 1 i+1 i−1 i+1 βiS (B) = 2n 2n  − for n ≤ i ≤ 2n and j = i + 2  i i+2   0 otherwise In particular, setting N = Syzn (B) gives an nth syzygy module with the same graded Betti numbers (after a degree shift) as M above. STILLMAN’S QUESTION AND HERZOG’S CONJECTURE 3 2. The BGG Correspondence We briefly recall the Bernstein-Gel’fand-Gel’fand (BGG) correspondence. We refer the reader to [3] or [4] for more details. Let K be a field of characteristic 0. Fix a positive integer n. Let S = K[x1 , . . . , x2n ] denote a standard graded polynomial ring over K. Set W = S1 and let V = HomK (W, K) denote the vectorspace dual of W . Let E = ∧V be the exterior algebra of V . Let e1 , . . . , e2n be a dual basis to x1 , . . . , x2n . Thus E = Khe1 , . . . , e2n i, which we also view as a positively graded ring with deg(ei ) = 1 to keep the notation in this paper consistent. (In [3] and [4], E is viewed as a negatively graded ring.) Let L denote the functor from the category of graded E-modules to the category of linear free complexes of S-modules defined as follows: Given any graded Emodule P , we define L(P ) to be S ⊗K P viewed as a complex of graded free S-modules, where L(P )i = S ⊗K P−i and the differential is induced by L(P ) : · · · → S ⊗K P−i s⊗p d →i P S ⊗K P1−i → · · · 7→ sxj ⊗ ej p We extend L to a functor on complexes of E-modules by taking the total complex of the resulting double complex. There is an adjoint function R from complexes of S-modules to complexes of E-modules, defined in a similar way, which creates an equivalence of bounded derived categories of graded S-modules and that of graded E-modules. An important property of the equivalence for the purpose of this paper is that under this equivalence the functor L identifies finitely generated E-modules with linear free S-complexes. 3. Counterexamples to Herzog’s Conjecture First we recall a well-known fact about exterior algebras: Theorem 3.1 (cf. [11, Theorem 5.2]). Let f ∈ E2 = ∧2 V be a generic element. Then the linear transformation f· ∧r V → ∧r+2 V is injective for r ≤ n − 1 and is surjective for r ≥ n − 1. As remarked in [9], it suffices to pick f = e1 e2 + · · · + e2n−1 e2n . We now come to the first main result. By convention, we set m k = 0 if k < 0 or if k > m. Theorem 3.2. Fix n > 0 and a field K of characteristic 0. Let S = K[x1 , . . . , x2n ] be a standard graded polynomial ring. There exists a finitely generated, graded Smodule M , generated by degree 0 elements such that reg(M ) = 1 and the graded Betti numbers are  2n 2n  if i = j and 0 ≤ i ≤ n  i+1 − i−1 2n 2n βi,j (M ) = if j = i + 1 and n + 1 ≤ i ≤ 2n i − i+2   0 otherwise. 2n 2n In particular, if we set N = Syzn (M ), then βilin (N ) = n+i − n+i+2 , contradicting Conjecture 1.2 for 0 ≤ i ≤ n − 2. Proof. Let f ∈ E2 be a general element. Consider the graded complex of free E-modules f G• : E(2 − 2n) −→ E(−2n). 4 JASON MCCULLOUGH Let F• = L(G• ). Since L(E) is the usual Koszul complex K• (x, S)[2n] with a shift in homological degree, we have F• = cone(K• (x, S)[2] → K• (x, S)). The complex F• fits in a short exact sequence 0 → K• (x, S)[1] → F• → K• (x, S) → 0. Since H0 (K• (x, S)) ∼ = K and Hi (K• (x, S)) = 0 for all i > 0, if follows that ( K for i = 0, −1 Hi (F• ) = 0 otherwise. By Theorem 3.1, every possible cancelation in F• that can occur does occur. Thus the minimal subcomplex F• ′ of F• has graded Betti table -1 0: 1 1: - 0 2n - 1 2n 2 ··· − 1 ··· - n-1 2n − n−2 - 2n n n 2n − n ··· 2n · · · n+2 2n-1 2n 2n 1 Note that under the BGG correspondence, the first linear strand corresponds to L(AnnE (f )) and the second linear strand is L(E/(f )). In particular, Since Hi (F′• ) = 0 for i > 0, if we set F′≥0 to be the truncation of F′• , then F′≥0 is the minimal free resolution of an S-module M . (We shift graded degrees to ensure that M is generated in degree 0.) In particular, M now has graded Betti table of the form 0 ··· n · · · 2n-1 2n 1 n-1 2n 2n 0: 2n 2n − 1 · · · − 2 n n−2 2n 2n ··· 2n 1 1: n − n+2 Setting N = Syzn (M ), we see that N is an nth syzygy with linear free resolution and graded Betti numbers as prescribed above. 4. Peeva-Stillman Question for Exterior Algebras Now let E be the exterior algebra of a finite dimensional K-vector space. Here we set the exterior variables to have degree +1 rather than −1, as in the previous section. While resolutions over E need not be finite, since E is a Koszul algebra, the regularity regE (M ) of a finitely generated E-module is finite. Here regE (M ) is defined as regE (M ) := sup{j − i\| TorE i (M, K)j 6= 0}. Irena Peeva posed the following variant of Stillman’s Question at the Joint Introductory Workshop at MSRI in the Fall of 2012: The following theorem gives the promised negative answer to Question 1.1. Theorem 4.1. Let K be a field of characteristic 0. Fix any positive integer n. Let V be V the K-vector space with basis e1 , e2 , . . . , e2n and consider the exterior algebra E = V . Set f = e1 e2 + e3 e4 + · · · + e2n−1 e2n . Then regE (E/(f )) = n. Proof. Consider the minimal free resolution of E/(f ) over E. Note that we may identify Syz2 (E/(f )) with AnnE (f ). By Theorem 3.1, AnnE (f ) has no elements of degree < n. Thus regE (E/(f )) ≥ n. On the other hand, by the proof of STILLMAN’S QUESTION AND HERZOG’S CONJECTURE 5 Theorem 3.2, we see that L(E/(f )) has a linear free S-resolution. By the Dictionary Theorem [3, Theorem 7.7], HomK (E/(f ), K) ∼ = AnnE (f ) has a linear free E-resolution. Therefore regE (E/(f )) = n. Remark 4.2. The previous theorem also holds over any field of positive characteristic p > 0, but it is no longer true that Ann(f ) has a linear free resolution since in particular f p = 0. 5. An Example Finally we show how one can construct nth syzygy modules with linear free resolutions of length n and with graded Betti numbers smaller than those predicted by Herzog’s Conjecture in Macaulay2. Here we make use of the BGG package written by Abo, Decker, Eisenbud, Schreyer, Smith and Stillman. Here we demonstrate the case n = 4. i1 : n = 4; i2 : E = QQ[e_1..e_(2n),SkewCommutative=>true]; -- an exterior algebra in 2n variables i3 : S = QQ[x_1..x_(2n)]; -- a polynomial ring in 2n variables i4 : I = ideal sum for i from 1 to n list e_(2i-1)e_(2i) o4 = ideal(e e + e e + e e + e e ) 1 2 3 4 5 6 7 8 o4 : Ideal of E i5 : betti res I -- The resolution over E of E/I. reg(E/I) = n o5 = total: 0: 1: 2: 3: 4: 0 1 1 . . . . 1 2 3 4 5 6 7 8 9 1 42 288 1155 3520 9009 20384 42042 80640 . . . . . . . . . 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 288 1155 3520 9009 20384 42042 80640 o5 : BettiTally i6 : J = ann I; o6 : Ideal of E i7 : P = J / ideal product flatten entries vars E; 6 JASON MCCULLOUGH i8 : loadPackage "BGG" o8 = BGG o8 : Package i9 : F = bgg(2n-2,P,S); 8 27 o9 : Matrix S <--- S i10 : M = coker F ** S^{-(2n-1)}; -- M is an S-module i11 : betti res M -- the linear strand of the resolution of M is L(J/(e_1...e_2n)) 0 1 2 3 4 5 6 o11 = total: 8 27 48 42 42 48 27 0: 8 27 48 42 . . . 1: . . . . 42 48 27 7 8 . 8 8 1 . 1 o11 : BettiTally i12 : G = bgg(n-1,comodule I,S); 42 o12 : Matrix S 48 <--- S i13 : N = coker G * S^{-n}; -- N is the nth syzygy module of M, with a shift in grading i14 : betti res N -- its resolution is L(E/I) under the BGG correspondence 0 1 2 3 4 o14 = total: 42 48 27 8 1 0: 42 48 27 8 1 o14 : BettiTally It is clear that this is not the most extreme counterexample to Herzog’s Conjecture. Note that the module N above has rank 14 and is a 4th syzygy module over S. By a Theorem of Bruns (cf. [6, Corollary 3.13]), there exists a free submodule F ⊂ N of rank 10 such N/F is a still a 4th syzygy module. The module N/F would then have Betti table 0 1 2 3 4 0: 32 48 27 8 1 STILLMAN’S QUESTION AND HERZOG’S CONJECTURE 7 It would be interesting to know in general what the minimal Betti numbers are for a nth syzygy module with linear free resolution. Acknowledgements The author thanks Srikanth Iyengar, Irena Peeva, and Mark Walker for useful conversations concerning this paper and also Jerzy Weyman for pointing out reference [11]. References 1. Tigran Ananyan and Melvin Hochster, Small subalgebras of polynomial rings and stillman’s conjecture, arXiv: 1610.09268, 2017. 2. Aldo Conca, Hans-Christian Herbig, and Srikanth Iyengar, Koszul property for the moment map of some classical representations, arXiv: 1705.02688, 2017. 3. David Eisenbud, The geometry of syzygies, Graduate Texts in Mathematics, vol. 229, SpringerVerlag, New York, 2005, A second course in commutative algebra and algebraic geometry. 4. David Eisenbud, Gunnar Fløystad, and Frank-Olaf Schreyer, Sheaf cohomology and free resolutions over exterior algebras, Trans. Amer. Math. Soc. 355 (2003), no. 11, 4397–4426. MR 1990756 5. David Eisenbud and Jee Koh, Some linear syzygy conjectures, Adv. Math. 90 (1991), no. 1, 47–76. 6. E. Graham Evans and Phillip Griffith, Syzygies, London Mathematical Society Lecture Note Series, vol. 106, Cambridge University Press, Cambridge, 1985. 7. Mark L. Green, Koszul cohomology and the geometry of projective varieties, J. Differential Geom. 19 (1984), no. 1, 125–171. 8. Jürgen Herzog, The linear strand of a graded free resolution, unpublished notes, 1998. 9. Srikanth Iyengar and Mark Walker, Examples of finite free complexes of small rank and homology, arXiv: 1706.02156, 2017. 10. Jason McCullough and Alexandra Seceleanu, Bounding projective dimension, Commutative Algebra. Expository Papers Dedicated to David Eisenbud on the Occasion of His 65th Birthday (I. Peeva, Ed.), Springer-Verlag London Ltd., London, 2013. 11. Guillermo Moreno-Socı́as and Jan Snellman, Some conjectures about the Hilbert series of generic ideals in the exterior algebra, Homology Homotopy Appl. 4 (2002), no. 2, part 2, 409–426, The Roos Festschrift volume, 2. 12. Irena Peeva and Mike Stillman, Open problems on syzygies and Hilbert functions, J. Commut. Algebra 1 (2009), no. 1, 159–195. 13. V. Reiner and V. Welker, Linear syzygies of Stanley-Reisner ideals, Math. Scand. 89 (2001), no. 1, 117–132. 14. Tim Römer, Bounds for Betti numbers, J. Algebra 249 (2002), no. 1, 20–37. E-mail address: jmccullo@iastate.edu Iowa State University, Department of Mathematics, Ames, IA 50011	0math.AC
Recruitment Market Trend Analysis with Sequential Latent Variable Models Chen Zhu1 , Hengshu Zhu1,2 , Hui Xiong3∗, Pengliang Ding1 , Fang Xie1 1 Baidu Inc., 2 Baidu Research-Big Data Lab, 3 Rutgers University, {zhuchen02, zhuhengshu, dingpengliang, xiefang}@baidu.com, hxiong@rutgers.edu arXiv:1712.02975v1 [cs.AI] 8 Dec 2017 ABSTRACT Recruitment market analysis provides valuable understanding of industry-specific economic growth and plays an important role for both employers and job seekers. With the rapid development of online recruitment services, massive recruitment data have been accumulated and enable a new paradigm for recruitment market analysis. However, traditional methods for recruitment market analysis largely rely on the knowledge of domain experts and classic statistical models, which are usually too general to model large-scale dynamic recruitment data, and have difficulties to capture the fine-grained market trends. To this end, in this paper, we propose a new research paradigm for recruitment market analysis by leveraging unsupervised learning techniques for automatically discovering recruitment market trends based on large-scale recruitment data. Specifically, we develop a novel sequential latent variable model, named MTLVM, which is designed for capturing the sequential dependencies of corporate recruitment states and is able to automatically learn the latent recruitment topics within a Bayesian generative framework. In particular, to capture the variability of recruitment topics over time, we design hierarchical dirichlet processes for MTLVM. These processes allow to dynamically generate the evolving recruitment topics. Finally, we implement a prototype system to empirically evaluate our approach based on real-world recruitment data in China. Indeed, by visualizing the results from MTLVM, we can successfully reveal many interesting findings, such as the popularity of LBS related jobs reached the peak in the 2nd half of 2014, and decreased in 2015. Keywords Trend Analysis, Recruitment Market, Latent Variable Model 1. INTRODUCTION The scarcity of skilled talents has stimulated the global recruitment industry in the past few years. An article from ∗ Corresponding Author. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. KDD ’16, August 13-17, 2016, San Francisco, CA, USA c 2016 ACM. ISBN 978-1-4503-4232-2/16/08. . . $15.00 DOI: http://dx.doi.org/10.1145/2939672.2939689 Forbes reported that, US corporations spend nearly $ 72 billion each year on a variety of recruiting services, and the worldwide number is likely three times bigger [4]. Along this line, a growing challenge is to provide an effective trend analysis of recruitment market, such as forecasting recruitment demand and predicting market status. Both employers and job seekers can benefit from the study of recruitment market trends. Moreover, at the macro level, this analysis can also provide valuable understanding of industry-specific economic growth for business analysts. With the rapid development of online recruitment services, such as Linkedin [7], Dice [3], and Lagou [6], massive recruitment posting data have been accumulated. For example, as of the end of 2015, there are more than 1.3 million job positions from 100K+ companies across more than 20 Internet-related industries, such as mobile Internet, online-to-offline (O2O), and cloud computing, available at Lagou [6]-a Chinese tech hiring service website. These huge data enable a new paradigm for studying recruitment market trends in a holistic and fine-grained manner. Recruitment market analysis is a classic topic in human capital economics, where recruitment market is either treated as a factor of macro economic phenomenons or the analysis is focused on advising people to make the best job decisions in a general economic framework [26, 27, 29]. However, previous studies rely largely on the knowledge of domain experts and classic statistical models, and thus usually too general to capture the high variability of recruitment market (e.g., the evolution of recruitment topics). Also, these studies have limited efforts on understanding the fine-grained market trends, such as forecasting the recruitment situation for a specific company in the next few months. Therefore, it is very appealing to design a new research paradigm for recruitment market analysis through large-scale analysis on massive recruitment data. Along this line, there are some major challenges. First, how to model the intrinsically sequential dependency of recruitment states (e.g., Hiring Freeze) forming the market trend? Second, how to model the semantic relationship between the market trend and job postings from different companies? Finally, how to model the variability of recruitment postings for a long time period? To tackle these challenges, we propose an unsupervised learning approach for recruitment market trend analysis, which can automatically discern the underlying trend of recruitment market. First, we develop a novel sequential market trend latent variable model (MTLVM), which is designed for capturing the temporal dependencies of recruitment states of companies and is able to automatically learn Figure 1: The trend of the number of job postings related to different skill requirements over two years, which indicates the demands of recruitment market. the latent recruitment topics and demands from recruitment data within a Bayesian generative framework. To be more specific, we assume that the current recruitment state of a specific company is influenced by its state in previous epoch, and it will impel the company to review appropriate recruitment demands (e.g., R&D recruitment). Then, different recruitment demands will generate different recruitment topics (e.g., experienced algorithm engineer), which will finally generate the recruitment postings. In particular, to capture the variability of recruitment topics over time, we design a hierarchical dirichlet processes for MTLVM, which can dynamically generate recruitment topics. Finally, we implement an intelligent prototype system to empirically evaluate our approach based on a real-world recruitment data set collected from China for the time period from 2014 to 2015. Indeed, by visualizing the results from MTLVM, we can successfully observe many interesting discoveries, such as the popularity of LBS related jobs reaches the peak in the 2nd half of 2014, and decreases in 2015. Generally, the contributions of this paper can be summarized as follows. • To the best of our knowledge, this paper is the first attempt to leverage unsupervised learning approach for automatically modeling the trend of recruit market. This work provides a new research paradigm for recruitment market analysis. • We propose a sequential latent variable model, named MTLVM, for learning the latent recruitment states, demands, and topics simultaneously. Particularly, MTLVM can dynamically generate recruitment topics by integrating hierarchical dirichlet processes. • We develop a prototype system for empirically evaluate our approach. Indeed, by visualizing the results obtained from MTLVM, we can successfully observe many interesting and useful findings. 2. OVERVIEW In this section, we first introduce some preliminaries of recruitment market modeling, and then formally present the overview of our model MTLVM. (a) Jan, 2014-Jun, 2014 (b) Jul, 2014-Dec, 2014 (c) Jan, 2015-Jun, 2015 (d) Jul, 2015-Nov, 2015 Figure 2: The word cloud representation of job postings related to “Mobile Software Engineer” with respect to different epochs in our data set, where the size of each keyword is proportional to its frequency. 2.1 Preliminaries Recent years have witnessed the rapid development of online recruitment services, which have already become the most important venue of talent seeking, especially for hightech companies. Therefore, the job posting data from these services can help researchers better understand the trend of recruitment market from the perspectives of not only an individual company but also the whole industry. Intuitively, different job postings can indicate different recruitment demands of companies, such as R&D related positions, which usually change over different epochs. For example, Figure 1 demonstrates the trend of the number of job postings related to different skill requirements during January 2014 to November 2015 based on our real-world data set. We can observe that skill “Information Retrieval” becomes less popular, compared with other skills, such as “Data Mining”, and “Machine Learning”. Meanwhile, “Spark”, an emerging technique for big data processing, has attracted more and more attention during these years. Indeed, such evolution of recruitment demands is inherently determined by the change of latent recruitment states of companies at different epochs, which have strong sequential dependency. For example, Alibaba, one of the largest E-Commerce companies in China, hugely enlarged the recruitment in 2014, which is followed by the recruitment state “Hiring Freeze” in 2015. As a result, its recruitment demands related to ECommerce largely shrank in 2015. To capture the change of recruitment states, and model the semantic relationship between recruitment demand and state, our model MTLVM follows the Beysian latent variable model with first-order Markov assumption, where current state is only determined by the state at previous epoch. Indeed, by further analyzing the descriptions of job postings, we observe that the detail of similar recruitment demands (e.g., Recruiting Mobile Software Engineer) will be influenced by the corresponding recruitment states. Thus the corresponding demands usually have high variability, and generate different recruitment topics. For example, α ρc & ce,t γ1 Gc De,t,m γ2 & H Q0 ϕe,t,m,n Χe,t,m,n γ0 0HW of unique recruitment states is C. Specifically, at the first epoch, the corresponding ce,1 for all companies is sampled from a uniform distribution. For the following epochs, if company e releases some job postings at previous epoch t − 1, namely that χe,t−1 exists, its current recruitment state ce,t is sampled from a multinomial distribution determined by previous state ce,t−1 . Otherwise, ce,t is drawn from the overall recruitment state at epoch t − 1, i.e., P (ce,t ) = PC c=1 P (ce,t \|ct−1 = c)P (ct−1 = c), where P (ct−1 = c) is the average P (ce0 ,t−1 ) of company e0 who has market state c at epoch t − 1. In addition, we name the chains consisting of neighbouring χe,t−1 that belong to the same company as a data chain. Therefore, if a company only occasionally releases jobs, it may have more than one data chain according to our formulation. Furthermore, we define the generative process of job posting χe,t,m ∈ χe,t of company e at epoch t as follows. First, a recruitment demand De,t,m is generated from the latent factor Gc which is sampled from a Dirichlet Process and determined by the current recruitment state ce,t . Then, we sample a recruitment topic φe,t,m,n from the demand De,t,m for each observation χe,t,m,n . Finally, each observation χe,t,m,n is generated from a multinomial distribution determined by corresponding topic φe,t,m,n . Specifically, Figure 3 shows the graphical representation of MTLVM. Figure 3: The graphical representation of MTLVM. 3. Figure 2 shows the word cloud representation 1 of job postings related to “Mobile Software Engineer” with respect to different epochs. We can observe that, “Mobile Game Development” is a hot topic in the second half of 2014, while “Android based Web Technology” becomes popular in the first half 2015. To model the semantic relationships among recruitment states, recruitment demands, and recruitment topics, our model, MTLVM, follows the idea of Hierarchical Dirichlet Processes, an infinity version of topic modeling, to model the job postings. Therefore the topic number can be automatically determined. In this section, we will introduce the technical details of our model MTLVM. And we illustrate the important mathematical notations in Table 1. 2.2 3.1 MODELING THE TREND OF RECRUITMENT MARKET Model Inference According to the introduction in Section 2, we can summarize the parameterizations of MTLVM as follows, ρc \|α ∼ Diri(α), ce,t \|{ρc }C c=1 , ce,t−1 Overview of MTLVM Formally, we regard the m-th job posting of company e at epoch t as a bag of words χe,t,m = {χe,t,m,n }n , where χe,t,m,n is the basic observation in job postings (e.g., keywords in job description), and \|χe,t,m \| = Ne,t,m . For modeling the trend of recruitment market, we first divide all job postings into different data units {χe,t }e,t with respect to companies and timestamps, which contain job postings of company e at epoch t. As introduced above, we assume that the current recruitment state of a specific company is influenced by its state in previous epoch, and it will impel the company to review appropriate recruitment demands. Then, different recruitment demands will generate different recruitment topics, which will finally generate the recruitment postings. Therefore, we define a parameter ce,t to represent the recruitment state of company e at epoch t, where the number 1 All the words are originally in Chinese, and are automatically translated by a commercial translation tool [1]. ∼ Multi(ρce,t−1 ), Q0 \|H, γ0 ∼ DP(γ0 , H), Gc \|Q0 , γ1 ∼ DP(γ1 , Q0 ), De,t,m \|{Gc }C c=1 , γ2 , ce,t ∼ DP(γ2 , Gce,t ), φe,t,m,n \|Ge,t,m ∼ De,t,m , χe,t,m,n \|φe,t,m,n ∼ Multi(φe,t,m,n ). Following the above parameterizations, we can get the joint probability distribution of χ and c as P (χ, c\|Λ) = Y P (ce,t \|ce,t−1 , Λ)P (χe,t \|Λ, ce,t ) , (1) e,t where Λ is a set of hyper-parameters, including α, H, γ0 , γ1 , γ2 , and c0 . Specifically, c0 is a default initial recruitment state and ρc0 is fixed to (1/C, ..., 1/C). Indeed, the above equation can be divided into two parts, i.e., P (ce,t \|ce,t−1 , Λ) that follows a multinomial distribution Multi(ρce,t−1 ), and P (χe,t \|Λ, ce,t ) that follows Dirichlet Processes, respectively. Specifically, P (χe,t \|Λ, ce,t ) can be computed by Table 1: Mathematical Notations. Me,t P (χe,t \|Λ, ce,t ) = Y P (De,t,m \|Λ, Gce,t ) × Symbol Description χe,t,m,n The n-th tokens at m-th job posting of company e at epoch t. The m-th job posting of company e at epoch t. The observation unit containing job postings of company e at epoch t. The entire data set of job postings. The recruitment state of company e at epoch t. The hyperparameter of the Dirichlet prior on ρc . The transition matrix of recruitment state c. The probability measure representing the recruitment strategy of state c. The probability measure representing the recruitment demand of χe,t,m . The recruitment topic of the n-th tokens at m-th job posting of company e at epoch t. The base measure for Dirichlet Process generating Gc . The concentration parameter for Dirichlet Process generating Q0 . The concentration parameter for Dirichlet Process generating Gc . The concentration parameter for Dirichlet Process generating De,t,m . The base measure for Dirichlet Process generating Q0 . A set of hyperparameters, including α, H, γ0 , γ1 , γ2 , and c0 . The default initial recruitment state. The number of tokens at m-th job posting of company e at epoch t. The number of job postings of company e at epoch t. The number of unique recruitment states. m=1 χe,t,m χe,t Ne,t,m Y P (φe,t,m,n \|Λ, De,t,m )P (χe,t,m,n \|φe,t,m,n ) . (2) n=1 χ ce,t α ρ Gc Therefore, the objective of learning MTLVM is to find a set of optimal parameters in P (ce,t \|ce,t−1 , Λ), P (De,t,m \|Λ, Gce,t ), P (φe,t,m,n \|Λ, De,t,m ), and P (χe,t,m,n \|φe,t,m,n ), which can maximize the probability of Equation 1. In this paper, we propose a two-step framework to learn our model by a Gibbs Sampling method. In the first step, we introduce how to optimize the transition matrix ρ, which is constituted of P (ce,t \|ce,t−1 , Λ), given P (χe,t \|Λ, ce,t ). Depending on equation 1, we could get the conditional distribution of ce,t P (ce,t = c\|c−e,t , χ, Λ) = P (ce,t = c, χe,t \|χ−e,t , Λ) P (χe,t \|c−e,t , χ−e,t , Λ) −e,t ∝ P (χe,t \|ce,t = c, Λ)P (ce,t = c\|c De,t,m φe,t,m,n Q0 γ0 γ1 (3) γ2 , Λ). H Since P (χe,t \|ce,t = c, Λ) is given, the only challenge is to calculate P (ce,t = c\|c−e,t , Λ). Here we follow the inference in [16] and give it directly. −e,t P (ce,t = c\|c −e,t +α e,t−1 ,c) −e,t +Cα q(c e,t−1 ) q(c where −e,t q(c) −e,t +I(ct−1 =c=ct+1 )+α e,t+1 ) −e,t q(c) +I(ct−1 =c)+Cα Me,t , (4) C means the number of recruitment states c ap- The Application of MTLVM After learning stage, MTLVM can be used for predicting the future trend of recruitment market, e.g., recruitment states, demands, topics, and basic observations. Specifically, given a company e, we can estimate its current recruitment state ce,t by ce,t = arg max P (c\|χe,t , ce,t−1 , ρ, Λ), c c0 Ne,t,m , χ, Λ) = P (χe,t \|ce,t = c, Λ) · q(c,c −e,t pearing except ce,t , and q(c means the number of pair e,t−1 ,c) ce,t−1 , c appearing except ce,t . In the second step, we introduce how to compute the parameters related to Dirichlet Process in Equation 2. Indeed, this task can be regarded as an analog of the Chinese Restaurant Process (CRP) [8], and the metaphor can be explained as follows. We have C cuisine styles (i.e., recruitment state) and a franchise (i.e., company) with Me,t restaurants (i.e., job postings). Everyday, the franchise will change its cuisine style according to the cuisine styles on last day. In Particular, the menus of different restaurants may be different, even if they share the same cuisine style. At each table of each restaurant, the dish (i.e., topic) is determined by the first customer (i.e., the basic observation in job postings) sitting there, and it is shared among all customers who sit at that table. When a new customer enters the restaurant, she can sit at an occupied table or a new table. If she chooses a new table, she can order a new dish from the menu. According to the above metaphor of CRP, we can easily obtained a Gibbs sampling scheme for posterior sampling given χ [28]. The detailed definition and inference can be found in Appendix. 3.2 Λ (5) where ρ is the transition matrix, and P (ce,t \|χe,t , ce,t−1 , ρ, Λ) ∝ P (ce,t , χe,t \|ce,t−1 , ρ, Λ) = P (χe,t \|ce,t , Λ)P (ce,t \|ce,t−1 , ρ). (6) Therefore, the probability of recruitment state at the next epoch ce,t+1 can be obtained by P (ce,t+1 ) = Multi(ρce,t ). Furthermore, the recruitment topics can be obtained in the same way introduced in Section 3.1. Thus, the probability of a basic observation χe,t+1,k (e.g., keywords in job description) from company e appearing at epoch t + 1 can be computed by P (χe,t+1,k \|Λ) = C X P (χe,t+1,k \|ce,t+1 , Λ)P (ce,t+1 ), (7) ce,t+1 =1 where P (χe,t+1,k \|ce,t+1 ) can be obtained by Equation 2. 4. EXPERIMENTS In this section, we will study the performance of our model MTLVM on a huge data set collected from a major online recruitment website in China. Furthermore, we have developed a web-based prototype system to empirically evaluate our model. This system can visualize the results of our model, provide in-depth analysis of recruitment market analysis, and help people understand the high variability of recruitment market. Figure 4 shows a screenshot of this prototype system. In this system, we show the trend analysis of the entire recruitment market and the detail evolution of recruitment state of companies. Table 3: Average VM/CM comparison. MTLVM (C=5) MTLVM (C=10) MTLVM (C=20) LDA K VM CM 105 100 118 100 0.464 0.682 0.688 0.637 5.044 6.783 6.279 6.722 learning recruitment states and recruitment topics). Specially, following [28], we set a symmetric Dirichlet distribution with parameters of 0.5 for the prior H over topic distributions. For simplicity, γ0 , γ1 , and γ2 are set to 1 directly, and another hyperparameter α is also be set to 1 empirically. 4.2 Figure 4: A screenshot of our system for recruitment market analysis. All of following visualization results in this section can be obtained by this prototype system. Table 2: The statistics of our real-world data set. Raw Data Filtered Data # # # # 4.1 job postings unique companies data units data chains 934,761 68,302 191,549 116,392 257,166 997 13,209 2,557 Data Set and Experimental Setup The data set used in our experiments is collected from a major online recruitment website in China and contains 934,761 job postings from 68,302 companies released from January 2014 to November 2015. Specially, Figure 5(a) to 5(d) demonstrate some statistics of our data set. As mentioned above, “data unit” in Figure 5(c) means a job posting set χe,t released by company e at epoch t and “data chain” means chains consisting of neighbouring χe,t−1 that belong to the same company. From these statistics we can observe that most of companies only randomly release very few job postings, and therefore cannot represent the trend of recruitment market. To avoid such bias, we only conserve companies which have released more than 100 job postings. Table 2 shows the detailed statistics of our raw data set and the filtered data set. By the above pre-process, we filtered about 72% original data; and the number of companies declined by 98.6%. However, the average number of job postings per company increases from 13.7 to 258, and the average length of chain also increases from 1.65 to 5.16, which make it more reasonable for training MTLVM. In particular, in each job posting, the keywords in job description (e.g., job responsibility, skill requirements) are treated as basic observations, and all stop words are removed to guarantee the modeling performance. Note that, our model is trained with original Chinese words, and all experimental results were translated into English by a commercial translation tool [1] for facilitating demonstration. In the following subsections, we will comprehensively study the performance of MTLVM in term of trend analysis (e.g., Evaluation of Recruitment Topics How to quantitatively evaluate the performance of latent variable models is always an open problem. Although perplexity and held-out likelihood are common measures for evaluating prediction results, they cannot demonstrate how coherent and meaningful the latent factors (e.g., recruitment states and topics) are. Therefore, in this paper, we follow the measures introduced in [31], which is inspired by [30, 13], for evaluating MTLVM. Specifically, we picked up top 10 keywords for each learned recruitment topic and asked 4 senior experts of human resource to evaluate its value. These experts are first required to judge whether this topic is valuable. If a topic is valuable, they need continue to judge how many words are relevant in the top 10 keyword list. Based on these manually labeled results, the metrics Validity Measure (VM) and Coherence Measure (CM) are defined as VM = # of relevant words # of valid topics , CM = . # of topics # of words in valid topics Besides, to evaluate how the number of recruitment states affects the performance, we train MTLVM with different settings of state number, i.e., C = 5, 10, 20 respectively. Furthermore, we select widely used topic model Latent Dirichlet Allocation (LDA) [12] as baseline. After convergence, the numbers of topic in all our models, i.e., K, are automatically determined as about 100. Therefore, we set K as 100 for LDA. Table 3 shows the average results of VM and CM. We can observe that in terms of VM, both of MTLVM (C=20) and MTLVM (C=10) outperform LDA a lot, and MTLVM (C=20) has the best performance. In terms of CM, the performance of MTLVM (C=10) is the best, while that of MTLVM (C=20) is worse than LDA. It may be because that too many states will make the model relatively sparse, and thus will make relevant words scattered in different topics. In particular, the performance of MTLVM (C=5) is the worst, which may be because that few states cannot accurately describe the market trend well. Overall, since MTLVM (C=10) has the most balanced results, we set state number C = 10 in all of following experiments. 4.3 Evaluation of Recruitment States Here, we will empirically evaluate the learned recruitment states from several aspects. Figure 6 shows the trend of popularity of recruitment states discovered by MTLVM over time. It is obvious that (a) (b) (c) (d) Figure 5: The distribution of (a) the number of companies that release job posting in different epochs, (b) the number of companies with respect to the number of their job postings, (c) the number of data units with respect to the number of contained job postings, and (d) the number of data chains with respect to their length. Table 4: The probabilities of top recruitment topics of selected recruitment states (e.g., #1, #3, #4, #9). The corresponding word cloud representations of these topics are shown in figure 8. state #3 state #4 state #5 state #9 top top top top 4.3.1 Figure 6: The trend of popularity of recruitment states discovered by our model over time. Figure 7: The transition matrix of recruitment states, where the i, j element means the transition probability from i-th state to j-th state, and deeper color means higher probability. these states change over time dramatically. Specifically, state #3 kepng a relative high popularity over the long period, while popularity of #6 is always low. Meanwhile, the popularity of state #9 is rising dramatically since February 2015. Several other states, such as state #2, #7, and #10, represent totally opposite trends. Furthermore, Figure 7 shows the transition matrix of recruitment states, where the element of i-th row and j-th column represents the transition probability from state i to state j. We can observe that, all states have the highest transition probabilities to themselves, which is due to the momentum of recruitment market. Also, the color of 3-th, and 9-th columns is relatively deeper, which indicates the importance of states #3, and #9. All of above results show that our model MTLVM has the ability to capture the high variability of recruitment market by discovering these latent recruitment states. #1 #2 #3 #4 0.22493 0.12399 0.10064 0.08932 0.22172 0.17637 0.11360 0.08725 0.18185 0.10126 0.07735 0.07242 0.19441 0.12350 0.10018 0.08021 Recruitment State Inspection Here, we will test whether the recruitment states discovered by our model are comprehensible. To solve this problem, we select 4 representative recruitment states according to above analysis and show their top 4 topics in figure 8 by word cloud representations, where the larger words have higher generative probabilities. Meanwhile, Table 4 shows the corresponding generative probabilities of these topics. We can find that the topic about “programming” always has very high probability in every state. In particular, the top #1 topics in state #1, state #3, and state #9 are the same. That means the demands for R&D related positions are always exuberant, since the high-probability words, “linux” and “mysql”, directly indicate the fundamental skill requirements of R&D. Actually, the salary of software engineer has kept rising for a long time, which can support this discovery of our model. These states also show that the work about games and front-end are also very popular, which is consistent with our real-world observations. Next, we •further illustrated in top figure Stateinspect #3. Itstates is obvious that the #18. topic is about R&D. The top #3 topic, containing “data”, “analysis”, “research”, and “algorithm”, indicates the demand of recruiting senior researchers or algorithmic engineers. The top #4 topic may be about propaganda (e.g., public relationship), because it contains several paperwork and advertising related words, such as “compose”, “edit”, “WeChat” and “blog”(e.g., social network based advertising). Besides, what makes this state different is the top #2 topic, which contains “human”, “resource”, “office”, and “assist”. All of them are about administrative management obviously. Actually, in our model, this topic only exists in this state. Since research development, propagandism, and administrative management are essential to any companies, we can conclude that this state covers the fundamental talent demands for most by high-tech companies. And (a) state #3 (b) state #4 (c) state #5 (d) state #9 Figure 8: The word cloud representations of top 4 topics of selected recruitment states (e.g., #3, #4, #5, #9), where the size of each keyword is proportional to its generative provability. The generative probabilities of these recruitment topics are shown in table 4. figure 6 also indicates state #3 is the most popular. • State #4. This state is closely related to state #5. We can find that the top #2, #3, and #4 topics are relative normal. And the top #1 topic, which contains “data”, “programming”, and “algorithm”, is apparently about R&D, too. However, what makes this topic different is word “lbs”. Location-based Service (LBS), such as map navigation or map localization, is the cornerstone of O2O, which is a kind of business model that uses online and mobile to drive offline local sales and become more and more popular since 2014. In figure 6, we notice that the popularity of this topic increased in 2014 and declined in 2015. This may be because the industry doorsill of this field is related high, so only a few large companies have capability to get into this field. Actually, only the largest IT companies in China, such as Baidu, Tencent, and sogou, provide such service now. • State #5. This is a very Chinese-style recruitment state. In top #4 topic, we find word “O2O” as well as “merchant” and “business”. Actually, with the proposal of “Internet Plus” [5], thousands of start-up companies, focusing on O2O, sprung up across the country. Meanwhile, the others topics in this state are related normal in terms of technology. The top #1 and #3 topics are just about some popular programming language, web-framework, and database. It may indicate that “O2O” is just a business concept rather than a technology concept. • State #9. As shown in figure 6, this state exploded since February 2015. In the word cloud representations of its topics, the top #4 is related meaningful. These high frequency words (“Baidu”, “data”, “mining”, “large”, “distributed”) indicate this is a big data related topic (e.g., large-scale machine learning and data mining). Its trend directly reveals that high-tech companies, such as Baidu, have paid more attention to big data related fields. 4.3.2 Visualization of Trend over Companies Here, we evaluate our model by checking the trend of several representative companies from a few important fields, such as Baidu is a famous high-tech company and Alibaba is the largest E-Commerce company in China. We visualize the evolution of the recruitment states of these companies in figure 9. From figure 9, we can first observe that state #3 is a very common state among most of companies. It is consistent with the analysis from topics of this state. Besides, we find that state #4, which is relevant to LBS, appears relatively frequently among Baidu, Sogou, and Tencent in 2014. Actually, all of these companies provide map service in China, especially Baidu. Baidu Map is the most popular navigation tool in China and many O2O companies use LBS API provided by Baidu to improve their services. So Baidu has remarkable strengths in LBS and has paid much attention to it indeed. Furthermore, Tencent, as one of the largest IT companies in China, its business is very scattered and covers many fields, such as game, social network, media, and entertainment. This kind of business strategy is directly reflected in Figure 9, where the recruitment state of Tencent changes frequently. Meanwhile, Baidu, Alibaba and Sogou (another search engine) prefer state #9, which is relevant to big data (e.g., large-scale machine learning and data mining), in 2015. Considering that their core business (search engine and Ecommerce) has several data-related practical applications (advertising and recommender system), the preference is also reasonable. In addition, we happened to find an interesting company, Zuora, whose state is almost state #9. Actually, Zuora is an enterprise software company that aim to automate billing, commerce, and finance operations with a subscription business model. Such business model is naturally related to big data processing and thus has need for senior data-related talents. Furthermore, we can observe that state #5, which is related to O2O, appears in company Tuniu, Qunar, and Feiniu frequently. Indeed, all of these companies aim to connect offline merchants and online customers, which is high consistent with O2O. Both of Tuniu and Qunar aim to provide one-stop travel booking service, such as hotel, ticket, car rental, and so on. And Feiniu is a B2C E-commerce wesite which is invested by a large retail corporation. The goal of its establishment is to fuse traditional offline service channel with online sale channel. 4.4 Evaluation of of Model Application Here, we will evaluate the proposed model by predicting basic observations in future. As introduced in Section 3, we make prediction by calculating the probability of basic observations with equation 7. So in this experiment, we will compare the likelihood of overall observations in test set to prove the performance of our model. The test set is built up by the following approach. We extract the recruitment postings in November 2015, the companies of which had also released some jobs in October 2015. It means all of companies in test set have recruitment states in the previous time span. In the end, the test set contains 350 companies and 4002 recruitment postings. Besides, all of recruitment postings between January 2014 and October 2015 are treated as train data. We select two model as baselines. One is dynamic Topic Model (DTM) [11]. DTM is a classic topic model for analyzing evolution of topics. It assumes topics evolve smoothly with respect to time and thus chains topics in adjacent epochs by state space models. Its performance was proved by predicting the next year of Science given all the articles from the previous years in [11]. It is obvious that DTM cannot be used for prediction directly, but due to its assumption, the topics extracted by it can model future data better than static topic models in some cases. In this experiment, we follow this method to prove the assumption of DTM is not very solid in our problem and indicate the necessity of our model. The code of DTM was got from [2]. The number of topics is set to 100 and all of other parameters are set to default values. In addition, we developed a simple sequence approach, called Byesian multivariate Hidden Markov Model (B-mHMM), as the other baseline. Compared with MTLVM, this baseline associates states with words directly. The joint probability distribution of B-mHMM is as follows Y Y P (ce,t \|ce,t−1 , ρ) P (χ, c\|ρ, ι) = e,t P (χe,t,m,n \|ce,t , ι) , (8) m,n where the first term is the same to equation 1. And as shown in second term, we assume that given the recruitment state, Table 5: The prediction performance of MTLVM and baseline methods in terms of log likelihood. log likelihood MTLVM B-mHMM DTM -2095737.793 -2133504.721 -2599529.728 all of observations χe,t,m,n are conditionally independent in B-mHMM and P (χe,t,m,n \|ce,t , ι) = M ulti(ιce,t ). By this baseline, we aim to prove the latent hierarchical structure of our model are meaningful for modeling recruitment markets. Because of the similarity of B-mHMM and MTLVM, The details and inference of B-mHMM follow those of MTLVM and are omitted here. Table 5 shows the log likelihood of prediction with respect to different models. The larger number means better performance. Both of B-mHMM and MTLVM outperform DTM largely. It indicates it is reasonable to employ latent recruitment states to model the trend of recruitment markets. And the performance of MTLVM is also better than that of BmHMM. All of these clearly validate the effectiveness of the proposed model. 5. RELATED WORK Generally, the related works of this paper can be grouped into two categories, namely recruitment market analysis, and sequential latent variable model. 5.1 Recruitment Market Analysis Traditionally, recruitment market analysis can be regarded as a classic topic in human capital economics, which attracts generations of researchers to contribute ever since Adam Smith. From the Macro perspective, labor is always a crucial element in studying gross economy growth, money market, exchange market and the equilibrium [26], ever since Solow proposed his growth model. Therefore, economists usually study topics about, for example, the demographic structure and participation rate of labor, the relation between inflation and unemployment rate, and how labor contributes in gross productivity or expenditures, etc [27]. In another Micro perspective, which is more relevant to our paper, all studies are set off from a basic market cleaning framework [29], where all employees choose their best balance between leisure and work, while all employers hire with budget constrain, and consequently the wage is derived as the marginal labor cost. Later researches improve our understandings by releasing constraints [19] (e.g., acknowledging the market with segmentations/barriers and as non-cleaning), or by detailed investigations (e.g., forming better utility functions of employees, and studying the actions of employees with game theory). Recently, several researchers in computer science try to employ data mining technology to solve these problems, such as offer categorization [24], job skill analysis [22]. However, previous research in economics efforts relies largely on the knowledge of domain experts or classic statistical models, and thus are usually too general to capture the high variability of recruitment market, and neglect the finegrained market trend. On the other hand, the recent research in computer science still focuses on those traditional human resource problems. Therefore, in this paper we pro- Figure 9: Visualization of the change of recruitment states over several companies, with different colors representing different states. pose a new research paradigm for recruitment market analysis by leveraging unsupervised learning approach. 5.2 Sequential Latent Variable Model Indeed, our novel sequential latent variable model MTLVM can be regarded as a combination of both Hidden Markov Model (HMM) and Hierarchical Dirichlet Processes (HDP) within a Bayesian generative framework, which can intrinsically capture the sequential dependency and variability of latent variable (e.g., recruitment states and topics). Specially, HMM based sequential latent variable models have been successfully applied to problems in a variety of fields, such as signal processing and speech recognition [25, 21], biometrics [15], genetics [14, 23], economics [18], and mobile Internet mining [20, 9, 33]. For much of their history, HMMs have been implemented by using recursive algorithms developed for parameter estimation [10], which are viewed as “black boxes” by many statisticians. In recent years, some researchers proposed to use Bayesian methods to simulate HMM parameters from the posterior distribution, which can provide more scalable and stable process of parameter estimation for HMM [16, 17]. Compared with the traditional maximum-likelihood estimation (MLE) based HMM learning solution, the Bayesian methods can directly maximize the probability of the hidden variables given the observed data by integrating over all possible parameter values rather than searching for an optimal set of parameter values. To this end, the model proposed in this paper also follows a Bayesian generative framework. Latent Dirichlet Allocation (LDA) [12] based latent variable models, have become one of the most powerful tools for mining textual data. However, most topic models [11, 31, 32] need a predefined parameter to indicate the number of topics, and thus fail to capture the variability of topics. To this end, the Hierarchical Dirichlet Processes (HDP) [28] is proposed as an infinity version of topic model, which can automatically learn the number of topics. Therefore, in this paper we propose to ingrate HDP into our MTLVM for capturing the variability of latent recruitment topics. 6. CONCLUSION In this paper, we provided a large-scale data driven analysis of recruitment market trends. Specifically, we developed a novel sequential latent variable model, named MTLVM, which is designed for capturing the temporal dependencies of corporate recruitment states and is able to automatically learn the latent recruitment topics within a Bayesian generative framework. Moreover, to capture the variability of recruitment topics over time, we designed hierarchical dirichlet processes for MTLVM. These processes allow to dynamically generate recruitment topics. Finally, we implemented a prototype system to empirically evaluate our approach based on large-scale real-world recruitment data. The results showed that our approach could effectively discover recruitment market trends and provide guidances for both job recruiters and job seekers. 7. APPENDIX Here, we will describe an analog of Chinese Restaurant Process (CRP) for P (χe,t \|Λ, ce,t ) and corresponding inference in detail. Specifically, we first define {ψe,t,m,z }z as variables sampled from Gce,t for each job posting and \|ψe,t,m \| = Ze,t,m . Each φe,t,m,n is linked with a ψe,t,m,z and ze,t,m,n is the index of ψe,t,m,z for φe,t,m,n . In other words, we have φe,t,m,n = ψe,t,m,ze,t,m,n . Besides, let ie,t,m,z denote the number of {φe,t,m,n }n linked with ψe,t,m,z . According to CRP, we can integrate out the De,t,m and get the conditional distribution of φe,t,m,n as follows. φe,t,m,n \|φe,t,m,1 , ..., φe,t,m,n−1 , {Gc }c , ce,t , γ2 ∼ Ze,t,m X z=1 γ2 ie,t,m,z δψ + Gc , n − 1 + γ2 e,t,m,z n − 1 + γ2 e,t (9) where δψ is a probability measure concentrated at ψ. If a new ψ is sampled, it indicates the second term on the right-hand side of equation 9 is chosen. Then we need to add Ze,t,m by 1, sample a new ψ by equation 10, and allocate φe,t,m,n to it. If the sampled value is an existed φ, we just need allocate φe,t,m,n to it. Second, we define {ηc,s }s as variables sampled from G0 for each recruitment state and \|ηc \| = Sc . Each ψe,t,m,z is linked with a ηc,s and se,t,m,z is the index of ηc,s for ψe,t,m,z . In other words, ψe,t,m,z = ηc,se,t,m,z . Besides, let jc,s denote the number of {ψe,t,m,z }e,t,m , which is linked with ηc,s and ce,t = c. Similarly, we can integrate out Gc and get the conditional distribution of ψe,t,m,z as follows. ψe,t,m,z \|ψ1,1,1,1 , ..., ψe,t,m,z−1 , Q0 , γ1 ∼ Psc j j Psc c,s δηc,s + Psc c,s Q0 . s=1 j +γ j +γ s=1 c,s 1 s=1 c,s The sampling process is similar to φ. 1 (10) Third, we let {θk }k denote the variables sampled from H and \|θ\| is K. Each ηc,s is linked with a θk and kc,s is the index of θk for ηc,s , i.e., ηc,s = θkc,s . And we also let ok denote the number of {ηc,s }c,s . Now we can write the conditional distribution of ηc,s directly as ηc,s \|η1,1 , ..., ηc,s−1 , H, γ0 ∼ PK k=1 PK ok δ ok +γ0 θk + k=1 PK ok H. ok +γ0 (11) k=1 Next, we will describe a Gibbs sampling method yielded from above. Specifically, we follow the inference method in [28] and sample z, s, k and θ, rather than dealing with φ, ψ and η directly. Sampling z, s, and k. Relying on equation 9, we can easily compute the conditional distribution of ze,t,m,n by P (ze,t,m,n = z\|z −e,t,m,n , s, k, θ, χ) ∝ ( new, γ2 P (χe,t,m,n \|θke,se,t,m,z ) used, i−e,t,m,n e,t,m,z P (χe,t,m,n \|θke,se,t,m,z ) (12) where i−e,t,m,n is ie,t,m,z except the variable φe,t,m,n . The e,t,m,z likelihood of ze,t,m,n = z is simply P (χe,t,m,n \|θke,se,t,m,z ) given all of other variables. And the prior probability that ze,t,m,n samples an existed ψe,t,m,z is proportional to i−e,t,m,n e,t,m,z is ie,t,m,z . Its prior probability for a new ψ is proportional to γ2 . The process of sampling s and k is similar to that of sampling z. Sampling θ. Given z, s, and k, {θk }k are mutually independent. So the conditional distribution for each θk is only related with all of χ that linked with it. It follows P (θk \|z, t, k, θ−k , χ) ∝ h(θk ) · Q c,s:kc,s =k Q e,t:ce,t =c Q m,z:se,t,m,z =s Q n:ze,t,m,n =z P (χe,t,m,n \|θk ), (13) where h(θk ) is the density of measure H at parameter θk . 8. REFERENCES [1] Baidu translation. http://fanyi.baidu.com/. [2] Code of dynamic topic model. https://code.google.com/archive/p/ princeton-statistical-learning/downloads. [3] Dice. http://www.dice.com/. [4] Forbes article. http://suo.im/gxxz0. [5] Internet plus. https://en.wikipedia.org/wiki/Internet Plus. [6] Lagou. http://www.lagou.com/. [7] Linkedin. http://www.linkedin.com/. [8] D. J. Aldous. Exchangeability and related topics. Springer, 1985. [9] T. Bao, H. Cao, E. Chen, J. Tian, and H. Xiong. An unsupervised approach to modeling personalized contexts of mobile users. Knowledge and Information Systems, 31(2):345–370, 2012. [10] L. E. Baum, T. Petrie, G. Soules, and N. Weiss. A maximization technique occurring in the statistical analysis of probabilistic functions of markov chains. The annals of mathematical statistics, 41(1):164–171, 1970. [11] D. M. Blei and J. D. Lafferty. Dynamic topic models. In Proceedings of the 23rd international conference on Machine learning, pages 113–120. ACM, 2006. [12] D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation. the Journal of machine Learning research, 3:993–1022, 2003. [13] J. Chang, S. Gerrish, C. Wang, J. L. Boyd-Graber, and D. M. Blei. Reading tea leaves: How humans interpret topic models. In Advances in neural information processing systems, pages 288–296, 2009. [14] G. A. Churchill. Stochastic models for heterogeneous dna sequences. Bulletin of mathematical biology, 51(1):79–94, 1989. [15] D. R. Fredkin and J. A. Rice. Bayesian restoration of single-channel patch clamp recordings. Biometrics, pages 427–448, 1992. [16] S. Goldwater and T. Griffiths. A fully bayesian approach to unsupervised part-of-speech tagging. In Annual meeting-association for computational linguistics, volume 45, page 744. Citeseer, 2007. [17] S. Guha, L. Yi, and D. Neuberg. Bayesian hidden markov modeling of array cgh data. Journal of the American Statistical Association, 103:485–497, 2008. [18] J. D. Hamilton. A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica: Journal of the Econometric Society, pages 357–384, 1989. [19] T. Hayashi and T. Sakai. Nash implementation of competitive equilibria in the job-matching market. International Journal of Game Theory, 38(4):453–467, 2009. [20] B. Huai, E. Chen, H. Zhu, H. Xiong, T. Bao, Q. Liu, and J. Tian. Toward personalized context recognition for mobile users: a semisupervised bayesian hmm approach. ACM Transactions on Knowledge Discovery from Data (TKDD), 9(2):10, 2014. [21] B.-H. Juang and L. R. Rabiner. Hidden markov models for speech recognition. Technometrics, 33(3):251–272, 1991. [22] C. Litecky, A. J. Igou, and A. Aken. Skills in the management oriented is and enterprise system job markets. In Proceedings of the 50th annual conference on Computers and People Research, pages 35–44. ACM, 2012. [23] J. S. Liu, A. F. Neuwald, and C. E. Lawrence. Markovian structures in biological sequence alignments. Journal of the American Statistical Association, 94(445):1–15, 1999. [24] E. Malherbe, M. Cataldi, and A. Ballatore. Bringing order to the job market: Efficient job offer categorization in e-recruitment. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1101–1104. ACM, 2015. [25] L. R. Rabiner. A tutorial on hidden markov models and selected applications in speech recognition. Proceedings of the IEEE, 77(2):257–286, 1989. [26] D. Romer and C. Chow. Advanced Macroeconomic Theory. Mcgraw-hill, 1996. [27] C. Shapiro and J. E. Stiglitz. Equilibrium unemployment as a worker discipline device. The American Economic Review, 74(3):433–444, 1984. [28] Y. W. Teh, M. I. Jordan, M. J. Beal, and D. M. Blei. Hierarchical dirichlet processes. Journal of the american statistical association, 101(476), 2006. [29] H. R. Varian and J. Repcheck. Intermediate microeconomics: a modern approach, volume 6. WW Norton & Company New York, NY, 2010. [30] P. Xie and E. P. Xing. Integrating document clustering and topic modeling. arXiv preprint arXiv:1309.6874, 2013. [31] H. Zhang, G. Kim, and E. P. Xing. Dynamic topic modeling for monitoring market competition from online text and image data. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1425–1434. ACM, 2015. [32] C. Zhu, H. Zhu, Y. Ge, E. Chen, Q. Liu, T. Xu, and H. Xiong. Tracking the evolution of social emotions with topic models. Knowledge and Information Systems, pages 1–28, 2015. [33] H. Zhu, C. Liu, Y. Ge, H. Xiong, and E. Chen. Popularity modeling for mobile apps: A sequential approach. Cybernetics, IEEE Transactions on, 45(7):1303–1314, 2015.	2cs.AI
Matrix Completion and Related Problems via Strong Duality Maria-Florina Balcan∗ Yingyu Liang† David P. Woodruff‡ Hongyang Zhang§ arXiv:1704.08683v3 [cs.DS] 19 Jan 2018 Abstract This work studies the strong duality of non-convex matrix factorization problems: we show that under certain dual conditions, these problems and its dual have the same optimum. This has been well understood for convex optimization, but little was known for non-convex problems. We propose a novel analytical framework and show that under certain dual conditions, the optimal solution of the matrix factorization program is the same as its bi-dual and thus the global optimality of the non-convex program can be achieved by solving its bi-dual which is convex. These dual conditions are satisfied by a wide class of matrix factorization problems, although matrix factorization problems are hard to solve in full generality. This analytical framework may be of independent interest to non-convex optimization more broadly. We apply our framework to two prototypical matrix factorization problems: matrix completion and robust Principal Component Analysis (PCA). These are examples of efficiently recovering a hidden matrix given limited reliable observations of it. Our framework shows that exact recoverability and strong duality hold with nearly-optimal sample complexity guarantees for matrix completion and robust PCA. ∗ Carnegie Mellon University. Email: ninamf@cs.cmu.edu University of Wisconsin-Madison. Email: yliang@cs.wisc.edu ‡ Carnegie Mellon University. Email: dwoodruf@cs.cmu.edu § Corresponding author. Carnegie Mellon University. Email: hongyanz@cs.cmu.edu † 1 Introduction Non-convex matrix factorization problems have been an emerging object of study in theoretical computer science [JNS13, Har14, SL15, RSW16], optimization [WYZ12, SWZ14], machine learning [BNS16, GLM16, GHJY15, JMD10, LLR16, WX12], and many other domains. In theoretical computer science and optimization, the study of such models has led to significant advances in provable algorithms that converge to local minima in linear time [JNS13, Har14, SL15, AAZB+ 16, AZ16]. In machine learning, matrix factorization serves as a building block for large-scale prediction and recommendation systems, e.g., the winning submission for the Netflix prize [KBV09]. Two prototypical examples are matrix completion and robust Principal Component Analysis (PCA). This work develops a novel framework to analyze a class of non-convex matrix factorization problems with strong duality, which leads to exact recoverability for matrix completion and robust Principal Component Analysis (PCA) via the solution to a convex problem. The matrix factorization problems can be stated as finding a target matrix X∗ in the form of X∗ = AB, by minimizing the objective function H(AB) + 1 2 n1 ×r and B ∈ Rr×n2 with a known value of r min{n , n }, where 1 2 2 kABkF over factor matrices A ∈ R H(·) is some function that characterizes the desired properties of X∗ . Our work is motivated by several promising areas where our analytical framework for non-convex matrix factorizations is applicable. The first area is low-rank matrix completion, where it has been shown that a low-rank matrix can be exactly recovered by finding a solution of the form AB that is consistent with the observed entries (assuming that it is incoherent) [JNS13, SL15, GLM16]. This problem has received a tremendous amount of attention due to its important role in optimization and its wide applicability in many areas such as quantum information theory and collaborative filtering [Har14, ZLZ16, BZ16]. The second area is robust PCA, a fundamental problem of interest in data processing that aims at recovering both the low-rank and the sparse components exactly from their superposition [CLMW11, NNS+ 14, GWL16, ZLZC15, ZLZ16, YPCC16], where the low-rank component corresponds to the product of A and B while the sparse component is captured by a proper choice of function H(·), e.g., the `1 norm [CLMW11, ABHZ16]. We believe our analytical framework can be potentially applied to other non-convex problems more broadly, e.g., matrix sensing [TBSR15], dictionary learning [SQW17a], weighted low-rank approximation [RSW16, LLR16], and deep linear neural network [Kaw16], which may be of independent interest. Without assumptions on the structure of the objective function, direct formulations of matrix factorization problems are NP-hard to optimize in general [HMRW14, ZLZ13]. With standard assumptions on the structure of the problem and with sufficiently many samples, these optimization problems can be solved efficiently, e.g., by convex relaxation [CR09, Che15]. Some other methods run local search algorithms given an initialization close enough to the global solution in the basin of attraction [JNS13, Har14, SL15, GHJY15, JGN+ 17]. However, these methods have sample complexity significantly larger than the information-theoretic lower bound; see Table 1 for a comparison. The problem becomes more challenging when the number of samples is small enough that the sample-based initialization is far from the desired solution, in which case the algorithm can run into a local minimum or a saddle point. Another line of work has focused on studying the loss surface of matrix factorization problems, providing positive results for approximately achieving global optimality. One nice property in this line of research is that there is no spurious local minima for specific applications such as matrix completion [GLM16], matrix sensing [BNS16], dictionary learning [SQW17a], phase retrieval [SQW16], linear deep neural networks [Kaw16], etc. However, these results are based on concrete forms of objective functions. Also, even when any local minimum is guaranteed to be globally optimal, in general it remains NP-hard to escape high-order saddle points [AG16], and additional arguments are needed to show the achievement of a local minimum. Most importantly, all existing results rely on strong assumptions on the sample size. 1 Strong Duality Primal Problem (1) Bi-Dual Problem (2) X AB Surface of (1) Surface of (2) By Theorem 3.3 + Dual Certificate Common Optimal Solution Figure 1: Strong duality of matrix factorizations. 1.1 Our Results Our work studies the exact recoverability problem for a variety of non-convex matrix factorization problems. The goal is to provide a unified framework to analyze a large class of matrix factorization problems, and to achieve efficient algorithms. Our main results show that although matrix factorization problems are hard to optimize in general, under certain dual conditions the duality gap is zero, and thus the problem can be converted to an equivalent convex program. The main theorem of our framework is the following. Theorems 3.3 (Strong Duality. Informal). Under certain dual conditions, strong duality holds for the non-convex optimization problem 1 e B) e = (A, argmin F (A, B) = H(AB) + kABk2F , H(·) is convex and closed, (1) 2 A∈Rn1 ×r ,B∈Rr×n2 where “the function H(·) is closed” means that for each α ∈ R, the sub-level set {X ∈ Rn1 ×n2 : H(X) ≤ α} is a closed set. In other words, problem (1) and its bi-dual problem e = argmin H(X) + kXkr∗ , X (2) X∈Rn1 ×n2 eB e = X, e where kXkr∗ is a convex function defined have exactly the same optimal solutions in the sense that PA r 1 2 2 2 by kXkr∗ = maxM hM, Xi − 2 kMkr and kMkr = i=1 σi (M) is the sum of the first r largest squared singular values. Theorem 3.3 connects the non-convex program (1) to its convex counterpart via strong duality; see Figure 1. We mention that strong duality rarely happens in the non-convex optimization region: low-rank matrix approximation [OW92] and quadratic optimization with two quadratic constraints [BE06] are among the few paradigms that enjoy such a nice property. Given strong duality, the computational issues of the original problem can be overcome by solving the convex bi-dual problem (2). The positive result of our framework is complemented by a lower bound to formalize the hardness of the above problem in general. Assuming that the random 4-SAT problem is hard (see Conjecture 1) [RSW16], we give a strong negative result for deterministic algorithms. If also BPP = P (see Section 6 for a discussion), then the same conclusion holds for randomized algorithms succeeding with probability at least 2/3. Theorem 6.1 (Hardness Statement. Informal). Assuming that random 4-SAT is hard on average, there is a problem in the form of (1) such that any deterministic algorithm achieving (1 + )OPT in the objective function value with ≤ 0 requires 2Ω(n1 +n2 ) time, where OPT is the optimum and 0 > 0 is an absolute constant. If BPP = P, then the same conclusion holds for randomized algorithms succeeding with probability at least 2/3. Our framework only requires the dual conditions in Theorem 3.3 to be verified. We will show that two prototypical problems, matrix completion and robust PCA, obey the conditions. They belong to the linear inverse problems of form (1) with a proper choice of function H(·), which aim at exactly recovering a hidden matrix X∗ with rank(X∗ ) ≤ r given a limited number of linear observations of it. For matrix completion, the linear measurements are of the form {X∗ij : (i, j) ∈ Ω}, where Ω is the support set which is uniformly distributed among all subsets of [n1 ] × [n2 ] of cardinality m. With strong 2 Table 1: Comparison of matrix completion methods. Here κ = σ1 (X∗ )/σr (X∗ ) is the condition number of e obeys kX e − X∗ kF ≤ , n(1) = max{n1 , n2 } and X∗ ∈ Rn1 ×n2 , is the accuracy such that the output X n(2) = min{n1 , n2 }. The first line of ours is an information-theoretic upper bound and the second line is a polynomial-time approach. Work Sample Complexity µ-Incoherence ∗ [JNS13] O κ4 µ2 r4.5 n(1) log n(1) log rkX kF Condition (3) ∗ 2 ∗ n kX k kX kF [Har14] O µrn(1) (r + log (1) F Condition (3) σ2 r O(max{µ6 κ16 r4 , µ4 κ4 r6 }n(1) log2 n(1) ) n o qn O(rn(1) κ2 max µ log n(2) , n(1) µ2 r6 κ4 [GLM16] [SL15] [ZWL15] [KMO10a] [Gro11] [Che15] Ours Lower Bound1 [CT10] O µ2 r4 κ2 (1) κ2 max(µ, log n + µr log kX∗ kF (1) )) µ ∗ n(2) kX kF Condition (3) kX∗ kF n(1) log 1 3n O µr log n log (1) (1) n o qn qn O n(2) r n(1) κ2 max µ log n(2) , µ2 r n(1) κ4 (2) (2) 2 √ Condition (3) (2) O(µr2 n [ZL16] [GLZ17] kX∗i: k2 ≤ O(µrn(1) log n(1) ) O(µrn(1) log2 n(1) ) O(µrn(1) log n(1) ) 2 O(κ µrn(1) log(n(1) ) log2κ (n(1) )) Ω(µrn(1) log n(1) ) Condition (3) Condition (3) Similar to (3) and (13) Conditions (3) and (13) Condition (3) Condition (3) Condition (3) Condition (3) duality, we can either study the exact recoverability of the primal problem (1), or investigate the validity of its convex dual (or bi-dual) problem (2). Here we study the former with tools from geometric functional analysis. Recall that in the analysis of matrix completion, one typically requires a µ-incoherence condition for a given rank-r matrix X∗ with skinny SVD UΣVT [Rec11, CT10]: r r µr µr T T , and kV ei k2 ≤ , for all i (3) kU ei k2 ≤ n1 n2 where ei ’s are vectors with i-th entry equal to 1 and other entries equal to 0. The incoherence condition claims that information spreads throughout the left and right singular vectors and is quite standard in the matrix completion literature. Under this standard condition, we have the following results. Theorems 4.1, 4.3, and 4.2 (Matrix Completion. Informal). X∗ ∈ Rn1 ×n2 is the unique matrix of rank at most r that is consistent with the m measurements with high probability, provided m = O(µ(n1 + n2 )r log(n1 + n2 )) and X∗ satisfies incoherence (3). In addition, there exists a convex optimization for matrix completion in the form of (2) that exactly recovers X∗ with high probability, provided that m = O(κ2 µ(n1 + n2 )r log(n1 + n2 ) log2κ (n1 + n2 )), where κ is the condition number of X∗ . To the best of our knowledge, our result is the first to connect convex matrix completion to non-convex matrix completion, two parallel lines of research that have received significant attention in the past few years. Table 1 compares our result with prior results. For robust PCA, instead of studying exact recoverability of problem (1) as for matrix completion, we investigate problem (2) directly. The robust PCA problem is to decompose a given matrix D = X∗ + S∗ into the sum of a low-rank component X∗ and a sparse component S∗ [ANW12]. We obtain the following theorem for robust PCA. 1 This lower bound is information-theoretic. 3 Theorems 5.1 (Robust PCA. Informal). There exists a convex optimization formulation for robust PCA in ∗ n1 ×n2 and S∗ ∈ Rn1 ×n2 with high the form of problem (2) that exactly recovers the incoherent matrix X ∈ R min{n ,n } 1 2 probability, even if rank(X∗ ) = Θ µ log2 max{n and the size of the support of S∗ is m = Θ(n1 n2 ), 1 ,n2 } where the support set of S∗ is uniformly distributedq among all sets of cardinality m, and the incoherence σr (X∗ ). parameter µ satisfies constraints (3) and kX∗ k∞ ≤ nµr 1 n2 The bounds in Theorem 5.1 match the best known results in the robust PCA literature when the supports of S∗ are uniformly sampled [CLMW11], while our assumption is arguably more intuitive; see Section 5. Note that our results hold even when X∗ is close to full rank and a constant fraction of the entries have noise. Independently of our work, Ge et al. [GJY17] developed a framework to analyze the loss surface of low-rank problems, and applied the framework to matrix completion and robust PCA. Their bounds are: for matrix completion, the sample complexity is O(κ6 µ4 r6 (n1 + n2 ) log(n1 + n2)); forrobust PCA, the outlier entries 1 n2 are deterministic and the number that the method can tolerate is O nµrκ 5 . Zhang et al. [ZWG17] also studied the robust PCA problem using non-convex optimization, where the outlier entries are deterministic 1 n2 and the number of outliers that their algorithm can tolerate is O nrκ . The strong duality approach is unique to our work. 1.2 Our Techniques Reduction to Low-Rank Approximation. Our results are inspired by the low-rank approximation problem: min A∈Rn1 ×r ,B∈Rr×n2 1 e − ABk2 . k−Λ F 2 (4) We know that all local solutions of (4) are globally optimal (see Lemma 3.1) and that strong duality holds for e ∈ Rn1 ×n2 [GRG16]. To extend this property to our more general problem (1), our main any given matrix −Λ insight is to reduce problem (1) to the form of (4) using the `2 -regularization term. While some prior work attempted to apply a similar reduction, their conclusions either depended on unrealistic conditions on local solutions, e.g., all local solutions are rank-deficient [HYV14, GRG16], or their conclusions relied on strong assumptions on the objective functions, e.g., that the objective functions are twice-differentiable [HV15]. e For concrete Instead, our general results formulate strong duality via the existence of a dual certificate Λ. applications, the existence of a dual certificate is then converted to mild assumptions, e.g., that the number of measurements is sufficiently large and the positions of measurements are randomly distributed. We will illustrate the importance of randomness below. e may not exist in the deterministic world. The Blessing of Randomness. The desired dual certificate Λ A hardness result [RSW16] shows that for the problem of weighted low-rank approximation, which can be cast in the form of (1), without some randomization in the measurements made on the underlying low rank matrix, it is NP-hard to achieve a good objective value, not to mention to achieve strong duality. A similar phenomenon was observed for deterministic matrix completion [HM12]. Thus we should utilize such randomness to analyze the existence of a dual certificate. For matrix completion, the assumption that the measurements are random is standard, under which, the angle between the space Ω (the space of matrices which are consistent with observations) and the space T (the space of matrices which are low-rank) is small with high probability, namely, X∗ is almost the unique low-rank matrix that is consistent with the measurements. Thus, our dual certificate can be represented as another form of a convergent Neumann series concerning the projection operators on the spaces Ω and T . The remainder of the proof is to show that such a construction obeys the dual conditions. To prove the dual conditions for matrix completion, we use the fact that the subspace Ω and the complement space T ⊥ are almost orthogonal when the sample size is sufficiently large. This implies 4 the projection of our dual certificate on the space T ⊥ has a very small norm, which exactly matches the dual conditions. Non-Convex Geometric Analysis. Strong duality implies that the primal problem (1) and its bi-dual problem (2) have exactly the same solutions Ω" e e e in the sense that AB = X. Thus, to show exact recoverability of linear inverse problems such as matrix completion and robust PCA, it suffices to study either the non-convex primal problem (1) or its convex counterpart (2). Here we do the former analysis for matrix completion. We mention 0 # − {& ∗ } that traditional techniques [CT10, Rec11, CRPW12] for convex optimization break down for our non-convex problem, since the subgradient of a non-convex objective function may not even exist [BV04]. Instead, we apply tools from geometric functional analysis [Ver09] to analyze the Descent Cone ) (& ∗ ) * geometry of problem (1). Our non-convex geometric analysis is in stark contrast to prior techniques of convex geometric analysis [Ver15] where Figure 2: Feasibility. convex combinations of non-convex constraints were used to define the Minkowski functional (e.g., in the definition of atomic norm) while our method uses the non-convex constraint null(A) itself. For matrix completion, problem (1) has two hard constraints: a) the rank of the output matrix should be no larger than r, as implied by the form of AB; b) the output matrix should be consistent with the sampled measurements, i.e., PΩ (AB) = PΩ (X∗ ). We study the feasibility condition of problem (1)∗ )from a geometric -. (& ∗ eB e = X is the unique feasible solution to problem (1) if and only if starting from X∗ , the0 perspective: A rank of X∗ + D increases for all directions D’s in the constraint set Ω⊥ = {D ∈ Rn1 ×n2 : PΩ (X∗ + D) = /01 ×03 45 ∗ 45 ) ∩ /01 ×03that . (&requirement PΩ (X∗ )} (a.k.a. the feasibility condition). This can be geometrically interpreted as-the the descent cone DS (X∗ ) = {t(X − X∗ ) ∈ Rn1 ×n2 : rank(X) ≤ r, t ≥ 0} and the constraint set Ω⊥ must intersect uniquely at 0 (see Figure 2), which means X∗ is the unique matrix that satisfies the constraints a) and b). This is shown by the following tangent cone argument. Let S be the set of all matrices with rank at most r around the underlying matrix X∗ . In the tangent cone argument, by definition, DS (X∗ ) is a subset of the tangent cone of S at X∗ . The latter cone of interest has a very nice form, namely, it is just the space T mentioned above (the space of matrices which are low-rank). Now leverage results from prior work which imply T ∩ Ω⊥ = {0} with a large enough sample size. Namely, among all matrices of the form X∗ + D, D = 0 is the only matrix such that rank(X∗ + D) ≤ r and X∗ + D is consistent with the observations. Using this argument, we can show that the sample size needed for exact recovery in matrix completion matches the known lower bound up to a constant factor. Putting Things Together. We summarize our new analytical framework with the following figure. Geometric Analysis Exact Recovery by Non-Convex Problem (1) with Optimal Sample Complexity Non-Convex Problem (1) Exact Recovery by Convex Problem (2) (NP-hard in general) Randomness Construction of Dual Certificate Reduction to Strong Duality Low-Rank Approximation Other Techniques. An alternative method is to investigate the exact recoverability of problem (2) via standard convex analysis. We find that the sub-differential of our induced function k · kr∗ is very similar to 5 that of the nuclear norm. With this observation, we prove the validity of robust PCA in the form of (2) by combining this property of k · kr∗ with standard techniques from [CLMW11]. 2 Preliminaries We will use calligraphy to represent a set, bold capital letters to represent a matrix, bold lower-case letters to represent a vector, and lower-case letters to represent scalars. Specifically, we denote by X∗ ∈ Rn1 ×n2 the underlying matrix. We use X:t ∈ Rn1 ×1 (Xt: ∈ R1×n2 ) to indicate the t-th column (row) of X. The entry in the i-th row, j-th column of X is represented by Xij . The condition number of X is κ = σ1 (X)/σr (X). We let n(1) = max{n1 , n2 } and n(2) = min{n1 , n2 }. For a function H(M) on an input matrix M, its conjugate function H ∗ is defined by H ∗ (Λ) = maxM hΛ, Mi − H(M). Furthermore, let H ∗∗ denote the conjugate function of H ∗ . We will frequently use rank(X) ≤ r to constrain the rank of X. This can be equivalently represented as X = AB, qPby restricting the number of columns of A and rows of B to be r. For norms, we denote by 2 kXkF = ij Xij the Frobenius norm of matrix X. Let σ1 (X) ≥ σ2 (X) ≥ ... ≥ σr (X) be the non-zero P singular values of X. The nuclear norm (a.k.a. trace norm) of X is defined by kXk∗ = ri=1 σi (X), and the operator norm of X is kXk = σ1 (X). Denote and P by kXk∞ = maxij \|Xij \|. For two matrices nA ×n 1 2 : B of equal dimensions, we denote by hA, Bi = ij Aij Bij . We denote by ∂H(X) = {Λ ∈ R H(Y) ≥ H(X) + hΛ, Y − Xi for any Y} the sub-differential of function H evaluated at X. We define ( 0, if X ∈ C; the indicator function of convex set C by IC (X) = For any non-empty set C, denote by +∞, otherwise. cone(C) = {tX : X ∈ C, t ≥ 0}. We denote by Ω the set of indices of observed entries, and Ω⊥ its complement. Without confusion, Ω also indicates the linear subspace formed by matrices with entries in Ω⊥ being 0. We denote by PΩ : Rn1 ×n2 → Rn1 ×n2 the orthogonal projector of subspace Ω. We will consider a single norm for these operators, namely, the operator norm denoted by kAk and defined by kAk = supkXkF =1 kA(X)kF . For any orthogonal projection operator PT to any subspace T , we know that kPT k = 1 whenever dim(T ) 6= 0. For distributions, denote by N (0, 1) a standard Gaussian random variable, Uniform(m) the uniform distribution of cardinality m, and Ber(p) the Bernoulli distribution with success probability p. 3 `2 -Regularized Matrix Factorizations: A New Analytical Framework In this section, we develop a novel framework to analyze a general class of `2 -regularized matrix factorization problems. Our framework can be applied to different specific problems and leads to nearly optimal sample complexity guarantees. In particular, we study the `2 -regularized matrix factorization problem (P) 1 F (A, B) = H(AB) + kABk2F , H(·) is convex and closed. 2 A∈Rn1 ×r ,B∈Rr×n2 min We show that under suitable conditions the duality gap between (P) and its dual (bi-dual) problem is zero, so problem (P) can be converted to an equivalent convex problem. 3.1 Strong Duality b 2 − hY, b ABi for a fixed Y, b leading to the objective We first consider an easy case where H(AB) = 12 kYk F 1 b 2 function 2 kY − ABkF . For this case, we establish the following lemma. 6 b − ABk2 is globally optimal, b any local minimum of f (A, B) = 1 kY Lemma 3.1. For any given matrix Y, F 2 b The objective function f (A, B) around any saddle point has a negative second-order given by svdr (Y). directional curvature. Moreover, f (A, B) has no local maximum.2 The proof of Lemma 3.1 is basically to calculate the gradient of f (A, B) and let it equal to zero; see b b − ABk2 for some Y Appendix B for details. Given this lemma, we can reduce F (A, B) to the form 12 kY F plus an extra term: 1 1 1 F (A, B) = kABk2F + H(AB) = kABk2F + H ∗∗ (AB) = max kABk2F + hΛ, ABi − H ∗ (Λ) Λ 2 2 2 (5) 1 1 = max k − Λ − ABk2F − kΛk2F − H ∗ (Λ) , max L(A, B, Λ), Λ 2 Λ 2 1 1 2 2 where we define L(A, B, Λ) , 2 k − Λ − ABkF − 2 kΛkF − H ∗ (Λ) as the Lagrangian of problem (P),3 and the second equality holds because H is closed and convex w.r.t. the argument AB. For any fixed value of Λ, by Lemma 3.1, any local minimum of L(A, B, Λ) is globally optimal, because minimizing L(A, B, Λ) is equivalent to minimizing 12 k − Λ − ABk2F for a fixed Λ. e such that (A, e B, e Λ) e is a primal-dual saddle The remaining part of our analysis is to choose a proper Λ e e B). e point of L(A, B, Λ), so that minA,B L(A, B, Λ) and problem (P) have the same optimal solution (A, For this, we introduce the following condition, and later we will show that the condition holds with high probability. e B) e to problem (P), there exists an Λ e ∈ ∂X H(X)\| e e such that Condition 1. For a solution (A, X=AB eB eB eT = Λ eB eT −A and e T (−A e B) e =A e T Λ. e A (6) Explanation of Condition 1. We note that ∇A L(A, B, Λ) = ABBT + ΛBT and ∇B L(A, B, Λ) = e Λ)\| e e in (6), then ∇A L(A, B, AT AB + AT Λ for a fixed Λ. In particular, if we set Λ to be the Λ e =0 A=A e e e e and ∇B L(A, B, Λ)\|B=B e = 0. So Condition 1 implies that (A, B) is either a saddle point or a local e e minimizer of L(A, B, Λ) as a function of (A, B) for the fixed Λ. The following lemma states that if it is a local minimizer, then strong duality holds. e B) e be a global minimizer of F (A, B). If there exists a dual certificate Lemma 3.2 (Dual Certificate). Let (A, e e is a local minimizer of L(A, B, Λ) e for the fixed Λ, e e then strong Λ satisfying Condition 1 and the pair (A, B) eB e = svdr (−Λ). e duality holds. Moreover, we have the relation A e B, e Λ) e is a primal-dual saddle Proof Sketch. By the assumption of the lemma, we can show that (A, point to the Lagrangian L(A, B, Λ); see Appendix C. To show strong duality, by the fact that F (A, B) = e = argmaxΛ L(A, e B, e Λ), we have F (A, e B) e = L(A, e B, e Λ) e ≤ L(A, B, Λ), e maxΛ L(A, B, Λ) and that Λ e e e for any A, B, where the inequality holds because (A, B, Λ) is a primal-dual saddle point of L. So on the e B) e ≤ minA,B L(A, B, Λ) e ≤ maxΛ minA,B L(A, B, Λ). one hand, minA,B maxΛ L(A, B, Λ) = F (A, On the other hand, by weak duality, we have minA,B maxΛ L(A, B, Λ) ≥ maxΛ minA,B L(A, B, Λ). Therefore, minA,B maxΛ L(A, B, Λ) = maxΛ minA,B L(A, B, Λ), i.e., strong duality holds. Therefore, eB e = argminAB L(A, B, Λ) e = argminAB 1 kABk2 + hΛ, e ABi − H ∗ (Λ) e = argminAB 1 k − Λ e − A F 2 2 e as desired. ABk2F = svdr (−Λ), This lemma then leads to the following theorem. 2 e is of full rank and does not Prior work studying the loss surface of low-rank matrix approximation assumes that the matrix Λ have the same singular values [BH89]. In this work, we generalize this result by removing these two assumptions. 3 One can easily check that L(A, B, Λ) = minM L0 (A, B, M, Λ), where L0 (A, B, M, Λ) is the Lagrangian of the constraint optimization problem minA,B,M 12 kABk2F + H(M), s.t. M = AB. With a little abuse of notation, we call L(A, B, Λ) the Lagrangian of the unconstrained problem (P) as well. 7 % −, + small . O %& %' = ψ !"#$ %& % $ large % , (dual certificate) Figure 3: Geometry of dual condition (7) for general matrix factorization problems. e B) e the optimal solution of problem (P). Define a matrix space Theorem 3.3. Denote by (A, e T + YB, e X ∈ Rn2 ×r , Y ∈ Rn1 ×r }. T , {AX Then strong duality holds for problem (P), provided that e ∈ ∂H(A e B) e , Ψ, (1) Λ e =A e B, e (2) PT (−Λ) e < σr (A e B). e (3) kPT ⊥ Λk (7) e so that the conditions in Lemma 3.2 hold. Λ e should Proof. The proof idea is to construct a dual certificate Λ satisfy the following: (a) (b) (c) e ∈ ∂H(A e B), e Λ (by Condition 1) eB e + Λ) e B e T = 0 and A e T (A eB e + Λ) e = 0, (A (by Condition 1) eB e = svdr (−Λ). e A (by the local minimizer assumption and Lemma 3.1) (8) eA e † )M(I − B eB e † ) and so kPT ⊥ Mk ≤ kMk, It turns out that for any matrix M ∈ Rn1 ×n2 , PT ⊥ M = (I − A eB e and V the a fact that we will frequently use in the sequel. Denote by U the left singular space of A right singular space. Then the linear space T can be equivalently represented as T = U + V. Therefore, eB e + Λ) e B e T = 0 and A e T (A eB e + Λ) e = 0 imply T ⊥ = (U + V)⊥ = U ⊥ ∩ V ⊥ . With this, we note that: (b) (A T ⊥ ⊥ ⊥ e e e e e e e e e e e e AB + Λ ∈ Null(A ) = Col(A) and AB + Λ ∈ Row(B) (so AB + Λ ∈ T ), and vice versa. And (c) eB e = svdr (−Λ) e implies that for an orthogonal decomposition −Λ e =A eB e + E, where A eB e ∈ T , and E ∈ A ⊥ e e e e e e = svdr (−Λ). e T , we have kEk < σr (AB). Conversely, kEk < σr (AB) and condition (b) imply AB e e e e e e Therefore, the dual conditions in (8) are equivalent to (1) Λ ∈ ∂H(AB) , Ψ; (2) PT (−Λ) = AB; (3) e < σr (A e B). e kPT ⊥ Λk To show the dual condition in Theorem 3.3, intuitively, we need to show that the angle θ between subspace T and Ψ is small (see Figure 3) for a specific function H(·). In the following (see Section E), we will demonstrate applications that, with randomness, obey this dual condition with high probability. 4 Matrix Completion In matrix completion, there is a hidden matrix X∗ ∈ Rn1 ×n2 with rank r. We are given measurements {X∗ij : (i, j) ∈ Ω}, where Ω ∼ Uniform(m), i.e., Ω is sampled uniformly at random from all subsets of [n1 ] × [n2 ] of cardinality m. The goal is to exactly recover X∗ with high probability. Here we apply our unified framework in Section 3 to matrix completion, by setting H(·) = I{M:PΩ (M)=PΩ (X∗ )} (·). A quantity governing the difficulties of matrix completion is the incoherence parameter µ. Intuitively, matrix completion is possible only if the information spreads evenly throughout the low-rank matrix. This intuition is captured by the incoherence conditions. Formally, denote by UΣVT the skinny SVD of a fixed 8 n1 × n2 matrix X of rank r. Candès et al. [CLMW11, CR09, Rec11, ZLZ16] introduced the µ-incoherence n(1) condition (3) to the low-rank matrix X. For conditions (3), it can be shown p that 1 ≤ µ ≤ r . The condition holds for many random matrices with incoherence parameter µ about r log n(1) [KMO10a]. We have two positive results. The first result is an information-theoretic upper bound: with the standard incoherence condition (3), X∗ is the unique matrix of rank at most r that is consistent with the observations. The proof is deferred to Appendix D. Theorem 4.1 (Information-Theoretic Upper Bound). Let Ω ∼ Uniform(m) be the support set uniformly distributed among all sets of cardinality m. Suppose that m ≥ cµn(1) r log n(1) for an absolute constant c. Then X∗ is the unique n1 × n2 matrix of rank at most r with µ-incoherence condition (3) such that PΩ (X) = PΩ (X∗ ), with probability at least 1 − n−10 (1) . Proof Sketch. We consider the feasibility of the matrix completion problem: Find a matrix X ∈ Rn1 ×n2 such that PΩ (X) = PΩ (X∗ ), rank(X) ≤ r. (9) Our proof first identifies a feasibility condition for problem (9), and then shows that X∗ is the only matrix which obeys this feasibility condition when the sample size is large enough. More specifically, we note that X∗ obeys the conditions in problem (9). Therefore, X∗ is the only matrix which obeys condition (9) if and only if X∗ + D does not follow the condition for all D, i.e., DS (X∗ ) ∩ Ω⊥ = {0}, where DS (X∗ ) is the descent cone of all low-rank matrices. We note that the descent cone DS (X∗ ) is contained in the subspace T by the tool of geometry functional analysis. Thus by a well-known fact that T ∩ Ω⊥ = {0} when the sample size is large, the proof is completed. We describe a simple finite-time inefficient algorithm given Theorem 4.1 in Section H. This positive result matches a lower bound from prior work, which claims that the sample complexity in Theorem 4.1 is optimal. Theorem 4.2 (Information-Theoretic Lower Bound. [CT10], Theorem 1.7). Denote by Ω ∼ Uniform(m) the support set uniformly distributed among all sets of cardinality m. Suppose that m ≤ cµn(1) r log n(1) for an absolute constant c. Then there exist infinitely many n1 × n2 matrices X0 of rank at most r obeying µ-incoherence (3) such that PΩ (X0 ) = PΩ (X∗ ), with probability at least 1 − n−10 (1) . Our second positive result converts the feasibility problem in Theorem 4.1 to a convex optimization problem, which can be efficiently solved. Theorem 4.3 (Efficient Matrix Completion). Let Ω ∼ Uniform(m) be the support set uniformly distributed among all sets of cardinality m. Suppose X∗ has condition number κ = σ1 (X∗ )/σr (X∗ ). Then there are absolute constants c and c0 such that with probability at least 1 − c0 n−10 (1) , the output of the convex problem e = argmin kXkr∗ , s.t. PΩ (X) = PΩ (X∗ ), X (10) X e = X∗ , provided that m ≥ cκ2 µrn(1) log2κ (n(1) ) log(n(1) ) and X∗ obeys is unique and exact, i.e., X µ-incoherence (3). e B) e = argminA,B 1 kABk2 , s.t.PΩ (AB) = Proof Sketch. We have shown in Theorem 4.1 that the problem (A, F 2 eB e = X∗ , with the optimal sample complexity. So if strong duality PΩ (X∗ ), exactly recovers X∗ , i.e., A holds, this non-convex optimization problem can be equivalently converted to the convex program (10). Then Theorem 4.3 is straightforward from strong duality. It now suffices to apply our unified framework in Section 3 to prove the strong duality. We show that e B) e be a the dual condition in Theorem 3.3 holds with high probability by the following arguments. Let (A, global solution to problem (10). For H(X) = I{M∈Rn1 ×n2 : PΩ M=PΩ X∗ } (X), we have e B) e = {G ∈ Rn1 ×n2 : hG, A e Bi e ≥ hG, Yi, for any Y ∈ Rn1 ×n2 s.t. PΩ Y = PΩ X∗ } Ψ = ∂H(A = {G ∈ Rn1 ×n2 : hG, X∗ i ≥ hG, Yi, for any Y ∈ Rn1 ×n2 s.t. PΩ Y = PΩ X∗ } = Ω, 9 eB e = X∗ . Then we only need to show where the third equality holds since A e ∈ Ω, (1) Λ e =A e B, e (2) PT (−Λ) e < 2 σr (A e B). e (3) kPT ⊥ Λk 3 (11) It is interesting to see that dual condition (11) can be satisfied if the angle θ between subspace Ω and subspace T is very small; see Figure 3. When the sample size \|Ω\| becomes larger and larger, the angle θ becomes smaller and smaller (e.g., when \|Ω\| = n1 n2 , the angle θ is zero as Ω = Rn1 ×n2 ). We show that the sample size m ≥ Ω(κ2 µrn(1) log2κ (n(1) ) log(n(1) )) is a sufficient condition for condition (11) to hold. 5 Robust Principal Component Analysis In this section, we develop our theory for robust PCA based on our framework. In the problem of robust PCA, we are given an observed matrix of the form D = X∗ + S∗ , where X∗ is the ground-truth matrix and S∗ is the corruption matrix which is sparse. The goal is to recover the hidden matrices X∗ and S∗ from the observation D. We set H(X) = λkD − Xk1 . To make the information spreads evenly throughout the matrix, the matrix cannot have one entry whose absolute value is significantly larger than other entries. For the robust PCA problem, Candès et al. [CLMW11] introduced an extra incoherence condition (Recall that X∗ = UΣVT is the skinny SVD of X∗ ) r µr T kUV k∞ ≤ . (12) n1 n2 In this work, we make the following incoherence assumption for robust PCA instead of (12): r µr ∗ σr (X∗ ). kX k∞ ≤ n1 n2 (13) Note that condition (13) is very similar to the incoherence condition (12) for the robust PCA problem, but the two notions are incomparable. Note that condition (13) has an intuitive explanation, namely, that the entries must scatter almost uniformly across the low-rank matrix. We have the following results for robust PCA. Theorem 5.1 (Robust PCA). Suppose X∗ is an n1 × n2 matrix of rank r, and obeys incoherence (3) and (13). Assume that the support set Ω of S∗ is uniformly distributed among all sets of cardinality m. Then with probability at least 1 − cn−10 (1) , the output of the optimization problem e S) e = argmin kXkr∗ + λkSk1 , (X, s.t. D = X + S, X,S with λ = σr (X∗ ) √ n(1) e = X∗ and S e = S∗ , provided that rank(X∗ ) ≤ ρr n(2) is exact, namely, X µ log2 n and m ≤ (1) ρs n1 n2 , where c, ρr , and ρs are all positive absolute constants, and function k · kr∗ is given by (14). The bounds on the rank of X∗ and the sparsity of S∗ in Theorem 5.1 match the best known results for robust PCA in prior work when we assume the support set of S∗ is sampled uniformly [CLMW11]. 6 Computational Aspects Computational Efficiency. We discuss our computational efficiency given that we have strong duality. We note that the dual and bi-dual of primal problem (P) are given by (see Appendix I) 10 (Dual, D1) (Bi-Dual, D2) 1 −H ∗ (Λ) − kΛk2r , 2 Λ∈Rn1 ×n2 max min M∈Rn1 ×n2 H(M) + kMkr∗ , where kΛk2r = r X σi2 (Λ), i=1 where kMkr∗ (14) 1 = maxhM, Xi − kXk2r . X 2 Problems (D1) and (D2) can be solved efficiently due to their convexity. In particular, Grussler et al. [GRG16] provided a computationally efficient algorithm to compute the proximal operators of functions 12 k · k2r and k · kr∗ . Hence, the Douglas-Rachford algorithm can find global minimum up to an error in function value in time poly(1/) [HY12]. Computational Lower Bounds. Unfortunately, strong duality does not always hold for general non-convex problems (P). Here we present a very strong lower bound based on the random 4-SAT hypothesis. This is by now a fairly standard conjecture in complexity theory [Fei02] and gives us constant factor inapproximability of problem (P) for deterministic algorithms, even those running in exponential time. If we additionally assume that BPP = P, where BPP is the class of problems which can be solved in probabilistic polynomial time, and P is the class of problems which can be solved in deterministic polynomial time, then the same conclusion holds for randomized algorithms. This is also a standard conjecture in complexity theory, as it is implied by the existence of certain strong pseudorandom generators or if any problem in deterministic exponential time has exponential size circuits [IW97]. Therefore, any subexponential time algorithm achieving a sufficiently small constant factor approximation to problem (P) in general would imply a major breakthrough in complexity theory. The lower bound is proved by a reduction from the Maximum Edge Biclique problem [AMS11]. The details are presented in Appendix G. Theorem 6.1 (Computational Lower Bound). Assume Conjecture 1 (the hardness of Random 4-SAT). Then there exists an absolute constant 0 > 0 for which any deterministic algorithm achieving (1 + )OPT in the objective function value for problem (P) with ≤ 0 , requires 2Ω(n1 +n2 ) time, where OPT is the optimum. If in addition, BPP = P, then the same conclusion holds for randomized algorithms succeeding with probability at least 2/3. Proof Sketch. Theorem 6.1 is proved by using the hypothesis that random 4-SAT is hard to show hardness of the Maximum Edge Biclique problem for deterministic algorithms. We then do a reduction from the Maximum Edge Biclique problem to our problem. The complete proofs of other theorems/lemmas and related work can be found in the appendices. Acknowledgments. We thank Rong Ge, Jason D. Lee, Zhouchen Lin, Tengyu Ma, Benjamin Recht and Tuo Zhao for useful discussions. This work was supported in part by NSF grants NSF CCF-1422910, NSF CCF-1535967, NSF CCF-1451177, NSF IIS-1618714, NSF CCF-1527371, a Sloan Research Fellowship, a Microsoft Research Faculty Fellowship, DMS-1317308, Simons Investigator Award, Simons Collaboration Grant, and ONR-N00014-16-1-2329. We would like to thank Rina Foygel for finding a bug in the proof of Theorem 4.3 in a previous version. 11 A Other Related Work Non-convex matrix factorization is a popular topic studied in theoretical computer science [JNS13, Har14, SL15, RSW16], machine learning [BNS16, GLM16, GHJY15, JMD10, LLR16], and optimization [WYZ12, SWZ14]. We review several lines of research on studying the global optimality of such optimization problems. Global Optimality of Matrix Factorization. While lots of matrix factorization problems have been shown to have no spurious local minima, they either require additional conditions on the local minima, or are based on particular forms of the objective function. Specifically, Burer and Monteiro [BM05] showed that one can minimize F (AAT ) for any convex function F by solving for A directly without introducing any local minima, provided that the rank of the output A is larger than the rank of the true minimizer Xtrue . However, such a condition is often impossible to check as rank(Xtrue ) is typically unknown a priori. To resolve the issue, Bach et al. [BMP08] and Journée et al. [JBAS10] proved that X = AAT is a global minimizer of F (X), if A is a rank-deficient local minimizer of F (AAT ) and F (X) is a twice differentiable convex function. Haeffele and Vidal [HV15] further extended this result by allowing a more general form of objective function F (X) = G(X) + H(X), where G is a twice differentiable convex function with compact level set and H is a proper convex function such that F is lower semi-continuous. However, a major drawback of this line of research is that these result fails when the local minimizer is of full rank. Matrix Completion. Matrix completion is a prototypical example of matrix factorization. One line of work on matrix completion builds on convex relaxation (e.g., [SS05, CR09, CT10, Rec11, CRPW12, NW12]), which does not achieve the optimal sample complexity. Recently, Ge et al. [GLM16] showed that matrix completion has no spurious local optimum, when \|Ω\| is sufficiently large and the matrix Y is incoherent. The result is only for positive semi-definite matrices and their sample complexity is not optimal. Another line of work is built upon good initialization for global convergence. Recent attempts showed that one can first compute some form of initialization (e.g., by singular value decomposition) that is close to the global minimizer and then use non-convex approaches to reach global optimality, such as alternating minimization, block coordinate descent, and gradient descent [KMO10b, KMO10a, JNS13, Kes12, Har14, BKS16, ZL15, ZWL15, TBSR15, CW15, SL15]. In our result, in contrast, we can reformulate non-convex matrix completion problems as equivalent convex programs, which guarantees global convergence from any initialization. Robust PCA. Robust PCA is also a prototypical example of matrix factorization. The goal is to recover both the low-rank and the sparse components exactly from their superposition [CLMW11, NNS+ 14, GWL16, ZLZC15, ZLZ16, YPCC16]. It has been widely applied to various tasks, such as video denoising, background modeling, image alignment, photometric stereo, texture representation, subspace clustering, and spectral clustering. There are typically two settings in the robust PCA literature: a) the support set of the sparse matrix is uniformly sampled [CLMW11, ZLZ16]; b) the support set of the sparse matrix is deterministic, but the non-zero entries in each row or column of the matrix cannot be too large [YPCC16, GJY17]. In this work, we discuss the first case. Our framework provides results that match the best known work in setting (b) [CLMW11]. Other Matrix Factorization Problems. Matrix sensing is another typical matrix factorization problem [CRPW12, JNS13, ZWL15]. Bhojanapalli et al. [BNS16] and Tu et al. [TBSR15] showed that the matrix recovery model minA,B 12 kA(AB − Y)k2F , achieves optimality for every local minimum, if the operator A satisfies the restricted isometry property. They further gave a lower bound and showed that the unstructured operator A may easily lead to a local minimum which is not globally optimal. Some other matrix factorization problems are also shown to have nice geometric properties such as the property that all local minima are global minima. Examples include dictionary learning [SQW17a], phase 12 retrieval [SQW16], and linear deep neural networks Q [Kaw16]. In multi-layer linear neural networks where the ∗ goal is to learn a multi-linear projection X = i Wi , each Wi represents the weight matrix that connects the hidden units in the i-th and (i + 1)-th layers. The study of such linear models is central to the theoretical understanding of the loss surface of deep neural networks with non-linear activation functions [Kaw16, CHM+ 15]. In dictionary learning, we aim to recover a complete (i.e., square and invertible) dictionary matrix A from a given signal X in the form of X = AB, provided that the representation coefficient B is sufficiently sparse. This problem centers around solving a non-convex matrix factorization problem with a sparsity constraint on the representation coefficient B [BMP08, SQW17a, SQW17b, ABGM14]. Other high-impact examples of matrix factorization models range from the classic unsupervised learning problems like PCA, independent component analysis, and clustering, to the more recent problems such as non-negative matrix factorization, weighted low-rank matrix approximation, sparse coding, tensor decomposition [BCMV14, AGH+ 14], subspace clustering [ZLZG15, ZLZG14], etc. Applying our framework to these other problems is left for future work. Atomic Norms. The atomic norm is a recently proposed function for linear inverse problems [CRPW12]. Many well-known norms, e.g., the `1 norm and the nuclear norm, serve as special cases of atomic norms. It has been widely applied to the problems of compressed sensing [TBSR13], low-rank matrix recovery [CR13], blind deconvolution [ARR14], etc. The norm is defined by the Minkowski functional associated with the convex hull of a set A: kXkA = inf{t > 0 : X ∈ tA}. In particular, if we set A to be the convex hull of the infinite set of unit-`2 -norm rank-one matrices, then k · kA equals to the nuclear norm. We mention that our objective term kABkF in problem (1) is similar to the atomic norm, but with slight differences: unlike the atomic norm, we set A to be the infinite set of unit-`2 -norm rank-r matrices for rank(X) ≤ r. With this, we achieve better sample complexity guarantees than the atomic-norm based methods. B Proof of Lemma 3.1 b any local minimum of f (A, B) = 1 kY b − ABk2 is globally Lemma 3.1 (Restated). For any given matrix Y, F 2 optimal. The objective function f (A, B) around any saddle point has a negative second-order directional curvature. Moreover, f (A, B) has no local maximum. Proof. (A,B) is a critical point of f (A, B) if and only if ∇A f (A, B) = 0 and ∇B f (A, B) = 0, or equivalently, b T and AT AB = AT Y. b ABBT = YB (15) Note that for any fixed matrix A (resp. B), the function f (A, B) is convex in the coefficients of B (resp. A). To prove the desired lemma, we have the following claim. Claim 1. If two matrices A and B define a critical point of f (A, B), then the global mapping M = AB is of the form b M = PA Y, with A satisfying bY b T = AA† Y bY b T AA† = Y bY b T AA† . AA† Y (16) Proof. If A and B define a critical point of f (A, B), then (15) holds and the general solution to (15) satisfies B = (AT A)† AT Y + (I − A† A)L, (17) b = AA† Y b = PA Y b by the property of the Moorefor any matrix L. So M = AB = A(AT A)† AT Y † T † T Penrose pseudo-inverse: A = (A A) A . 13 By (15), we also have b T AT ABBT AT = YB b T. or equivalently MMT = YM b (B) can be rewritten as Plugging in the relation M = AA† Y, bY b T AA† = Y bY b T AA† . AA† Y bY b T AA† is symmetric. Thus Note that the matrix AA† Y bY b T = AA† Y bY b T AA† , AA† Y as desired. To prove Lemma 3.1, we also need the following claim. Claim 2. Denote by I = {i1 , i2 , ..., ir } any ordered r-index set (ordered by λij , j ∈ [r] from the largest bY b T ∈ Rn1 ×n1 with p distinct values. Let UI = to the smallest) and λi , i ∈ [n1 ], the eigenvalues of Y bY bT ∈ [ui1 , ui2 , ..., uir ] denote the matrix formed by the orthonormal eigenvectors U = [u1 , u2 , ..., un1 ] of Y n ×n 1 1 R associated with the ordered p eigenvalues, whose multiplicities are m1 , m2 , ..., mp (m1 + m2 + ... + mp = n1 ). Then two matrices A and B define a critical point of f (A, B) if and only if there exists an ordered r-index set I, an invertible matrix C, and an r × n matrix L such that b + (I − A† A)L, A = (UD):I C and B = A† Y (18) where D is a p-block-diagonal matrix with each block equal to the orthogonal projector of dimension mi . For such a critical point, we have b AB = PA Y, 1 f (A, B) = 2 ! bY bT) − tr(Y X i∈I λi = 1X λi . 2 (19) i6∈I bY b T is a real symmetric covariance matrix. So it can always be represented as UΛUT , Proof. Note that Y bY b T and Λ ∈ Rn1 ×n1 is a where U ∈ Rn1 ×n1 is an orthonormal matrix consisting of eigenvectors of Y T b b diagonal matrix with non-increasing eigenvalues of YY . If A and B satisfy (18) for some C, L, and I, then b T ABBT = YB and b AT AB = AT Y, which is (15). So A and B define a critical point of f (A, B). For the converse, notice that PUT A = UT A(UT A)† = UT AA† U = UT PA U, or equivalently, PA = UPUT A UT . Thus (16) yields UPUT A UT UΛUT = UΛUT UPUT A UT , or equivalently, PUT A Λ = ΛPUT A . Notice that Λ ∈ Rn1 ×n1 is a diagonal matrix with p distinct eigenvalues bY b T . So PUT A is a block-diagonal matrix with p blocks, each of which is an orthogonal projector of of Y 14 dimension mi , corresponding to the eigenvalues λi , i ∈ [p]. Therefore, there exists an index set I such that PUT A = D:I DT:I , where D is a block-diagonal matrix. It follows that PA = UPUT A UT = UD:I DT:I UT = (UD):I (UD)T:I . Since the column space of A coincides with the column space of (UD):I , A is of the form A = (UD):I C, b + A(I − A† A)L = PA Y b and and B is given by (17). Thus AB = A(AT A)† AT Y 1 b f (A, B) = kY − ABk2F 2 1 b b 2 − PA Yk = kY F 2 1 b 2 = kPA⊥ Yk F 2 X 1 = λi . 2 i6∈I So the local minimizer of f (A, B) is given by (18) with I = Φ, which is globally optimal according to bY b T . We then show that when I (19), where Φ is the index set corresponding to the r largest eigenvalues of Y consists of other combinations of indices of eigenvalues, i.e., I = 6 Φ, the corresponding pair (A, B) given by (18) is a strict saddle point. Claim 3. If I = 6 Φ, then the pair (A, B) given by (18) is a strict saddle point. Proof. Let i ∈ Φ but i 6∈ I, and denote by UD = R. It is enough to slightly perturb the column space of A towards the direction of an eigenvector of λi . More precisely, fix two indices i and j such that i ∈ Φ, e :j = (1 + 2 )−1/2 (R:j + R:i ). Notice that i 6∈ I. Thus i 6∈ I, and j is the largest index in I. For any , let R eTR e e e e e† e†e R :j :j = I. Let A = RI C and B = A Y + (I − A A)L. A direct calculation shows that e B) e = f (A, B) − 2 (λi − λj )/(2 + 22 ). f (A, Hence, e B) e − f (A, B) f (A, 1 = − (λi − λj ) < 0. 2 →0 2 lim Note that all critical points of f (A, B) are in the form of (18), and if I 6= Φ, the pair (A, B) given by (18) is a strict saddle point, while if I = Φ, then the pair (A, B) given by (18) is a local minimum. We conclude that f (A, B) has no local maximum. The proof is completed. C Proof of Lemma 3.2 e B) e be a global minimizer of F (A, B). If there exists a dual certificate Λ e Lemma 3.2 (Restated). Let (A, e e e e as in Condition 1 such that the pair (A, B) is a local minimizer of L(A, B, Λ) for the fixed Λ, then strong eB e = svdr (−Λ). e duality holds. Moreover, we have the relation A 15 e − ABk2 + e B) e is a local minimizer of L(A, B, Λ) e = 1k − Λ Proof. By the assumption of the lemma, (A, F 2 e where c(Λ) e is a function that is independent of A and B. So according to Lemma 3.1, (A, e B) e = c(Λ), e namely, (A, e B) e globally minimizes L(A, B, Λ) when Λ is fixed to Λ. e Furthermore, argminA,B L(A, B, Λ), ∗ e ∈ ∂X H(X)\| e e implies that A eB e ∈ ∂Λ H (Λ)\| e by the convexity of function H, meaning that Λ X=AB Λ=Λ e B, e Λ). So Λ e = argmaxΛ L(A, e B, e Λ) due to the concavity of L(A, e B, e Λ) w.r.t. variable Λ. 0 ∈ ∂Λ L(A, e e e Thus (A, B, Λ) is a primal-dual saddle point of L(A, B, Λ). e = We now prove the strong duality. By the fact that F (A, B) = maxΛ L(A, B, Λ) and that Λ e B, e Λ), we have argmaxΛ L(A, e B) e = L(A, e B, e Λ) e ≤ L(A, B, Λ), e ∀A, B. F (A, e B, e Λ) e is a primal-dual saddle point of L. So on the one hand, we where the inequality holds because (A, have e B) e ≤ min L(A, B, Λ) e ≤ max min L(A, B, Λ). min max L(A, B, Λ) = F (A, A,B Λ A,B Λ A,B On the other hand, by weak duality, min max L(A, B, Λ) ≥ max min L(A, B, Λ). A,B Λ Λ A,B Therefore, minA,B maxΛ L(A, B, Λ) = maxΛ minA,B L(A, B, Λ), i.e., strong duality holds. Hence, eB e = argmin L(A, B, Λ) e A AB 1 e − ABk2 − 1 kΛk e 2 − H ∗ (Λ) e = argmin k − Λ F F 2 2 AB 1 e − ABk2 = argmin k − Λ F 2 AB e = svdr (−Λ), as desired. D Proof of Theorem 4.1 Theorem 4.1 (Information-Theoretic Upper Bound. Restated). Let Ω ∼ Uniform(m) be the support set, which is uniformly distributed among all sets of cardinality m. Suppose that m ≥ cµn(1) r log n(1) for an absolute constant c. Then X∗ is the unique n1 × n2 matrix of rank at most r with µ-incoherence (3) such that PΩ (X) = PΩ (X∗ ), with probability at least 1 − n−10 (1) . Proof. We note that the sampling model Uniform(m) is equivalent to the sampling model Ber(p) with m p = Θ n1 n2 , which we will frequently use in the sequel (see Appendix J). We consider the feasibility of the matrix completion problem: Find a matrix X ∈ Rn1 ×n2 such that PΩ (X) = PΩ (X∗ ), rank(X) ≤ r. (20) Our proof first identifies a feasibility condition for problem (20), and then shows that X∗ is the only matrix that obeys this feasibility condition when the sample size is large enough. We denote by S = {X ∈ Rn1 ×n2 : rank(X) ≤ r}, 16 and define DS (X∗ ) = {t(X − X∗ ) ∈ Rn1 ×n2 : rank(X) ≤ r, t ≥ 0}. We have the following proposition for the feasibility of problem (20). Proposition D.1 (Feasibility Condition). X∗ is the unique feasible solution to problem (20) if DS (X∗ ) ∩ Ω⊥ = {0}. Proof. Notice that problem (20) is equivalent to another feasibility problem Find a matrix D ∈ Rn1 ×n2 such that rank(X∗ + D) ≤ r, D ∈ Ω⊥ . Suppose that DS (X∗ ) ∩ Ω⊥ = {0}. Since rank(X∗ + D) ≤ r implies D ∈ DS (X∗ ), and note that D ∈ Ω⊥ , we have D = 0, which means X∗ is the unique feasible solution to problem (20). The remainder of the proof is to show DS (X∗ ) ∩ Ω⊥ = {0}. To proceed, we note that the “escaping through a mesh” techniques for matrix sensing do not work for matrix completion since Ω is not drawn from the Grassmanian according to the Haar measure. To address this issue, we instead need the following lemmas. The first lemma claims that the tangent cone of the set S evaluated at X∗ is slightly larger than the cone cone(S − {X∗ }). Lemma D.2 ([Jah07], Theorem 4.8). Let S be a non-empty subset of a real normed space. If S is star-shaped w.r.t. some X∗ ∈ S, i.e., t(S − {X∗ }) ⊆ S − {X∗ } for all t ∈ [0, 1], then it follows cone(S − {X∗ }) ⊆ T (S, X∗ ), where T (S, X∗ ) is the tangent cone of the set S at point X∗ defined by T (S, X∗ ) = {Ξ ∈ Rn1 ×n2 : ∃Xn ⊆ S, (an ) ⊆ R+ s.t. Xn → X∗ , an (Xn − X∗ ) → Ξ}. The second lemma states that the tangent cone of the set S evaluated at X∗ can be represented in a closed form. Lemma D.3 ([SU15], Theorem 3.2). Let X∗ = UΣVT be the skinny SVD of matrix X∗ . The tangent cone T (S, X∗ ) of the set S = {X ∈ Rn1 ×n2 : rank(X) ≤ r} at X∗ is a linear subspace given by T (S, X∗ ) = {ULT + MVT : L ∈ Rn2 ×r , M ∈ Rn1 ×r } , T . Now we are ready to prove Theorem 4.1. By Lemma D.2 and D.3, we have DS (X∗ ) = cone(S − {X∗ }) ⊆ T (S, X∗ ) = T , where the first equality holds by the definition of DS (X∗ ). So if T ∩ Ω⊥ = {0}, then DS (X∗ ) ∩ Ω⊥ = {0}, meaning that X∗ is the unique feasible solution to the problem (20). Thus the rest of proof is to find a sufficient condition for T ∩ Ω⊥ = {0}. We have the following lemma. Lemma D.4. Assume that Ω ∼ Ber(p) and the incoherence condition (3) holds. Then with probability at √ least 1 − n−10 1 − p + p, provided that p ≥ C0 −2 (µr log n(1) )/n(2) , where C0 (1) , we have kPΩ⊥ PT k ≤ is an absolute constant. 17 Proof. If Ω ∼ Ber(p), we have, by Theorem E.2, that with high probability kPT − p−1 PT PΩ PT k ≤ , provided that p ≥ C0 −2 µr log n(1) . n(2) Note, however, that since I = PΩ + PΩ⊥ , PT − p−1 PT PΩ PT = p−1 (PT PΩ⊥ PT − (1 − p)PT ) and, therefore, by the triangle inequality kPT PΩ⊥ PT k ≤ p + (1 − p). Since kPΩ⊥ PT k2 ≤ kPT PΩ⊥ PT k, the proof is completed. We note that kPΩ⊥ PT k < 1 implies Ω⊥ ∩ T = {0}. The proof is completed. E Proof of Theorem 4.3 e B) e = argminA,B 1 kABk2 , s.t.PΩ (AB) = PΩ (X∗ ), We have shown in Theorem 4.1 that the problem (A, F 2 eB e = X∗ , with the optimal sample complexity. So if strong duality holds, this exactly recovers X∗ , i.e., A non-convex optimization problem can be equivalently converted to the convex program (10). Then Theorem 4.3 is straightforward from strong duality. It now suffices to apply our unified framework in Section 3 to prove the strong duality. We show that the e B) e be a global solution to problem (10). dual condition in Theorem 3.3 holds with high probability. Let (A, For H(X) = I{M∈Rn1 ×n2 : PΩ M=PΩ X∗ } (X), we have e B) e = {G ∈ Rn1 ×n2 : hG, A e Bi e ≥ hG, Yi, for any Y ∈ Rn1 ×n2 s.t. PΩ Y = PΩ X∗ } Ψ = ∂H(A = {G ∈ Rn1 ×n2 : hG, X∗ i ≥ hG, Yi, for any Y ∈ Rn1 ×n2 s.t. PΩ Y = PΩ X∗ } = Ω, eB e = X∗ . Then we only need to show where the third equality holds since A (1) (2) (3) e ∈ Ω, Λ e =A e B, e PT (−Λ) (21) e < 2 σr (A e B). e kPT ⊥ Λk 3 We have the following lemma. Lemma E.1. If we can construct an Λ such that (a) Λ ∈ Ω, (b) e Bk e F ≤ kPT (−Λ) − A (c) 1 e B), e kPT ⊥ Λk < σr (A 3 s r e B), e σr (A 3n2(1) e such that Eqn. (21) holds with probability at least 1 − n−10 . then we can construct an Λ (1) Proof. To prove the lemma, we first claim the following theorem. 18 (22) Theorem E.2 ([CR09], Theorem 4.1). Assume that Ω is sampled according to the Bernoulli model with success probability p = Θ( n1mn2 ), and incoherence condition (3) holds. Then there is an absolute constant CR such that for β > 1, we have r βµn(1) r log n(1) −1 kp PT PΩ PT − PT k ≤ CR , , m q βµn(1) r log n(1) −β with probability at least 1 − 3n provided that CR < 1. m e − Λ ∈ Ω be the perturbation matrix between Λ and Λ e Suppose that Condition (22) holds. Let Y = Λ −1 e e e e e such that PT q (−Λ) = AB. Such a Y exists by setting Y = PΩ PT (PT PΩ PT ) (PT (−Λ) − AB). So e B). e We now prove Condition (3) in Eqn. (21). Observe that kPT YkF ≤ 3nr2 σr (A (1) e ≤ kPT ⊥ Λk + kPT ⊥ Yk kPT ⊥ Λk 1 e B) e + kPT ⊥ Yk. ≤ σr (A 3 (23) e B). e So we only need to show kPT ⊥ Yk ≤ 13 σr (A Before proceeding, we begin by introducing a normalized version QΩ : Rn1 ×n2 → Rn1 ×n2 of PΩ : QΩ = p−1 PΩ − I. With this, we have PT PΩ PT = pPT (I + QΩ )PT . Note that for any operator P : T → T , we have X P −1 = (PT − P)k whenever kPT − Pk < 1. k≥0 So according to Theorem E.2, the operator p(PT PΩ PT )−1 can be represented as a convergent Neumann series X p(PT PΩ PT )−1 = (−1)k (PT QΩ PT )k , k≥0 because kPT QΩ PT k ≤ < also note that 1 2 once m ≥ Cµn(1) r log n(1) for a sufficiently large absolute constant C. We p(PT ⊥ QΩ PT ) = PT ⊥ PΩ PT , 19 because PT ⊥ PT = 0. Thus e B))k e kPT ⊥ Yk = kPT ⊥ PΩ PT (PT PΩ PT )−1 (PT (−Λ) − A e B))k e = kPT ⊥ QΩ PT p(PT PΩ PT )−1 ((PT (−Λ) − A X e B))k e =k (−1)k PT ⊥ QΩ (PT QΩ PT )k ((PT (−Λ) − A k≥0 ≤ X e B))k e F k(−1)k PT ⊥ QΩ (PT QΩ PT )k ((PT (−Λ) − A k≥0 ≤ kQΩ k X e B))k e F kPT QΩ PT kk kPT (−Λ) − A k≥0 4 e B)k e F kPT (−Λ) − A p n n s r 1 2 e B) e ≤Θ σr (A m 3n2(1) ≤ 1 e B) e ≤ σr (A 3 with high probability. The proof is completed. It thus suffices to construct a dual certificate Λ such that all conditions in (22) hold. To this end, partition Ω = Ω1 ∪ Ω2 ∪ ... ∪ Ωb into b partitions of size q. By assumption, we may choose 128 1 q≥ Cβκ2 µrn(1) log n(1) and b ≥ log2κ 242 n2(1) κ2 3 2 for a sufficiently large constant C. Let Ωj ∼ Ber(q) denote the set of indices corresponding to the jeB e and set Λk = n1 n2 Pk PΩ (Wj−1 ), Wk = A eB e − PT (Λk ) for th partitions. Define W0 = A j j=1 q k = 1, 2, ..., b. Then by Theorem E.2, n1 n2 n1 n2 kWk kF = Wk−1 − PT PΩk (Wk−1 ) = PT PΩk PT (Wk−1 ) PT − q q F F 1 ≤ kWk−1 kF . 2κ q r √ e B) e ≤ ee So it follows that kWb kF ≤ (2κ)−b kW0 kF ≤ (2κ)−b rσ1 (A 2 2 σr (AB). 24 n(1) The following lemma together implies the strong duality of (10) straightforwardly. Lemma E.3. Under the assumptions of Theorem 4.3, the dual certification Wb obeys the dual condition (22) with probability at least 1 − n−10 (1) . Proof. It is well known that for matrix completion, the Uniform model Ω ∼ Uniform(m) is equivalent to the Bernoulli model Ω ∼ Ber(p), where each element in [n1 ] × [n2 ] is included with probability p = Θ(m/(n1 n2 )) independently; see Section J for a brief justification. By the equivalence, we can suppose Ω ∼ Ber(p). To prove Lemma E.3, as a preliminary, we need the following lemmas. Lemma E.4 ([Che15], Lemma 2). Suppose Z is a fixed matrix. Suppose Ω ∼ Ber(p). Then with high probability,   s log n(1) log n(1) k(I − p−1 PΩ )Zk ≤ C00  kZk∞ + kZk∞,2  , p p 20 where C00 > 0 is an absolute constant and   s sX   X 2 2 = max max Zib , max Zaj . j  i  a kZk∞,2 b Lemma E.5 ([CLMW11], Lemma 3.1). Suppose Ω ∼ Ber(p) and Z is a fixed matrix. Then with high probability, kZ − p−1 PT PΩ Zk∞ ≤ kZk∞ , provided that p ≥ C0 −2 (µr log n(1) )/n(2) for some absolute constant C0 > 0. Lemma E.6 ([Che15], Lemma 3). Suppose that Z is a fixed matrix and Ω ∼ Ber(p). If p ≥ c0 µr log n(1) /n(2) for some c0 sufficiently large, then with high probability, r 1 n(1) 1 k(p−1 PT PΩ − PT )Zk∞,2 ≤ kZk∞ + kZk∞,2 . 2 µr 2 Observe that by Lemma E.5, kWj k∞ j 1 e Bk e ∞, kA ≤ 2 and by Lemma E.6, kWj k∞,2 ≤ 1 2 r n(1) 1 kWj−1 k∞ + kWj−1 k∞,2 . µr 2 So kWj k∞,2 j r n(1) 1 e Bk e ∞ + 1 kWj−1 k∞,2 ≤ kA 2 µr 2 j r j n(1) 1 1 e e e Bk e ∞,2 . ≤j kABk∞ + kA 2 µr 2 Therefore, kPT ⊥ Λb k ≤ b X n1 n2 k PT ⊥ PΩj Wj−1 k q j=1 = b X kPT ⊥ ( j=1 ≤ b X j=1 k( n1 n2 PΩj Wj−1 − Wj−1 )k q n1 n2 PΩj − I)(Wj−1 )k. q 21 Let p denote Θ q n1 n2 . By Lemma E.4, kPT ⊥ Λb k b X 0 log n(1) ≤ C0 kWj−1 k∞ + C00 p s b log n(1) X kWj−1 k∞,2 p j=1 j=1 s # " r j b j b j X X log n n log n 1 1 1 (1) (1) e e (1) e Bk e ∞ + C0 e Bk e ∞,2 ≤ C00 kA j kABk∞ + kA 0 p 2 p 2 µr 2 j=1 j=1 s s r log n log n n log n(1) (1) e e (1) (1) e e e Bk e ∞,2 . ≤ C00 kABk∞ + 2C00 kABk∞ + C00 kA p p µr p eB e = X∗ , we note the facts that (we assume WLOG n2 ≥ n1 ) Setting A r r µr µr ∗ ∗ T T T ∗ σ1 (X ) ≤ κσr (X∗ ), kX k∞,2 = max kei UΣV k2 ≤ max kei Ukσ1 (X ) ≤ i i n1 n1 and that kX∗ k∞ = maxhX∗ , ei eTj i = maxhUΣVT , ei eTj i = maxheTi UΣ, eTj Vi ij ≤ Substituting p = Θ F ij ij max keTi UΣVT k2 keTj Vk2 ij κ2 µrn(1) log(n(1) ) log2κ (n(1) ) n1 n2 µrκ σr (X∗ ). ≤ max kX k∞,2 keTj Vk2 ≤ √ j n1 n2 ∗ e < 1 σr (X∗ ). The proof is completed. , we obtain kPT ⊥ Λk 3 Proof of Theorem 5.1 Theorem 5.1 (Robust PCA. Restated). Suppose X∗ is n1 × n2 , obeys incoherence (3) and (13). Assume that the support set Ω of S∗ is uniformly distributed among all sets of cardinality m. Then with probability at least 1 − cn−10 (1) , the output of the optimization problem e S) e = argmin kXkr∗ + λkSk1 , (X, s.t. D = X + S, (24) X,S with λ = σr (X∗ ) √ n(1) n (2) e = X∗ and S e = S∗ , provided that rank(X∗ ) ≤ ρr is exact, namely, X µ log2 n and m ≤ (1) ρs n1 n2 , where c, ρr , and ρs are all positive absolute constants, and function k · kr∗ is given by (14). F.1 Subgradient of the r∗ Function Lemma F.1. Let UΣVT be the skinny SVD of matrix X∗ of rank r. The subdifferential of k · kr∗ evaluated at X∗ is given by ∂kX∗ kr∗ = {X∗ + W : UT W = 0, WV = 0, kWk ≤ σr (X∗ )}. Proof. Note that for any fixed function f (·), the set of all optimal solutions of the problem f ∗ (X∗ ) = maxhX∗ , Yi − f (Y) Y 22 (25) form the subdifferential of the conjugate function f ∗ (·) evaluated at X∗ . Set f (·) to be 12 k · k2r and notice that the function 12 k · k2r is unitarily invariant. By Von Neumann’s trace inequality, the optimal solutions to problem (25) are given by [U, U⊥ ]Diag([σ1 (Y), ..., σr (Y), σr+1 (Y), ..., σn(2) (Y)])[V, V⊥ ]T , where n(2) {σi (Y)}i=r+1 can be any value no larger than σr (Y) and {σi (Y)}ri=1 are given by the optimal solution to the problem r r X 1X 2 ∗ max σi (X )σi (Y) − σi (Y). 2 {σi (Y)}ri=1 i=1 The solution is unique such that σi (Y) = σi F.2 (X∗ ), i=1 i = 1, 2, ..., r. The proof is complete. Dual Certificates Lemma F.2. Assume that kPΩ PT k ≤ 1/2 and λ < σr (X∗ ). Then (X∗ , S∗ ) is the unique solution to problem (5.1) if there exists a pair (W, F) for which X∗ + W = λ(sign(S∗ ) + F + PΩ K), where W ∈ T ⊥ , kWk ≤ σr (X∗ ) , 2 F ∈ Ω⊥ , kFk∞ ≤ 12 , and kPΩ KkF ≤ 14 . Proof. Let (X∗ + H, S∗ − H) be any optimal solution to problem (24). By the definition of the subgradient, the inequality follows kX∗ + Hkr∗ + λkS∗ − Hk1 ≥ kX∗ kr∗ + λkS∗ k1 + hX∗ + W∗ , Hi − λhsign(S∗ ) + F∗ , Hi = kX∗ kr∗ + λkS∗ k1 + hX∗ − λsign(S∗ ), Hi + hW∗ , Hi − λhF∗ , Hi = kX∗ kr∗ + λkS∗ k1 + hX∗ − λsign(S∗ ), Hi + σr (X∗ )kPT ⊥ Hk∗ + λkPΩ⊥ Hk1 = kX∗ kr∗ + λkS∗ k1 + hλF + λPΩ K − W, Hi + σr (X∗ )kPT ⊥ Hk∗ + λkPΩ⊥ Hk1 λ λ σr (X∗ ) = kX∗ kr∗ + λkS∗ k1 + kPT ⊥ Hk∗ + kPΩ⊥ Hk1 − kPΩ HkF . 2 2 4 We note that kPΩ HkF ≤ kPΩ PT HkF + kPΩ PT ⊥ HkF 1 ≤ kHkF + kPT ⊥ HkF 2 1 1 ≤ kPΩ HkF + kPΩ⊥ HkF + kPT ⊥ HkF , 2 2 which implies that λ4 kPΩ HkF ≤ λ4 kPΩ⊥ HkF + λ2 kPT ⊥ HkF ≤ λ4 kPΩ⊥ Hk1 + λ2 kPT ⊥ Hk∗ . Therefore, λ σr (X∗ ) − λ kPT ⊥ Hk∗ + kPΩ⊥ Hk1 2 4 ∗) − λ σ (X λ r ≥ kX∗ + Hkr∗ + λkS∗ − Hk1 + kPT ⊥ Hk∗ + kPΩ⊥ Hk1 , 2 4 where the second inequality holds because (X∗ + H, S∗ − H) is optimal. Thus H ∈ T ∩ Ω. Note that kPΩ PT k < 1 implies T ∩ Ω = {0}. This completes the proof. kX∗ + Hkr∗ + λkS∗ − Hk1 ≥ kX∗ kr∗ + λkS∗ k1 + According to Lemma F.2, to show the exact recoverability of problem (24), it is sufficient to find an appropriate W for which   W ∈ T ⊥,    kWk ≤ σr (X∗ ) , 2 (26) ∗ ∗ ))k ≤ λ ,  kP (X + W − λsign(S Ω F  4   kP ⊥ (X∗ + W)k ≤ λ . ∞ Ω 2 23 F.3 Dual Certification by Least Squares and the Golfing Scheme The remainder of the proof is to construct W such that the dual condition (26) holds true. Before introducing our construction, we assume Ω ∼ Ber(p), or equivalently Ω⊥ ∼ Ber(1 − p), where p is allowed be as large as an absolute constant. Note that Ω⊥ has the same distribution as that of Ω1 ∪ Ω2 ∪ ... ∪ Ωj0 , where the Ωj ’s are drawn independently with replacement from Ber(q), j0 = dlog n(1) e, and q obeys p = (1 − q)j0 (q = Ω(1/ log n(1) ) implies p = O(1)). We construct W based on such a distribution. Our construction separates W into two terms: W = WL + WS . To construct WL , we apply the golfing scheme introduced by [Gro11, Rec11]. Specifically, WL is constructed by an inductive procedure: Yj = Yj−1 + q −1 PΩj PT (X∗ − Yj−1 ), Y0 = 0, WL = PT ⊥ Yj0 . To construct WS , we apply the method of least squares by [CLMW11], which is X WS = λPT ⊥ (PΩ PT PΩ )k sign(S∗ ). (27) (28) k≥0 Note that kPΩ PT k ≤ 1/2. Thus kPΩ PT PΩ k ≤ 1/4 and the Neumann series in (28) is well-defined. Observe that PΩ WS = λ(PΩ − PΩ PT PΩ )(PΩ − PΩ PT PΩ )−1 sign(S∗ ) = λsign(S∗ ). So to prove the dual condition (26), it suffices to show that σr (X∗ ) , 4 (a) kWL k ≤ λ , 4 λ ≤ , 4 (b) kPΩ (X∗ + WL )kF ≤ (c) kPΩ⊥ (X∗ + WL )k∞ σr (X∗ ) , 4 λ (e) kPΩ⊥ WS k∞ ≤ . 4 (29) (d) kWS k ≤ F.4 (30) Proof of Dual Conditions Since we have constructed the dual certificate W, the remainder is to show that W obeys dual conditions (29) and (30) with high probability. We have the following. Lemma F.3. Assume Ωj ∼ Ber(q), j = 1, 2, ..., j0 , and j0 = 2dlog n(1) e. Then under the other assumptions of Theorem 5.1, WL given by (27) obeys dual condition (29). Proof. Let Zj = PT (X∗ − Yj ) ∈ T . Then we have Zj = PT Zj−1 − q −1 PT PΩj PT Zj−1 = (PT − q −1 PT PΩj PT )Zj−1 , and Yj = Pj k=1 q −1 P Z Ωk k−1 ∈ Ω⊥ . We set q = Ω(−2 µr log n(1) /n(2) ). 24 Proof of (a). It holds that L kW k = kPT ⊥ Yj0 k ≤ j0 X kq −1 PT ⊥ PΩk Zk−1 k k=1 = ≤ j0 X k=1 j0 X kPT ⊥ (q −1 PΩk Zk−1 − Zk−1 )k kq −1 PΩk Zk−1 − Zk−1 k k=1  ≤ C00  log n(1) q j0 X s kZk−1 k∞ + k=1 log n(1) q j0 X  kZk−1 k∞,2  . (by Lemma E.4) k=1 We note that by Lemma E.5, kZk−1 k∞ k−1 1 kZ0 k∞ , ≤ 2 and by Lemma E.6, kZk−1 k∞,2 1 ≤ 2 r n(1) 1 kZk−2 k∞ + kZk−2 k∞,2 . µr 2 Therefore, kZk−1 k∞,2 k−1 r n(1) 1 1 kZ0 k∞ + kZk−2 k∞,2 ≤ 2 µr 2 k−1 r k−1 n(1) 1 1 ≤ (k − 1) kZ0 k∞ + kZ0 k∞,2 , 2 µr 2 and so we have kWL k  s ! k−1 k−1r j0 k−1 j0 X X log n log n n 1 1 1 (1) (1) (1) ≤ C00  kZ0 k∞+ kZ0 k∞+ kZ0 k∞,2  (k − 1) q 2 q 2 µr 2 k=1 k=1   s s log n n log n log n (1) (1) (1) (1) ≤ 2C00  kX∗ k∞ + kX∗ k∞ + kX∗ k∞,2  q qµr q r √ n(2) n(2) n1 n2 ∗ ∗ ∗ ≤ 2C0 kX k∞ + kX k∞ + kX k∞,2 (since q = Ω(µr log n(1) )/n(2) ) µr µr µr σr (X∗ ) ≤ , (by incoherence (13)) 4 where we have used the fact that kX∗ k∞,2 ≤ p n(1) kX∗ k∞ ≤ 25 r µr σr (X∗ ). n(2) Proof of (b). Because Yj0 ∈ Ω⊥ , we have PΩ (X∗ + PT ⊥ Yj0 ) = PΩ (X∗ − PT Yj0 ) = PΩ Zj0 . It then follows from Theorem E.2 that kZj0 kF ≤ tj0 kX∗ kF √ ≤ tj0 n1 n2 kX∗ k∞ r µr j0 √ σr (X∗ ) ≤t n1 n2 n1 n2 λ ≤ . (tj0 ≤ e−2 log n(1) ≤ n−2 (1) ) 8 Proof of (c). By definition, we know that X∗ + WL = Zj0 + Yj0 . Since we have shown kZj0 kF ≤ λ/8, it suffices to prove kYj0 k∞ ≤ λ/8. We have kYj0 k∞ ≤ q −1 ≤ q −1 j0 X k=1 j0 X kPΩk Zk−1 k∞ k−1 kX∗ k∞ (by Lemma E.5) k=1 r n(2) 2 µr σr (X∗ ) (by incoherence (13)) ≤ C0 µr log n(1) n1 n2 λ ≤ , 8 µr(log n(1) )2 1/4 if we choose = C for an absolute constant C. This can be true once the constant ρr is n(2) sufficiently small. We now prove that WS given by (28) obeys dual condition (30). We have the following. Lemma F.4. Assume Ω ∼ Ber(p). Then under the other assumptions of Theorem 5.1, WS given by (28) obeys dual condition (30). Proof. According to the standard de-randomization argument [CLMW11], it is equivalent to studying the case when the signs δij of S∗ij are independently distributed as   w.p. p/2, 1, δij = 0, w.p. 1 − p,   −1, w.p. p/2. Proof of (d). Recall that WS = λPT ⊥ X (PΩ PT PΩ )k sign(S∗ ) k≥0 = λPT ⊥ sign(S∗ ) + λPT ⊥ X (PΩ PT PΩ )k sign(S∗ ). k≥1 √ To bound the first term, we have ksign(S∗ )k ≤ 4 n(1) p [Ver10]. So kλPT ⊥ sign(S∗ )k ≤ λksign(S∗ )k ≤ √ 4 pσr (X∗ ) ≤ σr (X∗ )/8. 26 P We now bound the second term. Let G = k≥1 (PΩ PT PΩ )k , which is self-adjoint, and denote by Nn1 and Nn2 the 12 -nets of Sn1 −1 and Sn1 −1 of sizes at most 6n1 and 6n2 , respectively [Led05]. We know that [Ver10] kG(sign(S∗ ))k = hG(yxT ), sign(S∗ )i sup x∈Sn2 −1 ,y∈Sn1 −1 ≤4 hG(yxT ), sign(S∗ )i. sup x∈Nn2 ,y∈Nn1 Consider the random variable X(x, y) = hG(yxT ), sign(S∗ )i which has zero expectation. By Hoeffding’s inequality, we have 2t2 2t2 ≤ 2 exp − . Pr(\|X(x, y)\| > t\|Ω) ≤ 2 exp − kGk2 kG(xyT )k2F Therefore, by a union bound, ∗ n1 +n2 Pr(kG(sign(S ))k > t\|Ω) ≤ 2 × 6 t2 exp − 8kGk2 Note that conditioned on the event {kPΩ PT k ≤ σ}, we have kGk = t2 Pr(λkG(sign(S∗ ))k > t) ≤ 2×6n1 +n2 exp − 2 8λ 1 − σ2 σ2 P . k≥1 (PΩ PT PΩ )k ≤ σ2 . 1−σ 2 So 2 ! Pr(kPΩ PT k ≤ σ)+Pr(kPΩ PT k > σ). The following lemma guarantees that event {kPΩ PT k ≤ σ} holds with high probability for a very small absolute constant σ. Lemma F.5 ([CLMW11], Cor 2.7). Suppose that Ω ∼ Ber(p) and incoherence (3) holds. Then with 2 −2 probability at least 1 − n−10 (1) , kPΩ PT k ≤ p + , provided that 1 − p ≥ C0 µr log n(1) /n(2) for an absolute constant C0 . Setting t = σr (X∗ ) , 8 this completes the proof of (d). P Proof of (e). Recall that WS = λPT ⊥ k≥0 (PΩ PT PΩ )k sign(S∗ ) and so PΩ⊥ WS = λPΩ⊥ (I − PT ) X (PΩ PT PΩ )k sign(S∗ ) k≥0 X = −λPΩ⊥ PT (PΩ PT PΩ )k sign(S∗ ). k≥0 Then for any (i, j) ∈ Ω⊥ , we have * S Wij = hWS , ei eTj i = + X λsign(S∗ ), − (PΩ PT PΩ )k PΩ PT (ei eTj ) . k≥0 Let X(i, j) = − P k≥0 (PΩ PT PΩ )k PΩ PT (ei eTj ). By Hoeffding’s inequality and a union bound, ! Pr S sup \|Wij \| ij > t\|Ω ≤2 X ij 27 2t2 exp − 2 λ kX(i, j)k2F . We note that conditioned on the event {kPΩ PT k ≤ σ}, for any (i, j) ∈ Ω⊥ , kX(i, j)kF ≤ ≤ = ≤ ≤ Then unconditionally, ! Pr S sup \|Wij \| ij >t 1 σkPT (ei eTj )kF 1 − σ2 q 1 σ 1 − kPT ⊥ (ei eTj )k2F 1 − σ2 q 1 σ 1 − k(I − UUT )ei k22 k(I − VVT )ej k22 1 − σ2 s 1 µr µr 1− σ 1− 1− 1 − σ2 n1 n2 r µr µr 1 σ + . 2 1−σ n1 n2 2t2 (1 − σ 2 )2 n1 n2 ≤ 2n1 n2 exp − 2 2 λ σ µr(n1 + n2 ) Pr(kPΩ PT k ≤ σ) + Pr(kPΩ PT k > σ). By Lemma F.5 and setting t = λ/4, the proof of (e) is completed. G Proof of Theorem 6.1 Our computational lower bound for problem (P) assumes the hardness of random 4-SAT. Conjecture 1 (Random 4-SAT). Let c > ln 2 be a constant. Consider a random 4-SAT formula on n variables in which each clause has 4 literals, and in which each of the 16n4 clauses is picked independently with probability c/n3 . Then any algorithm which always outputs 1 when the random formula is satisfiable, 0 and outputs 0 with probability at least 1/2 when the random formula is unsatisfiable, must run in 2c n time on some input, where c0 > 0 is an absolute constant. Based on Conjecture 1, we have the following computational lower bound for problem (P). We show that problem (P) is in general hard for deterministic algorithms. If we additionally assume BPP = P, then the same conclusion holds for randomized algorithms with high probability. Theorem 6.1 (Computational Lower Bound. Restated). Assume Conjecture 1. Then there exists an absolute constant 0 > 0 for which any algorithm that achieves (1 + )OPT in objective function value for problem (P) with ≤ 0 , and with constant probability, requires 2Ω(n1 +n2 ) time, where OPT is the optimum. If in addition, BPP = P, then the same conclusion holds for randomized algorithms succeeding with probability at least 2/3. Proof. Theorem 6.1 is proved by using the hypothesis that random 4-SAT is hard to show hardness of the Maximum Edge Biclique problem for deterministic algorithms. Definition 1 (Maximum Edge Biclique). The problem is Input: An n-by-n bipartite graph G. Output: A k1 -by-k2 complete bipartite subgraph of G, such that k1 · k2 is maximized. 28 [GL04] showed that under the random 4-SAT assumption there exist two constants 1 > 2 > 0 such that no efficient deterministic algorithm is able to distinguish between bipartite graphs G(U, V, E) with \|U \| = \|V \| = n which have a clique of size ≥ (n/16)2 (1 + 1 ) and those in which all bipartite cliques are of size ≤ (n/16)2 (1 + 2 ). The reduction uses a bipartite graph G with at least tn2 edges with large probability, for a constant t. Given a given bipartite graph G(U, V, E), define H(·) as follows. Define the matrix Y and W: Yij = 1 if edge (Ui , Vj ) ∈ E, Yij = 0 if edge (Ui , Vj ) 6∈ E; Wij = 1 if edge (Ui , Vj ) ∈PE, and Wij = poly(n) if 2 (Y − (AB) )2 . edge (Ui , Vj ) 6∈ E. Choose a large enough constant β > 0 and let H(AB) = β ij Wij ij ij 2 Now, if there exists a biclique in G with at least (n/16) (1 + 2 ) edges, then the number of remaining edges is at most tn2 − (n/16)2 (1 + 1 ), and so the solution to min H(AB) + 12 kABk2F has cost at most β[tn2 − (n/16)2 (1 + 1 )] + n2 . On the other hand, if there does not exist a biclique that has more than (n/16)2 (1 + 2 ) edges, then the number of remaining edges is at least (n/16)2 (1 + 2 ), and so any solution to min H(AB) + 21 kABk2F has cost at least β[tn2 − (n/16)2 (1 + 2 )]. Choose β large enough so that β[tn2 − (n/16)2 (1 + 2 )] > β[tn2 − (n/16)2 (1 + 1 )] + n2 . This combined with the result in [GL04] completes the proof for deterministic algorithms. To rule out randomized algorithms running in time 2α(n1 +n2 ) for some function α of n1 , n2 for which α = o(1), observe that we can define a new problem which is the same as problem (P) except the input description of H is padded with a string of 1s of length 2(α/2)(n1 +n2 ) . This string is irrelevant for solving problem (P) but changes the input size to N = poly(n1 , n2 ) + 2(α/2)(n1 +n2 ) . By the argument in the previous paragraph, any deterministic algorithm still requires 2Ω(n) = N ω(1) time to solve this problem, which is super-polynomial in the new input size N . However, if a randomized algorithm can solve it in 2α(n1 +n2 ) time, then it runs in poly(N ) time. This contradicts the assumption that BPP = P. This completes the proof. H Matrix Completion by Information-Theoretic Upper Bound Theorem 4.1 formulates matrix completion as a feasibility problem. However, it is a priori unclear if there is an algorithm for finding X∗ with O(µn(1) r log n(1) ) sample complexity and incoherence (3) via solving the feasibility problem. To answer this question, we mention that matrix completion can be solved in finite time under these minimum assumptions, namely, we note that the feasibility problem is equivalent to finding a zero of the polynomial X (eTi ABej − X∗ij )2 = 0 (i,j)∈Ω w.r.t. the (n1 + n2 )r unknowns A and B. Since A can be assumed to be orthogonal, if the entries of X∗ can be written down with poly(n) bits, then kBkF ≤ exp(poly(n)), which means if one rounds each of the entries of B to the nearest additive grid multiple of 1/ exp(poly(n)), then we will get a rank-k matrix B where each entry represents the true entry of the optimal B up to additive 1/ exp(poly(n)) error (of course one cannot write down B in some cases if the entries are irrational). Such an A and B can be found in exp((n1 + n2 )r) time [Ren92a, Ren92b, BPR96]. This gives an exponential time algorithm to solve the feasibility problem in Theorem 4.1 for matrix completion. I Dual and Bi-Dual Problems In this section, we derive the dual and bi-dual problems of non-convex program (P). According to (5), the primal problem (P) is equivalent to 1 1 min max k − Λ − ABk2F − kΛk2F − H ∗ (Λ). A,B Λ 2 2 29 Therefore, the dual problem is given by 1 1 max min k − Λ − ABk2F − kΛk2F − H ∗ (Λ) Λ A,B 2 2 n (2) 1 X 2 1 = max σi (−Λ) − kΛk2F − H ∗ (Λ) Λ 2 2 i=r+1 1 = max − kΛk2r − H ∗ (Λ), Λ 2 where kΛk2r = Pr 2 i=1 σi (Λ). (D1) The bi-dual problem is derived by 1 min max0 − kΛk2r − H ∗ (Λ0 ) + hM, Λ0 − Λi M Λ,Λ 2 1 2 0 ∗ 0 = min max hM, −Λi − k − Λkr + max hM, Λ i − H (Λ ) M −Λ Λ0 2 = min kMkr∗ + H(M), (D2) M where kMkr∗ = maxX hM, Xi − 21 kXk2r is a convex function, and H(M) = maxΛ0 [hM, Λ0 i − H ∗ (Λ0 )] holds by the definition of conjugate function. Problems (D1) and (D2) can be solved efficiently due to their convexity. In particular, [GRG16] provided a computationally efficient algorithm to compute the proximal operators of functions 12 k · k2r and k · kr∗ . Hence, the Douglas-Rachford algorithm can find the global minimum up to an error in function value in time poly(1/) [HY12]. J Equivalence of Bernoulli and Uniform Models Lemma J.1. Let n be the number of Bernoulli trials and suppose that Ω ∼ Ber(m/n). Then with probability at least 1 − n−10 , \|Ω\| = Θ(m), provided that m ≥ c log n for an absolute constant c. Proof. By the scalar Chernoff bound, with > 0 we have Pr(\|Ω\| ≤ m − n) ≤ exp −2 n2 /(2m) , (31) Pr(\|Ω\| ≥ m + n) ≤ exp −2 n2 /(3m) . (32) and Taking = m/(2n) and m ≥ c1 log n in (31) for an appropriate absolute constant c1 , we have Pr(\|Ω\| ≤ m/2) ≤ exp(−m/4) ≤ n−10 . 2 (33) Taking = m/n and m ≥ c2 log n in (32) for an appropriate absolute constant c2 , we have Pr(\|Ω\| ≥ 2m) ≤ exp(−m/3) ≤ n−10 . 2 (34) Given (33) and (34), we conclude that m/2 < \|Ω\| < 2m with probability at least 1 − n−10 , provided that m ≥ c log n for an absolute constant c. 30 References [AAZB+ 16] Naman Agarwal, Zeyuan Allen-Zhu, Brian Bullins, Elad Hazan, and Tengyu Ma. Finding approximate local minima for nonconvex optimization in linear time. arXiv preprint arXiv:1611.01146, 2016. [ABGM14] Sanjeev Arora, Aditya Bhaskara, Rong Ge, and Tengyu Ma. More algorithms for provable dictionary learning. arXiv preprint arXiv:1401.0579, 2014. [ABHZ16] Pranjal Awasthi, Maria-Florina Balcan, Nika Haghtalab, and Hongyang Zhang. Learning and 1-bit compressed sensing under asymmetric noise. In Annual Conference on Learning Theory, pages 152–192, 2016. [AG16] Anima Anandkumar and Rong Ge. Efficient approaches for escaping higher order saddle points in non-convex optimization. arXiv preprint arXiv:1602.05908, 2016. [AGH+ 14] Animashree Anandkumar, Rong Ge, Daniel J Hsu, Sham M Kakade, and Matus Telgarsky. Tensor decompositions for learning latent variable models. Journal of Machine Learning Research, 15(1):2773–2832, 2014. [AMS11] Christoph Ambühl, Monaldo Mastrolilli, and Ola Svensson. Inapproximability results for maximum edge biclique, minimum linear arrangement, and sparsest cut. SIAM Journal on Computing, 40(2):567–596, 2011. [ANW12] Alekh Agarwal, Sahand Negahban, and Martin J Wainwright. Noisy matrix decomposition via convex relaxation: Optimal rates in high dimensions. The Annals of Statistics, pages 1171–1197, 2012. [ARR14] Ali Ahmed, Benjamin Recht, and Justin Romberg. Blind deconvolution using convex programming. IEEE Transactions on Information Theory, 60(3):1711–1732, 2014. [AZ16] Zeyuan Allen-Zhu. Katyusha: The first direct acceleration of stochastic gradient methods. arXiv preprint arXiv:1603.05953, 2016. [BCMV14] Aditya Bhaskara, Moses Charikar, Ankur Moitra, and Aravindan Vijayaraghavan. Smoothed analysis of tensor decompositions. In ACM Symposium on Theory of Computing, pages 594–603, 2014. [BE06] Amir Beck and Yonina C Eldar. Strong duality in nonconvex quadratic optimization with two quadratic constraints. SIAM Journal on Optimization, 17(3):844–860, 2006. [BH89] Pierre Baldi and Kurt Hornik. Neural networks and principal component analysis: Learning from examples without local minima. Neural Networks, 2(1):53–58, 1989. [BKS16] Srinadh Bhojanapalli, Anastasios Kyrillidis, and Sujay Sanghavi. Dropping convexity for faster semi-definite optimization. In Annual Conference on Learning Theory, pages 530–582, 2016. [BM05] Samuel Burer and Renato DC Monteiro. Local minima and convergence in low-rank semidefinite programming. Mathematical Programming, 103(3):427–444, 2005. [BMP08] Francis Bach, Julien Mairal, and Jean Ponce. Convex sparse matrix factorizations. arXiv preprint arXiv:0812.1869, 2008. 31 [BNS16] Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro. Global optimality of local search for low rank matrix recovery. In Advances in Neural Information Processing Systems, pages 3873–3881, 2016. [BPR96] Saugata Basu, Richard Pollack, and Marie Francoise Roy. On the combinatorial and algebraic complexity of quantifier elimination. J. ACM, 43(6):1002–1045, 1996. [BV04] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. [BZ16] Maria-Florina Balcan and Hongyang Zhang. Noise-tolerant life-long matrix completion via adaptive sampling. In Advances in Neural Information Processing Systems, pages 2955–2963, 2016. [Che15] Yudong Chen. Incoherence-optimal matrix completion. IEEE Transactions on Information Theory, 61(5):2909–2923, 2015. [CHM+ 15] Anna Choromanska, Mikael Henaff, Michael Mathieu, Gérard Ben Arous, and Yann LeCun. The loss surfaces of multilayer networks. In International Conference on Artificial Intelligence and Statistics, 2015. [CLMW11] Emmanuel J. Candès, Xiaodong Li, Yi Ma, and John Wright. Robust principal component analysis? Journal of the ACM, 58(3):11, 2011. [CR09] Emmanuel J. Candès and Ben Recht. Exact matrix completion via convex optimization. Foundations of Computational Mathematics, 9(6):717–772, 2009. [CR13] Emmanuel J. Candès and Benjamin Recht. Simple bounds for recovering low-complexity models. Mathematical Programming, pages 1–13, 2013. [CRPW12] Venkat Chandrasekaran, Benjamin Recht, Pablo A Parrilo, and Alan S Willsky. The convex geometry of linear inverse problems. Foundations of Computational Mathematics, 12(6):805– 849, 2012. [CT10] Emmanuel J. Candès and Terence Tao. The power of convex relaxation: Near-optimal matrix completion. IEEE Transactions on Information Theory, 56(5):2053–2080, 2010. [CW15] Yudong Chen and Martin J. Wainwright. Fast low-rank estimation by projected gradient descent: General statistical and algorithmic guarantees. arXiv preprint arXiv:1509.03025, 2015. [Fei02] Uriel Feige. Relations between average case complexity and approximation complexity. In Proceedings of the 17th Annual IEEE Conference on Computational Complexity, Montréal, Québec, Canada, May 21-24, 2002, page 5, 2002. [GHJY15] Rong Ge, Furong Huang, Chi Jin, and Yang Yuan. Escaping from saddle points – online stochastic gradient for tensor decomposition. In Annual Conference on Learning Theory, pages 797–842, 2015. [GJY17] Rong Ge, Chi Jin, and Zheng Yi. No spurious local minima in nonconvex low rank problems: A unified geometric analysis. arXiv preprint: 1704.00708, 2017. [GL04] Andreas Goerdt and André Lanka. An approximation hardness result for bipartite clique. In Electronic Colloquium on Computational Complexity, Report, volume 48, 2004. 32 [GLM16] Rong Ge, Jason D Lee, and Tengyu Ma. Matrix completion has no spurious local minimum. In Advances in Neural Information Processing Systems, pages 2973–2981, 2016. [GLZ17] David Gamarnik, Quan Li, and Hongyi Zhang. Matrix completion from O(n) samples in linear time. arXiv preprint arXiv:1702.02267, 2017. [GRG16] Christian Grussler, Anders Rantzer, and Pontus Giselsson. Low-rank optimization with convex constraints. arXiv preprint arXiv:1606.01793, 2016. [Gro11] D. Gross. Recovering low-rank matrices from few coefficients in any basis. IEEE Transactions on Information Theory, 57(3):1548–1566, 2011. [GWL16] Quanquan Gu, Zhaoran Wang, and Han Liu. Low-rank and sparse structure pursuit via alternating minimization. In International Conference on Artificial Intelligence and Statistics, pages 600–609, 2016. [Har14] Moritz Hardt. Understanding alternating minimization for matrix completion. In IEEE Symposium on Foundations of Computer Science, pages 651–660, 2014. [HM12] Moritz Hardt and Ankur Moitra. Algorithms and hardness for robust subspace recovery. arXiv preprint: 1211.1041, 2012. [HMRW14] Moritz Hardt, Raghu Meka, Prasad Raghavendra, and Benjamin Weitz. Computational limits for matrix completion. In Annual Conference on Learning Theory, pages 703–725, 2014. [HV15] Benjamin D Haeffele and René Vidal. Global optimality in tensor factorization, deep learning, and beyond. arXiv preprint arXiv:1506.07540, 2015. [HY12] Bingsheng He and Xiaoming Yuan. On the O(1/n) convergence rate of the douglas–rachford alternating direction method. SIAM Journal on Numerical Analysis, 50(2):700–709, 2012. [HYV14] Benjamin Haeffele, Eric Young, and Rene Vidal. Structured low-rank matrix factorization: Optimality, algorithm, and applications to image processing. In International Conference on Machine Learning, pages 2007–2015, 2014. [IW97] Russell Impagliazzo and Avi Wigderson. P = BPP if E requires exponential circuits: Derandomizing the XOR lemma. In ACM Symposium on the Theory of Computing, pages 220–229, 1997. [Jah07] Johannes Jahn. Introduction to the theory of nonlinear optimization. Springer Berlin Heidelberg, 2007. [JBAS10] Michel Journée, Francis Bach, P-A Absil, and Rodolphe Sepulchre. Low-rank optimization on the cone of positive semidefinite matrices. SIAM Journal on Optimization, 20(5):2327–2351, 2010. [JGN+ 17] Chi Jin, Rong Ge, Praneeth Netrapalli, Sham M Kakade, and Michael I Jordan. How to escape saddle points efficiently. arXiv preprint arXiv:1703.00887, 2017. [JMD10] Prateek Jain, Raghu Meka, and Inderjit S Dhillon. Guaranteed rank minimization via singular value projection. In Advances in Neural Information Processing Systems, pages 937–945, 2010. [JNS13] Prateek Jain, Praneeth Netrapalli, and Sujay Sanghavi. Low-rank matrix completion using alternating minimization. In ACM Symposium on Theory of Computing, pages 665–674, 2013. 33 [Kaw16] Kenji Kawaguchi. Deep learning without poor local minima. arXiv preprint arXiv:1605.07110, 2016. [KBV09] Yehuda Koren, Robert Bell, and Chris Volinsky. Matrix factorization techniques for recommender systems. IEEE Computer, 42(8):30–37, 2009. [Kes12] Raghunandan Hulikal Keshavan. Efficient algorithms for collaborative filtering. PhD thesis, Stanford University, 2012. [KMO10a] Raghunandan H Keshavan, Andrea Montanari, and Sewoong Oh. Matrix completion from a few entries. IEEE Transactions on Information Theory, 56(6):2980–2998, 2010. [KMO10b] Raghunandan H Keshavan, Andrea Montanari, and Sewoong Oh. Matrix completion from noisy entries. Journal of Machine Learning Research, 11:2057–2078, 2010. [Led05] Michel Ledoux. The concentration of measure phenomenon. Number 89. American Mathematical Society, 2005. [LLR16] Yuanzhi Li, Yingyu Liang, and Andrej Risteski. Recovery guarantee of weighted low-rank approximation via alternating minimization. In International Conference on Machine Learning, pages 2358–2367, 2016. [NNS+ 14] Praneeth Netrapalli, UN Niranjan, Sujay Sanghavi, Animashree Anandkumar, and Prateek Jain. Non-convex robust PCA. In Advances in Neural Information Processing Systems, pages 1107–1115, 2014. [NW12] Sahand Negahban and Martin J Wainwright. Restricted strong convexity and weighted matrix completion: Optimal bounds with noise. Journal of Machine Learning Research, 13:1665–1697, 2012. [OW92] Michael L Overton and Robert S Womersley. On the sum of the largest eigenvalues of a symmetric matrix. SIAM Journal on Matrix Analysis and Applications, 13(1):41–45, 1992. [Rec11] Benjamin Recht. A simpler approach to matrix completion. Journal of Machine Learning Research, 12:3413–3430, 2011. [Ren92a] James Renegar. On the computational complexity and geometry of the first-order theory of the reals, part I: introduction. preliminaries. the geometry of semi-algebraic sets. the decision problem for the existential theory of the reals. J. Symb. Comput., 13(3):255–300, 1992. [Ren92b] James Renegar. On the computational complexity and geometry of the first-order theory of the reals, part II: the general decision problem. preliminaries for quantifier elimination. J. Symb. Comput., 13(3):301–328, 1992. [RSW16] Ilya Razenshteyn, Zhao Song, and David P. Woodruff. Weighted low rank approximations with provable guarantees. In ACM Symposium on Theory of Computing, pages 250–263, 2016. [SL15] Ruoyu Sun and Zhi-Quan Luo. Guaranteed matrix completion via nonconvex factorization. In IEEE Symposium on Foundations of Computer Science, pages 270–289, 2015. [SQW16] Ju Sun, Qing Qu, and John Wright. A geometric analysis of phase retrieval. In IEEE International Symposium on Information Theory, pages 2379–2383, 2016. 34 [SQW17a] Ju Sun, Qing Qu, and John Wright. Complete dictionary recovery over the sphere I: Overview and the geometric picture. IEEE Transactions on Information Theory, 63(2):853–884, 2017. [SQW17b] Ju Sun, Qing Qu, and John Wright. Complete dictionary recovery over the sphere II: Recovery by Riemannian trust-region method. IEEE Transactions on Information Theory, 63(2):885–914, 2017. [SS05] Nathan Srebro and Adi Shraibman. Rank, trace-norm and max-norm. In International Conference on Computational Learning Theory, pages 545–560. Springer, 2005. [SU15] Reinhold Schneider and André Uschmajew. Convergence results for projected line-search methods on varieties of low-rank matrices via Lojasiewicz inequality. SIAM Journal on Optimization, 25(1):622–646, 2015. [SWZ14] Yuan Shen, Zaiwen Wen, and Yin Zhang. Augmented lagrangian alternating direction method for matrix separation based on low-rank factorization. Optimization Methods and Software, 29(2):239–263, 2014. [TBSR13] Gongguo Tang, Badri Narayan Bhaskar, Parikshit Shah, and Benjamin Recht. Compressed sensing off the grid. IEEE Transactions on Information Theory, 59(11):7465–7490, 2013. [TBSR15] Stephen Tu, Ross Boczar, Mahdi Soltanolkotabi, and Benjamin Recht. Low-rank solutions of linear matrix equations via procrustes flow. arXiv preprint arXiv:1507.03566, 2015. [Ver09] Roman Vershynin. Lectures in geometric functional analysis. pages 1–76, 2009. [Ver10] Roman Vershynin. Introduction to the non-asymptotic analysis of random matrices. arXiv preprint: 1011.3027, 2010. [Ver15] Roman Vershynin. Estimation in high dimensions: A geometric perspective. In Sampling theory, a renaissance, pages 3–66. Springer, 2015. [WX12] Yu-Xiang Wang and Huan Xu. Stability of matrix factorization for collaborative filtering. In International Conference on Machine Learning, pages 417–424, 2012. [WYZ12] Zaiwen Wen, Wotao Yin, and Yin Zhang. Solving a low-rank factorization model for matrix completion by a nonlinear successive over-relaxation algorithm. Mathematical Programming Computation, 4(4):333–361, 2012. [YPCC16] Xinyang Yi, Dohyung Park, Yudong Chen, and Constantine Caramanis. Fast algorithms for robust PCA via gradient descent. In Advances in neural information processing systems, pages 4152–4160, 2016. [ZL15] Qinqing Zheng and John Lafferty. A convergent gradient descent algorithm for rank minimization and semidefinite programming from random linear measurements. In Advances in Neural Information Processing Systems, pages 109–117, 2015. [ZL16] Qinqing Zheng and John Lafferty. Convergence analysis for rectangular matrix completion using Burer-Monteiro factorization and gradient descent. arXiv preprint arXiv:1605.07051, 2016. 35 [ZLZ13] Hongyang Zhang, Zhouchen Lin, and Chao Zhang. A counterexample for the validity of using nuclear norm as a convex surrogate of rank. In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, volume 8189, pages 226–241, 2013. [ZLZ16] Hongyang Zhang, Zhouchen Lin, and Chao Zhang. Completing low-rank matrices with corrupted samples from few coefficients in general basis. IEEE Transactions on Information Theory, 62(8):4748–4768, 2016. [ZLZC15] Hongyang Zhang, Zhouchen Lin, Chao Zhang, and Edward Chang. Exact recoverability of robust PCA via outlier pursuit with tight recovery bounds. In AAAI Conference on Artificial Intelligence, pages 3143–3149, 2015. [ZLZG14] Hongyang Zhang, Zhouchen Lin, Chao Zhang, and Junbin Gao. Robust latent low rank representation for subspace clustering. Neurocomputing, 145:369–373, 2014. [ZLZG15] Hongyang Zhang, Zhouchen Lin, Chao Zhang, and Junbin Gao. Relations among some low rank subspace recovery models. Neural Computation, 27:1915–1950, 2015. [ZWG17] Xiao Zhang, Lingxiao Wang, and Quanquan Gu. A nonconvex free lunch for low-rank plus sparse matrix recovery. arXiv preprint arXiv:1702.06525, 2017. [ZWL15] Tuo Zhao, Zhaoran Wang, and Han Liu. A nonconvex optimization framework for low rank matrix estimation. In Advances in Neural Information Processing Systems, pages 559–567, 2015. 36	8cs.DS
arXiv:1403.3148v4 [cs.SI] 15 Nov 2016 Heat Kernel Based Community Detection Kyle Kloster David F. Gleich Purdue University West Lafayette, IN Purdue University West Lafayette, IN kkloste@purdue.edu dgleich@purdue.edu ABSTRACT The heat kernel is a type of graph diffusion that, like the much-used personalized PageRank diffusion, is useful in identifying a community nearby a starting seed node. We present the first deterministic, local algorithm to compute this diffusion and use that algorithm to study the communities that it produces. Our algorithm is formally a relaxation method for solving a linear system to estimate the matrix exponential in a degree-weighted norm. We prove that this algorithm stays localized in a large graph and has a worst-case constant runtime that depends only on the parameters of the diffusion, not the size of the graph. On large graphs, our experiments indicate that the communities produced by this method have better conductance than those produced by PageRank, although they take slightly longer to compute. On a real-world community identification task, the heat kernel communities perform better than those from the PageRank diffusion. Categories and Subject Descriptors G.2.2 [Discrete mathematics]: Graph theory—Graph algorithms; I.5.3 [Pattern recognition]: Clustering—Algorithms General Terms Algorithms,Theory Keywords heat kernel; local clustering 1. INTRODUCTION The community detection problem is to identify a set of nodes in a graph that are internally cohesive but also separated from the remainder of the network. One popular way to capture this idea is through the conductance measure of a set. We treat this idea formally in the next section, but Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. KDD’14, August 24–27, 2014, New York, NY, USA. Copyright is held by the owner/author(s). Publication rights licensed to ACM. ACM 978-1-4503-2956-9/14/08 ...$15.00. http://dx.doi.org/10.1145/2623330.2623706. informally, the conductance of a set is a ratio of the number of edges leaving the set to the number of edges touched by the set of vertices. If this value is small, then it indicates a set with many internal edges and few edges leaving. In many surveys and empirical studies [17, 35, 49], the conductance measure surfaces as one of the most reliable measures of a community. Although this measure has been critized for producing cavemen-type communities [24], empirical properties of real-world communities correlate highly with sets produced by algorithms that optimize conductance [1]. Furthermore, state-of-the-art methods for identifying overlapping sets of communities use conductance to find realworld communities better than any alternative [54]. Virtually all of the rigorous algorithms that identify sets of small conductance are based on min-cuts [6, 48], eigenvector computations [4, 21], or local graph diffusions [5, 14]. (One notable exception is the graclus method [17] that uses a relationship between kernel k-means and a variant of conductance.) In this paper, we study a new algorithm that uses a heat kernel diffusion [14] to identify small-conductance communities in a network. (The heat kernel is discussed formally in Section 3.) Although the properties of this diffusion had been analyzed in theory in Chung’s work [14], that work did not provide an efficient algorithm to compute the diffusion. Recently, Chung and Simpson stated a randomized Monte Carlo method to estimate the diffusion [16]. This paper introduces an efficient and deterministic method to estimate this diffusion. We use it to study the properties of the small conductance sets identified by the method as communities. For this use, a deterministic approach is critical as we need to differentiate between subtle properties of the diffusion. Our primary point of comparison is the well-known personalized PageRank diffusion [5], which has been used to establish new insights into the properties of communities in large scale networks [35]. Thus, we wish to understand how the communities produced by the heat kernel compare to those produced by personalized PageRank. The basic operation of our algorithm is a coordinate relaxation step. This has been used to efficiently compute personalized PageRank [5, 23, 38] where it is known as the “push” operation on a graph; the term “coordinate relaxation” is the classical name for this operation, which dates back to the Gauss-Seidel method. What distinguishes our approach from prior work is the use of coordinate relaxation on an implicitly constructed linear system that will estimate the heat kernel diffusion, which is formally the exponential of the random walk transition matrix. We began looking into this problem recently in a workshop paper [28], where we showed that this style of algorithm successfully estimates a related quantity. This paper tackles a fundamentally new direction and, although similar in techniques, required an entirely new analysis. In particular, we are able to estimate this diffusion in constant time in a degree-weighted norm that depends only on the parameters of the diffusion, and not the size of the graph. A Python implementation of our algorithm to accomplish this task is presented in Figure 2. In the remainder of the paper, we first review these topics formally (Sections 2 and 3); present our algorithm (Section 4) and discuss related work (Section 5). Then we show how our new approach differs from and improves upon the personalized PageRank diffusion in synthetic and real-world problems. Summary of contributions. · We propose the first local, deterministic method to accurately compute a heat kernel diffusion in a graph. The code is simple and scalable to any graph where out-link access is inexpensive. · Our method is always localized, even in massive graphs, because it has a provably constant runtime in a degreeweighted norm. · We compare the new heat kernel diffusion method to the venerable PageRank diffusion in synthetic, small, and large networks (up to 2 billion edges) to demonstrate the differences. On large networks such as Twitter, our method tends to produce smaller and tighter communities. It is also more accurate at detecting ground-truth communities. We make our experimental codes available in the spirit of reproducible research: https://www.cs.purdue.edu/homes/dgleich/codes/hkgrow 2. PRELIMINARIES We begin by fixing our notation. Let G = (V, E) be a simple, undirected graph with n = \|V \| vertices. Fix an ordering of the vertices from 1 to n such that we can refer to a vertex by it’s numeric ID. For a vertex v ∈ V , we denote by dv the degree of node v. Let A be the associated adjacency matrix. Because G is undirected, A is symmetric. Furthermore, let D be the diagonal matrix of degrees (Dii = di ) and P = (D−1 A)T = AD−1 be the random walk transition matrix. Finally, we denote by ei the vector (of an appropriate length) of zeros having a single 1 in entry i, and by e the vector of all 1s. Conductance. Given a subset S ⊆ V , we denote by vol(S) the sum of the degrees of all vertices in S, and by ∂(S), the boundary of S, the number of edges with one endpoint inside of S and one endpoint outside of S. With this notation, the conductance of a set S is given by φ(S) := ∂(S) min{vol(S), vol(V − S)} Conceptually, φ(S) is the probability that a random walk of length one will land outside of S, given that we start from a node chosen uniformly at random inside S. Matrix exponential. The heat kernel of a graph involves the matrix exponential, and we wish to briefly mention some facts about this operation. See Higham [22] for a more in-depth treatment. Consider a general matrix G. The exponential function of a matrix is not just the exponential applied element-wise, but rather is given by substituting the matrix into the Taylor series expansion of the exponential function: ∞ X 1 k exp {G} = G . k! k=0 That said, the exponential of a diagonal matrix is the exponential function applied to the diagonal entries. This phenomenon occurs because powers of a diagonal matrix are simply the diagonal elements raised to the given power. For any matrix G, if GT G = GGT (which is a generalization of the notion of a symmetric matrix, called a normal matrix), then G has an eigenvalue decomposition G = XΛX−1 and there is a simple, albeit inefficient, means of computing the exponential: exp {G} = X exp {Λ} X−1 . Computing the matrix exponential, or even the product of the exponential with a vector, exp {G} z, is still an active area of research [2]. 3. FINDING SMALL CONDUCTANCE COMMUNITIES WITH DIFFUSIONS A graph diffusion is any sum of the following form: f= ∞ X αk Pk s k=0 P where k αk = 1 and s is a stochastic vector (that is, it is non-negative and sums to one.) Intuitively, a diffusion captures how a quantity of si material on node i flows through the graph. The terms αk provide a decaying weight that ensures that the diffusion eventually dissipates. In the context of this paper, we are interested in the diffusions of single nodes or neighborhood P sets of a single vertex; and so in these cases s = ei or s = i∈S ei /\|S\| for a small set S. We call the origins of the diffusion the seeds. Given an estimate of a diffusion f from a seed, we can produce a small conductance community from it using a sweep procedure. This involves computing D−1 f, sorting the nodes in descending order by their magnitude in this vector, and computing the conductance of each prefix of the sorted list. Due to the properties of conductance, there is an efficient means of computing all of these conductances. We then return the set of smallest conductance as the community around the seeds. The personalized PageRank diffusion. One of the most well-known instances of this framework is the personalized PageRank diffusion. Fix α ∈ (0, 1). Then p is defined: p = (1 − α) ∞ X αk Pk s. (1) k=0 The properties of this diffusion have been studied extensively. In particular, Andersen et al. [5] establish a local Cheeger inequality using a particular algorithm called “push” that locally distributes mass. The local Cheeger inequality informally states that, if the seed is nearby a set with small conductance, then the result of the sweep procedure is a set with a related conductance. Moreover, they show that their “push” algorithm estimates f with an error ε in a degree1 weighted norm by looking at (1−α)ε edges. The heat kernel diffusion. Another instance of the same framework is the heat kernel diffusion [14, 15]. It sim- ply replaces the weights αk with tk /k!: ! ∞ X tk k −t (P) s = exp {−t(I − P)} s. h=e k! k=0 satisfying the degree weighted bound: kD−1 exp {−t(I − P)} s − D−1 xk∞ < ε. (2) While it was known that estimating h gave rise to a similar type of local Cheeger inequality [15]; until Chung and Simpson’s Monte Carlo approach [16], no methods were known to estimate this quantity efficiently. Our new algorithm is a deterministic approach that is more suitable for comparing the properties of the diffusions. It terminates after exploring 2Net edges (Theorem 1), where N is a parameter that grows ε slowly1 with ε. Heat kernels compared to PageRank. These different sets of coefficients simply assign different levels of importance to walks of varying lengths: the heat kernel coefficients tk decay much more quickly than αk , and so the heat kerk! nel more heavily weights shorter walks in the overall sum (depicted in Figure 1). This property, in turn, will have important consequences when we study these methods in large graphs in Section 6. 0 10 Weight α=0.99 By using standard properties of the matrix exponential, we can factor exp {−t(I − P)} = e−t exp {tP)} and scale by et so that the above problem is equivalent to computing y satisfying kD−1 (exp {tP} s − y)k∞ < et ε. The element-wise characterization is that y must satisfy: \|et hi − yi \| < et εdi for all i. A similar weighted objective was used in the push algorithm for PageRank [5]. Outline of algorithm. To accomplish this, we first approximate exp {tP} with its degree N Taylor polynomial, TN (tP), and then we compute TN (tP)s. But we use a large, implicitly formed linear system to avoid explicitly evaluating the Taylor polynomial. Once we have the linear system, we state a relaxation method in the spirit of Gauss-Seidel and the PageRank push algorithm in order to compute an accurate approximation of h. 4.1 Taylor Polynomial for exp{X} Determining the exponential of a matrix is a sensitive computation with a rich history [41, 42]. For a general matrix G, an approximation via the Taylor polynomial, exp {G} = ∞ X 1 Gk k! k=0 −5 10 t=1 t=5 0 α=0.85 t=15 20 40 60 80 100 Length Figure 1: Each curve represents the coefficients of (AD−1 )k in a sum of walks. The dotted blue lines give αk , and the red give tk /k!, for the indicated values of α and t. (3) ≈ N X 1 Gk , k! k=0 can be inaccurate when kGk is large and G has mixed signs, as large powers Gk can contain large, oppositely signed numbers that cancel properly only in exact arithmetic. However, we intend to compute exp {tP} s, where P, t, and s are nonnegative, so the calculation does not rely on any delicate cancellations. Furthermore, our approximation need not be highly precise. We therefore use the polynomial P k tk exp {tP} s ≈ TN (tP) = N k=0 k! P for our approximation. For details on choosing N , see Section 4.5. For now, assume that we have chosen N such that kD−1 exp {tP} s − D−1 TN (tP)sk∞ < ε/2. 4. ALGORITHM The overall idea of the local clustering algorithm is to approximate a heat kernel vector of the form h = exp {−t(I − P)} s so that we can perform a sweep over h. Here we describe a coordinate-relaxation method, which we call hk-relax, for approximating h. This algorithm is rooted in our recent work on computing an accurate column of exp {P} [28]; but is heavily tuned to the objective below. Thus, while the overall strategy is classical – just as the PageRank push method is a classic relaxation method – the simplifications and efficient implementation are entirely novel. In particular, the new objective in this paper enables us to get a constant runtime bound independent of any property of the graph, which differs markedly from both of our previous methods [28, 27]. Our objective. Recall that the final step of finding a small conductance community involves dividing by the degree of each node. Thus, our goal is to compute x ≈ h 1 Since the original publication of this work we have proven an upperbound on N in terms of ε and t [26]. (4) This way, if we compute y ≈ TN (tP)s satisfying kD−1 TN (tP)s − D−1 yk∞ < ε/2, then by the triangle inequality we will have kD−1 exp {tP} s − D−1 yk∞ < ε, (5) our objective. 4.2 Error weights Using a degree N Taylor polynomial, hk-relax ultimately approximates h by approximating each term in the sum of the polynomial times the vector s: s + 1t Ps + · · · + tN PN s. N! The total error of our computed solution is then a weighted k sum of the errors at each individual term, tk! Pk s. We show in Lemma 1 that these weights are given by the polynomials ψk (t), which we define now. For a fixed degree N Taylor P tk polynomial of the exponential, TN = N k=0 k! , we define ψk := N−k X m=0 k! tm (m+k)! for k = 0, · · · , N . (6) These polynomials ψk (t) are closely related to the φ functions central to exponential integrators in ODEs [39]. Note that ψ0 = TN . To guarantee the total error satisfies the criterion (3) then, it is enough to show that the error at each Taylor term satisfies an ∞-norm inequality analogous to (3). This is discussed in more detail in Section A. 4.3 Deriving a linear system To define the basic step of the hk-relax algorithm and to show how the ψk influence the total error, we rearrange the Taylor polynomial computation into a linear system. Denote by vk the kth term of the vector sum TN (tP)s: N TN (tP)s = s + 1t Ps + · · · + tN! PN s = v0 + v1 + · · · + vN . (7) (8) k+1 t t Note that vk+1 = (k+1)! Pk+1 = (k+1) Pvk . This identity implies that the terms vk exactly satisfy the linear system      I s v0 −t   v1   0   1 P I        ..   ..   .. −t   .  =  .  . (9)  . P 2       .   .   .. . .      . .  . I −t P N I vN 0 Let v = [v0 ; v1 ; · · · ; vN ]. An approximate solution v̂ to P (9) would have block components v̂k such that N k=0 v̂k ≈ t TN (tP)s, our desired approximation to e h. In practice, we update only a single length n solution vector, adding all updates to that vector, instead of maintaining the N + 1 different block vectors v̂k as part of v̂; furthermore, the block matrix and right-hand side are never formed explicitly. With this block system in place, we can describe the algorithm’s steps. # G is graph as dictionary - of - sets , # seed is an array of seeds , # t , eps , N , psis are precomputed x = {} # Store x , r as dictionaries r = {} # initialize residual Q = collections . deque () # initialize queue for s in seed : r [( s ,0)] = 1./ len ( seed ) Q . append (( s ,0)) while len ( Q ) > 0: (v , j ) = Q . popleft () # v has r [( v , j )] ... rvj = r [( v , j )] # perform the hk - relax step if v not in x : x [ v ] = 0. x [ v ] += rvj r [( v , j )] = 0. mass = ( t * rvj /( float ( j )+1.))/ len ( G [ v ]) # for neighbors of v for u in G [ v ]: next = (u , j +1) # in the next block if j +1 == N : # last step , add to soln x [ u ] += rvj / len ( G ( v )) continue if next not in r : r [ next ] = 0. thresh = math . exp ( t )* eps * len ( G [ u ]) thresh = thresh /( N * psis [ j +1])/2. if r [ next ] < thresh and \ r [ next ] + mass >= thresh : Q . append ( next ) # add u to queue r [ next ] = r [ next ] + mass Figure 2: Pseudo-code for our algorithm as working python code. The graph is stored as a dictionary of sets so that the G[v] statement returns the set of neighbors associated with vertex v. The solution is the vector x indexed by vertices and the residual vector is indexed by tuples (v, j) that are pairs of vertices and steps j. A fully working demo may be downloaded from github https://gist.github.com/dgleich/7d904a10dfd9ddfaf49a. 4.4 The hk-relax algorithm Given a random walk transition matrix P, scalar t > 0, and seed vector s as inputs, we solve the linear system from (9) as follows. Denote the initial solution vector by y and the initial nN × 1 residual by r(0) = e1 ⊗ s. Denote by r(i, j) the entry of r corresponding to node i in residual block j. The idea is to iteratively remove all entries from r that satisfy r(i, j) ≥ et εdi . 2N ψj (t) (10) To organize this process, we begin by placing the nonzero entries of r(0) in a queue, Q(r), and place updated entries of r into Q(r) only if they satisfy (10). Then hk-relax proceeds as follows. 1. At each step, pop the top entry of Q(r), call it r(i, j), and subtract that entry in r, making r(i, j) = 0. 2. Add r(i, j) to yi . t 3. Add r(i, j) j+1 Pei to residual block rj+1 . 4. For each entry of rj+1 that was updated, add that entry to the back of Q(r) if it satisfies (10). Once all entries of r that satisfy (10) have been removed, the resulting solution vector y will satisfy (3), which we prove in Section A, along with a bound on the work required to achieve this. We present working code for our method in Figure 2 that shows how to optimize this computation using sparse data structures. These make it highly efficient in practice. 4.5 Choosing N The last detail of the algorithm we need to discuss is how to pick N . In (4) we want to guarantee the accuracy of D−1exp {tP} s − D−1 TN (tP)s. By using D−1 exp {tP} = exp tPT D−1 and D−1 TN (tP) = TN (tPT )D−1 , we can get a new upperbound on kD−1 exp {tP} s−D−1 TN (tP)sk∞ by noting n o k exp tPT D−1 s − TN (tPT )D−1 sk∞ n o ≤ k exp tPT − TN (tPT )k∞ kD−1 sk∞ . Since s is stochastic, we have kD−1 sk∞ ≤ kD−1 sk1 ≤ 1. From [36] we know that the norm k exp tPT −TN (tPT )k∞ is bounded by ktPT kN+1 (N + 2) (N + 2) tN+1 ∞ ≤ . (N + 1)! (N + 2 − t) (N + 1)! (N + 2 − t) (11) So to guarantee (4), it is enough to choose N that implies (N+2) tN +1 < ε/2. Such an N can be determined effi(N+1)! (N+2−t) ciently simply by iteratively computing terms of the Taylor polynomial for et until the error is less than the desired error for hk-relax. In practice, this required a choice of N no greater than 2t log( 1ε ), which we think can be made rigorous. 4.6 Outline of convergence result The proof proceeds as follows. First, we relate the error vector of the Taylor approximation E1 = TN (tP)s − x, to the error vector from solving the linear system described in Section 4.4, E2 = TN (tP)s − y. Second, we express the error vector E2 in terms of the residual blocks of the linear system (9); this will involve writing E2 as a sum of residual blocks rk with weights ψk (tP). Third, we use the previous results to upperbound kD−1 TN (tP)s − P −1 D−1 xk∞ with N rk k∞ , and use this to show k=0 ψk (t)kD that kD−1 TN (tP)s − D−1 xk∞ < ε/2 is guaranteed by the stopping criterion of hk-relax, (3). Finally, we prove that performing steps of hk-relax until the stopping criterion is attained requires work bounded by 2Nψε1 (t) ≤ 2N et /ε. 5. RELATED WORK We wish to highlight a few ideas that have recently emerged in the literature to clarify how our method differs. We discuss these in terms of community detection, the matrix exponential, fast diffusion methods, and relaxation methods. Community detection and conductance. Conductance often appears in community detection and is known to be one of the most important measures of a community [49]. The personalized PageRank method is one of the most scalable methods to find sets of small conductance, although recent work opened up a new possibility with localized maxflow algorithms [48]. For the PageRank algorithm we use as a point of comparison, Zhu et al. [56] recently provided an improved bound on the performance of this algorithm when finding sets with high internal conductance. The internal conductance of a set is the minimum conductance of the subgraph induced by the set and we would expect that realworld communities have large internal conductance. Due to the similarity between our algorithm and the personalized PageRank diffusion, we believe that a similar result likely holds here as well. The matrix exponential in network analysis. Recently, the matrix exponential has frequently appeared as a tool in the network analysis literature. It has been used to estimate node centrality [18, 20, 19], for link-prediction [31], in graph kernels [30], and – as already mentioned – clustering and community detection [14]. Many of these studies involve fast ways to approximate the entire matrix exponential, instead of a single column as we study here. For instance, Sui et al. [51] describe a low-parameter decomposition of a network that is useful both for estimating Katz scores [25] and the matrix exponential. Orecchia and Mahoney [46] show that the heat kernel diffusion implicitly approximates a diffusion operator using a particular type of generalized entropy, which provides a principled rationale for its use. Fast methods for diffusions. Perhaps the most related work is a recent Monte Carlo algorithm by Chung and Simpson [16] to estimate the heat kernel diffusion via a random walk sampling scheme. This approach involves directly simulating a random walk with transition probabilities that mirror the Taylor series expansion of the exponential. In comparison, our approach is entirely deterministic and thus more useful to compare between diffusions, as it eliminates the algorithmic variance endemic to Monte Carlo simulations. A similar idea is used by Borgs et al. [13] to achieve a randomized sublinear time algorithm to estimate the largest PageRank entries, and in fact, Monto Carlo methods frequently feature in PageRank computations due to the relationship between the diffusion and a random walk [7, 13, 8, 9]. Most other deterministic approaches for the matrix exponential involve at least one matrix-vector product [47, 2]. Relaxation methods. The algorithm we use is a coordinate relaxation method similar to Gauss-Seidel and GaussSouthwell. If we applied it to a symmetric positive definite matrix, then it would be a coordinate descent method [37]. It has been proposed for PageRank in a few difference cases [23, 38, 5]. The same type of relaxation method has also been used to estimate the Katz diffusion [12]. We recently used it to estimate a column of the matrix exponential exp {P} ei in a strict, 1-norm error and were able to prove a sublinear convergence bound by assuming a very slowly growing maximum degree [28] or a power-law degree distribution [27]. This paper, in comparision, treats the scaled exponential exp {−t(I − P)} ei in a degree-weighted norm; it also shows a constant runtime independent of any network property. 6. EXPERIMENTAL RESULTS Here we compare hk-relax with a PageRank-based local clustering algorithm, pprpush [5]. Both algorithms accept as inputs a symmetric graph A and seed set s. The parameters required are t and ε, for hk-relax, and α and ε for pprpush. Both algorithms compute their respective diffusion ranks starting from the seed set, then perform a sweepcut on the resulting ranks. The difference in the algorithms lies solely in the diffusion used and the particular parameters. We conducted the timing experiments on a Dual CPU system with the Intel Xeon E5-2670 processor (2.6 GHz, 8 cores) with 16 cores total and 256 GB of RAM. None of the experiments needed anywhere near all the memory, nor any of the parallelism. Our implementation uses Matlab’s sparse matrix data structure through a C++ mex interface. It uses C++ unordered maps to store sparse vectors and is equivalent to the code in Figure 2. 6.1 Synthetic results In this section, we study the behavior of the PageRank and heat kernel diffusions on the symbolic image graph of a chaotic function f [50]. The graphs that result are loosely reminiscent of a social network because they have pockets of structure, like communities, and also chaotic behaviour that results in a small-world like property. The symbolic image of a function f is a graph where each node represents a region of space, and edges represent the action of the function in that region of space. We consider two-dimensional functions f (x, y) in the unit square [0, 1]2 so that we can associate each node with a pixel of an image and illustrate the vectors as images. In Figure 3 (left), we illustrate how the graph construction works. In the remaining examples, we let f be a map that results from a chaotic, nonlinear dynamical system [50] (we use the T10 construction with k = 0.22, η = 0.99 and sample 1000 points from each region of space, then symmetrize the result and discard weights). We discretize space in a 512 × 512 grid, which results in a 262, 144 node graph with 2M edges. In Figure 3 (right), we also show the PageRank vector with uniform teleportation as an image to “illustrate the structure” in the function f . Next, in Figure 4, we compare the vectors and sets identified by both diffusions starting from a single seed node. We chose the parameters α = 0.85 and t = 3 so the two methods perform the same amount of work. These results are what would be expected from Figure 1, and what many of the remaining experiments show: PageRank diffuses to a larger region of the graph whereas the heat-kernel remains more focused in a sub-region. PageRank, then, finds a large community with about 5,000 nodes whereas the heat kernel finds a small community with around 452 nodes with slightly worse conductance. This experiment suggests that, if these results also hold in real-world networks, then because real-world communities are often small [35], the heat kernel diffusion should produce more accurate communities in real-world networks. (a) ppr vector α = 0.85, ε = 10−5 (b) ppr set, φ = 0.31, size = 5510 1 y 0 f(x,y) 0 x 1 Figure 3: (Left) An illustration of the symbolic image of a function f as a graph. Each large, blue node represents a region of space. Thick, blue edges represent how the thin, red values of the function behave in that region. (Right) The global PageRank vector is then an image that illustrates features of the chaotic map f and shows that it has pockets of structure. (c) hk vector t = 3, ε = 10−5 (d) hk set, φ = 0.34, size = 455 Figure 4: When we compare the heat-kernel and PageRank diffusions on the symbolic image of the Chirikov map (see Figure 3), pprgrow finds a larger set with slightly better conductance, whereas hkgrow finds a tighter set with about the same conductance. In real-world networks, these smaller sets are more like real-world communities. 6.2 Runtime and conductance We next compare the runtime and conductance of the algorithms on a suite of social networks. For pprpush, we fix α = 0.99, then compute PageRank for multiple values of ε = 10−2 , 10−3 , 10−4 , 10−5 , and output the set of best conductance obtained. (This matches the way this method is commonly used in past work.) For hk-relax, we compute the heat kernel rank for four different parameter sets, and output the set of best conductance among them: (t, ε) = (10, 10−4 ); (20, 10−3 ); (40, 5 · 10−3 ); (80, 10−2 ). We also include in hk-relax an early termination criterion, in the case that the degrees of the nodes which are relaxed, P the sum of 1.5 dil , exceeds n . However, even the smaller input graphs (on which the condition is more likely to be met because of the smaller value of n1.5 ) do not appear to have reached this threshold. Furthermore, the main theorem of this paper imP plies that the quantity dil cannot exceed 2Nψε1 (t) . The datasets we use are summarized in Table 1; all datasets are modified to be undirected and a single connected component. These datasets were originally presented in the following papers [3, 10, 11, 32, 33, 34, 35, 40, 43, 44, 52, 55]. To compare the runtimes of the two algorithms, we display in Figure 5 for each graph the 25%, 50%, and 75% percentiles of the runtimes from 200 trials performed. For a given graph, each trial consisted of choosing a node of that graph uniformly at random to be a seed, then calling both the PageRank and the heat kernel algorithms. On the larger datasets, which have a much broader spectrum of node degrees and therefore greater variance, we instead performed 1,000 trials. Additionally, we display the 25%, 50%, and 75% percentiles of the conductances achieved during the exact same set of trials. The trendlines of these figures omit some of the trials in order to better show the trends, but all of the median results are plotted (as open circles). The figures show that in small graphs, hk-relax is faster than pprpush, but gets larger (worse) conductance. The picture reverses for large graphs where hk-relax is slower but finds smaller (better) conductance sets. Cluster size and conductance. We highlight the individual results of the previous experiment on the symmetrized twitter network. Here, we find that hk-relax finds sets of better conductance than pprpush at all sizes of communities in the network. See Figure 6. 6.3 Clusters produced vs. ground-truth We conclude with an evaluation of identifying groundtruth communities in the com-dblp, com-lj, com-amazon, com-orkut, com-youtube, and com-friendster datasets [40, 55]. In this experiment, for each dataset we first located 100 known communities in the dataset of size greater than 102 2 We thank AmirMahdi Ahmadinejad for pointing out a small mistake in our original implementation of this experiment – though we reported in Table 2 in our original publication [29] that we carried out this experiment on 100 commmunities of size greater than 10, Ahmadinejad noted that in practice we had used some communities of size less than 10. The results shown here in Table 2 are from the corrected experiment, and in fact are even more favorable for our hk-relax algorithm than our original reported results. 10,680 17,903 19,365 22,963 25,000 27,519 33,696 36,458 119,130 190,914 226,413 513,969 5,363,260 41,652,230 65,608,366 334,863 317,080 1,134,890 3,997,962 3,072,441 hkgrow 50% 25% 75% pprgrow 50% 25% 75% \|E\| 24,316 196,972 96,616 48,436 56,071 116,181 180,811 171,735 704,267 607,610 716,460 3,190,452 50,030,085 1,202,513,046 1,806,067,135 925,872 1,049,866 2,987,624 34,681,189 117,185,083 1.5 Runtime (s) pgp-cc ca-AstroPh-cc marvel-comics-cc as-22july06 rand-ff-25000-0.4 cond-mat-2003-cc email-Enron-cc cond-mat-2005-fix-cc soc-sign-epinions-cc itdk0304-cc dblp-cc flickr-bidir-cc ljournal-2008 twitter-2010 friendster com-amazon com-dblp com-youtube com-lj com-orkut 2 1 0.5 0 5 6 7 log10(\|V\|+\|E\|) 8 Conductances: hk vs. ppr 0 −1 10 hkgrow 50% 25% 75% pprgrow 50% 25% 75% −2 10 Table 2: The result of evaluating the heat kernel (hk) vs. PageRank (pr) on finding real-world communities. The heat kernel finds smaller, more accurate, sets with slightly worse conductance. data F1 -measure conductance set size hk pr hk pr hk pr amazon dblp youtube lj orkut friendster 0.608 0.364 0.128 0.138 0.183 0.125 0.415 0.273 0.078 0.104 0.116 0.112 0.124 0.238 0.477 0.564 0.824 0.866 0.050 0.144 0.361 0.524 0.736 0.860 145 156 137 156 95 96 5,073 4,529 5,833 299 476 218 . Given one such community, using every single node as an individual seed, we looked at the sets returned by hk-relax with t = 5, ε = 10−4 and pprpush using the standard procedure. We picked the set from the seed that had the highest F1 measure. (Recall that the F1 measure is a harmonic mean of precision and recall.) We report the mean of the F1 measure, conductance, and set size, where the average is taken over all 100 trials in Table 2. These results show that hk-relax produces only slightly inferior conductance scores, but using much smaller sets with substantially better F1 measures. This suggests that hk-relax better captures the properties of real-world communities than the PageRank diffusion in the sense that the tighter sets produced by the heat kernel are better focused around real-world communities than are the larger sets produced by the PageRank diffusion. 7. CONCLUSIONS These results suggest that the hk-relax algorithm is a viable companion to the celebrated PageRank push algorithm and may even be a worthy competitor for tasks that require accurate communities of large graphs. Furthermore, 9 10 log10(Conductances) Graph Runtime: hk vs. ppr Table 1: Datasets \|V \| 5 6 7 log10(\|V\|+\|E\|) 8 9 Figure 5: (Top figure) Runtimes of the hk-relax vs. ppr-push, shown with percentile trendlines from a select set of experiments. (Bottom) Conductances of hk-relax vs. ppr-push, shown in the same way. we suspect that the hk-relax method will be useful for the myriad other uses of PageRank-style diffusions such as linkprediction [31] or even logic programming [53]. In the future, we plan to explore this method on directed networks as well as better methods for selecting the parameters of the diffusion. It is also possible that our new ideas may translate to faster methods for non-conservative diffusions such as the Katz [25] and modularity methods [45], and we plan to explore these diffusions as well. Acknowledgements This work was supported by NSF CAREER Award CCF1149756. 8. REFERENCES [1] B. Abrahao, S. Soundarajan, J. Hopcroft, and R. Kleinberg. On the separability of structural classes of communities. In KDD, pages 624–632, 2012. [2] A. H. Al-Mohy and N. J. Higham. Computing the action of the matrix exponential, with an application to exponential integrators. SIAM J. Sci. Comput., 33(2):488–511, March 2011. [3] R. Alberich, J. Miro-Julia, and F. Rosselló. Marvel universe looks almost like a real social network. arXiv preprint cond-mat/0202174, 2002. [4] N. Alon and V. D. Milman. λ1 , isoperimetric inequalities for graphs, and superconcentrators. J. of Comb. Theory, Series B, 38(1):73–88, 1985. 0.9 conductance 0.8 0.7 0.6 0.5 0.4 0.3 hk ppr 0.2 0 1 density 0.1 2 3 4 log10(clustersize) 5 hk ppr 0 0.2 0.4 0.6 conductance 0.8 1 Figure 6: The top figure shows a scatter plot of conductance vs. community size in the twitter graph for the two community detection methods; the bottom figure shows a kernel density estimate of the conductances achieved by each method, which shows that hk-relax is more likely to return a set of lower conductance. [5] R. Andersen, F. Chung, and K. Lang. Local graph partitioning using PageRank vectors. In FOCS, 2006. [6] R. Andersen and K. Lang. An algorithm for improving graph partitions. In SODA, pages 651–660, January 2008. [7] K. Avrachenkov, N. Litvak, D. Nemirovsky, and N. Osipova. Monte carlo methods in pagerank computation: When one iteration is sufficient. SIAM J. Numer. Anal., 45(2):890–904, February 2007. [8] B. Bahmani, K. Chakrabarti, and D. Xin. Fast personalized pagerank on MapReduce. In SIGMOD, pages 973–984, New York, NY, USA, 2011. ACM. [9] B. Bahmani, A. Chowdhury, and A. Goel. Fast incremental and personalized pagerank. Proc. VLDB Endow., 4(3):173–184, Dec. 2010. [10] M. Boguñá, R. Pastor-Satorras, A. Dı́az-Guilera, and A. Arenas. Models of social networks based on social distance attachment. Phys. Rev. E, 70(5):056122, Nov 2004. [11] P. Boldi, M. Rosa, M. Santini, and S. Vigna. Layered label propagation: A multiresolution coordinate-free ordering for compressing social networks. In WWW, pages 587–596, March 2011. [12] F. Bonchi, P. Esfandiar, D. F. Gleich, C. Greif, and L. V. Lakshmanan. Fast matrix computations for pairwise and columnwise commute times and Katz scores. Internet Mathematics, 8(1-2):73–112, 2012. [13] C. Borgs, M. Brautbar, J. Chayes, and S.-H. Teng. Multi-scale matrix sampling and sublinear-time pagerank computation. Internet Mathematics, Online, 2013. [14] F. Chung. The heat kernel as the PageRank of a graph. PNAS, 104(50):19735–19740, 2007. [15] F. Chung. A local graph partitioning algorithm using heat kernel pagerank. Internet Mathematics, 6(3):315–330, 2009. [16] F. Chung and O. Simpson. Solving linear systems with boundary conditions using heat kernel pagerank. In Algorithms and Models for the Web Graph, pages 203–219. Springer, 2013. [17] I. S. Dhillon, Y. Guan, and B. Kulis. Weighted graph cuts without eigenvectors a multilevel approach. IEEE Trans. Pattern Anal. Mach. Intell., 29(11):1944–1957, November 2007. [18] E. Estrada. Characterization of 3d molecular structure. Chemical Physics Letters, 319(5-6):713–718, 2000. [19] E. Estrada and D. J. Higham. Network properties revealed through matrix functions. SIAM Review, 52(4):696–714, 2010. [20] A. Farahat, T. LoFaro, J. C. Miller, G. Rae, and L. A. Ward. Authority rankings from HITS, PageRank, and SALSA: Existence, uniqueness, and effect of initialization. SIAM Journal on Scientific Computing, 27(4):1181–1201, 2006. [21] M. Fiedler. Algebraic connectivity of graphs. Czechoslovak Mathematical Journal, 23(98):298–305, 1973. [22] N. J. Higham. Functions of Matrices: Theory and Computation. SIAM, 2008. [23] G. Jeh and J. Widom. Scaling personalized web search. In WWW, pages 271–279. ACM, 2003. [24] U. Kang and C. Faloutsos. Beyond ‘caveman communities’: Hubs and spokes for graph compression and mining. In ICDM, pages 300–309, Washington, DC, USA, 2011. IEEE Computer Society. [25] L. Katz. A new status index derived from sociometric analysis. Psychometrika, 18(1):39–43, March 1953. [26] K. Kloster. Graph diffusions and matrix functions: Fast algorithms and localization results. PhD thesis, Purdue University, 2016. [27] K. Kloster and D. F. Gleich. A fast relaxation method for computing a column of the matrix exponential of stochastic matrices from large, sparse networks. arXiv, cs.SI:1310.3423, 2013. [28] K. Kloster and D. F. Gleich. A nearly-sublinear method for approximating a column of the matrix exponential for matrices from large, sparse networks. In Algorithms and Models for the Web Graph, page in press, 2013. [29] K. Kloster and D. F. Gleich. Heat kernel based community detection. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1386–1395. ACM, 2014. [30] R. I. Kondor and J. D. Lafferty. Diffusion kernels on graphs and other discrete input spaces. In ICML, pages 315–322, San Francisco, CA, USA, 2002. Morgan Kaufmann Publishers Inc. [31] J. Kunegis and A. Lommatzsch. Learning spectral graph transformations for link prediction. In ICML, pages 561–568, 2009. [32] H. Kwak, C. Lee, H. Park, and S. Moon. What is Twitter, a social network or a news media? In WWW, pages 591–600, 2010. [33] J. Leskovec, D. Huttenlocher, and J. Kleinberg. Signed networks in social media. In CHI, pages 1361–1370, 2010. [34] J. Leskovec, J. Kleinberg, and C. Faloutsos. Graph evolution: Densification and shrinking diameters. ACM Trans. Knowl. Discov. Data, 1:1–41, March 2007. [35] J. Leskovec, K. J. Lang, A. Dasgupta, and M. W. Mahoney. Community structure in large networks: Natural cluster sizes and the absence of large well-defined clusters. Internet Mathematics, 6(1):29–123, September 2009. [36] M. Liou. A novel method of evaluating transient response. APPENDIX Proceedings of the IEEE, 54(1):20–23, 1966. A. CONVERGENCE THEORY [37] Z. Q. Luo and P. Tseng. On the convergence of the coordinate descent method for convex differentiable Here we state our main result bounding the work required by minimization. J. Optim. Theory App., 72(1):7–35, 1992. hk-relax to approximate the heat kernel with accuracy as de10.1007/BF00939948. scribed in (3). [38] F. McSherry. A uniform approach to accelerated PageRank Theorem 1. Let P, t, ψk (t), and r be as in Section 4. If steps computation. In WWW, pages 575–582, 2005. of hk-relax are performed until all entries of the residual satisfy [39] B. V. Minchev and W. M. Wright. Norges et εdi , then hk-relax produces an approximation x of r(i, j) < 2Nψ teknisk-naturvitenskapelige universitet. 2005. j (t) [40] A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, and h = exp {−t(I − P)} s satisfying B. Bhattacharjee. Measurement and analysis of online social networks. In Proceedings of the 7th ACM SIGCOMM kD−1 exp {−t(I − P)} s − D−1 xk∞ < ε, conference on Internet measurement, pages 29–42. ACM, and the amount of work required is bounded by 2007. [41] C. Moler and C. Van Loan. Nineteen dubious ways to 2N (et − 1) 2N ψ1 (t) ≤ . work(ε) ≤ compute the exponential of a matrix. SIAM Review, ε εt 20(4):801–836, 1978. [42] C. Moler and C. Van Loan. Nineteen dubious ways to Producing x satisfying compute the exponential of a matrix, twenty-five years kD−1 exp {−t(I − P)} s − D−1 xk∞ < ε later. SIAM Review, 45(1):3–49, 2003. [43] M. Newman. Network datasets. is equivalent to producing y satisfying http://www-personal.umich.edu/~mejn/netdata/, 2006. kD−1 exp {tP)} s − D−1 yk∞ < et ε. [44] M. E. J. Newman. The structure of scientific collaboration networks. PNAS, 98(2):404–409, 2001. We will show that the error vector in the hk-relax steps, TN (tP)s− [45] M. E. J. Newman. Modularity and community structure in y, satisfies kD−1 TN (tP)s − D−1 yk∞ < et ε/2. networks. PNAS, 103(23):8577–8582, 2006. The following lemma expresses the error vector, TN (tP)s − y, [46] L. Orecchia and M. W. Mahoney. Implementing as a weighted sum of the residual blocks rk in the linear system regularization implicitly via approximate eigenvector (9), and shows that the polynomials ψk (t) are the weights. computation. In ICML, pages 121–128, 2011. [47] L. Orecchia, S. Sachdeva, and N. K. Vishnoi. Lemma 1. Let ψk (t) be defined as in Section 4. Then in the Approximating the exponential, the Lanczos method and notation of Section 4.3, we can express the error vector of hkan Otilde(m)-time spectral algorithm for balanced relax in terms of the residual blocks rk as follows separator. In STOC, pages 1141–1160, 2012. N [48] L. Orecchia and Z. A. Zhu. Flow-based algorithms for local X TN (tP)s − y = ψk (tP)rk (12) graph clustering. In SODA, pages 1267–1286, 2014. k=0 [49] S. E. Schaeffer. Graph clustering. Computer Science Review, 1(1):27–64, 2007. Proof. Consider (9). Recall that v = [v0 ; v1 ; · · · ; vN ] and [50] D. L. Shepelyansky and O. V. Zhirov. Google matrix, let S be the (N + 1) × (N + 1) matrix of 0s with first subdiagonal dynamical attractors, and ulam networks. Phys. Rev. E, 1 equal to [ 11 , 12 , · · · , N ]. Then we can rewrite this linear system 81(3):036213, March 2010. more conveniently as [51] X. Sui, T.-H. Lee, J. Whang, B. Savas, S. Jain, K. Pingali, (I − S ⊗ (tP))v = e1 ⊗ s. (13) and I. Dhillon. Parallel clustered low-rank approximation of graphs and its application to link prediction. In Let vk be the true solution vectors that the v̂k are approximating H. Kasahara and K. Kimura, editors, Languages and in (13). We showed in Section 4.3 that the error TN (tP)s − y Compilers for Parallel Computing, volume 7760 of Lecture is in fact the sum of the errors vk − v̂k . Now we will express Notes in Computer Science, pages 76–95. Springer Berlin TN (tP)s − y in terms of the residual partitions, i.e. rk . Heidelberg, 2013. At any given step we have r = e1 ⊗ s − (I − S ⊗ (tP))v̂, so pre[52] C. (The Cooperative Association for Internet Data multiplying by (I − S ⊗ (tP))−1 yields (I − S ⊗ (tP))−1 r = v − v̂, Analyais). Network datasets. because (I − S ⊗ (tP))v = e1 ⊗ s exactly, by definition of v. Note http://www.caida.org/tools/measurement/skitter/router_topology/, that (I − S ⊗ (tP))−1 r = v − v̂ is the error vector for the linear 2005. Accessed in 2005. system (13). From this, an explicit computation of the inverse [53] W. Y. Wang, K. Mazaitis, and W. W. Cohen. (see [28] for details) yields Programming with personalized pagerank: A locally     groundable first-order probabilistic logic. In Proceedings of ! v0 − v̂0 r0 N X the 22Nd ACM International Conference on Conference on     .. Sk ⊗ (tP)k  ..  . (14)  = Information & Knowledge Management, CIKM ’13, . . k=0 pages 2129–2138, New York, NY, USA, 2013. ACM. vN − v̂N rN [54] J. J. Whang, D. F. Gleich, and I. S. Dhillon. Overlapping For our purposes, the full block vectors v, v̂, r and their individual community detection using seed set expansion. In CIKM, partitions are unimportant: we want only their sum, because pages 2099–2108, New York, NY, USA, 2013. ACM. P TN (tP)s − y = N ), as previously discussed. [55] J. Yang and J. Leskovec. Defining and evaluating network k=0 (vk − v̂k P communities based on ground-truth. In Proceedings of the Next we use (14) to express N k=0 (vk − v̂k ), and hence ACM SIGKDD Workshop on Mining Data Semantics, TN (tP)s − y, MDS ’12, pages 3:1–3:8, New York, NY, USA, 2012. ACM. [56] Z. A. Zhu, S. Lattanzi, and V. Mirrokni. A local algorithm in terms of the residual blocks rk . We accomplish this by examfor finding well-connected clusters. In ICML, pages ining the coefficients of an arbitrary block rk in (14), which in 396–404, 2013. turn requires analyzing the powers of (S ⊗ tP). Fix a residual block rj−1 and consider the product with a single term (Sk ⊗ (tP)k ). Since rj−1 is in block-row j of r, it multiplies with only the block-column j of each term (Sk ⊗ (tP)k ), so we want to know what the blocks in block-column j of (Sk ⊗ (tP)k ) look like. From [28] we know that ( (j−1)! ej+m if 0 ≤ m ≤ N + 1 − j m S ej = (j−1+m)! 0 otherwise. This means that block-column j of (Sm ⊗ (tP)m ) contains only a (j−1)! single nonzero block, (j−1+m)! tm Pm , for all 0 ≤ m ≤ N + 1 − j. Hence, summing the n × n blocks in block-column j of each power (Sm ⊗ (tP)m ) yields N+1−j X m=0 (j − 1)!tm Pm (j − 1 + m)! (15) kD−1 rk k∞ < = m=0 k=0 = N X k!tm Pm (k + m)! ! ≤ which is bounded above by −1 kD ψk (tP)rk , as desired. P Now that we’ve shown TN (tP)s − y = N k=0 ψk (tP)rk we can upperbound kD−1 TN (tP)s − D−1 yk∞ . To do this, we first show that (16) To prove this claim, recall that P = AD . Since ψk (t) is a polynomial, we have D−1 ψk (tP) = D−1 ψk (tAD−1 ) = ψk (tD−1 A)D−1 , which equals ψk (tPT )D−1 . Taking the infinity norm and applying the triangle inequality to the definition of ψk (t) gives us kD−1 ψk (tP)rk k∞ ≤ N−k X m=0 tm k! k(PT )m (D−1 rk )k∞ . (m + k)! we have k(PT )m (D−1 rk )k∞ ≤ equals k(D−1 rk )k∞ because PT Then k(PT )m k∞ k(D−1 rk )k∞ which is row-stochastic. Returning to our original objective, we then have kD−1 ψk (tP)rk k∞ ≤ N−k X m=0 tm k! k(D−1 rk )k∞ (m + k)! = ψk (t)kD−1 rk k∞ , proving the claim in (16). This allows us to continue: TN (tP)s − y = N X ψk (tP)rk so k=0 kD−1 TN (tP)s − D−1 yk∞ ≤ N X kD−1 ψk (tP)rk k∞ k=0 ≤ N X ψk (t)kD−1 rk k∞ . k=0 The stopping criterion for hk-relax requires that every entry of the residual satisfies r(i, j) < to satisfying r(i,j) di < et ε . 2Nψj (t) et εdi , 2Nψj (t) which is equivalent This condition guarantees that et ε/2. ψk (t) et ε 2N ψk (t) Finally, we have that −1 TN (tP)s − D yk∞ < et ε/2 implies kD−1 (exp {−t(I − P)} s − x)k∞ < ε/2, completing our proof of the first part of Theorem 1. k=0 −1 ψk (t)kD−1 rk k∞ N X k=0 rk kD−1 ψk (tP)rk k∞ ≤ ψk (t)kD−1 rk k∞ . N X k=0 m=0 N−k X N X Thus we have kD−1 TN (tP)s − D−1 yk∞ ≤ as the matrix coefficient of rj−1 in the right-hand side expression P PN m T of N ⊗ (tP)m )r. k=0 (vk − v̂k ) = (e ⊗ I)( m=0 S Substituting k = j − 1 on the right-hand side of (15) yields ! N N X X T m m (vk − v̂k ) = (e ⊗ I) S ⊗ (tP) r k=0 et ε . 2Nψk (t) Bounding work. It remains to bound the work required to perform steps of hkrelax until the stopping criterion is achieved. Because the stopping criterion requires each residual entry to satisfy r(i, j) et ε < di 2N ψj (t) we know that each step of hk-relax must operate on an entry r(i, j) larger than this threshold. That is, each step we relax an r(i,j) et ε ≥ 2Nψ . entry of r satisfying d (t) i j Consider the solution vector y ≈ TN (tP)s and note that each entry of y is really a sum of values that we have deleted from r. But r always contains only nonnegative values: the seed vector s is nonnegative, and each step involves setting entry r(i, j) = 0 and adding a scaled column of P, which is nonnegative, to r. Hence, PT kyk1 equals the sum of the r(i, j) added to y, l=1 r(il , jl ) = kyk1 . Finally, since the entries of y are nondecreasing (because they only change if we add positive values r(i, j) to them), we know that kyk1 ≤ kTN (tP)sk1 = ψ0 (t) ≤ et . This implies, then, that T X r(il , jl ) ≤ et . l=1 Using the fact that the values of r(i, j) that we relax must satisfy r(i,j) et ε ≥ 2Nψ we have that d (t) i j T X dil et ε ≤ et . 2N ψjl (t) l=1 Simplifying yields T X l=1 dil 2N ≤ . ψjl (t) ε By Lemma 1 we know ψk (t) ≤ ψ1 (t) for each k ≥ 1, so we can PT PT di di 2N l l l=1 ψ (t) ≤ l=1 ψ (t) ≤ ε , giving us lowerbound jl 1 T X l=1 dil ≤ 2N ψ1 (t) . ε Finally, note that the dominating suboperation in each step of hk-relax consists of relaxing r(i, j) and spreading this “heat kernel rank” to the neighbors of node i in block rj+1 . Since node i has di neighbors, this step consists of di adds, and so the work P performed by hk-relax is T l=1 dil , which is exactly the quantity we bounded above. This figure "chirikov-hk-set.png" is available in "png" format from: http://arxiv.org/ps/1403.3148v4 This figure "chirikov-hk-vec.png" is available in "png" format from: http://arxiv.org/ps/1403.3148v4 This figure "chirikov-pr-set.png" is available in "png" format from: http://arxiv.org/ps/1403.3148v4 This figure "chirikov-pr-vec.png" is available in "png" format from: http://arxiv.org/ps/1403.3148v4 This figure "chirikov-pr.png" is available in "png" format from: http://arxiv.org/ps/1403.3148v4	8cs.DS
Modeling The Intensity Function Of Point Process Via Recurrent Neural Networks Shuai Xiao1 , Junchi Yan23 , Stephen M. Chu3 , Xiaokang Yang1 , Hongyuan Zha4 arXiv:1705.08982v1 [cs.LG] 24 May 2017 1 Shanghai Jiao Tong University 2 East China Normal University 3 IBM Research – China 4 Georgia Tech {benjaminforever,xkyang}@sjtu.edu.cn, jcyan@sei.ecnu.edu.cn, schu@us.ibm.com, zha@cc.gatech.edu Abstract Event sequence, asynchronously generated with random timestamp, is ubiquitous among applications. The precise and arbitrary timestamp can carry important clues about the underlying dynamics, and has lent the event data fundamentally different from the time-series whereby series is indexed with fixed and equal time interval. One expressive mathematical tool for modeling event is point process. The intensity functions of many point processes involve two components: the background and the effect by the history. Due to its inherent spontaneousness, the background can be treated as a time series while the other need to handle the history events. In this paper, we model the background by a Recurrent Neural Network (RNN) with its units aligned with time series indexes while the history effect is modeled by another RNN whose units are aligned with asynchronous events to capture the long-range dynamics. The whole model with event type and timestamp prediction output layers can be trained end-to-end. Our approach takes an RNN perspective to point process, and models its background and history effect. For utility, our method allows a black-box treatment for modeling the intensity which is often a pre-defined parametric form in point processes. Meanwhile end-to-end training opens the venue for reusing existing rich techniques in deep network for point process modeling. We apply our model to the predictive maintenance problem using a log dataset by more than 1000 ATMs from a global bank headquartered in North America. Introduction Event sequence is becoming increasingly available in a variety of applications such as e-commerce transactions, social network activities, conflicts, and equipment failures etc. The event data can carry rich information not only about the event attribute (e.g. type, participator) but also the timestamp {zi , ti }N i=1 indicating when the event occurs. Being treated as a random variable when the event is stochastically generated in an asynchronous manner, the timestamp makes the Correspondence author is Junchi Yan. This research was partially supported by The National Key Research and Development Program of China (2016YFB1001003), NSFC (61602176, 61672231, 61527804, 61521062), STCSM (15JC1401700, 14XD1402100), China Postdoctoral Science Foundation Funded Project (2016M590337), the 111 Program (B07022) and NSF (IIS-1639792, DMS-1620345). Copyright c 2017, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. event sequence fundamentally different from the time series (Montgomery, Jennings, and Kulahci 2015) with equal and fixed time interval, whereby the time point only serves as a role for index {yt }Tt=1 . A major line of research (Aalen, Borgan, and Gjessing 2008) has been devoted to study event sequence, especially exploring the timestamp information to model the underlying dynamics of the system, whereby point process (Snyder and Miller 2012) has been a powerful and compact framework in this direction. Recently there are many machine learning based models for scalable point process modeling. We attribute the progressions in this direction in part to the smart mathematical reformulations and optimization techniques e.g. (Lewis and Mohler 2011; Zhou, Zha, and Song 2013b; Zhou, Zha, and Song 2013a), as well as novel parametric forms for the conditional intensity function (Shen et al. 2014; Ertekin, Rudin, and McCormick 2015; Xu et al. 2016) as carefully designed by researchers’ prior knowledge to capture the character of the dataset in study. However, one major limitation of the parametric forms of point process is due to their specialized and restricted expression capability for arbitrary distributed event data which trends to be oversimplified or even infeasible for capturing the problem complexity in real applications. Moreover, it runs at the risk of model underfitting due to the misjudgement on model choice. Recent works e.g. (Zhou, Zha, and Song 2013b) start to turn to non-parametric form to fit the structure of a point process, but their method is under the Hawkes process formulation,while this formulation runs at the risk of model mischoice. In this paper, we view the conditional intensity of a point process as a nonlinear mapping between the predicted transient occurrence intensity of events with different types, and the model input information of event participators, event profile and the system history. Such a nonlinear mapping is expected to be complex and flexible enough to model various characters of real event data for its application utility. In fact, deep learning models, such as Convolutional Neural Networks (CNNs) (LeCun et al. 1998), Recurrent Neural Networks (RNNs) (Pascanu, Mikolov, and Bengio 2013) have attracted wide attention in recent vision, speech and language communities, and many of them has dominated the competing results on perceptual benchmark tasks e.g. (Russakovsky et al. 2015). In particular, we turn to RNNs as a Figure 1: Time series and event sequence can be synergically modeled. The former can be used to timely capture the recent window for the time-varying features, while the latter can capture the long-range dependency over time. Note the dependency in an event sequence can be easily captured by an event sequence LSTM with less than 5 steps, while it takes too much more steps if using a time series with a fixed time interval e.g. 7 days in the figure (Note many unit steps in the top time series are omitted in the figure for clarity). natural way to encode such nonlinear and dynamic mapping, in an effort for modeling an end-to-end nonlinear intensity mapping without any prior knowledge. Key idea and highlights Our model interprets the conditional intensity function of a point process as a nonlinear mapping, which is synergetically established by a composite neural network with two RNNs as its building blocks. As illustrated in Fig.1, time series (top row) and event sequence (bottom row) are distinct to each other in that time series is more suitable to carry the synchronously (i.e. in a fixed pace) and regularly updated or constant profile features, while the event sequence can compactly catch event driven, more abrupt information, which can affect the condition intensity function over longer time period. More specifically, the highlights of this paper are: 1) We first make an observation that many conditional intensity functions can be viewed as an integration of two effects: i) spontaneous background component inherently affected by the internal (time-varying) attributes of the individual and the event type; ii) effects from history events. Meanwhile, most information in real world can also be covered by continuously updated features like age, temperature, and asynchronous event data such as clinical records, failures. This motivates us to devise a general approach. 2) Then we use an RNN whose units are aligned with the time points of a time series as its units, and an RNN whose units are aligned with events. The time series RNN can timely track the spontaneous background while the event sequence RNN is used to efficiently capture the long-range dependency over history with arbitrary time intervals. This allows to fit arbitrary dynamics of point process which otherwise will be very difficult or often impossible to be specified by a parameterized model under certain assumptions. 3) To our best knowledge, this is the first work to fully in- terpret and instantiate the conditional intensity function with fused time series and event sequence RNNs. This opens up the room for connecting the neural network techniques to traditional point process that emphasizes more on specific model driven by domain knowledge. More importantly the introduction of a full RNN treatment lessen the efforts for the design of (semi-)parametric point process model and its complex learning algorithms which often call for special tricks that prohibiting the wide use for practitioners. In contrast, neural networks and specifically RNN, is becoming off-the-shelf tools and getting widely used recently. 4) Our model is simple, general and can be end-to-end trained. We target to a predictive maintenance problem. The data is from a global bank headquartered in North America, consisting decades of thousands of event logs for a large number of Automated Teller Machines (ATMs). The stateof-the-art performance on failure type and timestamp prediction corroborates its suitability to real-world applications. Related Work and Motivation We view the related concepts and work in this section, which is mainly focused on Recurrent Neural Networks (RNNs) and their applications in time series and sequences data, respectively. Then we give our point of view on existing point process methods and their connection to RNNs. All these observations indeed motivate the work of this paper. Recurrent neural network The building blocks of our model are the Recurrent Neural Networks (RNNs) (Elman 1990; Pascanu, Mikolov, and Bengio 2013) and its modern variant Long Short-Term Memory (LSTM) units (Hochreiter and Schmidhuber 1997; Graves 2013). RNNs are dynamical systems whose next state and output depend on the present network state and input, which are more general models than the feed-forward networks. RNNs have long been explored in perceptual applications for many decades, however it can be very difficult to train RNNs to learn long-range dynamics perhaps in part due to the vanishing and exploding gradients problem. LSTMs provide a solution by incorporating memory units that allow the network to learn when to forget previous hidden states and when to update hidden states given new information. Recently, RNNs and LSTMs have been successfully applied in large-scale vision (Gregor et al. 2015), speech (Graves, rahman Mohamed, and Hinton 2014) and language (Sutskever, Vinyals, and Le. 2014) problems. RNNs for series data From application perspective, we view RNNs works by two scenarios as particularly considered in this paper: i) RNNs for synchronized series with evenly spaced interval e.g. time series or indexed sequence with pure order information e.g. language; ii) asynchronous sequence with random timestamp e.g. event data. i) Synchronized series: RNNs have been a long time a natural tool for standard time series modeling and prediction (Connor, Martin, and Atlas 1994; Han et al. 2004; Chandra and Zhang 2012; Chen, Chang, and Chang 2013), whereby the indexed series data point is fed as input to an (unfold) RNN. In a broader sense, video frames can also be treated as time series and RNN are widely used in recent visual analytics works (Jain et al. 2016; Tripathi et al. 2016) and so for speech (Graves, rahman Mohamed, and Hinton 2014). RNNs are also intensively adopted for sequence modeling tasks (Chung et al. 2014; Bengio et al. 2015) when only order information is considered. ii) Asynchronous event: In contrast, event sequence with timestamp about their occurrence, which are asynchronously and randomly distributed over the continuous time space, is another typical input type for RNNs (Du et al. 2016; Choi et al. 2016; Esteban et al. 2016) and (Che et al. 2016) (despite its title for ’time series’). One key differentiation against the first scenario is that the timestamp or time duration between events (together with other features) is taken as input to the RNNs. By doing so, (long-range) event dependency can be effectively encoded. Point process Point process has been a principled framework for modeling event data (Aalen, Borgan, and Gjessing 2008). The dynamics of the point process can be well captured by its conditional intensity function whose definition is briefly reviewed here: for a short time window [t, t + dt), λ(t) represents the rate for the occurrence of a new event conditioned on the history Ht = {zi , ti \|ti < t}: λ(t) = lim ∆t→0 E(N (t + ∆t) − N (t)\|Ht ) E(dN (t)\|Ht ) = ∆t dt where E(dN (t)\|Ht ) is the expectation of the number of events happened in the interval (t, t + dt] given the historical observations Ht . The conditional intensity function has played a central role in point processes and many popular processes vary on how it is parameterized. 1) Poisson process (Kingman 1992): the homogeneous Poisson process has a very simple form for its intensity function: λ(t) = λ0 . Poisson process and its time-varying generalization are both assumed to be independent of the history. 2) Reinforced poisson processes (Pemantle 2007; Shen et al. 2014): the model captures the ‘rich-get-richer’ mechanism characterized by a compact intensity function, which is recently used for popularity prediction (Shen et al. 2014). 3) Hawkes process (Hawkes 1971): Hawkes process has received wide attention recently in social network analysis (Zhou, Zha, and Song 2013a), viral diffusion(Yang and Zha 2013) and criminology (Lewis et al. 2010) etc. It explicitly uses a triggering term to model the excitation effect from history events and is originally motivated to analyze the earthquake and its aftershocks(Ogata 1988). 4) Reactive point process (Ertekin, Rudin, and McCormick 2015): it can be regarded as a generalization for the Hawkes process by adding a self-inhibiting term to account for the inhibiting effects from history events. 5) Self-correcting process (Isham and Westcott 1979): its background part increases steadily, while it is decreased by a constant e−α < 1 every time a new event appears. We reformulate these intensity functions in their general form in Table 1. It tries to separate the spontaneous background component and history event effect explicitly. Predictive maintenance Predictive maintenance (Mobley 2002) is a sound testbed for our model which refers to a practice that involves equipment risk prediction to allow for proactive scheduling of corrective maintenance. Such an early identification of potential concerns helps deploy limited resources more cost effectively, reduce operations costs Table 1: Conditional intensity functions of point processes. Model Poisson process Reinforced poisson process Hawkes process Reactive point process Self-correcting process Background History event effect µ(t) 0 µ(t) µ(t) 0 0 P γ(t) P ti <t δ(ti < t) P P ti <t γ(t, ti ) ti <t γ1 (t, ti ) − P ti <t γ2 (t, ti ) exp(µt − ti <t γ(t, ti ) Note:δ(t) is Dirac function, γ(t, ti ) is time-decaying kernel and µ(t) can be constant or time-varying function. and maximize equipment uptime (Grall et al. 2002). Predictive maintenance is adopted in a wide variety of applications such as fire inspection (Madaio et al. 2016), data center (Sirbu and Babaoglu 2015) and electrical grid (Ertekin, Rudin, and McCormick 2015) management. For its practical importance in different scenarios and relative rich event data for modeling, we target our model to a real-world dataset of more than 1,000 automated teller machines (ATMs) from a global bank headquartered in North America. Network Structure and End-to-End Learning Taking a sequence {x}Tt=1 as input, the RNN generates the hidden states {h}Tt=1 and outputs a sequence (Elman 1990; Pascanu, Mikolov, and Bengio 2013). Specifically, we implement our RNN with Long Short Term Memory (LSTM) (Hochreiter and Schmidhuber 1997; Graves 2013) for its popularity and well-known capability for efficient longrange dependency learning. In fact other RNN variant e.g. Gated Recurrent Units (GRU) (Chung et al. 2014) can also be alternative choice. We reiterate the formulation of LSTM: it ft ct ot ht = σ(Wi xt + Ui ht−1 + Vi ct−1 + bi ), = σ(Wf xt + Uf ht−1 + Vf ct−1 + bf ), = ft ct−1 + it tanh(Wc xt + Uc ht−1 + bc ), = σ(Wo xt + Uo ht−1 + Vo ct + bo ), = ot tanh(ct ) where denotes element-wise multiplication and the recurrent activation σ is the Logistic Sigmod function. The above system can be reduced into an LSTM equation: (ht , ct ) = LSTM(xt , ht−1 + ct−1 ) We consider two types of input: i) continuously and evenly distributed time-series data e.g. temperature; ii) event data whose occurrence time interval is random. The network is comprised by two RNNs using evenly spaced time series {yt }Tt=1 to model the background intensity of events occurrence and event sequence {zi , ti }N i=1 to capture long-range event dependency. As a result, we have: (hyt , cyt ) = LSTMy (yt , hyt−1 + cyt−1 ), (1) (hzt , czt ) (2) et Ut ut st = LSTMz (zt , hzt−1 tanh(Wf [hyt , hzt ] + czt−1 ), = + bf ), = softMax(WU et + bU ), = softMax(Wu [et , Ut ] + bu ) = Ws et + bs , (3) (4) (5) (6) where U and u denotes the main type and subtype of events respectively. s is the timestamp associated with each event. The total loss is the sum of the time prediction loss and the cross-entropy loss for event type: N X −WUj log(Utj ) − wuj log(ujt ) − log f (sjt \|hjt−1 ) (7) j=1 where N is the number of training samples indexed by j, and sjt is the timestamp for the coming event, while hjt−1 is the history information. The underlying rationale for the third term is that we not only encourage correct classification of the coming event type, but also reinforce the corresponding timestamp of the event shall be close to the ground truth. We adopt a Gaussian penalty function with a fixed σ 2 = 10: ! 1 −(sjt − s̃jt )2 j j f (st \|ht−1 ) = √ exp 2σ 2 2πσ Figure 2: Our network can be trained end-to-end. Time series and event sequence are fed into two RNNs (LSTM) which are connected to an embedding mapping layer that fuses the information from two LSTMs. Then three prediction layers are used to output the predicted main type, subtype of events, and the associated timestamp. Cross-entropy with the time penalty loss by Eq.7 and square loss are respectively used for event type and timestamp prediction. total max min mean std Training set Table 2: Statistics of main/sub-type of event count per ATM, and timestamp interval in days for all ATMs (in brackets). Ticket Error PRT CNG IDC COMM LMTP MISC 2226(–) 28434(–) 9204(–) 7767(–) 4082(–) 3371(–) 2525(–) 1485(–) 10(137.04) 168(153.90) 88(210.13) 50(200.07) 116(206.61) 47(202.79) 81(207.93) 32(204.41) 0(1.21) 0(0.10) 0(0.10) 0(0.10) 0(0.10) 0(0.10) 0(0.10) 0(0.10) 2.09(31.70) 26.70(6.31) 8.64(12.12) 7.29(15.49) 3.83(23.85) 3.16(22.35) 2.37(22.86) 1.39(24.27) 1.85(25.14) 18.38(9.74) 11.37(21.41) 6.59(23.87) 5.84(30.71) 3.90(29.36) 4.41(34.56) 2.54(34.38) Testing set The output s̃jt from the timestamp prediction layer is fed to the classification loss layer to compute the above penalty given the actual timestamp sjt for sample i. Following the importance weighting methodology for skewed data of model training (Rosenberg 2012), the weight parameters W, w for both main-type and subtype are set as the inverse of the sample number ratio in that type against the total size of samples, in order to weight more on those classes with fewer training samples. For the loss of independent main-type or subtype prediction as shown in Fig.3(b), we set the weight parameter w and W to zero respectively. We adopt RMSprop gradients (Dauphin et al. 2015) which have been shown to work well on training deep networks to learn these parameters. data Ticket Error PRT CNG IDC COMM LMTP MISC 1164(–) 11799(–) 4089(–) 3134(–) 1645(–) 1366(–) 939(–) 626(–) 15(148.00) 104(193.75) 60(205.48) 66(196.93) 35(205.75) 53(205.17) 20(186.87) 21(190.75) 0(0.13) 1(0.10) 0(0.10) 0(0.10) 0(0.10) 0(0.10) 0(0.10) 0(0.10) 2.52(26.30) 25.59(6.47) 8.86(11.45) 6.79(15.90) 3.57(25.15) 2.96(22.88) 2.04(26.63) 1.36(25.14) 2.41(25.22) 17.71(10.09) 10.77(20.44) 6.91(25.22) 4.29(31.62) 4.08(30.09) 2.76(36.96) 2.59(34.91) Experiments on Real-world Data We use failure prediction for predictive ATMs maintenance as a typical example of event based point process modeling. We have no prior knowledge on the dynamics of the complex system and the task can involve arbitrarily working schedules and heterogeneous mix of conditions. It takes much cost or even impractical to devise specialized models. Problem and real data description In maintenance support services, when a device fails, the equipment owner raises a maintenance service ticket and technician will be assigned to repair the failure. In fact, the history log and relevant profile information about the equipment can be indicative signals for the coming failures. The studied dataset is comprised of the event logs involving error reporting and failure tickets, which is originally collected from a large number of ATMs owned by an anonymous global bank headquartered in North America. The bank is also a customer of the technical support service department of a Fortune 500 IT company. ATM models The training data consists of 1085 ATMs and testing data has 469 ATMs, in total 1557 Wincor ATMs that cover 5 ATM machine models: ProCash 2100 RL (980, 430), 1500 RL (19, 5), 2100 FL (53, 21), 1500 FL (26, 10), type and 2250XE RL (7, 3). The numbers in the bracket indicate the number of machines for training and testing. Event type There are two main types ‘ticket’ and ‘error’ from Sep. 2014 to Mar. 2015. Statistics is presented in Table 2. Moreover ‘error’ is divided into 6 subtypes regarding in which component the error occurs: 1) printer (PRT), 2) cash dispenser module (CNG), 3) internet data center (IDC), 4) communication part (COMM), 5) printer monitor (LMTP), 6) miscellaneous e.g. hip card module, usb (MISC). Features The input features for the two RNNs are: 1) Time series RNN: For each sub-window of length 7 days, for the time series RNN, we extract features including: i) the inventory information: ATM models, age, location, etc; ii) Event statistics, including tickets events from maintenance records, and errors from system log. Their occurrence frequencies are used as features. The concatenation of the above two categories of features serves as the features for each sub-window i.e. time series point. 2) Event sequence RNN: event type and the time interval between two events. Model setting We use a single layer LSTM of size 32 with Sigmoid gate activations and tanh activation for hidden representation. The embedding layer is fully connected and it uses tanh activation and outputs a 16 dimensional vector. One-hot or embedding can be used for event type representation. For a large number of types, embedding representation is compact and efficient. For time series RNN, we set the length of each sub-window (i.e. the evenly spaced time interval) to be 7 days and the number of sub-window to be 5. In this way, our observation length is 35 days for time series. For event-dependency, the length of event sequence can be arbitrarily long. Here we take it by 7. We also test degraded versions of our model as follows: 1) Time series RNN: the input is event sequence (the right half in the yellow part of Fig.2) is removed. Note this design is in spirit similar to many LSTM models (Jain et al. 2016; Tripathi et al. 2016) used for video analytics, whereby the frame sequence can be treated as time series as the input to LSTMs. 2) Event (sequence) RNN: the RNN whose input is time series (the left half in the yellow part of Fig.2) is removed; 3) Intensity RNN: two RNN are fused as shown in Fig.2. For the above three methods, the output layer is directly the fine-grained subtype of events with no hierarchical structure as shown in the top left part of Fig.2) in dark green. We also term three ‘hierarchical’ versions whose two hierarchical prediction layers in Fig.2 are used: 4) Time series hRNN, 5) Event (sequence) hRNN, 6) Intensity hRNN. In addition, we compare three major peer methods. For Logistic model, the input are the concatenation of feature vectors for all active time series RNN sub-windows (set to 5 in this paper). For RMTPP and Hawkes process, we train the model on the event sequences with associated information. In fact, RMTPP will further process the event data into the similar input information to our event RNN. 1) Logistic model: we use Logistic regression for event timestamp prediction and use another independent Logistic classification model for event type prediction. 2) Recurrent Marked Temporal Point Processes (RMTPP): (Du et al. 2016) uses neural network to model the event dependency flexibly. The method can only sample transient time series features when an event happens and use partially parametric form for the base intensity. 3) Hawkes Process: To enable multi-type event prediction, we use a Multi-dimensional Hawkes process. Similar to (Zhou, Zha, and Song 2013a), we also add a sparsity regularization term on the mutual infection matrix but the lowrank assumption is removed as we only have 6 subtypes. Evaluation metrics We use several popular prediction metrics for performance evaluation. For the coming event type prediction, we adopt Precision, Recall, F1 Score and Confusion matrix over 2 main types (‘error’, ‘ticket’) as well Figure 3: Hierarchical layer and Flat independent layer. as Confusion matrix over 6 subtypes under ‘error’. Note all these metrics are computed for each type, and then are averaged over all types. For event time prediction, we use the Mean Absolute Error (MAE) which measures the absolute difference between the predicted time point and the actual one. These settings are similar to (Du et al. 2016). To evaluate the type and timestamp prediction jointly, we devise two more strict metrics. For type prediction, we narrow down to test the samples whose timestamp prediction error MAE< 3 days and we compute the new F1 score+. For timestamp, we recompute the new MAE+ only for the samples whose coming event is correctly predicted. Platform The code is based on Theano running on a Linux server with 32G memory, 2 CPUs with 6 cores for each: Intel(R) Xeon(R) CPU E5-2603 v3 @ 1.60GHz. We also use 4 GPU:GeForce GTX TITAN X for acculturation. Results and discussion All evaluations are performed on the testing dataset distinctive to training set whose statistics are shown in Table 2. Averaged performance Table 3 shows the averaged performance among various types of events. As shown in Fig.3, we test two architectures of the event type prediction layer, i.e. hierarchical predictor (Fig.3(a)) and flat independent predictors (Fig.3(b)). The main type includes ‘ticket’ and ‘error’ and the subtype include ‘ticket’ and the other six subtypes under ‘error’ as we describe earlier in the paper. Confusion matrix The confusion matrix for the six subtypes under ‘error’ event, as well as for the two main types ‘ticket’ and ‘error’ are shown in Fig.4 by various methods. We make observations and analysis based on the results: 1) As shown in 3, for main-type, the flat architecture that directly predicts the main types outperforms the hierarchical one in different settings of the input RNN as well as varying evaluation metrics. This can be explained that the loss function focuses on the main-type misclassification only. While for the subtype prediction, the hierarchical layer performs better since it fuses the output from the main-type prediction layer and the embedding layer as shown in Fig.3(a). 2) No surprisingly, for both event type and timestamp prediction, our main approach, i.e. intensity RNN that fuses two RNNs outperforms its counterparts time series RNN and event sequence RNN by a notable margin. While the event RNN also often performs better than the time series counterpart. This suggests at least in the studied dataset, history event effects are important for the future event occurrence. 3) Our main method intensity RNN is almost always superior against other methods except for the main-type pre- Table 3: Ablation test of our method and peer methods i.e. multi-dimensional Hawkes process, recurrent Hawkes process and Logistic classification (for type) and regression (for event timestamp). Numbers are averaged over types. subtype precision Time series RNN Event sequence RNN Intensity RNN Hawkes process Logistic prediction RMTPP 0.673 0.671 0.714 0.457 0.883 0.581 0.554 0.570 0.620 0.387 0.385 0.574 0.582 0.623 0.642 — — — 0.590 0.614 0.664 —— — recall Time series RNN Event sequence RNN Intensity RNN Hawkes process Logistic prediction RMTPP 0.853 0.821 0.905 0.493 0.795 0.691 0.522 0.543 0.614 0.394 0.273 0.583 0.738 0.770 0.805 — — — 0.608 0.621 0.661 — — — F1 score Time series RNN Event sequence RNN Intensity RNN Hawkes process Logistic prediction RMTPP 0.707 0.703 0.765 0.473 0.832 0.584 0.533 0.555 0.616 0.386 0.269 0.572 0.571 0.651 0.662 — — — 0.605 0.610 0.663 — — — MAE (in days) Time series RNN Event sequence RNN Intensity RNN Hawkes process Logistic prediction RMTPP 4.37 4.24 4.13 5.26 4.52 4.28 4.48 4.42 4.20 5.46 4.61 4.32 4.26 4.21 4.02 — — — 4.41 4.37 4.13 — — — F1 Score+ hierarchical output main-type subtype main-type Time series RNN Event sequence RNN Intensity RNN Hawkes process Logistic prediction RMTPP 0.768 0.705 0.825 0.467 0.846 0.584 0.547 0.597 0.661 0.451 0.286 0.619 0.572 0.639 0.684 — — — 0.603 0.646 0.708 — — — MAE+ model Time series RNN Event sequence RNN Intensity RNN Hawkes process Logistic prediction RMTPP 4.21 4.16 4.12 5.42 4.5 4.26 3.78 3.84 3.57 3.93 4.24 3.99 4.05 4.12 4.21 — — — 3.97 4.01 4.11 — — — diction task, whereby the Logistic classification model performs better. However for more challenging tasks i.e. subtype prediction and event timestamp prediction, our method significantly outperforms especially for subtype prediction task. Interestingly, all point process based models obtain better results on this task which suggests the point process models are more promising compared with classical classification models. Indeed, our methodology provides an endto-end learning mechanism without any pre-assumption in modeling point process. All these empirical results on realworld tasks suggest the efficacy of our approach. Conclusion We use Fig.5 to conclude and further position our model in the development of (implicit and explicite) modeling the intensity function of point process. In fact, Hawkes process uses a full explicit parametric model and RMTPP misses the dense time series features to model time-varying base intensity and assumes a partially parametric form for it. We make a further step by a full implicit mapping model. Our model (a) Intensity hRNN (b) Time series hRNN (c) Event hRNN (d) Intensity RNN (e) Time series RNN (f) Event RNN (g) Hawkes process (j) i-hRNN (o) (k) t-hRNN e-RNN (p) (h) Logistic (l) e-hRNN (m) (q) Logistic Hawkes (i) i-RNN (r) RMTPP (n) t-RNN RMTPP Figure 4: Confusion matrixes for sub/main type. Three methods in top and middle row use the hierarchical and flat structure as in Fig.3 respectively. Zoom in for better view. Figure 5: The evolving of point process modeling. (see Fig.2) is simple, general and can be learned end-to-end with standard backpropagation and opens up new possibilities for borrowing the advances in neural network learning to the area of point process modeling and applications. The representative study in this paper has clearly suggests its high potential to real-world problems, even we have no domain knowledge on the problem at hand. This is in contrast to existing point process models where an assumption about the dynamics is often need to be specified beforehand. References [Aalen, Borgan, and Gjessing 2008] Aalen, O.; Borgan, O.; and Gjessing, H. 2008. Survival and event history analysis: a process point of view. Springer Science & Business Media. [Bengio et al. 2015] Bengio, S.; Vinyals, O.; Jaitly, N.; and Shazeer, N. 2015. Scheduled sampling for sequence prediction with recurrent neural networks. In NIPS. [Chandra and Zhang 2012] Chandra, R., and Zhang, M. 2012. Cooperative coevolution of elman recurrent neural networks for chaotic time series prediction. Neurocomputing 86:116–123. [Che et al. 2016] Che, Z.; Purushotham, S.; Cho, K.; Sontag, D.; and Liu, Y. 2016. Recurrent neural networks for multivariate time series with missing values. In arXiv:1606.01865. [Chen, Chang, and Chang 2013] Chen, P. A.; Chang, L. C.; and Chang, F. J. 2013. Reinforced recurrent neural networks for multistep-ahead flood forecasts. Journal of Hydrology 497:71–79. [Choi et al. 2016] Choi, E.; Bahadori, M. T.; Schuetz, A.; Stewart, W. F.; and Sun, J. 2016. Doctor ai: Predicting clinical events via recurrent neural networks. In arXiv:1511.05942v10. [Chung et al. 2014] Chung, J.; Gulcehre, C.; Cho, K.; and Bengio, Y. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. In arXiv:1412.3555. [Connor, Martin, and Atlas 1994] Connor, J. T.; Martin, R. D.; and Atlas, L. E. 1994. Recurrent neural networks and robust time series prediction. IEEE Transactions on Neural Networks 5(2):240–254. [Dauphin et al. 2015] Dauphin, Y. N.; de Vries, H.; Chung, J.; and Bengio, Y. 2015. Rmsprop and equilibrated adaptive learning rates for non-convex optimization. In arXiv:1502.04390. [Du et al. 2016] Du, N.; Dai, H.; Trivedi, R.; Upadhyay, U.; GomezRodriguez, M.; and Song, L. 2016. Recurrent marked temporal point processes: Embedding event history to vectore. In KDD. [Elman 1990] Elman, J. L. 1990. Finding structure in time. Cognitive Science 14:179–211. [Ertekin, Rudin, and McCormick 2015] Ertekin, S.; Rudin, C.; and McCormick, T. H. 2015. Reactive point processes: A new approach to predicting power failures in underground electrical systems. The Annals of Applied Statistics 9(1):122–144. [Esteban et al. 2016] Esteban, C.; Staeck, O.; Yang, Y.; and Tresp, V. 2016. Predicting clinical events by combining static and dynamic information using recurrent neural networks. In arXiv:1602.02685. [Grall et al. 2002] Grall, A.; Dieulle, L.; Berenguer, C.; and Roussignol, M. 2002. Continuous-time predictive-maintenance scheduling for a deteriorating system. IEEE transactions on Reliability 51(2):141–150. [Graves, rahman Mohamed, and Hinton 2014] Graves, A.; rahman Mohamed, A.; and Hinton, G. 2014. Towards end-to-end speech recognition with recurrent neural networks. In ICML. [Graves 2013] Graves, A. 2013. Generating sequences with recurrent neural networks. In arXiv:1308.0850. [Gregor et al. 2015] Gregor, K.; Danihelka, I.; Graves, A.; Rezende, D.; and Wierstra, D. 2015. Draw: A recurrent neural network for image generation. In ICML. [Han et al. 2004] Han, M.; Xi, J.; Xu, S.; and Yin, F.-L. 2004. Prediction of chaotic time series based on the recurrent predictor neural network. IEEE Transactions on Signal Processing 52(12):3409–3416. [Hawkes 1971] Hawkes, A. G. 1971. Spectra of some self-exciting and mutually exciting point processes. Biometrika. [Hochreiter and Schmidhuber 1997] Hochreiter, S., and Schmidhuber, J. 1997. Long short-term memory. Neural computation 9(8):1735–1780. [Isham and Westcott 1979] Isham, V., and Westcott, M. 1979. A self-correcting pint process. Advances in Applied Probability 37:629–646. [Jain et al. 2016] Jain, A.; Singh, A.; Koppula, H. S.; Soh, S.; and Saxena, A. 2016. Recurrent neural networks for driver activity anticipation via sensory-fusion architecture. In ICRA. [Kingman 1992] Kingman, J. F. C. 1992. Poisson processes. volume 3. Oxford university press. [LeCun et al. 1998] LeCun, Y.; Bottou, L.; Bengio, Y.; and Haffner, P. 1998. Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11):2278–2324. [Lewis and Mohler 2011] Lewis, E., and Mohler, E. 2011. A nonparametric em algorithm for multiscale hawkes processes. Journal of Nonparametric Statistics. [Lewis et al. 2010] Lewis, E.; Mohler, G.; Brantingham, P. J.; and Bertozzi, A. 2010. Self-exciting point process models of insurgency in iraq. UCLA CAM Reports 10 38. [Madaio et al. 2016] Madaio, M.; Chen, S.; Haimson, O.; Zhang, W.; Cheng, X.; and Hinds-Aldrich, M. 2016. Firebird: Predicting fire risk and prioritizing fire inspections in atlanta. In KDD. [Mobley 2002] Mobley, R. K. 2002. An introduction to predictive maintenance. Butterworth-Heinemann. [Montgomery, Jennings, and Kulahci 2015] Montgomery, D. C.; Jennings, C. L.; and Kulahci, M. 2015. Introduction to time series analysis and forecasting. John Wiley & Sons. [Ogata 1988] Ogata, Y. 1988. Statistical models for earthquake occurrences and residual analysis for point processes. J. Amer. Statist. Assoc. 83(401):9–27. [Pascanu, Mikolov, and Bengio 2013] Pascanu, R.; Mikolov, T.; and Bengio, Y. 2013. On the difficulty of training recurrent neural networks. In ICML. [Pemantle 2007] Pemantle, R. 2007. A survey of random processes with reinforcement. Probability Survey 4(0):1–79. [Rosenberg 2012] Rosenberg, A. 2012. Classifying skewed data: Importance weighting to optimize average recall. In INTERSPEECH, 2242–2245. [Russakovsky et al. 2015] Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.; Berg, A. C.; and Fei-Fei, L. 2015. Imagenet large scale visual recognition challenge. International Journal of Computer Vision 115(3):211–252. [Shen et al. 2014] Shen, H.; Wang, D.; Song, C.; and Barabási, A. 2014. Modeling and predicting popularity dynamics via reinforced poisson processes. In AAAI. [Sirbu and Babaoglu 2015] Sirbu, A., and Babaoglu, O. 2015. A holistic approach to log data analysis in high-performance computing systems: The case of ibm blue gene/q. Euro-Par 2015: Parallel Processing Workshopss 631–643. [Snyder and Miller 2012] Snyder, D. L., and Miller, M. I. 2012. Random point processes in time and space. Springer Science & Business Media. [Sutskever, Vinyals, and Le. 2014] Sutskever, I.; Vinyals, O.; and Le., Q. V. 2014. Sequence to sequence learning with neural networks. In NIPS. [Tripathi et al. 2016] Tripathi, S.; Lipton, Z. C.; Belongie, S.; and Nguyen, T. 2016. Context matters: Refining object detection in video with recurrent neural networks. In BMVC. [Xu et al. 2016] Xu, H.; Wu, W.; Nemati, S.; and Zha, H. 2016. Patient flow prediction via discriminative learning of mutuallycorrecting processes. IEEE transactions on Knowledge and Data Engineering. [Yang and Zha 2013] Yang, S.-h., and Zha, H. 2013. Mixture of mutually exciting processes for viral diffusion. In ICML. [Zhou, Zha, and Song 2013a] Zhou, K.; Zha, H.; and Song, L. 2013a. Learning social infectivity in sparse low-rank networks using multi-dimensional hawkes processe. In AISTATS. [Zhou, Zha, and Song 2013b] Zhou, K.; Zha, H.; and Song, L. 2013b. Learning triggering kernels for multi-dimensional hawkes processes. In ICML.	2cs.AI
arXiv:1404.1957v3 [math.PR] 29 Oct 2015 The Annals of Applied Probability 2015, Vol. 25, No. 6, 3511–3570 DOI: 10.1214/14-AAP1081 c Institute of Mathematical Statistics, 2015 ERGODIC CONTROL OF MULTI-CLASS M/M/N + M QUEUES IN THE HALFIN–WHITT REGIME By Ari Arapostathis∗,1, Anup Biswas∗,2 and Guodong Pang†,3 The University of Texas at Austin∗ and Pennsylvania State University† We study a dynamic scheduling problem for a multi-class queueing network with a large pool of statistically identical servers. The arrival processes are Poisson, and service times and patience times are assumed to be exponentially distributed and class dependent. The optimization criterion is the expected long time average (ergodic) of a general (nonlinear) running cost function of the queue lengths. We consider this control problem in the Halfin–Whitt (QED) regime, that is, the number of servers n and the total offered load r scale √ like n ≈ r + ρ̂ r for some constant ρ̂. This problem was proposed in [Ann. Appl. Probab. 14 (2004) 1084–1134, Section 5.2]. The optimal solution of this control problem can be approximated by that of the corresponding ergodic diffusion control problem in the limit. We introduce a broad class of ergodic control problems for controlled diffusions, which includes a large class of queueing models in the diffusion approximation, and establish a complete characterization of optimality via the study of the associated HJB equation. We also prove the asymptotic convergence of the values for the multi-class queueing control problem to the value of the associated ergodic diffusion control problem. The proof relies on an approximation method by spatial truncation for the ergodic control of diffusion processes, where the Markov policies follow a fixed priority policy outside a fixed compact set. Received April 2014; revised November 2014. Supported in part by the Office of Naval Research Grant N00014-14-1-0196. 2 Supported in part by an award from the Simons Foundation (# 197982) to The University of Texas at Austin and in part by the Office of Naval Research through the Electric Ship Research and Development Consortium. 3 Supported in part by the Marcus Endowment Grant at the Harold and Inge Marcus Department of Industrial and Manufacturing Engineering at Penn State. AMS 2000 subject classifications. Primary 60K25; secondary 68M20, 90B22, 90B36. Key words and phrases. Multi-class Markovian queues, reneging/abandonment, Halfin–Whitt (QED) regime, diffusion scaling, long time-average control, ergodic control, stable Markov optimal control, spatial truncation, asymptotic optimality. 1 This is an electronic reprint of the original article published by the Institute of Mathematical Statistics in The Annals of Applied Probability, 2015, Vol. 25, No. 6, 3511–3570. This reprint differs from the original in pagination and typographic detail. 1 2 A. ARAPOSTATHIS, A. BISWAS AND G. PANG CONTENTS 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1. Contributions and comparisons . . . . . . . . . . . . . . . 1.2. Organization . . . . . . . . . . . . . . . . . . . . . . . . . 1.3. Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. The controlled system in the Halfin–Whitt regime . . . . . . 2.1. The multi-class Markovian many-server model . . . . . . 2.2. The ergodic control problem in the Halfin–Whitt regime 3. A broad class of ergodic control problems for diffusions . . . 3.1. The controlled diffusion model . . . . . . . . . . . . . . . 3.2. Structural assumptions . . . . . . . . . . . . . . . . . . . 3.3. Piecewise linear controlled diffusions . . . . . . . . . . . . 3.4. Existence of optimal controls . . . . . . . . . . . . . . . . 3.5. The HJB equation . . . . . . . . . . . . . . . . . . . . . . 3.6. Technical proofs . . . . . . . . . . . . . . . . . . . . . . . 4. Approximation via spatial truncations . . . . . . . . . . . . . 5. Asymptotic convergence . . . . . . . . . . . . . . . . . . . . . . 5.1. The lower bound . . . . . . . . . . . . . . . . . . . . . . . 5.2. The upper bound . . . . . . . . . . . . . . . . . . . . . . . 6. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 4 6 6 8 8 10 14 14 17 18 22 26 31 42 48 49 53 61 61 61 1. Introduction. One of the classical problems in queueing theory is to schedule the customers/jobs in a network in an optimal way. These problems are known as the scheduling problems which arise in a wide variety of applications, in particular, whenever there are different customer classes present in the network and competing for the same resources. The optimal scheduling problem has a long history in the literature. One of the appealing scheduling rules is the well-known cµ rule. This is a static priority policy in which it is assumed that each class-i customer has a marginal delay cost ci and an average service time 1/µi , and the classes are prioritized in the decreasing order of ci µi . This static priority rule has proven asymptotically optimal in many settings [4, 28, 32]. In [11], a single-server Markov modulated queueing network is considered and an averaged cµ-rule is shown asymptotically optimal for the discounted control problem. An important aspect of queueing networks is abandonment/reneging, that is, customers/jobs may choose to leave the system while being in the queue before their service. Therefore, it is important to include customer abandonment in modeling queueing systems. In [5, 6], Atar et al. considered a multi-class M/M/N +M queueing network with customer abandonment and proved that a modified priority policy, referred to as cµ/θ rule, is asymptotically optimal for the long run average cost in the fluid scale. Dai and Tezcan [13] showed the asymptotic optimality of a static priority policy on a finite ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 3 time interval for a parallel server model under the assumed conditions on the ordering of the abandonment rates and running costs. Although static priority policies are easy to implement, it may not be optimal for control problems of many multi-server queueing systems. For the same multi-class M/M/N + M queueing network, discounted cost control problems are studied in [3, 7, 22], and asymptotically optimal controls for these problems are constructed from the minimizer of a Hamilton–Jacobi–Bellman (HJB) equation associated with the controlled diffusions in the Halfin–Whitt regime. In this article, we are interested in an ergodic control problem for a multi-class M/M/N + M queueing network in the Halfin–Whitt regime. The network consists of a single pool of n statistically identical servers and a buffer of infinite capacity. There are d customer classes and arrivals of jobs/customers are d independent Poisson processes with parameters λni , i = 1, . . . , d. The service rate for class-i customers is µni , i = 1, . . . , d. Customers may renege from the queue if they have not started to receive service before their patience times. Class-i customers renege from the queue at rates γin > 0, i = 1, . . . , d. The scheduling policies are work-conserving, that is, no server stays idle if any of the queues is nonempty. We assume the system operates in the Halfin–Whitt regime, where the arrival rates and the number of servers are scaled appropriately in a manner that the traffic intensity of the system satisfies ! d X √ λni −→ ρ̂ ∈ R. n 1− nµni n→∞ i=1 In this regime, the system operations achieve both high quality (high server levels) and high efficiency (high servers’ utilization), and hence it is also referred to as the Quality-and-Efficiency-Driven (QED) regime; see, for example, [7, 16, 17, 19, 21] on the many-server regimes. We consider an ergodic cost function given by Z T 1 n r(Q̂ (s)) ds , lim sup E T →∞ T 0 where the running cost r is a nonnegative, convex function with polynomial growth and Q̂n = (Q̂n1 , . . . , Q̂nd )T is the diffusion-scaled queue length process. It is worth mentioning that in addition to the running cost above which is based on the queue-length, we can add an idle-server cost provided that it has at most polynomial growth. For such, a running cost structure the same analysis goes through. The control is the allocation of servers to different classes of customers at the service completion times. The value function is defined to be the infimum of the above cost over all admissible controls (among all work-conserving scheduling policies). In this article, we are interested in the existence and uniqueness of asymptotically optimal stable 4 A. ARAPOSTATHIS, A. BISWAS AND G. PANG stationary Markov controls for the ergodic control problem, and the asymptotic behavior of the value functions as n tends to infinity. In [7], Section 5.2, it is stated that analysis of this type of problems is important for modeling call centers. 1.1. Contributions and comparisons. The usual methodology for studying these problems is to consider the associated continuum model, which is the controlled diffusion limit in a heavy-traffic regime, and to study the ergodic control problem for the controlled diffusion. Ergodic control problems governed by controlled diffusions have been well studied in literature [1, 9] for models that fall in these two categories: (a) the running cost is near-monotone, which is defined by the requirement that its value outside a compact set exceeds the optimal average cost, thus penalizing unstable behavior (see Assumption 3.4.2 in [1] for details), or (b) the controlled diffusion is uniformly stable, that is, every stationary Markov control is stable and the collection of invariant probability measures corresponding to the stationary Markov controls is tight. However, the ergodic control problem at hand does not fall under any of these frameworks. First, the running cost we consider here is not near-monotone because the total queue length can be 0 when the total number of customers in the system are O(n). On the other hand, it is not at all clear that the controlled diffusion is uniformly stable (unless one imposes nontrivial hypotheses on the parameters), and this remains an open problem. One of our main contributions in this article is that we solve the ergodic control problem for a broad class of nondegenerate controlled diffusions, that in a certain way can be viewed as a mixture of the two categories mentioned above. As we show in Section 3, stability of the diffusion under any optimal stationary Markov control occurs due to certain interplay between the drift and the running cost. The model studied in Section 3 is far more general than the queueing problem described, and thus it is of separate interest for ergodic control. We present a comprehensive study of this broad class of ergodic control problems that includes existence of a solution to the ergodic HJB equation, its stochastic representation and verification of optimality (Theorem 3.4), uniqueness of the solution in a certain class (Theorem 3.5), and convergence of the vanishing discount method (Theorem 3.6). These results extend the well-known results for near-monotone running costs. The assumptions in these theorems are verified for the multiclass queueing model and the corresponding characterization of optimality is obtained (Corollary 3.1), which includes growth estimates for the solution of the HJB. We also introduce a new approximation technique, spatial truncation, for the controlled diffusion processes; see Section 4. It is shown that if we freeze the Markov controls to a fixed stable Markov control outside a compact set, then we can still obtain nearly optimal controls in this class of Markov ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 5 controls for large compact sets. We should keep in mind that this property is not true in general. This method can also be thought of as an approximation by a class of controlled diffusions that are uniformly stable. We remark that for a fixed control, the controlled diffusions for the queueing model can be regarded as a special case of the piecewise linear diffusions considered in [14]. It is shown in [14] that these diffusions are stable under constant Markov controls. The proof is via a suitable Lyapunov function. We conjecture that uniform stability holds for the controlled diffusions associated with the queueing model. For the same multi-class Markovian model, Gamarnik and Stolyar show that the stationary distributions of the queue lengths are tight under any work-conserving policy [15], Theorem 2. We also wish to remark here that we allow ρ̂ to be negative, assuming abandonment rates are strictly positive, while in [15], ρ̂ > 0 and abandonment rates can be zero. Another important contribution of this work is the convergence of the value functions associated with the sequence of multi-class queueing models to the value of the ergodic control problem, say ̺∗ , corresponding to the controlled diffusion model. It is not obvious that one can have asymptotic optimality from the existence of optimal stable controls for the HJB equations of controlled diffusions. This fact is relatively straightforward when the cost under consideration is discounted. In that situation, the tightness of paths on a finite time horizon is sufficient to prove asymptotic optimality [7]. But we are in a situation where any finite time behavior of the stochastic process plays no role in the cost. In particular, we need to establish the convergence of the controlled steady states. Although uniform stability of stationary distributions for this multi-class queueing model in the case where ρ̂ > 0 and abandonment rates can be zero is established in [15], it is not obvious that the stochastic model considered here has the property of uniform stability. Therefore, we use a different method to establish the asymptotic optimality. First, we show that the value functions are asymptotically bounded below by ̺∗ . To study the upper bound, we construct a sequence of Markov scheduling policies that are uniformly stable (see Lemma 5.1). The key idea used in establishing such stability results is a spatial truncation technique, under which the Markov policies follow a fixed priority policy outside a given compact set. We believe these techniques can also be used to study ergodic control problems for other many-server queueing models. The scheduling policies we consider in this paper allow preemption, that is, a customer in service can be interrupted for the server to serve a customer of a different class and her service will be resumed later. In fact, the asymptotic optimality is shown within the class of the work-conserving preemptive policies. In [7], both preemptive and nonpreemptive policies are studied, where a nonpreemptive scheduling control policy is constructed from the HJB equation associated with preemptive policies and thus is shown to be 6 A. ARAPOSTATHIS, A. BISWAS AND G. PANG asymptotically optimal. However, as far as we know, the optimal nonpreemptive scheduling problem under the ergodic cost remains open. For a similar line of work in uncontrolled settings, we refer the reader to [16, 19]. Admission control of the single class M/M/N + M model with an ergodic cost criterion in the Halfin–Whitt regime is studied in [26]. For controlled problems and for finite server models, asymptotic optimality is obtained in [12] in the conventional heavy-traffic regime. The main advantage in [12] is the uniform exponential stability of the stochastic processes, which is obtained by using properties of the Skorohod reflection map. A recent work studying ergodic control of a multi-class single-server queueing network is [25]. To summarize our main contributions in this paper: – We introduce a new class of ergodic control problems and a framework to solve them. – We establish an approximation technique by spatial truncation. – We provide, to the best of our knowledge, the first treatment of ergodic control problems at the diffusion scale for many server models. – We establish asymptotic optimality results. 1.2. Organization. In Section 1.3, we summarize the notation used in the paper. In Section 2, we introduce the multi-class many server queueing model and describe the Halfin–Whitt regime. The ergodic control problem under the heavy-traffic setting is introduced in Section 2.2, and the main results on asymptotic convergence are stated as Theorems 2.1 and 2.2. Section 3 introduces a class of controlled diffusions and associated ergodic control problems, which contains the queueing models in the diffusion scale. The key structural assumptions are in Section 3.2 and these are verified for a generic class of queueing models in Section 3.3, which are characterized by piecewise linear controlled diffusions. Section 3.4 concerns the existence of optimal controls under the general hypotheses, while Section 3.5 contains a comprehensive study of the HJB equation. Section 3.6 is devoted to the proofs of the results in Section 3.5. The spatial truncation technique is introduced and studied in Section 4. Finally, in Section 5 we prove the results of asymptotic optimality. 1.3. Notation. The standard Euclidean norm in Rd is denoted by \|·\|. The set of nonnegative real numbers is denoted by R+ , N stands for the set of natural numbers, and I denotes the indicator function. By Zd+ we denote the set of d-vectors of nonnegative integers. The closure, the boundary and the complement of a set A ⊂ Rd are denoted by A, ∂A and Ac , respectively. The open ball of radius R around 0 is denoted by BR . Given two real numbers a and b, the minimum (maximum) is denoted by a ∧ b (a ∨ b), respectively. ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 7 Define a+ := a ∨ 0 and a− := −(a ∧ 0). The integer part of a real number a is denoted by ⌊a⌋. We use the notation ei , i = 1, . . . , d, to denote the vector with ith entry equal to 1 and all other entries equal to 0. We also let e := (1, . . . , 1)T . Given any two vectors x, y ∈ Rd the inner product is denoted by x · y. By δx we denote the Dirac mass at x. For any function f : Rd → R and domain D ⊂ R we define the oscillation of f on D as follows: osc(f ) := sup{f (x) − f (y) : x, y ∈ D}. D For a nonnegative function g ∈ C(Rd ), we let O(g) denote the space of \|f (x)\| < ∞. This is a Banach space functions f ∈ C(Rd ) satisfying supx∈Rd 1+g(x) under the norm \|f (x)\| kf kg := sup . x∈Rd 1 + g(x) We also let o(g) denote the subspace of O(g) consisting of those functions f satisfying lim sup \|x\|→∞ \|f (x)\| = 0. 1 + g(x) By a slight abuse of notation, we also denote by O(g) and o(g) a generic member of these spaces. For two nonnegative functions f and g, we use the notation f ∼ g to indicate that f ∈ O(g) and g ∈ O(f ). We denote by Lploc (Rd ), p ≥ 1, the set of real-valued functions that are p d d locally p-integrable and by Wk,p loc (R ) the set of functions in Lloc (R ) whose p ith weak derivatives, i = 1, . . . , k, are in Lloc (Rd ). The set of all bounded k,α continuous functions is denoted by Cb (Rd ). By Cloc (Rd ) we denote the set of functions that are k-times continuously differentiable and whose kth derivatives are locally Hölder continuous with exponent α. We define Cbk (Rd ), k ≥ 0, as the set of functions whose ith derivatives, i = 1, . . . , k, are continuous and bounded in Rd and denote by Cck (Rd ) the subset of Cbk (Rd ) with compact support. For any path X(·), we use the notation ∆X(t) to denote the jump at time t. Given any Polish space X , we denote by P(X ) the set of probability measures on X and we endow P(X ) with the Prokhorov metric. For ν ∈ P(X ) and a Borel measurable map f : X → R, we often use the abbreviated notation Z f dν. ν(f ) := X The quadratic variation of a square integrable martingale is denoted by h·, ·i and the optional quadratic variation by [·, ·]. For presentation purposes we use the time variable as the subscript for the diffusion processes. Also κ1 , κ2 , . . . and C1 , C2 , . . . are used as generic constants whose values might vary from place to place. 8 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Fig. 1. A schematic model of the system. 2. The controlled system in the Halfin–Whitt regime. 2.1. The multi-class Markovian many-server model. Let (Ω, F, P) be a given complete probability space and all the stochastic variables introduced below are defined on it. The expectation w.r.t. P is denoted by E. We consider a multi-class Markovian many-server queueing system which consists of d customer classes and n parallel servers capable of serving all customers (see Figure 1). The system buffer is assumed to have infinite capacity. Customers of class i ∈ {1, . . . , d} arrive according to a Poisson process with rate λni > 0. Customers enter the queue of their respective classes upon arrival if not being processed. Customers of each class are served in the first-come-first-serve (FCFS) service discipline. While waiting in queue, customers can abandon the system. The service times and patience times of customers are classdependent and both are assumed to be exponentially distributed, that is, class i customers are served at rate µni and renege at rate γin . We assume that customer arrivals, service and abandonment of all classes are mutually independent. The Halfin–Whitt regime. We study this queueing model in the Halfin– Whitt regime [or the Quality-and-Efficiency-Driven (QED) regime]. Consider a sequence of such systems indexed by n, in which the arrival rates λni and the number of servers n both increase appropriately. Let rni := λni /µni be the mean offered load of classPi customers. The traffic intensity of the nth system is given by ρn = n−1 di=1 rni . In the Halfin–Whitt regime, the ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 9 parameters are assumed to satisfy the following: as n → ∞, (2.1) λni → λi > 0, µni → µi > 0, n √ λni − nλi √ n(µni − µi ) → µ̂i , → λ̂i , n λi rni → ρi := < 1, n µi This implies that √ n(1 − ρn ) → ρ̂ := d X γin → γi > 0, ρi = 1. i=1 d X ρi µ̂i − λ̂i i=1 µi ∈ R. The above scaling is common in multi-class multi-server models [7, 22]. Note that we do not make any assumption on the sign of ρ̂. State descriptors. Let Xin = {Xin (t) : t ≥ 0} be the total number of class i customers in the system, Qni = {Qni (t) : t ≥ 0} the number of class i customers in the queue and Zin = {Zin (t) : t ≥ 0} the number of class i customers in service. The following basic relationships hold for these processes: for each t ≥ 0 and i = 1, . . . , d, (2.2) Xin (t) = Qni (t) + Zin (t), Qni (t) ≥ 0, Zin (t) ≥ 0 and e · Z n (t) ≤ n. We can describe these processes using a collection {Ani , Sin , Rin , i = 1, . . . , d} of independent rate-1 Poisson processes. Define Ãni (t) := Ani (λni t), Z t S̃in (t) := Sin µni Zin (s) ds , 0 Z t n n n n R̃i (t) := Ri γi Qi (s) ds . 0 Then the dynamics take the form (2.3) Xin (t) = Xin (0) + Ãni (t) − S̃in (t) − R̃in (t), t ≥ 0, i = 1, . . . , d. Scheduling control. Following [7, 22], we only consider work-conserving policies that are nonanticipative and allow preemption. When a server becomes free and there are no customers waiting in any queue, the server stays idle, but if there are customers of multiple classes waiting in the queue, the 10 A. ARAPOSTATHIS, A. BISWAS AND G. PANG server has to make a decision on the customer class to serve. Service preemption is allowed, that is, service of a customer class can be interrupted at any time to serve some other class of customers and the original service is resumed at a later time. A scheduling control policy determines the processes Z n , which must satisfy the constraints in (2.2) and the work-conserving constraint, that is, e · Z n (t) = (e · X n (t)) ∧ n, t ≥ 0. Define the action set An (x) as An (x) := {a ∈ Zd+ : a ≤ x and e · a = (e · x) ∧ n}. Thus, we can write Z n (t) ∈ An (X n (t)) for each t ≥ 0. We also assume that all controls are nonanticipative. Define the σ-fields Ftn := σ{X n (0), Ãni (t), S̃in (t), R̃in (t) : i = 1, . . . , d, 0 ≤ s ≤ t} ∨ N and Gtn := σ{δÃni (t, r), δS̃in (t, r), δR̃in (t, r) : i = 1, . . . , d, r ≥ 0}, where δÃni (t, r) := Ãni (t + r) − Ãni (t), Z t n n n n n δS̃i (t, r) := Si µi Zi (s) ds + µi r − S̃in (t), 0 δR̃in (t, r) := Rin γin Z 0 t Qni (s) ds + γin r − R̃in (t), and N is the collection of all P-null sets. The filtration {Ftn , t ≥ 0} represents the information available up to time t while Gtn contains the information about future increments of the processes. We say that a working-conserving control policy is admissible if: (i) Z n (t) is adapted to Ftn , (ii) Ftn is independent of Gtn at each time t ≥ 0, (iii) for each i = 1, . . . , d, and t ≥ 0, the process δS̃in (t, ·) agrees in law with Sin (µni ·), and the process δR̃in (t, ·) agrees in law with Rin (γin ·). We denote the set of all admissible control policies (Z n , F n , G n ) by Un . 2.2. The ergodic control problem in the Halfin–Whitt regime. Define the diffusion-scaled processes X̂ n = (X̂1n , . . . , X̂dn )T , Q̂n = (Q̂n1 , . . . , Q̂nd )T and Ẑ n = (Ẑ1n , . . . , Ẑdn )T , ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 11 by (2.4) 1 X̂in (t) := √ (Xin (t) − ρi nt), n 1 Q̂ni (t) := √ Qni (t), n 1 Ẑin (t) := √ (Zin (t) − ρi nt) n for t ≥ 0. By (2.3), we can express X̂in as Z t Z t n n n n n n X̂i (t) = X̂i (0) + ℓi t − µi Ẑi (s) ds − γi Q̂ni (s) ds 0 0 (2.5) n n n + M̂A,i (t) − M̂S,i (t) − M̂R,i (t), where ℓn = (ℓn1 , . . . , ℓnd )T is defined as 1 ℓni := √ (λni − µni ρi n), n and (2.6) 1 n M̂A,i (t) := √ (Ani (λni t) − λni t), n Z t Z t 1 n n n n n n M̂S,i (t) := √ Si µi Zi (s) ds − µi Zi (s) ds , n 0 0 Z t Z t 1 n n n n n n M̂R,i (t) := √ Ri γi Qi (s) ds − γi Qi (s) ds n 0 0 are square integrable martingales w.r.t. the filtration {Ftn }. Note that √ 1 (λ̂i − ρi µ̂i ) . ℓni = √ (λni − λi n) − ρi n(µni − µi ) −→ ℓi := n→∞ n µi Define S := {u ∈ Rd+ : e · u = 1}. For Z n ∈ Un we define, for t ≥ 0 and for adapted Û n (t) ∈ S, (2.7) Q̂n (t) := (e · X̂ n (t))+ Û n (t), Ẑ n (t) := X̂ n (t) − (e · X̂ n (t))+ Û n (t). If Q̂n (t) = 0, we define Û n (t) := ed = (0, . . . , 0, 1)T . Thus, Ûin represents the fraction of class-i customers in the queue when the total queue size is positive. As we show later, it is convenient to view Û n (t) as the control. Note that the controls are nonanticipative and preemption is allowed. 12 A. ARAPOSTATHIS, A. BISWAS AND G. PANG 2.2.1. The cost minimization problem. We next introduce the running cost function for the control problem. Let r : Rd+ → R+ be a given function satisfying (2.8) c1 \|x\|m ≤ r(x) ≤ c2 (1 + \|x\|m ) for some m ≥ 1, and some positive constants ci , i = 1, 2. We also assume that r is locally Lipschitz. This assumption includes linear and convex running cost functions. For example, if we let hi be the holding cost rate for class i customers, then some of the typical running cost functions are the following: r(x) = d X hi xm i , m ≥ 1. i=1 These running cost functions evidently satisfy the condition in (2.8). Given the initial state X n (0) and a work-conserving scheduling policy n Z ∈ Un , we define the diffusion-scaled cost function as Z T 1 n n n (2.9) r(Q̂ (s)) ds , J(X̂ (0), Ẑ ) := lim sup E T →∞ T 0 where the running cost function r satisfies (2.8). Note that the running cost is defined using the scaled version of Z n . Then the associated cost minimization problem becomes (2.10) V̂ n (X̂ n (0)) := ninf n J(X̂ n (0), Ẑ n ). Z ∈U We refer to V̂ n (X̂ n (0)) as the diffusion-scaled value function given the initial state X̂ n (0) in the nth system. From (2.7), it is easy to see that by redefining r as r(x, u) = r((e · x)+ u) we can rewrite the control problem as ˜ X̂ n (0), Û n ), V̂ n (X̂ n (0)) = inf J( where (2.11) ˜ X̂ n (0), Û n ) := lim sup 1 E J( T →∞ T Z T 0 r(X̂ n (s), Û n (s)) ds , and the infimum is taken over all admissible pairs (X̂ n , Û n ) satisfying (2.7). For simplicity, we assume that the initial condition X̂ n (0) is deterministic and X̂ n (0) → x as n → ∞ for some x ∈ Rd . ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 13 2.2.2. The limiting controlled diffusion process. As in [7, 22], one formally deduces that, provided X̂ n (0) → x, there exists a limit X for X̂ n on every finite time interval, and the limit process X is a d-dimensional diffusion process with independent components, that is, (2.12) dXt = b(Xt , Ut ) dt + Σ dWt, with initial condition X0 = x. In (2.12), the drift b(x, u) : Rd × S → Rd takes the form (2.13) with b(x, u) = ℓ − R(x − (e · x)+ u) − (e · x)+ Γu, ℓ := (ℓ1 , . . . , ℓd )T , R := diag(µ1 , . . . , µd ), Γ := diag(γ1 , . . . , γd ). The control Ut lives in S and is nonanticipative, W (t) is a d-dimensional standard Wiener process independent of the initial condition X0 = x, and the covariance matrix is given by ΣΣT = diag(2λ1 , . . . , 2λd ). A formal derivation of the drift in (2.13) can be obtained from (2.5) and (2.7). A detailed description of equation (2.12) and related results are given in Section 3. Let U be the set of all admissible controls for the diffusion model (for a definition see Section 3). 2.2.3. The ergodic control problem in the diffusion scale. Define r̃ : Rd+ × Rd+ → R+ by r̃(x, u) := r((e · x)+ u), where r is the same function as in (2.9). In analogy with (2.11) we define the ergodic cost associated with the controlled diffusion process X and the running cost function r̃(x, u) as Z T 1 U J(x, U ) := lim sup Ex r̃(Xt , Ut ) dt , U ∈ U. T →∞ T 0 We consider the ergodic control problem (2.14) ̺∗ (x) = inf J(x, U ). U ∈U We call ̺∗ (x) the optimal value at the initial state x for the controlled diffusion process X. It is shown later that ̺∗ (x) is independent of x. A detailed treatment and related results corresponding to the ergodic control problem are given in Section 3. We next state the main results of this section, the proof of which can be found in Section 5. 14 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Theorem 2.1. Let X̂ n (0) → x ∈ Rd as n → ∞. Also assume that (2.1) and (2.8) hold. Then lim inf V̂ n (X̂ n (0)) ≥ ̺∗ (x), n→∞ where ̺∗ (x) is given by (2.14). Theorem 2.2. Suppose the assumptions of Theorem 2.1 hold. In addition, assume that r in (2.9) is convex. Then lim sup V̂ n (X̂ n (0)) ≤ ̺∗ (x). n→∞ Thus, we conclude that for any convex running cost function r, Theorems 2.1 and 2.2 establish the asymptotic convergence of the ergodic control problem for the queueing model. 3. A broad class of ergodic control problems for diffusions. 3.1. The controlled diffusion model. The dynamics are modeled by a controlled diffusion process X = {Xt , t ≥ 0} taking values in the d-dimensional Euclidean space Rd , and governed by the Itô stochastic differential equation (3.1) dXt = b(Xt , Ut ) dt + σ(Xt ) dWt . All random processes in (3.1) live in a complete probability space (Ω, F, P). The process W is a d-dimensional standard Wiener process independent of the initial condition X0 . The control process U takes values in a compact, metrizable set U, and Ut (ω) is jointly measurable in (t, ω) ∈ [0, ∞) × Ω. Moreover, it is nonanticipative: for s < t, Wt − Ws is independent of Fs := the completion of σ{X0 , Ur , Wr , r ≤ s} relative to (F, P). Such a process U is called an admissible control, and we let U denote the set of all admissible controls. We impose the following standard assumptions on the drift b and the diffusion matrix σ to guarantee existence and uniqueness of solutions to equation (3.1). (A1) Local Lipschitz continuity: The functions b = [b1 , . . . , bd ]T : Rd × U → Rd and σ = [σ ij ] : Rd → Rd×d are locally Lipschitz in x with a Lipschitz constant CR > 0 depending on R > 0. In other words, for all x, y ∈ BR and u ∈ U, \|b(x, u) − b(y, u)\| + kσ(x) − σ(y)k ≤ CR \|x − y\|. We also assume that b is continuous in (x, u). ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 15 (A2) Affine growth condition: b and σ satisfy a global growth condition of the form \|b(x, u)\|2 + kσ(x)k2 ≤ C1 (1 + \|x\|2 ) ∀(x, u) ∈ Rd × U, where kσk2 := trace(σσ T ). (A3) Local nondegeneracy: For each R > 0, it holds that d X i,j=1 −1 2 \|ξ\| aij (x)ξi ξj ≥ CR ∀x ∈ BR , for all ξ = (ξ1 , . . . , ξd )T ∈ Rd , where a := σσ T . In integral form, (3.1) is written as Z t Z t (3.2) σ(Xs ) dWs . b(Xs , Us ) ds + Xt = X0 + 0 0 The third term on the right-hand side of (3.2) is an Itô stochastic integral. We say that a process X = {Xt (ω)} is a solution of (3.1), if it is Ft -adapted, continuous in t, defined for all ω ∈ Ω and t ∈ [0, ∞), and satisfies (3.2) for all t ∈ [0, ∞) a.s. It is well known that under (A1)–(A3), for any admissible control there exists a unique solution of (3.1) [1], Theorem 2.2.4. We define the family of operators Lu : C 2 (Rd ) → C(Rd ), where u ∈ U plays the role of a parameter, by (3.3) Lu f (x) := 12 aij (x)∂ij f (x) + bi (x, u)∂i f (x), u ∈ U. We refer to Lu as the controlled extended generator of the diffusion. In ∂ (3.3) and elsewhere in this paper, we have adopted the notation ∂i := ∂x i 2 and ∂ij := ∂x∂i ∂xj . We also use the standard summation rule that repeated subscripts and superscripts are summed from 1 through d. In other words, the right-hand side of (3.3) stands for d d X 1 X ij ∂f ∂2f bi (x, u) a (x) (x) + (x). 2 ∂xi ∂xj ∂xi i=1 i,j=1 Of fundamental importance in the study of functionals of X is Itô’s formula. For f ∈ C 2 (Rd ) and with Lu as defined in (3.3), it holds that Z t LUs f (Xs ) ds + Mt , (3.4) a.s., f (Xt ) = f (X0 ) + 0 where Mt := Z 0 t h∇f (Xs ), σ(Xs ) dWs i 16 A. ARAPOSTATHIS, A. BISWAS AND G. PANG is a local martingale. Krylov’s extension of Itô’s formula [27], page 122, d extends (3.4) to functions f in the local Sobolev space W2,p loc (R ), p ≥ d. Recall that a control is called Markov if Ut = v(t, Xt ) for a measurable map v : R+ × Rd → U, and it is called stationary Markov if v does not depend on t, that is, v : Rd → U. Correspondingly, (3.1) is said to have a strong solution if given a Wiener process (Wt , Ft ) on a complete probability space (Ω, F, P), there exists a process X on (Ω, F, P), with X0 = x0 ∈ Rd , which is continuous, Ft -adapted, and satisfies (3.2) for all t a.s. A strong solution is called unique, if any two such solutions X and X ′ agree P-a.s., when viewed as elements of C([0, ∞), Rd ). It is well known that under assumptions (A1)–(A3), for any Markov control v, (3.1) has a unique strong solution [20]. Let USM denote the set of stationary Markov controls. Under v ∈ USM , the process X is strong Markov, and we denote its transition function by Pvt (x, ·). It also follows from the work of [8, 31] that under v ∈ USM , the transition probabilities of X have densities which are locally Hölder continuous. Thus, Lv defined by Lv f (x) := 12 aij (x)∂ij f (x) + bi (x, v(x))∂i f (x), v ∈ USM , for f ∈ C 2 (Rd ), is the generator of a strongly-continuous semi-group on Cb (Rd ), which is strong Feller. We let Pvx denote the probability measure and Evx the expectation operator on the canonical space of the process under the control v ∈ USM , conditioned on the process X starting from x ∈ Rd at t = 0. We need the following definition. Definition 3.1. A function h : Rd × U → R is called inf-compact on a set A ⊂ Rd if the set Ā ∩ {x : minu∈U h(x, u) ≤ β} is compact (or empty) in Rd for all β ∈ R. When this property holds for A ≡ Rd , then we simply say that h is inf-compact. Recall that control v ∈ USM is called stable if the associated diffusion is positive recurrent. We denote the set of such controls by USSM , and let µv denote the unique invariant probability measure on Rd for the diffusion under the control v ∈ USSM . We also let M := {µv : v ∈ USSM }. Recall that v ∈ USSM if and only if there exists an inf-compact function V ∈ C 2 (Rd ), a bounded domain D ⊂ Rd , and a constant ε > 0 satisfying Lv V(x) ≤ −ε ∀x ∈ D c . We denote by τ (A) the first exit time of a process {Xt , t ∈ R+ } from a set A ⊂ Rd , defined by τ (A) := inf{t > 0 : Xt ∈ / A}. ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 17 The open ball of radius R in Rd , centered at the origin, is denoted by BR , c ). and we let τR := τ (BR ), and τ̆R := τ (BR We assume that the running cost function r(x, u) is nonnegative, continuous and locally Lipschitz in its first argument uniformly in u ∈ U. Without loss of generality, we let κR be a Lipschitz constant of r(·, u) over BR . In summary, we assume that (A4) r : Rd ×U → R+ is continuous and satisfies, for some constant CR > 0 \|r(x, u) − r(y, u)\| ≤ CR \|x − y\| ∀x, y ∈ BR , ∀u ∈ U, and all R > 0. In general, U may not be a convex set. It is therefore often useful to enlarge the control set to P(U). For any v(du) ∈ P(U) we can redefine the drift and the running cost as Z Z (3.5) b̄(x, v) := b(x, u)v(du) and r̄(x, v) := r(x, u)v(du). U U It is easy to see that the drift and running cost defined in (3.5) satisfy all the aforementioned conditions (A1)–(A4). In what follows, we assume that all the controls take values in P(U). These controls are generally referred to as relaxed controls. We endow the set of relaxed stationary Markov controls with the following topology: vn → v in USM if and only if Z Z Z Z f (x) g(x, u)v(du\|x) dx f (x) g(x, u)vn (du\|x) dx −→ Rd n→∞ Rd U L1 (Rd ) L2 (Rd ) U (Rd for all f ∈ ∩ and g ∈ Cb × U). Then USM is a compact metric space under this topology [1], Section 2.4. We refer to this topology as the topology of Markov controls. A control is said to be precise if it takes value in U. It is easy to see that any precise control Ut can also be understood as a relaxed control by Ut (du) = δUt . Abusing the notation, we denote the drift and running cost by b and r, respectively, and the action of a relaxed control on them is understood as in (3.5). 3.2. Structural assumptions. Assumptions 3.1 and 3.2, described below, are in effect throughout the analysis, unless otherwise stated. Assumption 3.1. For some open set K ⊂ Rd , the following hold: (i) The running cost r is inf-compact on K. (ii) There exist inf-compact functions V ∈ C 2 (Rd ) and h ∈ C(Rd × U), such that (3.6) Lu V(x) ≤ 1 − h(x, u) ∀(x, u) ∈ Kc × U, Lu V(x) ≤ 1 + r(x, u) ∀(x, u) ∈ K × U. 18 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Without loss of generality, we assume that V and h are nonnegative. Remark 3.1. In the statement of Assumption 3.1, we refrain from using any constants in the interest of notational economy. There is no loss of generality in doing so, since the functions V and h can always be scaled to eliminate unnecessary constants. The next assumption is not a structural one, but rather the necessary requirement that the value of the ergodic control problem is finite. Otherwise, the problem is vacuous. For U ∈ U, define Z T 1 U ̺U (x) := lim sup Ex (3.7) r(Xs , Us ) ds . T →∞ T 0 Assumption 3.2. x ∈ Rd . There exists U ∈ U such that ̺U (x) < ∞ for some Assumption 3.2 alone does not imply that ̺v < ∞ for some v ∈ USSM . However, when combined with Assumption 3.1, this is the case as the following lemma asserts. Lemma 3.1. Let Assumptions 3.1 and 3.2 hold. Then there exists u0 ∈ USSM such that ̺u0 < ∞. Moreover, there exists a nonnegative inf-compact function V0 ∈ C 2 (Rd ), and a positive constant η such that (3.8) Lu0 V0 (x) ≤ η − r(x, u0 (x)) ∀x ∈ Rd . Conversely, if (3.8) holds, then Assumption 3.2 holds. Proof. The first part of the result follows from Theorem 3.1(e) and (3.23) whereas the converse part follows from Lemma 3.2. These proofs are stated later in the paper. Remark 3.2. There is no loss of generality in using only the constant η in Assumption 3.2, since V0 can always be scaled to achieve this. We also observe that for K = Rd the problem reduces to an ergodic control problem with near-monotone cost, and for K = ∅ we obtain an ergodic control problem under a uniformly stable controlled diffusion. 3.3. Piecewise linear controlled diffusions. The controlled diffusion process in (2.12) belongs to a large class of controlled diffusion processes, called piecewise linear controlled diffusions [14]. We describe this class of controlled diffusions and show that it satisfies the assumptions in Section 3.2. ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 19 Definition 3.2. A square matrix R is said to be an M -matrix if it can be written as R = sI − N for some s > 0 and nonnegative matrix N with property that ρ(N ) ≤ s, where ρ(N ) denotes the spectral radius of N . Let Γ = [γ ij ] be a given matrix whose diagonal elements are positive, = 0 for i = 1, . . . , d − 1, and the remaining elements are in R. (Note that for the queueing model, Γ is a positive diagonal matrix. Our results below hold for the more general Γ.) Let ℓ ∈ Rd and R be a nonsingular M -matrix. Define γ id (3.9) b(x, u) := ℓ − R(x − (e · x)+ u) − (e · x)+ Γu, with u ∈ S := {u ∈ Rd+ : e · u = 1}. Assume that eT R ≥ 0T . We consider the following controlled diffusion in Rd : (3.10) dXt = b(Xt , Ut ) dt + Σ dWt, where Σ is a constant matrix such that ΣΣT is invertible. It is easy to see that (3.10) satisfies conditions (A1)–(A3). Analysis of these types of diffusion approximations is an established tradition in queueing systems. It is often easy to deal with the limiting object and it also helps to obtain information on the behavior of the actual queueing model. We next introduce the running cost function. Let r : Rd × S → [0, ∞) be locally Lipschitz with polynomial growth and (3.11) c1 [(e · x)+ ]m ≤ r(x, u) ≤ c2 (1 + [(e · x)+ ]m ), for some m ≥ 1 and positive constants c1 and c2 that do not depend on u. Some typical examples of such running costs are + m r(x, u) = [(e · x) ] d X i=1 hi um i with m ≥ 1, for some positive vector (h1 , . . . , hd )T . Remark 3.3. The controlled dynamics in (3.9) and running cost in (3.11) are clearly more general than the model described in Section 2.2. In (3.10), X denotes the diffusion approximation for the number customers in the system in the Halfin–Whitt regime and its ith component X i denotes the diffusion approximation of the number of class i customers. Therefore, (e · X)+ denotes the total number of customers in the queue. For R and Γ diagonal as in (2.13), the diagonal entries of R and Γ denote the service 20 A. ARAPOSTATHIS, A. BISWAS AND G. PANG and abandonment rates, respectively, of the customer classes. The ith coordinate of U denotes the fraction of class-i customers waiting in the queue. Therefore, the vector-valued process Xt − (e · Xt )+ Ut denotes the diffusion approximation of the numbers of customers in service from different customer classes. Proposition 3.1. Let b and r be given by (3.9) and (3.11), respectively. Then (3.10) satisfies Assumptions 3.1 and 3.2, with h(x) = c0 \|x\|m and (3.12) K := {x : δ\|x\| < (e · x)+ } for appropriate positive constants c0 and δ. Proof. We recall that if R is a nonsingular M -matrix, then there exists a positive definite matrix Q such that QR + RT Q is strictly positive definite [14]. Therefore, for some positive constant κ0 it holds that 2 κ0 \|y\|2 ≤ y T [QR + RT Q]y ≤ κ−1 0 \|y\| ∀y ∈ Rd . The set K in (3.12), where δ > 0 is chosen later, is an open convex cone, and the running cost function r is inf-compact on K. Let V be a nonnegative function in C 2 (Rd ) such that V(x) = [xT Qx]m/2 for \|x\| ≥ 1, where the constant m is as in (3.11). Let \|x\| ≥ 1 and u ∈ S. Then m[xT Qx]m/2−1 T x [QR + RT Q]x 2 + m[xT Qx]m/2−1 Qx · (R − Γ)(e · x)+ u m/2−1 κ0 T 2 + ≤ ℓ · ∇V(x) − m[x Qx] \|x\| − C\|x\|(e · x) 2 ∇V(x) · b(x, u) = ℓ · ∇V(x) − κ0 , then on Kc ∩ {\|x\| ≥ 1} for some positive constant C. If we choose δ = 4C we have the estimate mκ0 T (3.13) [x Qx]m/2−1 \|x\|2 . ∇V(x) · b(x, u) ≤ ℓ · ∇V(x) − 4 Note that ℓ · V is globally bounded for m = 1. For any m ∈ (1, ∞), it follows by (3.13) that mκ0 T ∇V(x) · b(x, u) ≤ m(ℓT Qx)[xT Qx]m/2−1 − [x Qx]m/2−1 \|x\|2 4 (3.14) m/2 T mκ0 (λ(Q))m/2 m m\|ℓ Q\|(λ(Q)) \|x\|m−1 − \|x\| ≤ λ(Q) 4λ(Q) for x ∈ Kc ∩ {\|x\| ≥ 1}, where λ(Q) and λ(Q) are the smallest and largest eigenvalues of Q, respectively. We use Young’s inequality \|a\|m m − 1 m/(m−1) + \|b\| , a, b ≥ 0, \|ab\| ≤ m m ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 21 in (3.14) to obtain the bound (3.15) ∇V(x) · b(x, u) ≤ κ1 − mκ0 (λ(Q))m/2 \|x\|m 8λ(Q) for some constant κ1 > 0. A similar calculation shows for some constant κ2 > 0 it holds that (3.16) ∇V(x) · b(x, u) ≤ κ2 (1 + [(e · x)+ ]m ) ∀x ∈ K ∩ {\|x\| ≥ 1}. Also note that we can select κ3 > 0 large enough such that (3.17) 1 mκ0 \|trace(ΣΣT ∇2 V(x))\| ≤ κ3 + (λ(Q))m/2 \|x\|m . 2 16λ(Q) Hence, by (3.13)–(3.17) there exists κ4 > 0 such that mκ0 (λ(Q))m/2 \|x\|m IKc (x) + κ2 [(e · x)+ ]m IK (x) (3.18) Lu V(x) ≤ κ4 − 16λ(Q) for all x ∈ Rd . It is evident that we can scale V, by multiplying it with a constant, so that (3.18) takes the form (3.19) Lu V(x) ≤ 1 − c0 \|x\|m IKc (x) + c1 [(e · x)+ ]m IK (x) ∀x ∈ Rd . By (3.11), the running cost r is inf-compact on K. It then follows from (3.11) and (3.19) that (3.6) is satisfied with h(x) := c0 \|x\|m . We next show that (3.10) satisfies Assumption 3.2. Let u0 (·) ≡ ed = (0, . . . , 0, 1)T . Then we can write (3.10) as (3.20) dXt = (ℓ − R(Xt − (e · Xt )+ u0 ) − (e · x)+ Γu0 ) dt + Σ dWt. It is shown in [14] that the solution Xt in (3.20) is positive recurrent and, therefore, u0 is a stable Markov control. This is done by finding a suitable Lyapunov function. In particular, in [14], Theorem 3, it is shown that there exists a positive definite matrix Q̃ such that if we define (3.21) ψ(x) := (e · x)2 + κ̃[x − ed φ(e · x)]T Q̃[x − ed φ(e · x)], for some suitably chosen constant κ̃ and   y, φ(y) = − 21 δ̃,  smooth, a function φ ∈ C 2 (R), given by if y ≥ 0, if y ≤ −δ̃, if − δ̃ < y < 0, where δ̃ > 0 is a suitable constant and 0 ≤ φ′ (y) ≤ 1, then it holds that (3.22) Lu0 ψ(x) ≤ −κ̃1 \|x\|2 , 22 A. ARAPOSTATHIS, A. BISWAS AND G. PANG for \|x\| large enough and positive constant κ̃1 . We define V0 := eaψ where a is to be determined later. Note that \|∇ψ(x)\| ≤ κ̃2 (1 + \|x\|) for some constant κ̃2 > 0. Hence, a straightforward calculation shows that if we choose a small enough, then for some constant κ̃3 > 0 it holds that Lu0 V0 (x) ≤ (−κ̃1 a\|x\|2 + a2 kΣk2 κ̃2 (1 + \|x\|)2 )V0 (x) ≤ −κ̃3 \|x\|2 V0 (x), for all \|x\| large enough. Since V0 (x) > [(e · x)+ ]m , m ≥ 1, for all large enough \|x\| we see that V0 satisfies (3.8) with control u0 . Hence, Assumption 3.2 holds by Lemma 3.1. 3.4. Existence of optimal controls. Definition 3.3. Recall the definition of ̺U in (3.7). For β > 0, we define Uβ := {U ∈ U : ̺U (x) ≤ β for some x ∈ Rd }. We also let UβSM := Uβ ∩ USM , and ̺ˆ∗ := inf{β > 0 : Uβ 6= ∅}, ̺∗ := inf{β > 0 : UβSM 6= ∅}, ̺˜∗ := inf{π(r) : π ∈ G}, where Z d G := π ∈ P(R × U) : Rd ×U u L f (x)π(dx, du) = 0 ∀f ∈ Cc∞ (Rd ) , and Lu f (x) is given by (3.3). It is well known that G is the set of ergodic occupation measures of the controlled process in (3.1), and that G is a closed and convex subset of P(Rd × U) [1], Lemmas 3.2.2 and 3.2.3. We use the notation πv when we want to indicate the ergodic occupation measure associated with the control v ∈ USSM . In other words, πv (dx, du) := µv (dx)v(du\|x). Lemma 3.2. If (3.8) holds for some V0 and u0 , then we have πu0 (r) ≤ η. Therefore, ̺ˆ∗ < ∞. Proof. Let (Xt , u0 (Xt )) be the solution of (3.1). Recall that τR is the first exit time from BR for R > 0. Then by Itô’s formula Z T ∧τR u0 u0 Ex [V0 (XT ∧τR )] − V0 (x) ≤ ηT − Ex r(Xs , u0 (Xs )) ds . 0 ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 23 Therefore, letting R → ∞ and using Fatou’s lemma, we obtain the bound Z T u0 Ex r(Xs , u0 (Xs )) ds ≤ ηT + V0 (x) − min V0 , Rd 0 and thus 1 lim sup Eux0 T →∞ T Z T 0 r(Xs , u0 (Xs )) ds ≤ η. In the analysis, we use a function h̃ ∈ C(Rd × U) which, roughly speaking, is of the same order as r in K × U and lies between r and a multiple of r + h on Kc × U, with K as in Assumption 3.1. The existence of such a function is guaranteed by Assumption 3.1 as the following lemma shows. Lemma 3.3. Define H := (K × U) ∪ {(x, u) ∈ Rd × U : r(x, u) > h(x, u)}, where K is the open set in Assumption 3.1. Then there exists an inf-compact function h̃ ∈ C(Rd × U) which is locally Lipschitz in its first argument uniformly w.r.t. its second argument, and satisfies (3.23) r(x, u) ≤ h̃(x, u) ≤ k0 (1 + h(x, u)IHc (x, u) + r(x, u)IH (x, u)) 2 for all (x, u) ∈ Rd × U, and for some positive constant k0 ≥ 2. Moreover, (3.24) Lu V(x) ≤ 1 − h(x, u)IHc (x, u) + r(x, u)IH (x, u) for all (x, u) ∈ Rd × U, where V is the function in Assumption 3.1. Proof. If f (x, u) denotes the right-hand side of (3.23), with k0 = 4, then f (x, u) − r(x, u) > h(x, u)IHc (x, u) + r(x, u)IH (x, u) ≥ h(x, u)IKc (x) + r(x, u)IK (x), since r(x, u) > h(x, u) on H \ (K × U). Therefore, by Assumption 3.1, the set {(x, u) : f (x, u) − r(x, u) ≤ n} is bounded in Rd × U for every n ∈ N. Hence, there exists an increasing sequence of open balls Dn , n = 1, 2, . . . , centered at 0 in Rd such that f (x, u) − r(x, u) ≥ n for all (x, u) ∈ Dnc × U. Let g : Rd → R be any nonnegative smooth function such that n − 1 ≤ g(x) ≤ n for x ∈ Dn+1 \ Dn , n = 1, 2, . . . , and g(x) = 0 on D1 . Clearly, h̃ := r + g is continuous, inf-compact, locally Lipschitz in its first argument, and satisfies (3.23). That (3.24) holds is clear from (3.6) and the fact that H ⊇ K × U. 24 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Remark 3.4. It is clear from the proof of Lemma 3.3 that we could fix the value of the constant k0 in (3.23), say k0 = 4. However, we keep the variable k0 because this provides some flexibility in the choice of h̃, and also in order to be able to trace it along the different calculations. Remark 3.5. Note that if h ≥ r and r is inf-compact, then H = K × U and h̃ := r satisfies (3.23). Note also, that in view of (3.11) and Proposition 3.1, for the multi-class queueing model we have r(x, u) ≤ c2 (1 + [(e · x)+ ]m ) c2 dm−1 (1 + (1 ∧ c0 )\|x\|m ) 1 ∧ c0 c2 dm−1 1 + m m c ≤ 1 + c0 \|x\| IK (x) + m [(e · x) ] IK (x) 1 ∧ c0 δ m−1 1 c2 d r(x, u)IK (x) 1 + h(x)IKc (x) + ≤ 1 ∧ c0 c1 δm ≤ ≤ ≤ c2 dm−1 (1 + h(x)IKc (x) + r(x, u)IK (x)) 1 ∧ c0 ∧ c1 δm c2 dm−1 (1 + h(x)IHc (x, u) + r(x, u)IH (x, u)). 1 ∧ c0 ∧ c1 δm Therefore, h̃(x, u) := c2 + c2 dm−1 \|x\|m satisfies (3.23). Remark 3.6. We often use the fact that if g ∈ C(Rd × U) is bounded below, then the map P(Rd × U) ∋ ν 7→ ν(g) is lower semi-continuous. This easily follows from two facts: (a) g can be expressed as an increasing limit of bounded continuous functions, and (b) if g is bounded and continuous, then π 7→ π(g) is continuous. Theorem 3.1. Let β ∈ (ˆ ̺∗ , ∞). Then: (a) For all U ∈ Uβ and x ∈ Rd such that ̺U (x) ≤ β, then Z T 1 U lim sup Ex (3.25) h̃(Xs , Us ) ds ≤ k0 (1 + β). t→∞ T 0 (b) ̺ˆ∗ = ̺∗ = ̺˜∗ . (c) For any β ∈ (̺∗ , ∞), we have UβSM ⊂ USSM . (d) The set of invariant probability measures Mβ corresponding to controls in UβSM satisfies Z h̃(x, v(x))µv (dx) ≤ k0 (1 + β) ∀µv ∈ Mβ . Rd ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 25 In particular, UβSM is tight in P(Rd ). (e) There exists (Ṽ , ̺˜) ∈ C 2 (Rd ) × R+ , with Ṽ inf-compact, such that (3.26) min[Lu Ṽ (x) + h̃(x, u)] = ̺˜. u∈U Proof. Using Itô’s formula, it follows by (3.24) that (3.27) 1 U (E [V(XT ∧τR )] − V(x)) T x Z T ∧τR 1 U h(Xs , Us )IHc (Xs , Us ) ds ≤ 1 − Ex T 0 Z T ∧τR 1 U + Ex r(Xs , Us )IH (Xs , Us ) ds . T 0 Since V is inf-compact, (3.27) together with (3.23) implies that Z T 1 U 2 lim sup E h̃(Xs , Us ) ds k0 T →∞ T x 0 (3.28) Z T 1 U ≤ 2 + 2 lim sup Ex r(Xs , Us ) ds . T →∞ T 0 Part (a) then follows from (3.28). Now fix U ∈ Uβ and x ∈ Rd such that ̺U (x) ≤ β. The inequality in (3.25) U : t ≥ 1}, defined by implies that the set of mean empirical measures {ζx,t Z t 1 U U IA×B (Xs , Us ) ds ζx,t(A × B) := Ex t 0 for any Borel sets A ⊂ Rd and B ⊂ U, is tight. It is the case that any limit point of the mean empirical measures in P(Rd × U) is an ergodic occupation measure [1], Lemma 3.4.6. Then in view of Remark 3.6 we obtain (3.29) U π(r) ≤ lim sup ζx,t (r) ≤ β t→∞ for some ergodic occupation measure π. Therefore, ̺˜∗ ≤ ̺ˆ∗ . Disintegrating the measure π as π(dx, du) = v(du\|x)µv (dx), we obtain the associated control v ∈ USSM . From ergodic theory [33], we also know that Z T 1 lim sup Evx r(Xs , v(Xs )) ds = πv (r) for almost every x. T →∞ T 0 It follows that ̺∗ ≤ ̺˜∗ , and since it is clear that ̺ˆ∗ ≤ ̺∗ , equality must hold among the three quantities. If v ∈ UβSM , then (3.28) implies that (3.29) holds with U ≡ v and π ≡ πv . Therefore, parts (c) and (d) follow. 26 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Existence of (Ṽ , ̺˜), satisfying (3.26), follows from Assumption 3.2 and [1], Theorem 3.6.6. The inf-compactness of Ṽ follows from the stochastic representation of Ṽ in [1], Lemma 3.6.9. This proves (e). Existence of a stationary Markov control that is optimal is asserted by the following theorem. Theorem 3.2. Let G denote the set of ergodic occupation measures corresponding to controls in USSM , and Gβ those corresponding to controls in UβSM , for β > ̺∗ . Then: (a) The set Gβ is compact in P(Rd ) for any β > ̺∗ . (b) There exists v ∈ USM such that ̺v = ̺∗ . Proof. By Theorem 3.1(d), the set Gβ is tight for any β > ̺∗ . Let {πn } ⊂ Gβ , for some β > ̺∗ , be any convergent sequence in P(Rd ) such that πn (r) → ̺∗ as n → ∞ and denote its limit by π∗ . Since G is closed, π∗ ∈ G, and since the map π → π(r) is lower semi-continuous, it follows that π∗ (r) ≤ ̺∗ . Therefore, Gβ is closed, and hence compact. Since π(r) ≥ ̺∗ for all π ∈ G, the equality π∗ (r) = ̺∗ follows. Also v is obtained by disintegrating π∗ . Remark 3.7. The reader might have noticed at this point that Assumption 3.1 may be weakened significantly. What is really required is the existence of an open set Ĥ ⊂ Rd × U and inf-compact functions V ∈ C 2 (Rd ) and h ∈ C(Rd × U), satisfying (H1) inf {u : (x,u)∈Ĥ} r(x, u) −→ ∞. (H2) \|x\|→∞ u L V(x) ≤ 1 − h(x, u)IĤc (x, u) + r(x, u)IĤ (x, u) ∀(x, u) ∈ Rd × U. In (H1), we use the convention that the ‘inf’ of the empty set is +∞. Also note that (H1) is equivalent to the statement that {(x, u) : r(x, u) ≤ c} ∩ Ĥ is bounded in Rd × U for all c ∈ R+ . If (H1)–(H2) are met, we define H := Ĥ ∪ {(x, u) ∈ Rd × U : r(x, u) > h(x, u)}, and following the proof of Lemma 3.3, we assert the existence of an inf-compact h̃ ∈ C(Rd × U) satisfying (3.23). In fact, throughout the rest of the paper, Assumption 3.1 is not really invoked. We only use (3.24), the inf-compact function h̃ satisfying (3.23), and, naturally, Assumption 3.2. 3.5. The HJB equation. For ε > 0, let rε (x, u) := r(x, u) + εh̃(x, u). By Theorem 3.1(d), for any π ∈ Gβ , β > ̺∗ , we have the bound (3.30) π(rε ) ≤ β + εk0 (1 + β). ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 27 Therefore, since rε is near-monotone, that is, lim inf min rε (x, u) > inf π(rε ), \|x\|→∞ u∈U π∈G there exists πε ∈ arg minπ∈G π(rε ). Let π∗ ∈ G be as in the proof of Theorem 3.2. The sub-optimality of π∗ relative to the running cost rε and (3.30) imply that πε (r) ≤ πε (rε ) (3.31) ≤ π∗ (rε ) ≤ ̺∗ + εk0 (1 + ̺∗ ) ∀ε > 0. It follows from (3.31) and Theorem 3.1(d) that {πε : ε ∈ (0, 1)} is tight. Since πε 7→ πε (r) is lower semi-continuous, if π̄ is any limit point of πε as ε ց 0, then taking limits in (3.31), we obtain (3.32) π̄(r) ≤ lim sup πε (r) ≤ ̺∗ . εց0 Since G is closed, π̄ ∈ G, which implies that π̄(r) ≥ ̺∗ . Therefore, equality must hold in (3.32), or in other words, π̄ is an optimal ergodic occupation measure. Theorem 3.3. There exists a unique function V ε ∈ C 2 (Rd ) with V ε (0) = 0, which is bounded below in Rd , and solves the HJB (3.33) min[Lu V ε (x) + rε (x, u)] = ̺ε , u∈U where ̺ε := inf π∈G π(rε ), or in other words, ̺ε is the optimal value of the ergodic control problem with running cost rε . Also a stationary Markov control vε is optimal for the ergodic control problem relative to rε if and only if it satisfies (3.34) Hε (x, ∇V ε (x)) = b(x, vε (x)) · ∇V ε (x) + rε (x, vε (x)) where (3.35) Hε (x, p) := min[b(x, u) · p + rε (x, u)]. u∈U Moreover: (a) for every R > 0, there exists kR such that (3.36) osc V ε ≤ kR ; BR a.e. in Rd , 28 A. ARAPOSTATHIS, A. BISWAS AND G. PANG (b) if vε is a measurable a.e. selector from the minimizer of the Hamiltonian in (3.35), that is, if it satisfies (3.33), then for any δ > 0, Z τ̆δ V ε (x) ≥ Evxε (rε (Xs , vε (Xs )) − ̺ε ) ds + inf V ε ; Bδ 0 (c) for any stationary control v ∈ USSM and for any δ > 0, Z τ̆δ ε v ε V (x) ≤ Ex (rε (Xs , v(Xs )) − ̺ε ) ds + V (Xτ̆δ ) , 0 where τ̆δ is hitting time to the ball Bδ . Theorem 3.4. following hold: Let V ε , ̺ε , and vε , for ε > 0, be as in Theorem 3.3. The (a) The function V ε converges to some V∗ ∈ C 2 (Rd ), uniformly on compact sets, and ̺ε → ̺∗ , as ε ց 0, and V∗ satisfies min[Lu V∗ (x) + r(x, u)] = ̺∗ . (3.37) u∈U Also, any limit point v∗ (in the topology of Markov controls) as ε ց 0 of the set {vε } satisfies Lv∗ V∗ (x) + r(x, v∗ (x)) = ̺∗ a.e. in Rd . (b) A stationary Markov control v is optimal for the ergodic control problem relative to r if and only if it satisfies a.e. in Rd , (3.38) H(x, ∇V∗ (x)) = b(x, v(x)) · ∇V∗ (x) + r(x, v(x)) where H(x, p) := min[b(x, u) · p + r(x, u)]. u∈U Moreover, for an optimal v ∈ USM , we have Z T 1 v r(Xs , v(Xs )) ds = ̺∗ lim Ex T →∞ T 0 ∀x ∈ Rd . (c) The function V∗ has the stochastic representation Z τ̆δ v Ex V∗ (x) = lim S inf (r(Xs , v(Xs )) − ̺∗ ) ds δց0 v∈ (3.39) = lim δց0 Ev̄x β β>0 USM Z 0 τ̆δ 0 (r(Xs , v∗ (Xs )) − ̺∗ ) ds for any v̄ ∈ USM that satisfies (3.38). ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 29 (d) If U is a convex set, u 7→ {b(x, u) · p + r(x, u)} is strictly convex whenever it is not constant, and u 7→ h̃(x, u) is strictly convex for all x, then any measurable minimizer of (3.33) converges pointwise, and thus in USM , to the minimizer of (3.37). Theorem 3.4 guarantees the existence of an optimal stable control, which is made precise by (3.38), for the ergodic diffusion control problem with the running cost function r. Moreover, under the convexity property in part (d), the optimal stable control can be obtained as a pointwise limit from the minimizing selector of (3.33). For instance, if we let + r(x, u) = (e · x) d X hi um i , m > 1, i=1 then by choosing h and h̃ + \|u\|2 as in Proposition 3.1, we see that the approximate value function V ε and approximate control vε converge to the desired optimal value function V∗ and optimal control v∗ , respectively. Concerning the uniqueness of the solution to the HJB equation in (3.37), recall that in the near-monotone case the existing uniqueness results are as follows: there exists a unique solution pair (V, ̺) of (3.37) with V in the class of functions C 2 (Rd ) which are bounded below in Rd . Moreover, it satisfies V (0) = 0 and ̺ ≤ ̺∗ . If the restriction ̺ ≤ ̺∗ is removed, then in general, there are multiple solutions. Since in our model r is not near-monotone in Rd , the function V∗ is not, in general, bounded below. However, as we show later in Lemma 3.10 the negative part of V∗ grows slower than V, that is, it holds that V∗− ∈ o(V), with o(·) as defined in Section 1.3. Therefore, the second part of the theorem that follows may be viewed as an extension of the well-known uniqueness results that apply to ergodic control problems with near-monotone running cost. The third part of the theorem resembles the hypotheses of uniqueness that apply to problems under a blanket stability hypothesis. Theorem 3.5. (3.40) Let (V̂ , ̺ˆ) be a solution of min[Lu V̂ (x) + r(x, u)] = ̺ˆ, u∈U such that V̂ − ∈ o(V) and V̂ (0) = 0. Then the following hold: (a) Any measurable selector v̂ from the minimizer of the associated Hamiltonian in (3.38) is in USSM and ̺v̂ < ∞. (b) If ̺ˆ ≤ ̺∗ then necessarily ̺ˆ = ̺∗ and V̂ = V∗ . (c) If V̂ ∈ O(minu∈U h̃(·, u)), then ̺ˆ = ̺∗ and V̂ = V∗ . 30 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Applying these results to the multi-class queueing diffusion model, we have the following corollary. Corollary 3.1. For the queueing diffusion model with controlled dynamics given by (3.10), drift given by (3.9), and running cost as in (3.11), there exists a unique solution V , satisfying V (0) = 0, to the associated HJB in the class of functions C 2 (Rd ) ∩ O(\|x\|m ), whose negative part is in o(\|x\|m ). This solution agrees with V∗ in Theorem 3.4. Proof. Existence of a solution V follows by Theorem 3.4. Select V ∼ \|x\|m as in the proof of Proposition 3.1. That the solution V is in the stated class then follows by Lemma 3.10 and Corollary 4.1 that appear later in Sections 3.6 and 4, respectively. With h ∼ \|x\|m as in the proof of Proposition 3.1, it follows that minu∈U h̃(x, u) ∈ O(\|x\|m ). Therefore, uniqueness follows by Theorem 3.5. We can also obtain the HJB equation in (3.37) via the traditional vanishing discount approach as the following theorem asserts. Similar results are shown for a one-dimensional degenerate ergodic diffusion control problem in [29] and certain multi-dimensional ergodic diffusion control problems (allowing degeneracy and spatial periodicity) in [2]. Theorem 3.6. Let V∗ and ̺∗ be as in Theorem 3.4. For α > 0, we define Z ∞ −αt U e r(Xt , Ut ) dt . Vα (x) := inf Ex U ∈U 0 The function Vα − Vα (0) converges, as α ց 0, to V∗ , uniformly on compact subsets of Rd . Moreover, αVα (0) → ̺∗ , as α ց 0. The proofs of the Theorems 3.3–3.6 are given in Section 3.6. The following result, which follows directly from (3.31), provides a way to find ε-optimal controls. Proposition 3.2. Let {vε } be the minimizing selector from Theorem 3.3 and {µvε } be the corresponding invariant probability measures. Then almost surely for all x ∈ Rd , Z Z T 1 vε lim Ex r(x, vε (x))µvε (dx) r(Xs , vε (Xs )) ds = T →∞ T Rd 0 ≤ ̺∗ + εk0 (1 + ̺∗ ). ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 31 3.6. Technical proofs. Recall that rε (x, u) = r(x, u) + εh̃(x, u), with h̃ as in Lemma 3.3. We need the following lemma. For α > 0 and ε ≥ 0, we define Z ∞ ε U −αt Vα (x) := inf Ex (3.41) e rε (Xt , Ut ) dt , U ∈U 0 where we set r0 ≡ r. Clearly, when ε = 0, we have Vα0 ≡ Vα . We quote the following result from [1], Theorem 3.5.6, Remark 3.5.8. Lemma 3.4. Provided ε > 0, then Vαε defined above is in C 2 (Rd ) and is the minimal nonnegative solution of min[Lu Vαε (x) + rε (x, u)] = αVαε (x). u∈U The HJB in Lemma 3.4 is similar to the equation in [7], Theorem 3, which concerns the characterization of the discounted control problem. Lemma 3.5. Let u be any precise Markov control and Lu be the corresponding generator. Let ϕ ∈ C 2 (Rd ) be a nonnegative solution of Lu ϕ − αϕ = g, d where g ∈ L∞ loc (R ). Let κ : R+ → R+ be any nondecreasing function such that kgkL∞ (BR ) ≤ κ(R) for all R > 0. Then for any R > 0 there exists a constant D(R) which depends on κ(4R), but not on u, or ϕ, such that osc ϕ ≤ D(R) 1 + α inf ϕ . BR B4R Proof. Define g̃ := α(g − 2κ(4R)) and ϕ̃ := 2κ(4R) + αϕ. Then g̃ ≤ 0 in B4R and ϕ̃ solves Lu ϕ̃ − αϕ̃ = g̃ in B4R . Also kg̃kL∞ (B4R ) ≤ α(2κ(4R) + kgkL∞ (B4R ) ) ≤ 3α(2κ(4R) − kgkL∞ (B4R ) ) = 3 inf \|g̃\| B4R ≤ 3\|B4R \|−1 kg̃kL1 (B4R ) . Hence by [1], Theorem A.2.13, there exists a positive constant C̃H such that sup ϕ̃(x) ≤ C̃H inf ϕ̃(x), x∈B3R x∈B3R 32 A. ARAPOSTATHIS, A. BISWAS AND G. PANG implying that α sup ϕ(x) ≤ C̃H 2κ(4R) + inf αϕ(x) . (3.42) x∈B3R x∈B3R We next consider the solution of Lu ψ = 0 in B3R , ψ=ϕ on ∂B3R . Then Lu (ϕ − ψ) = αϕ + g in B3R . If ϕ(x̂) = inf x∈B3R ϕ(x), then applying the maximum principle ([1], Theorem A.2.1, [18]) it follows from (3.42) that sup \|ϕ − ψ\| ≤ Ĉ(1 + αϕ(x̂)). (3.43) x∈B3R Again ψ attains its minimum at the boundary ([1], Theorem A.2.3, [18]). Therefore, ψ − ϕ(x̂) is a nonnegative function, and hence by the Harnack inequality, there exists a constant CH > 0 such that ψ(x) − ϕ(x̂) ≤ CH (ψ(x̂) − ϕ(x̂)) ≤ CH Ĉ(1 + αϕ(x̂)) ∀x ∈ B2R . Thus, combining the above display with (3.43) we obtain osc ϕ ≤ sup(ϕ − ψ) + sup ψ − ϕ(x̂) ≤ Ĉ(1 + CH )(1 + αϕ(x̂)). B2R B2R B2R This completes the proof. Lemma 3.6. Let Vαε be as in Lemma 3.4. Then for any R > 0, there exists a constant kR > 0 such that osc Vαε ≤ kR for all α ∈ (0, 1] and ε ∈ [0, 1]. BR Proof. Recall that µu0 is the stationary probability distribution for the process under the control u0 ∈ USSM in Lemma 3.1. Since u0 is sub-optimal for the α-discounted criterion in (3.41), and Vαε is nonnegative, then for any ball BR , using Fubini’s theorem, we obtain Z ε Vαε (x)µu0 (dx) µu0 (BR ) inf Vα ≤ BR Rd ≤ Z Rd Eux0 Z ∞ 0 −αt e rε (Xt , u0 (Xt )) dt µu0 (dx) 1 µu (rε ) α 0 1 ≤ (η + εk0 (1 + η)), α = ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 33 where for the last inequality we used Lemma 3.2 and Theorem 3.1(a). Therefore, we have the estimate α inf Vαε ≤ BR η + εk0 (1 + η) . µu0 (BR ) The result then follows by Lemma 3.5. We continue with the proof of Theorem 3.3. Proof of Theorem 3.3. Consider the function V̄αε := Vαε − Vαε (0). In view of Lemma 3.5 and Lemma 3.6, we see that V̄αε is locally bounded uniformly in α ∈ (0, 1] and ε ∈ (0, 1]. Therefore, by standard elliptic theory, V̄αε and its first- and second-order partial derivatives are uniformly bounded in Lp (B), for any p > 1, in any bounded ball B ⊂ Rd , that is, for some constant CB depending on B and p, kV̄αε kW2,p (B) ≤ CB ([18], Theorem 9.11, page 117). Therefore, we can extract a subsequence along which V̄αε converges. Then the result follows from Theorems 3.6.6, Lemma 3.6.9 and Theorem 3.6.10 in [1]. The proof of (3.36) follows from Lemma 3.5 and Lemma 3.6. Remark 3.8. In the proof of the following lemma, and elsewhere in the paper, we use the fact that if U ⊂ USSM is a any set of controls such that the corresponding set {µv : v ∈ U } ⊂ M of invariant probability measures is tight then the map v 7→ πv from the closure of U to P(Rd × U) is continuous, and so is the map v 7→ µv . In fact, the latter is continuous under the total variation norm topology [1], Lemma 3.2.6. We also recall that G and M are closed and convex subsets of P(Rd × U) and P(Rd ). Therefore,{πv : v ∈ Ū } is compact in G. Note also that since U is compact, tightness of a set of invariant probability measures is equivalent to tightness of the corresponding set of ergodic occupation measures. Lemma 3.7. If {vε : ε ∈ (0, 1]} is a collection of measurable selectors from the minimizer of (3.33), then the corresponding invariant probability measures {µε : ε ∈ (0, 1]} are tight. Moreover, if vεn → v∗ along some subsequence εn ց 0, then the following hold: (a) µεn → µv∗ as εn ց 0, (b) Rv∗ is a stable Markov control, (c) Rd r(x, v∗ (x))µv∗ (dx) = limεց0 ̺ε = ̺∗ . Proof. By (3.25) and (3.31), the set of ergodic occupation measures corresponding to {vε : ε ∈ (0, 1]} is tight. By Remark 3.8, the same applies to the set {µε : ε ∈ (0, 1]}, and also part (a) holds. Part (b) follows from the equivalence of the existence of an invariant probability measure for a 34 A. ARAPOSTATHIS, A. BISWAS AND G. PANG controlled diffusion and the stability of the associated stationary Markov control (see [1], Theorem 2.6.10). Part (c) then follows since equality holds in (3.32). We continue with the following lemma that asserts the continuity of the mean hitting time of a ball with respect to the stable Markov controls. Lemma 3.8. Let {vn : n ∈ N} ⊂ UβSM , for some β > 0, be a collection of Markov controls such that vn → v̂ in the topology of Markov controls as n → ∞. Let µn , µ̂ be the invariant probability measures corresponding to the controls vn , v̂, respectively. Then for any δ > 0, it holds that Evxn [τ̆δ ] −→ Ev̂x [τ̆δ ] ∀x ∈ Bδc . n→∞ Proof. Define H(x) := minu∈U h̃(x, u). It is easy to see that H is infcompact and locally Lipschitz. Therefore, by Theorem 3.1(d) we have sup µn (H) ≤ k0 (1 + β), n∈N and since µn → µ̂, we also have µ̂(H) ≤ k0 (1 + β). Then by [1], Lemma 3.3.4, we obtain Z τ̆δ Z τ̆δ sup Evxn H(Xs ) ds + Ev̂x (3.44) H(Xs ) ds < ∞. n∈N 0 0 Let R be a positive number greater than \|x\|. Then by (3.44), there exists a positive k such that Z τ̆δ Z τ̆δ 1 k Evx I{H>R} (Xs ) ds ≤ Evx H(Xs )I{H>R} (Xs ) ds ≤ R R 0 0 for v ∈ {{vn }, v̂}. From this assertion and (3.44), we see that Z τ̆δ v sup Ex I{H>R} (Xs ) ds −→ 0. v∈{{vn },v̂} R→∞ 0 Therefore, in order to prove the lemma it is enough to show that, for any R > 0, we have Z τ̆δ Z τ̆δ vn v̂ Ex I{H≤R} (Xs ) ds −→ Ex I{H≤R} (Xs ) ds . 0 n→∞ 0 But this follows from [1], Lemma 2.6.13(iii). ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 35 Lemma 3.9. Let (V ε , ̺ε ) be as in Theorem 3.3, and vε satisfy (3.35). There exists a subsequence εn ց 0, such that V εn converges to some V∗ ∈ C 2 (Rd ), uniformly on compact sets, and V∗ satisfies min[Lu V∗ (x) + r(x, u)] = ̺∗ . (3.45) u∈U Also, any limit point v∗ (in the topology of Markov controls) of the set {vε }, as ε ց 0, satisfies (3.46) Lv∗ V∗ (x) + r(x, v∗ (x)) = ̺∗ a.e. in Rd . Moreover, V∗ admits the stochastic representation Z τ̆δ v V∗ (x) = S inf Ex (r(Xs , v(Xs )) − ̺∗ ) ds + V∗ (Xτ̆δ ) v∈ (3.47) = Evx∗ β β>0 USM Z 0 τ̆δ 0 (r(Xs , v∗ (Xs )) − ̺∗ ) ds + V∗ (Xτ̆δ ) . It follows that V∗ is the unique limit point of V ε as ε ց 0. Proof. From (3.36), we see that the family {V ε : ε ∈ (0, 1]} is uniformly locally bounded. Hence, applying the theory of elliptic PDE, it follows that d {V ε : ε ∈ (0, 1]} is uniformly bounded in W2,p loc (R ) for p > d. Consequently, 1,γ ε {V : ε ∈ (0, 1]} is uniformly bounded in Cloc for some γ > 0. Therefore, along some subsequence εn ց 0, V εn → V∗ ∈ W2,p ∩ C 1,γ , as n → ∞, uniformly on compact sets. Also, limεց0 ̺ε = ̺∗ by Lemma 3.6(c). Therefore, passing to the limit we obtain the HJB equation in (3.45). It is straightforward to verify that (3.46) holds [1], Lemma 2.4.3. By Theorem 3.3(c), taking limits as ε ց 0, we obtain Z τ̆δ Evx (3.48) V∗ (x) ≤ S inf (r(Xs , v(Xs )) − ̺∗ ) ds + V∗ (Xτ̆δ ) . v∈ β β>0 USM 0 Also by Theorem 3.3(b) we have the bound V ε (x) ≥ −̺ε Evxε [τ̆δ ] + inf V ε . Bδ Using Lemma 3.8 and taking limits as εn ց 0, we obtain the lower bound (3.49) V∗ (x) ≥ −̺∗ Evx∗ [τ̆δ ] + inf V∗ . Bδ By Lemma 3.7(c) and Theorem 3.1(d), v∗ ∈ USSM , and πv∗ (h̃) ≤ k0 (1 + ̺∗ ). Define Z τ̆δ v∗ ϕ(x) := Ex (3.50) h̃(Xs , v∗ (Xs )) ds . 0 36 A. ARAPOSTATHIS, A. BISWAS AND G. PANG For \|x\| > δ, we have Evx∗ [I{τR <τ̆δ } ϕ(XτR )] = Evx∗ I{τR <τ̆δ } Z τ̆δ τR ∧τ̆δ h̃(Xs , v∗ (Xs )) ds . Therefore, by the dominated convergence theorem and the fact that ϕ(x) < ∞ we obtain Evx∗ [ϕ(XτR )I{τR <τ̆δ } ] −→ 0. Rր∞ By (3.48) and (3.49), we have \|V∗ \| ∈ O(ϕ). Thus (3.49) and (3.50) imply that lim inf Evx∗ [V∗ (XτR )I{τR <τ̆δ } ] = 0, Rր∞ and thus lim inf Evx∗ [V∗ (XτR ∧τ̆δ )] = Evx∗ [V∗ (Xτ̆δ )]. (3.51) Rր∞ Applying Itô’s formula to (3.46), we obtain Z τ̆δ ∧τR v∗ (3.52) V∗ (x) = Ex (r(Xs , v∗ (Xs )) − ̺∗ ) ds + V∗ (Xτ̆δ ∧τR ) . 0 Taking limits as R → ∞, and using the dominated convergence theorem, we obtain (3.47) from (3.48). Recall the definition of o(·) from Section 1.3. We need the following lemma. Lemma 3.10. Let V∗ be as in Lemma 3.9. It holds that V∗− ∈ o(V). Proof. Let v∗ be as in Lemma 3.9. Applying Itô’s formula to (3.24) with u ≡ v∗ we obtain Z τ̆δ v∗ Ex h(Xs , v∗ (Xs ))IHc (Xs , v∗ (Xs )) ds (3.53) 0 ≤ Evx∗ Z τ̆δ 0 r(Xs , v∗ (Xs ))IH (Xs , v∗ (Xs )) ds + Evx∗ [τ̆δ ] + V(x). Therefore, adding the term Z τ̆δ v∗ Ex r(Xs , v∗ (Xs ))IH (Xs , v∗ (Xs )) ds − (1 + 2̺∗ )Evx∗ [τ̆δ ] 0 ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 37 to both sides of (3.53) and using the stochastic representation of V∗ we obtain Z τ̆δ −1 v∗ F (x) := 2k0 Ex h̃(Xs , v∗ (Xs )) ds − 2(1 + ̺∗ )Evx∗ [τ̆δ ] (3.54) 0 ≤ 2V∗ (x) + V(x) − 2 inf V∗ . Bδ From the stochastic representation of V∗ we have V∗− (x) ≤ ̺∗ Evx∗ [τ̆δ ]−inf Bδ V∗ . For any R > δ, we have Z τ̆δ v∗ c (3.55) Ex h̃ Ex [τ̆R ] ∀x ∈ BR . h̃(Xs , v∗ (Xs )) ds ≥ inf c BR ×U 0 It is also straightforward to show that lim\|x\|→∞ EExx[τ̆[τ̆Rδ ]] = 1. Therefore, since h̃ is inf-compact, it follows by (3.54) and (3.55) that the map x 7→ Evx∗ [τ̆δ ] is in o(F ), which implies that V∗− ∈ o(F ). On the other hand, by (3.54) we obtain F (x) ≤ V(x) − 2 supBδ V∗ for all x such that V∗ (x) ≤ 0, which implies that the restriction of F to the support of V∗− is in O(V). It follows that V∗− ∈ o(V). We next prove Theorem 3.4. Proof of Theorem 3.4. Part (a) is contained in Lemma 3.9. To prove part (b), let v̄ be any control satisfying (3.38). By Lemma 3.10 the map V + 2V∗ is inf-compact and by Theorem 3.4 and (3.24) it satisfies Lv̄ (V + 2V∗ )(x) ≤ 1 + 2̺∗ − r(x, v̄(x)) − h(x, v̄(x))IHc (x, v̄(x)) ≤ 2 + 2̺∗ − 2k0−1 h̃(x, v̄(x)) ∀x ∈ Rd . This implies that v̄ ∈ USSM . Applying Itô’s formula, we obtain Z T 1 v̄ lim sup Ex (3.56) h̃(Xs , v̄(Xs )) ds ≤ k0 (1 + ̺∗ ). T →∞ T 0 Therefore, πv̄ (h̃) < ∞. By (3.24), we have Z t v̄ v̄ Ex [V(Xt )] ≤ V(x) + t + Ex r(Xs , v̄(Xs )) ds , 0 and since r ≤ h̃, this implies by (3.56) that 1 (3.57) lim sup Ev̄x [V(XT )] ≤ 1 + k0 (1 + ̺∗ ). T →∞ T Since V∗− ∈ o(V), it follows by (3.57) that 1 lim sup Ev̄x [V∗− (XT )] = 0. T →∞ T 38 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Therefore, by Itô’s formula, we deduce from (3.37) that Z T 1 v̄ (3.58) r(Xs , v̄(Xs )) ds ≤ ̺∗ . lim sup Ex T →∞ T 0 On the other hand, since the only limit point of the mean empirical measures v̄ , as t → ∞, is π , and π (r) = ̺ , then in view of Remark 3.6, we obtain ζx,t v̄ v̄ ∗ v̄ (r) ≥ ̺ . This proves that equality holds in (3.58) and that lim inf t→∞ ζx,t ∗ the “lim sup” may be replaced with “lim.” Conversely, suppose v ∈ USM is optimal but does not satisfy (3.38). Then there exists R > 0 and a nontrivial nonnegative f ∈ L∞ (BR ) such that fε (x) := IBR (x)(Lv V ε (x) + rε (x, v(x)) − ̺ε ) converges to f , weakly in L1 (BR ), along some subsequence ε ց 0. By applying Itô’s formula to (3.33), we obtain Z T ∧τR 1 v 1 v ε ε (E [V (XT ∧τR )] − V (x)) + Ex rε (Xs , v(Xs )) ds T x T 0 (3.59) Z T ∧τR 1 v fε (Xs , v(Xs )) ds . ≥ ̺ε + Ex T 0 Define, for some δ > 0, G(x) := Evx Z 0 τ̆δ rε (Xs , v(Xs )) ds . Since V is bounded from below, by Theorem 3.3(c) we have V ε ∈ O(G). Invoking [1], Corollary 3.7.3, we obtain ε 1 v ε Ex [V (XT )] = 0, T →∞ T lim and lim Evx [V ε (XT ∧τR )] = Evx [V ε (XT )]. R→∞ Therefore, taking limits in (3.59), first as R ր ∞, and then as T → ∞, we obtain (3.60) πv (rε ) ≥ ̺ε + πv (fε ). Taking limits as ε ց 0 in (3.60), since µv has a strictly positive density in BR , we obtain πv (r) ≥ ̺∗ + πv (f ) > ̺∗ , which is a contradiction. This completes the proof of part (b). ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 39 The first equality (3.39) follows by Lemma 3.9, taking limits as δ ց 0. To show that the second equality holds for any optimal control, suppose v̄ satisfies (3.38). By (3.24) we have, for δ > 0 and \|x\| > δ, Z τR ∧τ̆δ (1 + r(Xs , v̄(Xs ))) ds . Ev̄x [V(XτR )I{τR <τ̆δ } ] ≤ V(x) + sup V − + Ev̄x Bδ 0 It follows that (see [1], Lemma 3.3.4) lim sup Ev̄x [V(XτR )I{τR <τ̆δ } ] < ∞, R→∞ and since V∗− ∈ o(V) we must have lim sup Ev̄x [V∗− (XτR )I{τR <τ̆δ } ] = 0. R→∞ By the first equality in (3.47), we obtain V∗+ ∈ O(ϕ), with ϕ as defined in (3.50) with v∗ replaced by v̄. Thus, in analogy to (3.51), we obtain lim inf Ev̄x [V∗ (XτR ∧τ̆δ )] = Ev̄x [V∗ (Xτ̆δ )]. Rր∞ The rest follows as in the proof of Lemma 3.9 via (3.52). We next prove part (d). We assume that U is a convex set and that c(x, u, p) := {b(x, u) · p + r(x, u)} is strictly convex in u if it is not identically a constant for fixed x and p. We fix some point ū ∈ U. Define B := {x ∈ Rd : c(x, ·, p) = c(x, ū, p) for all p}. It is easy to see that on B both b and r do not depend on u. It is also easy to check that B is a closed set. Let (V∗ , v∗ ) be the limit of (V ε , vε ), where V∗ is the solution to (3.37) and v∗ is the corresponding limit of vε . We have already shown that v∗ is a stable Markov control. We next show that it is, in fact, a precise Markov control. By our assumption, vε is the unique minimizing selector in (3.34) and, moreover, vε is continuous in x. By the definition of rε it is clear that the restriction of vε to B does not depend on ε. Let vε (x) = v ′ (x) on B. Using the strict convexity property of c(x, ·, ∇V∗ ) it is easy to verify that vε converges to the unique minimizer of (3.37) on B c . In fact, since B c is open, then for any sequence xε → x ∈ B c it holds that vε (xε ) → v∗ (x). This follows from the definition of the minimizer and the uniform convergence of ∇V ε to ∇V∗ . Therefore, we see that v∗ is a precise Markov control, v∗ = v ′ on B, and vε → v∗ pointwise as ε → 0. It is also easy to check that pointwise convergence implies convergence in the topology of Markov controls. We now embark on the proof of Theorem 3.5. 40 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Proof of Theorem 3.5. The hypothesis that V̂ − ∈ o(V) implies that the map V + 2V̂ is inf-compact. Also by (3.24) and (3.40), it satisfies Lv̂ (V + 2V̂ )(x) ≤ 1 + 2ˆ ̺ − r(x, v̂(x)) − h(x, v̂(x))IHc (x, v̂(x)) ≤ 2 + 2ˆ ̺ − 2k0−1 h̃(x, v̂(x)) ∀x ∈ Rd . R Therefore, h̃(x, v̂(x)) dπv̂ < ∞ from which it follows that ̺v̂ < ∞. This proves part (a). By (3.24), we have Z t v̂ v̂ r(Xs , v̂(Xs )) ds , Ex [V(Xt )] ≤ V(x) + t + Ex 0 and since ̺v̂ < ∞, this implies that 1 lim sup Ev̂x [V(XT )] ≤ 1 + ̺v̂ . (3.61) T →∞ T Since V̂ − ∈ o(V), it follows by (3.61) that lim sup T →∞ 1 v̂ − E [V̂ (XT )] = 0. T x Therefore, by Itô’s formula, we deduce from (3.40) that Z T 1 v̂ 1 v̂ + E [V̂ (XT )] + Ex r(Xs , v̂(Xs )) ds = ̺ˆ. lim sup T x T T →∞ 0 This implies that ̺v̂ ≤ ̺ˆ and since by hypothesis ̺ˆ ≤ ̺∗ we must have ̺ˆ = ̺∗ . Again by (3.24), we have Z τR ∧τ̆δ v̂ − v̂ Ex [V(XτR )I{τR <τ̆δ } ] ≤ V(x) + sup V + Ex (1 + r(Xs , v̂(Xs ))) ds . Bδ 0 It follows by [1], Lemma 3.3.4, that lim sup Ev̂x [V(XτR )I{τR <τ̆δ } ] < ∞, R→∞ and since V̂ − ∈ o(V) we must have (3.62) lim sup Ev̂x [V̂ − (XτR )I{τR <τ̆δ } ] = 0. R→∞ Using (3.62) and following the steps in the proof of the second equality in (3.47), we obtain Z τ̆δ v̂ V̂ (x) ≥ Ex (r(Xs , v̂(Xs )) − ̺∗ ) ds + inf V̂ Bδ 0 ≥ V∗ (x) − sup V∗ + inf V̂ . Bδ Bδ 41 ERGODIC CONTROL IN THE HALFIN–WHITT REGIME Taking limits as δ ց 0, we have V∗ ≤ V̂ . Since Lv̂ (V∗ − V̂ ) ≥ 0 and V∗ (0) = V̂ (0), we must have V̂ = V∗ on Rd , and the proof of part (b) is complete. R To prove part (c) note that by part (a) we haveR ̺v̂ < ∞. Therefore, h̃ dπv̂ ≤ ∞ by Theorem 3.1(a), which implies that \|V̂ \| dµv̂ ≤ ∞ by the hypothesis. Therefore, Ev̂x (\|V̂ (Xt )\|) converges as t → ∞ by [23], Proposition 2.6, which of course implies that 1t Ev̂x (\|V̂ (Xt )\|) tends to 0 as t → ∞. Similarly, we deduce that 1t Evx∗ (\|V̂ (Xt )\|) as t → ∞. Applying Itô’s formula to (3.40), with u ≡ v∗ , we obtain ̺ˆ ≤ ̺∗ . Another application with u ≡ v̂ results in ̺ˆ = ̺v̂ . Therefore, ̺ˆ = ̺∗ . The result then follows by part (b). We finish this section with the proof of Theorem 3.6. Proof of Theorem 3.6. We first show that limαց0 αVα (0) = ̺∗ . Let Ṽ(t, x) := e−αt V(x), and τn (t) := τn ∧ t. Applying Itô’s formula to (3.24), we obtain Z τn (t) U U Ex [Ṽ(τn (t), Xτn (t) )] ≤ V(x) − Ex αṼ(s, Xs ) ds 0 It follows that Z U Ex (3.63) τn (t) −αs e 0 ≤ + EU x Z + EU x Z τn (t) −αs e 0 τn (t) −αs e 0 (1 − h(Xs , Us ))IHc (Xs , Us ) ds (1 + r(Xs , Us ))IH (Xs , Us ) ds . h(Xs , Us )IHc (Xs , Us ) ds 1 + V(x) + EU x α Z τn (t) 0 e−αs r(Xs , Us )IH (Xs , Us ) ds . Taking limits first as n ր ∞ and then as t ր ∞ in (3.63), and evaluating U at an optimal α-discounted control vα∗ , relative to r we obtain the estimate, using also (3.23), Z ∞ 2 ∗ e−αs h̃(Xs , vα∗ (Xs )) ds ≤ + V(x) + 2Vα (x). (3.64) 2k0−1 Exvα α 0 By (3.23) and (3.64), it follows that Z ∗ ε vα Vα (x) ≤ Vα (x) ≤ Ex 0 ∞ −αs e rε (Xs , vα∗ (Xs )) ds ≤ Vα (x) + εk0 (α−1 + V(x) + Vα (x)). 42 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Multiplying by α and taking limits as α ց 0 we obtain lim sup αVα (0) ≤ ̺ε ≤ (1 + εk0 ) lim sup αVα (0) + εk0 . αց0 αց0 The same inequalities hold for the “lim inf.” Therefore, limαց0 αVα (0) = ̺∗ . Let Ṽ := lim (Vα − Vα (0)). αց0 (Note that a similar result as Lemma 3.5 holds.) Then Ṽ satisfies Z τ̆δ [ β v Ṽ (x) ≤ lim Ex (r(Xs , v(Xs )) − ̺∗ ) ds ∀v ∈ USM . δց0 0 β>0 This can be obtained without the near-monotone assumption on the running cost; see, for example, [1], Lemma 3.6.9 or Lemma 3.7.8. It follows from (3.39) that Ṽ ≤ V∗ . On the other hand, since Lv∗ (Ṽ − V∗ ) ≥ 0, and Ṽ (0) = V∗ (0), we must have Ṽ = V∗ by the strong maximum principle. 4. Approximation via spatial truncations. We introduce an approximation technique which is in turn used to prove the asymptotic convergence results in Section 5. Let v0 ∈ USSM be any control such that πv0 (r) < ∞. We fix the control v0 on the complement of the ball B̄l and leave the parameter u free inside. In other words, for each l ∈ N we define b(x, u), if (x, u) ∈ B̄l × U, bl (x, u) := b(x, v0 (x)), otherwise, r(x, u), if (x, u) ∈ B̄l × U, rl (x, u) := r(x, v0 (x)), otherwise. We consider the family of controlled diffusions, parameterized by l ∈ N, given by (4.1) dXt = bl (Xt , Ut ) dt + σ(Xt ) dWt , with associated running costs rl (x, u). We denote by USM (l, v0 ) the subset of USM consisting of those controls v which agree with v0 on B̄lc . Let η0 := πv0 (r). It is well known that there exists a nonnegative solution ϕ0 ∈ d W2,p loc (R ), for any p > d, to the Poisson equation (see [1], Lemma 3.7.8(ii)) Lv0 ϕ0 (x) = η0 − h̃(x, v0 (x)) x ∈ Rd , which is inf-compact, and satisfies, for all δ > 0, Z τ̆δ v0 ϕ0 (x) = Ex (h̃(Xs , v0 (Xs )) − η0 ) ds + ϕ0 (Xτ̆δ ) 0 ∀x ∈ Rd . ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 43 We recall the Lyapunov function V from Assumption 3.1. We have the following theorem. Theorem 4.1. Let Assumptions 3.1 and 3.2 hold. Then for each l ∈ N d l there exists a solution V l in W2,p loc (R ), for any p > d, with V (0) = 0, of the HJB equation min[Lul V l (x) + rl (x, u)] = ̺l , (4.2) u∈U Lul where is the elliptic differential operator corresponding to the diffusion in (4.1). Moreover, the following hold: (i) (ii) 2ϕ0 (x) (iii) (iv) ̺l is nonincreasing in l; there exists a constant C0 , independent of l, such that V l (x) ≤ C0 + for all l ∈ N; (V l )− ∈ o(V + ϕ0 ) uniformly over l ∈ N; the restriction of V l on Bl is in C 2 . Proof. As earlier, we can show that Z ∞ l U −αs Vα (x) := inf Ex e rl (Xs , Us ) ds U ∈U 0 is the minimal nonnegative solution to (4.3) min[Lul Vαl (x) + rl (x, u)] = αVαl (x), u∈U d and Vαl ∈ W2,p loc (R ), p > d. Moreover, any measurable selector from the minimizer in (4.3) is an optimal control. A similar estimate as in Lemma 3.5 holds and, therefore, there exists a subsequence {αn }, along which Vαl n (x) − Vαl n (0) d l converges to V l in W2,p loc (R ), p > d, and αn Vαn (0) → ̺l as αn ց 0, and (V l , ̺l ) satisfies (4.2) (see also [1], Lemma 3.7.8). To show that πvl (r) = ̺l , v l is a minimizing selector in (4.2), we use the following argument. Since πv0 (r) < ∞, we claim that there exists a nonnegative, inf-compact function g ∈ C(Rd ) such that πv0 (g · (1 + r)) < ∞. Indeed, this is true since integrability and uniform integrability of a function under any given measure are equivalent (see also the proof of [1], Lemma 3.7.2). Since every control in USM (l, v0 ) agrees with v0 on Blc , then for any x0 ∈ B̄lc the map Z τ̆l v v 7→ Ex0 g(Xs )(1 + r(Xs , v(Xs ))) ds 0 is constant on USM (l, v0 ). By the equivalence of (i) and (iii) in Lemma 3.3.4 of [1], this implies that sup v∈USM (l,v0 ) πv (g · (1 + r)) < ∞ ∀l ∈ N, 44 A. ARAPOSTATHIS, A. BISWAS AND G. PANG and thus r is uniformly integrable with respect to the family {πv : v ∈ USM (l, v0 )} for any l ∈ N. It then follows by [1], Theorem 3.7.11, that (4.4) ̺l = inf v∈USM (l,v0 ) πv (r), l ∈ N. This yields part (i). Moreover, in view of Lemmas 3.5 and 3.6, we deduce that for any δ > 0 it holds that supBδ \|V l \| ≤ κδ , where κδ is a constant independent of l ∈ N. It is also evident by (4.4) that ̺l is decreasing in l and ̺l ≤ η0 for all l ∈ N. Fix δ such that minu∈U h̃(x, u) ≥ 2η0 on Bδc . Since ϕ0 is nonnegative, we obtain Z τ̆δ v0 (4.5) ∀x ∈ Rd . (h̃(Xs , v0 (Xs )) − η0 ) ds ≤ ϕ0 (x) Ex 0 Using an analogous argument as the one used in the proof of [1], Lemma 3.7.8, we have Z τ̆δ l v (4.6) V (x) ≤ Ex (rl (Xs , v(Xs )) − ̺l ) ds + κδ ∀v ∈ USM (l, v0 ). 0 Thus, by (4.5) and (4.6), and since by the choice of δ > 0, it holds that r ≤ h̃ ≤ 2(h̃ − η0 ) on Bδc , we obtain Z τ̆δ V l (x) ≤ Evx0 2(h̃(Xs , v0 (Xs )) − η0 ) ds + κδ 0 (4.7) ≤ κδ + 2ϕ0 (x) ∀x ∈ Rd . This proves part (ii). Now fix l ∈ N. Let vαl be a minimizing selector of (4.3). Note then that l vα ∈ USM (l, v0 ). Therefore, vαl is a stable Markov control. Let vαl n → v l in the topology of Markov controls along the same subsequence as above. Then it is evident that v l ∈ USM (l, v0 ). Also from Lemma 3.8, we have vl l Exαn [τ̆δ ] −→ Evx [τ̆δ ] ∀x ∈ Bδc , ∀δ > 0. αn ց0 Using [1], Lemma 3.7.8, we obtain the lower bound l V l (x) ≥ −̺l Evx [τ̆δ ] − κδ . (4.8) By [1], Theorem 3.7.12(i) (see also (3.7.50) in [1]), it holds that Z τ̆δ l vl l l V (x) = Ex (rl (Xs , v (Xs )) − ̺l ) ds + V (Xτ̆δ ) 0 (4.9) ≥ l Evx Z 0 τ̆δ l l rl (Xs , v (Xs )) ds − ̺l Evx [τ̆δ ] − κδ ∀x ∈ Blc . ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 45 By (3.23), we have 2k0−1 h̃(x, u)IH (x, u) ≤ 1 + r(x, u)IH (x, u). Therefore, using the preceding inequality and (4.9), we obtain l (4.10) V l (x) + (1 + ̺l )Evx [τ̆δ ] + κδ Z τ̆δ 2 vl l l ≥ Ex h̃(Xs , v (Xs ))IH (Xs , v (Xs )) ds . k0 0 By (3.24), (4.9) and the fact that V is nonnegative, we have Z τ̆δ 2 vl l Ex h̃(Xs , v l (Xs ))IHc (Xs , v l (Xs )) ds − V(x) − Evx [τ̆δ ] k0 0 Z τ̆δ vl l l ≤ Ex (4.11) r(Xs , v (Xs ))IH (Xs , v (Xs )) ds 0 l l ≤ V (x) + ̺l Evx [τ̆δ ] + κδ . Combining (4.7), (4.10) and (4.11), we obtain Z τ̆δ l l h̃(Xs , v l (Xs )) ds ≤ k0 (1 + ̺l )Evx [τ̆δ ] Evx 0 + k0 V(x) + 2k0 (ϕ0 (x) + κδ ) 2 for all l ∈ N. As earlier, using the inf-compact property of h̃ and the fact that ̺l ≤ η0 is bounded, we can choose δ large enough such that Z τ̆δ vl vl l (4.12) η0 Ex [τ̆δ ] ≤ Ex h̃(Xs , v (Xs )) ds ≤ k0 V(x) + 4k0 (ϕ0 (x) + κδ ) 0 for all l ∈ N. Since h̃ is inf-compact, part (iii) follows by (4.8) and (4.12). Part (iv) is clear from regularity theory of elliptic PDE [18], Theorem 9.19, page 243. Similar to Theorem 3.3, we can show that oscillations of {V l } are uniformly bounded on compacts. Therefore, if we let l → ∞ we obtain a HJB equation (4.13) min[Lu V̂ (x) + r(x, u)] = ̺ˆ, u∈U with V̂ ∈ C 2 (Rd ) and liml→∞ ̺l = ̺ˆ. By Theorem 4.1, we have the bound (4.14) V̂ (x) ≤ C0 + 2ϕ0 (x), 46 A. ARAPOSTATHIS, A. BISWAS AND G. PANG for some positive constant C0 . This of course, implies that V̂ + (x) ≤ C0 + 2ϕ0 (x). Moreover, it is straightforward to show that for any v ∈ USSM with ̺v < ∞, we have 1 lim sup Evx [V(Xt )] < ∞. t→∞ t Therefore, if in addition, we have 1 lim sup Evx [ϕ0 (Xt )] < ∞, t→∞ t then it follows by Theorem 4.1(iii) that (4.15) 1 lim sup V̂ − (Xt ) −→ 0. t→∞ t→∞ t Theorem 4.2. Suppose that ϕ0 ∈ O(minu∈U h̃(·, u)). Then, under the assumptions of Theorem 4.1, we have liml→∞ ̺l = ̺ˆ = ̺∗ , and V̂ = V∗ . Moreover, V∗ ∈ O(ϕ0 ). Proof. Let {v̂l } be any sequence of measurable selectors from the minimizer of (4.2) and {πl } the corresponding sequence of ergodic occupation measures. Since by Theorem 3.1 {πl } is tight, then by Remark 3.8 if v̂ is a limit point of a subsequence {v̂l }, which we also denote by {v̂l }, then π̂ = πv̂ is the corresponding limit point of {πl }. Therefore, by the lower semi-continuity of π → π(r) we have ̺ˆ = lim πl (r) ≥ π̂(r) = ̺v̂ . l→∞ It also holds that (4.16) Lv̂ V̂ (x) + r(x, v̂(x)) = ̺ˆ, a.s. By (4.15), we have lim inf T →∞ 1 v̂ E [V̂ (XT )] = 0, T x and hence applying Itô’s rule on (4.16) we obtain ̺v̂ ≤ ̺ˆ. On the other hand, if v∗ is an optimal stationary Markov control, then by the hypothesis ϕ0 ∈ O(h̃), the fact that πv∗ (h̃) < ∞, (4.14) and [23], Proposition 2.6, we deduce that Evx∗ [V̂ + (Xt )] converges as t → ∞, which of course together with (4.15) implies that 1t Ev̂x [V̂ (Xt )] tends to 0 as t → ∞. Therefore, evaluating (4.13) at v∗ and applying Itô’s rule we obtain ̺v∗ ≥ ̺ˆ. Combining the two estimates, we have ̺v̂ ≤ ̺ˆ ≤ ̺∗ , and thus equality must hold. Here, we have used the fact that there exists an optimal Markov control for r by Theorem 3.4. ERGODIC CONTROL IN THE HALFIN–WHITT REGIME Next, we use the stochastic representation in (4.9) Z τ̆δ l v̂l l (4.17) V (x) = Ex (r(Xs , v̂l (Xs )) − ̺l ) ds + V (Xτ̆δ ) , 0 47 x ∈ Bδc . ̺v Fix any x ∈ Bδc . Since USM0 is compact, it follows that for each δ and R with ̺v 0 < δ < R, the map Fδ,R (v) : USM0 → R+ defined by Z τ̆δ ∧τR v Fδ,R (v) := Ex r(Xs , v(Xs )) ds 0 is continuous. Therefore, the map F̄δ := limRր∞ Fδ,R is lower semi-continuous. It follows that Z τ̆δ Z τ̆δ v̂ v̂l r(Xs , v̂(Xs )) ds ≤ lim Ex (4.18) Ex r(Xs , v̂l (Xs )) ds . l→∞ 0 0 On the other hand, since h̃ is inf-compact, it follows by (4.12) that τ̆δ is uniformly integrable with respect to the measures {Pv̂xl }. Therefore, as also shown in Lemma 3.8, we have lim Ev̂xl [τ̆δ ] = Ev̂x [τ̆δ ]. (4.19) l→∞ Since V l → V̂ , uniformly on compact sets, and ̺l → ̺∗ , as l → ∞, it follows by (4.17)–(4.19) that Z τ̆δ V̂ (x) ≥ Ev̂x (r(Xs , v̂(Xs )) − ̺∗ ) ds + V̂ (Xτ̆δ ) , x ∈ Bδc . 0 Therefore, by Theorem 3.4(b), for any δ > 0 and x ∈ Bδc we obtain Z τ̆δ v̂ V∗ (x) ≤ Ex (r(Xs , v̂(Xs )) − ̺∗ ) ds + V∗ (Xτ̆δ ) 0 ≤ V̂ (x) + Ev̂x [V ∗ (Xτ̆δ )] − Ev̂x [V̂ (Xτ̆δ )], and taking limits as δ ց 0, using the fact that V̂ (0) = V∗ (0) = 0, we obtain V∗ ≤ V̂ on Rd . Since Lv̂ (V∗ − V̂ ) ≥ 0, we must have V∗ = V̂ . By Theorem 4.1(ii), we have V∗ ∈ O(ϕ0 ). Remark 4.1. It can be seen from the proof of Theorem 4.2 that the assumption ϕ0 ∈ O(h̃) can be replaced by the weaker hypothesis that 1 v∗ T Ex [ϕ0 (XT )] → 0 as T → ∞. It is easy to see that if one replaces rl by  1   r(x, u) + f (u), for x ∈ B̄l , l rl (x, u) =   r(x, v0 (x)) + 1 f (v0 (x)), otherwise, l Remark 4.2. 48 A. ARAPOSTATHIS, A. BISWAS AND G. PANG for some positive valued continuous function f , the same conclusion of Theorem 4.2 holds. If we consider the controlled dynamics given by (3.20), with running cost as in (3.11), then there exists a function V ∼ \|x\|m satisfying (3.6). This fact is proved in Proposition 3.1. There also exists a Lyapunov function V0 ∈ O(\|x\|m ), satisfying the assumption in Theorem 4.2, relative to any control v0 with πv0 (h̃) < ∞, where h̃ is selected as in Remark 3.5. Indeed, in order to construct V0 we recall the function ψ in (3.21). Let V0 ∈ C 2 (Rd ) be any function such that V0 = ψ m/2 on the complement of the unit ball centered at the origin. Observe that for some positive constants κ1 and κ2 it holds that κ1 \|x\|2 ≤ ψ(x) ≤ κ2 \|x\|2 . Then a straightforward calculation from (3.22) shows that (3.8) holds with the above choice of V0 . By the stochastic representation of ϕ0 , it follows that ϕ0 ∈ O(V0 ). We have proved the following corollary. Corollary 4.1. For the queueing diffusion model with controlled dynamics given by (3.20), and running cost given by (3.11), there exists a solution (up to an additive constant) to the associated HJB in the class of functions in C 2 (Rd ) whose positive part grows no faster than \|x\|m and whose negative part is in o(\|x\|m ). We conclude this section with the following remark. Remark 4.3. Comparing the approximation technique introduced in this section with that in Section 3, we see that the spatial truncation technique relies on more restrictive assumption on the Lyapunov function V0 and the running cost function (Theorem 4.2). In fact, the growth of h̃ also restricts the growth of r by (3.23). Therefore, the class of ergodic diffusion control problems considered in this section is more restrictive. For example, if the running cost r satisfies (3.11) and h̃ ∼ \|x\|m , then it is not obvious that one can obtain a Lyapunov function V0 with growth at most of order \|x\|m . For instance, if the drift has strictly sub-linear growth, then it is expected that the Lyapunov function should have growth larger than \|x\|m . Therefore, the class of problems considered in Section 3 is larger than those considered in this section. 5. Asymptotic convergence. In this section, we prove that the value of the ergodic control problem corresponding to the multi-class M/M/N + M queueing network asymptotically converges to ̺∗ , the value of the ergodic control for the controlled diffusion. ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 49 Recall the diffusion-scaled processes X̂ n , Q̂n and Ẑ n defined in (2.4), and from (2.5) and (2.6) that Z t Z t n n n n n n X̂i (t) = X̂i (0) + ℓi t − µi Ẑi (s) ds − γi Q̂ni (s) ds (5.1) 0 0 n n n + M̂A,i (t) − M̂S,i (t) − M̂R,i (t), n (t), M̂ n (t) and M̂ n (t), i = 1, . . . , d, as defined in (2.6), are where M̂A,i S,i R,i square integrable martingales w.r.t. the filtration {Ftn } with quadratic variations λni t, n Z µni t n n hM̂S,i i(t) = Z (s) ds, n 0 i Z γin t n n Q (s) ds. hM̂R,i i(t) = n 0 i n hM̂A,i i(t) = 5.1. The lower bound. In this section, we prove Theorem 2.1. Proof of Theorem 2.1. Recall the definition of V̂ n in (2.10), and consider a sequence such that supn V̂ n (X̂ n (0)) < ∞. Let ϕ ∈ C 2 (Rd ) be any function satisfying ϕ(x) := \|x\|m for \|x\| ≥ 1. As defined in Section 1.3, ∆X(t) denotes the jump of the process X at time t. Applying Itô’s formula on ϕ (see, e.g., [24], Theorem 26.7), we obtain from (5.1) that Z t Θn1 (X̂1n (s), Ẑ1n (s))ϕ′ (X̂1n (s)) ds E[ϕ(X̂1n (t))] = E[ϕ(X̂1n (0))] + E 0 +E Z 0 +E t Θn2 (X̂1n (s), Ẑ1n (s))ϕ′′ (X̂1n (s)) ds X s≤t ∆ϕ(X̂1n (s)) − ϕ′ (X̂1n (s−)) · ∆X̂1n (s) 1 ′′ n n n − ϕ (X̂ (s−))∆X̂1 (s)∆X̂1 (s) , 2 where Θn1 (x, z) := ℓn1 − µn1 z − γ1n (x − z), µn1 z + γ1n (x − z) λn1 1 n n √ + µ ρ1 + . Θ2 (x, z) := 2 1 n n 50 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Since {ℓn1 } is a bounded sequence, it is easy to show that for all n there exist positive constants κi , i = 1, 2, independent of n, such that Θn1 (x, z)ϕ′ (x) ≤ κ1 (1 + \|(e · x)+ \|m ) − κ2 \|x\|m , κ2 Θn2 (x, z)ϕ′′ (x) ≤ κ1 (1 + \|(e · x)+ \|m ) + \|x\|m , 4 provided that x − z ≤ (e · x)+ and √zn ≤ 1. We next compute the terms corresponding to the jumps. For that, first we see that the jump size is of order √1n . We can also find a positive constant κ3 such that sup \|ϕ′′ (y)\| ≤ κ3 (1 + \|x\|m−2 ) ∀x ∈ Rd . \|y−x\|≤1 Using Taylor’s approximation, we obtain the inequality ∆ϕ(X̂1n (s)) − ϕ′ (X̂1n (s−)) · ∆X̂1n (s) ≤ 1 sup \|ϕ′′ (y)\|[∆(X̂1n (s))]2 . 2 \|y−X̂ n (s−)\|≤1 1 Hence, combining the above facts we obtain X E ∆ϕ(X̂1n (s)) − ϕ′ (X̂1n (s−)) · ∆X̂1n (s) s≤t (5.2) 1 ′′ n n n − ϕ (X̂1 (s−))∆X̂1 (s)∆X̂1 (s) 2 X ≤E κ3 (1 + \|X̂1n (s−)\|m−2 )(∆(X̂1n (s)))2 s≤t t λn1 µn1 Z1n (s) γ1n Qn1 (s) = κ3 E + + ds n n n 0 Z t κ2 ≤E κ4 + \|X̂1n (s)\|m + κ5 ((e · X̂ n (s))+ )m ds , 4 0 Z (1 + \|X̂1n (s)\|m−2 ) for some suitable positive constants κ4 and κ5 , independent of n, where in the second inequality we use the fact that the optional martingale [X̂1n ] is the sum of the squares of the jumps, and that [X̂1n ] − hX̂1n i is a martingale. Therefore, for some positive constants C1 and C2 it holds that 0 ≤ E[ϕ(X̂1n (t))] (5.3) ≤ E[ϕ(X̂1n (0))] + C1 t − + C2 E Z t 0 κ2 E 2 Z 0 t \|X̂1n (s)\|m ds ((e · X̂ (s)) ) ds . n + m ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 51 By (2.8), we have r(Q̂n (s)) ≥ c1 ((e · X̂ n (s))+ )m , dm which, combined with the assumption that supn V̂ n (X̂ n (0)) < ∞, implies that Z T 1 + m n ((e · X̂ (s)) ) ds < ∞. sup lim sup E n T →∞ T 0 In turn, from (5.3) we obtain sup lim sup n T →∞ 1 E T Z T 0 \|X̂1n (s)\|m ds < ∞. Repeating the same argument for coordinates i = 2, . . . , d, we obtain Z T 1 (5.4) \|X̂ n (s)\|m ds < ∞. sup lim sup E n T →∞ T 0 We introduce the process  n  X̂i (t) − Ẑin (t) , n n + Ui (t) :=  (e · X̂ (t)) ed , i = 1, . . . , d, if (e · X̂ n (t))+ > 0, otherwise. Since Z n is work-conserving, it follows that U n takes values in S, and Uin (t) represents the fraction of class i customers in queue. Define the mean empirical measures Z T 1 n n n ΦT (A × B) := E IA×B (X̂ (s), U (s)) ds T 0 for Borel sets A ⊂ Rd and B ⊂ S. From (5.4), we see that the family {ΦnT : T > 0, n ≥ 1} is tight. Hence, for any sequence Tk → ∞, there exists a subsequence, also denoted by Tk , such that ΦnTk → π n , as k → ∞. It is evident that {π n : n ≥ 1} is tight. Let π n → π along some subsequence, with π ∈ P(Rd × S). Therefore, it is not hard to show that Z n n r̃(x, u)π(dx, du), lim V̂ (X̂ (0)) ≥ n→∞ Rd ×U where, as defined earlier, r̃(x, u) = r((e · x)+ u). To complete the proof of the theorem, we only need to show that π is an ergodic occupation measure for the diffusion. For that, consider f ∈ Cc∞ (Rd ). Recall that [X̂in , X̂jn ] = 0 52 A. ARAPOSTATHIS, A. BISWAS AND G. PANG for i 6= j [30], Lemmas 9.2 and 9.3. Therefore, using Itô’s formula and the definition of ΦnT , we obtain 1 E[f (X̂ n (T ))] T 1 = E[f (X̂ n (0))] T ! Z d X (5.5) Ani (x, u) · fxi (x) + Bin (x, u)fxi xi (x) ΦnT (dx, du) + Rd ×U + i=1 d " X 1 X fxi (X̂ n (s−)) · ∆X̂in (s) E ∆f (X̂ n (s)) − T s≤T i=1 # d 1 X n n n fxi xj (X̂ (s−))∆X̂i (s)∆X̂j (s) , − 2 i,j=1 where Ani (x, u) := ℓni − µni (xi − (e · x)+ ui ) − γin (e · x)+ ui , λn µn xi + (γin − µni )(e · x)+ ui 1 n √ µi ρi + i + i . Bin (x, u) := 2 n n We first bound the last term in (5.5). Using Taylor’s formula, we see that n ∆f (X̂ (s)) − − d X i=1 ∇f (X̂ n (s−)) · ∆X̂ n (s) d 1 X fxi xj (X̂ n (s−))∆X̂in (s)∆X̂jn (s) 2 i,j=1 = d kkf kC 3 X √ \|∆X̂in (s)∆X̂jn (s)\| n i,j=1 for some positive constant k, where we use the fact that the jump size is √1 . Hence, using the fact that independent Poisson processes do not have n simultaneous jumps w.p.1, using the identity Q̂ni = X̂in − Ẑin , we obtain " d X 1 X ∇f (X̂ n (s−)) · ∆X̂ n (s) E ∆f (X̂ n (s)) − T i=1 s≤T (5.6) # d 1 X − fxi xj (X̂ n (s−))∆X̂in (s)∆X̂jn (s) 2 i,j=1 ERGODIC CONTROL IN THE HALFIN–WHITT REGIME kkf k 3 ≤ √C E T n "Z T 0 53 d n X λ i=1 # n Z n (s) n Qn (s) µ γ i + i i + i i ds . n n n Therefore, first letting T → ∞ and using (5.2) and (5.4) we see that the expectation on the right-hand side of (5.6) is bounded above. Therefore, as n → ∞, the left-hand side of (5.6) tends to 0. Thus, by (5.5) and the fact that f is compactly supported, we obtain Z Lu f (x)π(dx, du) = 0, Rd ×U where Lu f (x) = λi ∂ii f (x) + (ℓi − µi (xi − (e · x)+ ui ) − γi (e · x)+ ui )∂i f (x). Therefore, π ∈ G. 5.2. The upper bound. The proof of the upper bound in Theorem 2.2 is a little more involved than that of the lower bound. Generally, it is very helpful if one has uniform stability across n ∈ N (see, e.g., [12]). In [12], uniform stability is obtained from the reflected dynamics with the Skorohod mapping. However, here we establish the asymptotic upper bound by using the technique of spatial truncation that we have introduced in Section 4. Let vδ be any precise continuous control in USSM satisfying vδ (x) = u0 = (0, . . . , 0, 1) for \|x\| > K > 1. First, we construct a work-conserving admissible policy for each n ∈ N (see [7]). Define a measurable map ̟ : {z ∈ Rd+ : e · z ∈ Z} → Zd+ as follows: for z = (z1 , . . . , zd ) ∈ Rd , let ! d X (zi − ⌊zi ⌋) . ̟(z) := ⌊z1 ⌋, . . . , ⌊zd−1 ⌋, ⌊zd ⌋ + i=1 Note that \|̟(z) − z\| ≤ 2d. Define uh (x) := ̟((e · x − n)+ vδ (x̂n )), x ∈ Rd , x1 − ρ1 n xd − ρd n n √ x̂ := ,..., √ , n n n √ o An := x ∈ Rd+ : sup \|xi − ρi n\| ≤ K n . i We define a state-dependent, work-conserving policy as follows:  n X − uh (X n ), if X n ∈ An ,    i ! + i−1 X Zin [X n ] := (5.7) n n  , otherwise. Xj X ∧ n−   i j=1 54 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Therefore, whenever the state of the system is in Acn , the system works under the fixed priority policy with the least priority given to class-d jobs. First, we show that this is a well-defined policy for all large n. It is enough to show that Xin − uh (X n ) ≥ 0 for all i when X n ∈ An . If not, then for some i, 1 ≤ i ≤ d, we must have Xin − uh (X n ) < 0 and so Xin < (e · X n − n)+ + d. Since X n ∈ An , we obtain √ −K n + ρi n ≤ Xin < (e · X n − n)+ + d !+ d X n (Xi − ρi n) = +d i=1 √ ≤ dK n + d. But this cannot hold for large n. Hence, this policy is well defined for all large n. Under the policy defined in (5.7), X n is a Markov process and its generator given by Ln f (x) = d X i=1 + λni (f (x + ei ) − f (x)) + d X i=1 d X i=1 µni Zin [x](f (x − ei ) − f (x)) γin Qni [x](f (x − ei ) − f (x)), x ∈ Zd+ , where Z n [x] is as above and Qn [x] := x − Z n [x]. It is easy to see that, for x∈ / An , ! + #+ " i−1 X n xj . Qi [x] = xi − n − j=1 Lemma 5.1. Let X n be the Markov process corresponding to the above control. Let q be an even positive integer. Then there exists n0 ∈ N such that Z T 1 q n sup lim sup E \|X̂ (s)\| ds < ∞, n≥n0 T →∞ T 0 where X̂ n = (X̂1n , . . . , X̂dn )T is the diffusion-scaled process corresponding to the process X n , as defined in (2.4). Proof. The proof technique is inspired by [6], Lemma 3.1. Define fn (x) := d X i=1 βi (xi − ρi n)q , ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 55 where βi , i = 1, . . . , d, are positive constants to be determined later. We first show that for a suitable choice of βi , i = 1, . . . , d, there exist constants Ci , i = 1, 2, independent of n ≥ n0 , such that Ln fn (x) ≤ C1 nq/2 − C2 fn (x), (5.8) x ∈ Zd+ . Choose n large enough so that the policy is well defined. We define Yin := xi − ρi n. Note that Also, (a ± 1)q − aq = ±qa · aq−2 + O(aq−2 ), µni Zin [x] = µni xi Ln fn (x) = d X i=1 − − ≤ (5.9) i=1 d X i=1 i=1 d X d X i=1 + βi µni xi [qYin \|Yin \|q−2 + O(\|Yin \|q−2 )] βi (γin − µni )Qni [x][qYin \|Yin \|q−2 + O(\|Yin \|q−2 )] βi (λni + µni xi + \|γin − µni \|Qin [x])O(\|Yin \|q−2 ) i=1 ≤ Then a ∈ R. βi λni [qYin \|Yin \|q−2 + O(\|Yin \|q−2 )] d X d X + − µni Qni [x]. βi qYin \|Yin \|q−2 (λni − µni xi − (γin − µni )Qni [x]) βi (λni + (µni + \|γin − µni \|)(Yin + ρi n))O(\|Yin \|q−2 ) d X i=1 βi qYin \|Yin \|q−2 (λni − µni xi − (γin − µni )Qni [x]), where in the last inequality we use the fact that Qni [x] ≤ xi for x ∈ Zd+ . Let √ δin := λni − µni ρi n = O( n). The last estimate is due to the assumptions in (2.1) concerning the parameters in the Halfin–Whitt regime. Then (5.10) d X i=1 βi qYin \|Yin \|q−2 (λni − µni xi − (γin − µni )Qni [x]) = −q d X i=1 βi µni \|Yin \|q + d X i=1 βi qYin \|Yin \|q−2 (δin − (γin − µni )Qni [x]). 56 A. ARAPOSTATHIS, A. BISWAS AND G. PANG If x ∈ An and n is large, then Qni [x] = uh (x) = ̟((e · x − n)+ vδ (x̂n )) √ ≤ (e · x − n)+ + d ≤ 2dK n. Let x ∈ Acn . We use the fact that for any a, b ∈ R it holds that a+ − b+ = ξ[a − b] for some ξ ∈ [0, 1]. Also, !+ #+ " i−1 X nρj = 0, i = 1, . . . , d. nρi − n − j=1 Thus, we obtain maps ξ, ξ̃ : Rd → [0, 1]d such that !+ #+ " i−1 X nρj − Qni [x] −Qni [x] = nρi − n − j=1 = ξi (x)(nρi − xi ) − ξ˜i (x) i−1 X (xj − nρj ), x ∈ Acn . j=1 Hence, from (5.10) we obtain d X i=1 βi qYin \|Yin \|q−2 (λni − µni xi − (γin − µni )Qni [x]) d d X X √ βi ((1 − ξi (x))µni + ξi (x)γin )\|Yin \|q βi \|Yin \|q−1 − q ≤ O( n)q i=1 i=1 +q d X i=1 βi Yin \|Yin \|q−2 δin − (γin − µni )ξ̃i (x) where we used the fact that on An we have !+ #+ " i−1 X √ = O( n) xj xi − n − j=1 i−1 X Yjn j=1 ! , ∀i. Observe that there exists ϑ > 0, independent of n due to (2.1), such that (1 − ξi (x))µni + ξi (x)γin ≥ min(µni , γin ) ≥ ϑ for all n ∈ N, all x ∈ Rd , and all i = 1, . . . , d. As a result, we obtain d X i=1 βi qYin \|Yin \|q−2 (λni − µni xi − (γin − µni )Qni [x]) 57 ERGODIC CONTROL IN THE HALFIN–WHITT REGIME d d X X √ n q−1 βi \|Yin \|q βi \|Yi \| − qϑ ≤ O( n)q (5.11) i=1 i=1 +q d X i=1 βi Yin \|Yin \|q−2 δin − (γin − µni )ξ̃i (x) i−1 X j=1 ! Ynj . We next estimate the last term on the right-hand side of (5.11). Let κ := ϑ . Using Young’s inequality, we obtain the estisupn,i \|γin − µni \|, and ε1 := 8κ mate q i−1 i−1 X 1 X n n q−1 j n q Yj . \|Yi \| Yn ≤ ε1 \|Yi \| + q−1 ε 1 j=1 j=1 Therefore, q d X i=1 βi Yin \|Yin \|q−2 −(γin − µni )ξ̃i (x) ≤ qκ d X ε1 βi \|Yin \|q ≤ qκ d X ε1 βi \|Yin \|q i=1 i=1 + + i−1 βi X ε1q−1 j=1 βi q−1 ε1q−1 i−1 X Yjn j=1 Yjn d ! q! i−1 X \|Yjn \|q j=1 ! ! d i−1 qϑ X βi q−1 X n q n q = βi \|Yi \| + q d \|Yj \| . 8 ε1 i=1 j=1 We choose β1 = 1 and for i ≥ 2, we define βi by εq1 βi := q min βj . d j≤i−1 With this choice of βi it follows from above that q d X i=1 βi Yin \|Yin \|q−2 −(γin − µni )ξ̃i (x) i−1 X j=1 Ynj ! Using the preceding inequality in (5.11), we obtain (5.12) d X i=1 d qϑ X ≤ βi \|Yin \|q . 4 i=1 βi qYin \|Yin \|q−2 (λni − µni xi − (γin − µni )Qni [x]) d d X √ 3 X βi \|Yin \|q . ≤ O( n)q βi \|Yin \|q−1 − qϑ 4 i=1 i=1 58 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Combining (5.9) and (5.12), we obtain Ln fn (x) ≤ (5.13) d X i=1 d X √ O(n)O(\|Yin \|q−2 ) O( n)O(\|Yin \|q−1 ) + i=1 d 3 X βi \|Yin \|q . − qϑ 4 i=1 By Young’s inequality, for any ε > 0, we have the bounds √ √ O( n)O(\|Yin \|q−1 ) ≤ ε[O(\|Yin \|q−1 )]q/(q−1) + ε(1−q) [O( n)]q , O(n)O(\|Yin \|q−2 ) ≤ ε[O(\|Yin \|q−2 )]q/(q−2) + ε(1−q/2) [O(n)]q/2 . Thus, choosing ε properly in (5.13) we obtain (5.8). We proceed to complete the proof of the lemma by applying (5.8). First, we observe that E[sups∈[0,T ] \|X n (s)\|p ] is finite for any p ≥ 1 as this quantity is dominated by the Poisson arrival process. Therefore, from (5.8) we see that Z T n n n Ln fn (X (s)) ds E[fn (X (T ))] − fn (X (0)) = E 0 ≤ C1 n q/2 T − C2 E Z T 0 fn (X (s)) ds , n which implies that # "Z d d T X X βi (X̂in (0))q . C2 E βi (X̂in (s))q ds ≤ C1 T + 0 i=1 i=1 Hence, the proof follows by dividing both sides by T and letting T → ∞. Proof of Theorem 2.2. Let r be the given running cost with polynomial growth with exponent m in (2.8). Let q = 2(m + 1). Recall that r̃(x, u) = r((e · x)+ u) for (x, u) ∈ Rd × S. Then r̃ is convex in u and satisfies (3.11) with the same exponent m. For any δ > 0, we choose vδ ∈ USSM such that vδ is a continuous precise control with invariant probability measure µδ and Z (5.14) r̃(x, vδ (x))µδ (dx) ≤ ̺∗ + δ. Rd We also want the control vδ to have the property that vδ (x) = (0, . . . , 0, 1) outside a large ball. To obtain such vδ , we see that by Theorems 4.1, 4.2 and ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 59 Remark 4.2 we can find vδ′ and a ball Bl for l large, such that vδ′ ∈ USSM , vδ′ (x) = ed for \|x\| > l, vδ′ is continuous in Bl , and Z δ r̃(x, vδ′ (x))µ′δ (dx) − ̺∗ < , 2 d R where µ′δ is the invariant probability measure corresponding to vδ′ . We note that vδ′ might not be continuous on ∂Bl . Let {χn : n ∈ N} be a sequence c of cut-off functions such that χn ∈ [0, 1], it vanishes on Bl−(1/n) , and it n n takes the value 1 on Bl−(2/n) . Define the sequence vδ (x) := χ (x)vδ′ (x) + (1 − χn (x))ed . Then vδn → vδ′ , as n → ∞, and the convergence is uniform on the complement of any neighborhood of ∂Bl . Also by Proposition 3.1 the corresponding invariant probability measures µnδ are exponentially tight. Thus, Z Z ′ ′ r̃(x, vδn (x))µnδ (dx) −→ 0. r̃(x, vδ (x))µδ (dx) − Rd Rd n→∞ Combining the above two expressions, we can easily find vδ which satisfies (5.14). We construct a scheduling policy as in Lemma 5.1. By Lemma 5.1, we see that for some constant K1 it holds that Z T 1 q n (5.15) \|X̂ (s)\| ds < K1 , q = 2(m + 1). sup lim sup E n≥n0 T →∞ T 0 Define vh (x) := ̟((e · x − n)+ vδ (x̂n )), √ v̂h (x̂n ) := ̟( n(e · x̂n )+ vδ (x̂n )). Since vδ (x̂n ) = (0, . . . , 0, 1) when \|x̂n \| ≥ K, it follows that Qn [X n ] = X n − Z n [X n ] = vh (X n ) P n for large n, provided that d−1 i=1 Xi ≤ n. Define ( d−1 ) X √ x̂ni > ρd n . Dn := x : i=1 Then 1 n r(Q̂ (t)) = r √ v̂h (X̂ (t)) + r(X̂ n (t) − Ẑ n (t))I{X̂ n (t)∈Dn } n 1 n − r √ v̂h (X̂ (t)) I{X̂ n (t)∈Dn } . n n 60 A. ARAPOSTATHIS, A. BISWAS AND G. PANG Define, for each n, the mean empirical measure ΨnT by Z T 1 IA (X̂ n (t)) dt . ΨnT (A) := E T 0 By (5.15), the family {ΨnT : T > 0, n ≥ 1} is tight. We next show that Z Z T 1 n (5.16) lim lim sup E r((e · x)+ vδ (x))µδ (dx). r(Q̂ (t)) dt = n→∞ T →∞ T Rd 0 For each n, select a sequence {Tkn : k ∈ N} along which the “lim sup” in (5.16) is attained. By tightness, there exists a limit point Ψn of ΨnT n . Since Ψn has k support on a discrete lattice, we have Z Z 1 1 n r √ v̂h (x) Ψn (dx). r √ v̂h (x) ΨT n (dx) −→ k k→∞ d n n d R R Therefore, 1 lim sup E T →∞ T Z T 0 1 r √ v̂h (x) Ψn (dx) ± E n , r(Q̂ (t)) dt ≶ n d R n Z where E n = lim sup T →∞ 1 E T Z 0 T 1 r(Q̂n (t)) + r √ v̂h (X̂ n (t)) I{X̂ n (t)∈Dn } dt . n By (5.15), the family {Ψn : n ≥ 1} is tight. Hence, it has a limit Ψ. By definition, we have 1 2d √ v̂h (x) − (e · x)+ vδ (x) ≤ √ . n n Thus, using the continuity property of r and (2.8) it follows that Z Z 1 n r √ v̂h (x) Ψ (dx) −→ r((e · x)+ vδ (x))Ψ(dx), n→∞ n d d R R along some subsequence. Therefore, in order to complete the proof of (5.16) we need to show that lim sup E n = 0. n→∞ Since the policies are work-conserving, we observe that 0 ≤ X̂ n − Ẑ n ≤ (e · X̂ n )+ , and therefore for some positive constants κ1 and κ2 , we have 1 n r √ v̂h (X̂ (t)) ∨ r(X̂ n (t) − Ẑ n (t)) ≤ κ1 + κ2 [(e · X̂ n )+ ]m . n ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 61 Given ε > 0 we can choose n1 so that for all n ≥ n1 , Z T 1 + m n √ √ [(e · X̂ (s)) ] I{\|X̂ n (s)\|>(ρd / d) n} ds ≤ ε, lim sup E T →∞ T 0 p where we use (5.15). We observe that Dn ⊂ {\|x̂n \| > ρd d/n}. Thus, (5.16) holds. In order to complete the proof, we only need to show that Ψ is the invariant probability measure corresponding to vδ . This can be shown using the convergence of generators as in the proof of Theorem 2.1. 6. Conclusion. We have answered some of the most interesting questions for the ergodic control problem of the Markovian multi-class many-server queueing model. This current study has raised some more questions for future research. One of the interesting questions is to consider nonpreemptive policies and try to establish asymptotic optimality in the class of nonpreemptive admissible polices [7]. It will also be interesting to study a similar control problem when the system has multiple heterogeneous agent pools with skill-based routing. It has been observed that customers’ service requirements and patience times are nonexponential [10] in some situations. It is therefore important and interesting to address similar control problems under general assumptions on the service and patience time distributions. Acknowledgements. We thank the anonymous referee for many helpful comments that have led to significant improvements in our paper. Ari Arapostathis acknowledges the hospitality the Department of Industrial and Manufacturing Engineering in Penn State while he was visiting at the early stages of this work. Guodong Pang acknowledges the hospitality of the Department of Electrical and Computer Engineering at University of Texas at Austin while he was visiting for this work. Part of this work was done while Anup Biswas was visiting the Department of Industrial and Manufacturing Engineering in Penn State. Hospitality of the department is acknowledged. REFERENCES [1] Arapostathis, A., Borkar, V. S. and Ghosh, M. K. (2012). Ergodic Control of Diffusion Processes. Encyclopedia of Mathematics and Its Applications 143. Cambridge Univ. Press, Cambridge. MR2884272 [2] Arisawa, M. and Lions, P.-L. (1998). On ergodic stochastic control. Comm. Partial Differential Equations 23 2187–2217. MR1662180 [3] Atar, R. (2005). Scheduling control for queueing systems with many servers: Asymptotic optimality in heavy traffic. Ann. Appl. Probab. 15 2606–2650. MR2187306 [4] Atar, R. and Biswas, A. (2014). Control of the multiclass G/G/1 queue in the moderate deviation regime. Ann. Appl. Probab. 24 2033–2069. MR3226171 [5] Atar, R., Giat, C. and Shimkin, N. (2010). The cµ/θ rule for many-server queues with abandonment. Oper. Res. 58 1427–1439. MR2560545 62 A. ARAPOSTATHIS, A. BISWAS AND G. PANG [6] Atar, R., Giat, C. and Shimkin, N. (2011). On the asymptotic optimality of the cµ/θ rule under ergodic cost. Queueing Syst. 67 127–144. MR2771197 [7] Atar, R., Mandelbaum, A. and Reiman, M. I. (2004). Scheduling a multi class queue with many exponential servers: Asymptotic optimality in heavy traffic. Ann. Appl. Probab. 14 1084–1134. MR2071417 [8] Bogachev, V. I., Krylov, N. V. and Röckner, M. (2001). On regularity of transition probabilities and invariant measures of singular diffusions under minimal conditions. Comm. Partial Differential Equations 26 2037–2080. MR1876411 [9] Borkar, V. S. (1989). Optimal Control of Diffusion Processes. Pitman Research Notes in Mathematics Series 203. Longman, Harlow. MR1005532 [10] Brown, L., Gans, N., Mandelbaum, A., Sakov, A., Shen, H., Zeltyn, S. and Zhao, L. (2005). Statistical analysis of a telephone call center: A queueingscience perspective. J. Amer. Statist. Assoc. 100 36–50. MR2166068 [11] Budhiraja, A., Ghosh, A. and Liu, X. (2014). Scheduling control for Markovmodulated single-server multiclass queueing systems in heavy traffic. Queueing Syst. 78 57–97. MR3238008 [12] Budhiraja, A., Ghosh, A. P. and Lee, C. (2011). Ergodic rate control problem for single class queueing networks. SIAM J. Control Optim. 49 1570–1606. MR2817491 [13] Dai, J. G. and Tezcan, T. (2008). Optimal control of parallel server systems with many servers in heavy traffic. Queueing Syst. 59 95–134. MR2430811 [14] Dieker, A. B. and Gao, X. (2013). Positive recurrence of piecewise Ornstein– Uhlenbeck processes and common quadratic Lyapunov functions. Ann. Appl. Probab. 23 1291–1317. MR3098433 [15] Gamarnik, D. and Stolyar, A. L. (2012). Multiclass multiserver queueing system in the Halfin-Whitt heavy traffic regime: Asymptotics of the stationary distribution. Queueing Syst. 71 25–51. MR2925789 [16] Gamarnik, D. and Zeevi, A. (2006). Validity of heavy traffic steady-state approximation in generalized Jackson networks. Ann. Appl. Probab. 16 56–90. MR2209336 [17] Garnett, O., Mandelbaum, A. and Reiman, M. I. (2002). Designing a call center with impatient customers. Manuf. Serv. Oper. Manag. 4 208–227. [18] Gilbarg, D. and Trudinger, N. S. (1983). Elliptic Partial Differential Equations of Second Order, 2nd ed. Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences] 224. Springer, Berlin. MR0737190 [19] Gurvich, I. (2014). Diffusion models and steady-state approximations for exponentially ergodic Markovian queues. Ann. Appl. Probab. 24 2527–2559. MR3262510 [20] Gyöngy, I. and Krylov, N. (1996). Existence of strong solutions for Itô’s stochastic equations via approximations. Probab. Theory Related Fields 105 143–158. MR1392450 [21] Halfin, S. and Whitt, W. (1981). Heavy-traffic limits for queues with many exponential servers. Oper. Res. 29 567–588. MR0629195 [22] Harrison, J. M. and Zeevi, A. (2004). Dynamic scheduling of a multiclass queue in the Halfin–Whitt heavy traffic regime. Oper. Res. 52 243–257. MR2066399 [23] Ichihara, N. and Sheu, S.-J. (2013). Large time behavior of solutions of Hamilton– Jacobi–Bellman equations with quadratic nonlinearity in gradients. SIAM J. Math. Anal. 45 279–306. MR3032978 [24] Kallenberg, O. (2002). Foundations of Modern Probability, 2nd ed. Springer, New York. MR1876169 ERGODIC CONTROL IN THE HALFIN–WHITT REGIME 63 [25] Kim, J. and Ward, A. R. (2013). Dynamic scheduling of a GI/GI/1 + GI queue with multiple customer classes. Queueing Syst. 75 339–384. MR3110643 [26] Koçağa, Y. L. and Ward, A. R. (2010). Admission control for a multi-server queue with abandonment. Queueing Syst. 65 275–323. MR2652045 [27] Krylov, N. V. (1980). Controlled Diffusion Processes. Applications of Mathematics 14. Springer, New York. Translated from the Russian by A. B. Aries. MR0601776 [28] Mandelbaum, A. and Stolyar, A. L. (2004). Scheduling flexible servers with convex delay costs: Heavy-traffic optimality of the generalized cµ-rule. Oper. Res. 52 836–855. MR2104141 [29] Ocone, D. and Weerasinghe, A. (2003). Degenerate variance control in the onedimensional stationary case. Electron. J. Probab. 8 no. 24, 27 pp. (electronic). MR2041825 [30] Pang, G., Talreja, R. and Whitt, W. (2007). Martingale proofs of many-server heavy-traffic limits for Markovian queues. Probab. Surv. 4 193–267. MR2368951 [31] Stannat, W. (1999). (Nonsymmetric) Dirichlet operators on L1 : Existence, uniqueness and associated Markov processes. Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 28 99–140. MR1679079 [32] van Mieghem, J. A. (1995). Dynamic scheduling with convex delay costs: The generalized cµ rule. Ann. Appl. Probab. 5 809–833. MR1359830 [33] Yosida, K. (1980). Functional Analysis, 6th ed. Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences] 123. Springer, Berlin. MR0617913 A. Arapostathis A. Biswas Department of Electrical and Computer Engineering The University of Texas at Austin 1616 Guadalupe St., UTA 7.508 Austin, Texas 78701 USA E-mail: ari@ece.utexas.edu anupbiswas@utexas.edu G. Pang The Harold and Inge Marcus Department of Industrial and Manufacturing Engineering College of Engineering Pennsylvania State University University Park, Pennsylvania 16802 USA E-mail: gup3@psu.edu	3cs.SY
1 Role Playing Learning for Socially Concomitant Mobile Robot Navigation arXiv:1705.10092v1 [cs.RO] 29 May 2017 Mingming Li, Rui Jiang, Shuzhi Sam Ge, Fellow, IEEE, and Tong Heng Lee, Member, IEEE Abstract—In this paper, we present the Role Playing Learning (RPL) scheme for a mobile robot to navigate socially with its human companion in populated environments. Neural networks (NN) are constructed to parameterize a stochastic policy that directly maps sensory data collected by the robot to its velocity outputs, while respecting a set of social norms. An efficient simulative learning environment is built with maps and pedestrians trajectories collected from a number of real-world crowd data sets. In each learning iteration, a robot equipped with the NN policy is created virtually in the learning environment to play itself as a companied pedestrian and navigate towards a goal in a socially concomitant manner. Thus, we call this process Role Playing Learning, which is formulated under a reinforcement learning (RL) framework. The NN policy is optimized end-toend using Trust Region Policy Optimization (TRPO), with consideration of the imperfectness of robot’s sensor measurements. Simulative and experimental results are provided to demonstrate the efficacy and superiority of our method. l Index Terms—Socially concomitant navigation, mobile robot, neural network, reinforcement learning I. I NTRODUCTION The capability to navigate in densely populated and dynamic environments is one of the most important features that enable the deployment of mobile robots in unstructured environment, such as schools, shopping malls and transportation hubs. The key difference between the problem of navigating among humans and the traditional path planning and obstacle avoidance problems is that humans tend to smoothly evade each other interactively and cooperatively, rather than remaining static or maintaining an indifferent trajectory dynamics. In other words, there are social norms that need to be understood and complied to achieve maximum comfort of all involved pedestrians during navigation. We refer to this as the problem of social navigation, which aims to model such social norms and develop a robotic navigation policy that is socially acceptable to the pedestrians around. For social navigation, the traditional approaches based on Dynamic Window Approach (DWA) [1] or potential fields [2], [3] are usually of limited efficacy as pedestrians are simply regarded as uncooperative obstacles. An illustrative example is the freezing robot problem (FRP) [4], [5], where a mobile robot will be stuck in a narrow corridor when facing a crowd of people if it lacks the ability to predict the joint collision avoidance behaviors of human pedestrians. To this M. Li, R. Jiang, S. S. Ge and T. H. Lee are with the Department of Electrical and Computer Engineering, and the Social Robotics Lab, Smart System Institute (SSI), National University of Singapore, Singapore 117576 (e-mail: li mingming@u.nus.edu; rui jiang@u.nus.edu; samge@nus.edu.sg; eleleeth@nus.edu.sg). end, researches have been done to understand the principles of humans’ joint collision avoidance strategies and one of the pioneering works are the social force model (SFM) [6], [7]. Other joint collision avoidance model such as reciprocal velocity obstacles (RVO) have been proposed in [8], [9], [10], with an underlying assumption that all involved agents adopt the same collision avoidance strategies. These ideas are also applied to visual tracking of pedestrians [11], [12]. More recently, several attempts are made to learn probabilistic models of pedestrians’ trajectories during joint collision avoidance, based on which the robot’s navigation decision is generated such that it is able to behave naturally and correctly in similar situations [13], [5], [14], [15]. In this paper, we propose to augment the dimensions of human-robot interaction in social navigation by further endowing robot with appropriate group behaviors when it is travelling with a human companion. This capability is highly desirable for assistive mobile robots [16], [17], [18], which serve as assistants and companions and are expected to travel along with theirx human partners in not only home environment but also possibly crowded public areas. In other words, apart from understanding the collision avoidance behaviors of pedestrians, the robot also needs to consider the motion of its companion so as to maintain a sense of affinity when they are travelling together towards a certain goal. We call this socially concomitant navigation (SCN) and it is more challenging than the aforementioned social navigation problem, where the robot is assumed to travel alone with a simpler pursuit of reaching a specific goal while being free of collision. To address the problem of SCN, we develop a new learning scheme called Role Playing Learning (RPL). Particularly, we formulate such problem under the framework of Partially Observable Markov Decision Process (POMDP) and reinforcement learning (RL). A neural network (NN) is used to parameterize the navigation policy of the robot, which is optimized to gives proper steering commands for the next time instance based on the robot’s current and previous observations to its surroundings. To facilitate the RL process, we create a simulative navigation environment by mirroring a collections of real world pedestrians data sets and develop an on-policy optimization method called Partially Observable Trust Region Policy Optimization (PO-TRPO). In each run in an optimization iteration, the robot will attempt to play itself as a companion of a randomly chosen pedestrian by executing the NN navigation policy. The NN policy is then optimized using PO-TRPO based on a batch of collected trajectories. Compared to the existing analytically derived or data-driven approaches, our RPL scheme has the following advantages: 2 1) RPL scheme is less restrictive. It does not rely on the assumption that the robot and other agents (pedestrians) share the same decision-making models [8], [9], [10], [5], [14] or that the navigation goals of pedestrians are known [5], [14]. 2) The formulation of RPL scheme is more generalizable and flexible. Our formulation contain no manuallydefined feature and domain knowledge (e.g., statistics of pedestrians’ behaviors). It is not hardware-specific and can be easily modified to incorporate kinematics of different mobile robot platforms, sensor specifications and navigation objectives. In addition, unlike [15], [14], the learned navigation policy operates without assess to the global map of the environment. Therefore, it is not environment-specific and is well generalizable to unmet real-world scenarios. 3) We explicitly consider the noise and limitation of the robot’s sensor measurements. Most approaches for social navigation assume that the robot has full and accurate knowledge of interested variables, such as positions or distance of pedestrians and obstacles [8], [9], [10], [14]. On the contrary, our RPL schemes is rooted from the situation where the robot can only perceive those lie within its sensor’s Field of View (FoV), with the existence of measurement noise. 4) As a RL-based approach, RPL is efficient. Although RPL aims at solving tasks that involve interaction among robot, humans and physical environment, it does not require participation of human in both data collection and learning, which is known to be tedious and timeconsuming. Instead, the learning process is safely automated in a simulative yet realistic environment with no human intervention. We evaluate the performance of our approach in both simulations and real-world experiments, by comparing it with a baseline planner based on RVO [8] and humans, repectively. We also show that, with some tricks, the learned navigation policy can still be effective when the navigation scenario is reduced to the aforementioned social navigation, which means the robot is travelling without human companion. The remainder of this paper is organized as follows. Related work is first reviewed in Section II. In Section III, the problem of SCN is formulated as a POMDP and associated definitions are given. RPL scheme and PO-TRPO algorithm are described in Section IV. Sections V and VI provide extensive results of simulation and experiment, followed by some concluding remarks in Section VII. II. R ELATED W ORK The problems of robot navigation in populated and dynamic environment can be addressed from a number of angles, which can be largely classified into two groups as in the following subsections. A. Interactive Behaviors Models Many researches have been proposed to describe the interactive navigation behaviors of humans by fitting a computational model to the observed pedestrians trajectories [19]. In this way, the robot’s path planner is able to understand pedestrians’ intention during joint collision avoidance and actively calculate an optimal route towards its goal. In the field of robotics, a majority of work in this direction is done via inverse reinforcement learning (IRL) [20], which learns a cost function that explains the observed behaviors. For example, maximum entropy IRL [21] is adopted in a number works [22], [23], [24], [25], [26] for discrete human behavior prediction and route planning. However, discrete representation is less desirable when modeling trajectories, which are in nature continuous and has higher order dynamics, such as velocities and acceleration. Instead, [15] adopts Maximum-A-Posteriori Bayesian IRL [27] to learn appropriate navigation behavior of a specific mobile robot from a set of demonstration trajectories. Note that, the demonstration data in [15] is specific to configurations of the robot and its sensor and has to be collected via human operation, which could be time-consuming. On the other hand, [13], [14] learns probabilistic models of composite trajectories of pedestrians from video data by maximum entropy learning and IRL. To better capture the characteristics of observed trajectories, they propose to develop their models based on a set of features that are hand-crafted according to the domain knowledge from psychological studies. In addition, those features contain velocities and accelerations of pedestrians, which, in practice, are hard to precisely measure. Besides, interacting Gaussian process (IGP) is derived in [5] to model the joint trajectories of pedestrian while explicitly considering the effects of observation noise. Nevertheless, the design of IGP also requires several hand-crafted kernels that are formulated based on the priori information in a specific application scenario. Other than researchers in robotics, the community of computer vision also possess great interest in pedestrian modeling. One of the important topics is trajectory prediction in video space. In [11], Linear Trajectory Avoidance (LTA) is developed as a dynamic model for pedestrians in video space for shortterm trajectory prediction and it is integrated into visual tracking system. Gaussian process is adopted in [28] to learn the motion pattern of pedestrians. Recently, Social LSTM is proposed in [29] for human trajectory prediction in crowd space. Similarly, the feature of social sensitivity is developed in [30] to analyze trajectories of pedestrians and bicyclists. While the above methods can effectively predict the navigation intention of pedestrians in videos, it is still unclear how to apply these model to navigation of robot in real scenarios. B. Steering Models In contrast to learning behavior models of pedestrians, a more direct perspective is to develop a steering model that outputs the immediate navigation actions given the robot’s current observation to the environment. One of the pioneering work in this direction is the social force model (SFM) [6], which uses energy/potential functions to encode the social status of pedestrian. Then, the navigation motivation of a pedestrian can be derived by taking the gradients of these energy functions. Following this idea, subsequent work [31], [32], [33], [12] 3 propose to infer the optimal parameters of the energy function by fitting them to video data. However, they are likely to produce suboptimal results if the demonstration data from humans are imperfect. In [34], the authors integrate a people tracker and an iterative A∗ planner, with which the robot actively follows the pedestrian travelling in a similar direction to navigate through crowded environment. [35] follow the same idea and formulate the choice of a pedestrian to follow as a Multi-Policy Decision Making process. On the other hands, [36] develops a hierarchical POMDP for predictive navigation in dynamic environment. The idea is to predict the motion of pedestrians and generate a environment-specific cost map for path planning and obstacle avoidance. Other than navigating in a pedestrian-aware manner, several reactive collision avoidance techniques have also been developed, such as DWA [1], [37], velocity obstacles [38] and reciprocal velocity obstacles (RVO) [8], [9], [10]. The common idea of these methods is to treat pedestrians as moving obstacles and reactively update the planner every short periods to achieve collision avoidance. As mentioned in Section I, these methods are less effective for social navigation as they lack predictive abilities and are based on some restrictive assumptions, such as accurate knowledge of moving agents’ velocities [37] and that all agents adopt the identical collision avoidance strategy [8], [9], [10]. Our proposed navigation policy belongs to the steering models. It takes an observation vector as input and outputs the navigation action through a stochastic neural networks. During RPL, our policy is optimized by the PO-TRPO algorithm, which is derived based on the recent advances in deep reinforcement learning (DRL) [39], [40]. DRL exploits the massive representation power of deep neural networks (DNN) [41] to build a complex yet sophisticated decision model, with which an agent can directly learn from raw signals instead of carefully crafted feature and tends to act more intelligently. Recently, there are several attempts in using DNN and DRL for robot navigation. For example, an end-to-end motion planner is learned in [42] to map raw sensor data of a laser range finder onto steering commands of a mobile robot. In [43], a decentralized multi-agent collision avoidance policy is learned via DRL, which can be thought as a DRL version of the original RVO approach [8]. Finally, a target-driven visual navigation policy for home environment is learned in [44] via DRL. They create a set 3D virtual home environments for effective and efficient training of the agent, which shares a similar idea with our proposed RPL scheme. III. P ROBLEM F ORMULATION To formulate the problem of socially concomitant navigation, we gives the following rules of SCN: 1) The robot should reach its goal as fast as possible; 2) The robot should not collide with any of the pedestrians or its companion, or run into any obstacle; 3) The robot should not run too far away from its companion. The above rules serve as a generic description of the robot’s desired performance during navigation. To give concrete definitions, consider the navigation process as an infinite-horizon discounted POMDP in discrete time, defined by the tuple (S, A, F, O, p0 , r, γ). S is a finite set of states s reflecting the navigation status of the robot. A is a finite set of actions a. In this paper, it is defined as a twosome of the translational and rotational velocities of a synchro-drive mobile robot, i.e., a = [vT , vR ]. F : S × A → S is state-transition mapping, which is characterized by the dynamics of the robot, the other humans and the environment. Without loss of generality, we assume deterministic state transition, i.e., si+1 = F (si , ai ), where si , ai are the state and action taken at time ti . O is the set of the robot’s observation o to the state s and β(o\|s) denotes the conditional observation probability distribution. Note that, in practice, the robot’s observation has only incomplete access to s or is subject to certain measurement noise, which implies o 6= s. p0 : S → R is the initial state distribution, i.e., s0 ∼ p0 . r : S → R is a scalar reward given to the robot and γ ∈ (0, 1] is the reward discount factor. Robot motion dynamics: In this paper, synchro-drive mobile robots are considered, whose motion equation can be approximated by assuming the robot’s velocities to be constant within a certain short time period [ti , ti+1 ] [1] with length ∆t = ti+1 − ti . Particularly, let φr (ti ) and ρr (ti ) = [xr (ti ), yr (ti )] denote the robot’s heading and its positions in a 2D Cartesian space at time ti , respectively. vT (ti ) ∈ [0, v̄T ] and vR (ti ) ∈ [−v̄R , v̄R ] represent the robot’s translational and rotational velocities. Define ∆xr = xr (ti+1 ) − xr (ti ) and ∆yr = yr (ti+1 ) − yr (ti ). When the robot has nonzero rotational velocity, i.e., vR (ti ) 6= 0, we have vT (ti )(sin φr (ti ) − sin(φr (ti ) + vR (ti )∆t)) (1) vR (ti ) vT (ti )(cos φr (ti ) − cos(φr (ti ) + vR (ti )∆t)) ∆yr = (2) vR (ti ) ∆xr = − Otherwise, when vR (ti ) = 0, ∆xr = vT (ti ) cos φr (ti ) (3) ∆yr = vT (ti ) sin φr (ti ) (4) With the above formulations, our goal is optimizing a stochastic navigation policy Pθ : O × A → [0, 1] with parameters θ in order to maximize the expected discounted reward: η(Pθ ) = Eτ [ ∞ X γ i r(si , ai )] (5) i=0 where τ = (s0 , o0 , a0 , s1 , o1 , a1 , · · · ) denotes the whole trajectory and ai ∼ Pθ (ai \|oi ). The specific definitions of the above ingredients for SCN will be elaborated as follows: State: Given ρr and φr , define the distance d and direction φ of a point ρ = [x, y] to the robot as follows: p d(ρ) = (x − xr )2 + (y − yr )2 (6) y − yr ) − φr (7) φ(ρ) = arctan( x − xr Then, the robot’s distance to the goal located at ρg = [xg , yg ] are computed as dg = d(ρg ) and φg = φ(ρg ) denotes the offset angle between the robot’s current heading φr and its goal. Similarly, we can define the twosomes (djped , φjped ), 4 (djcom , φjcom ) or (djobs , φjobs ) to describe the relative position of a pedestrians ρjped , a companion ρjcom or an obstacle ρjobs to the robot. With such definitions, the state s is defined to incorporate the information related to the robot’s navigation status as follows: s = [dg , φg , a, pped , pcom , pobs ] (8) where a is the current action vector and n n ped ped ] , φped pped = [d1ped , φ1ped , · · · , dped (9) pcom = [dcom , φcom ] (10) L− L− R− R− L+ L+ R+ R+ pobs = [dF obs , dobs , φobs , dobs , φobs , dobs , φobs , dobs , φobs ](11) The vector pped includes the distances and directions of nped closest pedestrians while pcom includes those of the robot’s companion. The vector pobs is a compact description to the robot’s perception of the surrounding environment. Particularly, the boundaries of the occupied space (obstacles) in the environment are represented as a finite point set Z = {ρ1obs , ρ2obs , · · · , ρjobs , · · · }. Then, the 9 variables in pobs are defined based on the following assumption Assumption 1: An obstacle ρobs ∈ Z has no effect on the robot’s navigation decision if it satisfies d(ρobs ) > d¯obs , where d¯obs is a predefined finite constant. By Assumption 1, it is sufficient to consider only obstacles in Z that are closed enough to the robot, whose distances are less than d¯obs . In practice, this limit may correspond to the robot’s perception range. Let Z̄ = {ρ\|ρ ∈ Z, dρ ≤ d¯obs } (12) The components in vector pobs are described as follows: The distance to the nearest obstacle located at heading of the robot, i.e., dF obs = min ρ∈Z̄ and \|φ(ρ)\|≤ǫρ d(ρ) (13) where ǫρ is a small constant. L R R For dL obs , φobs , dobs , φobs , they represent the distance and direction of the closest and farthest obstacles on the robot’s R left (ρL obs ) and right side (ρobs ), respectively, which are defined mathematically as follows: − ρL obs = arg min d(ρ) (14) ρ∈Z̄ and φ(ρ)>ǫρ − ρR obs = arg min d(ρ) Fig. 1: Illustration of the state variables in Eq. (8). The blue, yellow and green circles represent the robot, its companion (Com.) and the pedestrians (Ped.) respectively. The red dashed circle with a radii d¯obs represents the boundary of the set Z̄ in Eq. 12. The black arrow shows the current heading of the robot. Considering the robot’s current position as the origin, the polar coordinates of the pedestrians, the companion, the closest( and the farthest) obstacles in each direction are compactly represented as vectors pped , pcom , pobs . account in order to develop a robust and practical navigation system. To this end, we define o as the robot’s observation to the true state s as follows: o = [dg , φg , a, p̂ped , p̂com , p̂obs ] By Eq. (18), we assume that the robot has accurate information about the goal position and its current taken action (i.e., the velocity commands output to the robot’s motor) while its observations to pped , pcom , pobs may be imperfect. Particularly, consider the Field of Views (FoVs) for the robot’s pedestrian and obstacle detectors illustrated as Fig. 2. Mathematically, let finite point sets Fped and Fobs denote the current FoVs of pedestrian and obstacle detectors, charac− + + − + terized by threesomes (φ+ ped , φped , dped ) and (φobs , φobs , dobs ), respectively. The robot’s observations to the pedestrians’ relative positions are obtained as (15) nped nped ] , φ̂ped p̂ped = [dˆ1ped , φ̂1ped , · · · , dˆped ρ∈Z̄ and φ(ρ)<−ǫρ + ρL obs = arg max d(ρ) (16) = arg max d(ρ) dˆjped (17) ρ∈Z̄ and φ(ρ)<−ǫρ Then, the variables in pobs can be simply determined as the distance and directions of the above points according to Eqs. (6) and (7). Figure. 1 provides a comprehensive illustration of the state variables pped , pcom and pobs . Observation: As discussed in the previous sections, sensors mounted on the robot are always subject to various kinds of limitation and measurement noise, which must be taken into (19) where ρ∈Z̄ and φ(ρ)>ǫρ + ρR obs (18) = ( djped + d˜ped , if ρjped ∈ Fped d+ else ped , (20) φjped , if ρjped ∈ Fped π, if else (21) and φ̂jped = for j = 1, · · · , nped , with d˜ped being the measurement noise/error. 5 Fig. 2: Field of Views of the pedestrians (green) and obstacles (blue) detectors. The arrow (φ = 0) points towards the + current heading of the robot. The constants φ+ ped , φobs and − − φped , φobs denote the maximum and minimum offset angles + in the corresponding FoVs. Finally, d+ ped and dobs represent the maximum detection ranges for the pedestrian and obstacle detectors, respectively. The values of these constants should be determined according the specific configurations of the robot’s sensor and the corresponding detection algorithms. Any pedestrian/obstacle outside the FoVs is not observable and therefore will be omitted. sensors, such as range sensors [45], [46], RGB-D [47], Timeof-Flight (ToF) [48] and omnidirectional cameras [49], as long as the interested positions can be extracted/estimated from the sensor’s raw measurements. Remark 3: The mathematical definitions of the variables in observations p̂ped , p̂com , p̂obs are given for better understanding and are required only in the simulative RPL process. In practice, it is clear that these values can be directly measured via the robot’s onboard sensors without accessing the actual 2-D Cartesian coordinates [x, y] of the considered point sets (e.g., Z, Fped and Fobs ). For example, consider a robot equipped with a laser range finder. These distances and offset angles can be easily obtained from the returned ranges array[50]. Reward function: A scalar reward will be given to the robot as an award of reaching the goal or a penalty of colliding with obstacles/pedestrians/companions or losing its companions. Particularly, at time ti , the process of SCN will be terminated if any of the following three termination conditions is true. 1) Goal Reaching Condition dg (ti ) ≤ 0.8 (29) 2) Collision Conditions Similarly, define p̂obs = ˆL− L− ˆR− R− ˆL+ L+ ˆR+ R+ [dˆF obs , dobs , φ̂obs , dobs , φ̂obs , dobs , φ̂obs , dobs , φ̂obs ] (22) Compared to the states in (13) to (17), only the obstacles within Fobs are observable. Thus, dˆF obs is formulated as dˆF obs = min ρ∈Z̄∩Fobs and \|φ(ρ)\|≤ǫρ d(ρ) + d˜obs (23) where d˜obs is the measurement noise/error for obstacle detection. The closest observed obstacles on the robot’s left and right sides are defined in a similar way as: − ρ̂L obs = arg min d(ρ) + d˜obs (24) ρ∈Z̄∩Fobs and φ(ρ)>ǫρ − ρ̂R obs = d(ρ) + d˜obs arg min (25) ρ∈Z̄∩Fobs and φ(ρ)<−ǫρ + ρ̂L obs = arg max d(ρ) + d˜obs (26) ρ∈Z̄∩Fobs and φ(ρ)>ǫρ + ρ̂R obs = arg max d(ρ) + d˜obs (27) ρ∈Z̄∩Fobs and φ(ρ)<−ǫρ Then, their distance and directions to the robot are calculated using Eqs. (6) and (7). For observation to the robot’s companions, we rely on the following assumptions. com Assumption 2: The companions ρ1com , · · · , ρncom are always observable to the robot. Then, p̂com = [dˆcom , φcom ], where dˆcom = dcom + d˜com (28) Remark 1: By Eqs. (20) and (23) to (28), it is implied that the observation/measurement noises d˜ped , d˜obs and d˜com are additive and independent in different observations. A typical example of such noise is the Additive Gaussian White Noise (AGWN). Remark 2: Our general formulations of states (8) and observations (18) are applicable to various types of onboard min djped (ti ) ≤ 0.4 (30) dcom (ti ) − L− F min(dobs (ti ), dobs (ti ), dR obs (ti )) ≤ 0.4 (31) ≤ 0.2 (32) j 3) Stray Condition dcom (ti ) ≥ 2 Based on the above three will be given to the robot as  10000,    −10000, r=    −10\|vR \|, (33) terminal conditions, a reward r follows: if (29) if (30) or (31) or ((32) or (33) else (34) Clearly, a positive reward will be given to the robot if it reaches its goal and it will receive a large negative reward if it collides with anything or be stray from its companion. Otherwise, the robot will receive an intermediate reward −10\|vR \|, which penalizes the robot for its rotational velocity to encourage a smoother trajectory with less turning behaviors. IV. ROLE P LAYING L EARNING In this section, we described the RPL scheme to learn an effective navigation policy Pθ (a\|o) for SCN in an efficient data-driven manner. The core idea is to transform the crowd trajectories data collected from real-world into a simulative and dynamic navigation environment, where the robot can play itself as a virtual pedestrian and iteratively improve the performance of Pθ (a\|o) via Partially Observable Trusted Region Policy Optimization (PO-TRPO). Consider a set of simulative navigation environment E = {E1 , · · · , Ej , · · · }. Each environment Ej = (Tj , Mj ) contains a set of pedestrian trajectories Tj = {ρk0:Tk } and a binary map Mj that annotates the 2-D Cartesian coordinates of 6 obstacles/occupied space in the environment. With E, the abstract process of RPL is described by the following pseudo codes in Algorithm 1. Algorithm 1 Role Playing Learning Initialize navigation policy Pθ for Iter = 0, 1, · · · , MaxIter do while Number of collected sample time steps ≤ Batch size do Randomly choose an environment Ej from E and then a trajectory ρk0:Tk from Tj . Initialize the robot’s position at ρk0 and initial velocities [vT , vR ] = [0, 0]. Set ρg = ρkTk . Choose the robot’s heading such that φg = 0. Choose SCN Mode with probability 0.5 if SCN Mode then Assign ρkT0 :Tk as the trajectory of the robot’s companion, where T0 = arg minT ′ kρkT ′ − ρkTk k ≥ 0.6 else Create a synthesized companion that moves along the robot end if Assign all other trajectories in Tj as pedestrians. while None of the termination conditions in (29) to (33) is satisfied do Update the states and observations of the robot according to Eqs. (8) and (18). Let the robot execute its policy Pθ . Update the robot’s position according to dynamics (1) to (4) Calculate the current reward from Eq. (34) Update the positions of the companion and pedestrians according to the trajectories in Tj . end while end while Update Pθ using PO-TRPO. end for Companion Synthesization in non-SCN mode: As described in Algorithm 1, RPL actually incorporates two different navigation scenarios: the SCN proposed in this paper and the traditional social navigation scenario, where the robot has no human companion. This helps develop a navigation policy adaptable to both situations, with no restrictive assumption on the existence of companion. Particularly, the companion position vector pcom and its observation p̂com are synthesized, with dcom = dˆcom = 0.8 for every time step while φcom = φg . It is clear that the synthesized pcom is equivalent to the situation where the companion is travelling non-distractively along the robot with a constant distance and guarantee that termination conditions (31) and (33) are always false. On the other hand, in SCN mode, the companion is assigned with a truncated trajectory ρkT0 :Tk such that the initial robotcompanion distance is sufficiently large. In this paper, we construct a deep policy neural network to parameterize the navigation policy Pθ , whose structure is shown in Fig. 3. The policy network Pθ is to be trained with the Trust Region Policy Optimization (TRPO) [40] method. Fig. 3: Structure of the deep policy network Pθ . At time ti , the observation vector oi is input to the feature network, which is a feedforward multi-layer perceptron (MLP). The output of the feature network is then fed to a LSTM network [51], a recurrent network for aggregation of the information collected through the navigation process. The LSTM network’s outputs are assigned as the mean vector µ ∈ R2 of the diagonal Gaussian unit N (µ, Σ) on the right. The covariance matrix Σ = σ 2 I ∈ R2×2 , however, is independent of oi amd it is designed to be gradually decreasing during training and fixed during tests and experiments. Finally, the actions ai = [vT , VR ] are drawn according to N (µ, Σ). However, the original TRPO method is derived based on fully observable MDP, which can not be directly applied to our problem due to the imperfect observation in our formulation and practice. Thus, we proposed to extend the original TRPO algorithm as PO-TRPO, which will be described in the following subsections. A. Trusted Region Policy Optimization The TRPO [40] algorithm is an effective on-policy optimization method for large nonlinear policies and tends to give monotonic improvement during the iterative optimization process. To be specific, a fully observable MDP is considered by TRPO and therefore the policy to be optimized is formulated as Pξ∗ (a\|s), where ξ is the parameter vector of the policy P ∗ . Note that, Pξ∗ (a\|s) determines the action a directly from the true state s, which differs from our observation-based policy Pθ (a\|o). Let us consider the following standard definitions of the state-action value function Qξ (si , ai ), the value function Vξ (si ) and the advantage function Aξ (si , ai ): Qξ (si , ai ) = Esi+1 ,ai+1 ,··· [ ∞ X γ l r(si+l )], (35) l=0 Vξ (si ) = Eai ,si+1 ,··· [ ∞ X γ l r(si+l )], (36) l=0 Aξ (si , ai ) = Qξ (si , ai ) − Vξ (si ) (37) ai ∼ Pξ∗ (a\|s), si+1 = F (si , ai ) (38) where In addition, define νξ as the discounted visitation frequencies νξ (s) = p(s0 = s) + γp(s1 = s) + γ 2 p(s2 = s) + · · · (39) where s0 ∼ p0 , ai and si≥1 are generated according to Pξ∗ and F . Let ξ − denote the old parameters in last iteration. TRPO proposes to optimize the parameters ξ iteratively regarding the following objective function: 7 Algorithm 2 PO-TRPO maximize subject to Es∼νξ− ,a∼q∗ [ Pξ∗ (a\|s) q(a\|s) Aξ (a\|s)] ∗ Es∼νξ− [DKL (Pξ− (·\|s) k Pξ∗ (·\|s))] ≤ (40) − ǫ (41) where q ∗ (a\|s) is the importance sampling distribution and DKL (Pξ∗− k Pξ∗ ) is the Kullback-Leibler divergence between the old and current policies. Compute the estimated advantages Âi for all time steps using GAE with the estimated value function V̂ζ . Update θ with objective function (49) and constraints (50) Update ζ with objective function (47) and constraints (48) V. S IMULATION B. Partially Observable TRPO As mentioned, our navigation problem is considered as a POMDP. The policy Pθ (ai \|oi ) depends on the observation oi instead of the true state. Therefore, we write the objective function (40) and the constraint (41) as maximize subject to Es∼νθ− ,a,o∼q [ P o β(o\|s)Pθ (a\|o) Aθ− (a\|s)] q(a,o\|s) Es∼νθ− [DKL (Pθ− (·\|o) k Pθ (·\|o))] ≤ ǫ (42) (43) For PO-TRPO, samples are collected by executing the old policy Pθ− (a\|o) to generate a set of trajectories, such as s0 , o0 , a0 , s1 , o1 , a1 , · · · , sT −1 , oT −1 , aT −1 , sT . Therefore, q(ai , oi \|si ) = β(oi \|si )Pθ− (ai \|oi ) (44) where i = 0, · · · , T − 1. Next, for a trajectory s0:T , we use the generalized advantage estimation (GAE) [39] to construct an empirical estimation Â of the advantage function Aθ− (ai \|si ) as the following: Âi = T −i X V (γλ)l δi+l (45) l=0 where δiV = ri + γ V̂ζ (si+1 ) − V̂ζ (si ) (46) As a data-driven approach, our deep neural network policy requires a massive amount of data to learn the socially concomitant navigation behavior. In this section, we describe how to construct a simulative environment according to the proposed RPL scheme. Particularly, the environments, the deep neural network policy and the PO-TRPO algorithm (Algorithm 2) are developed under the framework of RLLAB [53]. We make use of trajectories of interacting pedestrians collected from five different data sets, which includes the ETH and Hotel video clips from the ETH Walking Pedestrians (EWAP) [11], the motion capture (MC) data set from [14], as well as the Zara and UCY video clips from [32]. Note that, the Zara and UCY data sets have multiple subsets: Zara01, Zara02, Zara03, UCY01 and UCY03. Thus, there are totally 8 different RPL environments, i.e., E = {E1 , · · · , E8 }. The details of these 8 environments are summarized in Tab. I. TABLE I: Details of the 8 RPL environments Name No. Trajectories Name No. Trajectories ETH 365 Zara02 204 Hotel 420 Zara03 137 MC 324 UCY01 413 Zara01 148 UCY03 434 Each trajectory in these environment provides the ID and a sequences of 2-D Cartesian positions of a pedestrian with a sampling period ∆t = 0.1 second. In addition, eight binary grid maps M1 , · · · , M8 representing the occupied space/static obstacles are given. However, these maps are kept unknown to the robot throughout training and evaluation. They are only used to simulate the robot’s perception to the environment as the state pobs and observation p̂obs . Without loss of generality, 1ζ we use the ETH data set as the evaluation environment and Finally, as the conditional observation probability distribuall other data sets in the Tab. I as training environments. In tion β(o\|s) is independent of parameters θ and time, we obtain other words, the learned policy’s performance will be assessed an estimation of the objective function (42) and the constraints in an RPL environment that is excluded during training, (43) by replacing the expectations with sample averages as: which reflects whether it can properly generalize to uncovered PK PTk Pθ (aki \|oki ) k 1 (49) situations. maximize Jθ = PK T k Âi k=1 i=0 Pθ− (ak i \|oi ) k=1 k As some of the trajectories in these environments are θ− subject to D̄KL (Pθ− , Pθ ) ≤ ǫ (50) of people who were wandering or remained approximately stationary, they are excluded from the candidates of the robot’s where companion but will still be considered as pedestrians when the Tk K X X 1 θ− k k D̄KL (Pθ− , Pθ ) = PK DKL (Pθ− (·\|oi ) k Pθ (·\|oi ))] robot is navigating in the same environment. We use a feed-forward neural network with 2 hidden layers T k=1 k k=1 i=0 (51) as the feature network in our NN policy, containing 256 and which has the same form as the one obtained in [39], except 64 Tanh units, respectively. Its output is then fed to a LSTM that the policy Pθ (a\|o) is conditioned on observation o instead. network with 64 units. The variance of of the Gaussian output Finally, the constrained optimization problem described in unit σ is chosen to be linearly decaying from 0.5 to 0.05 in 100 (49) and (50) is solved by conjugate gradient algorithm [52] training iterations, which effectively encourages exploration algorithm. To summarize, the pseudo code for PO-TRPO during the early stage of learning and ensures convergence update in Algorithm 1 is given as below: of the navigation policy. For GAE, a 3 layer feed-forward and V̂ζ (si ) is the estimation of the value function (36) with parameters ζ (and ζ − being the old parameters). By collecting a set of K trajectories {sk0:Tk , ok0:Tk , ak0:Tk }K k=1 , V̂ζ is obtained by solving the following constrained regression problem [39]: PK PTk PTk −i l k 2 minimize J1ζ = k=1 i=0 kV̂ζ (ski )− l=0 γ ri+l k (47) PK PTk kVζ (ski )−Vζ− (ski )k ≤ ǫ1 (48) subject to k=1 i=0 2J − 8 network with 256,64,16 Tanh units are used, with γ = 0.995 and λ = 0.96. The update step size for policy network is adaptively chosen as ǫ = 0.01/σ. For GAE update, a fixed step size ǫ1 = 0.1 is used. The update batch size (Batch size in Algorithm 1) is 50000. In RPL, we consider at most 3 pedestrians (i.e., np = 3). Thus, the state pped and observation p̂ped will only describe the 3 closest pedestrians and omit the others. For the situation where less that 3 pedestrians are perceived, dummy static pedestrians will be created in the remote corner of the environment so as to maintain the dimensions of pped and p̂ped . Considering a Kobuki Turtlebot 2 with a Hokuyo URG04LX laser range finder [50] mounted on its top, we specify the sensor limitation of the robot in simulation as follows: 2π (52) φ+ = φ+ ped obs = 3 2π φ− = φ− (53) ped obs = − 3 The measurement noises d˜ped , d˜com and d˜obs are modeled 2 2 2 by zero-mean Gaussian N (0, σped ), N (0, σcom ) and N (0, σobs ) with their variances specified as follows: σped σcom σobs = 0.01djped (54) = = 0.01djcom 0.01djobs (55) (56) Finally, the maximum translational and rotational velocities are assigned as 0.7m/s and π3 rad/s, i.e., 0 ≤ vT ≤ 0.7 and \|vR \| ≤ π3 and ∆t = 0.1. An example of our RPL environment constructed from the ETH data set is illustrated in Fig. 4 (a) Real-world Environment A. Results We trained our deep policy network for 1200 iterations with the data from RPL environments except for the held-out ETH environment. The curve of average discounted return obtained from each batch of trajectories is visualized in Fig. 5 Fig. 5: Average discounted return as RPL progresses We compare the performance of our policy with a planner based on RVO [8], where the robot, its companion and the surrounding pedestrians are treated agents. In every time steps, the positions and velocities of all agents are given to the planner. Note that, for fair comparison, the agents’ positions are subject to noise described in (54) and (55). For observations to obstacles, we assume the planner has full and perfect knowledge as required in the original RVO algorithm. With this protocol, we update the robot’s position according to the planner’s output and update the positions of the other agents according to their own trajectories in the RPL environments. The same termination conditions in Section III are applied to the robot directed by the RVO-based planner to determine whether the robot has conducted an successful navigation. For both of our policy and the RVO-based planner, we conduct 300 trials in the evaluation environment and compute the rates (in percentages) of different terminal conditions (RG: the robot reaches the goal successfully; HC/HP/HO: the robot hits a companion/pedestrian/obstacle; and LC: the robot loses its companion). The performance statistics of our policy and the RVO-based planner in SCN scenarios are listed in Tab. II. (b) Simulative Environment for RPL Fig. 4: An illustrative example our RPL simulative environment. The black curve represents the trajectory of the robot navigating toward its goal (the red dot). The yellow curve denotes the trajectory of the robot’s companion. Besides, there are a number of blue curves representing the pedestrians perceived by the robot and the green lines denotes the fences around the entrance of the university (bottom center). Note that, all trajectories of pedestrians are not synthesized but captured from the video. Thus, the robot can be thought as playing a role as an extra person in an realistic environment. TABLE II: Rates of different terminal conditions of our policy and RVO-based planner in SCN scenarios Terminal Condition Our policy RVO RG 77 29.7 LC 5.7 47 HC 6.7 0.3 HP 9.6 23 HO 1 0 It can bee seen from Tab. II that our policy performs much better than the RVO-based planner in SCN. The RVO-based planner has a much lower success rate (29.7%) while its rate of LC is 47%, suggesting that it frequently losses its companion in SCN. Clearly, this is due to the fact that RVO is in nature a collision avoidance algorithm. Thus, it simply takes the robot’s 9 companion as another normal agent and the robot tends to stay far behind its companion to avoid collision instead of actively following it. On the contrary, our policy achieves a much higher success rate (77%). This indicates that it learns to effectively balance the objectives of SCN so that the robot is able to reach the prescribed goal while maintaining its distance to its companion and avoiding collision with other agents in the environment. In addition to SCN, the scenarios without companion are also tested, which, as analyzed in the previous sections, reduces to the traditional social navigation scenarios. The comparative results are shown in Tab. III. TABLE III: Rates of different terminal conditions of our policy and RVO-based planner in traditional social navigation scenarios Terminal Condition Our policy RVO RG 84 80 HP 13.7 18 Fig. 6: The narrow corridor where experiments are performed HO 2.3 2 For situations without companion, our policy still outperforms the RVO-based planner with higher success rate (84% to 80%) and lower HP rate (13.7% to 18%). Finally, it is worth noting that the RVO-based planner requires velocities of the companion/pedestrians and an accurate global map of the static obstacles. Conversely, our policy depends only on position measurements that are directly accessible from the robot’s onboard sensors, which is therefore much simpler and more practical. (a) Trajectories of the robot (moving from left to right) and other two pedestrians (moving from right to left). VI. E XPERIMENTS In experiments, we assess the performance of our developed navigation policy by comparing it with humans in the same scenarios. Particularly, a robot and a human are to repeat each specific navigation scenario for 10 times, respectively. Then, the the following two metrics are calculated: 1) Average minimum distance to the pedestrians (D̄ped ): the average of the minimum distance between the robot/compared human to other pedestrians throughout a trajectory . 2) Average maximum distance to the companion (D̄com ): the average of the maximum distance between the robot/compared human to its/his companion throughout a trajectory. We use the same mobile platform (a synchron-drive Turtlebot 2 with a Kobuki base) and the same laser range finder (Hokuyo URG-04LX) simulated in last section. For pedestrian detection and localization, we adopt the ROS-compatible leg tracker in [54]. We use an ultra wideband (UWB) indoor positioning system to localize the companion and the navigation goal, which can then be easily mapped to the observations p̂com and dg , φg based on the odometry of the robot. Finally, a laptop is placed onboard as the processing unit and the policy is operated with a period of 0.1 second. The experiments are conducted in a narrow corridor with width of 1.56 meters as shown in Fig. 6, which is a typical scenario that requires pedestrians to navigate cooperatively. (b) Human control experiment in a similar navigation scenario. The black trajectory is from left to right and the other two are from right to left. Fig. 7: Comparison between the robot with our policy and human control experiment in a social navigation scenario A. Scenario 1: Traditional Social Navigation In this subsection, we examine our method’s performance in traditional social navigation scenario. Particularly, the robot is required to pass the corridor with two oncoming pedestrians and arrive at a goal that is 7 meters ahead. In addition, a control experiment of 3 humans (one as the compared human and the other two as pedestrians) is conducted in the same space. The metric D̄ped is computed. Example trajectories of the robot and the human control are shown in Fig. 7. In the robotic experiments, the trajectories of pedestrians are obtained from the robot’s laser range finder while the robot’s trajectory is based on its own odometry sensor. On the other hand, all trajectories in the human control experiments are captured using the UWB localization system. From Fig. 7, it is clear that the robot with our policy is 10 able understand human’s cooperative behavior for collision avoidance and navigate in an appropriate manner such that both itself and the other two pedestrians can successfully pass through the corridor. Specifically, when observing the two pedestrians (blue and purple) 4 meters ahead. The robot started to approach the wall on its left side so as to create free space on the right for the pedestrians to smoothly walk through. By comparing both figures in Fig. 7, we can see that the robot is as proactive as human since both black trajectories in Fig. 7(a) and Fig. 7(b) started to make space for the oncoming pedestrians at the early stage of cooperative avoidance process. As for the performance metrics, the average minimal distance to pedestrians for our robot is D̄ped = 0.35m. Although it is smaller than that of the human control experiments (D̄ped =0.56m), this value still indicates a safe and decent navigation behavior of our robot as its radius is only 0.17m. (a) Trajectories of the robot and its companion (moving from left to right) and a pedestrian (moving from right to left). B. Scenario 2: Socially Concomitant Navigation In this subsection, the scenario of SCN is studied. A human companion initially standing in front of the robot will start to walk through the same corridor while another pedestrian is passing from the other end. As described in the previous sections, the robot with our policy should closely navigate with its companion and avoid the oncoming pedestrian cooperatively. An additional metric D̄com is used to evaluate the performance of our policy by comparing with the statistics obtained from another 10 human control experiments. Example trajectories are shown in Fig. 8 and the performance metrics D̄ped and D̄com are summarized in Tab. IV. TABLE IV: Performance metrics of the robot and human controls in SCN scenarios Robot Compared Human D̄ped (m) 0.49 0.37 D̄com (m) 1.05 1 As shown in Fig. 8 and Tab. IV, the robot is able to achieve both objectives of SCN. On one hand, it is effectively engaged into the joint collision avoidance process. The resulted behavior is similar to that observed in the last subsection and the robot even has a slightly larger D̄ped . On the other hand, the average maximum distance D̄com is 1.05m, which is within the limit (2m) we specified in the learning process and nearly the same as that of the compared human, showing that the robot can actively navigate along with its companion instead of deviating to other areas or lagging itself behind. This shows that the robot driven by our policy is able to understand the pace of its companion and achieve a similar sense of companionship in terms of distance. In sum, the above results demonstrate the practical efficacy of our methods for both the traditional social navigation and the more complicated SCN scenarios. It proves that the policy learned from our RPL simulative environment is transferable to uncovered real-world situations. VII. C ONCLUSIONS In this paper, the problem of socially concomitant navigation (SCN) has been investigated and formulated under a POMDP (b) Human control experiment in a similar SCN. The black (compared human) and orange (companion) trajectories are from left to right and the blue (pedestrian) trajectory is from right to left. Fig. 8: Comparison between the robot with our policy and human control experiment in a SCN scenario framework, with explicit considerations of the limitation and inaccuracy of mobile robots’ onboard sensors. The Partially Observable TRPO (PO-TRPO) algorithm has been proposed for optimization of navigation policies. The Role Playing Learning (RPL) scheme has been developed to enable efficient and safe reinforcement learning of navigation policies by mirroring a large amount of real-world pedestrian trajectories into simulative environments. Comparative simulation and experiment studies have demonstrated the efficacy and superiority of our policy in both SCN and traditional social navigation scenarios. R EFERENCES [1] D. F. W. B. S. Thrun, D. Fox, and W. Burgard, “The dynamic window approach to collision avoidance,” IEEE Transactions on Robotics and Automation, vol. 4, p. 1, 1997. [2] Y. K. Hwang and N. Ahuja, “A potential field approach to path planning,” IEEE Transactions on Robotics and Automation, vol. 8, no. 1, pp. 23–32, 1992. [3] S. S. Ge and Y. J. Cui, “New potential functions for mobile robot path planning,” IEEE Transactions on robotics and automation, vol. 16, no. 5, pp. 615–620, 2000. [4] P. Trautman and A. Krause, “Unfreezing the robot: Navigation in dense, interacting crowds,” in Intelligent Robots and Systems (IROS), 2010 IEEE/RSJ International Conference on, 2010, pp. 797–803. [5] P. Trautman, J. Ma, R. M. Murray, and A. Krause, “Robot navigation in dense human crowds: Statistical models and experimental studies of human–robot cooperation,” The International Journal of Robotics Research, vol. 34, no. 3, pp. 335–356, 2015. [6] D. Helbing and P. Molnar, “Social force model for pedestrian dynamics,” Physical review E, vol. 51, no. 5, p. 4282, 1995. [7] D. Helbing, I. Farkas, and T. Vicsek, “Simulating dynamical features of escape panic,” Nature, vol. 407, no. 6803, pp. 487–490, 2000. 11 [8] J. Van Den Berg, S. J. Guy, M. Lin, and D. Manocha, “Reciprocal nbody collision avoidance,” in Robotics research, 2011, pp. 3–19. [9] J. Van Den Berg, P. Abbeel, and K. Goldberg, “Lqg-mp: Optimized path planning for robots with motion uncertainty and imperfect state information,” The International Journal of Robotics Research, vol. 30, no. 7, pp. 895–913, 2011. [10] J. Van den Berg, M. Lin, and D. Manocha, “Reciprocal velocity obstacles for real-time multi-agent navigation,” in Robotics and Automation, 2008. ICRA 2008. IEEE International Conference on, 2008, pp. 1928–1935. [11] S. Pellegrini, A. Ess, K. Schindler, and L. Van Gool, “You’ll never walk alone: Modeling social behavior for multi-target tracking,” in 2009 IEEE 12th International Conference on Computer Vision, 2009, pp. 261–268. [12] K. Yamaguchi, A. C. Berg, L. E. Ortiz, and T. L. Berg, “Who are you with and where are you going?” in Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, 2011, pp. 1345–1352. [13] M. Kuderer, H. Kretzschmar, C. Sprunk, and W. Burgard, “Feature-based prediction of trajectories for socially compliant navigation.” in Robotics: science and systems, 2012. [14] H. Kretzschmar, M. Spies, C. Sprunk, and W. Burgard, “Socially compliant mobile robot navigation via inverse reinforcement learning,” The International Journal of Robotics Research, p. 0278364915619772, 2016. [15] B. Kim and J. Pineau, “Socially adaptive path planning in human environments using inverse reinforcement learning,” International Journal of Social Robotics, vol. 8, no. 1, pp. 51–66, 2016. [16] A. Bicchi, A. Fagiolini, and L. Pallottino, “Towards a society of robots,” IEEE Robotics & Automation Magazine, vol. 17, no. 4, pp. 26–36, 2010. [17] H.-M. Gross, C. Schroeter, S. Mueller, M. Volkhardt, E. Einhorn, A. Bley, C. Martin, T. Langner, and M. Merten, “Progress in developing a socially assistive mobile home robot companion for the elderly with mild cognitive impairment,” in Intelligent Robots and Systems (IROS), 2011 IEEE/RSJ International Conference on, 2011, pp. 2430–2437. [18] H. Wang and X. P. Liu, “Adaptive shared control for a novel mobile assistive robot,” IEEE/ASME Transactions on Mechatronics, vol. 19, no. 6, pp. 1725–1736, 2014. [19] B. D. Argall, S. Chernova, M. Veloso, and B. Browning, “A survey of robot learning from demonstration,” Robotics and autonomous systems, vol. 57, no. 5, pp. 469–483, 2009. [20] P. Abbeel and A. Y. Ng, “Apprenticeship learning via inverse reinforcement learning,” in Proceedings of the twenty-first international conference on Machine learning, 2004, p. 1. [21] B. D. Ziebart, A. L. Maas, J. A. Bagnell, and A. K. Dey, “Maximum entropy inverse reinforcement learning.” in AAAI, vol. 8, 2008, pp. 1433– 1438. [22] N. D. Ratliff, J. A. Bagnell, and M. A. Zinkevich, “Maximum margin planning,” in Proceedings of the 23rd international conference on Machine learning, 2006, pp. 729–736. [23] B. D. Ziebart, N. Ratliff, G. Gallagher, C. Mertz, K. Peterson, J. A. Bagnell, M. Hebert, A. K. Dey, and S. Srinivasa, “Planning-based prediction for pedestrians,” in Intelligent Robots and Systems, 2009. IROS 2009. IEEE/RSJ International Conference on, 2009, pp. 3931– 3936. [24] P. Henry, C. Vollmer, B. Ferris, and D. Fox, “Learning to navigate through crowded environments,” in Robotics and Automation (ICRA), 2010 IEEE International Conference on, 2010, pp. 981–986. [25] P. Vernaza and D. Bagnell, “Efficient high dimensional maximum entropy modeling via symmetric partition functions,” in Advances in Neural Information Processing Systems, 2012, pp. 575–583. [26] K. M. Kitani, B. D. Ziebart, J. A. Bagnell, and M. Hebert, “Activity forecasting,” in European Conference on Computer Vision. Springer, 2012, pp. 201–214. [27] J. Choi and K.-E. Kim, “Map inference for bayesian inverse reinforcement learning,” in Advances in Neural Information Processing Systems, 2011, pp. 1989–1997. [28] K. Kim, D. Lee, and I. Essa, “Gaussian process regression flow for analysis of motion trajectories,” in Computer vision (ICCV), 2011 IEEE international conference on, 2011, pp. 1164–1171. [29] A. Alahi, K. Goel, V. Ramanathan, A. Robicquet, L. Fei-Fei, and S. Savarese, “Social lstm: Human trajectory prediction in crowded spaces,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 961–971. [30] A. Robicquet, A. Sadeghian, A. Alahi, and S. Savarese, “Learning social etiquette: Human trajectory understanding in crowded scenes,” in European Conference on Computer Vision, 2016, pp. 549–565. [31] A. Johansson, D. Helbing, and P. K. Shukla, “Specification of the social force pedestrian model by evolutionary adjustment to video tracking [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] [54] data,” Advances in complex systems, vol. 10, no. supp02, pp. 271–288, 2007. A. Lerner, Y. Chrysanthou, and D. Lischinski, “Crowds by example,” in Computer Graphics Forum, vol. 26, no. 3. Wiley Online Library, 2007, pp. 655–664. D. Helbing and A. Johansson, “Pedestrian, crowd and evacuation dynamics,” in Encyclopedia of Complexity and Systems Science, 2009, pp. 6476–6495. J. Müller, C. Stachniss, K. Arras, and W. Burgard, “Socially inspired motion planning for mobile robots in populated environments,” in Proc. of International Conference on Cognitive Systems, 2008. D. Mehta, G. Ferrer, and E. Olson, “Autonomous navigation in dynamic social environments using multi-policy decision making,” in Intelligent Robots and Systems (IROS), 2016 IEEE/RSJ International Conference on, 2016, pp. 1190–1197. A. F. Foka and P. E. Trahanias, “Probabilistic autonomous robot navigation in dynamic environments with human motion prediction,” International Journal of Social Robotics, vol. 2, no. 1, pp. 79–94, 2010. M. Seder and I. Petrovic, “Dynamic window based approach to mobile robot motion control in the presence of moving obstacles,” in Robotics and Automation, 2007 IEEE International Conference on, 2007, pp. 1986–1991. P. Fiorini and Z. Shiller, “Motion planning in dynamic environments using velocity obstacles,” The International Journal of Robotics Research, vol. 17, no. 7, pp. 760–772, 1998. J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel, “Highdimensional continuous control using generalized advantage estimation,” arXiv preprint arXiv:1506.02438, 2015. J. Schulman, S. Levine, P. Moritz, M. I. Jordan, and P. Abbeel, “Trust region policy optimization,” CoRR, abs/1502.05477, 2015. Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015. M. Pfeiffer, M. Schaeuble, J. Nieto, R. Siegwart, and C. Cadena, “From perception to decision: A data-driven approach to end-toend motion planning for autonomous ground robots,” arXiv preprint arXiv:1609.07910, 2016. Y. F. Chen, M. Liu, M. Everett, and J. P. How, “Decentralized noncommunicating multiagent collision avoidance with deep reinforcement learning,” arXiv preprint arXiv:1609.07845, 2016. Y. Zhu, R. Mottaghi, E. Kolve, J. J. Lim, A. Gupta, L. Fei-Fei, and A. Farhadi, “Target-driven visual navigation in indoor scenes using deep reinforcement learning,” arXiv preprint arXiv:1609.05143, 2016. D.-G. Choi, Y. Bok, J.-S. Kim, and I. S. Kweon, “Extrinsic calibration of 2-d lidars using two orthogonal planes,” IEEE Transactions on Robotics, vol. 32, no. 1, pp. 83–98, 2016. L. M. Miller and T. D. Murphey, “Optimal planning for target localization and coverage using range sensing,” in Automation Science and Engineering (CASE), 2015 IEEE International Conference on, 2015, pp. 501–508. F. Endres, J. Hess, J. Sturm, D. Cremers, and W. Burgard, “3-d mapping with an rgb-d camera,” IEEE Transactions on Robotics, vol. 30, no. 1, pp. 177–187, 2014. S. Foix, G. Alenya, J. Andrade-Cetto, and C. Torras, “Object modeling using a tof camera under an uncertainty reduction approach,” in Robotics and Automation (ICRA), 2010 IEEE International Conference on, 2010, pp. 1306–1312. M. Liu and R. Siegwart, “Topological mapping and scene recognition with lightweight color descriptors for an omnidirectional camera,” IEEE Transactions on Robotics, vol. 30, no. 2, pp. 310–324, 2014. L. Kneip, F. Tâche, G. Caprari, and R. Siegwart, “Characterization of the compact hokuyo urg-04lx 2d laser range scanner,” in Robotics and Automation, 2009. ICRA’09. IEEE International Conference on, 2009, pp. 1447–1454. S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997. J. Nocedal and S. J. Wright, “Conjugate gradient methods,” Numerical optimization, pp. 101–134, 2006. Y. Duan, X. Chen, R. Houthooft, J. Schulman, and P. Abbeel, “Benchmarking deep reinforcement learning for continuous control,” arXiv preprint arXiv:1604.06778, 2016. A. Leigh, J. Pineau, N. Olmedo, and H. Zhang, “Person tracking and following with 2d laser scanners,” in Robotics and Automation (ICRA), 2015 IEEE International Conference on, 2015, pp. 726–733.	2cs.AI
arXiv:1301.2418v1 [math.AC] 11 Jan 2013 BORNES EFFECTIVES DES FONCTIONS D’APPROXIMATION DES SOLUTIONS FORMELLES D’ÉQUATIONS BINOMIALES GUILLAUME ROND Résumé. The aim of this paper is to give an effective version of the Strong Artin Approximation Theorem for binomial equations. First we give an effective version of the Greenberg Approximation Theorem for polynomial equations, then using the Weierstrass Preparation Theorem, we apply this effective result to binomial equations. We prove that the Artin function of a system of binomial equations is bounded by a doubly exponential function in general and that it is bounded by an affine function if the the order of the approximated solutions is bounded. 1. Introduction Dans [Gr], M. Greenberg a montré le théorème suivant (k est un corps quelconque) : Théorème 1.1. [Gr] Soit I un idéal de k[t, X] (X := (X1 , ..., Xn )). Alors il existe une fonction β : N −→ N telle que : ∀i ∈ N, ∀xj (t) ∈ k[[t]], 1 ≤ j ≤ n tels que f (t, x(t)) ∈ (t)β(i) ∀f ∈ I ∃xj (t) ∈ k[[t]], 1 ≤ j ≤ n, tels que f (t, x(t)) = 0 ∀f ∈ I et xj (t) − xj (t) ∈ (t)i ∀j. De plus il existe deux constantes a et b telles que ∀i ∈ N β(i) ≤ ai + b. La plus petite fonction β vérifiant cette propriété est appelée fonction d’approximation de Greenberg ou fonction de Artin-Greenberg de I. Dans [Ar], M. Artin a généralisé ce résultat au cas où t est remplacé par un nombre fini quelconque de variables en montrant l’existence d’une fonction d’approximation dans ce cas. On peut énoncer son théorème sous la forme suivante (dans le cas particulier où t est remplacé par deux variables t et z, mais seul ce cas nous intéressera par la suite) : Théorème 1.2. [Ar] Soit I un idéal de k[t, z, X] (X := (X1 , ..., Xn )). Alors il existe une fonction β : N −→ N telle que : ∀i ∈ N, ∀xj (t, z) ∈ k[[t, z]], 1 ≤ j ≤ n, tels que f (t, z, x(t, z)) ∈ (t, z)β(i) ∀f ∈ I ∃xj (t, z) ∈ k[[t, z]], 1 ≤ j ≤ n tels que f (t, z, x(t, z)) = 0 ∀f ∈ I et xj (t, z)−xj (t, z) ∈ (t, z)i ∀j. La plus petite fonction vérifiant cette propriété est appelée fonction de Artin de I. On peut remarquer que si I est un idéal de k[X] on peut parler de sa fonction de Artin-Greenberg et de sa fonction de Artin en faisant référence à la fonction de Artin-Greenberg de Ik[t, X] et à la fonction de Artin de Ik[t, z, X]. Malheureusement, la preuve donnée par M. Artin (qui utilise essentiellement le théorème des fonctions implicites et le théorème de préparation de Weierstrass) n’apporte que très peu d’information sur la nature de la croissance de cette fonction (exceptée que celle-ci est constructive [La]), et pendant longtemps s’est posée la question de savoir si toute fonction 2000 Mathematics Subject Classification. Primary 13B40 ; Secondary 13P99, 14B12. 1 2 GUILLAUME ROND de Artin était bornée par une fonction affine. On peut mentionner qu’il existe une manière différente de celle de M. Artin pour montrer l’existence d’une fonction d’approximation qui utilise les ultraproduits [B-D-L-v.D] et qui n’est pas constructive et donc qui n’apporte aucune information sur ces fonctions d’approximation. Dans [Ro2] est présenté un exemple d’un idéal dont la fonction de Artin n’est pas bornée par une fonction affine (c’est le seul exemple connu jusqu’à présent, l’exemple donné dans [Ro3] concernant le cas où les variables t et z sont remplacées par trois variables). Dans cet exemple il est montré que la fonction de Artin considérée est minorée par une fonction polynomiale de degré 2. Néanmoins on ne sait rien de plus sur cette fonction de Artin et on ne connait toujours aucune borne générale sur aucun exemple autre que quelques cas connus où la fonction de Artin est majorée par une fonction affine. Pour mieux comprendre la croissance des fonctions de Artin, il serait intéressant d’avoir des exemples pour lesquels on ait des bornes effectives de leur fonction de Artin. On peut remarquer que l’exemple cité de [Ro2] a la particularité d’être un idéal principal engendré par un binôme. Le but de ce travail est justement de donner des bornes générales pour la fonction de Artin d’un idéal binomial I. Le principe général est le suivant : il est plus facile d’essayer de donner des bornes sur les fonctions de Artin d’une famille d’idéaux que sur un idéal en particulier. En effet, la preuve de M. Artin et ses avatars utilisent toujours une récurrence sur la hauteur de I : si I est premier, soit on peut utiliser le théorème des fonctions implicites et on construit la solution approchée cherchée, soit on remplace I par I + (δ) où δ est un mineur de la matrice jacobienne de I bien choisi et on augmente ainsi la hauteur de I. Cependant, le nouvel idéal obtenu n’est plus premier a priori et l’utilisation du théorème des fonctions implicites nécessitant de travailler avec un idéal premier, il faut remplacer I par ses idéaux premiers associés et de borner la fonction de Artin de I par celles de ses idéaux premiers associés. Il se trouve que l’on connait des bornes sur le degré de générateurs de ces idéaux premiers en fonction du degré des générateurs de I. Il est donc assez naturel d’essayer de trouver une fonction qui majore toutes les fonctions de Artin des idéaux engendré par des polynômes de degré d fixé. Cette stratégie fonctionne comme l’a montré M. Artin, i.e. on peut choisir la même fonction β dans le théorème 1.2 pour tous les idéaux engendrés par des polynômes de degré inférieur à une valeur fixée (c.f. [Ar]). Néanmoins, il n’est pas possible d’avoir plus d’informations sur cette fonction de Artin "uniforme" que le fait que celle-ci soit constructible (cf. [La]). Dans notre travail, nous montrons néanmoins que cette stratégie donne une borne effective des fonctions de Artin-Greenberg, c’est-à-dire donne une borne effective sur les coefficients a et b du théorème 1.1 en fonction du degré d des générateurs de I et du nombre n de variables Xi (c.f. théorème 3.1). Celle-ci est assez grande (polynomiale en d, doublement exponentielle en n) mais a le mérite d’être uniforme. Ensuite, grâce au lemme 4.1, nous utilisons le théorème de préparation de Weierstrass de manière à ramener la majoration de la fonction de Artin d’un idéal binomial à deux majorations (cf. section 4) : • La majoration de la fonction de Artin d’un idéal binomial dont les solutions approchées évitent le lieu singulier (et donc dans ce cas cette fonction est majorée par une fonction affine). • La majoration des fonctions de Artin-Greenberg d’une famille d’idéaux (JD )D∈N engendrés par des polynômes dépendant d’un nombre croissant de variables mais dont le degré est borné par le degré des générateurs de I. Pour cela on utilise donc le théorème 3.1. On en déduit deux choses (sur un corps algébriquement clos de caractéristique nulle) : si l’on borne l’ordre des solutions approchées xj (t, z), 1 ≤ j ≤ n, alors la fonction de Artin d’un idéal binomial est majorée par une fonction affine, et en général la fonction de Artin d’un BORNES EFFECTIVES DES FONCTIONS D’APPROXIMATION DES SOLUTIONS FORMELLES D’ÉQUATIONS BINOMIALES 3 idéal binomial est majorée par une fonction doublement exponentielle (c.f. théorème 4.2). Malheureusement on n’a aucune manière de savoir si cette dernière borne est raisonnable ou pas. En effet la famille d’idéaux (JD ) n’est pas quelconque, mais semble tout de même assez difficile à appréhender : ces idéaux ne sont pas réduits en général et il est vite impossible de calculer leur radical ou une décomposition primaire de ceux-ci du fait du nombre rapidement important de variables qui entrent en jeu. 2. Rappels sur certaines bornes effectives en algèbre commutative Nous allons commencer par rappeler quelques résultats classiques en algèbre commutative effective que nous allons utiliser librement dans la suite. Théorème 2.1. [He][Te] Soit I un idéal de k[u], u := (u1 , ..., un ), I = (f1 , ..., fp ) avec deg (fr ) ≤ d pour 1 ≤ r ≤ p. Soit I = Q1 ∩ · · · ∩ Qs une décomposition primaire minimale √ √ √ (i.e telle que Q ,..., Q soient tous distincts). On a alors I = P l s 1 ∩ · · · ∩ Ps avec √ Pl := Ql pour 1 ≤ l ≤ s. On a alors les résultats suivants : √ e a) Soit e := min{n, p}(n + 2)(d + 1)min{n,p}+1 ≤ (n + 2)2 (d + 1)n+1 . On a I ⊂ I. b) On s ≤ dmin{n,p} . c) Il existe une fonction λ1 (n, d), polynomiale en d et de degré exponentiel en n, tel que chaque Pl est engendré par des polynômes de degré inférieur ou égal à λ1 (n, d). d) Il existe une fonction λ2 (n, d), polynomiale en d et de degré exponentiel en n, tel que chaque Ql est engendré par des polynômes de degré inférieur ou égal à λ1 (n, d). Proposition 2.2. [S] Soient I1 ,..., Iq , q idéaux de k[u1 , ..., un ] engendrés par des polynômes de degré inférieur ou égal à d. Alors I := I1 ∩ · · · ∩ Iq est engendré par des polynômes de n−1 degré inférieur ou égal à n((q − 1)d)2 +d Démonstration. Écrivons Ii = (fi,1 , ..., fi,si ) où deg (fi,j ) ≤ d pour tous entiers i et j. Tout système de génerateurs de I, noté g (1) , ..., g (s) correspond à un système générateur, noté (s) (1) {ui,j , ..., ui,j } pour 1 ≤ i ≤ q et 1 ≤ j ≤ si , du k[u1 , ..., un ]-module défini par les équations (f1, 1 U1, 1 + · · · + f1,s1 U1,s1 ) − (fi,1 Ui, 1 + · · · + fi,si Ui,si ) = 0, 2 ≤ i ≤ q (l) (l) et relié par g (l) = fi,1 ui, 1 + · · · + fi,si ui,si pour 1 ≤ l ≤ s. D’après la proposition 55 de [S], (1) (s) (l) n−1 il existe un tel système {ui,j , ..., ui,j } tel que deg (ui,j ) ≤ n((q − 1)d)2 l. On en déduit le résultat. pour tous i, j et 3. Borne effective de la fonction de Artin-Greenberg dans le cas polynomial Dans cette partie nous allons redémontrer le théorème de Greenberg en suivant essentiellement sa preuve mais en faisant attention à la complexité de chaque étape. Dans toute cette partie k désigne un corps de caractéristique nulle. Théorème 3.1. [Gr][Ar] Pour tous n, d, i ∈ N, il existe β : N −→ N tel que pour tout idéal I de k[t, x], avec x = (x1 , ..., xn ), tel que deg (fr ) ≤ d pour tout r, pour tout i ∈ N et pour tout x(t) ∈ k[[t]]n tel que f (t, x(t)) ∈ (t)β(i) pour tout f ∈ I, il existe x(t) ∈ k[[t]]n tel que f (t, x(t)) = 0 pour tout f ∈ I et x(t) − x(t) ∈ (t)i . De plus β peut être choisie affine, de la forme i 7−→ a(n, d)i + b(n, d) où a(n, d) et b(n, d) sont bornés par une fonction polynomiale en d de degré exponentiel en n. 4 GUILLAUME ROND Définition 3.2. Nous noterons dans la suite β(n, d, i) le plus petit entier β(i) qui vérifie les conclusions du théorème 3.1. Nous allons noter βh (n, d, i) le plus petit entier qui vérifie les conclusions du théorème 3.1 pour tout idéal I quelconque de hauteur h engendré par des polynômes de degré inférieur ou égal à d, et βhp (n, d, i) le plus petit entier qui vérifie les conclusions du théorème 3.1 pour tout idéal I premier de hauteur h engendré par des polynômes de degré inférieur ou égal à d. Théorème 3.3. Pour tous n, d, d′ , i, j ∈ N, nous avons les relations suivantes : (1) p βn+1 (n, d, i) ≤ 2 (2) βk (n, d, i) ≤ (n + 3)2 (d + 1)2n+3 max {βhp (n, λ1 (n + 1, d), i)} (3) βkp (n, d, i) ≤ (e′ + 1)βk+1 (n, k(d − 1), i) + 1 n n+2 . e′ := (n + 3)2 1 + λ2 (n + 1, d) + (n + 1)((dn+1 − 2)λ2 (n + 1, d))2 h>k Ce théorème implique directement le théorème 3.1. Nous allons donner une preuve du thèorème 3.1 en montrant au fur et à mesure les inégalités (1), (2) et (3) du théorème 3.3. • Si ht(I) = n + 1 et I est premier, alors I est un idéal maximal. Supposons qu’il existe x(t) ∈ k[[t]]x tel que f (t, x(t)) ∈ (t)2 pour tout f ∈ I, et notons ϕ : k[t, x] −→ k le k-homomorphisme défini par ϕ(h) = h(0, x(0)). Alors I ⊂ Ker(ϕ), mais Ker(ϕ) étant un idéal propre de k[t, x] et I maximal, nous avons I = Ker(ϕ). Donc en particulier t ∈ I, mais clairement t ∈ / Ker(ϕ′ ) ce qui contredit l’existence de x(t). Donc on peut prendre ici β = 2, le théorème 3.1 est alors valable car l’hypothèse n’est jamais vérifiée, et on a l’égalité (1). • Supposons que ht(I) = k, où I = (f1 , ..., fp ), et que le théorème 3.1 est vrai pour tout idéal de hauteur strictement plus grande que k. √ Soit I = P1 ∩ · · · ∩ Ps la décomposition primaire du radical de I, où ht(Pl ) = hl . Notons Ps e := (n + 1)(n + 3)(d + 1)n+2 . Alors l’entier β ′ := e. l=1 βhpl (n, λ1 (n + 1, d), i) satisfait ′ les conditions du théorème 3.1 pour I. En effet soit x(t) ∈ k[[t]]n tel que f (t, x(t)) ∈ (t)β p β (n,λ(n+1,d),i) pour tout f ∈ I. Donc il existe un entier l tel que g(t, x(t)) ∈ (t) hl pour tout g ∈ Pl . En effet, dans le cas contraire, pour tout l, il existerait gl ∈ Pl tel que β p (n,λ(n+1,d),i) gl (t, x(t)) ∈ / (t) hl . Notons alors g := (g1 ...gs )e . On a g ∈ I par définition de ′ e et des Pl , mais g(t, x(t)) ∈ / (t)β ce qui contredit ce qui précède, et donc il existe un entier l tel que g(t, x(t)) ∈ (t)βhl (n,λ(n+1,d),i) pour tout g ∈ Pl . Donc il existe x(t) ∈ k[[t]]n tel que g(t, x(t)) = 0 pour tout g ∈ Pl et tel que x(t) − x(t) ∈ (t)i . Comme I ⊂ Pl , on obtient la conclusion voulue. On a montré ainsi l’inégalité (2). • Nous supposons donc maintenant que I = (f1 , ..., fp ) est premier de hauteur k. Alors k[t, x]I est régulier, et on peut supposer que f1 , ..., fk engendrent Ik[t, x]I . Il existe donc un 1 ,...,fk ) , noté δ, tel que δ ∈ / I (cf. par exemple mineur d’ordre k de la matrice jacobienne ∂(f∂(t,x) proposition 1 [Wa]). Notons J := (f1 , ..., fp , δ). On a alors ht(J) = k + 1. Remarquons que deg (δ) ≤ k(d − 1). Notons alors I ′ := (f1 , ..., fk ) et soit I ′ = Q1 ∩ · · · ∩ Qq une décomposition primaire réduite BORNES EFFECTIVES DES FONCTIONS D’APPROXIMATION DES SOLUTIONS FORMELLES D’ÉQUATIONS BINOMIALES 5 de I ′ . Nous allons renuméroter les Ql de telle sorte que Ql ⊂ I pour 1 ≤ l ≤ s et Ql 6⊂ I pour l > s. Or Ik[t, x]I = I ′ k[t, x]I et Q1 k[t, x]I ∩ · · · ∩ Qs k[t, x]I est une décomposition primaire réduite de Ik[t, x]I (cf. Theorem 17, chapter 4 [Z-S]), donc s = 1 et I = Q1 . Soit J ′ := Q2 ∩ · · · ∩ Qq si q 6= 1 et J ′ = A si q = 1. On pose alors J = J ′ si I 6⊂ J ′ et J := A si I ⊂ J ′ . On a donc I ′ = I ∩ J et J 6⊂ I. Chaque idéal Ql est engendré par des polynômes de degré inférieur ou égal à λ2 (n + 1, d). Donc J ′ est engendré par des polynômes de degré inférieur ou égal à (n + 1)((q − 2)λ2 (n + √ n 1, d))2 + λ2 (n + 1, d). On a δ ∈ I + J (cf. lemma 7 [Wa]). Notons n n+2 . e′ := (n + 3)2 1 + λ2 (n + 1, d) + (n + 1)((dn+1 − 2)λ2 (n + 1, d))2 ′ Alors δ e ∈ I + J car q ≤ dn+1 . Soit α := (e + 1)′ βk+1 (n, k(d − 1), i) + 1 Soit x(t) ∈ k[[t]]n tel que f (t, x(t)) ∈ (t)α pour tout f ∈ I. Cas 1 : Si δ(t, x(t)) ∈ (t)βk+1 (n,k(d−1),i) , comme α ≥ βk+1 (n, k(d − 1), i), par définition de βk+1 , il existe x(t) ∈ k[[t]]n tel que f (t, x(t)) = δ(t, x(t)) = 0 pour tout f ∈ I et x(t) − x(t) ∈ (t)i , et on a la conclusion voulue. Cas 2 : Supposons maintenant que δ(t, x(t)) ∈ / (t)βk+1 (m,k(d−1),i) . En dérivant la relation ′ fr (t, x(t)) ∈ (t)(e +1)βk+1 (n,k(d−1),i)+1 par rapport à t, on obtient X ∂fr ′ ∂fr ∂xλ (t) (t, x(t)) (t, x(t)) = − mod. (t)(e +1)βk+1 (n,k(d−1),i) . ∂t ∂xλ ∂t λ On en déduit l’existence d’un mineur d’ordre k de la matrice jacobienne de f , encore noté ∂(f1 ,...,fk ) δ, qui ne fait intervenir que des dérivées partielles par rapport aux xλ (i.e. δ = ∂(x 1 ,...,xk ) quitte à renommer les variables xλ ), tel que δ(t, x(t)) ∈ / (t)βk+1 (n,k(d−1),i) . En particulier δ∈ / (f1 , ..., fp ). Par définition de α et en utilisant le théorème des fonctions implicites de Tougeron (c.f. théorème 3.2 [To] ou Lemma 5.11 [Ar]), il existe x(t) ∈ k[[t]]n tel que fr (t, x(t)) = 0, ′ pour 1 ≤ r ≤ k, et x(t) − x(t) ∈ (t)e βk+1 (n,k(d−1),i) . On a alors δ(t, x(t)) − δ(t, x(t)) ∈ ′ (t)e βk+1 (n,k(d−1),i) et e′ ≥ 1. Donc δ(t, x(t)) ∈ / (t)βk+1 (n,k(d−1),i) . Pp ′ ′ Or δ e ∈ I + J, donc δ e = r=1 hr fr + h0 où h0 ∈ J. Comme fr (t, x(t)) − fr (t, x(t)) ∈ ′ (t)e βk+1 (n,k(d−1),i) et fr (t, x(t)) ∈ (t)α , pour 1 ≤ r ≤ p, on voit que ′ fr (t, x(t)) ∈ (t)e βk+1 (n,k(d−1),i) pour 1 ≤ r ≤ p ′ (car α ≥ e′ βk+1 (n, k(d − 1), i)). Donc nécessairement h0 (t, x(t)) ∈ / (t)e βk+1 (n,k(d−1),i) , d’où ′ h0 (t, x(t)) 6= 0. Or h0 fr ∈ I ∩J = I = (f1 , ..., fk ) pour 1 ≤ r ≤ p. On en déduit fr (t, x(t)) = 0 pour 1 ≤ r ≤ p. Comme βk+1 (n, k(d − 1), i) ≥ i, on a x(t) − x(t) ∈ (t)i ce qui conclut la démonstration. 6 GUILLAUME ROND 4. Fonction de Artin d’un idéal binomial Pour étudier la fonction de Artin d’un idéal binomial, nous allons mettre les xi (t, z) sous forme de Weierstrass et nous allons utiliser le lemme suivant : Lemme 4.1. Soit k un corps quelconque. Soient P = u(x)(a0 (e x)+a1 (e x)xn +· · ·+ad−1 (e x)xd−1 n + d e−1 e xn ) et Q = v(x)(b0 (e x) + b1 (e x)xn + · · · + be−1 (e x)xn + xn ) ∈ k[[x1 , ..., xn ]] deux polynômes de Weierstrass en xn (avec x = (x1 , ..., xn ) et x e := (x1 , ..., xn−1 )) tels que P − Q ∈ (x)i avec i−d i > d. Alors d = e, u(x)−v(x) ∈ (x) et aj (e x)−bj (e x) ∈ (e x)i−d+ord(P )−j pour 0 ≤ j ≤ d−1. Démonstration. Tout d’abord comme i > d, on voit que e = d car (P −Q)(0, ..., 0, xn ) ∈ (xn )i et le terme constant de u est égal au terme constant de v. De même ord(P ) = ord(Q). D’autre part, si P − Q ∈ (x)i alors u−1 P − u−1 Q ∈ (x)i . On peut donc supposer que u = 1. La division de Weirstrass de xdn par P par rapport à la variable xn est la suivante : x)xd−1 − · · · − a0 (e x)). xdn = 1.P + R(x) = P + (−ad−1 (e n Considérons la division de Weierstrass de xdn par Q par rapport à la variable xn : xdn = C(x)Q + R′ (x) = C(x)Q + Dd−1 (e x)xd−1 + · · · + D0 (e x). n Par unicité dans le théorème de division de Weierstrass, on a C(x) = v −1 (x) et Dj (e x) = −bj (e x) pour 0 ≤ j ≤ d − 1. Cette division peut se faire de manière algorithmique. En effet on construit les suites (Ck (x))k et (Rk (x))k par récurrence de la manière qui suit : on pose xdn = C0 (x)Q + R0 avec C0 (x) = 1. Puis par induction, pour k ≥ 0, si xdn = Ck (x)Q+Rk (x), on considère le plus petit monôme de Rk divisible par xdn , noté Mk , et on pose k Q. Alors la suite (ord(Rk+1 (x) − Rk (x)))k est strictement croissante Rk+1 (x) := Rk (x) − M xd n et (Rk (x))k converge vers R′ (x) pourP la topologie (x)-adique. De même la suite (deg (Mk ))k k est strictement croisssante et C(x) = k M xd . En particulier, comme R0 (x)−R(x) = Q−P ∈ n (x)i , M0 est un monôme de degré supérieur ou égal à i, et C(x) − 1 ∈ (x)i−d . On en déduit que R(x) − R′ (x) ∈ (x)i , donc aj (e x) − bj (e x) ∈ (x)i−d+ord(P )−j , pour 0 ≤ j ≤ d − 1. Nous allons maintenant étudier la fonction de Artin d’un idéal binomial. Soit k un corps algébriquement clos et soit I un idéal de k[[t, z]][X1 , ..., Xn ] engendré par f1 ,..., fp avec fk := ak X αk + bk X βk où ak , bk ∈ k et αk , βk ∈ Nn pour 1 ≤ k ≤ p. Considérons xj (t, z) ∈ k[[t, z]], 1 ≤ j ≤ n, tels que fk (xj (t, z)) ∈ (t, z)i , 1 ≤ k ≤ p. On peut supposer, quitte à faire un changement linéaire de coordonnées en t et z, que les xj sont régulières en la variable z. On a donc xj (t, z) = uj (t, z)(xj,0 (t) + xj,1 (t)z + · · · + xj,dj −1 (t)z dj −1 + z dj ) où dj est l’ordre de P xj (t, z) avec xj,l ∈ (t)k[[t]], pour 1 ≤ j ≤ n et 0 ≤ l ≤ dj . Supposons que i > j αk,j dj pour tout k. En utilisant le lemme 4.1, on a alors X X βk,j dj , 1 ≤ k ≤ p, αk,j dj = Dk := j j fk (u1 (t, z), ..., un (t, z)) ∈ (t, z)i−Dk , 1 ≤ k ≤ p et Pd (xj,l (t)) ∈ (t)i−d où Pd (Xj,l ) ∈ k[Xj,l , 1≤j≤n, 0≤l≤dj −1 ] est le coefficient de z d dans Y βk,j αk,j Y Xj,0 + · · · + Xj,dj −1 z dj −1 + z dj Xj,0 + · · · + Xj,dj −1 z dj −1 + z dj + j j BORNES EFFECTIVES DES FONCTIONS D’APPROXIMATION DES SOLUTIONS FORMELLES D’ÉQUATIONS BINOMIALES 7 Notons D := maxk Dk . On obtient alors deux systèmes d’équations indépendants l’un de l’autre : (1) fk (u1 (t, z), ..., un (t, z)) ∈ (t, z)i−D , 1 ≤ k ≤ p (2) Pd (xj,l (t)) ∈ (t)i−D , 0 ≤ d ≤ D − 1. On va chercher la fonction d’approximation de Artin de ces deux systèmes, le premier ayant une fonction de Artin bornée par une fonction affine puisque ui (0, 0) 6= 0 pour tout i et que le lieu singulier d’une variété torique est inclus dans l’union des axes de coordonnées, et le second étant un système à coefficients dans k[[t]]. Théorème 4.2. Soit k un corps algébriquement clos de caractéristique nulle. Alors on a les propriétés suivantes : i) Pour tout d′ ∈ N et pour tout d := (d1 , ...., dn ) ∈ Nn , il existe ad,d′ , bd,d′ vérifiant la propriété suivante : Soit I un idéal binomial de k[U1 , ..., Un ] engendré par des binômes f1 , ..., fp de degré inférieur à d′ . Soit i ∈ N et soient x1 (t, z),..., xn (t, z) ∈ k[[t, z]] tels que ord(xj (t, z)) = dj et fk (xj (t, z)) ∈ (t, z)ad,d′ i+bd,d′ , pour tous k. Alors il existe xj (t, z) ∈ k[[t, z]], tels que fk (xj (t, z)) = 0 pour tout k et tels que xj (t, z) − xj (t, z) ∈ (t, z)i pour tout j. ii) Pour tout d′ ∈ N il existe une fonction doublement exponentielle en i, notée βd′ , telle que pour tout idéal binomial I de k[U1 , ..., Un ] engendré par des binômes de degré inférieur à d′ , la fonction de Artin de Ik[[t, z]][U ] est bornée par βd′ . Démonstration. Nous allons d’abord montrer i) dont on déduira ensuite ii). Supposons, comme précédemment, que I est engendré par les fk := ak X αk + bk X βk , où ak , bk ∈ k et αk , βk ∈ Nn pour 1 ≤ √ k ≤ p. √ L’idéal I étant un idéal binomial, I est encore un idéal binomial et, si I = I1 ∩ · · · ∩ Iq est √ I, alors les I sont des idéaux binomiaux [E-S]. une décomposition primaire minimale de k √ e Soit e ∈ N tel que I ⊂ I. Supposons que fl (u1 (t, z), ..., un (t, z)) ∈ (t, z)qei , 1 ≤ l ≤ √p où uj,0 := uj (0, 0) 6= 0 pour tout j. Alors g(u1 (t, z), ..., un (t, z)) ∈ (t, z)qi pour tout g ∈ I, et donc il existe un entier k tel que g(u1 (t, z), ..., un (t, z)) ∈ (t, z)i pour tout g ∈ Ik . Comme Ik est un idéal premier binomial de k[U ] alors (k[U ]/Ik )(U1 −u1,0 ,...,Un −un,0 ) est régulier car u1,0 ...un,0 6= 0 (le lieu singulier d’une variété torique est toujours inclus dans l’union des hyperplans de coordonnées). Donc k −→ (k[U ]/Ik )(U1 −u1,0 ,...,Un −un,0 ) est lisse, donc k[[t, z]] −→ (k[[t, z]][U ]/Ik )(t,z,U1 −u1,0 ,...,Un−un,0 ) est lisse ; ceci implique qu’il existe uj (t, z) ∈ k[[t, z]], 1 ≤ j ≤ n, tels que g(u1 (t, z), ..., un (t, z)) = 0 pour tout g ∈ Ik et uj (t, z) − uj (t, z) ∈ (t, z)i . Soit i −→ ai + b la fonction de Artin de l’idéal de k[[t]][Xj,l ] engendré par les Pd (Xj,l ), 0 ≤ d ≤ D − 1. Posons ad := max{qe, a} et bd := b + D. D’après les théorèmes 2.1 et 3.1, on voit que ad et bd peuvent être bornées par une fonction dépendant uniquement de d′ . Soit i ∈ N, on a alors ad i + bd > D. D’après ce qui précède, si fk (xj (t, z)) ∈ (t, z)ad i+bd ∀k, alors fk (u1 (t, z), ..., un (t, z)) ∈ (t, z)qei , 1 ≤ k ≤ p 8 GUILLAUME ROND et Pd (xj,l (t)) ∈ (t)ai+b , 0 ≤ d ≤ D. Donc il existe uj (t, z) ∈ k[[t, z]], 1 ≤ j ≤ n, tels que g(u1 (t, z), ..., un (t, z)) = 0 pour tout g ∈ I et uj (t, z) − uj (t, z) ∈ (t, z)i , et il existe xj,l (t) ∈ k[[t]] tels que Pd (xj,l (t)) = 0, 0 ≤ d ≤ D, et xj,l (t) − xj,l (t) ∈ (t)i . On pose alors xj = uj (t, z)(xj, 0 + xj,1 t + · · · + xj,dj −1 tdj −1 + tdj ). On a bien fk (xj (t, z)) = 0, 1 ≤ k ≤ p, et xj (t, z)− xj (t, z) ∈ (t, z)i , 1 ≤ j ≤ n. Ceci prouve i). On peut remarquer que aPet b sont bornées par une fonction polynomiale en maxk {\|αk \|, \|βk \|} de degré exponentiel en j dj , d’après le théorème 3.1. Donc il existe une constante C telle P j dj . Notons aEd et bEd les plus petites constantes satisfaisant i) pour l’idéal que ad , bd ≤ C C engendré par les fk , k ∈ E, Poù E est un sous-ensemble de {1, ..., p}. Là encore il existe une C constante CE telle que CE E aEd , bEd ≤ C C P j dj j∈E dj majore aEd et bEd . Donc en posant C := maxE CE on a pour tout sous-ensemble E de {1, ..., p}. C ni Soit xj (t, z) ∈ k[[t, z]], 1 ≤ j ≤ n, tels que fk (xj (t, z)) ∈ (t, z)C (i+1) , 1 ≤ k ≤ p. Si dj := ord(xj (t, z)) ≥ i, alors on pose xj (t, z) = 0, sinon on pose xj (t, z) = xj (t, z). Alors fk (xj (t, z)) = 0 ou fk (xj (t, z)) = fk (xj (t, z)) selon l’entier k. On peut donc remplacer les xj (t, z) par les xj (t, z) et supposer que ord(xj (t, z)) < i pour tout j. On supposer que P P ni C C (i+1) > D := maxk { j αk,j dj , j βk,j dj } en choisissant i assez grand. Comme dj < i, 1 ≤ j ≤ n, on a fk (xj (t, z)) ∈ (t, z)ad i+bd , 1 ≤ k ≤ p. On applique alors i), et on a l’existence de xj (t, z) ∈ k[[t, z]] tels que fk (xj (t, z)) = 0, 1 ≤ k ≤ p, et xj (t, z) − xj (t, z) ∈ (t, z)i , 1 ≤ j ≤ n. Ceci prouve ii). Exemple 4.3. Dans [Ro2], il est montré que la fonction de Artin du polynôme X 2 − ZY 2 n’est pas bornée par une fonction affine. On voit que la non-linéarité de la fonction de Artin dans ce cas provient du fait que les fonctions de Artin-Greenberg des systèmes Pd (Xj,l ) = 0, 0 ≤ d ≤ D, sont bornées par des fonctions affines dont les coefficients croissent vite en fonction de l’ordre des xj,l (t). Exemple 4.4. On peut remarquer que la famille de solutions approchées de l’équation X 2 − Y 3 dans [Ro1] sont des solutions dont l’ordre est borné. Exemple 4.5. Soit f := X 2 − Y 3 . Si x(t, z), y(t, z) ∈ k[[t, z]] vérifient ord(x(t, z)) = 3 et ord(y(t, z)) = 2, on peut écrire x(t, z) = u(t, z)(x0 (t) + x1 (t)z + x2 (t)z 2 + z 3 ) y(t, z) = v(t, z)(y0 (t) + y1 (t)z + z 2 ). Si x2 (t, z) − y 3 (t, z) ∈ (t, z)i avec i > 6, alors u2 (t, z) − v 3 (t, z) ∈ (t, z)i−6 et les xj (t) et yj (t) sont solutions du système suivant modulo (t)i−6 :  2x2 − 3y1 = 0    2 2  x + 2x  1 − 3y1 − 3y0 = 0 2   3 2x0 + 2x1 x2 − y1 − 6y0 y1 = 0 x21 + 2x0 x2 − 3y0 y12 − 3y02 = 0     2x0 x1 − 3y02 y1 = 0    x20 − y03 = 0 On peut vérifier à l’aide de Macaulay2 [M2] que l’idéal de k[x0 , x1 , x2 , y0 , y1 ] défini par ces polynômes n’est pas réduit. Ceci semble assez général. On peut remarquer que les variétés BORNES EFFECTIVES DES FONCTIONS D’APPROXIMATION DES SOLUTIONS FORMELLES D’ÉQUATIONS BINOMIALES 9 algébriques définies par Pd (Xj,l ) = 0, 0 ≤ d ≤ D, sont très proches des espaces de jets de la variété définie par I qui ne sont pas réduits en général. Exemple 4.6. On considère ici le problème suivant : soient p et q deux entiers premiers entre eux. Si x(t, z) est une série formelle telle que xp (t, z) est proche d’une puissance qième, est-ce que x(t, z) est proche d’une puissance q-ième ? Y a-t-il une fonction qui mesure le rapport entre la distance de xp (t, z) à une puissance q-ième et celle de x(t, z) à une puissance q-ième ? On a la réponse suivante : Proposition 4.7. Soient p et q premiers entre eux. Il existe une fonction β : N −→ N telle que pour tout x ∈ k[[t, z]], si il existe u(t, z) ∈ k[[t, z]] telle que xp (t, z) − uq (t, z) ∈ (t, z)β(i) , alors il existe v(t, z) ∈ k[[t, z]] telle que x(t, z) − v q (t, z) ∈ (t, z)i . On peut choisir pour β une fonction doublement exponentielle en i. Pour tout d ∈ N, si on se restreint à tous les x(t, z) dont l’ordre vaut d, alors on peut choisir pour β une fonction affine. Démonstration. Soit β la fonction de Artin de X p − Y q . Supposons que xp (t, z) − uq (t, z) ∈ (t, z)β(i) . Alors il existe x′ (t, z), u′ (t, z) ∈ k[[t, z]] tels que x′p (t, z) − u′q (t, z) = 0 et x(t, z) − x′ t(t, z), u(t, z) − u′ (t, z) ∈ (t, z)i . Comme p et q sont premiers entre eux et que k[[t, z]] est factoriel il existe v(t, z) ∈ k[[t, z]] tel que x′ (t, z) = v q (t, z) et u′ (t, z) = v p (t, z). On a alors le résultat avec le théorème 4.2. Références [Ar] M. Artin, Algebraic approximation of structures over complete local rings, Publ. Math. IHES, 36, (1969), 23-58. [B-D-L-v.D] J. Becker, J. Denef, L. Lipshitz, L. van den Dries, Ultraproducts and Approximation in Local Rings I, Inventiones math., 51, 189-203 (1979). [E-S] D. Eisenbud, B. Sturmfels, Binomial ideals, Duke Math. J., 84, (1996), no. 1, 1-45. [Gr] M. J. Greenberg, Rational points in henselian discrete valuation rings, Publ. Math. IHES, 31, (1966), 59-64. [He] G. Hermann, Die Frage der endlich vielen Schritte in der Theorie der Polynomideale, Math. Ann., 95, (1926), no. 1, 736-788. [La] D. Lascar, Caractère effectif des théorèmes d’approximation d’Artin, C. R. Acad. Sci. Paris Sér. A-B, 287, (1978), no. 14, A907-A910. [M2] D. R. Grayson, M. E. Stillman, Macaulay 2, a software system for research in algebraic geometry, Available at ttp ://www.math.uiuc.edu/Macaulay2/ [Ro1] G. Rond, À propos de la fonction de Artin en dimension N ≥ 2, C. R. Math. Acad. Sci. Paris, 340, no. 8, (2005), 577-580. [Ro2] G. Rond, Sur la linéarité de la fonction de Artin, Annales Scientifiques de l’École Normale Supérieure, vol. 38, no. 6, (2005), 979-988. [Ro3] G. Rond, Lemme d’Artin-Rees, théorème d’Izumi et fonction de Artin, Journal of Algebra, vol. 299, no. 1, (2006), 245-275. [S] A. Seidenberg, Constructions in Algebra, Trans. A.M.S., 197, (oct 1974) ; 273-313. [Te] B. Teissier, Résultats récents d’algèbre commutative effective, Séminaire Bourbaki, Vol. 1989/90, Astérisque, No. 189-190, (1990), Exp. No. 718, 107-131. [To] J.-C. Tougeron, Idéaux de fonctions différentiables, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 71. Springer-Verlag, Berlin-New York, (1972). [Wa] J. J. Wavrik, A theorem on solutions of analytic equations with applications to deformations of complex structures, Math. Ann., 216, (1975), 127-142. [Z-S] O. Zariski, P. Samuel, Commutative Alegbra I, D. Van Nostrand Company, Inc., Princeton, New Jersey, 1958. 10 GUILLAUME ROND IML, Campus de Luminy, Case 907, 13288 Marseille Cedex 9, France E-mail address: rond@iml.univ-mrs.fr	0math.AC
Estimation in the convolution structure density model. Part II: adaptation over the scale of anisotropic classes. arXiv:1704.04420v1 [math.ST] 14 Apr 2017 O.V. Lepski ∗ T. Willer ∗ Institut de Mathématique de Marseille Aix-Marseille Université 39, rue F. Joliot-Curie 13453 Marseille, France e-mail: oleg.lepski@univ-amu.fr e-mail: thomas.willer@univ-amu.fr Abstract: This paper continues the research started in Lepski and Willer (2016). In the framework of the convolution structure density model on Rd , we address the problem of adaptive minimax estimation with Lp –loss over the scale of anisotropic Nikol’skii classes. We fully characterize the behavior of the minimax risk for different relationships between regularity parameters and norm indexes in the definitions of the functional class and of the risk. In particular, we show that the boundedness of the function to be estimated leads to an essential improvement of the asymptotic of the minimax risk. We prove that the selection rule proposed in Part I leads to the construction of an optimally or nearly optimally (up to logarithmic factor) adaptive estimator. AMS 2000 subject classifications: 62G05, 62G20. Keywords and phrases: deconvolution model, density estimation, oracle inequality, adaptive estimation, kernel estimators, Lp –risk, anisotropic Nikol’skii class. 1. Introduction In the present paper we will be interested in the adaptive estimation in the convolution structure density model. Our considerations here continue the research started in Lepski and Willer (2016). Thus, we observe i.i.d. vectors Zi ∈ Rd , i = 1, . . . , n, with a common probability density p satisfying the following structural assumption p = (1 − α)f + α[f ⋆ g], f ∈ Fg (R), α ∈ [0, 1], (1.1) where α ∈ [0, 1] and g : Rd → R are supposed to be known and f : Rd → R is the function to be estimated. Recall that for two functions f, g ∈ L1 Rd Z f (x − z)g(z)νd (dz), x ∈ Rd , f ⋆ g (x) = Rd and for any α ∈ [0, 1], g ∈ L1 Rd and R > 1, n o Fg (R) = f ∈ B1,d (R) : (1 − α)f + α[f ⋆ g] ∈ P Rd . ∗ This work has been carried out in the framework of the Labex Archimède (ANR-11-LABX-0033) and of the A*MIDEX project (ANR-11-IDEX-0001-02), funded by the ”Investissements d’Avenir” French Government program managed by the French National Research Agency (ANR). 1 imsart-generic ver. 2014/10/16 file: DSDM-adap-ArXivl.tex date: April 17, 2017 Furthermore P Rd denotes the set of probability densities on Rd , Bs,d (R) is the ball of radius R > 0 in Ls Rd := Ls Rd , νd , 1 ≤ s ≤ ∞ and νd is the Lebesgue measure on Rd . At last, for any Pd R U ∈ L1 Rd let Ǔ (t) := Rd U (x)e−i j=1 xj tj νd (dx), t ∈ Rd , be the Fourier transform of U . The convolution structure density model (1.1) will be studied for an arbitrary g ∈ L1 Rd and f ∈ Fg (R). Then, except in the case α = 0, the function f is not necessarily a probability density. We want to estimate f using the observations Z (n) = (Z1 , . . . , Zn ). By estimator, we mean any d (n) n Z -measurable map fˆ : R → Lp R . The accuracy of an estimator fˆ is measured by the Lp –risk 1/p , p ∈ [1, ∞), Rn(p) [fˆ, f ] := Ef kfˆ − f kpp where Ef denotes the expectation with respect to the probability measure Pf of the observations Z (n) = (Z1 , . . . , Zn ). Also, k · kp , p ∈ [1, ∞), is the Lp -norm on Rd . The objective is to construct an estimator of f with a small Lp –risk. 1.1. Adaptive estimation (p) Let F be a given subset of Lp Rd . For any estimator f˜n define its maximal risk by Rn f˜n ; F = (p) supf ∈F Rn f˜n ; f and its minimax risk on F is given by φn (F) := inf Rn(p) f˜n ; F . f˜n (1.2) Here, the infimum is taken over all possible estimators. An estimator whose maximal risk is bounded by φn (F) up to some constant factor is called minimax on F. Let Fϑ , ϑ ∈ Θ be a collection of subsets of Lp Rd , νd , where ϑ is a nuisance parameter which may have a very complicated structure. The problem of adaptive estimation can be formulated as follows: is it possible to construct a single estimator fˆn which would be simultaneously minimax on each class Fϑ , ϑ ∈ Θ, i.e. (p) ˆ lim sup φ−1 n (Fϑ )Rn fn ; Fϑ < ∞, ∀ϑ ∈ Θ? n→∞ We refer to this question as the problem of minimax adaptive estimation over the scale {Fϑ , ϑ ∈ Θ}. If such an estimator exists, we will call it optimally adaptive. Using the modern statistical language we call the estimator fˆn nearly optimally adaptive if (p) ˆ fn ; Fϑ < ∞, ∀ϑ ∈ Θ. lim sup φ−1 n (Fϑ )Rn n→∞ ln n We will be interested in adaptive estimation over the scale ~ L ~ ~r, L, ~ ∩ Fg,∞ (R, Q), ϑ = β, ~ R, Q , Fϑ = N~r,d β, ~ L ~ is the anisotropic Nikolskii class, see Definition 1 below. As it was explained in where N~r,d β, ~ L ~ ~r, L ~ , β, ~ ∈ (0, ∞)d × [1, ∞]d × (0, ∞)d Part I, the adaptive estimation over the scale N~r,d β, can be viewed as the adaptation to anisotropy and inhomogeneity of the function to be estimated. Recall also that Fg,∞(R, Q) := f ∈ Fg (R) : (1 − α)f + α[f ⋆ g] ∈ B∞,d (Q) , 2 so f ∈ Fg,∞ (R, Q) simply means that the common density of observations p is uniformly bounded by Q. It is easy to see that if α = 1 and kgk∞ < ∞ then Fg,∞ (R, Q) = Fg (R) for any Q ≥ Rkgk∞ . Let us briefly discuss another example. Let r > 1 and L < ∞ be arbitrary but a priory chosen numbers. Assume that the considered collection of anisotropic Nikol’skii classes obeys the following ~ ∈ (0, L]d . Suppose also that kgks < ∞, where 1/s = 1 − 1/r. restrictions: ~r ∈ [r, ∞]d and L ~ L ~ ∩ Fg,∞ (R, Q) = Then, there exists Q0 completely determined by r, L and R such that N~r,d β, ~ L ~ ∩ Fg (R) for any Q > Q0 kgks . N~r,d β, Additionally, we will study the adaptive estimation over the collection ~ L ~ ~r, L, ~ ∩ Fg (R) ∩ B∞,d (Q), ϑ = β, ~ R, Q . Fϑ = N~r,d β, We will show that the boundedness of the underlying function allows to improve considerably the accuracy of estimation. 1.2. Historical notes The minimax adaptive estimation is a very active area of mathematical statistics and the interested reader can find a very detailed overview as well as several open problems in adaptive estimation in the recent paper, Lepski (2015). Below we will discuss only the articles whose results are relevant to our consideration, i.e. the density setting under Lp -loss, from a minimax perspective. Let us start with the following remark. If one assumes additionally that f, g ∈ P Rd the convolution structure density model can be interpreted as follows. The observations Zi ∈ Rd , i = 1, . . . , n, can be written as a sum of two independent random vectors, that is, Zi = Xi + ǫi Yi , i = 1, . . . , n, (1.3) where Xi , i = 1, . . . , n, are i.i.d. d-dimensional random vectors with common density f to be estimated. The noise variables Yi , i = 1, . . . , n, are i.i.d. d-dimensional random vectors with known common density g. At last εi ∈ {0, 1}, i = 1, . . . , n, are i.i.d. Bernoulli random variables with P(ε1 = 1) = α, where α ∈ [0, 1] is supposed to be known. The sequences {Xi , i = 1, . . . , n}, {Yi , i = 1, . . . , n} and {ǫi , i = 1, . . . , n} are supposed to be mutually independent. The observation scheme (1.3) can be viewed as the generalization of two classical statistical models. Indeed, the case α = 1 corresponds to the standard deconvolution model Zi = Xi + Yi , i = 1, . . . , n. Another ”extreme” case α = 0 correspond to the direct observation scheme Zi = Xi , i = 1, . . . , n. The ”intermediate” case α ∈ (0, 1), considered for the first time in Hesse (1995), is understood as partially contaminated observations. Direct case, α = 0 There is a vast literature dealing with minimax and minimax adaptive density estimation, see for example, Efroimovich (1986), Hasminskii and Ibragimov (1990), Golubev (1992), Donoho et al. (1996), Devroye and Lugosi (1997), Rigollet (2006), Rigollet and Tsybakov (2007), Samarov and Tsybakov (2007), Birgé (2008), Giné and Nickl (2009), Akakpo (2012), Gach et al. (2013), Lepski (2013), among many others. Special attention was paid to the estimation of densities with unbounded support, see Juditsky and Lambert–Lacroix (2004), Reynaud–Bouret et al. (2011). The most developed results can be found in Goldenshluger and Lepski (2011), Goldenshluger and Lepski (2014) and in Section 2 we will compare in detail our results with those obtained in these papers. 3 Intermediate case, α ∈ (0, 1) To the best of our knowledge, adaptive estimation in the case of partially contaminated observations has not been studied yet. We were able to find only two papers dealing with minimax estimation. The first one is Hesse (1995) (where the discussed model was introduced in dimension 1) in which the author evaluated the L∞ -risk of the proposed estimator over a functional class formally corresponding to the Nikol’skii class N∞,1 (2, 1). In Yuana and Chenb (2002) the latter result was developed to the multidimensional setting, i.e. to the minimax es~ timation on N∞,d 2, 1 . The most intriguing fact is that the accuracy of estimation in partially contaminated noise is the same as in the direct observation scheme. However none of these articles studied the optimality of the proposed estimators. We will come back to the aforementioned papers in Section 1.3.1 in order to compare the assumptions imposed on the noise density g. Deconvolution case, α = 1 First let us remark that the behavior of the Fourier transform of the function g plays an important role in all the works dealing with deconvolution. Indeed ill-posed problems correspond to Fourier transforms decaying towards zero. Our results will be established for ”moderately” ill posed problems, so we detail only results in papers studying that type of operators. This assumption means that there exist ~µ = (µ1 , . . . , µd ) ∈ (0, ∞)d and Υ1 > 0, Υ2 > 0 such that the Fourier transform ǧ of g satisfies: Υ1 d Y (1 + j=1 t2j )− µj 2 ≤ ǧ(t) ≤ Υ2 d Y (1 + t2j )− j=1 µj 2 , ∀t = (t1 , . . . , td ) ∈ Rd . (1.4) Some minimax and minimax adaptive results in dimension 1 over different classes of smooth functions can be found in particular in Stefanski and Carroll (1990), Fan (1991), Fan (1993), Pensky and Vidakovic (1999), Fan and Koo (2002), Comte and al. (2006), Hall and Meister (2007), Meister (2009), Lounici and Nickl (2011), Kerkyacharian et al. (2011). There are very few results in the multidimensional setting. It seems that Masry (1993) was the first paper where the deconvolution problem was studied for multivariate densities. It is worth noting that Masry (1993) considered more general weakly dependent observations and this paper formally does not deal with the minimax setting. However the results obtained in this paper could be formally compared with the estimation under L∞ -loss over the isotropic Hölder class of regularity 2, i.e. N∞,d ~2, 1 which is exactly the same setting as in Yuana and Chenb (2002) in the case of partially contaminated observations. Let us also remark that there is no lower bound result in Masry (1993). The most developed results in the deconvolution model were obtained in Comte and Lacour (2013) and Rebelles (2016) and in Section 2 we will compare in detail our results with those obtained in these papers. 1.3. Lower bound for the minimax Lp -risk We have seen that the problem of optimal adaptation over the collection Fϑ , ϑ ∈ Θ is formulated as the ”attainability” of the family of minimax risks φn (Fϑ ), ϑ ∈ Θ by a single estimator. Although it is not necessary, the following ”two-stage” approach is used for the majority of problems related to the minimax adaptive estimation. The first step consists in finding a lower bound for φn (Fϑ ) for any ϑ ∈ Θ while the second one consists in constructing an estimator ”attaining”, at least asymptotically, this bound. We adopt this strategy in our investigations and below we present several lower bound results recently obtained in Lepski and Willer (2017). 4 1.3.1. Assumptions on the function g imposed in Lepski and Willer (2017) Let J∗ denote the set of all subsets of {1, . . . , d}. Set J = J∗ ∪ ∅ and for any J ∈ J let \|J\| denote the cardinality of J while {j1 < · · · < j\|J\| } denotes its elements. \|J \| For any J ∈ J∗ define the operator DJ = ∂tj ∂···∂tj and let D∅ denote the identity operator. For 1 \|J \| any I, J ∈ J define DI,J = DI DJ and note that obviously DI,J = DJ,I . Assumption 1 (α 6= 1). DJ ǧ exists for any J ∈ J∗ and supJ∈J∗ DJ ǧ ∞ < ∞; Assumption 2 (α = 1). DJ ǧ exists for any J ∈ J∗ and supJ∈J∗ ǧ −1 DJ ǧ there exists ~ µ = (µ1 , . . . , µd ) ∈ (0, ∞)d and Υ > 0 such that µj Q \|ǧ(t)\| ≤ Υ dj=1 (1 + t2j )− 2 , ∀t = (t1 , . . . , td ) ∈ Rd . Assumption 3 (α = 1). g is a bounded function. Assumption 4 (α = 1). DI,J ǧ exists for any I, J ∈ J and supI,J∈J DI,J ǧ Z Y g(z) zj2 dz < ∞. sup J∈J∗ Rd ∞ 1 < ∞. Moreover, < ∞. Moreover j∈J It is worth noting that all the bounds in Lepski and Willer (2017) are obtained under Assumptions 1 and 2. Assumption 3 is used when the estimation of unbounded functions is considered; we come back to this assumption in Section 2.4.2. As to Assumption 4, it seems purely technical and does not appear in upper bound results. We also recall that the lower bounds in Lepski and Willer (2017) are proved under the condition: d g∈P R . 1.3.2. Some lower bounds from Lepski and Willer (2017) ~ (α) = ~ ~ (α) = (0, . . . , 0), α ∈ [0, 1), and introduce for any β~ ∈ (0, ∞)d , ~r ∈ [1, ∞]d Set µ µ, α = 1, µ d ~ ∈ (0, ∞) the following quantities. and L d X 2µj (α) + 1 1 = , β(α) βj j=1 d X 2µj (α) + 1 1 = , ω(α) βj rj j=1 L(α) = d Y 2µj (α)+1 βj Lj . (1.5) j=1 Define for any 1 ≤ s ≤ ∞ and α ∈ [0, 1] κα (s) = ω(α)(2 + 1/β(α)) − s τ (s) = 1 − 1/ω(0) + 1/(sβ(0)). (1.6) General case. Remind that z(α) = ω(α)(2 + 1/β(α))β(0)τ (∞) + 1, p∗ = maxl=1,...,d rl ∨ p. Set ̺(α) =                1−1/p 1−1/ω(α)+1/β(α) , β(α) 2β(α)+1 , τ (p)ω(α)β(0) , z(α) ∗ ω(α)(1−p /p) , κα (p∗ ) 5 κα (p) > pω(α); 0 < κα (p) ≤ pω(α); κα (p) ≤ 0, τ (p∗ ) > 0; κα (p) ≤ 0, τ (p∗ ) ≤ 0. (1.7) ∗ /p) Here and later we assume 0/0 = 0, which implies in particular that ω(α)(1−p = 0 if p∗ = p and κα (p∗ ) κα (p) = 0. Recall also that κα (p∗ )/p∗ = −1 if p∗ = ∞. Put at last  L(α)n−1 , κα (p) > 0;    L(α)n−1 ln(n), κα (p) ≤ 0, τ (p∗ ) ≤ 0; δn =   κα (p)  − κα (p) ≤ 0, τ (p∗ ) > 0. [L(0)] ω(α)pτ (p) L(α)n−1 ln(n), Theorem 1 (Lepski and Willer (2017)). Let L0 > 0 and 1 ≤ p < ∞ be fixed. ~ ∈ [L0 , ∞)d , ~µ ∈ (0, ∞)d , R > 1 and g ∈ P Rd , Then for any β~ ∈ (0, ∞)d , ~r ∈ [1, ∞]d , L ~ such that satisfying Assumptions 1–4, there exists c > 0 independent of L lim inf inf n→∞ sup f˜n f ∈N ~~ ~ r ,d β,L ∩Fg (R) δ n−̺(α) R(n) f˜n ; f ≥ c, p where the infimum is taken over all possible estimators. Following the terminology used in Lepski and Willer (2017), we will call the set of parameters satisfying κα (p) > pω(α) the tail zone, satisfying 0 < κα (p) ≤ pω(α) the dense zone and satisfying κα (p) ≤ 0 the sparse zone. In its turn, the latter zone is divided into two sub-domains: the sparse zone 1 corresponding to τ (p∗ ) > 0 and the sparse zone 2 corresponding to τ (p∗ ) ≤ 0. Bounded case. Introduce ρ(α) =                1−1/p 1−1/ω(α)+1/β(α) , β(α) 2β(α)+1 , τ (p)ω(α)β(0) , z(α) ω(α) p , κα (p) > pω(α); 0 < κα (p) ≤ pω(α); κα (p) ≤ 0, τ (∞) > 0; (1.8) κα (p) ≤ 0, τ (∞) ≤ 0. Theorem 2 (Lepski and Willer (2017)). Let L0 > 0 and 1 ≤ p < ∞ be fixed. ~ ∈ [L0 , ∞)d , Q > 0, ~µ ∈ (0, ∞)d and g ∈ P Rd , Then for any β~ ∈ (0, ∞)d , ~r ∈ [1, ∞]d , L ~ such that satisfying Assumptions 1 and 2 there exists c > 0 independent of L −ρ(α) (n) ˜ sup lim inf inf δ R f ; f ≥ c, n n p n→∞ f˜ n ~L ~ ∩P Rd ∩B∞,d (Q) f ∈N~r,d β, where the infimum is taken over all possible estimators. 1.4. Assumptions on the function g The selection rule from the family of linear estimators, the Lp -norm oracle inequalities obtained in Part I and all the adaptive results presented in the paper are established under the following condition imposed on the function g. Assumption 5. (1) if α 6= 1 then there exists ε > 0 such that 1 − α + αǧ(t) ≥ ε, ∀t ∈ Rd ; (2) if α = 1 then there exists µ ~ = (µ1 , . . . , µd ) ∈ (0, ∞)d and Υ0 > 0 such that µj Q \|ǧ(t)\| ≥ Υ0 dj=1 (1 + t2j )− 2 , ∀t = (t1 , . . . , td ) ∈ Rd . 6 Comparing this condition with Assumption 2 from Section 1.3.1, we can assert that both are coherent if α = 1. Indeed, in this case, we come the following assumption, which is well-known in the literature: µj µj Q Q Υ0 dj=1 (1 + t2j )− 2 ≤ \|ǧ(t)\| ≤ Υ dj=1 (1 + t2j )− 2 , ∀t ∈ Rd . referred to as a moderately ill-posed statistical problem, cf. (1.4). In particular, the assumption is checked for the centered multivariate Laplace law. Note first that Assumption 5 is in some sense weaker than Assumption 1 when α ∈ (0, 1), since it does not require regularity properties of the function g. Moreover both assumptions are not too restrictive. They are verified for many distributions, including centered multivariate Laplace and Gaussian ones. Note also that Assumption 5 always holds with ε = 1 − 2α if α < 1/2. Additionally, it holds with ε = 1 − α if ǧ is a real positive function. The latter is true, in particular, for any probability law obtained by an even number of convolutions of a symmetric distribution with itself. Next, our Assumption 5 is weaker than the conditions imposed in Hesse (1995) and Yuana and Chenb (2002). In these papers ǧ ∈ C(2) Rd , ǧ(t) 6= 0 for any t ∈ Rd and 1 − α + αǧ(t) ≥ 1 − α, ∀t ∈ Rd . 2. Adaptive estimation over the scale of anisotropic Nikol’skii classes We start this section by recalling the definition of the pointwise selection rule proposed in Part I. 2.1. Pointwise selection rule R Let K : Rd → R be a continuous function belonging to L1 Rd such that R K = 1. Set k d e , k ∈ Z and let Hd = ~h = (h1 , . . . , hd ) : hj ∈ H, j = 1, . . . , d . Recall that Hisotr = Q d d d ~ ~ H : h = (h, . . . , h), h ∈ H . Set V~h = j=1 hj and let for any h ∈ H K~h (t) = V~−1 K t1 /h1 , . . . , td /hd , t ∈ Rd . H = ~h ∈ h Later on for any u, v ∈ Rd the operations and relations u/v, uv, u ∨ v,u ∧ v, u ≥ v, au, a ∈ R, are understood in coordinate-wise sense. In particular u ≥ v means that uj ≥ vj for any j = 1, . . . , d. For any ~h ∈ (0, ∞)d let M ·, ~h satisfy the operator equation Z ~ g(t − y)M t, ~h dt, y ∈ Rd . (2.1) K~h (y) = (1 − α)M y, h + α Rd Introduce for any ~h ∈ Hd and x ∈ Rd fb~h (x) = n−1 n X i=1 bn x, ~h = U where M∞ = (2π)−d ε−1 Ǩ n 1X σ b2 x, ~h = M 2 Zi − x, ~h ; n M Zi − x, ~h , s i=1 b2 x, ~h 2λn ~h σ 4M∞ λn ~h + , Q n 3n dj=1 hj (hj ∧ 1)µj (α) 1 + Υ−1 0 k1 1α=1 1 α6=1 7 ∨ 1 and d X 1 + µj (α) ln(hj ) . λn ~h = 4 ln(M∞ ) + 6 ln (n) + (8p + 26) j=1 Let H be an arbitrary subset of For any ~h ∈ H and x ∈ Rd introduce h i bn x, ~η bn x, ~h ∨ ~η − 4U b~ (x) = sup fb~ (x) − fb~η (x) − 4U ; R h∨~ η h Hd . + η ∈H ~ bn∗ x, ~h = U sup η ∈H: ~ ~ η≥~h bn x, ~η , U (2.2) (2.3) h i b~ (x) + 8U bn∗ x, ~h . and define ~h(x) = arg inf ~h∈H R h Our final estimator is fbh(x) (x), x ∈ Rd and we will call (2.2) the pointwise selection rule. ~ Remark 1. Note that the estimator fb~h depends on H and later on we will consider two choices d of the parameter set H, namely H = Hd and H = Hisotr . So, to present our results we will write b f~h,H in order to underline the aforementioned dependence. The choice H = Hd will be used when d the adaptation is studied over anisotropic Nikol’skii classes while H = Hisotr will be used when the considered scale consists of isotropic classes. 2.2. Anisotropic Nikol’skii classes Let (e1 , . . . , ed ) denote the canonical basis of Rd . For some function G : Rd → R1 and real number u ∈ R define the first order difference operator with step size u in direction of the variable xj by ∆u,j G(x) = G(x + uej ) − G(x), j = 1, . . . , d. By induction, the k-th order difference operator with step size u in direction of the variable xj is defined as k X k k−1 l+k k ∆u,j G(x) = ∆u,j ∆u,j G(x) = (−1) ∆ul,j G(x). (2.4) l l=1 ~ = Definition 1. For given vectors ~r = (r1 , . . . , rd ) ∈ [1, ∞]d β~ = (β1 , . . . , βd ) ∈ (0, ∞)d and L d d 1 (L1 , . . . , Ld ) ∈ (0, ∞) we say that a function G : R → R belongs to the anisotropic Nikolskii class ~ L ~ if N~r,d β, (i) kGkrj ≤ Lj for all j = 1, . . . , d; (ii) for every j = 1, . . . , d there exists natural number kj > βj such that k j ∆u,j G rj ≤ Lj \|u\|βj , ∀u ∈ R, ∀j = 1, . . . , d. If βj = β ∈ (0, ∞), rj = r ∈ [1, ∞] and Lj = L ∈ (0, ∞) for any j = 1, . . . , d the corresponding Nikolskii class, denoted furthermore Nr,d (β, L), is called isotropic. 8 2.3. Construction of kernel K First, we recall that all results concerning the Lp risk of the pointwise selection rule, established in Part I, are proved under the following assumption imposed on the kernel K. Assumption 6. There exist k1 > 0 and k2 > 0 such that Z Ǩ(t) Rd d Y (1 + t2j ) j=1 µj (α) 2 dt ≤ k1 , Z Rd Ǩ(t) 2 d Y (1 + t2j )µj (α) dt ≤ k22 . (2.5) j=1 use the following specific kernel K in the definition of the estimator’s family Next, we will d ~ b f~h (·), h ∈ H [see, e.g., Kerkyacharian et al. (2001) or Goldenshluger and Lepski (2014)]. Let ℓ beR an integer number, and let K : R1 → R1 be a compactly supported continuous function satisfying R1 K(y)dy = 1, and K ∈ C(R1 ). Put Kℓ (y) = ℓ X ℓ i=1 1 y , (−1)i+1 K i i i (2.6) and add the following structural condition to Assumption 6. Q Assumption 7. K(x) = dj=1 Kℓ (xj ), ∀x ∈ Rd . The kernelRK constructed in this way is bounded, compactly supported, belongs to C(Rd )∩L1 (Rd ) and satisfies Rd K = 1. Some examples of kernels satisfying simultaneously Assumptions 6 and 7 can be found for instance in Comte and Lacour (2013). 2.4. Main results Introduce the following notations: δn = L(α)n−1 ln(n) and   [ln(n)]t(H) ,     ( 1   ln p (n) ∨ [ln(n)]t(H) , d − 1, H = Hd ; t(H) = bn (H) = 1 d  0, H = Hisotr ,  ln p (n),      1, κα (p) > pω(α); κα (p) = pω(α); κα (p) = 0; otherwise, 2.4.1. Bounded case The we address is the adaptive estimation over the collection of the functional classes first problem ~ ~ . N~r,d β, L ∩ Fg (R) ∩ B∞,d (Q) β,~ ~ r ,L,R,Q ~ As it was conjectured in Lepski and Willer (2017), the boundedness of the function belonging to ~ L ~ ∩ Fg (R) is a minimal condition allowing to eliminate the inconsistency zone. The results N~r,d β, obtained in Theorem 3 below together with those from Theorem 2 confirm this conjecture. Theorem 3. Let α ∈ [0, 1], ℓ ∈ N∗ and g ∈ L1 Rd , satisfying Assumption 5, be fixed. Let K satisfy Assumptions 6 and 7. 9 ~ ∈ (0, ℓ]d , ~r ∈ (1, ∞]d and L ~ ∈ [L0 , ∞)d 1) Then for any p ∈ (1, ∞), Q > 0, R > 0, L0 > 0, β ~ such that: there exists C < ∞, independent of L, −1 −ρ(α) (n) lim sup sup δn Rp fb~h,H ; f ≤ C, bn H d n→∞ ~L ~ ∩Fg (R)∩B∞,d (Q) f ∈N~r,d β, where ρ(α) is defined in (1.8). 2) For any p ∈ (1, ∞), Q > 0, R > 0, L0 > 0, β ∈ (0, ℓ], r ∈ [1, ∞] and L ∈ [L0 , ∞) there exists C < ∞, independent of L, such that: −1 −ρ(α) (n) d lim sup sup δn Rp fb~h,Hd ; f ≤ C, bn Hisotr isotr n→∞ f ∈Nr,d β,L ∩Fg (R)∩B∞,d (Q) Some remarks are in order. 10 . Our estimation procedure is completely data-driven, i.e. inde~ ~r, L, ~ R, Q, and the assertions of Theorem 3 are completely new if α 6= 0. Comparing pendent of β, the results obtained in Theorems 2 and 3 we can assert that our estimator is optimally-adaptive if κα (p) < 0 and nearly optimally adaptive if 0 < κα (p) < pω(α). The construction of an estimation procedure which would be optimally-adaptive when κα (p) ≥ 0 is an open problem, and we conjecture that the lower bounds for the asymptotics of the minimax risk found in Theorem 2 are sharp in order. This conjecture in the case α = 1 is partially confirmed by the results obtained in Comte and Lacour (2013) and Rebelles (2016). Since both articles deal with the estimation of unbounded functions we will discuss them in the next section. It is worth noting that all the previous statements are true not only for the convolution structure density model but also, in view of Theorem 2, for the observation scheme (1.3) as well. 20 . We note that the asymptotic of the minimax risk under partially contaminated observations, α ∈ (0, 1), is independent of α and coincides with the asymptotic of the risk in the direct observation model, α = 0. For the first time this phenomenon was discovered in Hesse (1995) and Yuana and Chenb (2002). In the very recent paper Lepski (2017), in the particular case ~r = (p, . . . , p), p ∈ (1, ∞) the optimally adaptive estimator was built. It is easy to check that independently of the value of β~ and ~ µ, the corresponding set of parameters belongs to the dense zone. Note however that our estimator is only optimally-adaptive in this zone, but it is applied to a much more general collection of functional classes. It is worth noting that the estimator procedure, used in Lepski (2017), has nothing in common with our pointwise selection rule. 30 . As to the direct observation scheme, α = 0, our results coincide with those obtained recently in Goldenshluger and Lepski (2014), when pω(0) > κ0 (p). However, for the tail zone pω(0) ≤ κ0 (p), our bound is slightly better since the bound obtained in the latter paper contains an additional d factor ln p (n). It is interesting to note that although both estimator constructions are based upon local selections from the family of kernel estimators, the selection rules are different. 40 . Let us finally discuss the results corresponding to the tail zone, κα (p) > pω(α). First, the lower bound for the minimax risk is given by [L(α)n−1 ]ρ(α) while the accuracy provided by our estimator is d−1 ln p (n)[L(α)n−1 ln(n)]ρ(α) . As we mentioned above, the passage from [L(α)n−1 ]ρ(α) to [L(α)n−1 ln(n)]ρ(α) seems to be an unavoidable payment for the application of a local selection scheme. It is interesting to note that d−1 the additional factor ln p (n) disappears in the dimension d = 1. First, note that if α = 0 the onedimensional setting was considered in Juditsky and Lambert–Lacroix (2004) and Reynaud–Bouret et al. 10 (2011). The setting of Juditsky and Lambert–Lacroix (2004) corresponds to r = ∞, while Reynaud–Bouret et al. (2011) deal with the case of p = 2 and τ (2) > 0. Both settings rule out the sparse zone. The rates of convergence found in these papers are easily recovered from our results corresponding to the tail and dense zones. Next, we remark that the aforementioned factor appears only when anisotropic functional classes are considered. Indeed, in view of the second assertion of Theorem 3 our estimator is nearly optimally adaptive on the tail zone in the isotropic case. The natural question arising in this context, d−1 is whether the ln p (n)-factor is an unavoidable payment for anisotropy of the underlying function or not? At last, we note that in the isotropic case our results remain true when the corresponding Nikol’skii class is defined in L1 -norm on Rd (r = 1). It is worth noting that the analysis of the proof of the theorem allows us to assert that if rj = 1, j ∈ J for some J 6= {1, . . . , d} the first statement remains true up to some logarithmic factor. However the asymptotic of the maximal risk of our estimator if rj = 1 for any j = 1, . . . , d remains unknown. 50 . We finish our discussion with the following remark. If α 6= 1 the assumption f ∈ Fg,∞ (R, Q) implies in many cases that f is uniformly bounded and, therefore, Theorem 3 is applicable. In particular it is always the case if the model (1.3) is considered. Indeed f, g ∈ P Rd in this case, which implies kf k∞ ≤ (1 − α)−1 kpk∞ ≤ (1 − α)−1 Q. Another case is kgk∞ < ∞ and recall that this assumption was used in the proofs of Theorems 1 and 2, Assumption 3. We obviously have that kf k∞ ≤ (1 − α)−1 Q + αRkgk∞ . More generally kf k∞ ≤ (1−α)−1 (Q+αD) if f ∈ Fg,∞ (R, Q) and kf ⋆gk∞ ≤ D. Since the definition of the Nikol’skii class implies that kf kr∗ ≤ L∗ , where r ∗ = supj=1,...,d rj and L∗ = supj=1,...,d Lj , the latter condition can be verified in particular if kgkq < ∞, 1/q = 1 − 1/r ∗ . All saying above explains why we study the estimation of unbounded functions only in the case α = 1. 2.4.2. Unbounded case, α = 1 The problem we address now is the adaptive estimation over the collection of functional classes ~ L ~ ∩ Fg,∞ (R, Q) ~ ~ . N~r,d β, β,~ r,L,R,Q As we already mentioned, if additionally kgk∞ < 0 then Fg,∞ (R, Q) = Fg (R) for any Q ≥ Rkgk∞ and, therefore, in view of Theorem 1 discussed in Section 1.3, there is no consistent estimator if either p = 1 or κα (p) ≤ 0, τ (p) ≤ 0, p∗ = p. Analyzing the proof of the latter theorem, we come to the following assertion. Conjecture 1. Let α = 1 and assume that Assumption 4 is fulfilled. Suppose additionally that Assumption 2 holds with minj=1,...,d µj > 1/p. Then, the assertion of Theorem 1 remains true if ~ L ~ L ~ ∩ Fg (R) by N~r,d β, ~ ∩ Fg,∞ (R, Q). one replaces N~r,d β, The latter result is formulated as a conjecture only because we will not prove it in the present paper. Its proof is postponed to Part III where the adaptive estimation over the collection ~ L ~ ∩ Fg,u (R, Q) ∩ Bq,d (Q) ~ ~ N~r,d β, β,~ r,L,R,Q,u,q introduced in Part I will be studied. For this reason, later on we will only consider the parameters ~ ~r belonging to the set Pp,~µ defined below. β, n o ~ ~r : κα (p) ≤ 0, τ (p) ≤ 0, max rj ≤ p . Pp,~µ = (0, ∞)d × [1, ∞]d \ β, j=1,...,d 11 For given p > 1 and ~ µ ∈ (0, ∞)d the latter set consists of the class parameters for which a uniform consistent estimation is possible. Theorem 4. Let ℓ ∈ N∗ and g ∈ L1 Rd , satisfying Assumption 5 be fixed and let K satisfy Assumptions 6 and 7. ~ ~r) ∈ Pp,~µ ∩ (0, ℓ]d × 1) Then for any p > [minj=1,... µj ]−1 , R, Q > 0, 0 < L0 ≤ L∞ < ∞, β, ~ ∈ [L0 , L∞ ]d there exists C < ∞, independent of L, ~ such that: (1, ∞]d and L lim sup n→∞ sup ~L ~ ∩Fg,∞ (R,Q) f ∈N~r,d β, bn H d −1 δn−̺(1) Rp(n) fb~h,Hd ; f ≤ C, where ̺(·) is defined in (1.7). 2) For any p > [minj=1,... µj ]−1 , R, Q > 0, 0 < L0 ≤ L∞ < ∞, (β, r) ∈ Pp,~µ ∩ (0, ℓ] × (1, ∞] and L ∈ [L0 , L∞ ] there exists C < ∞, independent of L, such that: lim sup n→∞ sup f ∈Nr,d β,L ∩Fg,∞ (R,Q) d bn Hisotr −1 δn−̺(1) R(n) fb~h,Hd p isotr ; f ≤ C. Some remarks are in order. 10 . Note that kgk1 < ∞, kgk∞ < ∞ implies that kgk2 < ∞ and, therefore the Parseval identity together with Assumption 5 allows us to assert that kgk∞ < ∞ ⇒ µj > 1/2, ∀j = 1, . . . , d. (2.7) Hence, the condition p > [minj=1,... µj ]−1 is automatically checked if p ≥ 2 and kgk∞ < ∞. Also, it is worth noting that considering the adaptation over the collection of isotropic classes, we do not require that the coordinates of ~µ would be the same. The latter is true for the second assertion of Theorem 3 as well. At last, analyzing the proof of the theorem, we can assert that the second assertion remains true under the slightly weaker assumption p > d(µ1 + · · · + µd )−1 . 20 . The assertion of Theorem 1 has no analogue in the existing literature except the results obtained in Comte and Lacour (2013) and Rebelles (2016). Comte and Lacour (2013) deals with the particular case p = 2, ~r = (2, . . . , 2) while Rebelles (2016) studied the case ~r = (p, . . . , p), p ∈ (1, ∞). It is easy to check that in both papers whatever the value of β~ and ~µ, the corresponding set of parameters belongs to the dense zone. Note also that the estimation procedures used in Comte and Lacour (2013) as well as in Rebelles (2016), if p ≥ 2, (both based on a global version of the Goldenshluger-Lepski method) are optimally-adaptive. They attain the asymptotic of minimax risks corresponding to the dense zone found in Theorem 1, while our method is only nearly optimally adaptive. However, it is well-known that the global selection from the family of standard kernel estimators leads to correct results only if ~r = (p, . . . , p) when the Lp -risk is considered, see, for instance Goldenshluger and Lepski (2011). On the other hand, estimation procedures based on a local selection scheme, which can be applied to the estimation of functions belonging to much more general functional classes, often do not lead to an optimally adaptive method. Fortunately, the loss of accuracy inherent to local procedures is logarithmic w.r.t. the number of observations. 30 . Together with Theorems 1 and 2, Theorems 3 and 4 provide the full classification of the asymptotics of the minimax risks over anisotropic/isotropic Nikolskii classes for the class parameters belonging to the sparse zone and, up to some logarithmic factor, belonging to the tail and dense zones as well as the boundaries. We mean that the results of these theorems are valid for any 12 ~ ∈ (0, ∞)d . Indeed, for given β~ and L ~ one can choose L0 = fixed β~ ∈ (0, ∞)d , ~r ∈ (1, ∞]d and L minj=1,...d Lj , L∞ = maxj=1,...d Lj and the number ℓ, used in the kernel construction (2.6), as any integer strictly larger than maxj=1,...d βj . 2.4.3. Open problems Let us briefly discuss some unresolved adaptive estimation problems in the convolution structure density model. Construction of an optimally-adaptive estimator As we already mentioned the proposed pointwise selection rule leads to an optimal adaptive estimator only for the class parameters belonging to the sparse zone (in both bounded and unbounded case). We conjecture that the construction of an optimally-adaptive estimator for all values of the nuisance parameters via pointwise selection is impossible, and other methods should be invented. It is worth noting that no optimally-adaptive estimator is known neither in the density model nor in the density deconvolution even in dimension 1. In dimension larger than 1, one of the intriguing questions is related to the eventual price to pay d−1 for anisotropy (ln p (n)-factor) discussed in the remark 40 after Theorem 3. Adaptive estimation of unbounded functions We were able to study the unbounded case only if α = 1. The estimation of unbounded densities under direct as well as partially contaminated observations remain open problems. We conjecture that the results obtained in the case α = 1 are not true anymore for α 6= 1 (neither upper bounds nor lower bound), but correct (or nearly correct) upper bounds for the asymptotics of the minimax risk can still be deduced from the oracle inequalities proved in Part I. In the case α = 1 there are at least two interesting problems. First, all our results are valid under the condition p > [minj=1,... µj ]−1 . How the absence of this assumption may have effects on the accuracy of estimation is absolutely unclear. Next, let us mention that the lower bound result proved in Theorem 1 holds only under the consideration of the convolution structure density model. Could the same bounds be established in the deconvolution model (1.3)? Adjustment of ”lower” and ”upper bound” assumptions to each other Comparing the assertions of Theorems 1 and 2 with those of Theorem 3 and 4, we remark that the obtention of the corresponding lower bounds for the minimax risk requires additional, rather restrictive, assumptions on the function g. Can they be weakened or even removed? 3. Proof of Theorems 3 and 4 The proofs are based on the application of Theorem 3 from Part I and on some auxiliary assertions presented below. In the subsequent proof c, c1 , c2 , C, C1 , C2 . . ., stand for constants that can depend on g, L0 , L∞ , ~ ~r, d and p, but are independent of L ~ and n. These constants can be different on different Q, R, β, appearances. 3.1. Important concepts from Part I and proof outline In this section we recall the definition of some important quantities that appeared in Theorem 3 of Part I and discuss the facts which should be established to make this theorem applicable. 13 I0 . Theorem 3 (Part I) deals with the minimax result over a class F being an arbitrary subset of Fg,u (R, D) ∩ Bq,d (D) defined in Section 2.3 of Part I. In Theorem 3 we will consider F = ~ L ~ ∩ B∞,d (Q) and, therefore, F ⊂ Fg,∞ (R, D) ∩ B∞,d (Q) with D = Q[1 − α + αkgk1 ]. This N~r,d β, makes Theorem 3 (Part I) with u = ∞ applicable in this case. ~ L ~ ~r and L ~ ∩ Fg,∞ (Q). We will show that for any β, ~ one In Theorem 4 we consider F = N~r,d β, ~ L ~ ⊂ Bq,d (D) and, therefore, Theorem 3 (Part I) is can find q > 1 and D > 0 such that N~r,d β, applicable with u = ∞. The latter inclusions are mostly based on the embedding of anisotropic Nikol’skii spaces used in the proof of Proposition 3 and on Lemma 1. II0 . The application of Theorem 3 (Part I) in the case u = ∞ requires to compute J ~h, v = j ∈ {1, . . . , d} : hj ∈ Vj (v) , Vj (v) = v ∈ H : Bj,∞,F (v) ≤ cv , X sj −sj ; v Bj,sj ,F hj Λ~s (v, F, ∞) = inf ~h∈H(v,2) Λ~s v, F = inf ~h∈H(v) X ¯ ~h,v) j∈J( v −sj ¯ ~h,v) j∈J( Bj,sj ,F hj sj +v −2 Fn2 ~h , 1/2 Qd P − 12 −µj (α) (nh) where remind Fn ~h = ln n + dj=1 \| ln hj \| and c > 0 is a universal j=1 j (hj ∧ 1) constant completely determined by the kernel Kℓ and the dimension d. In the next section we propose quite sophisticated constructions of vectors h(·, s) and ~h(·, s), s ∈ [1, ∞] and show, Propositions 1 and 2, that ~ h(v, 1) ∈ H(v), v ∈ [v, 1], ~ h(v, ∞) ∈ H(v, 2), v ∈ [v, v], ~h(v, ∞) ∈ H(v, 2). (3.1) Here v is defined in (3.15), v, v are defined in (3.16) and v ∈ {1, v1 , v3 , v, v ∧ v3 }, where v1 , v3 are defined in (3.18) and v is given in (3.19). In Proposition 3 we prove that for any ~h ∈ Hd Bj,r j ,N~ r,d and if τ (p∗ ) > 0 then additionally B hj ≤ C1 Lj hβj , j ~L ~ β, hj ≤ C1 Lj hγj , j ~~ j,qj ,N~r,d β,L j = 1, . . . , d. (3.2) j = 1, . . . , d, (3.3) ~ At last the definition of where ~γ and ~q are defined in (3.12) below and C1 is independent of L. ~ h(·, s) and h(·, s), s ∈ [1, ∞] together with (3.2) allows us to assert, see (3.26), that ~ J h(v, 1), v ⊇ J∞ , J ~ h(v, ∞), v ⊇ J∞ , J ~h(v, ∞), v ⊇ J∞ , ∀v > 0, (3.4) where J∞ = {j = 1, . . . , d : rj = ∞}. Thus, putting λ1 (v) = X j∈J¯∞ λ2 (v) = X j∈J¯∞ r β r v −rj Lj j hj (v, ∞) j j , λ= X j∈J¯∞ q γ q v−qj Lj j hj (v, ∞) j j ; d Y r β r (hj (v, 1))−1−2µj (α) , v −rj Lj j hj (v, 1) j j + v −2 (ln n/n) j=1 14 we obtain in view of (3.1), (3.2) and (3.4) that ~ L ~ , ∞ ≤ C1 λ1 (v), Λ~r v, N~r,d β, ~ L ~ Λ~r v, N~r,d β, ≤ C1 λ2 (v), ∀v ∈ [v, v]; (3.5) ∀v ∈ [v, 1] (3.6) To get (3.6) we have used that for all n large enough and all v ∈ [v, 1] Q ~ Fn h(v, 1)) ≤ C2 (ln n/n) dj=1 (hj (v, 1))−1−2µj (α) , ~ This follows from assertions (4.1) and (4.3) established in the proof where C2 is independent of L. of Proposition 1. We deduce from (3.5) and (3.6), the following bound. i h ~ L ~ L ~ , ∞ ∧ Λ~r v, N~r,d β, ~ (3.7) Λ~r v, N~r,d β, ≤ C1 λ2 (v)1[v,v] (v) + λ1 (v)1[v,v] (v) . Moreover, if τ (p∗ ) > 0 we get in view of (3.1), (3.3) and (3.4) ~ L ~ , ∞ ≤ C1 vp λ. vp Λ~q v, N~r,d β, (3.8) 3.2. Special set of bandwidths The bandwidth’s construction presented below as well as auxiliary statements from the next section will be exploited not only for proving Theorems 3 and 4, but also in the consideration forming Part III of this work. By this reason we formulate them in a bit more general form than what is needed for our current purposes. Set for any r, s ∈ [1, ∞] sω(α)(2+1/β(α)) − r, (s+ω(α)) −1 −d and let 20d max(2cKℓ kKℓ k∞ , kKℓ k1 ) κα (r, s) = Recall that c = that C1 appeared in (3.2)) α ∈ [0, 1]. L > 0 be any number satisfying (recall L ≤ 1 ∧ (C1−1 c) ∧ L0 . (3.9) Recall that δn = L(α)n−1 ln n and introduce for any v > 0, s ∈ [1, ∞] and j = 1, . . . , d LL−1 j e j (v, s) = η β1 j a−2 δn γ1 −2 j a LL−1 δn j b j (v, s) = η sω(α) (s+ω(α))βj rj sυ(α) (s+υ(α))γj qj 1 v βj v 1 γj − sω(α)(2+1/β(α)) (s+ω(α))β r j j sυ(α)(2+1/γ(α)) − (s+υ(α))γ q j j ; (3.10) , (3.11) where we have put p± = [supj∈J¯∞ rj ] ∨ p, J¯∞ is complimentary to J∞ and ( βj τ (p± ) p± , j ∈ J¯∞ , j ∈ J¯∞ , τ (rj ) , qj = , γj = ∞, j ∈ J∞ , βj , j ∈ J∞ . d X 2µj (α) + 1 1 := , γ(α) γj j=1 (3.12) d X 2µj (α) + 1 1 := . υ(α) γj qj j=1 The constant a > 0 will be chosen differently in accordance with some special relationships between ~ ~r, ~ the parameters β, µ, α and p. Determine hj (·, s) and hj (·, s), j = 1, . . . , d, from the relations e j (v, s) , v > 0; hj (v, s) = max h ∈ H : h ≤ η (3.13) b j (v, s) , v > 0, hj (v, s) = max h ∈ H : h ≤ η (3.14) ~ s) = h1 (·, s), . . . , hd (·, s) and ~h(·, s) = h1 (·, s), . . . , hd (·, s) . and set h(·, 15 3.3. Auxiliary statements All the results formulated below are proved in Section 4. Let − ω(α) ω(α)(2+1/β(α)) z(v) = 2 a−2 δn ω(α)+u v u+ω(α) , and remark that z(·) ≡ 2 if u = ∞. Note also that z(v) ≥ 2, ∀v ≥ a−2 δn u ∈ [1, ∞], 1 2+1/β(α) = v. (3.15) Introduce the following notations: µ(α) = minj=1,...,d µj (α), X= d X µj (α) 1 1 − = , 2β(1) 2β(0) βj Y = j=1 d X µj (α) 1 1 − = . 2ω(1) 2ω(0) βj rj j=1 Recall that z(α) = ω(α)(2 + 1/β(α))β(0)τ (∞) + 1 and define 1 v = (a−2 δn ) 1−1/ω(α)+1/β(α) , v = a−2 δn ω(α)τ (∞)β(0) z(α)+ω(α)/u . (3.16) Set u∗ = [−τ (∞)β(0)]−1 if τ (∞) < 0 and let u∗ = ∞ if τ (∞) ≥ 0. Put finally y = u∗ ∨ p∗ . ~ ~r, L0 , L∞ , ~ ~ ∈ [L0 , L∞ ]d . Then, Proposition 1. Let β, µ, α and p be given. Assume that L ~ such that for all n large enough 1) there exists a > 0 independent of L ~ h(v, 1) ∈ H(v), ∀v ∈ [v, 1], ~ and u such that for all n large enough 2) there exists a > 0 independent of L ~h(v, u) ∈ H v, z(v) if either κα (p∗ , u) < 0, τ (∞) ≥ 0 or κα (p∗ , u) < 0, τ (p∗ ) > 0, Y ≥ [X + 1]y−1 − 1/u. Remark 2. Note that if α 6= 1, the condition Y ≥ [X + 1]y−1 − 1/u simply means u ≤ u∗ ∨ p∗ , since X = Y = 0. On the other hand if α = 1 this condition holds if τ (∞) ≥ 0 whatever the values ~ µ of β, ~ and ~r, since Y > 0. Also, note that µ(1) + 1/u − 1/y ≥ 0 ⇒ Y ≥ [X + 1]y−1 − 1/u. (3.17) Indeed, since rj ≤ p∗ ≤ y for any j = 1, . . . , d we have Y − [X + 1]y−1 + 1/u ≥ µ(1)[1 − τ (y)] − 1/y + 1/u ≥ µ(1) − 1/y + 1/u and (3.17) follows. To get the last inequality we have used that τ (u∗ ) = 0 and that τ (·) is strictly decreasing, so τ (y) ≤ 0. In particular we deduce from (3.17) that the condition Y > [X +1]y−1 −1/u is always fulfilled in the case u = u∗ . Recall that v → 0, n → ∞, is defined in (3.15) and introduce the following quantities. v1 = a−2 δn 1 1−u/ω(0)+1/β(0) , v2 = a−2 δn uω(1) κ1 (p∗ ,u)(ω(1)+u) , v3 = a−2 δn where π(u) = [1/ω(0) − 1/u][1 + X] − 1/β(0)[Y + 1/u]. Define also v = v1{τ (p∗ )>0} + v2 1{τ (p∗ )≤0} 16 − Y +1/u π(u)∨0 , (3.18) (3.19) Note that v1 → ∞, n → ∞, if ∞ > u ≥ u∗ ∨ p∗ (it will be proved in Proposition 2 below). However v1 = 1 if u = ∞. As it is shown in the proof of Proposition 1, formulae (4.11), v < v for all n large enough. Also v2 → ∞, n → ∞, if κ1 (p∗ , u) < 0. At last v3 → ∞, n → ∞, since ω(0) > ω(1). Moreover v3 = ∞ if π(u) ≤ 0. Introduce finally   [v, 1], p∗ = ∞      [v, v1 ], α 6= 1, p∗ < ∞;    [v, v3 ], α = 1, p∗ < ∞, κα (p∗ , u) ≥ 0; Iu (α) =     α = 1, p∗ < ∞, κα (p∗ , u) < 0, Y ≥ [X + 1]y−1 − 1/u; [v, v],      [v, v ∧ v ], α = 1, p∗ < ∞, κ (p∗ , u) < 0, Y < [X + 1]y−1 − 1/u, 3 α ~ ~r, L0 , L∞ , ~ ~ ∈ [L0 , L∞ ]d , u ∈ [u∗ ∨ p∗ , ∞]. Proposition 2. Let β, µ, α and p be given and let L ~ Then, there exists a > 0 independent of L and u such that for all n large enough ~ h(v, u) ∈ H v, z(v) , v ∈ Iu (α). In the current paper we will use the statements of Proposition 1 and 2 only with u = ∞. In this context we remark that κα (·) ≡ κα (·, ∞). Proposition 3. Let ℓ ∈ N∗ , p > 1 and K satisfying Assumption 7 be fixed. Then for any β~ ∈ (0, ℓ]d , ~ ∈ (0, ∞)d one can find C1 > 0 independent of L ~ such that (3.2) holds. If ~r ∈ [1, ∞]d and L ∗ additionally τ (p ) > 0 then (3.3) is fulfilled as well. At last, (3.2) and (3.3) remain true if one replaces the quantity B by B∗ . The quantities Bj,s,F(·) and B∗j,s,F(·) are introduced in Part I but the reader can find them in the proof of the proposition. Let us also present the following auxiliary results which will be useful in the sequel. Their proofs are postponed to Appendix. Lemma 1. For any u ∈ [1, ∞] κα (p∗ , u) ≤ 0, τ (p∗ ) > 0, Y ≥ [X + 1]y −1 ∗ − 1/u, τ (p ) > 0, ⇒ ⇒ z(α) + ω(α)/u > 0; (3.20) z(α)/ω(α) − 1 + 2/u ≥ 0. (3.21) Let Y − [X + 1]y−1 > 0 and κ1 (p∗ , ∞) ≥ 0. Then there exists s > p∗ such that τ (s) > 0, s ≥ (1 + X)/Y. (3.22) We finish this section with the following observations which will be useful in the sequel. If κα (p∗ ) ≥ 0 one has ̺(α) = ^ β(α) 1 − 1/p := r(α), 1 − 1/ω(α) + 1/β(α) 2β(α) + 1 If κα (p∗ ) < 0 one has ρ(α) = r(α) ^ ω(α) p . ω(α)(1 − p∗ /p) 1{τ (p∗ )>0} + 1{τ (p∗ )≤0} ; ̺(α) = r(α) z(α) κα (p∗ ) ^ τ (p)ω(α)β(0) ω(α) ρ(α) = r(α) 1{τ (∞)>0} + 1 . z(α) p {τ (∞)≤0} 17 ^ τ (p)ω(α)β(0) (3.23) (3.24) (3.25) 3.4. Concluding remarks Let us collect some bounds for several terms appearing in Theorem 3 (Part I) and used in the proofs of Theorems 3 and 4 simultaneously. β1 j , j ∈ J . Then, (3.4) follows 10 . First we remark that hj (·, 1) ≡ hj (·, ∞) ≡ hj (·, ∞) ≤ LL−1 ∞ j from (3.2) and (3.9) because for any j ∈ J∞ and v > 0 B (3.26) ~ ~ hj (v, 1) = B ~ ~ hj (v, ∞) = B ~ ~ hj (v, ∞) ≤ cv. j,∞,N~r,d β,L j,∞,N~r,d β,L j,∞,N~r,d β,L ~ s), s ∈ {1, ∞} that 20 . We deduce from the definition of h(·, ω(α) λ1 (v) ≤ c1 δnω(α) v −ω(α)(2+1/β(α)) , v ∈ I∞ (α), λ2 (v) ≤ c1 δnω(α)+1 v − ω(α)(2+1/β(α)) ω(α)+1 , v ∈ [v, 1]. It yields together with (3.7) and the definitions of v and v, choosing v = v, ω(α) Z v ω(α)(2+1/β(α)) p− p−1 ~ ~ ~ ~ ω(α)+1 v Λ~r v, N~r,d β, L , ∞ ∧ Λ~r v, N~r,d β, L dv ≤ c2 δnω(α)+1 v 1{κα (p)>pω(α)} v ω(α) p− ω(α)(2+1/β(α)) ω(α)+1 1{κα (p)<pω(α)} + δnω(α) v p−ω(α)(2+1/β(α)) 1{κα (p)>0} ω(α) ω(α)+1 ω(α) ω(α) p−ω(α)(2+1/β(α)) 1{κα (p)<0} + ln (n) δn 1{κα (p)=pω(α)} + δn 1{κα (p)=0} +δn v +δnω(α)+1 v =: An + c2 δnω(α) v p−ω(α)(2+1/β(α)) 1{κα (p)<0} . (3.27) After elementary computations and taking into account (3.23), we obtain An ≤ c3 bpn (H)δnpρ(α) , An ≤ c3 bpn (H)δnp̺(α) . (3.28) These bounds are not surprising because ̺(α) = ρ(α) if κα (p) ≥ 0. At last, if τ (p∗ ) > 0, we get from (3.8) thanks to the definition of ~h(·, ∞) and the presentation proved in (4.6) with u = ∞ ω(α)τ (p)β(0) ~ L ~ , ∞ ≤ c4 δn z(α) . vp Λ~q v, N~r,d β, p−1 30 . At last, choosing v = v, we obtain ℓH (v) ≤ c6 δn1−1/ω(α)+1/β(α) ln n (3.23), (3.24) and (3.25): ℓH (v) ≤ c6 bpn (H)δnpρ(α) , ℓH (v) ≤ c6 bpn (H)δnp̺(α) . (3.29) t(H) , which yields by (3.30) 3.5. Proof of Theorem 3 As it has already been mentioned we will apply Theorem 3 (Part I) with u = ∞, q = ∞, D = Q[1 − α + αkgk1 ] ∨ Q and v = v. 10 . Consider the cases κα (p∗ ) ≥ 0, or κα (p∗ ) < 0, τ (∞) ≤ 0. Choose v = 1 and remark that the statements of Propositions 1 and 2 hold for any v ∈ [v, v]. Indeed, it suffices to note that I∞ (α) ⊇ [v, v] := [v, 1], because v1 , v2 , v3 > 1 and v ≥ 1 if τ (∞) < 0 since in this case v > 1 by (3.20). Then we can apply all the bounds obtained above, and in particular we get from (3.5) ~ L ~ , ∞ ≤ C1 λ1 (1) ≤ c5 δnω(α) ≤ c5 bpn (H)δnpρ(α) , Λ~r 1, N~r,d β, (3.31) 18 since ω(α) ≥ pρ(α) in both considered cases in view of the second equality in (3.23) and of (3.25). Applying the third assertion of Theorem 3 (Part I), we obtain from (3.27), (3.28), (3.31) and (3.30) 1 p (p) b p pρ(α) sup ≤ c7 bn (H)δnρ(α) , Rn [f~h(·) , f ] ≤ C (c2 + c3 + c5 + c6 )bn (H)δn ~L ~ ∩Fg (R) f ∈N~r,d β, and the assertion of Theorem 3 follows in both considered cases. 20 . Consider the case κα (p∗ ) < 0, τ (∞) > 0. Choose v = v and remark that the statements of Propositions 1 and 2 hold hold for any v ∈ [v, v]. Indeed, τ (∞) > 0 implies v < 1 and, therefore, v = v ∧ v3 = v. We deduce from (3.27), (3.28), (3.29) and (3.30), applying the first assertion of Theorem 3 (Part I) that 1 ω(α)τ (p)β(0) p z(α) p pρ(α) (p) b + (c3 + c6 )bn (H)δn sup Rn [f~h(·) , f ] ≤ C c8 δn ≤ c9 bn (H)δnρ(α) . (3.32) ~L ~ ∩Fg (R) f ∈N~r,d β, Here we have also used (3.25). This completes the proof of Theorem 3. 3.6. Proof of Theorem 4 In the following we assume p∗ < ∞, since p∗ = ∞ implies by definition of the anisotropic Nikol’skii ~ ~ class that N~r,d β, L ⊂ B∞,d (L∞ ). Hence, the results in that case follow from Theorem 3 since ̺(α) = ρ(α) when p∗ = ∞. Moreover, we remark that the imposed condition p > [minj=1,... µj ]−1 implies Y ≥ [X + 1]y−1 − 1/u in view of (3.17) proved in Remark 2. This, first, makes the second assertion of Proposition 1 applicable. Next, it allows (recall that p∗ < ∞ and α = 1) to rewrite I∞ (1) appeared in Proposition 2 as I∞ (1) = [v, v3 ]1{κ1 (p∗ )≥0} + [v, v]1{κ1 (p∗ )<0} . 10 . Consider the case κα (p∗ ) < 0, τ (p∗ ) > 0. ~ ∈ [L0 , L∞ ] we remark that in view of Nikol’skii (1977) [Theorem Taking into account that L ~ L ~ ⊂ Bp∗ ,d (c9 L∞ ), where c9 is independent of L. ~ Thus, Theorem 3 (Part 6.9.1, Section 6.9] N~r,d β, ∗ I) is applicable with u = ∞, q = p and D = c9 L∞ ∨ Q. Choose v = v and remark that the statements of Propositions 1 and 2 hold since v = v. The assertion of the theorem is obtained from (3.27), (3.28), (3.29), (3.30), (3.24) and the first assertion of Theorem 3 (Part I) by the same computations that led to (3.32). 20 . Consider the case κ1 (p∗ ) < 0, τ (p∗ ) ≤ 0. Recall that p∗ > p in this case because it is necessary for the existence of an uniformly consistent estimator. Since the definition of the anisotropic ~ L ~ ⊂ Bp∗ ,d (L∞ ), we assert that the second assertion of Theorem Nikol’skii class implies that N~r,d β, 3 (Part I) is applicable with u = ∞, q = p∗ and D = L∞ ∨ Q. Choose v = v2 and note that v = v2 in the considered case. Thus, we deduce from (3.27), (3.28), (3.30) and (3.24) sup ~L ~ ∩Fg,∞ (R,Q) f ∈N~r,d β, Rn(p) [fb~h(·) , f ] ≤C ω(1)− c′2 δn ω(1)κ1 (p,∞) κ1 (p∗ ,∞) + (c3 + c6 )bpn (H)δnp̺(α) ω(1)(p−p∗ ) κ1 (p∗ ,∞) + δn 1 p , and the assertion of the theorem follows in this case. 30 . It remains to study the case κ1 (p∗ ) ≥ 0. Let s be an arbitrary number satisfying (3.22) of Lemma 1. Since τ (s) > 0 and s > p∗ we can assert in view of Nikol’skii (1977) [Theorem 6.9.1, ~ L ~ ⊂ Bs,d (c9 L∞ ), where c9 is independent of L. ~ Thus, Theorem 3 (Part) is Section 6.9] N~r,d β, 19 applicable with u = ∞, q = s and D = c10 L∞ ∨ Q. Choosing v = v3 , we deduce from (3.27), (3.28), (3.30) and from the second assertion of Theorem 3 (Part 2) sup ~L ~ ∩Fg,∞ (R,Q) f ∈N~r,d β, Rn(p) [fb~h(·) , f ] 1 Y (s−p) p [1+X]/ω(0)−Y /β(0) p p̺(α) . ≤ C (c3 + c6 )bn (H)δn + δn (3.33) Since either p∗ /ω(0) = 1/β(0), κ1 (p∗ ) > 0 or p∗ /ω(0) > 1/β(0), κ1 (p∗ ) ≥ 0 and s > p∗ ≥ p we get Y (s − p) [ω(0) − ω(1)](s − p) = >0 [1 + X]/ω(0) − Y /β(0) κ1 (p∗ , ∞) + p∗ − ω(0)/β(0) Simple algebra shows that Y (s − p) p ≥ [1 + X]/ω(0) − Y /β(0) 2 + 1/β(1) ⇔ sY ω(1) ≥ p(X + 1)[2 + 1/β(1)]−1 Using again κ1 (p∗ ) ≥ 0 and p∗ ≥ p we obtain sY ω(1) 2 + 1/β(1) = sκ1 (p∗ )Y + sp∗ Y ≥ spY ≥ p(X + 1) since s satisfies (3.22) of Lemma 1. Thus, we have for all n large enough Y (s−p) p p̺(α) δn[1+X]/ω(0)−Y /β(0) ≤ δn2+1/β(1) ≤ bpn (H)δn and the assertion of the theorem in the case κ1 (p∗ ) ≥ 0 follows from (3.33) and the first equality in (3.23). Theorem 4 is proved. 4. Proofs of Propositions 1, 2 and 3 The proof of Lemma 2 is postponed to Appendix. ~ ~r, ~ Lemma 2. For any β, µ, p ≥ 1 and α ∈ [0, 1] the following is true. −1 1/γ(α) − 1/β(α) = τ (∞)β(0) 1/ω(α) − 1/υ(α) . 4.1. Proof of Proposition 1 We start the proof with several remarks which will be useful in the sequel. First, obviously there ~ ~r, µ ~ such that exists 0 < T := T β, ~ , p < ∞ independent of L lim (ln n)−1 sup n→∞ sup sup d n X α∈{0,1} s∈[1,∞] v∈[v,1∨v] j=1 o = T. ln hj (v, 1) + ln hj (v, s) (4.1) Next, for any s ∈ [1, ∞] and any v > 0 d ω(α) 2s−ω(α)/β(α) −1−2µj (α) ln n Y − 1 − 1 e j (v, s) = a2 L β(α) a−2 δn ω(α)+s v s+ω(α) = 2L β(α) a2 v 2 z−1 (v). η n (4.2) j=1 1) Let us proceed to the proof of the first assertion. First we remark that for all n ≥ 3 ~ h(v, 1) ∈ (0, 1]d , 20 ∀v ∈ [v, 1]. (4.3) Indeed for any v > 0 we have since L ≤ L0 , β r e j j j (v, 1) ≤ a−2 δn η ω(α) 1+ω(α) v rj − ω(α)(2+1/β(α)) 1+ω(α) , j ∈ J¯∞ . (4.4) Therefore, for any v ∈ [v, 1] one has in view of the definition of v β r e j j j (v, 1) ≤ a−2 δn η ω(α) 1+ω(α) v 1− ω(α)(2+1/β(α)) 1+ω(α) ≤ a−2 δn Note that for any j ∈ J∞ e j (v, 1) = LL−1 η j v β1 j ω(α) 1+ω(α) v 1− ω(α)(2+1/β(α)) 1+ω(α) = 1, j ∈ J¯∞ . 1 ≤ v βj ≤ 1, ∀v ≤ 1. e j (·, 1) by construction. and the proof of (4.3) is completed since hj (·, 1) ≤ η d+2 Pd µj (α) − 1 j=1 Set T0 = T + 2 e L β(α) and remark that in view of (4.1), (4.2) and (4.3) for all n large enough and any v ∈ [v, 1] ~ Gn h(v, 1) ≤ ≤ 1 (T + 2) ln n T0 L β(α) ln n ≤ 1+µj (α) Q Q 1+µj (α) e j (v, 1) n dj=1 hj (v, 1) n dj=1 η 1 ω(α) 2−ω(α)/β(α) 2 T0 L β(α) ln n 1+ω(α) δ 1+ω(α) v 1+ω(α) = T a . n 0 Qd 1+2µj (α) e j (v, 1) n j=1 η (4.5) Here we have taken into account that hj (v, s) ≥ e−1 η j (v, s). Since ω(α) 2 T0 a 1+ω(α) δnω(α)+1 v denoting a = p 2−ω(α)/β(α) 1+ω(α) ≤ T0 a2 v ⇔ v ≥ v, a/T0 we assert that ~ 1) ≤ av, Gn h(v, ∀v ∈ [v, 1]. The first assertion is established. 2) Before proving the second assertion, let us make several remarks. 10 . For any u ∈ [1, ∞] the following is true. b j (v, u) = LL−1 η j v β1 j b j (v, u) = LL−1 η j , j ∈ J∞ , γ1 j a−2 δn (p± )β(0) γω(α)τ [z(α)+ω(α)/u] j , j ∈ J¯∞ . (4.6) b j (v, u) since, remind γj = βj , qj = ∞ if The first equality follows directly from the definition of η j ∈ J∞ . Thus, let us prove the second equality. We have γ q b j j j (v, u) = LL−1 η j p± a−2 δn uυ(α) u+υ(α) v p± − uυ(α)(2+1/γ(α)) u+υ(α) , ∀j ∈ J¯∞ . Here we have used that qj = p± for any j ∈ J¯∞ . Using the definition of v we get γ q b j j j (v, u) η = p ± LL−1 j −2 a δn Using the definition of z(α) we obtain (∞)β(0) uυ(α) + ω(α)τ u+υ(α) z(α)+ω(α)/u 21 h p± − uυ(α)(2+1/γ(α)) u+υ(α) i , ∀j ∈ J¯∞ . A := uυ(α) ω(α)τ (∞)β(0) uυ(α)(2 + 1/γ(α)) ω(α)τ (∞)β(0)p± + p± − = u + υ(α) z(α) + ω(α)/u u + υ(α) z(α) + ω(α)/u uυ(α) 1 + ω(α)/u − ω(α)τ (∞)β(0) 1/γ(α) − 1/β(α) . + (u + υ(α))(z(α) + ω(α)/u) We obtain applying Lemma 2 A= uυ(α)ω(α)[1/s + 1/υ(α)] ω(α)τ (p± )β(0)p± ω(α)τ (∞)β(0)p± + = . z(α) + ω(α)/u (u + υ(α))(z(α) + ω(α)/u) z(α) + ω(α)/u The second formula in (4.6) is established. 20 . Next, let us prove that ~h(v, u) ∈ (0, 1]d , ∀u ∈ [1, ∞]. (4.7) If J∞ 6= ∅, which is equivalent to p∗ = ∞, the definition of v implies that v ≤ 1 for all n large enough, since τ (p∗ ) = τ (∞) > 0 and in view of (3.20). We deduce from the first equality in (4.6) b j (v, u) = LL−1 hj (v, u) ≤ η j v β1 1 ≤ v βj ≤ 1, j ∀j ∈ J∞ . and (4.7) is proved for any j ∈ J∞ . It remains to note that τ (p± ) ≥ τ (p∗ ) since p∗ ≥ p± and therefore, if τ (p∗ ) ≥ 0 we have b j (v, u) ≤ 1, hj (v, u) ≤ η ∀j ∈ J¯∞ , for all n large enough in view of (3.20) of Lemma 1, the second equality in (4.6) and since LL−1 j ≤ 1. This completes the proof of (4.7). 30 . For any u ∈ [1, ∞] one has a−2 δn d Y j=1 −1−2µj (α) bj η a−2 δn d Y j=1 (v, u) ≤ T −1 (α) a−2 δn −1 b −1 η (0) a−2 δn j (v, u) ≤ T 1− ω(α)τ (∞)β(0)/β(α)+1 1− z(α)+ω(α)/u ; (4.8) ω(α) z(α)+ω(α)/u (4.9) j (α) Q j (α) 1+2µ 1+2µ Q −1 −1 βj γj where we have denoted T α = inf L∈[L . d ~ j∈J∞ LLj j∈J¯∞ LLj 0 ,L∞ ] Indeed, we have in view of (4.6) and the definition of v d Y j=1 1+2µj (α) bj η d Y j=1 where we have put (v, u) ≥ T −1 (α) a−2 δn b j (v, u) ≥ T −1 (0) a−2 δn η 1 β∞ (α) = P j∈J∞ 1+2µ(α) , γ±1(α) βj γ ω(α)τ (p± )β(0) ω(α)τ (∞)β(0) + ± (α)[z(α)+ω(α)/u] β∞ (α)[z(α)+ω(α)/u] γ = ω(α)τ (p± )β(0) ω(α)τ (∞)β(0) + ± (0)[z(α)+ω(α)/u] β∞ (0)[z(α)+ω(α)/u] P j∈J¯∞ 1+2µ(α) . γj Note that for any α ∈ [0, 1] X (1 + 2µ(α))τ (rj ) τ (∞) τ (∞) 1 τ (∞) τ (p± ) + = = + . + γ± (α) β∞ (α) βj β∞ (α) β(α) ω(α)β(0) ¯ j∈J∞ 22 , and (4.8) and (4.9) are established. 40 . Simple algebra shows that for any u ∈ [1, ∞] a−2 δn 1− ω(α)τ (∞)β(0)/β(α)+1 z(α)+ω(α)/u = 2v2 z−1 (v), and we deduce from (4.8) for any u ∈ [1, ∞] (recall that z ≡ 2 if u = ∞) δn d Y j=1 −1−2µj (α) bj η (v, u) ≤ 2T −1 (α)a2 v2 z−1 (v). (4.10) Let us also prove that for any u ∈ [1, ∞] and all n large enough v > v := a−2 δn 1 2+1/β(α) ⇒ z(v) ≥ 2. (4.11) The latter inclusion follows from (3.15). Indeed, if τ (∞) ≤ 0 then v ≥ 1 ≥ v. If τ (∞) > 0 1 1 + ω(α)/u ω(α)τ (∞)β(0) − =− <0 z(α) + ω(α)/u 2 + 1/β(α) [z(α) + ω(α)/u][2 + 1/β(α)] in view of (3.20), so v > v. Note at last that for any u ∈ [1, ∞] vz−1 (v) = 2 a−2 δn ω(α)τ (u)β(0) z(α)+ω(α)/u . (4.12) 50 . Let us proceed to the proof of the second assertion. Let us choose a < aT (α)/(4T0 ) < 1. We have in view of (4.1), (4.8) and (4.10) similarly to (4.5) Fn2 ~h(v, u) ≤ Q d T0 δn j=1 −1 2 2 −1 2 2 −1 1+2µj (α) ≤ 2T0 T (α)a v z (v) ≤ a v z (v). (4.13) b j (v, u) η Thus to prove the assertion all we need to show is that ~h(v, u) ∈ H(v), i.e. Gn ~h(v, u) ≤ av. Let us distinguish three cases. 50 a. Let τ (∞) ≥ 0. We remark that the definition of v in this case yields v ≤ 1 for all n large enough and we obtain from (4.10) and (4.11) that δn d Y j=1 −1−2µj (α) bj η (v, u) ≤ T −1 (α)a2 v. (4.14) Then we have in view of (4.1), (4.7), (4.8) and (4.14) similarly to (4.5) Gn ~h(v, u) ≤ Q d T0 δn j=1 −1 2 1+2µj (α) ≤ T0 T (α)a v ≤ av. b j (v, u) η (4.15) 50 b. Let τ (∞) < 0, τ (p∗ ) > 0 and α 6= 1. Then by assumption u ≤ p, and thus τ (u) ≥ 0. We get from (4.10) and (4.12) δn d Y j=1 −1−2µj (α) bj η (v, u∗ ) ≤ 4T −1 (α)a2 v, 23 (4.16) so Gn ~h(v, u) ≤ av follows from (4.15) and (4.13). 50 c. Let τ (∞) < 0, τ (p∗ ) > 0, α = 1. We have as previously G2n ~h(u) ≤ (T + 2) ln n (T + 2) ln n 1+2µj (α) Qd n j=1 hj (v, u) n j=1 hj (v, u) T (0)a−2 δn 2 −1 4 2 −1 ≤ 2T0 T (1)T1 a v z (v) Qd . b j (v, u) j=1 η Qd (4.17) Here we have used (4.10) and put T1 = T −1 (0)L−1/β(0) . Our goal now is to show that for any u ∈ [1, ∞] and all n large enough −2 T (0)a δn z −1 (v) d Y j=1 b −1 η j (v, u) ≤ 1. (4.18) In view of (4.9) and of the definition of z(·) in order to establish (4.18) it suffices to show that z(1)/ω(1) − 1 + 2/u ≥ 0. Since we assumed τ (∞) < 0 and τ (p∗ ) > 0, then necessarily u∗ > p∗ since τ (u∗ ) = 0 and τ (·) is strictly decreasing. Hence, the required results follows from (3.21). Thus, (4.18) is proved. Then choosing a such that T0 (2T −1 (1)T1 )1/2 a2 ≤ a, we obtain from (4.17) and (4.18) that for all all n large enough Gn ~h(v, u) ≤ T0 (2T −1 (α)T1 )1/2 a2 v ≤ av. The second assertion is proved 4.2. Proof of Proposition 2 We start the proof with several remarks which will be useful in the sequel. 10 . Let us show that for all n large enough ~ h(v, u) ∈ (0, 1]d , ∀v ∈ Iu (α), ∀u ≥ u∗ ∨ p∗ . e j (·, u), j = 1, . . . , d, In view of the definition of η β r e j j j (v, u) = LL−1 η j rj −2 a δn uω(α) u+ω(α) v rj − uω(α)(2+1/β(α)) u+ω(α) , (4.19) j ∈ J¯∞ . (4.20) Therefore, for any v ∈ [v0 , 1] one has, taking into account that L ≤ L0 , β r e j j j (v, u) ≤ a−2 δn η uω(α) u+ω(α) v 1− uω(α)(2+1/β(α)) u+ω(α) ≤ a−2 δn uω(α) u+ω(α) 1− v0 uω(α)(2+1/β(α)) u+ω(α) = 1, j ∈ J¯∞ . It remains to note that v > v0 for all n large enough and, therefore, e j (v, u) ≤ 1, j ∈ J¯∞ , η ∀v ∈ [v, 1] ∩ Iu (α). e j (·, u), j = 1, . . . , d, We also have in view of the definition of η β1 j ≤ 1, e j (v, u) = LL−1 η v j 24 j ∈ J∞ , (4.21) for any v ≤ 1. This together with (4.21) proves (4.19) in the cases when Iu (α) = [v, 1]. Noting that p∗ < ∞ is equivalent to J∞ = ∅, we deduce from (4.20) for any v ≥ 1 β r e j j j (v, u) ≤ a−2 δn η uω(α) u+ω(α) v p∗ − uω(α)(2+1/β(α)) u+ω(α) Thus, if κα (p∗ , u) ≥ 0 then for any v ≥ 1 β r e j j j (v, u) ≤ a−2 δn η uω(α) u+ω(α) ≤ a−2 δn uω(α) u+ω(α) v −κα (p ∗ ,u) , j = 1, . . . , d. → 0, n → ∞, j = 1, . . . , d. This together with (4.21) yields (4.19) in the case κα (p∗ , u) ≥ 0, p∗ < ∞, whatever the value of α. Let α = 1, p∗ < ∞, κα (p∗ , u) < 0, τ (p∗ ) > 0. Then v = v and we have for any j = 1, . . . , d and v ∈ [1, v] in view of the definition of v β r β r e j j j (v, u) ≤ η e j j j (v, u) = a−2 δn η κ (p∗ ,u)ω(1)τ (∞)β(0) uω(1) − 1 z(1)+ω(1)/u u+ω(1) = a−2 δn p∗ τ (p∗ )ω(1) z(1)+ω(1)/u → 0, n → ∞, in view of (3.20). Hence, (4.19) holds in this case. Let α = 1, κα (p∗ , u) < 0, τ (p∗ ) ≤ 0. Then v = v2 and we have for any v ∈ [1, v2 ] in view of the definition of v2 β r β r e j j j (v, u) ≤ η e j j j (v2 , u) = a−2 δn η uω(1) u+ω(1) −κ1 (p∗ ,u) v2 = 1, j = 1, . . . , d. and, therefore (4.19) holds in this case. Let κα (p∗ , u) < 0, α 6= 1, u < ∞. First we note that τ (∞) < 0 and u ≥ u∗ ∨ p∗ imply 1 − u/ω(0) + 1/β(0) = 1 − u + uτ (u) ≤ 1 − u + uτ u∗ ∨ p∗ ≤ 1 − u < 0, since either τ (p∗ ) ≤ 0 or u∗ > p∗ and τ u∗ ∨ p∗ = 0. Thus v1 → ∞, n → ∞ and, therefore, for any v ∈ [1, v1 ] uω(0) β r β r −κ (p∗ ,u) e j j j (v, u) ≤ η e j j j (v1 , u) = a−2 δn u+ω(0) v1 0 , j = 1, . . . , d. η Note that 1 − u/ω(0) + 1/β(0) = κ0 (p∗ , u) 1/u + 1/ω(0) − (u − p∗ ) 1/u + 1/ω(0) and, therefore − κ0 (p∗ , u) uω(0) ≥− , 1 − u/ω(0) + 1/β(0) u + ω(0) − uω(0) −κ (p∗ ,u) which yields v1 0 ≤ a−2 δn u+ω(0) . It remains to note that if τ (∞) ≥ 0 then u∗ = ∞ and, therefore u = ∞. It implies v1 = 1 and Iu (α) = [v, 1] and this case has been already treated. This completes the proof of (4.19). ~ ~r, ~µ, p < ∞ independent of L ~ such that 20 . Remark that there obviously exists 0 < S := S β, −1 lim (ln n) n→∞ sup sup sup d X α∈{0,1} u∈[1,∞] v∈Iu (α) j=1 ln hj (v, u) = S. Hence, in view of (4.19) one has for all n large enough and v ∈ Iu (α) p (S + 2) ln n (S + 2) ln n ~ ~ Fn h(v, u) ≤ √ Q (4.22) . , Gn h(v, u) ≤ Q 1+µj (α) 1 +µ (α) n dj=1 hj (v, u) n dj=1 hj (v, u) 2 j 25 Pd − 1 e j (v, u) and setting S0 = S + 2 ed+2 j=1 µj L β(1) we Taking into account that hj (v, u) ≥ e−1 η obtain from (4.2) for any α ∈ [0, 1] and v ∈ Iu (α) (S + 2)n −1 ln (n) d Y j=1 1+2µj (α) hj (v, s) ≤ 2S0 a2 v 2 z−1 (v). (4.23) From now on we choose a ≤ a/(2S0 ) < 1. It yields in view of (4.22) and (4.23) ~ Fn2 h(v, u) ≤ a2 v 2 z−1 (v), ∀v ∈ Iu (α). (4.24) 30 . Since (4.24) holds, to finish the proof of Proposition (2) all we need to show is that Gn ~h(v, u) ≤ av, ∀v ∈ Iu (α). Let us distinguish three cases. 30 a. Let p∗ = ∞ or α 6= 1, u = ∞. First we note that in these cases Iu (α) = [v, 1]. Next in view of the second inequality in (4.22), (4.19), (4.23) and (4.24) we obtain ~ Gn h(v, u) ≤ n (S + 2) ln n 2 2 −1 1+2µj (α) ≤ a v z (v) ≤ av, ∀v ∈ Iu (α). j=1 hj (v, u) Qd (4.25) To get the last inequality we have used that a < 1, z(·) ≥ 2 and v ≤ 1. 30 b. Let α 6= 1, p∗ < ∞, u < ∞. We have in view of the second inequality in (4.22) and (4.23) h(v, u) ≤ a2 v 2 z−1 (v), ∀v ∈ Iu (0). Gn ~ uω(0) u−ω(0)−ω(0)/β(0) u+ω(0) , and since For any u 6= ∞, simple algebra shows that vz−1 (v) = a−2 δn u+ω(0) v ∗ u ≥ u , which implies u − ω(0) − ω(0)/β(0) > 0, the result follows from sup vz−1 (v) = v1 z−1 (v1 ) = 1. v∈Iu (0) 30 c. Let α = 1, p∗ < ∞. We have in view of the second inequality in (4.22) and (4.23) G2n ~ h(v, u) ≤ n (S + 2) ln n (S + 2) ln n 1+2µj (α) Qd n j=1 hj (v, u) j=1 hj (v, u) Qd ≤ S1 a2 a2 v 2 z−1 (v) Qd a−2 δn e j (v, u) j=1 η v ∈ Iu (1), , (4.26) where we have denoted S1 = S0 L−1/β(0) . Our goal now is to show that for all n large enough −2 sup a v∈Iu (1) δn z −1 (v) d Y j=1 We easily compute for any v > 0 a−2 δn z−1 (v) d Y j=1 −2 −1 e −1 η j (v, u) = z (v) a δn 26 e −1 η j (v, u) ≤ 1. uω(1) 1− (ω(1)+u)ω(0) v (4.27) uω(1)(2+1/β(1)) 1 − β(0) (u+ω(1))ω(0) . (4.28) Denoting the right hand side of the obtained inequality by P (v) we obviously have −2 sup a δn z −1 (v) v∈Iu (1) d Y j=1 e −1 η v) , j (v, u) ≤ max P (v), P (e (4.29) e ∈ {v3 , v, v ∧ v3 }. Remarking that z(v) = 2 we easily compute that for any u ∈ [1, ∞] where v P (v) = 2−1 a−2 δn 2+1/β(α)−1/β(0 2+1/β(α) → 0, n → ∞. (4.30) Moreover we obviously have P (v) = 2−1 a−2 δn 2uω(1)(Y +1/u) ω(1)+u v 2uω(1)π(u) u+ω(1) , v > 0. (4.31) e = v3 . 30 c1. Consider the case κ1 (p∗ , u) ≥ 0. Here v If π(u) ≤ 0 then v3 = ∞ and we deduce from (4.31) sup P (v) = P (v) → 0, n → ∞, (4.32) v≥v thanks to (4.30). If π(u) > 0 the definition of v3 implies that P (v3 ) = 1. (4.33) Both last results together with (4.29) and (4.30) prove (4.27) in the case κ1 (p∗ , u) ≥ 0. e = v. 30 c2. Consider the case κ1 (p∗ , u) < 0, Y ≥ [X + 1]y−1 − 1/u. Here v ∗ ∗ ∗ If τ (p ) > 0 then v = v. Moreover y = u since u = ∞ if τ (∞) ≥ 0 and τ (u∗ ) = 0 if τ (∞) < 0. Hence in view of (3.21) of Lemma 1 z(1)/ω(1) − 1 + 2/u ≥ 0. We have in view of the definition of v Note that, P (v) = 2−1 a−2 δn uω 2 (1)τ (∞)β(0)π(u) uω(1)(1/ω(1)−1/ω(0)+2/u) + [u+ω(1)][z(α)+ω(α)/u] ω(1)+u . (4.34) uω(1)(1/ω(1) − 1/ω(0) + 2/u) uω 2 (1)τ (∞)β(0)π(u) + ω(1) + u [u + ω(1)][z(1) + ω(1)/u] ω(1)τ (∞) ω(1)[1 − 1/u] ω(1)[1/ω(0) − 1/u] − =1− > 0. = 1− z(1) + ω(1)/u z(1) + ω(1)/u z(1) + ω(1)/u To get the last inequality we have used that 1− ω(1)[1 − 1/u] >0 z(1) + ω(1)/u ⇔ z(1)/ω(1) − 1 + 2/u > 0. Thus, we conclude that P (v) ≤ 1, which together with (4.30) implies (4.27) in the considered case. If τ (p∗ ) < 0 then v = v2 . Moreover y = p∗ . We have in view of the definition of v2 P (v2 ) = 2−1 a−2 δn uω(1)(1/ω(1)−1/ω(0)+2/u) [uω(1)]2 π(u) + ω(1)+u κ1 (p∗ ,u)[u+ω(1)]2 27 . (4.35) After routine computations we come to the following equality [uω(1)]2 π(u) uω(1)(1/ω(1) − 1/ω(0) + 2/u) + ω(1) + u κ1 (p∗ , u)[u + ω(1)]2 2uω(1)p∗ Y − (X + 1)(y)−1 + 1/u ≥ 0. = − κ1 (p∗ , u)[u + ω(1)] Hence, P (v2 ) ≤ 1 for all n large enough, which together with (4.30) allows us to assert (4.27) in the considered case. e = v ∧ v3 . 30 c3. Consider the case κ1 (p∗ , u) < 0, Y < [X + 1]y−1 − 1/u. Here v If π(u) ≤ 0 the required result follows from (4.32). If π(u) > 0 then by (4.31) P (·) is strictly increasing and, therefore, P v ∧ v3 ≤ P v3 = 1 in view of (4.33). This completes the proof (4.27). p Finally to conclude in the case 30 c, choosing a ≤ 1/S1 , we deduce from (4.26) and (4.27) that for all n large enough √ Gn ~ h(v, u) ≤ S1 aav ≤ av, v ∈ Iu (1). 4.3. Proof of Proposition 3 In view of Lemma 5 in Lepski (2015), if τ (p∗ ) > 0 then ~ L ~ ⊆ N~q,d ~γ , c2 L ~ , N~r,d β, (4.36) ~ Note also that γj ≤ βj for any j = 1, . . . , d. where c2 is independent on L. ~ ~r or ~γ , ~q and without further mentioning the couple ~γ , ~q is used 10 . Let ~π , ~s be either β, below under the condition τ (p∗ ) > 0. We obviously have for any ~h ∈ H Z Kℓ (u) f x + uhej − f (x) ν1 (du) bh,f,j (x) := sup h∈H: h≤h = sup h∈H: h≤h For j = 1, . . . , d we have Z R R Kℓ (u) ∆uh,j f (x) ν1 (du) . Z X ℓ ℓ 1 u ∆hu,j f (x) ν1 (du) Kℓ (u)∆uh,j f (x)ν1 (du) = (−1)i+1 Kℓ i i R i=1 i R Z Z ℓ X ℓ = (−1)ℓ−1 Kℓ (z) (−1)i+ℓ ∆izh,j f (x) ν1 (dz) = (−1)ℓ−1 Kℓ (z) ∆ℓzh,j f (x) ν1 (dz). i R R Z i=1 The last equality follows from the definition of the ℓ-th order difference operator (2.4). Hence, for any j ∈ J∞ we have in view of the definition of the Nikol’skii class (remind that γj = βj , j ∈ J∞ ) 28 kbh,f,j k∞ ≤ sup Z h∈H: h≤h R Kℓ (z) This yields for any h ∈ H ∆ℓzh,j f (·) ν (dz) ∞ 1 ≤ Lj sup h∈H: h≤h π hj j Z R Kℓ (z) \|z\|πj ν1 (dz). (h) ≤ c1 Lj hπj , B (4.37) ~L ~ j,∞,N~r,d β, and the first and the second assertions of the proposition are proved for any j ∈ J∞ . Let j ∈ J¯∞ . Choosing k from the relation ek = h (recall that h ∈ H), we have for any x ∈ Rd Z Z ℓ bh,f,j (x) = sup Kℓ (z) ∆zek ,j f (x) ν1 (dz) =: lim sup Kℓ (z) ∆ℓzek ,j f (x) ν1 (dz) . k≤k l→−∞ l≤k≤k R R We have in view of monotone convergence theorem and the triangle inequality Z Kℓ (z) ∆ℓzek ,j f (·) ν1 (dz) bh,f,j s = lim sup j l→−∞ l≤k≤k ≤ k X k=−∞ Z R R sj Kℓ (z) ∆ℓzek ,j f (·) ν1 (dz) . sj By the Minkowski inequality for integrals [see, e.g., (Folland 1999, Section 6.3)], we obtain bv,f,j sj ≤ Z k X k=−∞ R \|Kℓ (z)\| ∆ℓzek ,j f sj ν1 (dz), j = 1, . . . , d. ~ L ~ and (4.36), we have for any j = 1, . . . , d, Taking into account that f ∈ N~r,d β, bh,f,j sj ≤ Z k X ekπj ≤ c1 Lj hπj , \|Kℓ (z)\| \|z\| ν1 (dz) Lj βj R k=−∞ ∀h ∈ Hd . This proves the first and the second assertions of the proposition for any j ∈ J¯∞ . ~ L ~ and recall that 20 . Set F = N~r,d β, Z X X ∗ Kℓ (u) f x + uhej − f (x) ν1 (du) Bj,sj ,F (h) := sup ≤ sup f ∈F h∈H: h≤h R sj f ∈F h∈H: h≤h (4.38) bh,f,j sj Hence, the third assertion follows from (4.37) and (4.38). 5. Appendix 5.1. Proof of Lemma 1 Note that z(α) + ω(α)/s = ω(α)(2 + 1/β(α))β(0)τ (p∗ ) + 1 − ω(α)(2 + 1/β(α))(p∗ )−1 + ω(α)/s = ω(α)(2 + 1/β(α))β(0)τ (p∗ ) − (p∗ )−1 (1 + ω(α)/s)κα (p∗ , s), 29 . and (3.20) follows. On the other hand we have z(α)/ω(α) − 1 + 2/u = 2 + 2X β(0)τ (∞) + 2Y + 2/u and (3.21) is checked if τ (∞) ≥ 0 since X, Y ≥ 0. If τ (∞) < 0 and τ (p∗ ) > 0 then we note first that necessarily u∗ > p∗ since τ (u∗ ) = 0 and τ (·) is strictly decreasing. Hence y = u∗ and z(α)/ω(α) − 1 + 2/u = 2 + 2X β(0)τ (∞) + 2Y + 2/u = 2 Y − (X + 1)y−1 + 1/u ≥ 0 and (3.21) is established. Let us prove (3.22). First we note that (3.22) is obvious if τ (∞) ≥ 0 because in this case τ (s) > 0 for any s ≥ 1. Thus, from now on we will assume that τ (∞) < 0. Next, if u∗ > p∗ then (3.22) holds. Indeed, in this case 0 < Y − [X + 1]y−1 = Y − [X + 1](u∗ )−1 implies u∗ > (X + 1)/Y . Hence any number from the interval p∗ ∨ (X + 1)/Y, u∗ satisfies (3.22). At last, note that if p∗ ≥ u∗ we have 0≤ κα (p∗ , ∞) ω(α) = 2 + 2X − 2p∗ Y + 1/β(0) − p∗ /ω(0) = 2y (1 + X)y−1 − Y + 1/β(0) − p∗ /ω(0) < 0, since 1/β(0) ≤ p∗ /ω(0) in view of rj ≤ p∗ for any j = 1, . . . , d. The obtained contradiction completes the proof of (3.22). 5.2. Proof of Lemma 2 Indeed, 1/γ(α) − 1/β(α) = 1/γ± (α) − 1/β± (α) = X 1 + 2µj (α) τ (rj )/τ (p± ) − 1 βj j∈J± −1 X 1 + 2µj (α) −1 = β(0)τ (p± ) (1/rj − 1/p± ) = τ (p± )β(0) 1/ω(α) − 1/(β± (α)p± ) . βj j∈J± Moreover, in view of the latter inequality 1/ω(α) − 1/υ(α) = 1/ω(α) − 1/(p± γ± (α)) −1 = 1/ω(α) − 1/(p± β± (α)) − τ (p± )β(0)p± 1/ω(α) − 1/(β± (α)p± ) −1 = 1 − τ (p± )β(0)p± 1/ω(α) − 1/(β± (α)p± ) . −1 It remains to note that 1 − τ (p± )β(0)p± = τ (∞)/τ (p± ) and the lemma follows. References Akakpo, N. (2012). Adaptation to anisotropy and inhomogeneity via dyadic piecewise polynomial selection. Math. Methods Statist. 21, 1–28. 30 Birgé, L. (2008). Model selection for density estimation with L2 –loss. arXiv:0808.1416v2, http://arxiv.org Comte, F., Rozenholc, Y. and Taupin, M.-L. (2006). Penalized contrast estimator for adaptive density deconvolution. Canad. J. Statist., 34, 3, 431–452. Comte, F. and Lacour, C. (2013). Anisotropic adaptive kernel deconvolution. Ann. Inst. H. Poincaré Probab. Statist. 49, 2, 569–609. Fan, J. (1991). On the optimal rates of convergence for nonparametric deconvolution problems. Ann. Statist., 19, 3, 1257–1272. Fan, J. (1993). Adaptively local one-dimensional subproblems with application to a deconvolution problem. Ann. Statist., 21, 2, 600–610. Fan, J. and Koo, J. (2002). Wavelet deconvolution. IEEE Trans. Inform. Theory, 48, 734–747. Devroye, L. and Lugosi, G. (1997). Nonasymptotic universal smoothing factors, kernel complexity and Yatracos classes. Ann. Statist. 25, 2626–2637. Donoho, D. L., Johnstone, I. M., Kerkyacharian, G. and Picard, D. (1996). Density estimation by wavelet thresholding. Ann. Statist. 24, 508–539. Efroimovich, S. Yu. (1986). Non-parametric estimation of the density with unknown smoothness. Theory Probab. Appl. 30, 557–568. Folland, G. B. (1999). Real Analysis. Second edition. Wiley, New York. Gach, D., Nickl, R., and Spokoiny, V. (2013). Spatially adaptive density estimation by localised Haar projections. Ann. Inst. H.Poincaré (Probab-Stat) 49, 900–914. Giné E. and Nickl, R. (2009). An exponential inequality for the distribution function of the kernel density estimator, with application to adaptive estimation. Probab. Theory and Related Fields 143, 569–596. Goldenshluger, A. and Lepski, O.V. (2011). Bandwidth selection in kernel density estimation: oracle inequalities and adaptive minimax optimality. Ann. Statist. 39, 1608–1632. Goldenshluger, A. and Lepski, O.V. (2014). On adaptive minimax density estimation on Rd . Probab. Theory Related Fields, 159, 479–543. Golubev, G.K. (1992). Non-parametric estimation of smooth probability densities. Probl. Inform. Transm. 1, 52–62. de Guzman, M. (1975). Differentiation of Integrals in Rn . With appendices by Antonio Crdoba, and Robert Fefferman, and two by Roberto Moriyn. Lecture Notes in Mathematics, Vol. 481. Springer-Verlag, Berlin-New York. Hall, P. and Meister, A. (2007). A ridge-parameter approach to deconvolution. Ann. Statist., 35, 4, 1535–1558. Hasminskii, R. and Ibragimov, I. (1990). On density estimation in the view of Kolmogorov’s ideas in approximation theory. Ann. Statist. 18, 999–1010. Hesse, C. H. (1995). Deconvolving a density from partially contaminated observations. Journal of Multivariate Analysis 55, 246–260. Juditsky, A. and Lambert–Lacroix, S. (2004). On minimax density estimation on R. Bernoulli 10, 187–220. Kerkyacharian, G., Lepski, O. and Picard, D. (2001). Nonlinear estimation in anisotropic multi–index denoising. Probab. Theory Related Fields 121, 137–170. Kerkyacharian, G., Thanh, M. and Picard, D. (2011). Localized spherical deconvilution. Ann. Statist. 39, 2, 1042–1068. Lepski, O.V. (2013). Multivariate density estimation under sup-norm loss: oracle approach, adaptation and independence structure. Ann. Statist. 41, 2, 1005–1034. 31 Lepski, O.V. (2015). Adaptive estimation over anisotropic functional classes via oracle approach. Ann. Statist. 43, 3, 1178–1242. Lepski, O.V., and Willer T. (2017). Lower bounds in the convolution structure density model. Bernoulli, 23, 2, 884-926. Lepski, O.V., and Willer T. (2016). Estimation in the convolution structure density model. Part I: oracle inequalities. Annals of Statist., submitted. Lepski, O.V. (2017). A new approach to estimator selection. Bernoulli, to appear. arXiv:1603.03934v1, http://arxiv.org. Lounici, K. and Nickl, R. (2011). Global uniform risk bounds for wavelet deconvolution estimators. Ann. Statist. 39, 2, 201–231. Masry, E. (1993). Strong consistency and rates for deconvolution of multivariate densities of stationary processes. Stochastic processes and their applications, 47, 1, 53–74. Meister, A. (2009). Deconvolution problems in nonparametric statistics, 193, Lecture Notes in Statistics. Springer-Verlag, Berlin. Nikol’skii, S. M. (1977). Priblizhenie Funktsii Mnogikh Peremennykh i Teoremy Vlozheniya. (in Russian). [Approximation of functions of several variables and imbedding theorems.] Second edition, revised and supplemented. Nauka, Moscow. Pensky, M. and Vidakovic, B. (1999) Adaptive wavelet estimator for nonparametric density deconvolution. Ann. Statist., 27, 6, 2033–2053. Rebelles, G. (2016) Structural adaptive deconvolution under Lp -losses. Math. Methods Statist. 25, 1, 26-53. Reynaud-Bouret, P., Rivoirard, V. and Tuleau-Malot, C. (2011). Adaptive density estimation: a curse of support? J. Statist. Plann. Inference 141, 115-139. Rigollet, P. (2006). Adaptive density estimation using the blockwise Stein method. Bernoulli 12, 351–370. Rigollet, P. and Tsybakov, A. B. (2007). Linear and convex aggregation of density estimators. Math. Methods Statist. 16, 260–280. Samarov, A. and Tsybakov, A. (2007). Aggregation of density estimators and dimension reduction. Advances in Statistical Modeling and Inference, 233–251, Ser. Biostat., 3, World Sci. Publ., Hackensack, NJ. Stefanski, L. A. (1990). Rates of convergence of some estimators in a class of deconvolution problems. Statist. Probab. Lett., 9, 229–235. Stefanski, L. A. and Carroll, R.J.(1990). Deconvoluting kernel density estimators. Statistics, 21, 1696-184. Yuana, M. and Chenb, J. (2002). Deconvolving multidimensional density from partially contaminated observations. Journal of Statistical Planning and Inference. 104, 147–160. 32	10math.ST
D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA arXiv:1502.05990v5 [math.ST] 8 Nov 2017 Jie Yang1 , Liping Tong2 and Abhyuday Mandal3 1 University of Illinois at Chicago, 2 Advocate Health Care and 3 University of Georgia Abstract: Cumulative link models have been widely used for ordered categorical responses. Uniform allocation of experimental units is commonly used in practice, but often suffers from a lack of efficiency. We consider D-optimal designs with ordered categorical responses and cumulative link models. For a predetermined set of design points, we derive the necessary and sufficient conditions for an allocation to be locally D-optimal and develop efficient algorithms for obtaining approximate and exact designs. We prove that the number of support points in a minimally supported design only depends on the number of predictors, which can be much less than the number of parameters in the model. We show that a D-optimal minimally supported allocation in this case is usually not uniform on its support points. In addition, we provide EW D-optimal designs as a highly efficient surrogate to Bayesian D-optimal designs. Both of them can be much more robust than uniform designs. Key words and phrases: Approximate design, exact design, multinomial response, cumulative link model, minimally supported design, ordinal data. 1. Introduction In this paper we determine optimal and efficient designs for factorial experiments with qualitative factors and ordered categorical responses, or simply ordinal data. Design of experiment with multinomial response, and ordered categories in particular, is becoming increasingly popular in a rich variety of scientific disciplines, especially when human evaluations are involved (Christensen (2015)). Examples include a wine bitterness study (Randall (1989)), potato pathogen experiments (Omer et al. (2000)), a radish seedling’s damping-off study (Krause et al. (2001)), a polysilicon deposition study (Wu (2008)), beef cattle research (Osterstock et al. (2010)), and a toxicity study (Agresti (2013)). This research is motivated by an odor removal study conducted by the textile engineers at the University of Georgia. The scientists studied the manufacture of bio-plastics containing odorous volatiles, that need to be removed before commercialization. For that purpose, a 22 factorial experiment was conducted using algae and synthetic plastic resin blends. The factors were types of algae (x1 : raffinated or solvent extracted algae (−), catfish pond algae (+)) and synthetic resins (x2 : polyethylene (−), 2 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL polypropylene (+)). The response Y had three ordered categories: serious odor (j = 1), medium odor (j = 2), and almost no odor (j = 3). Following traditional factorial design theory, a pilot study with equal numbers (10 in this case) of replicates at each experimental setting was conducted, a uniform design. The results are summarized in Table 1, where yij represents the number of responses falling into the jth category under the ith experimental setting. As demonstrated later (Section 4), the best design identified by our research could improve the efficiency by 25% with only three experimental settings involved. Table 1: Pilot Study of Odor Removal Study Experimental setting i 1 2 3 4 Factor Algae x1 + + − − level Resin x2 + − + − Summarized responses (Y , odor) Serious Medium No odor yi1 yi2 yi3 2 6 2 7 2 1 0 0 10 0 2 8 For such kind of ordinal response Y with J categories and d predictors x = (x1 , . . . , xd )T , the most popular model in practice was first the proportional odds model (also known as cumulative logit model, see Liu and Agresti (2005) for a detailed review). McCullagh (1980) extended it to the cumulative link model (also known as ordinal regression model) g (P (Y ≤ j \| x)) = θj − β T x, j = 1, . . . , J − 1 (1.1) where g is a general link function, with the proportional odds model as a special case when g is the logit link. Examples include the complementary log-log link for the polysilicon deposition study (see Example 6) and the cauchit link for the toxicity study (see Example 9). We adopt the cumulative link model (1.1). When there are only two categories (J = 2), the cumulative link model (1.1) is essentially a generalized linear model for binary data (McCullagh and Nelder (1989); Dobson and Barnett (2008)). For optimal designs under generalized linear models, there is a growing body of literature (see Khuri et al. (2006), Atkinson et al. (2007), Stufken and Yang (2012), and references therein). In this case, it is known that the minimum number of experimental settings required by a nondegenerate Fisher information matrix is d + 1, which equals the number of parameters (Fedorov (1972); 2 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA 3 Yang and Mandal (2015)). A design with the least number of experimental settings, known as a minimally supported design, is of practical significance with a specified regression model due to the cost of changing settings. It is also known that the experimental units should be uniformly assigned when a minimally supported design is adopted for binary response, or under a univariate generalized linear model (Yang and Mandal (2015)). When J ≥ 3, the cumulative link model is a special case of the multivariate generalized linear model (McCullagh (1980)). The relevant results in the optimal design literature are meagre and restricted to the logit link function (Zocchi and Atkinson (1999); Perevozskaya et al. (2003)). Here we obtain theoretical results and efficient algorithms for general link functions and reveal that the optimal designs with J ≥ 3 are quite different from the cases with J = 2. We prove that the minimum number of experimental settings is still d + 1, but strictly less than the number of parameters d + J − 1 (Theorems 3 and 4). This counter-intuitive result is due to the multinomial-type responses: from a single experimental setup, the summarized responses have J − 1 degrees of freedom, requiring fewer distinct experimental settings in a minimally supported design. For the same reason, the allocation of replicates in a minimally supported design is usually not uniform (Section 5), which differs from the traditional factorial design theory. As with generalized linear models, the information matrix under cumulative link models depends on unknown parameters. Different approaches have been proposed to solve the dependence of optimal designs on unknown parameters, including local optimality (Chernoff (1953)), Bayesian approach (Chaloner and Verdinelli (1995)), a maximin approach (Pronzato and Walter (1988); Imhof (2001)), and a sequential procedure (Ford et al. (1989)). As pointed out by Ford, Torsney, and Wu (1992), locally optimal designs are not only important when good initial parameters are available from previous experiments, but can also be a benchmark for designs chosen to satisfy experimental constraints. We mainly focus on locally optimal designs. For situations where local values of the parameters are difficult to obtain, but the experimenter has an idea of the range of parameters with or without a prior distribution, we recommend EW optimal designs, where the Fisher information matrix is replaced by its expected values (Atkinson et al. (2007); Yang, Mandal, and Majumdar (2016)). We compare Bayesian Doptimal designs (Chaloner and Verdinelli (1995)) with EW D-optimal designs for ordinal data. As a surrogate for Bayesian designs, an EW design is much easier to find and retains high efficiency with respect to Bayesian criterion (Section 6). 3 4 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL Among various optimal design criteria, D-optimality, which maximizes the determinant of Fisher information matrix, is the most frequently used (Zocchi and Atkinson (1999)) and often performs well according to other criteria (Atkinson et al. (2007)). We study D-optimal designs. In the design literature, one type of experiment deals with quantitative or continuous factors only. Such a design problem includes the identification of a set of design points {xi }i=1,...,m and the corresponding weights {pi }i=1,...,m (see, for example, Atkinson et al. (2007) and Stufken and Yang (2012)). Numerical algorithms are typically used for cases with two or more factors (see, for example, Woods et al. (2006)). Another type of experiment employs qualitative or discrete factors, where the set of design points {xi }i=1,...,m is predetermined and only the weights {pi }i=1,...,m are to be optimized (see, for example, Yang and Mandal (2015)). One can pick grid points of continuous factors and turn the first kind of problem into the second. Tong, Volkmer, and Yang (2014, Section 5) also bridged the gap between the two types of problems in a way that results involving discrete factors can be applied to the cases with continuous factors. We concentrate on the second kind of design problems and assume that {xi }i=1,...,m are given and fixed. This paper is organized as follows. In Section 2, we describe the preliminary setup and obtain the Fisher information matrix for the cumulative link model with a general link, generalizing Perevozskaya et al. (2003). We also identify a necessary and sufficient condition for the Fisher information matrix to be positive definite, which determines the minimum number of experimental settings required. In Sections 3 and 4, we provide theoretical results and numerical algorithms for searching locally D-optimal approximate or exact designs. In Section 5, we identify analytic D-optimal designs for special cases to illustrate that a D-optimal minimally supported design is usually not uniform on its support points. In Section 6, we illustrate by examples that the EW D-optimal design can be highly efficient with respect to Bayesian D-optimality. We make concluding remarks in Section 7 and relegate additional proofs and results to the supplementary materials. 2. Fisher Information Matrix and Its Determinant Suppose there are m (m ≥ 2) predetermined experimental settings. For the ith experimental setting with corresponding predictors xi = (xi1 , . . . , xid )T ∈ Rd (d ≥ 1), there are ni experimental units assigned to it. Among the ni experimental units, the kth one generates a response Vik which belongs to one of J (J ≥ 2) ordered categories. As shown in Example 2, the dimension d of the predictors can be significantly larger than the number of factors considered in the experiment, which allows more flexible models. 4 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA 5 2.1 General setup In many applications, Vi1 , . . . , Vini are regarded as i.i.d. discrete random variables. Let πij = P (Vik = j), where i = 1, . . . , m; j = 1, . . . , J; and k = 1, . . . , ni . Let Yij = #{k \| Vik = j} be the number of Vik ’s falling into the jth category. Then (Yi1 , . . . , YiJ ) ∼ Multinomial(ni ; πi1 , . . . , πiJ ) . Assumption 1. 0 < πij < 1, i = 1, . . . , m; j = 1, . . . , J. Let γij = P (Vik ≤ j) = πi1 + · · · + πij , j = 1, . . . , J. Based on Assumption 1, 0 < γi1 < γi2 < · · · < γi,J−1 < γiJ = 1 for each i = 1, . . . , m. Consider independent multinomial observations (Yi1 , . . . , YiJ ), i = 1, . . . , m with corresponding predictors x1 , . . . , xm . Under a cumulative link model or ordinal regression model (McCullagh (1980); Agresti (2013); Christensen (2015)), there exists a link function g and parameters of interest θ1 , . . . , θJ−1 , β = (β1 , . . . , βd )T , such that g(γij ) = θj − xTi β, j = 1, . . . , J − 1. (2.1) This leads to m(J −1) equations in d + J −1 parameters (β1 , . . . , βd , θ1 , . . . , θJ−1 ). Assumption 2. The link g is differentiable and its derivative g ′ > 0. Assumption 2 is satisfied for commonly used link functions including logit (log(γ/(1 − γ)), probit (Φ−1 (γ)), log-log (− log(− log(γ))), complementary log-log (log(− log(1−γ))), and cauchit (tan(π(γ−1/2))) (McCullagh and Nelder (1989); Christensen (2015)). Some relevant formulas of these link functions are provided in the supplementary materials (Section S.1). According to Assumption 2, g is strictly increasing, and then θ1 < θ2 < · · · < θJ−1 . Example 1. Consider the logit link g(γ) = log(γ/(1 − γ)) with two predictors and three ordered categories. Model (2.1) consists of 2m equations g(γij ) = θj − xi1 β1 − xi2 β2 , i = 1, . . . , m; j = 1, 2 and parameters (β1 , β2 , θ1 , θ2 ). Under Assumptions 1 and 2, θ1 < θ2 . Example 2. Suppose the model consists of three covariates x1 , x2 , x3 and a few second-order predictors, g(γij ) = θj − xi1 β1 − xi2 β2 − xi3 β3 − xi1 xi2 β12 − x2i1 β11 − x2i2 β22 , where i = 1, . . . , m; j = 1, . . . , J − 1. Then the number of predictors is d = 6. Under the cumulative link model (2.1), the log-likelihood function (up P PJ to a constant) is l(β1 , . . . , βd , θ1 , . . . , θJ−1 ) = m j=1 Yij log(πij ), where i=1 5 6 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL πij = γij − γi,j−1 with γij = g −1 (θj − xTi β) for j = 1, . . . , J − 1 and γi0 = 0, γiJ = 1. Perevozskaya et al. (2003) obtained a detailed form of the Fisher information matrix for logit link and one predictor. Our result is for general link and d predictors; its proof is relegated to the supplementary materials (Section S.3). Theorem 1. Under Assumptions 1 and 2, the Fisher information matrix can be written as m X F= ni Ai (2.2) i=1 where Ai is the (d + J − 1) × (d + J − 1) matrix Ai1 Ai2 (ei xis xit )s=1,...d;t=1,...,d (−xis cit )s=1,...,d;t=1,...,J−1 = ATi2 Ai3 (−cis xit )s=1,...,J−1;t=1,...,d Ai3 and Ai3 is the (J − 1) × (J − 1) symmetric tri-diagonal matrix with diagonal entries ui1 , . . . , ui,J−1, and off-diagonal entries −bi2 , . . . , −bi,J−1 when J ≥ P 3, where ei = Jj=1 πij−1 (gij −gi,j−1 )2 > 0 with gij = (g −1 )′ (θj −xTi β) > 0 for −1 j = 1, . . . , J − 1 and gi0 = giJ = 0; cit = git [πit−1 (git − gi,t−1 ) − πi,t+1 (gi,t+1 − −1 −1 −1 2 git )]; uit = git (πit + πi,t+1 ) > 0; and bit = gi,t−1 git πit > 0. Ai3 contains only one entry ui1 when J = 2. As the Fisher information matrix, F is always positive semi-definite, \|F\| ≥ 0 (Fedorov (1972)). As a special case, Ai is the Fisher information at the experimental setting xi (also known as a design point or support point) and thus is positive semi-definite. 2.2 Determinant of Fisher information matrix Among different criteria for optimal designs, D-criterion looks for the allocation maximizing \|F\|, the determinant of F. Here, a D-optimal design with m predetermined design points x1 , . . . , xm could either be an integerPm valued allocation (n1 , n2 , . . . , nm ) maximizing \|F\| with fixed n = i=1 ni > 0, known as an exact design; or a real-valuedPallocation (p1 , p2 , . . . , pm ) m maximizing \|n−1 F\| with pi = ni /n ≥ 0 and i=1 pi = 1, known as an approximate design. Theorem 2. The determinant of the Fisher information matrix, X cα1 ,...,αm · nα1 1 · · · nαmm , \|F\| = α1 +···+αm =d+J−1 6 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA 7 is an order-(d + J − 1) homogeneous polynomial of (n1 , . . . , nm ) and X \|Aτ \| . (2.3) cα1 ,...,αm = τ ∈(α1 ,...,αm ) The proof of Theorem 2 is relegated to the supplementary materials (Section S.3). Given a map τ : {1, 2, . . . , d + J − 1} → {1, . . . , m}, Aτ in (2.3) is a (d + J − 1) × (d + J − 1) matrix whose kth row is the same as the kth row of Aτ (k) , k = 1, . . . , d + J − 1. We take τ ∈ (α1 , . . . , αm ) where αi = #{j : τ (j) = i} for each i = 1, . . . , m. In order to obtain analytic properties of \|F\|, we need some lemmas. The first of them covers Lemma 1 in Perevozskaya et al. (2003) as a special case: Lemma 1. Rank(Ai ) = Rank(Ai3 ) = J − 1. Furthermore, Ai3 is positive definite and J J−1 Y Y 2 πit−1 > 0 \|Ai3 \| = gis · s=1 −1 ′ where gis = (g ) (θs − xTi β) t=1 > 0 for s = 1, . . . , J − 1. Example 3. Suppose d = 2, J = 3, with link function g. According to Theorem 2, \|F\| is then an order-4 homogeneous polynomial of (n1 , . . . , nm ). Based on Lemma S.4 and Lemma S.5 in the supplementary materials (Section S.2), we can remove all the terms of the form n4i , n3i nj , or n2i n2j from \|F\|. Therefore, \|F\| = m X X cijk · n2i nj nk + i=1 j<k,j6=i,k6=i X cijkl · ni nj nk nl i<j<k<l for some coefficients cijk and cijkl . Based on Lemmas S.4 and S.5, in order to keep cα1 ,...,αm 6= 0, the largest possible αi is J − 1 and the fewest possible number of positive αi ’s is d + 1. Theorem 3. \|F\| > 0 only if m ≥ d + 1. To determine whether d + 1 experimental settings or support points are enough to keep the Fisher information matrix positive definite, we study the leading term of \|F\| with max1≤i≤m αi = J − 1.PFor example, ai0 = J − 1 m for some 1 ≤ i0 ≤ m. From Lemma S.5 and i=1 αi = d + J − 1, to have cα1 ,...,αm 6= 0, there must exist 1 ≤ i1 < i2 < · · · < id ≤ m which are different from i0 , such that, αi1 = · · · = αid = 1. A lemma provides an explicit formula for such a coefficient cα1 ,...,αm : 7 8 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL Lemma 2. Suppose αi0 = J − 1 and αi1 = · · · = αid = 1. Then cα1 ,...,αm = d Y eis · \|Ai0 3 \| · \|X1 [i0 , i1 , . . . , id ]\|2 s=1 where X1 = (1 X) is an m × (d + 1) matrix with 1 = (1, . . . , 1)T , X = (x1 , . . . , xm )T , and X1 [i0 , i1 , . . . , id ] is the sub-matrix consisting of the i0 th, i1 th, . . . , id th rows of X1 . The proof of Lemma 2 is in the supplementary materials (Section S.3). To find D-optimal allocations, we write \|F\| = f (n1 , . . . , nm ) for an order(d + J − 1) homogeneous polynomial function f . The D-optimal exact design problem is to find an integer-valued allocation (n1 , . . . , nm ) maximizing f (n1 , . . . , nm ) subject to ni ∈ {0, 1, . . . , n}, i = 1, . . . , m and n1 + · · · + nm = n with given positive integer n. Denote pi = ni /n, i = 1, . . . , m. According to Theorem 1, f (n1 , . . . , nm ) = m X i=1 ni Ai = n m X pi Ai = nd+J−1 f (p1 , . . . , pm ) (2.4) i=1 Due to (2.4), Theorems 2 and 3 can be directly applied to approximate design problems too: find a real-valued allocation (p1 , . . . , pm ) maximizing f (p1 , p2 , . . . , pm ) subject to 0 ≤ pi ≤ 1, i = 1, . . . , m and p1 + · · · + pm = 1. According to Lemma 1, \|Ai0 3 \| > 0. Thus cα1 ,...,αm in Lemma 2 is positive as long as X1 [i0 , . . . , id ] is of full rank. Theorem 3 implies that a minimally supported design contains at least d + 1 support points, while the following theorem states a necessary and sufficient condition for the minimum number of support points to be exactly d + 1: Theorem 4. f (p) > 0 for some p = (p1 , . . . , pm )T if and only if the extended design matrix X1 = (1 X) is of full rank d + 1. The minimal number of experimental settings required can thus be strictly less than the number of parameters. In the odor removal study, for example, the main-effects cumulative link model (2.1) involves four independent parameters – two β’s for the covariates (d = 2) and two θ’s for the intercepts (J − 1 = 2) – while a minimally supported design could involve only three experimental settings. For multinomial responses with J = 3 categories, we get two degrees of freedom from each experimental setting. Here the optimal allocation of experimental units is often not uniform (see Section 4), contrary to the case of binary responses (Yang, Mandal, and Majumdar (2016); Yang and Mandal (2015)). 8 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA 9 3. D-optimal Approximate Design A (locally) D-optimal approximate design is a real-valued allocation p = (p1 , . . . , pm )T maximizing f (p) = f (p1 , . . . , pm ) with pre-specified values of parameters. The solution always exists since f is continuous and the set of feasible Pm allocations S := {(p1 , . . . , pm )T ∈ Rm \| pi ≥ 0, i = 1, . . . , m; i=1 pi = 1} is convex and compact. A nontrivial D-optimal approximate design problem requires an assumption. Assumption 3. m ≥ d + 1 and Rank(X1 ) = d + 1. Assumption 3 is adopted throughout. With it, the set of valid allocations P S+ := {p = (p1 , . . . , pm )T ∈ S \| f (p) > 0} is nonempty. Since F = m i=1 ni Ai = Pm n i=1 pi Ai is linear in p and φ(·) = log \| · \| is concave on positive semi-definite matrices, f (p) = n1−d−J \|F\| is log-concave (Silvey (1980)) and thus S+ is also convex. Theorem 5. A feasible allocation p = (p1 , . . . , pm )T satisfies f (p) > 0 if and only if Rank(X1 [{i \| pi > 0}]) = d + 1, where X1 [{i \| pi > 0}] is the sub-matrix consisting of the {i \| pi > 0}th rows of X1 . As a direct conclusion of Theorem 5, S+ contains all p whose coordinates are all strictly positive. A special case is the uniform allocation pu = (1/m, . . . , 1/m)T . A necessary and sufficient condition for an approximate design to be Doptimal is of the general-equivalence-theorem type (Kiefer (1974); Pukelsheim (1993); Atkinson et al. (2007); Stufken and Yang (2012); Fedorov and Leonov (2014); Yang, Mandal, and Majumdar (2016)), which is convenient when searching for numerical solutions. Following Yang, Mandal, and Majumdar (2016), for a given p = (p1 , . . . , pm )T ∈ S+ and i ∈ {1, . . . , m}, we set 1−z 1−z 1−z 1−z p1 , . . . , pi−1 , z, pi+1 , . . . , pm (3.1) fi (z) = f 1 − pi 1 − pi 1 − pi 1 − pi with 0 ≤ z ≤ 1. Here fi (z) is well defined as long as pi < 1. Theorem 6. Suppose p = (p1 , . . . , pm )T ∈ S+ and i ∈ {1, . . . , m}. For 0 ≤ z ≤ 1, J−1 X aj z j (1 − z)J−1−j (3.2) fi (z) = (1 − z)d j=0 t−1 ) where a0 = fi (0), (aJ−1 , . . . , a1 = B−1 s,t=1,...,J−1 , and c = J−1 c, BJ−1 = (s T d+J−1 −d J−1 (c1 , . . . , cJ−1 ) with cj = (j + 1) j fi (1/(j + 1)) − j fi (0). )T Following the lift-one algorithm proposed in Yang, Mandal, and Majumdar (2016), we have parallel results and an algorithm for our case. For simplicity, we also call it the lift-one algorithm. 9 10 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL Theorem 7. Given an allocation p = (p∗1 , . . . , p∗m )T ∈ S+ , p is D-optimal if and only if for each i = 1, . . . , m, fi (z), 0 ≤ z ≤ 1 attains it maximum at z = p∗i . A lift-one algorithm 1◦ Start with an allocation p0 = (p1 , . . . , pm )T satisfying f (p0 ) > 0. 2◦ Set up a random order of i going through {1, 2, . . . , m}. 3◦ For each i, determine fi (z) according to Theorem 6, with J determinants fi (0), fi (1/2), fi (1/3), . . . , fi (1/J) calculated according to (3.1). 4◦ Use the quasi-Newton method with gradient defined in (S.13) to find z∗ maximizing fi (z) with 0 ≤ z ≤ 1. If fi (z∗ ) ≤ fi (0), let z∗ = 0. Take (i) p∗ = (p1 (1 − z∗ )/(1 − pi ), . . . , pi−1 (1 − z∗ )/(1 − pi ), z∗ , pi+1 (1 − z∗ )/(1 − (i) pi ), . . . , pm (1 − z∗ )/(1 − pi ))T , so f (p∗ ) = fi (z∗ ). (i) (i) 5◦ Replace p0 with p∗ , and f (p0 ) with f (p∗ ). (i) 6◦ Repeat 2◦ ∼ 5◦ until f (p0 ) = f (p∗ ) for each i. Theorem 8. When the lift-one algorithm converges, the resulting p maximizes f (p). Example 4. Odor removal study Here the response was ordinal in nature, serious odor, medium odor, and no odor. We fit the cumulative link model (2.1) to the data presented in Table 1. The estimated values of the model parameters are (βˆ1 , βˆ2 , θˆ1 , θˆ2 )T = (−2.44, 1.09, −2.67, −0.21)T . If a follow-up experiment is planned and the estimated parameter values are regarded as the true values, the D-optimal approximate allocation found by the lift-one algorithm is po = (0.4449, 0.2871, 0, 0.2680)T . The efficiency of the uniform pu = (1/4, 1/4, 1/4, 1/4)T is (f (pu )/f (po ))1/4 = 79.7%, which is far from satisfactory. Example 5. Wine bitterness study Christensen (2015, Table 1) aggregated the wine data from Randall (1989). It contains the output of a factorial experiment with two treatment factors each at two levels (Temperature x1 : cold (−) or warm (+); Contact x2 : no (−) or yes (+)) affecting wine bitterness. The response was ordinal with five levels (from “1” being least bitter to “5” being most bitter). The original design employed a uniform allocation pu = (1/4, 1/4, 1/4, 1/4)T . The estimated parameter values under the logit link are (βˆ1 , βˆ2 , θˆ1 , θˆ2 , θˆ3 , θˆ4 )T = (1.25, 0.76, −3.36, −0.76, 1.45, 2.99)T . If a follow-up experiment is planned regarding the estimated values of the parameters as the true values, then the D-optimal approximate allocation found by the lift-one algorithm is po = (0.2694, 0.2643, 0.2333, 0.2330)T . The efficiency of the original 10 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA 11 design pu is 99.9%. Nevertheless, the corresponding efficiency may drop to 80% if \|β1 \| and \|β2 \| are both larger than 3 (see Figure 1(a)). In that case, the D-optimal allocations are minimally supported, see Figure 1(b); this is discussed further in Section 5. (b) 4 4 (a) 0.8 0.8 p1=0 2 2 0.85 0.95 0.99 0.99 0.99 0.99 0.95 p1>0, p2>0, p3>0, p4>0 −2 0.9 −2 0.85 0.95 0 0 0.95 β2 0.85 β2 p3=0 0.9 0.9 0.9 −4 −4 −4 0.85 0.8 0.8 −2 0 2 4 p4=0 −4 β1 p2=0 −2 0 2 4 β1 Figure 1: Wine bitterness study with assumed true parameter values (β1 , β2 , −3.36, −0.76, 1.45, 2.99)T : (a) contour plot of efficiency of the original design; (b) regions for a D-optimal design to be minimally supported. In the examples we have studied, the lift-one algorithm often converges within a few iterations. The Yang, Mandal, and Majumdar (2016) lift-one algorithm is guaranteed to converge and can be applied if the lift-one algorithm here does not converge in a pre-specified number of iterations. 4. D-optimal Exact Design In the design literature, different discretization methods have been proposed to round an approximate design into an exact design for a given n, including the quota method (Kiefer (1971); Pukelsheim (1993)) and the efficient rounding procedure (Pukelsheim (1993); Pukelsheim and Rieder (1992)), which usually work well for large enough n but with no guarantee for small sample size (Imhof, Lopez-Fidalgo, and Wong (2001)). In this section, we provide a direct search for D-optimal exact designs. From Theorem 5, we have the result as follows: Corollary 1. \|F\| > 0 if and only if Rank(X1 [{i \| ni > 0}]) = d + 1. We assume n ≥ d + 1 throughout this section. To maximize f (n) = f (n1 , . . . , nm ) = \|F\|, we adopt the exchange algorithm idea of Fedorov (1972). It is 11 12 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL used here to adjust ni and nj simultaneously for randomly chosen (i, j), while keeping ni + nj = c as a constant. We start with an n = (n1 , . . . , nm )T satisfying f (n) > 0. Following Yang, Mandal, and Majumdar (2016), for 1 ≤ i < j ≤ m, let fij (z) = f (n1 , . . . , ni−1 , z, ni+1 , . . . , nj−1 , c − z, nj+1 , . . . , nm ) (4.1) where c = ni + nj , z = 0, 1, . . . , c, so fij (ni ) = f (n). From Theorem 2, Lemmas S.4 and S.5, we have the result as follows. Theorem 9. Suppose n = (n1 , . . . , nm )T satisfies f (n) > 0 and ni + nj ≥ J for given 1 ≤ i < j ≤ m. For z = 0, 1, . . . , ni + nj , fij (z) = J X cs z s (4.2) s=0 where c0 = fij (0), and c1 , . . . , cJ can be obtained using (c1 , . . . , cJ )T = B−1 J (d1 , T t−1 . . . , dJ ) with BJ = (s )st as a J × J matrix and ds = (fij (s) − fij (0))/s. The J × J matrix BJ in Theorem 9 shares the same form of BJ−1 in Theorem 6. According to Theorem 9, in other to maximize fij (z) with z = 0, 1, . . . , ni + nj , one can obtain the exact polynomial form of fij (z) by calculating fij (0), fij (1), . . . , fij (J). There is no practical need to find out the exact form of fij (z) if ni + nj < J since one can simply calculate fij (z) for each z. Following Yang, Mandal, and Majumdar (2016), an exchange algorithm (see the supplementary materials, Section S.5) based on Theorem 9 could be used to search for a D-optimal exact allocation. Example 4 Odor removal study (continued) To conduct a follow-up experiment with n experimental units using the exchange algorithm, we obtain the D-optimal exact designs across different n’s (Table 2). As expected, the D-optimal exact allocation (n1 , . . . , n4 )T is consistent with the D-optimal approximate allocation po = (p1 , . . . , p4 )T (last row of Table 2) for large n. The time costs in seconds (last column of Table 2) are recorded on a PC with 2GHz CPU and 8GB memory. If we rerun an experiment with n = 40, the D-optimal exact design is no = (18, 11, 0, 11)T , and the efficiency of the uniform design nu = (10, 10, 10, 10)T is (f (nu )/f (no ))1/4 = 79.7%. Example 6. Polysilicon deposition study Wu (2008) considered an experiment for studying the polysilicon deposition process with six 3-level factors, described in details by Phadke (1989). Due to the inconvenience of counting the number of surface defects, a major evaluating characteristic, they treated it as a 5-category ordinal variable: 1 for 0 ∼ 3 defects, 2 for 4 ∼ 30, 3 for 12 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA 13 Table 2: D-optimal Exact Designs and the Approximate Design for the Odor Removal Study n 3 10 40 100 1000 po n1 1 4 18 44 445 0.4449 n2 1 3 11 29 287 0.2871 n3 0 0 0 0 0 0 n−4 \|F\| 0.0002911 0.0003133 0.0003177 0.0003180 0.0003181 0.0003181 n4 1 3 11 27 268 0.2680 # iterations 1 3 3 4 5 5 Time(sec.) < 0.01 0.02 0.02 0.05 0.39 0.03 31 ∼ 300, 4 for 301 ∼ 1000, and 5 for 1001 and more. The original design, denoted by nu , includes 18 experimental settings based on an L18 orthogonal array. To apply a cumulative link model, we represent each 3-level factor, say A, with levels 1, 2, 3, by its linear component A1 taking values −1, 0, 1 and a quadratic component A2 taking values 1, −2, 1 (Wu and Hamada (2009)). Then the fitted model with complementary log-log link chosen by both AIC and BIC criteria (see, for example, Agresti (2013)) involves four cut-points (αˆ1 , αˆ2 , αˆ3 , αˆ4 ) = (−1.59, −0.58, 0.41, 1.22), and twelve other coefficients (βˆ11 , βˆ12 , βˆ21 , βˆ22 , . . . , βˆ62 ) = (1.45, −0.22, 1.35, 0.02, −0.12, −0.34, 0.19, 0.00, 0.22, 0.08, 0.05, 0.17). When the true parameter values were assumed to be the estimated ones, we used the exchange algorithm to find a D-optimal 18-run design, denoted by no (see the supplementary materials, Section S.6, for a list of the 18 experimental settings). Compared with no , the efficiency of the original design nu is (f (nu )/f (no ))1/16 = 73.1%. In order to check the efficiency of a rounded design, we used the lift-one algorithm to find that the D-optimal approximate design contains 100 positive pi ’s out of the 729 distinct experimental settings. In this case, both the quota method and the efficient rounding procedure end with the same rounded design nr (see Section S.6). Its efficiency is (f (nr )/f (no ))1/16 = 86.1%. 5. Minimally Supported Design It is of practical significance to have experiment run with the minimal number of different settings. For example, the 18 experimental settings in the polysilicon deposition study (Example 6) had to be run in a sequential way and only two settings were arranged on each day (Phadke (1989)). Less experimental settings often indicate less time and less cost. Another practical application of a minimally supported design is that an optimal allocation restricted to those support points can be obtained more easily or even analytically. According to Theorem 3, a minimally supported design contains at least d+1 support points. On the other hand, according to Theorem 5 and Corollary 1, a minimally supported design could contain exactly d + 1 support points if the extended design matrix X1 = (1 X) is of full rank. 13 14 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL Example 7. Let J = 2 with a binomial response. There are d + 1 parameters, θ1 , β1 , . . . , βd . For a general link function g satisfying Assumptions 1 and 2, 2 /[π (1 − π )], gi0 = gi2 = 0, gi1 = (g−1 )′ (θ1 − xTi β) > 0, ei = ui1 = ci1 = gi1 i1 i1 i = 1, . . . , m. Then Ai3 in Theorem 1 contains only the entry ui1 , and thus \|Ai3 \| = ui1 , or simply ei (Lemma 1 still holds). Assume further that the m × d design matrix X satisfies Assumption 3. According to Theorem 2, Lemmas S.4, S.5, and 2, given p = (p1 , . . . , pm )T , X f (p) = n−(d+1) \|F\| = \|X1 [i0 , i1 , . . . , id ]\|2 pi0 ei0 pi1 ei1 · · · pid eid 1≤i0 <i1 <···<id ≤m (5.1) Here (5.1) is essentially the same as Lemma 3.1 in Yang and Mandal (2015). Then a minimally supported design can contain d + 1 support points and a Doptimal one keeps equal weight 1/(d+1) on all support points (Yang and Mandal (2015, Theorem 3.2)). For univariate responses (including binomial ones) under a generalized linear model, a minimally supported design must keep equal weights on all its support points in order to keep D-optimality (Yang and Mandal (2015)). However, for multinomial responses with J ≥ 3, this is usually not the case. In this section, we use one-predictor (d = 1) and two-predictor (d = 2) cases for illustration. In order to check if a minimally supported design is D-optimal, we need a Karush-Kuhn-Tucker-type condition. Since f (p) is log-concave, the KarushKuhn-Tucker conditions (Karush (1939); Kuhn and Tucker (1951)) are also sufficient. Theorem 10. An allocation p = (p∗1 , . . . , p∗m )T satisfying f (p) > 0 is D-optimal if and only if there exists a λ ∈ R such that ∂f (p)/∂pi = λ if p∗i > 0 or ≤ λ if p∗i = 0, i = 1, . . . , m. 5.1 Minimally supported designs with one predictor We start with d = 1 and J ≥ 3. The corresponding parameters here are β1 and θ1 , . . . , θJ−1 . Consider designs supported on two points (m = 2, minimally supported), and invoke Theorem 2, Lemmas S.4 and S.5. Theorem 11. If d = 1, J ≥ 3, and m = 2, the objective function is f (p1 , p2 ) = n−2 \|F\| = J−1 X cs pJ−s ps2 1 (5.2) s=1 T t−1 ) where (c1 , . . . , cJ−1 )T = B−1 st as a (J − J−1 (d1 , . . . , dJ−1 ) , with BJ−1 = (s J 1) × (J − 1) matrix and ds = f (1/(s + 1), s/(s + 1)) · (s + 1) /s. 14 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA 15 Q QJ −1 2 · 2 Actually, according to Lemma 2, c1 = e2 J−1 g1s s=1 t=1 π1t (x1 − x2 ) , QJ−1 2 QJ −1 cJ−1 = e1 s=1 g2s · t=1 π2t (x1 − x2 )2 , where x1 , x2 are the predictor levels. Theorem 11 provides a way to find the exact form of f (p1 , p2 ) after calculating \|F\| for J − 1 different allocations. Then the D-optimal problem is to maximize an order-J polynomial f (z, 1 − z) for z ∈ [0, 1]. As a special case, the D-optimal allocation of J = 3 can be solved explicitly as follows: Corollary 2. If d = 1, J = 3, and m = 2, the objective function is f (p1 , p2 ) = p1 p2 (c1 p1 + c2 p2 ) (5.3) 2 g 2 (π π π )−1 (x − x )2 > 0, c = e g 2 g 2 (π π π )−1 where c1 = e2 g11 1 2 2 1 21 22 21 22 23 12 11 12 13 (x1 − x2 )2 > 0, and x1 , x2 are the two levels of the predictor. The D-optimal design p = (p∗1 , p∗2 ) is p c c1 − c2 + c21 − c1 c2 + c22 ∗ p p1 , p∗2 = (5.4) p1 = 2 2 2c1 − c2 + c1 − c1 c2 + c2 2c1 − c2 + c21 − c1 c2 + c22 Furthermore, p∗1 = p∗2 = 1/2 if and only if c1 = c2 . Under the setup of Corollary 2, p∗1 = p∗2 = 1/2 if β1 = 0. In general p∗1 6= p∗2 , and p∗1 > p∗2 if and only if c1 > c2 . The following result provides conditions for D-optimality of such a minimally supported design. Its proof is relegated to the supplementary materials (Section S.3). Corollary 3. Suppose d = 1, J = 3, m ≥ 3, and let x1 , . . . , xm be the m distinct levels of the predictor. A minimally supported design p = (p∗1 , p∗2 , 0, . . . , 0)T is D-optimal if and only if (1) p∗1 , p∗2 are defined as in (5.4), (2) si3 (p∗1 )2 + (si5 − 2c1 )p∗1 p∗2 + (si4 − c2 )(p∗2 )2 ≤ 0, i = 3, . . . , m, 2 g 2 (π π π )−1 (x − x )2 > 0, where c1 , c2 are as in Corollary 2, si3 = ei g11 1 i 12 11 12 13 −1 2 2 2 si4 = ei g21 g22 (π21 π22 π23 ) (x2 −xi ) > 0, si5 = e1 (u22 ui1 + u21 ui2 −2b22 bi2 )(x1 − x2 )(x1 − xi ) + e2 (u12 ui1 + u11 ui2 − 2b12 bi2 )(x2 − x1 )(x2 − xi ) + ei (u12 u21 + u11 u22 − 2b12 b22 )(xi − x1 )(xi − x2 ). Example 8. Consider d = 1, J = 3, and m = 3 with factor levels {−1, 0, 1}. Under the logit link g, the parameters β, θ1 , θ2 satisfy g(γ1j ) = θj +β, g(γ2j ) = θj , g(γ3j ) = θj − β, j = 1, 2. We investigate when a D-optimal design is minimally supported. According to Theorem 11, a D-optimal deign satisfies p1 = p3 = 1/2 if β = 0. Figure 2 shows cases with more general parameter values. In Figure 2(a), four regions in (θ1 , θ2 )-plane are occupied by minimally supported 15 16 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL (a) β=−2 (b) θ2=5 4 4 p2=0 p3=0 p1>0, p2>0, p3>0 p1=0 2 2 p3=0 0 θ1 0 θ2 p2=0 −2 −2 p2=0 p1=0 p1>0, p2>0, p3>0 −4 −4 p1>0, p2>0, p3>0 −4 −2 0 2 4 −4 −2 0 2 4 β θ1 Figure 2: Regions for a two-point design to be D-optimal with d = 1, J = 3, x ∈ {−1, 0, 1}, and logit link (note that θ1 < θ2 is required) designs (θ1 < θ2 is required). For example, regions labeled with p2 = 0 indicates a minimally supported design satisfying p2 = 0 is D-optimal given such a triple (θ1 , θ2 , β = −2). From Figure 2(b), a design supported on {−1, 1} (that is, p2 = 0) is D-optimal if β is not far from 0. Example 9. Toxicity study Agresti (2013, Table 8.7) reported data from a developmental toxicity study with one factor (concentration of diEGdiME at five levels: 0, 62.5, 125, 250, 500 mg/kg per day) and a 3-category ordinal response (status of mouse fetus: nonlive, malformation, or normal). In this case, d = 1, J = 3, and m = 5. We fit a cumulative link model with cauchit link chosen by both AIC and BIC criteria. The estimated parameter values are (βˆ1 , θˆ1 , θˆ2 )T = (−0.0176, −8.80, −5.34). If (βˆ1 , θˆ1 , θˆ2 )T is regarded as the true parameter value, then the D-optimal approximate allocation found by the lift-one algorithm is po = (0, 0, 0, 0.4285, 0.5715)T , which is minimally supported. Alternatively, for each pair of indices (i, j), 1 ≤ i < j ≤ 5, we obtain the best design (p∗i , p∗j ) supported only on xi , xj according to Corollary 2, then check whether (p∗i , p∗j ) is D-optimal using Corollary 3. Here po is the only minimally supported design that is also D-optimal. With respect to po , the efficiency of the original design (roughly a uniform one) is 52.6%. 5.2 Minimally supported designs with two predictors In this section, we consider experiments with two predictors (d = 2) and a three-category response (J = 3). The parameters are β1 , β2 , θ1 , θ2 . For cases with J ≥ 4, similar conclusions could be obtained, but with messier notation. 16 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA 17 According to Theorem 3, a minimally supported design needs three support points, for example, (xi1 , xi2 ), i = 1, 2, 3. Under Assumption 3, the 3 × 3 matrix X1 = (1 X) is of full rank. Following Theorem 2, Lemmas S.4, S.5, and 2, the objective function with (d, J, m) = (2, 3, 3) is f (p1 , p2 , p3 ) = \|X1 \|2 e1 e2 e3 · p1 p2 p3 (w1 p1 + w2 p2 + w3 p3 ) (5.5) 2 2 −1 > 0. Since f (p , p , p ) = 0 if p p p = 0, we where wi = e−1 1 2 3 1 2 3 i gi1 gi2 (πi1 πi2 πi3 ) need only consider p = (p1 , p2 , p3 )T satisfying 0 < p1 , p2 , p3 < 1. According to Theorem 10, p maximizes f (p1 , p2 , p3 ) only if ∂f ∂f ∂f = = ∂p1 ∂p2 ∂p3 (5.6) Following Tong, Volkmer, and Yang (2014), we obtain its analytic solution: Theorem 12. Without loss of generality, w1 ≥ w2 ≥ w3 > 0 . The allocation p = (p∗1 , p∗2 , p∗3 )T maximizing f (p1 , p2 , p3 ) in (5.5) exists and is unique. It satisfies 0 < p∗3 ≤ p∗2 ≤ p∗1 < 1 and can be obtained analytically as follows. (i) If w1 ≥ w2 = w3 , then p∗1 =p∆1 /(4w1 + ∆1 ), p∗2 = p∗3 = 2w1 /(4w1 + ∆1 ), where ∆1 = 2w1 − 3w2 + 4w12 − 4w1 w2 + 9w22 . A special case is p∗1 = p∗2 = p∗3 = 1/3 if w1 = w2 = w3 . (ii) If w1 = w2 > w3 , then p∗1 p = p∗2 = ∆2 /[2(∆2 + 2w1 )], p∗3 = 2w1 /(∆2 + 2w1 ), where ∆2 = 3w1 − 2w3 + 9w12 − 4w1 w3 + 4w32 . (iii) If w1 > w2 > w3 , then p∗1 = y1 /(y1 + y2 + 1), p∗2 = y2 /(y1 + y2 + 1), p∗3 = 1/(y1 + y2 + 1), where y1 = − b2 21/3 (3b1 − b22 ) A1/3 + , − 3 3A1/3 3 × 21/3 y2 = (w1 − w3 )y1 (w2 − w3 ) + (w1 − w2 )y1 with A = −27b0 + 9b1 b2 − 2b32 + 33/2 (27b20 + 4b31 − 18b0 b1 b2 − b21 b22 + 4b0 b32 )1/2 , bi = ci /c3 , i = 0, 1, 2, and c0 = w3 (w2 − w3 ) > 0, c1 = 3w1 w2 − w1 w3 − 4w2 w3 +2w32 > 0, c2 = 2w12 −4w1 w2 −w1 w3 +3w2 w3 , c3 = w1 (w2 −w1 ) < 0. The proof of Theorem 12 is relegated to the supplementary materials (Section S.3). Corollary 4. Suppose d = 2, J = 3, and m = 3. Then p = (1/3, 1/3, 1/3)T is D-optimal if and only if w1 = w2 = w3 , where w1 , w2 , w3 are defined as in (5.5). Example 10. Consider a 22 factorial design problem with a three-category response and four design points (1, 1), (1, −1), (−1, 1), (−1, −1), denoted by 17 18 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL 2 2 −1 (xi1 , xi2 ), i = 1, 2, 3, 4. Take wi = e−1 i gi1 gi2 (πi1 πi2 πi3 ) , i = 1, 2, 3, 4. There are five special cases: (i) if β1 = β2 = 0, then w1 = w2 = w3 = w4 ; (ii) if β1 = 0, β2 6= 0, then w1 = w3 , w2 = w4 , but w1 6= w2 ; (iii) if β1 6= 0, β2 = 0, then w1 = w2 , w3 = w4 , but w1 6= w3 ; (iv) if β1 = β2 6= 0, then w2 = w3 , but w1 , w2 , w4 are distinct; (v) if β1 = −β2 6= 0, then w1 = w4 , but w1 , w2 , w3 are distinct. Theorem 12 provides analytic forms of minimally supported designs with d = 2 and J = 3. Corollary 5. Suppose d = 2, J = 3, and m ≥ 4. Let (xi1 , xi2 ), i = 1, . . . , m be m distinct level combinations of the two predictors. With X1 = (1 X) an m × 3 matrix, a minimally supported design p = (p∗1 , p∗2 , p∗3 , 0, . . . , 0)T is D-optimal if and only if p∗1 , p∗2 , p∗3 are obtained according to Theorem 12, and \|X1 [1, 2, i]\|2 e1 e2 ei p∗1 p∗2 (w1 p∗1 + w2 p∗2 ) + \|X1 [1, 3, i]\|2 e1 e3 ei p∗1 p∗3 (w1 p∗1 + w3 p∗3 ) + \|X1 [2, 3, i]\|2 e2 e3 ei p∗2 p∗3 (w2 p∗2 + w3 p∗3 ) + Di p∗1 p∗2 p∗3 ≤ \|X1 [1, 2, 3]\|2 e1 e2 e3 p∗2 p∗3 (2w1 p∗1 + w2 p∗2 + w3 p∗3 ), for i = 4, . . . , m, 2 2 −1 where ej = uj1 + uj2 − 2bj2 , wj = e−1 j gj1 gj2 (πj1 πj2 πj3 ) , j = 1, . . . , m, Di = P {j,k,s,t}∈Ei ej ek (us1 ut2 + us2 ut1 − 2bs2 bt2 ) · \|X1 [j, k, s]\| · \|X1 [j, k, t]\| with the sum over Ei = {(1, 2, 3, i), (1, 3, 2, i), (1, i, 2, 3), (2, 3, 1, i), (2, i, 1, 3), (3, i, 1, 2)}. Example 11. Consider experiments with d = 2, J = 3, m = 4, and design points (1, 1), (1, −1), (−1, 1), (−1, −1). Figure 3 provides the boundary lines of regions of parameters (β1 , β2 , θ1 , θ2 ) for which the best three-point design is D-optimal. In particular, Figure 3(a) shows the region of (β1 , β2 ) for given θ1 , θ2 . It clearly indicates that the best three-point design tends to be D-optimal when the absolute values of β1 , β2 are large. The region tends to be larger as the absolute values of θ1 , θ2 increase. On the other hand, Figure 3(b) displays the region of (θ1 , θ2 ) for given β1 , β2 . The symmetry of the boundary lines about θ1 + θ2 = 0 is due to the logit link which is symmetric about 0. An interesting conclusion based on Corollary 5 is that in this case a three-point design can never be D-optimal if β1 = 0 or β2 = 0. Remark 1. Extra degrees of freedom play an important role against the uniformity of D-optimal allocation in a minimally supported design. For multinomialtype responses with J categories, the total degrees of freedom from m distinct experimental settings is m(J − 1), while a cumulative link model contains d + J − 1 parameters. For a minimally supported design, m = d + 1 and m(J − 1) = d + J − 1 if and only if J = 2 (see Example 7). Then the objective function f (p) ∝ pi0 pi1 · · · pid and the D-optimal allocation is pi0 = pi1 = 18 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA (b) 3 10 (a) θ1=1, θ2=2 θ1=3, θ2=5 θ1=.2, θ2=.5 β1=1, β2=1 β1=1, β2=3 β1=1, β2=.2 0 θ2 0 −3 −10 −2 −5 −1 β2 1 5 2 19 −3 −2 −1 0 1 2 3 −10 β1 −5 0 5 10 θ1 Figure 3: Boundary lines for a three-point design to be D-optimal with logit link: Region of (β1 , β2 ) for given (θ1 , θ2 ) is outside the boundary lines in Panel (a); Region of (θ1 , θ2 ) (with θ1 < θ2 ) for given (β1 , β2 ) is between the boundary lines and θ1 = θ2 in Panel (b) · · · = pid = 1/(d + 1). However, if J ≥ 3, the degrees of freedom is strictly larger than the number of parameters and there are “extra” degrees of freedom. In this case, distinct experimental settings may play different roles in estimating the parameters values. For example, if d = 1, J = 3, m = 2, the objective function f (p) = p1 p2 (c1 p1 + c2 p2 ) according to Corollary 2; if d = 2, J = 3, m = 3, f (p) ∝ p1 p2 p3 (w1 p1 +w2 p2 +w3 p3 ) according to equation (5.5). The D-optimality of a uniform allocation then depends on c1 = c2 or w1 = w2 = w3 , which is not true in general. 6. EW D-optimal Design The previous sections mainly focus on locally D-optimal designs which require assumed parameter values, (β1 , . . . , βd , θ1 , . . . , θJ−1 ). For many applications, the experimenter may have little information about the values of parameters. Then Bayes D-optimality (Chaloner and Verdinelli (1995)) which maximizes E(log \|F\|) given a prior distribution on parameters provides a reasonable solution. An alternative is EW D-optimality (Yang, Mandal, and Majumdar (2016); Atkinson et al. (2007)) which essentially maximizes log \|E(F)\|. According to Yang, Mandal, and Majumdar (2016)’s simulation study across different models and choices of priors, EW D-optimal designs are much easier to calculate and still highly efficient compared with Bayes designs. Based on Theorem 1, an EW D-optimal design that maximizes \|E(F)\| can be viewed as a locally D-optimal design with ei , cit , uit and bit replaced by their expectations. After the replacement, Lemma S.2 still holds. Therefore, almost all results in the previous sections can be applied directly to EW D-optimal designs. 19 20 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL The only exception is Lemma 1 which provides the formula for \|Ai3 \| in terms of gij and πij . In order to find EW D-optimal designs, \|Ai3 \| needs to be calculated in terms of uit and bit . For example, \|Ai3 \| = ui1 if J = 2, \|Ai3 \| = ui1 ui2 − b2i2 if J = 3, and \|Ai3 \| = ui1 ui2 ui3 − ui1 b2i3 − ui3 b2i2 if J = 4. Then the formulas of \|Ai3 \| in Lemma 2, c1 , c2 in Corollary 2, si3 , si4 , si5 in Corollary 3, wi in (5.5), and wj in Corollary 5 need to be written in terms of uit and bit . According to Lemma S.2, we only need to calculate E(uit ), i = 1, . . . , m; t = 1, . . . , J − 1 and E(bit ), i = 1, . . . , m; t = 2, . . . , J − 1 (if J ≥ 3). Then E(cit ) = P E(uit ) − E(bit ) − E(bi,t+1 ) and E(ei ) = J−1 t=1 E(cit ). After that, we can use the lift-one algorithm in Section 3 or the exchange algorithm in Section 4 to find EW D-optimal designs. Example 4 : Odor Removal Study (continued) Instead of assuming the parameter values (β1 , β2 , θ1 , θ2 ) = (−2.44, 1.09, −2.67, −0.21), consider true values of parameters that satisfy β1 ∈ [−3, −1], β2 ∈ [0, 2], θ1 ∈ [−4, −2], and θ2 ∈ [−1, 1]. We assume that the four parameters are independently and uniformly distributed within their intervals. We use R function constrOptim to maximize φ(p) = E(log \|F\|) and find the Bayes D-optimal allocation pb = (0.3879, 0.3264, 0.0000, 0.2857)T . The procedure costs 313 seconds computational time. In order to get the EW D-optimal design, we only need 5.43 seconds in total to calculate E(uit ), E(bit ), and find pe = (0.3935, 0.3259, 0, 0.2806)T using the lift-one algorithm. Even in terms of Bayes Optimality (Chaloner and Larntz (1989); Song and Wong (1998); Abebe et al. (2014)), the relative efficiency of pe with respect to pb is exp{(φ(pe ) − φ(pb ))/4} × 100% = 99.99%, while the relative efficiency of the uniform allocation pu = (0.25, 0.25, 0.25, 0.25)T is 87.67%. In order to check robustness towards misspecified parameter values, we let θ = (β1 , β2 , θ1 , θ2 )T run through all 0.1-grid points in [−3, −1]×[0, 2]×[−4, −2]× [−1, 1]. For each θ, we use the lift-one algorithm to find the D-optimal allocation pθ and the corresponding determinant f (pθ ) = \|F(pθ )\|, and then calculate the efficiency (f (p)/f (pθ ))1/4 for p = pb , pe , and pu , respectively. Table 3 shows the summary statistics of the efficiencies. It implies that pb and pe are comparable and both of them are much better than pu in terms of robustness. Table 3: Summary of Efficiency in Odor Removal Study Design Bayes pb EW pe Uniform pu Min. 0.8464 0.8465 0.7423 1st Quartile 0.9813 0.9802 0.8105 Median 0.9915 0.9917 0.8622 7. Discussion 20 Mean 0.9839 0.9838 0.8674 3rd Quartile 0.9964 0.9967 0.9249 Max. 1.0000 1.0000 0.9950 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA 21 In this paper, we use real experiments to illustrate how much improvements the experimenter could make. Compared with our D-optimal designs, the efficiencies of the original designs are often far from satisfactory: 79.7% in Example 4, 73.1% in Example 6, and 52.6% in Example 9. More interestingly, our D-optimal designs recommended for Example 4 and Example 9 are both minimally supported. We have two surprising findings that are different from the cases under univariate generalized linear models (Yang and Mandal, 2015): (1) the minimum number of experimental settings can be strictly less than the number of parameters, and (2) the allocation of experimental units on the support points of a minimally supported design is usually not uniform. Cumulative link models are widely used for modeling ordinal data. Nevertheless, there are other models used for multinomial-type responses, including baseline-category logit model for nominal response, adjacent-categories logit model for ordinal data, and continuation-ratio logit model for hierarchical response (see Liu and Agresti (2005), Agresti (2013) for a review). The methods developed in this paper could be extended for those models as well. For further extensions, our approaches could be used for planning experiments with more than one categorical response. For example, both the paper feeder experiment and the PCB experiment analyzed by Joseph and Wu (2004) involved multiple binomial responses. Supplementary Materials The proofs of Theorems 1, 2, 4, 5, and 12, Lemma 2, and Corollaries 3 and 5 are available in the Supplementary Materials. There are also tabularized formulas for commonly used link functions, additional lemmas for Section 2 and Section 5.2, maximization of fi (z) in Section 3, exchange algorithm for D-optimal exact allocation in Section 4, and more results for Example 6. Acknowledgements We thank Dr. Suraj Sharma for providing the details of the odor removal study, and Dr. John Stufken for valuable suggestions on an early version of this paper. We also thank an associate editor and the reviewers for comments and suggestions that substantially improved the quality of the manuscript. This research is in part supported by the LAS Award for Faculty of Science at UIC. References Abebe, H. T., Tan, F. E., Van Breukelen, G. J., Serroyen, J., and Berger, M. P. (2014). On the choice of a prior for Bayesian D-optimal designs for the logistic regression model with a single predictor. Communications in Statistics - Simulation and Computation, 43, 18111824. 21 22 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL Agresti, A. (2013). Categorical Data Analysis, Third Edition. Wiley, New Jersey. Atkinson, A. C., Donev, A. N. and Tobias, R. D. (2007). Optimum Experimental Designs, with SAS. Oxford University Press, New York. Chaloner, K. and Larntz, K. (1989). Optimal Bayesian design applied to logistic regression experiments. Journal of Statistical Planning and Inference, 21, 191-208. Chaloner, K. and Verdinelli, I. (1995). Bayesian experimental design: a review. Statistical Science, 10, 273-304. Chernoff, H. (1953). Locally optimal designs for estimating parameters. Annals of Mathematical Statistics, 24, 586-602. Christensen, link R. H. models – B. (2015). estimation Analysis with of the ordinal R-package data with ordinal. cumulative Available via http://cran.r-project.org/web/packages/ordinal/vignettes/clm_intro.pdf Dobson, A. J. and Barnett, A. (2008). An Introduction to Generalized Linear Models, Third Edition. Chapman & Hall/CRC, London. Fedorov, V. V. (1972). Theory of Optimal Experiments. Academic Press, New York. Fedorov, V. V. and Leonov, S. L. (2014). Optimal Design for Nonlinear Response Models. Chapman & Hall/CRC, New York. Ford, I., Titterington, D. M., and Kitsos, C. P. (1989). Recent advances in nonlinear experimental design. Technometrics, 31, 49-60. Ford, I., Torsney, B., and Wu, C. F. J. (1992). The use of a canonical form in the construction of locally optimal designs for non-linear problems. Journal of the Royal Statistical Society, Series B, 54, 569-583. Imhof, L. A. (2001). Maximin designs for exponential growth models and heteroscedastic polynomial models. Annals of Statistics, 29, 561-576. Imhof, L., Lopez-Fidalgo, J., and Wong, W.K. (2001). Efficiencies of rounded optimal approximate designs for small samples. Statistica Neerlandica, 55, 301-318. Joseph, V. R. and Wu, C. F. J. (2004). Failure amplification method: an information maximization approach to categorical response optimization (with discussions). Technometrics, 46, 1-31. Karush, W. (1939). Minima of functions of several variables with inequalities as side constraints, M.Sc. Dissertation, Department of Mathematics, University of Chicago. Kiefer, J. (1971). The role of symmetry and approximation in exact design optimality. In Statistical Decision Theory and Related Topics, S.S. Gupta and J. Yackel (eds), 109-118, Academic Press, New York. Kiefer, J. (1974). General equivalence theory for optimum designs (approximate theory). Annals of Statistics, 2, 849-879. Khuri, A. I., Mukherjee, B., Sinha, B. K. and Ghosh, M. (2006). Design issues for generalized linear models: A review. Statistical Science, 21, 376-399. 22 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA 23 Krause, M. S., Madden, L. V. and Hoitink, H. A. J. (2001). Effect of potting mix microbial carrying capacity on biological control of Rhizoctonia damping-off of radish and Rhizoctonia crown and root rot of Poinsettia. Phytopathology, 91, 1116-1123. Kuhn, H. W. and Tucker, A. W. (1951). Nonlinear programming. Proceedings of 2nd Berkeley Symposium, Berkeley: University of California Press, 481-492. Liu, I. and Agresti, A. (2005). The analysis of ordered categorical data: An overview and a survey of recent developments. Test, 14, 1-73. McCullagh, P. (1980). Regression models for ordinal data. Journal of the Royal Statistical Society, Series B, 42, 109-142. McCullagh, P. and Nelder, J. (1989). Generalized Linear Models, Second Edition. Chapman and Hall/CRC, Boca Raton. Omer, M. A., Johnson, D. A. and Rowe, R. C. (2000). Recovery of Verticillium dahliae from North American certified seed potatoes and characterization of strains by vegetative compatibility and aggressiveness. American Journal of Potato Research, 77, 325-331. Osterstock, J. B., MacDonald, J. C., Boggess, M. M. and Brown, M. S. (2010). Analysis of ordinal outcomes from carcass data in beef cattle research. Journal of Animal Science, 88, 3384-3389. Perevozskaya, I., Rosenberger, W. F. and Haines, L. M. (2003). Optimal design for the proportional odds model. The Canadian Journal of Statistics, 31, 225-235. Phadke, M. S. (1989). Quality Engineering using Robust Design. Prentice-Hall, Englewood Cliffs. Pronzato, L. and Walter, E. (1988). Robust experiment design via maximin optimization. Mathematical Biosciences, 89, 161-176. Pukelsheim, F. (1993). Optimal Design of Experiments. John Wiley & Sons, New York. Pukelsheim, F. and Rieder, S. (1992). Efficient rounding of approximate designs. Biometrika, 79, 763-770. Randall, J. (1989). The analysis of sensory data by generalised linear model. Biometrical Journal, 31, 781-793. Silvey, S. D. (1980). Optimal Design. Chapman and Hall, London. Song, D. and Wong, W. K. (1998). Optimal two-point designs for the Michaelis-Menten model with heteroscedastic errors. Communications in Statistics - Theory and Methods, 27, 15031516. Stufken, J. and Yang, M. (2012). Optimal designs for generalized linear models. In: Design and Analysis of Experiments, Volume 3: Special Designs and Applications, K. Hinkelmann (ed.). Wiley, New York. Tong, L., Volkmer, H. W., and Yang, J. (2014). Analytic solutions for D-optimal factorial designs under generalized linear models. Electronic Journal of Statistics, 8, 1322-1344. Woods, D. C., Lewis, S. M., Eccleston, J. A. and Russell, K. G. (2006). Designs for generalized linear models with several variables and model uncertainty. Technometrics, 48, 284-292. 23 24 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL Wu, F-C. (2008). Simultaneous optimization of robust design with quantitative and ordinal data. International Journal of Industrial Engineering: Theory, Applications and Practice, 15, 231-238. Wu, C. F. J. and Hamada, M. (2009). Experiments: Planning, Analysis, and Optimization, Second Edition. Wiley, New York. Yang, J. and Mandal, A. (2015). D-optimal factorial designs under generalized linear models. Communications in Statistics - Simulation and Computation, 44, 2264-2277. Yang, J., Mandal, A., and Majumdar, D. (2016). Optimal designs for 2k factorial experiments with binary response. Statistica Sinica, 26, 385-411. Zocchi S. S. and Atkinson, A. C. (1999). Optimum experimental designs for multinomial logistic models. Biometrics, 55, 437-444. University of Illinois at Chicago E-mail: jyang06@math.uic.edu Phone: (312) 413-3748 Fax: (312) 996-1491 Advocate Health Care E-mail: lipingtong@gmail.com University of Georgia E-mail: amandal@stat.uga.edu 24 SUPPLEMENTARY MATERIALS S1 D-OPTIMAL DESIGNS WITH ORDERED CATEGORICAL DATA Jie Yang1 , Liping Tong2 and Abhyuday Mandal3 1 University of Illinois at Chicago, 2 Advocate Health Care and 3 University of Georgia Supplementary Materials S.1 Commonly Used Link Functions for Cumulative Link Models Link function logit probit log-log c-log-log cauchit g(γ) γ log 1−γ Φ−1 (γ) − log[− log(γ)] log[− log(1 − γ)] tan[π(γ − 21 )] g −1 (η) (g −1 )′ (η) eη 1+eη eη (1+eη )2 Φ(η) exp{−e−η } 1 − exp{−eη } 1 arctan(η) + 21 π φ(η) exp{−η − e−η } exp{η − eη } 1 π(1+η2 ) where Φ−1 (·) is the cumulative distribution function of N(0, 1), φ(·) is the probability density function of N(0, 1), and “c-log-log” stands for complementary log-log. Example 1 (continued) For logit link g, g −1 (η) = eη /(1+eη ) and (g −1 )′ = g −1(1 − g −1 ). Thus gij = (g −1)′ (θj − xTi β) = γij (1 − γij ). With J = 3, we have πi1 + πi2 + πi3 = 1 for i = 1, . . . , m. Then for i = 1, . . . , m, gi1 = −1 πi1 (πi2 + πi3 ), gi2 = (πi1 + πi2 )πi3 , bi2 = πi1 πi3 πi2 (πi1 + πi2 )(πi2 + πi3 ), ui1 = −1 −1 2 πi1 πi2 (πi1 +πi2 )(πi2 +πi3 ) , ui2 = πi3 πi2 (πi1 +πi2 )2 (πi2 +πi3 ), ci1 = πi1 (πi1 + πi2 )(πi2 +πi3 ), ci2 = πi3 (πi1 +πi2 )(πi2 +πi3 ), ei = (πi1 +πi2 )(πi1 +πi3 )(πi2 +πi3 ). S.2 Additional Lemmas For Section 2: Since (Yi1 , . . . , YiJ ), i = 1, . . . , m are m independent random vectors, the log-likelihood function (up to a constant) of the cumulative link model is l(β1 , . . . , βd , θ1 , . . . , θJ−1 ) = m X J X i=1 j=1 S1 Yij log(πij ) S2 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL where πij = γij − γi,j−1 with γij = g −1 (θj − xTi β) for j = 1, . . . , J − 1 and γi0 = 0, γiJ = 1, i = 1, . . . , m. For s = 1, . . . , d, t = 1, . . . , J − 1, m X ∂l Yi1 = (−xis ) · · (g −1)′ (θ1 − xTi β) ∂βs π i1 i=1 Yi2 −1 ′ · (g ) (θ2 − xTi β) − (g −1 )′ (θ1 − xTi β) + πi2 YiJ −1 ′ T +··· + −(g ) (θJ−1 − xi β) πiJ m X ∂l Yit Yi,t+1 −1 ′ T = (g ) (θt − xi β) − ∂θt π πi,t+1 it i=1 Since Yij ’s come from multinomial distributions, we know E(Yij ) = ni πij , E(Yij2 ) = ni (ni − 1)πij2 + ni πij , and E(Yis Yit ) = ni (ni − 1)πis πit when s 6= t. Then we have the following lemma: Lemma S.1. Let F = (Fst ) be the (d+J −1)×(d+J −1) Fisher information matrix. (i) For 1 ≤ s ≤ d, 1 ≤ t ≤ d, X m J X (gij − gi,j−1)2 ∂l ∂l Fst = E = ni xis xit ∂βs ∂βt πij i=1 j=1 where gij = (g −1 )′ (θj −xTi β) > 0 for j = 1, . . . , J −1 and gi0 = giJ = 0. (ii) For 1 ≤ s ≤ d, 1 ≤ t ≤ J − 1, X m git − gi,t−1 gi,t+1 − git ∂l ∂l = ni (−xis )git − Fs,d+t = E ∂βs ∂θt π πi,t+1 it i=1 (iii) For 1 ≤ s ≤ J − 1, 1 ≤ t ≤ d, X m gis − gi,s−1 gi,s+1 − gis ∂l ∂l = ni (−xit )gis − Fd+s,t = E ∂θs ∂βt πis πi,s+1 i=1 (iv) For 1 ≤ s ≤ J − 1, 1 ≤ t ≤ J − 1,  P −1 −1 2  m ni gis (πis + πi,s+1 ), Pi=1 ∂l ∂l m −1 Fd+s,d+t = E = n g g (−πi,s∨t ),  i=1 i is it ∂θs ∂θt 0, where s ∨ t = max{s, t}. S2 if s = t if \|s − t\| = 1 if \|s − t\| ≥ 2 SUPPLEMENTARY MATERIALS S3 Perevozskaya et al. (2003) obtained a detailed form of Fisher information matrix for logit link and one predictor. Our expressions here are good for fairly general link and d predictors. To simplify the notations, we denote for i = 1, . . . , m, ei = J X (gij − gi,j−1)2 j=1 cit uit πij >0 git − gi,t−1 gi,t+1 − git = git , t = 1, . . . , J − 1 − πit πi,t+1 −1 = git2 (πit−1 + πi,t+1 ) > 0, t = 1, . . . , J − 1 bit = gi,t−1 git πit−1 > 0, t = 2, . . . , J − 1 (if J ≥ 3) (S.1) (S.2) (S.3) (S.4) Note that gij is defined in Lemma S.1 (i). Then we obtain the following lemma which plays a key role in calculating \|F\|. Lemma S.2. cit = uit − bit − bi,t+1 , i = 1, . . . , m; t = 1, . . . , J − 1; ei = PJ−1 PJ−1 t=1 (uit − 2bit ), i = 1, . . . , m, where bi1 = biJ = 0 for i = t=1 cit = 1, . . . , m. Lemma S.3. Rank((Ai1 Ai2 )) ≤ 1 where “=” is true if and only if xi 6= 0. Based on Lemmas 1 and S.3, we obtain the two lemmas below on cα1 ,...,αm which significantly simplify the structure of \|F\| as a polynomial of (n1 , . . . , nm ). Lemma S.4. If max1≤i≤m αi ≥ J, then \|Aτ \| = 0 for any τ ∈ (α1 , . . . , αm ) and thus cα1 ,...,αm = 0. Proof of Lemma S.4: Without any loss of generality, we assume α1 ≥ α2 ≥ · · · ≥ αm . Then max1≤i≤m αi ≥ J implies α1 ≥ J. In this case, for any τ ∈ (α1 , . . . , αm ), τ −1 (1) := {i \| τ (i) = 1} ⊂ {1, . . . , d + J − 1} and \|τ −1 (1)\| = α1 . If \|τ −1 (1)∩{1, . . . , d}\| ≥ 2, then \|Aτ \| = 0 due to Lemma S.3; otherwise {d+1, . . . , d+J −1} ⊂ τ −1 (1) and thus \|Aτ \| = 0 due to Lemma 1. Thus cα1 ,...,αm = 0 according to (2.3) provided in Theorem 2. Lemma S.5. If #{i : αi ≥ 1} ≤ d, then \|Aτ \| = 0 for any τ ∈ (α1 , . . . , αm ) and thus cα1 ,...,αm = 0. Proof of Lemma S.5: Without any loss of generality, we assume α1 ≥ α2 ≥ · · · ≥ αm . Then #{i : αi ≥ 1} ≤ d indicates αd+1 = · · · = αm = 0. Let τ : {1, 2, . . . , d + J − 1} → {1, . . . , m} satisfy τ ∈ (α1 , . . . , αm ). Then the (d + J − 1) × (d + J − 1) matrix Aτ can be written as Aτ 1 Aτ 2 Aτ 3 Aτ 4 S3 S4 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL = (eτ (s) xτ (s)s xτ (s)t )s=1,...d;t=1,...,d (−xτ (s)s cτ (s)t )s=1,...,d;t=1,...,J−1 (−cτ (d+s)s xτ (d+s)t )s=1,...,J−1;t=1,...,d Aτ 4 where the (J −1)×(J −1) matrix Aτ 4 is either a single entry uτ (d+1)1 (if J = 2) or symmetric tri-diagonal with diagonal entries uτ (d+1)1 , . . . , uτ (d+J−1),J−1 , upper off-diagonal entries −bτ (d+1)2 , . . . , −bτ (d+J−2),J−1 , and lower off-diagonal entries −bτ (d+2)2 , . . . , −bτ (d+J−1),J−1 . Note that Aτ is asymmetric in general. If #{i : αi ≥ 1} ≤ d−1, then there exists an i0 such that 1 ≤ i0 ≤ d and \|τ −1 (i0 ) ∩ {1, . . . , d}\| ≥ 2. In this case, \|Aτ \| = 0 according to Lemma S.3. If #{i : αi ≥ 1} = d, we may assume \|τ −1 (i) ∩ {1, . . . , d}\| = 1 for i = 1, . . . , d (otherwise \|Aτ \| = 0 according to Lemma S.3). Suppose α1 ≥ α2 ≥ · · · ≥ αk ≥ 2 > αk+1 . Then {d + 1, . . . , d + J − 1} ⊂ ∪ki=1 τ −1 (i) P and ki=1 (αi − 1) = J − 1. In order to show \|Aτ \| = 0, we first replace (1) (1) Aτ 1 with Aτ 1 = (eτ (s) xτ (s)t )s=1,...d; t=1,...,d and replace Aτ 2 with Aτ 2 = (1) (−cτ (s)t )s=1,...,d; t=1,...,J−1 . It changes Aτ into a new matrix Aτ . Note that Q (1) \|Aτ \| = ds=1 xτ (s)s · \|Aτ \|. According to Lemma S.2, the sum of the columns (1) of Aτ 2 is (−eτ (1) , . . . , −eτ (d) )T , and the elementwise sum of the columns of Aτ 4 is (cτ (d+1)1 , cτ (d+2)2 , . . . , cτ (d+J−1),J−1 )T . Secondly, for t = 1, . . . , d, we add x1t (−eτ (1) , . . . , −eτ (d) , cτ (d+1)1 , . . . , cτ (d+J−1),J−1 )T to the tth column of (1) (2) (1) (2) Aτ . We denote the resulting matrix by Aτ . Note that \|Aτ \| = \|Aτ \|. We (2) (2) consider the sub-matrix Aτ d which consists of the first d columns of Aτ . (2) For s ∈ τ −1 (1), the sth row of Aτ d is simply 0. For i = 2, . . . , k, the jth row (2) of Aτ d is proportional to (xi1 − x11 , xi2 − x12 , . . . , xid − x1d ) if j ∈ τ −1 (i). P (2) Therefore, Rank(Aτ d ) ≤ (d + J − 1) − α1 − ki=2 (αi − 1) = d − 1, which (1) (2) leads to \|Aτ \| = 0 and thus \|Aτ \| = 0, \|Aτ \| = 0. According to (2.3) in Theorem 2, cα1 ,...,αm = 0. Lemma S.6. F = F(p) is always positive semi-definite. It is positive definite if and only if p ∈ S+ . Furthermore, log f (p) is concave on S. For Section 5.2: The procedure seeking for analytic solutions here follows Tong, Volkmer, and Yang (2014). As a direct conclusion of the KarushKuhn-Tucker conditions (see also Theorem 10), a necessary condition for (p1 , p2 , p3 ) to maximize f (p1 , p2 , p3 ) in (5.5) is (5.6), which are equivalent to ∂f /∂p1 = ∂f /∂p3 and ∂f /∂p2 = ∂f /∂p3 . In terms of pi , wi ’s, they are (p3 − p1 )(p1 w1 + p2 w2 + p3 w3 ) = (w3 − w1 )p1 p3 (p3 − p2 )(p1 w1 + p2 w2 + p3 w3 ) = (w3 − w2 )p2 p3 S4 (S.5) (S.6) SUPPLEMENTARY MATERIALS S5 Denote y1 = p1 /p3 > 0 and y2 = p2 /p3 > 0. Since p1 + p2 + p3 = 1, it implies p3 = 1/(y1 + y2 + 1), p1 = y1 /(y1 + y2 + 1), and p2 = y2 /(y1 + y2 + 1). In terms of y1 , y2 , (S.5) and (S.6) are equivalent to (1 − y1 )(y1 w1 + y2 w2 + w3 ) = (w3 − w1 )y1 (1 − y2 )(y1 w1 + y2 w2 + w3 ) = (w3 − w2 )y2 (S.7) (S.8) Lemma S.7. Suppose 0 < w3 < w2 < w1 . If (p1 , p2 , p3 ) maximizes f (p1 , p2 , p3 ) in (5.5) under the constrains p1 , p2 , p3 ≥ 0 and p1 + p2 + p3 = 1, then 0 < p3 ≤ p2 ≤ p1 < 1. The proof of the lemma above is straightforward, because otherwise one could exchange pi , pj to strictly improve f (p1 , p2 , p3 ) . Now we are ready to get solutions to equations (S.7) and (S.8) case by case. (i) w1 = w3 . In that case, (S.7) implies y1 = 1. After plugging it into (S.8), the only positive p solution is y2 = (−3w1 + 2w2 + 9w12 − 4w1 w2 + 4w22 )/(2w2) . (ii) w2 = w3 . In that case, (S.8) implies y2 = 1. After plugging it into (S.7), the only positive p solution is y1 = (2w1 − 3w2 + 4w12 − 4w1 w2 + 9w22 )/(2w1 ) . (iii) w1 = w2 but w1 6= w3 . The ratio of (S.7) and (S.8) leads to y1 = y2 . After plugging it into (S.7), the only positive solution is y1 = (3w1 − p 2 2w3 + 9w1 − 4w1 w3 + 4w32)/(4w1 ) . (iv) w1 , w2 , w3 are distinct. Without any loss of generality, we assume 0 < w3 < w2 < w1 , because otherwise the previous elimination procedure in the order of p3 , p2 , p1 could be easily changed accordingly. Based on Lemma S.7, if (p1 , p2 , p3 ) maximizes f4 , then 0 < p3 ≤ p2 ≤ p1 < 1 and thus y1 ≥ y2 ≥ 1. The ratio of (S.7) and (S.8) leads to (1 − y1 )/(1 − y2 ) = (w3 − w1 )/(w3 − w2 ) · y1 /y2 , which implies y2 = (w1 − w3 )y1 . (w2 − w3 ) + (w1 − w2 )y1 (S.9) Note that (w2 − w3 ) + (w1 − w2 )y1 ≥ w1 − w3 > 0. After plugging (S.9) into (S.7), we get c0 + c1 y1 + c2 y12 + c3 y13 = 0 (S.10) where c0 = w3 (w2 − w3 ) > 0, c1 = 3w1 w2 − w1 w3 − 4w2 w3 + 2w32 > 0, c2 = 2w12 − 4w1 w2 − w1 w3 + 3w2 w3 , c3 = w1 (w2 − w1 ) < 0. S5 S6 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL Lemma S.8. Suppose 0 < w3 < w2 < w1 . Then equation (S.10) has one and only one solution y1∗ ≥ 1. Furthermore, y1∗ > 1. Proof of Lemma S.8: In order to locate the roots of equation (S.10), we let f1 (y1 ) = c0 + c1 y1 + c2 y12 + c3 y13 . Then f1 (1) = c0 + c1 + c2 + c3 = (w1 − w3 )2 > 0. On the other hand, the first derivative of f1 is f1′ (y1 ) = a0 +a1 y1 +a2 y12 , where a0 = 3w1 w2 − w1 w3 − 4w2 w3 + 2w32 = w1 (w2 − w3 ) + 2(w1 − w2 )w2 + 2(w2 − w3 )2 > 0, a1 = 2(2w12 − 4w1 w2 − w1 w3 + 3w2 w3 ), and a2 = 3w1 (w2 − w1 ) < 0. Therefore, a21 − 4a0 a2 > a21 ≥ 0 and f1′ (y1 ) = a2 (y1 − y11 )(y1 − y12 ), where p p −a1 + a21 − 4a0 a2 −a1 − a21 − 4a0 a2 y11 = < 0, y12 = > y11 2a2 2a2 It can be verified that y12 < 1 if and only if w1 < 2(w2 + w3 ). There are two cases: Case (i): If y12 < 1, then f1′ (y1 ) < 0 for all y1 > 1. That is, f1 (y1 ) strictly decreases after y1 = 1. Since f1 (1) > 0 and f1 (∞) = −∞, then there is one and only one solution in (1, ∞); Case (ii): If y12 ≥ 1, then f1′ (y1 ) ≥ 0 for y1 ∈ [1, y12 ] and f1′ (y1 ) < 0 for y1 ∈ (y12 , ∞). That is, f1 (y1 ) increases in [1, y12 ] and then strictly decreases in (y12 , ∞). Again, due to f1 (1) > 0 and f1 (∞) = −∞, there is one and only one solution in (1, ∞). In either case, the conclusion is justified. S.3 Additional Proofs Proof of Theorem 1 It is a direct conclusion of Lemmas S.1 and S.2. Examples of Ai3 in Theorem 1 include (ui1 ),     u −b 0 0 i1 i2 ui1 −bi2 0  −bi2 ui1 −bi2 ui2 −bi3 0    ui2 −bi3  ,  , −bi2  −bi2 ui2 0 −bi3 ui3 −bi4  0 −bi3 ui3 0 0 −bi4 ui4 for J = 2, 3, 4, or 5 respectively. Proof of Theorem 2 To study the structure of \|F\| as a polynomial (i) function of (n1 , . . . , nm ), we denote the (k, l)th entry of Ai by akl . Given a row map τ : {1, 2, . . . , d + J − 1} → {1, . . . , m}, we define a (d + J − 1) × (τ (k)) whose kth row is given by the kth row of (d + J − 1) matrix Aτ = akl S6 SUPPLEMENTARY MATERIALS S7 A Pτm(k) . For a power index (α1 , . . . , αm ) with αi ∈ {0, 1, . . . , d + J − 1} and i=1 αi = d + J − 1, we denote τ ∈ (α1 , . . . , αm ) if αi = #{j : τ (j) = i} for each i = 1, . . . , m. In terms of the construction of Aτ , it says that αi rows of Aτ are from the matrix Ai . According to the Leibniz formula for the determinant, \|F\| = m X X ni Ai = i=1 (−1) sgn(σ) σ∈Sd+J −1 d+J−1 m Y X k=1 (i) ni ak,σ(k) i=1 where σ is a permutation of {1, 2, . . . , d + J − 1}, and sgn(σ) is the sign or signature of σ. Therefore, cα1 ,...,αm = X (−1) sgn(σ) σ∈Sd+J −1 = X = d+J−1 Y ak,σ(k) d+J−1 Y ak,σ(k) (τ (k)) τ ∈(α1 ,...,αm ) k=1 X τ ∈(α1 ,...,αm ) σ∈Sd+J −1 X X (−1) sgn(σ) (τ (k)) k=1 \|Aτ \| τ ∈(α1 ,...,αm ) Proof of Lemma 2 To simplify the notations, we let is = s + 1, s = 0, . . . , d. That is, α1 = J − 1, α2 = · · · = αd+1 = 1. There are only two types of τ ∈ (α1 , . . . , αm ), such that, \|Aτ \| may not be 0. τ of type I: There exist 1 ≤ k ≤ d, 2 ≤ l ≤ d + 1, and 1 ≤ q ≤ J − 1, such that, τ (k) = 1 and τ (d + q) = l. Following a similar procedure as in the proof of Lemma S.5, we obtain \|Aτ \| = d+1 Y d ei · \|A13 \| · (−1) \|X1 [1, 2, . . . , d + 1]\| · (−1) sgn(τ ) d Y xτ (s)s · s=1 i=2 clq el τ of type II: τ (d + 1) = · · · = τ (d + J − 1) = 1 and {τ (1), . . . , τ (d)} = {2, . . . , d + 1}. It can be verified that \|Aτ \| = d+1 Y d ei · \|A13 \| · (−1) \|X1 [1, 2, . . . , d + 1]\| · (−1) sgn(τ ) d Y s=1 i=2 S7 xτ (s)s S8 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL According to Theorem 2, X X \|Aτ \| + \|Aτ \| cα1 ,...,αm = τ of type I τ of type II  d X d+1 d+1 X Y ei · \|A13 \| · (−1)d \|X1 [1, 2, . . . , d + 1]\| ·  = i=2 k=1 l=2 τ ∈Sd+1 :τ (k)=1,τ (d+1)=l (−1)sgn(τ ) d Y s=1 = xτ (s)s J−1 X clq q=1 el X + = (−1)sgn(τ ) d Y s=1 τ ∈Sd+1 :τ (d+1)=1  xτ (s)s  d Y d+1 Y ei · \|A13 \| · (−1) \|X1 [1, 2, . . . , d + 1]\| · d+1 Y ei · \|A13 \| · (−1)d \|X1 [1, 2, . . . , d + 1]\| · (−1)d \|X1 [1, 2, . . . , d + 1]\| d i=2 = X i=2 d+1 Y X τ ∈Sd+1 (−1) sgn(τ ) xτ (s)s s=1 ei · \|A13 \| · \|X1 [1, 2, . . . , d + 1]\|2 i=2 where Sd+1 is the set of permutations of {1, . . . , d + 1}. The general case with i0 , i1 , . . . , id can be obtained similarly. Proof of Theorem 4 Suppose Rank(X1 ) = d + 1. Then there exist i0 , . . . , id ∈ {1, . . . , m}, such that, \|X1 [i0 , i1 , . . . , id ]\| = 6 0. According to Lemma S.4, f (p) can be regarded as an order-(J − 1) polynomial of pi0 . Let pi0 = x ∈ (0, 1) and pi = (1−x)/(m−1) for i 6= i0 . Based on Lemma 2, f (p) can be written as d d+1 1−x 1−x J−2 J−1 + aJ−2 x fi0 (x) = aJ−1 x m−1 m−1 d+J−2 d+J−1 1−x 1−x + · · · + a1 x + a0 , where m−1 m−1 d Y X ei′s \|X1 [i0 , i′1 , . . . , i′d ]\|2 > 0 aJ−1 = \|Ai0 3 \| {i′1 ,...,i′d }⊂{1,...,m}\{i0 } s=1 Therefore, limx→1− (1 − x)−d x1−J fi0 (x) = (m − 1)−d aJ−1 > 0. That is, f (p) > 0 for pi0 = x close enough to 1 and pi = (1 − x)/(m − 1) for i 6= i0 . S8 SUPPLEMENTARY MATERIALS S9 In order to justify that the condition Rank(X1 ) = d+1 is also necessary, we only need to show that f (p) ≡ 0 if Rank(X1 ) ≤ d. Actually, for any (1) τ : {1, . . . , d + J − 1} → {1, . . . , m}, we construct Aτ as in the proof Qd (1) of Lemma S.5. Then \|Aτ \| = s=1 xτ (s)s · \|Aτ \|. Similar as in the proof of Lemma S.5, for t = 1, . . . , d, we add xτ (1)t (−eτ (1) , . . . , −eτ (d) , cτ (d+1)1 , (1) . . . , cτ (d+J−1),J−1 )T to the tth column of Aτ . We denote the resulting (3) (1) (3) matrix by Aτ . Note that \|Aτ \| = \|Aτ \|. We consider the sub-matrix (3) (3) Aτ d which consists of the first d columns of Aτ . For s ∈ τ −1 (τ (1)), (3) (3) the sth row of Aτ d is simply 0. For s = 2, . . . , k, the sth row of Aτ d is eτ (s) (xτ (s)1 − xτ (1)1 , . . . , xτ (s)d − xτ (1)d ). For s = 1, . . . , J − 1, the (d + s)th (3) row of Aτ d is −cτ (d+s)s (xτ (d+s)1 − xτ (1)1 , . . . , xτ (d+s)d − xτ (1)d ). We claim (3) (3) that Rank(Aτ d ) ≤ d − 1. Otherwise, if Rank(Aτ d ) = d, then there exist i1 , . . . , id ∈ {2, . . . , d + J − 1}, such that, the sub-matrix consisting of the (3) i1 th, . . . , id th rows of Aτ d is nonsingular. Then the sub-matrix consisting of the τ (1)th, τ (i1 )th, . . . , τ (id )th rows of X1 is nonsingular, which implies (3) Rank(X1 ) = d + 1. The contradiction implies Rank(Aτ d ) ≤ d − 1. Then (3) \|Aτ \| = 0 and thus \|Aτ \| = 0 for each τ . Based on Theorem 2, \|F\| ≡ 0 and thus f (p) ≡ 0. Proof of Theorem 5 Combining Theorem 1 and Theorem 4, it is straightforward that f (p) = 0 if Rank( X1 [{i \| pi > 0}]) ≤ d. We only need to show that f (p) > 0 if Rank(X1 [{i \| pi > 0}]) = d + 1. Due to Theorem 1, we only need to verify the case pi > 0, i = 1, . . . , m, because otherwise we may simply remove all support points with pi = 0. Suppose pi > 0, i = 1, . . . , m and Rank(X1 ) = d + 1. Then there exist i0 , . . . , id ∈ {1, . . . , m}, such that, \|X1 [i0 , . . . , id ]\| = 6 0. According to the proof of Theorem 4, for each i ∈ {i0 , . . . , id }, there exists an ǫi ∈ (0, 1), such that, f (p) > 0 as long as pi = x ∈ (1 − ǫi , 1) and pj = (1 − x)/(m − 1) for j 6= i. On the other hand, for each i ∈ / {i0 , . . . , id }, if we denote the jth row of X1 by αj , j = 1, . . . , m, then αi = a0 αi0 + · · · + ad αid for some real numbers a0 , . . . , ad . Since αi 6= 0, then at least one ai 6= 0. Without any loss of generality, we assume a0 6= 0. Then it can be verified that \|X1 [i, i1 , . . . , id ]\| = 6 0 too. Following the proof of Theorem 4 again, for such an i ∈ / {i0 , . . . , id }, there also exists an ǫi ∈ (0, 1), such that, f (p) > 0 as long as pi = x ∈ (1 − ǫi , 1) and pj = (1 − x)/(m − 1) for j 6= i. Let ǫ∗ = min{mini ǫi , (m − 1) mini pi , 1 − 1/m}/2. For i = 1, . . . , m, denote δi = (δi1 , . . . , δim )T ∈ S with δii = 1 − ǫ∗ and δij = ǫ∗ /(m − 1) for j 6= i. It can be verified that p = a1 δ1 + · · · + am δm with ai = (pi − ǫ∗ /(m − 1))/(1 − S9 S10 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL mǫ P ∗ /(m − 1)). By the choice of ǫ∗ , f (δi ) > 0, ai > 0, i = 1, . . . , m, and i ai = 1. Then f (p) > 0 according to Lemma S.6. Proof of Corollary 3 In order to check when a minimally supported design supported only on {x1 , x2 } is D-optimal, we add one more support point, that is, x3 . According to Theorem 2, Lemmas S.4, S.5, and 2, the objective function for a D-optimal approximate design on {x1 , x2 , x3 } is f (p1 , p2 , p3 ) = p1 p2 (c210 p1 + c120 p2 ) + p1 p3 (c201 p1 + c102 p3 ) + p2 p3 (c021 p2 + c012 p3 ) + c111 p1 p2 p3 , where c210 c120 c201 c102 c021 c012 c111 = = = = = = = 2 2 e2 g11 g12 (π11 π12 π13 )−1 (x1 − x2 )2 > 0 2 2 e1 g21 g22 (π21 π22 π23 )−1 (x1 − x2 )2 > 0 2 2 e3 g11 g12 (π11 π12 π13 )−1 (x1 − x3 )2 > 0 2 2 e1 g31 g32 (π31 π32 π33 )−1 (x1 − x3 )2 > 0 2 2 e3 g21 g22 (π21 π22 π23 )−1 (x2 − x3 )2 > 0 2 2 e2 g31 g32 (π31 π32 π33 )−1 (x2 − x3 )2 > 0 e1 (u22 u31 + u21 u32 − 2b22 b32 )(x1 − x2 )(x1 − x3 ) + e2 (u12 u31 + u11 u32 − 2b12 b32 )(x2 − x1 )(x2 − x3 ) + e3 (u12 u21 + u11 u22 − 2b12 b22 )(x3 − x1 )(x3 − x2 ) Based on Theorem 10, the design p = (p∗1 , p∗2 , 0)T is D-optimal if and only if ∂f (p)/∂f (p1 ) = ∂f (p)/∂f (p2 ) ≥ ∂f (p)/∂f (p3 ) Similar conclusions could be justified for x4 , . . . , xm if m ≥ 4. Proof of Theorem 12 According to the solutions provided by the software Mathematica, the largest root of equation (S.10) after simplification is b2 21/3 (3b1 − b22 ) A1/3 y1 = − − + (S.11) 3 3A1/3 3 × 21/3 where A = −27b0 + 9b1 b2 − 2b32 + 33/2 (27b20 + 4b31 − 18b0 b1 b2 − b21 b22 + 4b0 b32 )1/2 , and bi = ci /c3 , i = 0, 1, 2. Note that the calculation of A and thus y1 should be regarded as operations among complex numbers since the expression under square root could be negative. Nevertheless, y1 at the end would be a real number. Thus we are able to provide the analytic solution maximizing f (p1 , p2 , p3 ). Proof of Corollary 5 In order to check when a minimally supported design is D-optimal, we first add the four design points, that is, we consider S10 SUPPLEMENTARY MATERIALS S11 four design points (xi1 , xi2 ), i = 1, 2, 3, 4 and check when the D-optimal design could be constructed on the first three design points. Let X1 be defined as in Lemma 2. In this case, X1 is a 4 × 3 matrix. Following Theorem 2, Lemmas S.4, S.5, and 2, the objective function for a minimally supported design at (d, J, m) = (2, 3, 4) is f (p1 , p2 , p3 , p4 ) = + + + + c1111 p1 p2 p3 p4 \|X1 [1, 2, 3]\|2e1 e2 e3 · p1 p2 p3 (w1 p1 + w2 p2 + w3 p3 ) \|X1 [1, 2, 4]\|2e1 e2 e4 · p1 p2 p4 (w1 p1 + w2 p2 + w4 p4 ) \|X1 [1, 3, 4]\|2e1 e3 e4 · p1 p3 p4 (w1 p1 + w3 p3 + w4 p4 ) \|X1 [2, 3, 4]\|2e2 e3 e4 · p2 p3 p4 (w2 p2 + w3 p3 + w4 p4 ) 2 2 −1 where ei = ui1 + ui2 − 2bi2 , wi = e−1 i gi1 gi2 (πi1 πi2 πi3 ) , i = 1, 2, 3, 4, and X ei ej (uk1 ul2 + uk2ul1 − 2bk2 bl2 ) · \|X1[i, j, k]\| · \|X1[i, j, l]\| (S.12) c1111 = 1≤i<j≤4 with {i, j, k, l} = {1, 2, 3, 4} given 1 ≤ i < j ≤ 4. According to Theorem 10, a minimally supported design p = (p∗1 , p∗2 , p∗3 , 0)T in this case is D-optimal if and only if ∂f /∂p1 = ∂f /∂p2 = ∂f /∂p3 ≥ ∂f /∂p4 at p. Then ∂f /∂p1 = ∂f /∂p2 = ∂f /∂p3 at p is equivalent to (1) of Corollary 5, and ∂f /∂p4 ≤ ∂f /∂p1 at p leads to (2) of Corollary 5 since the forms of ∂f /∂pi at p, i = 1, 2, 3 will not change if more than four design points (i.e., m > 4) are added into consideration. Note that \|X1 [1, 2, 3]\|2e1 e2 e3 p∗2 p∗3 (2w1p∗1 + w2 p∗2 + w3 p∗3 ) in (2) of Corollary 5 is equal to ∂f /∂p1 at p. It could be replaced with \|X1 [1, 2, 3]\|2e1 e2 e3 p∗1 p∗3 (w1 p∗1 + 2w2 p∗2 +w3 p∗3 ) (i.e., ∂f /∂p2 ), or \|X1 [1, 2, 3]\|2e1 e2 e3 p∗1 p∗2 (w1 p∗1 +w2p∗2 +2w3p∗3 ) (i.e., ∂f /∂p3 ), since these three are all equal. S.4 Maximization of fi (z) in Section 3 According to Theorem 6, fi (z) is an order-(d + J − 1) polynomial of z. In other to determine its coefficients a0 , a1 , . . . , aJ−1 as in (3.2), we need to calculate fi (0), fi (1/2), fi (1/3), . . . , fi (1/J), which are J determinants defined in (3.1). Note that B−1 J−1 is a matrix determined by J − 1 only. For example, −1 B1 = 1 for J = 2,   3 −3 1 2 −1 4 − 32  , B2−1 = , B3−1 =  − 52 −1 1 1 1 −1 2 2 S11 S12 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL   4 −1 19  −7 11 2 6  7 −4 −1  2 1 1 1 −2 2 6 4 −6  − 13 3 B4−1 =  3  2 1 −6 for J = 3, 4, or 5 respectively. Once a0 , . . . , aJ−1 in (3.2) are determined, the maximization of fi (z) on z ∈ [0, 1] is numerically straightforward since it is a polynomial and its derivative fi′ (z) is given by (1−z)d J−1 X jaj z j−1 (1−z)J−1−j −(1−z)d−1 J−1 X (d+J −1−j)aj z j (1−z)J−1−j j=0 j=1 (S.13) S.5 Exchange algorithm for D-optimal exact allocation in Section 4 Exchange algorithm for D-optimal allocation (n1 , . . . , nm )T given n > 0: 1◦ Start with an initial design n = (n1 , . . . , nm )T such that f (n) > 0. 2◦ Set up a random order of (i, j) going through all pairs {(1, 2), (1, 3), . . . , (1, m), (2, 3), . . . , (m − 1, m)}. 3◦ For each (i, j), let c = ni + nj . If c = 0, let n∗ij = n. Otherwise, there are two cases. Case one: 0 < c ≤ J, we calculate fij (z) as defined in (4.1) for z = 0, 1, . . . , c directly and find z ∗ which maximizes fij (z). Case two: c > J, we first calculate fij (z) for z = 0, 1, . . . , J; secondly determine c0 , c1 , . . . , cJ in (4.2) according to Theorem 9; thirdly calculate fij (z) for z = J + 1, . . . , c based on (4.2); fourthly find z ∗ maximizing fij (z) for z = 0, . . . , c. For both cases, we define n∗ij = (n1 , . . . , ni−1 , z ∗ , ni+1 , . . . , nj−1, c − z ∗ , nj+1 , . . . , nm )T Note that f (n∗ij ) = fij (z ∗ ) ≥ f (n) > 0. If f (n∗ij ) > f (n), replace n with n∗ij , and f (n) with f (n∗ij ). 4◦ Repeat 2◦ ∼ 3◦ until convergence, that is, f (n∗ij ) = f (n) in step 3◦ for any (i, j). S12 SUPPLEMENTARY MATERIALS S.6 S13 More on Example 6: Polysilicon Deposition Study Table S.1 shows the list of experimental settings for the polysilicon deposition study. The factors are decomposition temperature(A), decomposition pressure(B), nitrogen flow (C), silane flow(D), setting time(E), cleaning method(F ). Column 1 provides original indices of experimental settings out of 729 distinct ones. For each experimental setting labelled “1” in a design, 9 responses are collected (Phadke (1989)) and assumed to be independent. Table S.1: Polysilicon Deposition Study: Experimental Settings for the Original, Rounded Approximate, and D-optimal Exact Designs Index 1 76 89 98 111 116 122 130 167 181 199 201 243 258 286 290 291 294 299 301 313 331 336 339 350 365 376 384 394 399 407 421 461 464 495 501 505 521 522 536 557 558 569 588 625 631 641 671 679 A 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 B 1 1 2 2 2 2 2 2 3 3 3 3 3 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 3 3 3 3 1 1 1 1 1 1 1 1 2 2 2 2 2 3 3 C 1 3 1 1 2 2 2 2 1 1 2 2 3 1 2 2 2 2 3 3 3 1 1 1 1 2 2 3 3 3 1 1 3 3 1 1 1 2 2 2 3 3 1 1 3 3 3 1 2 D 1 3 1 2 1 1 2 3 1 3 2 2 3 2 2 3 3 3 1 1 2 1 2 2 3 2 3 1 2 3 1 2 1 1 1 2 3 1 1 3 2 2 1 3 1 2 3 3 1 E 1 2 3 3 1 3 2 2 2 1 1 1 3 2 3 1 1 2 1 2 3 3 1 2 3 2 3 2 3 1 1 3 1 2 3 2 1 3 3 2 3 3 1 1 2 1 1 2 2 F 1 1 2 2 3 2 2 1 2 1 1 3 3 3 1 2 3 3 2 1 1 1 3 3 2 2 1 3 1 3 2 1 2 2 3 3 1 2 3 2 2 3 2 3 1 1 2 2 1 Original 1 1 1 0 0 0 1 0 0 0 0 1 1 1 0 1 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 1 1 0 0 0 0 0 1 0 1 0 0 1 0 1 0 1 1 Rounded 0 0 0 0 0 1 0 0 0 1 1 0 0 0 1 0 1 0 0 1 0 1 1 1 1 0 0 0 1 1 0 0 1 1 1 0 0 0 0 1 0 1 1 0 0 0 0 0 0 D-optimal 0 0 0 1 1 0 0 1 1 0 1 0 1 0 0 0 0 1 1 0 1 1 1 0 0 1 0 0 0 0 1 0 0 0 0 1 1 1 0 0 0 0 0 0 1 0 1 0 0 Table S.2 shows the model matrix for the D-optimal design no found for the polysilicon deposition study. In this table, each 3-level factor is represented by its linear component and quadratic component. Thus there are level combinations of 12 predictors. S13 S14 JIE YANG, LIPING TONG AND ABHYUDAY MANDAL Table S.2: Polysilicon Deposition Study: Model Matrix for the D-optimal Design Index 98 111 130 167 199 243 294 299 313 331 336 365 407 501 505 521 625 641 A1 −1 −1 −1 −1 −1 −1 0 0 0 0 0 0 0 1 1 1 1 1 A2 1 1 1 1 1 1 −2 −2 −2 −2 −2 −2 −2 1 1 1 1 1 B1 0 0 0 1 1 1 −1 −1 −1 0 0 0 1 −1 −1 −1 0 0 B2 −2 −2 −2 1 1 1 1 1 1 −2 −2 −2 1 1 1 1 −2 −2 C1 −1 0 0 −1 0 1 0 1 1 −1 −1 0 −1 −1 −1 0 1 1 C2 1 −2 −2 1 −2 1 −2 1 1 1 1 −2 1 1 1 −2 1 1 S14 D1 0 −1 1 −1 0 1 1 −1 0 −1 0 0 −1 0 1 −1 −1 1 D2 −2 1 1 1 −2 1 1 1 −2 1 −2 −2 1 −2 1 1 1 1 E1 1 −1 0 0 −1 1 0 −1 1 1 −1 0 −1 0 −1 1 0 −1 E2 1 1 −2 −2 1 1 −2 1 1 1 1 −2 1 −2 1 1 −2 1 F1 0 1 −1 0 −1 1 1 0 −1 −1 1 0 0 1 −1 0 −1 0 F2 −2 1 1 −2 1 1 1 −2 1 1 1 −2 −2 1 1 −2 1 −2	10math.ST
Contextual Bandits with Stochastic Experts Rajat Sen1 , Karthikeyan Shanmugam2 , and Sanjay Shakkottai1 1 arXiv:1802.08737v1 [stat.ML] 23 Feb 2018 2 The University of Texas at Austin IBM Research, Thomas J. Watson Center February 27, 2018 Abstract We consider the problem of contextual bandits with stochastic experts, which is a variation of the traditional stochastic contextual bandit with experts problem. In our problem setting, we assume access to a class of stochastic experts, where each expert is a conditional distribution over the arms given a context. We propose upper-confidence bound (UCB) algorithms for this problem, which employ two different importance sampling based estimators for the mean reward for each expert. Both these estimators leverage information leakage among the experts, thus using samples collected under all the experts to estimate the mean reward of any given expert. This leads to instance dependent regret bounds of µ)M log T /∆), where λ(µ µ) is a term that depends on the mean rewards of the experts, ∆ is the O (λ(µ smallest gap between the mean reward of the optimal expert and the rest, and M quantifies the inforµ) is typically O(log N ). mation leakage among the experts. We show that under some assumptions λ(µ We implement our algorithm with stochastic experts generated from cost-sensitive classification oracles and show superior empirical performance on real-world datasets, when compared to other state of the art contextual bandit algorithms. 1 Introduction Modern machine learning applications like recommendation engines [24, 11, 25], computational advertising [28, 10], A/B testing in medicine [29, 30] are inherently online. In these settings the task is to take sequential decisions that are not only profitable but also enable the system to learn better in future. For instance in a computational advertising system, the task is to sequentially place advertisements on users’ webpages with the dual objective of learning the preferences of the users and increasing the click-through rate on the fly. A key attribute of these systems is the well-known exploration (searching the space of possible decisions for better learning) and exploitation (taking decisions that are more profitable) trade-off. A principled method to capture this trade-off is the study of multi-armed bandit problems [12]. K-armed stochastic bandit problems have been studied for several decades. These are formulated as a sequential process, where at each time step any one of the K-arms can be selected. Upon selection of the k-th arm, the arm returns a stochastic reward with an expected reward of µk . Starting from the work of [21], a major focus has been on regret, which is the difference in the total reward that is accumulated from the genie optimal policy (one that always selects the arm with the maximum expected reward) from that of the chosen online policy. The current state-of-art algorithms achieve a regret of O((K/∆) log T ) [12, 7, 4, 5], which is order-wise optimal [21]. Here, ∆ corresponds to the gap in expected reward between the best arm and the next best one. Additional side information can be incorporated in this setting through the framework of contextual bandits. In the stochastic setting, it is assumed that at each time-step nature draws (x, r1 , ..., rK ) from a fixed but unknown distribution. Here, x ∈ X represents the context vector, while r1 , ..., rK are the rewards of the K-arms [22]. The context x is revealed to the policy-designer, after which she decides to choose an arm a ∈ {1, 2, ..., K}. Then, the reward ra is revealed to the policy-designer. In the computational advertising example, the context can be thought of as the browsing history, age, gender etc. of an user arriving in 1 the system, while r1 , ..., rK are generated according to the probability of the user clicking on each of the K advertisements. The task here is to learn a good mapping from the space of contexts X to the space of arms [K] = {1, 2, ..., K} such that when the decisions are taken according to that mapping, the mean reward observed is high. A popular model in the stochastic contextual bandits literature is the experts setting [3, 18, 22]. The task is to compete against the best expert in a class of experts Π = {π1 , ..., πN}, where each expert π ∈ Π is a function mapping X → [K]. The mean reward of an expert π is defined as E rπ(X) , where X is the random variable denoting the context and the expectation is taken over the unknown distribution over (x, r1 , ..., rK ). The best expert is naturally defined as the expert with the highest mean reward. The expected difference in rewards of a genie policy that always chooses the best expert and the online algorithm employed by the policy-designer is defined as the regret. This problem has been well-studied in the literature, where a popular approach is to reduce the contextual bandit problem to supervised learning techniques through argmin-oracles [8]. p KT polylog(N ) at time This leads to powerful algorithms with instance-independent regret bounds of O T [3, 18]. In practice the class of experts are generated online by training cost-sensitive classification oracles [3, 18]. Once trained, the resulting classifiers/oracles can provide reliable confidence scores given a new context, especially if they are well-calibrated [17]. These confidence scores effectively are a K-dimensional probability vector, where the k th entry is the probability of the classifier/oracle choosing the k th arm as the best, given a context. Motivated by this observation, we propose a variation of the traditional experts setting, which we term contextual bandits with stochastic experts. We assume access to a class of stochastic experts Π = {π1 , ..., πN }, which are not deterministic. Instead, each expert π ∈ Π, is a conditional probability distribution over the arms given a context. For an expert π ∈ Π the conditional distribution is denoted by πV \|X (v\|x) where V ∈ [K] is the random variable denoting the arm chosen and X is the context. An additional benefit is that this setting allows us to derive regret bounds in terms of closeness of these soft experts quantified by divergence measures, rather than in terms of the total number of arms K. As before, the task is to compete against the expert in the class with the highest mean reward. The expected reward of a stochastic expert π is defined as EX,V ∼π(V \|X) [rV ], i.e the mean reward observed when the arm is drawn from the conditional distribution π(V \|X). We propose upper-confidence (UCB) style algorithms for the contextual bandits with stochastic experts problem, that employ two importance sampling based estimators for the mean rewards under various experts. We prove instance-dependent regret guarantees for our algorithms. The main contributions of this paper are listed in the next section. 1.1 Main Contributions The contributions of this paper are three-fold: (i) (Importance Sampling based Estimators): The key components in our approach are two importance sampling based estimators for the mean rewards under all the experts. Both these estimators are based on the observation that samples collected under one expert can be reweighted by likelihood/importance ratios and averaged to provide an estimate for the mean reward under another expert. This sharing of information is termed as information leakage and has been utilized before under various settings [23, 27, 10]. The first estimator that we use is an adaptive variant of the well-known clipping technique, which was proposed in [27]. The estimator is presented in Eq. (3). However, we carefully adapt the clipping threshold in an online manner, in order to achieve regret guarantees. We also propose an importance sampling variant of the classical median of means estimator (see [26, 13]). This estimator is also designed to utilize the samples collected under all experts together to estimate the mean reward under any given expert. We define the estimator in Eq. (6). To the best of our knowledge, importance sampling has not been used in conjunction with the median of means technique in the literature before. We provide novel confidence guarantees for this estimator which depends on chi-square divergences between the conditional distributions under the various experts. This may be of independent interest. 2 (ii) (Instance Dependent Regret Bounds): We propose the contextual bandits with stochastic experts problem. We design two UCB based algorithms for this problem, based on the two importance sampling based estimators mentioned above. We show that utilizing the information leakage between the experts leads to regret guarantees that scale sub-linearly in N , the number of experts. The information leakage between any two experts in the first estimator is governed by a pairwise log-divergence measure (Def. 2). For the second estimator, chi-square divergences (Def. 3) characterize the leakage. µ We show that the regret of our UCB algorithm based on these two estimators scales as 1 : O λ(µ∆)M log T . Here, M is related to the largest pairwise divergence values under the two divergence measures used. ∆ µ) is a parameter that is the gap between the mean rewards of the optimal expert and the second best. λ(µ only depends on the gaps between mean rewards of the optimum experts and various sub-optimal ones. It is a normalized sum of difference in squares of the gaps of adjacent sub-optimal experts ordered by their gaps. Under the assumption that the suboptimal gaps (except that of the second best arm) are uniformly µ) is O(log N ) in expectation. We distributed in a bounded interval, we can show that the parameter λ(µ define this parameter explicitly in Section 6. For the clipped estimator we show that M = M 2 log2 (1/∆) where M is the largest pairwise log-divergence associated with the clipped estimator. For the median of means estimator, M = σ 2 where σ 2 is the largest pairwise chi-squared divergence. Naively treating each expert as an arm would lead to a regret scaling of O(N log T /∆). However, this ignores p information leakage. Existing instance-independent bounds for contextual bandits scale as KT poly log(N ) [3]. Our problem dependent bounds have a near optimal dependence on ∆ and does not depend on K, the numbers of arms. However, it depends on the divergence measure associated with the information leakage in the problem (M or σ parameters). Besides our analysis, we empirically show that this divergence based approach rivals or performs better than very efficient heuristics for contextual bandits (like bagging etc.) on real-world data sets. (iii) (Empirical Validation): We empirically validate our algorithm on three real world data-sets [20, 19, 1] against other state of the art contextual bandit algorithms [22, 3] implemented in Vowpal Wabbit [2]. In our implementation, we use online training of cost-sensitive classification oracles [8] to generate the class of stochastic experts. We show that our algorithms have better regret performance on these data-sets compared to the other algorithms. 2 Related Work Contextual bandits has been studied in the literature for several decades, starting with the simple setting of discrete contexts [12], to linear contextual bandits [16] and finally the general experts setting [18, 3, 22, 6, 9]. In this work, we focus on the experts setting. Contextual bandits with experts√was first studied in the adversarial setting, where there are algorithms with the optimal regret scaling O( KT log N ) [6]. In this paper, we are more interested in the stochastic version of the problem, where the context and the rewards of the arms are generated from an unknown but fixed distribution. The first strategies to be explored in this setting and epsilon-greedy [22] style strategies that achieve a √ were explore-then-commit regret scaling of O K log N T 2/3 in the instance-independent case. Following this there have been several p efforts to design adaptive algorithms that achieve a O( KT polylog(N )) instance-independent regret scaling. Notable among these are [18, 3]. These algorithms map the contextual bandit problem to supervised learning and assume access to cost-sensitive classification oracles. These algorithms have been heavily optimized in Vowpal Wabbit [2]. We study the contextual bandits with stochastic experts problem, where the experts are not deterministic functions mapping contexts to arms, but are conditional distributions over the arms given a context. 1 Tighter regret bounds are derived in Theorems 1 and 2. Here, we only mention the Corollaries of our approach, that are easy to state. 3 We show that we can achieve instance-dependent regret guarantees for this problem, that can scale as O ((M log N/∆) log T ) under some assumptions. Here, ∆ is the gap between the mean reward of the best expert and the second best and M is a divergence term between the experts. Our algorithms are based on importance sampling based estimators which leverage information leakage among stochastic experts. We use an adaptive clipped importance sampling estimator for the mean rewards of the experts, that was introduced in [27]. In [27], the estimator was studied in a best-arm/pure explore setting, while we study a cumulative regret problem where we need to adjust the parameters of the estimator in an online manner. In addition, we introduce an importance sampling based median of means style estimator in this paper, that can leverage the information leakage among experts. 3 Problem Setting and Definitions The general stochastic contextual bandit problem with K arms is defined as a sequential process for T discrete time-steps [22], where T is the time-horizon of interest. At each time t ∈ {1, 2, · · · , T } nature draws a vector (xt , r1 (t), ..., rK (t)) from an unknown but fixed probability distribution. Here, ri (t) ∈ [0, 1] is the reward of arm i. The context vector xt ∈ X is revealed to the policy-designer, whose task is then to choose an arm out the K possibilities. Only the reward rv(t) (t) of the chosen arm v(t), is then revealed to the policy-designer. We will use rv(t) in place of rv(t) (t) for notational convenience. ⇡(V \|X) X V Y Figure 1: Bayesian Network denoting the joint distribution of the random variables at a given time-step, under our contextual bandit setting. X denotes the context, V denotes the chosen arm, while Y denotes the reward from the chosen arm that also depends on the context observed. The distribution of the reward given the chosen arm and the context, and the marginal of the context remain fixed over all time slots. However, the conditional distribution of the chosen arm given the context is dependent on the stochastic expert at that time-step. Stochastic Experts: We consider a class of stochastic experts Π = {π1 , · · · , πN }, where each πi is a conditional probability distribution πV \|X (v\|x) where V ∈ [K] is the random variable denoting the arm chosen and X is the context. We will use the shorthand πi (V \|X) to denote the conditional distribution corresponding to expert i, for notational convenience. The observation model at each time step t is as follows: (i) A context xt is observed. (ii) The policy-designer chooses a stochastic expert πk(t) ∈ Π. An arm v(t) is drawn from the probability distribution πk(t) (V \|xt ), by the policy-designer. (iv) The stochastic reward yt = rv(t) is revealed. The joint distribution of the random variables X, V, Y denoting the context, arm chosen and reward observed respectively at time t, can be modeled by the Bayesian Network shown in Fig. 1. The joint distribution factorizes as follows, p(x, v, y) = p(y\|v, x)p(v\|x)p(x) (1), where p(y\|v, x) (the reward distribution given the arm and the context), and p(x) (marginal distribution of the context) is determined by the nature’s distribution and are fixed for all time-steps t = 1, 2, ..., T . On the other hand p(v\|x) (distribution of the arm 4 chosen given the context) depends on the expert selected at each round. At time t, p(v\|x) = πk(t) (v\|x) that is the conditional distribution encoded by the stochastic expert chosen at time t. Now we are at a position to define the objective of the problem. Regret: The objective in our contextual bandit problem is to perform as well as the best expert in the class of experts. We will define pk (x, v, y) , p(y\|v, x)πk (v\|x)p(x) as the distribution of the corresponding random variables when the expert chosen is πk ∈ Π. The expected reward of expert k is now denoted by, µk = Epk (x,v,y) [Y ], where Ep(.) denotes expectation with respect to distribution p(.). The best expert is given by k ∗ = arg maxk∈[N ] µk . The objective is to minimize the regret till time T , which is defined as PT R(T ) = t=1 µ∗ − µk(t) , where µ∗ = µk∗ . Note that this is analogous to the regret definition for the deterministic expert setting [22]. Let us define ∆k , µ∗ − µk as the optimality gap in terms of expected reward, for expert k. Let µ , {µ1 , ..., µN }. Now we will define some divergence metrics that will be important in describing our algorithms and theoretical guarantees. 3.1 Divergence Metrics In this section we will define some f -divergence metrics that will be important in analyzing our estimators. Similar divergence metrics were defined in [27] to analyze the clipped estimator in (3) in the context of a best arm identification problem. In addition to the divergence metric in [27], we will also define the chi-square divergence metric which will be useful in analyzing the median of means based estimator (6). First, we define conditional f -divergence. Definition 1. Let f (·) be a non-negative convex function such that f (1) = 0. For two joint distributions pX,Y (x, y) and qX,Y (x, y) (and the associated conditionals), the conditional f -divergence Df (pX\|Y kqX\|Y ) is given by: i h pX\|Y (X\|Y ) . Df (pX\|Y kqX\|Y ) = EqX,Y f qX\|Y (X\|Y ) Recall that πi is a conditional distribution of V given X. Thus, Df (πi kπj ) is the conditional f -divergence between the conditional distributions πi and πj . Note that in this definition the marginal distribution of X is the marginal of X given by nature’s inherent distribution over the contexts. In this work we will be concerned with two specific f -divergence metrics that are defined as follows. Definition 2. (Mij measure) [27] Consider the function f1 (x) = x exp(x − 1) − 1. We define the following log-divergence measure: Mij = 1 + log(1 + Df1 (πi kπj )), ∀i, j ∈ [N ]. The Mij -measures will be crucial in analyzing one of our estimators (clipped estimator) defined in Section 5. Definition 3. (σij measure) Df2 (πi kπj ) is known as the chi-square divergence between the respective con2 ditional distributions, where f2 (x) = x2 − 1. Let σij = 1 + Df2 (πi kπj ). The σij -measures are important in analyzing our second estimator (median of means) defined in Section 5. 4 A Meta-Algorithm In this section, we propose a general upper-confidence bound (UCB) style strategy that utilizes the structure of the problem to converge to the best expert much faster than a naive UCB strategy that treats each expert as an arm of the bandit problem. One of the key observations in this framework is that rewards collected under one expert can give us valuable information about the mean under another expert, owing to the Bayesian Network factorization of the joint distribution of X, V and Y . We propose two estimators for the mean rewards of different experts, that leverage this information leakage among experts, through importance sampling. These estimators are defined in Section 5. We propose a meta-algorithm (Algorithm 1) that is designed to use these estimators and the corresponding confidence intervals, to control regret. 5 Algorithm 1 D-UCB: Divergence based UCB for contextual bandits with stochastic experts 1: 2: 3: 4: 5: 6: 7: For time step t = 1, observe context x1 and choose a random expert π ∈ Π. Play an arm drawn from the conditional distribution π(V \|x1 ). for t = 2, ..., T do Observe context xt Let k(t) = arg maxk Uk (t − 1) , µ̂k (t − 1) + sk (t − 1). Select an arm v(t) from the distribution πk(t) (V \|xt ). Observe the reward Y (t). end for Here, µ̂k (t) denotes an estimate for the mean reward for expert k at time t, while sk (t) denotes the upper confidence bound for the corresponding estimator at time t. We propose two estimators that utilize all the samples observed under various experts to provide an estimate for the mean reward under expert k. The first estimator denoted by µ̂ck (t) (Section 5, Eq. (3)) is a clipped importance sampling estimator inspired by [27]. If this estimator is used, then sk (t) is set as in Equation. (4). The second estimator denoted by µ̂m k (t) (Section 5, Eq. (6)) is a median of means based importance sampling estimator. If this estimator is used, then sk (t) is set as in Equation. (7). 5 Estimators and Confidence Bounds In this section we define two estimators for estimating the mean rewards under a given expert. Both these estimators can effectively leverage the information leakage between samples collected under various experts, through importance sampling. One key observation that enables us in doing so is the following equation, πk (V \|X) . (2) µk = Epj (x,v,y) Y πj (V \|X) This has been termed as information leakage and has been leveraged before in the literature [27, 23, 10] in best-arm identification settings. Recall that the subscript pj (x, v, y) denotes that the expectation is taken under the joint distribution in (1), where p(v\|x) = πj (v\|x) i.e. under the distribution imposed by expert πj . However, even under this distribution we can technically estimate the mean reward under expert πk . The above equation is the motivation behind our estimators. Now, we will introduce our first estimator. Clipped Estimator: This estimator was introduced in [27] in the context of a pure exploration problem. Here, we analyze this estimator in a cumulative regret setting, where the parameters of the estimator need to be adjusted differently. Let ni (t) denote the number of times expert i has been invoked by Algorithm 1 till time t, for all i ∈ [N ]. We define the fraction νi (t) , ni (t)/t. We will also define Ti (t) as the subset of time-steps among {1, .., t}, in which the expert i was selected. Let µ̂ck (t) be the estimate of the mean reward of expert k from samples collected till time t. The estimator is given by, µ̂ck (t) = N 1 X X 1 πk (Vj (s)\|Xj (s)) πk (Vj (s)\|Xj (s)) Yj (s) ×1 ≤ 2 log(2/(t))Mkj . Zk (t) j=1 Mkj πj (Vj (s)\|Xj (s)) πj (Vj (s)\|Xj (s)) (3) s∈Tj (t) Here, Aj (s) is the value of the random variable A at time s drawn using expert j, where A can be the P r.v’s X,Y or V . We set Zk (t) = j nj (t)/Mkj . (t) is an adjustable term which controls the bias-variance trade-off for the estimator. 6 Intuition: The clipped estimator is a weighted average of the samples collected under different experts, where each sample is scaled by the importance ratio as suggested by (2). We also clip the importance ratios which are larger than a clipper level. This clipping introduces bias but decreases variance. The clipper level is carefully chosen to trade-off bias and variance. The clipper level values and the weights are dependent on the divergence terms Mkj ’s. When the divergence Mkj is large, it means that the samples from expert j is not valuable for estimating the mean for expert k. Therefore, a weight of 1/Mkj is applied. Similarly, the clipper level is set at 2 log(2/(t))Mkj to restrict the aditive bias to (t). The upper confidence term in Algorithm 1 for the estimator µ̂ck (t) is chosen as, sck (t) = at time t, where β(t) is such that, β(t) log(2/β(t)) √ = 3 β(t) 2 c1 t log t Zk (t) . (4) We set c1 = 16 in our analysis. The upper confidence bound is derived using Lemma 1. Lemma 1. Consider the estimator in Eq. (3). We have the following confidence bound at time t, 2 ! Zk (t) δ2 t c . P (µk − δ − (t)/2 ≤ µ̂k (t) ≤ µk + δ) ≥ 1 − 2 exp − 8(log(2/(t)))2 t The lemma is implied by Theorem 4 in [27]. We include the proof in Section A in the appendix. The lemma shows that the clipped estimator can pool samples from all experts, in order to estimate the mean under expert k. The variance of the estimator depends on Zk (t), which depends on the log-divergences and number of times each expert has been invoked. Median of Means Estimator: Now we will introduce our second estimator which is based on the wellknown median of means technique of estimation. Median of means estimators are popular for statistical estimation when the underlying distributions are heavy-tailed [13]. The estimator for the mean under the k th expert at time t is obtained through the following steps: (i) We divide the total samples into l(t) = bc2 log(1/δ(t))c groups, such that the fraction of samples from each expert is preserved. We choose c2 = 8 for our analysis. Let us index the groups as r = 1, 2..., l(t). This means that there are at least bni (t)/l(t)c samples from expert i in each group. (ii) We calculate the empirical mean of expert k from the samples in each group through importance sampling. (iii) The median of these means is our estimator. (r) Now we will setup some notation. P Let Tj ⊆ {1, 2, ..., t} be the indices of the samples from expert j that lie in group r. Let W (r, t) = k i ni (r, t)/σki , where ni (r, t) is the number of samples from expert i in group P r. Let n(r, t) = i ni (r, t). Then the mean of expert k estimated from group r is given by, (r) µ̂k (t) = N X X 1 1 πk (Vj (s)\|X(s)) Yj (s) . Wk (r, t) j=1 σ π kj j (Vj (s)\|Xj (s)) (r) (5) s∈Tj The median of means estimator for expert k is then given by, (1) (l(t)) µ̂m (t) . k (t) , median µ̂k (t), · · · , µ̂k (6) Intuition: The mean of every group is a weighted average of samples from each expert, rescaled by the importance ratios. This is similar to the clipped estimator in Eq. (3). However, here the importance ratios are not clipped at a particular level. In this estimator, the bias-variance trade-off is controlled by taking the median of means from l(t) groups. The number of groups l(t) needs to be carefully set in-order to control the bias-variance trade-off. The upper confidence bound used in conjunction with this estimator at time t is given by, 7 sm k (t) 1 = Wk (t) r c3 log(1/δ(t)) t (7) where Wk (t) = minr∈[l(t)] Wk (r, t)/n(r, t) and δ(t) is set as 1/t2 in our algorithm. This choice is inspired by the following lemma. Lemma 2. Let δ(t) ∈ (0, 1). Then the estimator in (6) has the following confidence bound, ! r c3 log(1/δ(t)) 1 m ≥ 1 − δ(t). (8) P \|µ̂k (t) − µk \| ≤ Wk (t) t We provide the proof of this lemma in Section B in the appendix. The constant c3 is 64. 6 Theoretical Results In this section, we provide instance dependent regret guarantees for Algorithm 1 for the two estimators proposed - a) The clipped estimator (3) and b) The median of means estimator (6). Let ∆ = mink6=k∗ ∆k be the gap in the expected reward between the optimum expert and the second best. We define a parameter µ), later in the section, that depends only on the gaps of the expected rewards of various experts from the λ(µ optimal one. µ)M 2 log2 (6/∆) log T /∆). SimiFor the Algorithm 1 that uses the clipped estimator, regret scales as O(λ(µ µ)σ 2 log T /∆). Here M is the maxlarly, for the case of the median of means estimator, regret scales as O(λ(µ 2 imum log-divergence and σ is the maximum chi-square divergence between two experts, respectively. When the gaps between the optimum expert and sub-optimal ones are distributed uniformly at random in µ) parameter is at most O(log N ) in expectation. In contrast, if the [δ2 , 1] (δ2 > 0), we show that the λ(µ experts were used as separate arms, a naive application of UCB-1 [5] bounds would yield a regret scaling of log T . This can be prohibitively large when the number of experts are large. O N ∆ For ease of exposition of our results, let us re-index the experts using indices {(1), (2), ..., (N )} such that 0 = ∆(1) ≤ ∆(2) ≤ ... ≤ ∆(N ) . The regret guarantees for our clipped estimator are provided under the following assumption. Assumption 1. Assume the log-divergence terms (2) are bounded for all i, j ∈ [N ]. Let M = maxi,j Mij . Now we are at a position to present one of our main theorems that provides regret guarantees for Algorithm 1 using the estimator (3). Theorem 1. Suppose Assumption 1 holds. Then the regret of Algorithm 1 at time T using estimator (3), is bounded as follows: ! N −1 N X C1 M 2 log2 (6/∆(k) ) log T C1 M 2 log2 (6/∆(N ) ) log T γ(∆(k) ) π2 X R(T ) ≤ + ∆(i) + 1− ∆(N ) 3 i=2 ∆(k) γ(∆(k+1) ) k=2 Here, C1 is an universal constant and γ(x) = 2 x . log2 (6/x) We defer the proof of Theorem 1 to Appendix A. We now present Theorem 2 that provides regret guarantees for Algorithm 1 using the estimator (6). The theorem holds under the following assumption. Assumption 2. Assume the chi-square terms (3) are bounded for all i, j ∈ [N ]. Let σ = maxi,j σij . Theorem 2. Suppose Assumption 2 holds. Then the regret of Algorithm 1 at time T using estimator (6), is bounded as follows: ! ! N −1 N X ∆2(k) C2 σ 2 log T C2 σ 2 log T π2 X R(T ) ≤ + 1− 2 + ∆(i) ∆(N ) ∆(k) ∆(k+1) 3 i=2 k=2 Here, C2 is an universal constant. 8 0.6 1.0 1.0 0.8 0.9 Average Loss 0.8 Average Loss Average Loss 1.0 0.6 0.4 0.7 0.2 0.4 0 500 1000 1500 Time-Steps: T (a) 0.8 0 2500 5000 7500 Time-Steps: T 10000 0.6 0 (b) 5000 10000 15000 Time-Steps: T (c) 20000 (d) Figure 2: In all these plots, the progressive validation loss [3] till time T has been plotted as a function of time T . (a) Performance of the algorithms on the Drug Consumption dataset [19]. (b) Performance of the algorithms on the Stream Analytics dataset [1]. (c) Performance of the algorithms on the Letters dataset [20]. (d) Legend. . The proof of Theorem 2 has been deferred to Appendix B. Now, we will delve deeper into the instance dependent terms in Theorems 1 and 2. of Theorem 1 and 2 imply the following corollary. The proofs PN −1 ∆2(k) µ) , 1 + k=2 1 − ∆2 Corollary 1. Let λ(µ . We have the following regret bounds: (k+1) (i) For Algorithm 1 with estimator (3), R(T ) ≤ O M 2 log2 (6/∆(2) ) log T ∆(2) (ii) Similarly for Algorithm 1 with estimator (6), R(T ) ≤ O σ 2 log T ∆(2) µ), ∆1 min λ(µ . (2) µ), ∆1 min λ(µ . (2) Corollary 1 leads us to our next result. In Corollary 2 we show that when the ∆ gaps are uniformly µ) is O(log N ), in expectation. distributed, then the λ(µ Corollary 2. Consider a generative model where ∆(3) ≤ ... ≤ ∆(N ) are the order statistics of N − 2 random variables drawn i.i.d uniform over the interval [∆(2) , 1]. Let p∆ denote the measure over these ∆’s. Then we have the following: 2 M log N log2 (1/∆(2) ) log T (i) For Algorithm 1 with estimator (3), Ep∆ [R(T )] = O . ∆(2) 2 log T (ii) For Algorithm 1 with estimator (6), Ep∆ [R(T )] = O σ log∆N . (2) Remark 1. Note that our guarantees do not have any term containing K - the number of arms. This dependence is implicitly captured in the divergence terms among the experts. In fact when the number of arms K is very large, we expect our divergence based algorithms to perform comparatively better than other algorithms, whose guarantees explicitly depend on K. This phenomenon is observed in practice in µ) our empirical validation on real world data-sets in Section 7. We also show empirically, that the term λ(µ grows very slowly with the number of experts on real-world data-sets. This empirical result is included in Appendix D. 7 Empirical Results In this section, we will empirically test our algorithms on three real-world multi-class classification datasets, against other state of the art algorithms for contextual bandits with experts. Any multi-class classification dataset can be converted into a contextual bandit scenario, where the features are the contexts. At each time-step, the feature (context) of a sample point is revealed, following which the contextual bandit algorithm chooses one of the K classes, and the reward observed is 1 if its the correct class otherwise it is 0. This is bandit feedback as the correct class is never revealed, if not chosen. This method has been widely used to benchmark contextual bandit algorithms [9, 3], and is in fact implemented in Vowpal Wabbit [2]. 9 Our algorithm is run in batches. At the starting of each batch, we add experts trained on prior data through cost-sensitive classification oracles [8] and also update the divergence terms between experts, which are estimated from data observed so far. During each batch, Algorithm 1 is deployed with the current set of experts. The pseudo-code for this procedure is provided in Algorithm 2. We use XgBoost [15] and Algorithm 2 Batched D-UCB with cost-sensitive classification experts 1: 2: 3: 4: 5: 6: 7: 8: Let Π = {π1 }, which is an expert that chooses arms randomly. For time steps t = 1 to 3K, choose an arm sampled from expert π1 . t = 3K + 1. Add experts to Π trained on observed data and update divergences. while t <= T do √ for s = t to t + O( t) do Deploy Algorithm 1 with experts in Π. end for √ Let t = t + O( t). Add experts to Π trained on observed data and update divergences. end while Logistic Regression in scikit-learn [14] with calibration, as the base classifiers for our cost-sensitive oracles. Bootstrapping is used to generate different experts. At the starting of each batch 4 new experts are added. The constants are set as c1 = 1, c2 = 4 and c3 = 2 in practice. All the settings are held fixed over all three data-sets, without any parameter tuning. We provide more details in Appendix D. In the appendix we also show that the gap dependent term in our theoretical bounds grows much slower compared to UCB-1 bounds (Fig. 3), as the number of experts increase in the stream analytics dataset [1]. An implementation of our algorithm can be found here 2 . We compare against Vopal Wabbit implementations of the following algorithms: (i) -greedy [22] - parameter set at ’–epsilon 0.06’. (ii) First (Greedily selects best expert) - parameter set at ’–first 100’. (iii) Online Cover [3] - parameter set at ’–cover 5’ (iv) Bagging (Simulates Thompson Sampling through bagged classifiers) - parameter set at ’–bag 7’. Drug Consumption Data: This dataset [19] is a part of UCI repository. It has data from 1885 respondents with 32 dimensional continuous features (contexts) and their history of drug use. There are 19 drugs under study (19 arms). For each entry, if the bandit algorithm selects the drug most recently used, the reward observed is 1, o.w. 0 reward is observed. The performance of the algorithms are shown in Fig. 2a. We see that D-UCB (Algorithm 2) with median of moments clearly performs the best in terms of average loss, followed by D-UCB with the clipped estimator. D-UCB with median of moments converges to an average loss of 0.4 within 1885 samples. Stream Analytics Data: This dataset [1] has been collected using the stream analytics client. It has 10000 samples with 100 dimensional mixed features (contexts). There are 10 classes (10 arms). For each entry, if the bandit algorithm selects the correct class, the reward observed is 1, o.w. 0 reward is observed. The performance of the algorithms are shown in Fig. 2b. In this data-set bagging performs the best closely followed by the two versions of D-UCB (Algorithm 2). Bagging is a strong competitor empirically, however this algorithm lacks theoretical guarantees. Bagging converges to an average loss of 8%, while D-UCB with median of moments converges to an average loss of 10%. Letters Data: This dataset [20] is a part of the UCI repository. It has 20000 samples of hand-written English letters, each with 17 hand-crafted visual features (contexts). There are 26 classes (26 arms) corresponding to 26 letters. For each entry, if the bandit algorithm selects the correct letter, the reward observed is 1, o.w. 0 reward is observed. The performance of the algorithms are shown in Fig. 2c. Both versions of D-UCB significantly outperform the others. The median of moments based version converges to an average loss of 0.62, while the clipped version converges to an average loss of 0.68. 2 https://github.com/rajatsen91/CB StochasticExperts 10 Conclusion We study the problem of contextual bandits with stochastic experts. We propose two UCB style algorithms, that use two different importance sampling estimators, which can leverage information leakage between the stochastic experts. We provide instance-dependent regret guarantees for our UCB based algorithms. Our algorithms show strong empirical performance on real-world datasets. We believe that this paper introduces an interesting problem setting for studying contextual bandits, and opens up opportunities for future research that may include better regret bounds for the problem and an instance-dependent lower-bound. Acknowledgment This work is partially supported by NSF SaTC 1704778, ARO W911NF-17-1-0359, and the US DoT supported D-STOP Tier 1 University Transportation Center. References [1] Stream analytics dataset. http://bit.ly/2HD8bNH. Accessed: 2018-2-10. [2] Vowpal wabbit. https://github.com/JohnLangford/vowpal_wabbit. Accessed: 2017-10-10. [3] Alekh Agarwal, Daniel Hsu, Satyen Kale, John Langford, Lihong Li, and Robert Schapire. Taming the monster: A fast and simple algorithm for contextual bandits. In International Conference on Machine Learning, pages 1638–1646, 2014. [4] Shipra Agrawal and Navin Goyal. Analysis of thompson sampling for the multi-armed bandit problem. In Conference on Learning Theory, pages 39–1, 2012. [5] Peter Auer. Using confidence bounds for exploitation-exploration trade-offs. Journal of Machine Learning Research, 3(Nov):397–422, 2002. [6] Peter Auer, Nicolo Cesa-Bianchi, Yoav Freund, and Robert E Schapire. The nonstochastic multiarmed bandit problem. SIAM journal on computing, 32(1):48–77, 2002. [7] Peter Auer and Ronald Ortner. Ucb revisited: Improved regret bounds for the stochastic multi-armed bandit problem. Periodica Mathematica Hungarica, 61(1-2):55–65, 2010. [8] Alina Beygelzimer and John Langford. The offset tree for learning with partial labels. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 129–138. ACM, 2009. [9] Alina Beygelzimer, John Langford, Lihong Li, Lev Reyzin, and Robert Schapire. Contextual bandit algorithms with supervised learning guarantees. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 19–26, 2011. [10] Léon Bottou, Jonas Peters, Joaquin Quiñonero-Candela, Denis X Charles, D Max Chickering, Elon Portugaly, Dipankar Ray, Patrice Simard, and Ed Snelson. Counterfactual reasoning and learning systems: The example of computational advertising. The Journal of Machine Learning Research, 14(1):3207– 3260, 2013. [11] Djallel Bouneffouf, Amel Bouzeghoub, and Alda Lopes Gançarski. A contextual-bandit algorithm for mobile context-aware recommender system. In International Conference on Neural Information Processing, pages 324–331. Springer, 2012. [12] Sébastien Bubeck, Nicolo Cesa-Bianchi, et al. Regret analysis of stochastic and nonstochastic multiarmed bandit problems. Foundations and Trends R in Machine Learning, 5(1):1–122, 2012. [13] Sébastien Bubeck, Nicolo Cesa-Bianchi, and Gábor Lugosi. Bandits with heavy tail. IEEE Transactions on Information Theory, 59(11):7711–7717, 2013. 11 [14] Lars Buitinck, Gilles Louppe, Mathieu Blondel, Fabian Pedregosa, Andreas Mueller, Olivier Grisel, Vlad Niculae, Peter Prettenhofer, Alexandre Gramfort, Jaques Grobler, et al. Api design for machine learning software: experiences from the scikit-learn project. arXiv preprint arXiv:1309.0238, 2013. [15] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794. ACM, 2016. [16] Wei Chu, Lihong Li, Lev Reyzin, and Robert E Schapire. Contextual bandits with linear payoff functions. In International Conference on Artificial Intelligence and Statistics, pages 208–214, 2011. [17] Ira Cohen and Moises Goldszmidt. Properties and benefits of calibrated classifiers. In PKDD, volume 3202, pages 125–136. Springer, 2004. [18] Miroslav Dudik, Daniel Hsu, Satyen Kale, Nikos Karampatziakis, John Langford, Lev Reyzin, and Tong Zhang. Efficient optimal learning for contextual bandits. arXiv preprint arXiv:1106.2369, 2011. [19] Elaine Fehrman, Awaz K Muhammad, Evgeny M Mirkes, Vincent Egan, and Alexander N Gorban. The five factor model of personality and evaluation of drug consumption risk. In Data Science, pages 231–242. Springer, 2017. [20] Peter W Frey and David J Slate. Letter recognition using holland-style adaptive classifiers. Machine learning, 6(2):161–182, 1991. [21] Tze Leung Lai and Herbert Robbins. Asymptotically efficient adaptive allocation rules. Advances in applied mathematics, 6(1):4–22, 1985. [22] John Langford and Tong Zhang. The epoch-greedy algorithm for multi-armed bandits with side information. In Advances in neural information processing systems, pages 817–824, 2008. [23] Finnian Lattimore, Tor Lattimore, and Mark D Reid. Causal bandits: Learning good interventions via causal inference. In Advances in Neural Information Processing Systems, pages 1181–1189, 2016. [24] Lei Li, Dingding Wang, Tao Li, Daniel Knox, and Balaji Padmanabhan. Scene: a scalable two-stage personalized news recommendation system. In Proceedings of the 34th international ACM SIGIR conference on Research and development in Information Retrieval, pages 125–134. ACM, 2011. [25] Lihong Li, Wei Chu, John Langford, and Robert E Schapire. A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th international conference on World wide web, pages 661–670. ACM, 2010. [26] Gábor Lugosi and Shahar Mendelson. Sub-gaussian estimators of the mean of a random vector. arXiv preprint arXiv:1702.00482, 2017. [27] Rajat Sen, Karthikeyan Shanmugam, Alexandros G. Dimakis, and Sanjay Shakkottai. Identifying best interventions through online importance sampling. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3057–3066, International Convention Centre, Sydney, Australia, 2017. PMLR. [28] Liang Tang, Romer Rosales, Ajit Singh, and Deepak Agarwal. Automatic ad format selection via contextual bandits. In Proceedings of the 22nd ACM international conference on Conference on information & knowledge management, pages 1587–1594. ACM, 2013. [29] Cem Tekin, Onur Atan, and Mihaela Van Der Schaar. Discover the expert: Context-adaptive expert selection for medical diagnosis. IEEE Transactions on Emerging Topics in Computing, 3(2):220–234, 2015. [30] Cem Tekin, Jinsung Yoon, and Mihaela van der Schaar. Adaptive ensemble learning with confidence bounds for personalized diagnosis. In AAAI Workshop: Expanding the Boundaries of Health Informatics Using AI, 2016. 12 A Clipped Estimator As mentioned in Section 5, the motivating equation guiding the design of our estimators is Eq. (2). This equation tells us that even when the statistics of the samples observed are governed by the distribution of (X, V, Y ) under expert j, we can infer the mean of expert k. Such observations were made in [23, 27] in the context of best arm identification problems. Suppose we observe t samples under expert πj . Guided by Eq. (2), one might come up with the following naive importance sampled estimator for the mean under expert k (µk ): t µ̂k = πk (Vj (s)\|Xj (s)) 1X Yj (s) . t s=1 πj (Vj (s)\|Xj (s)) However, it is not possible to derive good confidence interval for the above estimator because even though the reward variable Y is bounded, the reweighing term πk (Vj (s)\|Xj (s))/πj (Vj (s)\|Xj (s)) can be unbounded and in some case heavy-tailed. The key idea is to come up with robust estimators that have good variance properties. One approach of controlling the variance of such estimators is to clip that the samples that are too large. This leads to the following clipped estimator [27]: t µ̂k = πk (Vj (s)\|Xj (s)) 1X Yj (s) ×1 t s=1 πj (Vj (s)\|Xj (s)) πk (Vj (s)\|Xj (s)) ≤ ηkj . πj (Vj (s)\|Xj (s)) (9) The clipping makes the estimator biased, however it helps in controlling the variance. The clipper level ηkj which depends on the relationship between πk and πj needs to be set carefully to control the bias-variance trade-off. In [27], it has been shown that if the log-divergence measure Mkj (defined in (2)) is bounded, then a good choice is ηkj = 2 log(2/)Mkj , if we want an additive bias of at most (t)/2 (Theorem 3 in [27]). This idea can be generalized to estimating the mean of expert k, while observing samples from all the other experts. This leads to the clipped estimator in Eq. (3). Lemma 1 provides confidence guarantees for this estimator. The proof of this lemma follows from Theorem 4 in [27], but we include it here for completeness. In what follows, we will abbreviate Epj (.) [.] as Ej [.]. In this section let µ̂k (t) = µ̂ck (t). Proof of Lemma 1. Note that from Lemma 3 in [27] it follows that: Ej [µ̂k (t)] ≤ µk ≤ Ej [µ̂k (t)] + (t) . 2 (10) For the sake of analysis, let us consider the rescaled version µ̄k (t) = (Zk (t)/t)µ̂k (t) which can be written as: N πk (Vj (s)\|Xj (s)) πk (Vj (s)\|Xj (s)) 1XX 1 Yj (s) ×1 ≤ 2 log(2/(t))Mkj . µ̄k (t) = t j=0 Mkj πj (Vj (s)\|Xj (s)) πj (Vj (s)\|Xj (s)) (11) s∈Tj Since Yj (s) ≤ 1, we have every random variable in the sum in (11) bounded by 2 log(2/(t)) Let, µ̄k = E[µ̄k (t)]. Therefore by Chernoff bound, we have the following chain: δ2 t P (\|µ̄k (t) − µ̄k \| ≤ δ) ≤ 2 exp − 8(log(2/(t)))2 t t t δ2 t =⇒ P \|µ̄k (t) − µ̄k \|≤δ ≤ 2 exp − Zk (t) Zk (t) Zk (t) 8(log(2/(t)))2 2 t δ t =⇒ P \|µ̂k (t) − µ̂k \| ≤ δ ≤ 2 exp − Zk (t) 8(log(2/(t)))2 2 ! δ2 t Zk (t) =⇒ P (\|µ̂k (t) − µ̂k \| ≤ δ) ≤ 2 exp − 8(log(2/(t)))2 t 13 (12) Now we can combine Equations (12) and (10) to obtain: δ2 t P (µk − δ − (t)/2 ≤ µ̂k (t) ≤ µk + δ) ≥ 1 − 2 exp − 8(log(2/(t)))2 Zk (t) t 2 ! Now, we will prove Theorem 1. Note that we will re-index the experts such that 0 = ∆(1) ≤ ∆(2) ≤ ... ≤ ∆(N ) . Note that throughout this proof Uk (t), µ̂k (t) and sk (t) in Algorithm 1 are defined as in Equations (3) and (4) respectively. Before we proceed let us prove some key lemmas. First we prove that with high enough probability the upper confidence bound estimate for the optimal expert k ∗ is greater than the true mean µ∗ . Lemma 3. We have the following confidence bound at time t, 1 . t2 P (Uk∗ (t) ≤ µ∗ ) ≥ 1 − Proof. We have the following chain, 1 3β(t) P (Uk∗ (t) ≥ µ∗ ) = P (µ̂k∗ (t) ≥ µ∗ − sk∗ (t)) = P µ̂k∗ (t) ≥ µ∗ − ≥ 1 − 2. 2 t The last inequality is obtained by setting δ = β(t) and (t) = β(t) in Lemma 1. Next we prove that for a large enough time t, the UCB estimate of the k th expert is less than that of µ∗ . 2 2 k ) log T Lemma 4. We have the following confidence bound at time t > 144M log ∆(6/∆ , 2 k P (Uk (t) < µ∗ ) ≥ 1 − 1 . t2 Proof. We have the following chain, 3β(t) P (Uk (t) > µ∗ ) = P µ̂k (t) > µ∗ − 2 (i) ∆k ≤ P µ̂k (t) > µ∗ − 2 (ii) ∆k ≤ P µ̂k (t) > µk + 2 (iii) ≤ 1 . t2 2 2 k ) log T Here, (i) follows from the fact that Zkt(t) ≥ 1/M , t > 144M log ∆(6/∆ and the definition of β(t). (ii) is 2 k by definition of ∆k . Finally the concentration bound in (iii) follows from Lemma 1. Note that Lemmas 3 and 4 together imply, P (k(t) = k) ≤ for k 6= k ∗ and t > 144M 2 log2 (6/∆k ) log T ∆2k . 14 2 t2 (13) Proof of Theorem 1. Let Tk = 144M 2 log2 (6/∆(k) ) log T ∆2(k) for k = 2, .., N . The regret of the algorithm can be bounded as, R(T ) ≤ ∆(N ) TN + ≤ ∆(N ) TN + ≤ = N −3 X k=0 k=0 ∆(N −k−1) P (1{k(t) ∈ {(1), ..., (N − k − 1)}) + k=0 t=TN −k  TN −k−1 ∆(N −k−1) (TN −k−1 − TN −k ) + 144M 2 log2 (6/∆(N ) ) log T π2 + ∆(N ) 3 X N X ! ∆(i) P (1{k(t) = (i)}) i=N −k  N X 2∆(i)  t2 t=TN −k i=N −k ! N γ(∆(N −k−1) ) π2 X 1− + ∆(N ) TN + ∆(i) γ(∆(N −k) ) 3 i=2 ! N −1 N X X 144M 2 log2 (6/∆(k) ) log T γ(∆(k) ) ∆(i) + 1− . ∆(k) γ(∆(k+1) ) i=2 144M 2 log2 (6/∆(N −k−1) ) log T ∆(N −k−1) Here, γ(x) = B N −3 X N −3 TNX −k−1 X (14) k=2 2 x . log2 (6/x) Median of Means Estimator The median of means estimator is popular for estimating statistics under heavy-tailed distribution [13, 26]. We shall see that the median of means based estimator in Eq. (6) has good variance properties, when the chi-square divergence (Assumption 2) are bounded. Before proving Lemma 2, we will be establishing some intermediate results. Lemma 5. Consider the quantity µ̂rk (t) in Eq. (5). The variance of this quantity is upper bounded as follows: Var[µ̂rk (t)] ≤ 1 1 σ2 . ≤ 2 m Wk (t) m where m = bt/l(t)c. Proof. We have the following chain, Var[µ̂rk (t)] N X X 1 1 πk (Vj (s)\|X(s)) = 2 Var Yj (s) × π (V (s)\|X (s)) Wk (r, t)2 j=1 σkj j j j (r) s∈Tj ≤ 1 Wk (r, t)2 N X X j=1 s∈T (r) 1 2 Var σkj πk (Vj (s)\|X(s)) πj (Vj (s)\|Xj (s)) j m = Wk (r, t)2 1 = m PN 1 nj (r,t) j=1 n(r,t).σkj ≤ 2 1 1 . m Wk (t)2 (15) Now, we can apply Chebyshev to conclude that for all r ∈ [l(t)], 15 P \|µ̂rk (t) 1 − µk \| ≤ Wk (t) r 4 m ! ≥ 3 . 4 (16) Now we will prove Lemma 2. Proof of Lemma 2. In light of Eq. (16), the probability that the median is not within distance µk is bounded as, r ! 4 1 m P \|µ̂k (t) − µk \| > Wk (t) m 1 Wk (t) q 4 m of ≤ P (Bin(l(t), 1/4) > l(t)/2) ≤ δ(t). This concludes the proof. Note that we will re-index the experts such that 0 = ∆(1) ≤ ∆(2) ≤ ... ≤ ∆(N ) . Note that throughout this proof Uk (t), µ̂k (t) and sk (t) in Algorithm 1 are defined as in Equations (6) and (7) respectively. Before we proceed let us prove some key lemmas. Now we prove lemmas analogous to Lemmas 4 and 3. Lemma 6. We have the following confidence bound at time t, P (Uk∗ (t) ≤ µ∗ ) ≥ 1 − 1 . t2 The proof follows directly from Lemma 2. Lemma 7. We have the following confidence bound at time t > P (Uk (t) < µ∗ ) ≥ 1 − 256σ 2 log T ∆2k , 1 . t2 Proof. We have the following chain, 1 P (Uk (t) > µ ) = P µ̂k (t) > µ − Wk (t) (i) ∆k ∗ ≤ P µ̂k (t) > µ − 2 (ii) ∆k ≤ P µ̂k (t) > µk + 2 ∗ ∗ (iii) ≤ r 64 log t t ! 1 . t2 Here, (i) follows from the fact that Wk1(t) ≤ σ and t > concentration bound in (iii) follows from Lemma 2. 256σ 2 log T ∆2k . (ii) is by definition of ∆k . Finally the Note that Lemma 7 and 6 together imply that, P (k(t) = k) ≤ for k 6= k ∗ and t > 256σ 2 log T ∆2k . 16 2 t2 (17) Proof of Theorem 2. Let Tk = R(T ) ≤ ∆(N ) TN + N −3 TNX −k−1 X k=0 t=TN −k 256σ 2 log T ∆2(k) for k = 2, .., N . The regret of the algorithm can be bounded as, P (1{k(t) ∈ {(1), ..., (N − k − 1)}) × ∆(N −k−1) + ! N X ∆(i) P (1{k(t) = (i)}) i=N −k  N X 2∆(i)  ∆(N −k−1) (TN −k−1 − TN −k ) + ≤ ∆(N ) TN + t2 t=TN −k i=N −k k=0 ! ! N N −3 X ∆2(N −k−1) π2 X 512σ 2 log T + ∆(N ) TN + ≤ 1− ∆(i) ∆(N −k−1) ∆2(N −k) 3 i=2 k=0 ! ! N −1 N X 512σ 2 log T ∆2(k) 512σ 2 log T π2 X = . + ∆(i) + 1− 2 ∆(N ) 3 i=2 ∆(k) ∆(k+1) N −3 X  TN −k−1 X (18) k=2 C Instance Dependent Terms In this section we devote our attention to the instance dependent terms in Theorems 1 and 2. We will first prove Corollary 1. Proof of Corollary 1. We will prove the two statements about the two estimators separately, (i) Going back to Lemma 4 in the proof of Theorem 1, we get that, P(Uk (t) < µ∗ ) ≥ 1 − when t > 144M 2 log2 (6/∆(2) ) log T ∆2k 1 t2 . This simply follows from the fact that ∆(2) is the smallest gap. Therefore, the chain leading to Eq. (14) follows with the new definition of Tk = 144M 2 log2 (6/∆(2) ) log T ∆2(k) . Hence, the regret of Algorithm 1 under estimator (3) is bounded as follows: 144M 2 log2 (6/∆(2) ) log T R(T ) ≤ ∆(2) = 1+ N −1 X k=2 µ)M 2 log2 (6/∆(2) ) log T 144λ(µ π2 + ∆(2) 3 N X 1− ∆2(k) !! ∆2(k+1) ! ∆(i) . π2 + 3 N X ! ∆(i) i=2 (19) i=2 We can analyze the same terms in an alternate manner. From Eq. (14) in the proof of Theorem 1, it follows that the regret of Algorithm 1 under the clipped estimator is bounded by, ! π2 X R(T ) ≤ ∆(N ) T2 + ∆(i) . 3 i Using the definition of T2 in (14) we obtain: R(T ) ≤ 144M 2 log2 (6/∆(2) ) log T 1 × . ∆(2) ∆(2) Combining the above equation with (19) we get the desired result. 17 (ii) Theorem 2 immediately implies that µ)σ 2 log T π2 256λ(µ + R(T ) ≤ ∆(2) 3 N X ! ∆(i) , i=2 for the median of means estimator. We can alternately analyze the regret as follows. From Eq. (18) in the proof of Theorem 2, it follows that the regret of Algorithm 1 under the median of means estimator is bounded by, ! π2 X R(T ) ≤ ∆(N ) T2 + ∆(i) . 3 i Using the definition of T2 in (18) we obtain: R(T ) ≤ 256σ 2 log T 1 × . ∆(2) ∆(2) Combining the equations above we get the desired result. Now we will work under the assumption that the gaps in the means of the experts are generated according to the generative model in Corollary 2. µ)] = O(log N ). Proof of Corollary 2. In light of Corollary 1, we just need to prove that Ep∆ [λ(µ Now, we will assume that {∆(i) } for i = 3, ..., N , are order statistics of N − 2 i.i.d uniform r.vs over the interval [∆(2) , 1]. Note that by Jensen’s we have the following: " 2 # ∆(k) ∆(k) 2 1−E . ≤1−E ∆2(k+1) ∆(k+1) (20) Let X = ∆(k) and Y = ∆(k+1) for k ≥ 3. The joint pdf of X, Y is given by, (N − 2)! f (x, y) = (k − 1)!(N − 3 − k)! x − ∆(2) 1 − ∆(2) k−1 y − ∆(2) N −3−k 1 × 1− . 1 − ∆(2) (1 − ∆(2) )2 Therefore, we have the following chain: Z 1 Z y x − ∆(2) k−1 y − ∆(2) N −3−k X x (N − 2)! 1 = × 1− dxdy 2 Y y (k − 1)!(N − 3 − k)! 1 − ∆ 1 − ∆ (1 − ∆ (2) (2) (2) ) y=∆(2) x=∆(2) Z 1 Z b (1 − ∆(2) )a + ∆(2) (N − 2)! k−1 N −3−k = × (a) (1 − b) dadb (1 − ∆ )b + ∆ (k − 1)!(N − 3 − k)! (2) (2) b=0 a=0 Z 1 Z b a (N − 2)! k−1 N −3−k × (a) (1 − b) dadb ≥ b=0 a=0 b (k − 1)!(N − 3 − k)! k = . k+1 E 18 Combining this with Eq (20) yields, Ep∆ "N −1 X k=2 1− ∆2(k) !# ∆2(k+1) N −1 X k2 ≤1+ 1− (k + 1)2 k=3 N −1 X 2k + 1 =1+ (k + 1)2 k=3 ≤1+ N −1 X k=3 2 k+1 (21) ≤ 1 + 2 log N. D More on Empirical Results Instance Dependent Term In this section we provide more details about our empirical results under the following sub-headings. D-UCB Term UCB1 Term 200 150 100 50 0 25 50 75 100 Number of Experts: N Figure 3: We plot the instance-dependent terms from D-UCB bounds (the term in Theorem 2 involving the gaps (22)) and that of UCB-1 bounds (23) as the number of experts grows in the stream analytics dataset. It can be observed that the instance dependent term from D-UCB grows at a much slower pace with the number of experts, and in fact stops increasing after a certain point. Training of Stochastic Experts: In Algorithm 2, new experts are added before starting a new batch. These stochastic experts are classifying functions trained using cost-sensitive classification oracles on data observed so far, which uses the ideas in [8]. The key idea is to reduce the cost-sensitive classification problem into importance weighted classification, which can be solved using binary classifiers by providing weights to each samples. Suppose a context x is observed and Algorithm 2 chooses an expert πi and draws an arm a from the conditional distribution πi (V \|x). Suppose the reward observed is r(a). Then the training sample (x, a) with a sample weight of r(a)/πi (a\|x) is added to the dataset for training the next batch of experts. It has been shown that this importance weighing yields good classification experts. These classifiers can provide confidence scores for arms, given a context and hence can serve as stochastic experts. 4 different experts are added at the beginning of each batch, out of which three are trained by XgBoost as baseclassifier while one is trained by logistic regression. Diversity is maintained among the experts added by training them on bootstrapped versions of the data observed so far, and also through selecting different hyper-parameters. Note that the parameter selection scheme is not tuned per dataset, but is held fixed for all three datasets. 19 Estimating Divergence Parameters: Both our divergence metrics Mij ’s and σij ’s can be estimated from data observed so far, during a run of Algorithm 2. These divergences do not depend on the arm chosen, but only on the context distribution and the conditional distributions encoded by the expert. Therefore, they can be easily estimated from data observed. Suppose, n contexts have been observed so far {x1 , ..., xn }. We are interested in estimating σij that isPthe chi-square divergence between πi and πj . An estimator for n this would be the empirical mean (1/n) × k=1 Df2 (πi (.\|xk )kπj (.\|xk )). Note that the distribution over the arms πj (.\|xk ) is nothing but the confidence scores observed through evaluation of the classifying oracle πj on the features/context xi . In order to be robust, we use the median of means estimator instead of the simple empirical mean for estimating the divergences. Empirical Analysis of Instance Dependent terms: In this section we empirically validate that our instance dependent terms in Theorem 1 and 2 are indeed much smaller compared to corresponding terms in the UCB-1 [5] regret bounds, even in real problem where our generative assumptions do not hold. In order to showcase this, we plot the instance-dependent term in Theorem 2 which is given by, ! N −1 X ∆2(k) 1 1 1− 2 + (22) ∆(k) ∆(k+1) ∆(N ) k=2 along with the corresponding term in UCB-1 bounds given by, N X k=2 1 , ∆(k) (23) as the number of stochastic experts grow in the stream dataset experiments in Section 7. The true means of the experts have been estimated in hindsight over the whole dataset. The plot is shown in Fig. 3. It can be observed that the term in the bounds of D-UCB grows at a much slower pace, and in fact stops increasing with the number of experts after a certain point. 20	2cs.AI
TRANSACTIONS OF THE AMERICAN MATHEMATICAL SOCIETY Volume 00, Number 0, Pages 000–000 S 0002-9947(XX)0000-0 arXiv:1601.08231v5 [math.RA] 22 Jan 2018 GENERALIZED NIL-COXETER ALGEBRAS OVER DISCRETE COMPLEX REFLECTION GROUPS APOORVA KHARE Abstract. We define and study generalized nil-Coxeter algebras associated to Coxeter groups. Motivated by a question of Coxeter (1957), we construct the first examples of such finite-dimensional algebras that are not the ‘usual’ nil-Coxeter algebras: a novel 2-parameter type A family that we call N CA (n, d). We explore several combinatorial properties of N CA (n, d), including its Coxeter word basis, length function, and Hilbert–Poincaré series, and show that the corresponding generalized Coxeter group is not a flat deformation of N CA (n, d). These algebras yield symmetric semigroup module categories that are necessarily not monoidal; we write down their Tannaka– Krein duality. Further motivated by the Broué–Malle–Rouquier (BMR) freeness conjecture [J. reine angew. math. 1998], we define generalized nil-Coxeter algebras N CW over all discrete real or complex reflection groups W , finite or infinite. We provide a complete classification of all such algebras that are finite-dimensional. Remarkably, these turn out to be either the usual nilCoxeter algebras, or the algebras N CA (n, d). This proves as a special case – and strengthens – the lack of equidimensional nil-Coxeter analogues for finite complex reflection groups. In particular, generic Hecke algebras are not flat deformations of N CW for W complex. Contents 1. Introduction and main results 2971 2. Background and motivation 2977 3. Proof of Theorem A: Tannakian formalism for semigroup categories 2979 4. Proof of Theorem B: Distinguished basis of words 2980 5. Proof of Theorem C, primitive elements, and categorification 2987 6. Proof of Theorem D: Finite-dimensional generalized nil-Coxeter algebras2990 Acknowledgments 2996 References 2996 1. Introduction and main results Throughout this paper, k will denote a fixed unital commutative ground ring. In this paper we define and study generalized nil-Coxeter algebras associated to Coxeter groups, and more generally to all discrete complex reflection groups, 2010 Mathematics Subject Classification. 20F55 (Primary), 20F05, 20C08 (Secondary). Key words and phrases. Complex reflection group, generalized Coxeter group, generalized nilCoxeter algebra, length function. c XXXX American Mathematical Society 2971 2972 APOORVA KHARE finite or infinite. These are algebras that map onto the associated graded algebras of (generic) Hecke algebras over complex reflection groups and of Iwahori–Hecke algebras over Coxeter groups. As we discuss, working with these algebras allows for a broader class than the corresponding reflection groups. We begin with real groups. Coxeter groups and their associated Hecke algebras play an important role in representation theory, combinatorics, and mathematical physics. Each such group is defined by a Coxeter matrix, i.e. a symmetric “integer” matrix M := (mij )i,j∈I with I finite and mii = 2 6 mij 6 ∞ ∀i 6= j. The Artin >0 monoid BM associated to M is generated by {Ti : i ∈ I} modulo the braid relations Ti Tj Ti · · · = Tj Ti Tj · · · for all i 6= j with mij < ∞, with precisely mij factors on either side. The Artin group (or generalized braid group) BM is the group generated by these relations; typically we use {si : i ∈ I} to denote its generators. There are three well-studied algebras associated to the matrix M : the group algebra kW (M ) of the Coxeter group, the 0-Hecke algebra [Ca, No], and the nil-Coxeter algebra N C(M ) [FS] (also called the nilCoxeter algebra, nil Coxeter algebra, and nil Hecke ring in the literature). These are all free k-modules, with a “Coxeter word basis” {Tw : w ∈ W (M )} and length function `(Tw ) := `(w) in W (M ); in each of them the Ti satisfy a quadratic relation. In a sense, the usual nil-Coxeter algebras N C(M ) are better-behaved than all other generic Hecke algebras (in which Ti2 = ai Ti + bi for scalars ai , bi , see [Hum, Chapter 7]): the words Tw have unique lengths and form a monoid together with 0. Said differently, the algebras N C(M ) are the only generic Hecke algebras that are graded with Ti homogeneous of positive degree. Indeed, if deg Ti = 1 ∀i, then Q d −1 and di are N C(M ) has Hilbert–Poincaré polynomial i∈I [di ]q , where [d]q := qq−1 the exponents of W (M ). We now introduce the main objects of interest in the present work: generalized Coxeter matrices and their associated nil-Coxeter algebras (which are always Z>0 graded). Definition 1.1. Define a generalized Coxeter matrix to be a symmetric ‘integer’ matrix M := (mij )i,j∈I with I finite, 2 6 mij 6 ∞ ∀i 6= j, and mii < ∞ ∀i. Now fix such a matrix M . (1) Given an integer vector d = (di )i∈I with di > 2 ∀i, define M (d) to be the matrix replacing the diagonal in M by d. (2) The generalized Coxeter group W (M ) is the quotient of the braid group ii BM2 by the order relations sm = 1 ∀i, where M2 is the Coxeter matrix i M ((2, . . . , 2)). We remark that we used the familiar Artin group BM2 associated to the Coxeter group W (M2 ); we could just as well have written BM , since the diagonals of the matrices M2 and M play no role in BM2 = BM . (3) The braid diagram, or Coxeter graph of M (or of W (M )) has vertices indexed by I, and for each pair i 6= j of vertices, mij − 2 edges between them. (4) Define the corresponding generalized nil-Coxeter algebra as follows (with M2 as above): (1.2) >0 kBM khTi , i ∈ Ii 2 N C(M ) := = , (Ti Tj Ti · · · = Tj Ti Tj · · ·, Timii = 0, ∀i 6= j ∈ I) (Timii = 0 ∀i) \| {z } \| {z } mij times mij times GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2973 where we omit the braid relation Ti Tj Ti · · · = Tj Ti Tj · · · if mij = ∞. (5) Given d = (di )i∈I as above, define WM (d) := W (M (d)) and N CM (d) := N C(M (d)). We are interested in the family of (generalized) nil-Coxeter algebras for multiple reasons: category theory, real reflection groups, complex reflection groups, and deformation theory. We elaborate on these motivations in this section and the next. 1.1. Tannaka–Krein duality for semigroup categories. In [Kho], the representation categories Rep N C(An ) were used to categorify the Weyl algebra. For now we highlight two properties of generalized nil-Coxeter algebras N C(M ) which also have categorical content: (i) for no choice of coproduct on N C(M ) can it be a bialgebra (shown below); and (ii) every algebra N C(M ) is equipped with a cocommutative coproduct ∆ : Ti 7→ Ti ⊗ Ti for all i ∈ I. Viewed through the prism of representation categories, the coproduct in (ii) equips Rep N C(M ) with the structure of a symmetric semigroup category [ES, §13,14]. Note by (i) that the simple module k does not serve as a unit object, whence Rep N C(M ) is necessarily not monoidal. It is natural to apply the Tannakian formalism to such categories with “tensor” structure. We record the answer which, while not surprising once formulated, we were unable to find in the literature. Definition 1.3. A semigroup-tensor category is a semigroup category (C, ⊗) which is also additive and such that ⊗ is bi-additive. Theorem A. Let A be an associative unital algebra over a field k, C := Rep A, and F : C → Veck the forgetful functor. (1) Any semigroup-tensor structure on C together with a tensor structure on F equips A with a coproduct ∆ : A → A ⊗ A that is an algebra map. (2) If the semigroup-tensor structure on C is braided (respectively, symmetric), then (A, ∆) is a quasi-triangular (respectively, triangular) algebra with coproduct. This simply means there exists an invertible element R ∈ A ⊗ A satisfying the ‘hexagon relations’ (1 ⊗ ∆)R = R13 R12 , and such that ∆ 1 ⊗ 1. op = R∆R −1 (∆ ⊗ 1)R = R13 R23 , . Triangularity means further that RR21 = Notice that generalized nil-Coxeter algebras are indeed examples of such triangular algebras, with a (cocommutative) coproduct but no counit. Such algebras are interesting in the theory of PBW deformations of smash product algebras; see Section 2. We also show below how to obtain an “honest” symmetric tensor category from each algebra N C(M ), via a central extension. As noted above, Theorem A is in a sense “expected”, and serves to act more as motivation. That the algebras N C(M ) provide concrete examples of symmetric, nonmonoidal semigroup-tensor categories is novel. The main results below now focus on the algebras N C(M ) themselves. 1.2. Real reflection groups and novel family of finite-dimensional nilCoxeter algebras. Our next result constructs a novel family of generalized nilCoxeter algebras of type A, which are finite-dimensional. Classifying the finitedimensional objects in Coxeter-type settings in algebra and combinatorics has been 2974 APOORVA KHARE a subject of tremendous classical and modern interest, including Weyl, Coxeter, and complex reflection groups, their nil-Coxeter and associated Hecke algebras; but also finite type quivers, Kleinian singularities, the McKay–Slodowy correspondence, simple Lie algebras, etc. A very recent setting involves the classification of finite-dimensional Nichols algebras. Some of the prominent ingredients in the study of these algebras are common to the present work. See [GHV, HV1, HV2] for more details. Another famous recent classification is that of finite-dimensional pointed Hopf algebras [AnSc], which turn out to arise from generalized small quantum groups. With these motivations, our goal is to similarly classify all generalized nil-Coxeter algebras, and our next result presents the first novel family of such examples. We remark that in equation (1.2), in generalizing the “order relations” from Ti2 = 0 to Timii = 0 we were also motivated by another such setting: the classical work of Coxeter [Cox3], which investigated generalized Coxeter matrices M for which the group W (M ) is finite. Specifically, Coxeter considered the (type A) Artin braid group BAn−1 , and instead of quotienting by the relations s2i = 1 to obtain the symmetric group Sn , he worked with spi = 1 ∀i ∈ I. Coxeter was interested in computing for which (n, p) is the quotient group WAn−1 ((p, . . . , p)) a finite group, and what is its order. He showed (see also [As]) that WAn−1 ((p, . . . , p)) is finite if and only if n1 + p1 > 21 , in which case the size of the quotient group is 1−n 1 1 1 · n!/nn−1 . Coxeter’s result was extended by Koster in his thesis + − n p 2 [Ko] to completely classify the “generalized Coxeter groups” W (M ) that are finite. These turn out to be precisely the finite Coxeter groups and the Shephard groups. Parallel to the above classical works, we wish to understand for which matrices M is the algebra N C(M ) finitely generated as a k-module. If W (M ) is a Coxeter group, then dim N C(M ) = \|W (M )\|. Few other answers are known. For instance, Marin [Mar] has shown that the algebra N CA2 ((m, n)) is not finitely generated when m, n > 3. However, apart from the usual nil-Coxeter algebras, to our knowledge no other finitely generated algebras N C(M ) were known to date. In the following result, following Coxeter’s construction in type A above, we exhibit the first such finite-dimensional family of algebras N C(M ). Theorem B. Given integers n > 1 and d > 2, define the k-algebra (1.4) N CA (n, d) := N CAn ((2, . . . , 2, d)). In other words, N CA (n, d) is generated by T1 , . . . , Tn , with relations: (1.5) (1.6) (1.7) ∀ 0 < i < n; Ti Ti+1 Ti = Ti+1 Ti Ti+1 , Ti Tj = Tj Ti , T12 = ··· = 2 Tn−1 = Tnd ∀ \|i − j\| > 1; = 0. Then N CA (n, d) is a free k-module, with k-basis of n!(1 + n(d − 1)) generators {Tw : w ∈ Sn } t {Tw Tnk Tn−1 Tn−2 · · · Tm+1 Tm : w ∈ Sn , k ∈ [1, d−1], m ∈ [1, n]}. In particular, for all l ∈ [1, n − 1], the subalgebra Rl generated by T1 , . . . , Tl is isomorphic to the usual nil-Coxeter algebra N CAl ((2, . . . , 2)). GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2975 Remark 1.8. We adopt the following notation in the sequel without further reference: let (1.9) w◦ ∈ Sn+1 , w◦0 ∈ Sn , w◦00 ∈ Sn−1 denote the respective longest elements, where the symmetric group Sl+1 corresponds to the k-basis of the algebra Rl for l = n − 2, n − 1, n. The algebras N CA (n, d) have not been studied previously for d > 2, and we begin to explore their properties. When d = 2, N CA (n, d) specializes to the usual nilCoxeter algebra of type An . In this vein, we present three properties of N CA (n, d) akin to the usual nil-Coxeter algebras. Theorem C. Fix integers n > 1 and d > 2. (1) The algebra N CA (n, d) has a length function that restricts to the usual length function `An−1 on Rn−1 ' N CAn−1 ((2, . . . , 2)) (from Theorem B), and (1.10) `(Tw Tnk Tn−1 · · · Tm ) = `An−1 (w) + k + n − m, for all w ∈ Sn , k ∈ [1, d − 1], and m ∈ [1, n]. (2) There is a unique longest word Tw◦0 Tnd−1 Tn−1 · · · T1 of length ln,d := `An−1 (w◦0 ) + d + n − 2. (3) If k is a field, then N CA (n, d) is local, with unique maximal ideal m generated by T1 , . . . , Tn . For all k, the ideal m is nilpotent with m1+ln,d = 0. We also study the algebra N CA (n, d) in connection to Khovanov’s categorification of the Weyl algebra. See Proposition 5.6 below. 1.3. Complex reflection groups and BMR freeness conjecture. Determining the finite-dimensionality of the algebras N C(M ) is strongly motivated by the study of complex reflection groups and their Hecke algebras. Recall that such groups were enumerated by Shephard–Todd [ST]; see also [Coh, LT]. Subsequently, Popov [Pop1] classified the infinite discrete groups generated by affine unitary reflections; in the sequel we will term these infinite complex reflection groups. For more on these groups, see e.g. [BS, Hug1, Hug2, Mal, ORS, ReS] and the references therein. For complex reflection groups, an important program is the study of generic Hecke algebras over them, as well as the associated BMR freeness conjecture of Broué, Malle, and Rouquier [BMR1, BMR2] (see also the recent publications [Et, Lo, Mar, MP] and the thesis [Ch]). This conjecture connects the dimension of a generic Hecke algebra to the order of the underlying reflection group. Here we will study this connection for the corresponding nil-Coxeter algebras, which we define as follows given [Be, BMR2]. Definition 1.11. Suppose W is a discrete (finite or infinite) complex reflection group, together with a finite generating set of complex reflections {si : i ∈ I}, the ii order relations sm = 1 ∀i, a set of braid relations {Rj : j ∈ J} – each involving i words with at least two distinct reflections si – and for the infinite non-Coxeter complex reflection groups W listed in [Mal, Tables I, II], one more order relation R0m0 = 1. Now define I0 := I t {0} for these infinite non-Coxeter complex reflection 2976 APOORVA KHARE groups W , and I0 := I otherwise. Given an integer vector d ∈ NI0 with di > 2 ∀i, define the corresponding generalized nil-Coxeter algebra to be (1.12) N CW (d) := ({Rj0 , j ∈ khTi , i ∈ Ii , = 0 ∀i ∈ I, (R00 )d0 = 0) J}, Tidi where the braid relations Rj are replaced by the corresponding relations Rj0 in the alphabet {Ti : i ∈ I}, and similarly for R00 if R0m0 = 1 in W . There is also the notion of the corresponding braid diagram as in [BMR2, Tables 1–4] and [Mal, Tables I, II]; this is no longer always a Coxeter graph. Note by [Pop1, §1.6] that in the above definition, one has to work with a specific presentation for complex reflection groups, as there is no canonical (minimal) set of generating reflections. See [Ba] for related work. There is no known finite-dimensional generalized nil-Coxeter algebra associated to a finite complex reflection group. Indeed, Marin mentions in [Mar] that a key difference between real and complex reflection groups W is the lack of nil-Coxeter algebras of dimension precisely \|W \| for the latter. This was verified in some cases for complex reflection groups in loc. cit. Our final result shows this assertion – and in fact a stronger statement – for all discrete finite and infinite, real and complex reflection groups. Even stronger (a priori ): we provide a complete classification of finite-dimensional generalized nil-Coxeter algebras for all such groups. Notice by [Pop1, Theorem 1.4] that it suffices to consider only the groups whose braid diagram is connected. Theorem D. Suppose W is any irreducible discrete real or complex reflection group. In other words, W is a real reflection group with connected braid diagram, or a complex reflection group with connected braid diagram and presentation given as in [BMR2, Tables 1–4], [Mal, Tables I, II], or [Pop1, Table 2]. Also fix an integer vector d with di > 2 ∀i (including possibly for the additional order relation as in [Mal]). Then the following are equivalent: (1) The generalized nil-Coxeter algebra N CW (d) is finitely generated as a kmodule. (2) Either W is a finite Coxeter group and di = 2 ∀i, or W is of type An and d = (2, . . . , 2, d) or (d, 2, . . . , 2) for some d > 2. (3) The ideal m generated by {Ti : i ∈ I} is nilpotent. If these assertions hold, there exists a length function and a unique longest element in N CW (d), say of length l; now m1+l = 0. In other words, the only finite-dimensional examples (when k is a field) are the usual nil-Coxeter algebras, and the algebras N CA (n, d). Note also that all of the above results are characteristic-free. A key tool in proving both Theorems B and D is a diagrammatic calculus, which is akin to crystal theory from combinatorics and quantum groups. 1.4. Further questions and Organization of the paper. To our knowledge, the algebras N CA (n, d) for d > 2 are a novel construction – and in light of Theorem D, the only finite-dimensional generalized nil-Coxeter algebras other than the ‘usual’ ones. In particular, a further exploration of their properties is warranted. We conclude this section by discussing some further directions. GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2977 (1) Nil-Coxeter algebras are related to flag varieties [BGG, KK], categorification [Kho, KL], and symmetric function theory [BSS]. Also recall, the divided difference operator representation of the usual type A nil-Coxeter algebra N CA (n, 2) is used to define Schubert polynomials [FS, LS], and the polynomials the Ti simultaneously annihilate are precisely the symmetric polynomials. It will be interesting to determine if N CA (n, d), d > 2 has a similar “natural” representation as operators on a polynomial ring; and if so, to consider the polynomials one obtains analogously. (See [Mar] for a related calculation.) We observe here that for d > 2, the algebra N CA (n, d) does not “come from” a finite reflection group, as it is of larger dimension than the corresponding generalized Coxeter group, by equation (2.1) below. (2) Given both the connection to Coxeter groups as well as the crystal methods used below, it will be interesting to explore if the algebras N CA (n, d) are connected to crystals over some Lie (super)algebra. (3) Our proof of Theorem D below involves a case-by-case argument, running over all discrete complex reflection groups. A type-free proof of this result would be desirable. The paper is organized as follows. In Section 2 we elaborate on our motivations and make additional remarks. In the following four sections we prove, in turn, the four main theorems above. 2. Background and motivation In this section we elaborate on some of the aforementioned motivations for studying generalized nil-Coxeter algebras and their finite-dimensionality. First, these algebras are interesting from a categorical perspective, as their module categories are symmetric semigroup-tensor categories (see Definition 1.3) but are not monoidal. We will discuss in the next section a Tannaka–Krein duality for such categories, as well as a central extension to a symmetric tensor category. The second motivation comes from real reflection groups: we provide a novel family of finite-dimensional algebras N CA (n, d) of type A (akin to the work of Coxeter [Cox3] and Koster [Ko]). In this context, it is remarkable (by Theorem D) that the algebras N CA (n, d) and the usual nil-Coxeter algebras N CW ((2, . . . , 2)) are the only finite-dimensional examples. As Theorem C shows, the algebras N CA (n, d) for d > 2 are similar to their “usual” nil-Coxeter analogues for d = 2. Note however that these algebras also differ in key aspects. See Theorem 5.2 and Proposition 5.3, which show in particular that for N CA (n, d) with d > 2, there are multiple “maximal” words, i.e., words killed by left- and right-multiplication by every generator Ti . A more fundamental difference arises out of considerations of flat deformations, which we make precise in the remarks around equation (2.1) below. Our third motivation comes from complex reflection groups and is of much recent interest: the BMR freeness conjecture, which discusses the equality of dimensions of generic Hecke algebras and (the group algebra of) the underlying finite complex reflection group. In this paper we study the associated graded algebra, i.e. where all deformation parameters are set to zero. As shown by Marin [Mar] in some of the cases, non-Coxeter reflection groups do not come equipped with finite-dimensional nil-Coxeter analogues. We make this precise in a strong way in 2978 APOORVA KHARE Theorem D above, for all complex reflection groups W . In particular, Theorem D shows that generic Hecke algebras are not flat deformations of their underlying (associated graded) nil-Coxeter analogues for complex W . This is a property shared by the algebras N CA (n, d) for d > 2 (but not by Iwahori–Hecke algebras of Coxeter groups W = W (M ), which are flat deformations of N C(M )). Indeed, if Mn,d denotes the generalized Coxeter matrix corresponding to N CA (n, d), then we claim that: (2.1) ( (n + 1)!, if d > 2 is even, dim N CA (n, d) = n!(1 + n(d − 1)) > \|W (Mn,d )\| = 1, if d > 2 is odd. To see (2.1), if mij is odd for any generalized Coxeter matrix M = M (d), then si , sj are conjugate in W (M ), whence sgi = sgj = 1 in W (M ) for g = gcd(di , dj ). On the other hand, N CM (d) surjects onto the nil-Coxeter algebra N CM ((2, . . . , 2)) if di > 2 ∀i. Now if di , dj > 2 are coprime, say, then si generates the trivial subgroup of W (M ), while Ti does not vanish in N C(M ). The generic Hecke algebras discussed above fit in a broader framework of deformation theory, which provides a fourth motivation behind this paper (in addition to the question of flatness discussed above). The theory of flat/PBW deformations of associative algebras is an area of sustained activity, and subsumes Drinfeld Hecke/orbifold algebras [Dr], graded affine Hecke algebras [Lu], symplectic reflection algebras and rational Cherednik algebras [EG], infinitesimal and other Hecke algebras, and other programs in the literature. We also highlight the program of Shepler and Witherspoon; see [SW1, SW2] and the references therein. In all of these settings, a bialgebra A (usually a Hopf algebra) acts on a vector space V and hence on a quotient SV of its tensor algebra, and one characterizes the deformations of this smash-product algebra A n SV which are flat, also termed the “PBW deformations”. In this regard, the significance of the generalized nil-Coxeter algebras N C(M ) is manifold. First, the above bialgebra settings were extended in recent work [Kha] to the framework of “cocommutative algebras” A, which also include the algebras N CW (d). Moreover, we characterized the PBW deformations of A n Sym(V ), thereby extending in loc. cit. the PBW theorems in the previously mentioned works. The significance of our framework incorporating A = N CW (d) along with the previously studied algebras, is that the full Hopf/bialgebra structure of A – specifically, the antipode or even counit – is not required in order to characterize the flat deformations of A n Sym(V ). Coming to finite-dimensionality, it was shown in the program of Shepler and Witherspoon (see e.g. [SW2]), and then in [Kha], that when the algebra A with coproduct is finite-dimensional over a field k, it is possible to characterize the graded k[t]-deformations of A n Sym(V ), whose fiber at t = 1 has the PBW property. For A = N CW (d), this deformation-theoretic consideration directly motivates our classification result in Theorem D. We conclude with a third connection to the aforementioned active program on PBW deformations. We studied in [Kha] the case when (A, m, ∆) is local with ∆(m) ⊂ m ⊗ m. In this setting, if m is a nilpotent two-sided ideal, then one obtains a lot of information about the deformations of A n Sym(V ), including understanding the PBW deformations, as well as their center, abelianization, and GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2979 modules, especially the simple modules. Now if A = N CW (d) then m is generated by the Ti ; this explains the interest above in understanding when m is nilpotent. Theorem D shows that this condition is in fact equivalent to the generalized nilCoxeter algebra being finite-dimensional. 3. Proof of Theorem A: Tannakian formalism for semigroup categories The remainder of this paper is devoted to proving the four main theorems in the opening section. We begin by studying the representation category of N C(M ) for a generalized Coxeter matrix M . The first assertion is that this category can never be a monoidal category in characteristic zero, and it follows from the following result. Proposition 3.1. Suppose k is a field of characteristic zero and M is a generalized Coxeter matrix. Then N C(M ) is not a bialgebra. The result fails to hold in positive characteristic. Indeed, for any prime p > 2 the algebra (Z/pZ)[T ]/(T p ) is a bialgebra, with coproduct ∆(T ) := 1 ⊗ T + T ⊗ 1 and counit ε(T ) := 0. Proof. Note there is a unique possible counit, ε : Ti 7→ 0 ∀i ∈ I. Now suppose ∆ : N C(M ) → N C(M ) ⊗ N C(M ) is such that (id ⊗ε) ◦ ∆ = id = (ε ⊗ id) ◦ ∆ on N C(M ). Setting m := ker ε to be the ideal generated by {Ti : i ∈ I}, it follows that (3.2) ∆(Ti ) ∈ 1 ⊗ Ti + Ti ⊗ 1 + m ⊗ m. Note that m ⊗ m constitutes the terms of higher ‘total degree’ in ∆(Ti ), in the Z>0 -grading on N C(M ). Now if ∆ is multiplicative, then raising (3.2) to the mii th power yields: mX ii −1 mii 0 = ∆(Ti )mii = Tik ⊗ Timii −k + higher degree terms. k k=1 This is impossible as long as the image of Ti in N C(M ) is nonzero; assuming this, it follows ∆ cannot be multiplicative, hence not a coproduct on N C(M ). Finally, N C(M ) surjects onto the usual nil-Coxeter algebra N C(M2 ) with M2 = M ((2, . . . , 2)). As N C(M2 ) has a Coxeter word basis indexed by W (M ), it follows that Ti is indeed nonzero in N C(M ). As a consequence of Proposition 3.1 and the Tannakian formalism in [ES, Theorem 18.3], for any generalized Coxeter matrix M the module category Rep N C(M ) is necessarily not a tensor category. That said, the map ∆ : Ti 7→ Ti ⊗ Ti is a coproduct on N C(M ), i.e. a coassociative algebra map. The cocommutativity of ∆ implies Rep N C(M ) is a symmetric semigroup category. We now outline how to show the first theorem above, which seeks to understand Tannaka–Krein duality for such categories (possibly without unit objects). Proof of Theorem A. The proof of part (1) follows that of [ES, Theorem 18.3]; one now ignores the last statement in that proof. The additional data required in the two braided versions in part (2) can be deduced from the proof of [ES, Proposition 14.2]. 2980 APOORVA KHARE We conclude this section by passing from Rep N C(M ) to an “honest” tensor category – say with k a field. Alternately, via the Tannakian formalism in [ES, g Theorem 18.3], we produce a bialgebra N C(M ) that surjects onto N C(M ). Namely, g N C(M ) is generated by {Ti : i ∈ I} and an additional generator T∞ , subject to the braid relations on the former set, as well as 2 := T∞ , ∀i ∈ I. Timii = Ti T∞ = T∞ Ti = T∞ g Note that N C(M ) is no longer Z>0 -graded; but it is a central extension: g 0 → kT∞ → N C(M ) → N C(M ) → 0. Now asking for all Ti and T∞ to be grouplike yields a unique bialgebra structure g on N C(M ): e : Ti 7→ Ti ⊗ Ti , ∆ T∞ 7→ T∞ ⊗ T∞ , εe : Ti , T∞ 7→ 1, g and hence a monoidal category structure on Rep N C(M ), as claimed. 4. Proof of Theorem B: Distinguished basis of words We now prove our main theorems on the algebras N C(M ) – specifically, the family N CA (n, d) – beginning with Theorem B. Note that if d = 2 then the algebra N CA (n, d) is the usual nil-Coxeter algebra, while if n = 1 then the algebra is k[T1 ]/(T1d ). Theorems B and C are easily verified for these cases, e.g. using [Hum, Chapter 7]. Thus, we assume throughout their proofs below that n > 2 and d > 3. We begin by showing the k-rank of N CA (n, d) is at most n!(1 + n(d − 1)). Notice that N CA (n, d) is spanned by words in the Ti . We now claim that a word in the Ti is either zero in N CA (n, d), or equal by the braid relations to a word in which all occurrences of Tn are successive, in a monomial Tnk for some 1 6 k 6 d − 1. To show the claim, consider a word T := · · · Tna Tw Tnb · · · , where a, b > 0 and Tw = Ti1 · · · Tik is a word in T1 , . . . , Tn−1 . Rewrite T using the braid relations if required, so that w ∈ Sn has minimal length, say k. We may assume k > 0, else we would be done. Now using the braid relations Ti Tn = Tn Ti for i 6 n − 2, further assume that i1 = ik = n − 1 (otherwise the factors may be ‘moved past’ the Tn using the braid relations). Similarly, i2 = ik−1 = n − 2, and so on. Thus, if Tw 6= 0, then assume by the minimality of `(w) that Tw = Tn−1 Tn−2 · · · Tm+1 Tm Tm+1 · · · Tn−2 Tn−1 , for some 1 6 m 6 n − 1. We next claim that the following relation holds in the Artin braid group Bn , hence in N CAn (d) for any d: (4.1) Tn−1 · · · Tm · · · Tn−1 = Tm Tm+1 · · · Tn−2 Tn−1 Tn−2 · · · Tm+1 Tm . This is shown by descending induction on m 6 n − 1. Hence, (4.2) Tna · (Tn−1 · · · Tm · · · Tn−1 ) · Tnb = Tna · (Tm · · · Tn−2 Tn−1 Tn−2 · · · Tm ) · Tnb = (Tm · · · Tn−2 )Tna−1 (Tn Tn−1 Tn )Tnb−1 (Tn−2 · · · Tm ) = (Tm · · · Tn−2 )Tna−1 (Tn−1 Tn Tn−1 )Tnb−1 (Tn−2 · · · Tm ). If max(a, b) = 1 then the claim follows; if a > 1 then the last expression contains 2 the substring (Tn Tn−1 Tn )Tn−1 = Tn−1 Tn Tn−1 = 0; and similarly if b > 1. This shows the claim. GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2981 We now prove the upper bound on the k-rank. Notice that T1 , . . . , Tn−1 generate a subalgebra Rn−1 ⊂ N CA (n, d) in which the nil-Coxeter relations for WAn−1 = Sn are satisfied. Hence the map : N CAn−1 ((2, . . . , 2)) Rn−1 := hT1 , . . . , Tn−1 i is an algebra map. Now notice by equation (4.2) that every nonzero word in N CA (n, d) \ Rn−1 is of the form T = Tw Tnk Tw0 , where 1 6 k 6 d − 1, w, w0 ∈ WAn−1 , and hence Tw , Tw0 ∈ Rn−1 . By a similar reasoning as above, assuming w0 of minimal length in Sn−1 , we may rewrite T such that Tw0 = Tn−1 · · · Tm for some 1 6 m 6 n. Carrying out this operation yields Tw00 Tnk Tn−1 · · · Tm for some reduced word w00 ∈ WAn−1 (i.e., such that Tw00 is nonzero in N CAn−1 ((2, . . . , 2))). Thus, N CA (n, d) = Rn−1 + d−1 X n X k=1 m=1 Rn−1 · Tnk · (Tn−1 · · · Tm ). As Rn−1 has at most n! generators, it follows that N CA (n, d) has at most (1 + n(d − 1)) · n! generators, which shows the desired upper bound on its k-rank. The hard part of the proof involves showing that the words Tw Tnk Tn−1 · · · Tm form a k-basis of N CA (n, d). We will require the following technical lemma on the symmetric group and its nil-Coxeter algebra. A proof is included for completeness. Lemma 4.3. Suppose W = WAn−1 = Sn is the symmetric group, with simple reflections s1 , . . . , sn−1 labelled as usual. Then every element w of WAn−1 \WAn−2 = Sn \ Sn−1 can be written in reduced form as w = w0 sn−1 · · · sm0 , where w0 ∈ Sn−1 = WAn−2 and m0 ∈ [1, n − 1] are unique. Given such an element w ∈ Sn , we have in the usual nil-Coxeter algebra N CAn ((2, . . . , 2)): ( Tw0 Tn−1 · · · Tm−1 · Tn · · · Tm0 , if m0 < m, (4.4) Tn · Tw · Tn · · · Tm = 0 otherwise. Note that equation (4.4) can be thought of as a statement on lengths in the symmetric group. Proof. We first claim that w ∈ WAn−1 \ WAn−2 has a reduced expression in which sn−1 occurs exactly once. The proof is by induction on n: clearly the claim is true for n = 2. Now given the claim for n − 2 > 2, consider any reduced expression for w that contains a sub-word sn−1 w00 sn−1 , where w ∈ WAn−2 . By the induction hypothesis, w00 = w0 sn−2 · · · sm for some w0 ∈ WAn−3 and m ∈ [1, n − 1]. Hence if m 6 n − 2, then w = · · · sn−1 (w0 sn−2 sn−3 · · · sm ) sn−1 = · · · w0 (sn−1 sn−2 sn−1 )(sn−3 · · · sm ) · · · , and by the braid relations, this equals a reduced expression for w ∈ WAn−1 , with one less occurrence of sn−1 . A similar analysis works if m = n − 1. Repeatedly carrying out this procedure proves the claim. We can now prove the uniqueness of w0 , m0 as in the lemma. By the previous paragraph, write w ∈ WAn−1 \ WAn−2 as w = w1 sn−1 w2 , with w1 , w2 ∈ WAn−2 and w2 of smallest possible length, say w2 = si1 · · · sik for i1 , . . . , ik 6 n − 2. Using the braid relations, clearly i1 = n−2, hence i2 = n−3 (by minimality of k). Choose the smallest l > 3 such that il 6= n − 1 − l. We now produce a contradiction assuming that such an integer l exists. If il < n − 1 − l, then we may move sil past the preceding terms, contradicting the minimality of k. Clearly il 6= n − l, else w2 was 2982 APOORVA KHARE not reduced. Thus il > n − 1 − l, whence w2 is of the form sn−2 · · · sil sil −1 sil · · · . Now verify in WAn−1 that w = w1 sn−1 sn−2 · · · sil +1 sil sil −1 sil · · · = w1 sn−1 · · · sil +1 sil −1 sil sil −1 · · · = w1 sil −1 · sn−1 · sn−2 · · · sil +1 sil sil −1 · · · , which contradicts the minimality of k. Thus such an integer l cannot exist, which proves that w = w1 sn−1 · · · sm0 for some m0 ∈ [1, n − 1]. We next claim that the integer m0 is unique for w ∈ WAn \ WAn−1 . We first make the sub-claim that if w ∈ WAn−1 is reduced, then so is wsn sn−1 · · · sm . To see why, first recall [Hum, Lemma 1.6, Corollary 1.7], which together imply that if wα > 0 for any finite Coxeter group W , any w ∈ W , and any simple root α > 0, then `(wsα ) = `(w)+1. Now the sub-claim follows by applying this result successively to (wsn · · · sj+1 , αj ) for j = n, n−1, · · · , m. Next, define Cm := WAn−1 ·sn sn−1 · · · sm , with Cn+1 := WAn−1 . It follows by the sub-claim above that \|Cm \| = \|WAn−1 \| = n! for all m. Hence, (n + 1)! = \|WAn \| 6 n+1 X m=1 \|Cm \| 6 n+1 X n! = (n + 1)!. m=1 Fn+1 This shows that WAn = m=1 Cm , which proves the uniqueness of m in the above claim. Now write w1 in reduced form to obtain that w0 = wsm0 · · · sn−1 is also unique. It remains to show equation (4.4) in N CAn ((2, . . . , 2)). Using the above analysis, write Tw = Tw0 Tn−1 · · · Tm0 ; since Tn commutes with Tw0 , we may assume w0 = 1. First suppose m0 > m. Then it suffices to prove that (Tn · · · Tm0 )2 = 0 for all 1 6 m0 6 n. Without loss of generality we may work in the subalgebra generated by Tm0 , . . . , Tn , and hence suppose m0 = 1. We now prove by induction that (Tn · · · T1 )2 = 0. This is clear if n = 1, 2, and for n > 2, (Tn · · · T1 )2 = Tn Tn−1 Tn · Tn−2 · · · T1 · Tn−1 · · · T1 = Tn−1 Tn · (Tn−1 · · · T1 )2 = 0. Next suppose m0 < m; once again we may suppose m0 = 1. We prove the result by induction on n, the base case of n = 2 (and m = 2) being easy. Thus, for 1 < m 6 n, we compute: Tn · · · T1 · Tn · · · Tm = Tn Tn−1 Tn · Tn−2 · · · T1 · Tn−1 · · · Tm = Tn−1 Tn · (Tn−1 · · · T1 ) · (Tn−1 · · · Tm ) = Tn−1 Tn · (Tn−2 · · · Tm−1 ) · (Tn−1 · · · T1 ) = Tn−1 Tn−2 · · · Tm−1 · Tn Tn−1 · · · T1 . Remark 4.5. Notice that equation (4.4) holds in any algebra containing elements T1 , . . . , Tn that satisfy the braid relations and T12 = 0. In particular, (4.4) holds in N CA (n, d) for n > 1. Returning to the proof of Theorem B, we now introduce a diagrammatic calculus akin to crystal theory. We first write out the n = 2 case, in order to provide intuition for the case of general n. Let M be a free k-module, with basis given by the nodes in the graph in Figure 1 below. In the figure, the node 122 1 should be thought of as T1 T22 T1 (applied to the unit 1N CA (2,d) , i.e., to the generating basis vector corresponding to ∅), and similarly for the other nodes. The arrows denote the action of T1 and T2 ; all remaining GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2983 1 0 12d 1 122 1 ··· 1 1 0 2 ··· 2 22 1 2 2 2 2 2 1 2d 1 ∅ 1 21 12 1 22 2 ··· 2 1 122 2d 0 1 ··· 0 12d 2 w◦ Figure 1. Regular representation for N CA (2, d), with d0 = d − 1 generator actions on nodes yield zero. Now one verifies by inspection that the defining relations in N CA (2, d) are satisfied by this action on M . Therefore M is an N CA (2, d)-module of k-rank 4d − 2 = 2!(1 + 2(d − 1)). Since M is generated by the basis vector corresponding to the node ∅, we have a surjection : N CA (2, d) M that sends T1 , T2k , T1 T2k , T2k T1 , T1 T2k T1 to the corresponding basis vectors in the free k-module M . Now the result for n = 2 follows by the upper bound on the k-rank, proved above. The strategy is similar for general n, but uses the following more detailed notation. For each w ∈ Sl with l 6 n, let Tw denote the corresponding (welldefined) word in the alphabet {T1 , . . . , Tl−1 }, and let Rl−1 denote the subalgebra of N CA (n, d) generated by these letters. Now define a free k-module M of k-rank n!(1 + n(d − 1)), with basis elements the set of words (4.6) B := {B(w, k, m) : w ∈ Sn , k ∈ [1, d − 1], m ∈ [1, n]} t {B(w) : w ∈ Sn }. We observe here that the basis vectors B(w, k, m), B(w) are to be thought of as corresponding respectively to the words (4.7) Tw Tnk Tn−1 · · · Tm , Tw , w ∈ Sn , k ∈ [1, d − 1], m ∈ [1, n]. Definition 4.8. An expression for a word in N CA (n, d) of the form (4.7) will be said to be in standard form. We now define an N CA (n, d)-module structure on M , via defining a directed graph structure on B (or more precisely, on B t {0}) that we now describe. The following figure (Figure 2 below) may help in visualizing the structure. The figure should be thought of as analogous to the central hexagon and either of the two “arms” in Figure 1. We begin by explaining the figure. Each node (wkm) (or (w)) corresponds to the basis vector B(w, k, m) (or B(w)). Notice that the vectors {B(w, 1, m)} t {B(w)} are in bijection with the Coxeter word basis of the usual nil-Coxeter algebra N CAn ((2, . . . , 2)). Let V1 denote their span, of k-rank (n + 1)!. Now given 1 6 m 6 n and 1 6 k 6 d − 1 =: d0 , define Vk,m to be the span L of the basis elements {B(w, k, m) : w ∈ Sn }, of k-rank n!. Then M = V1 ⊕ k>1,m Vk,m . Note as a special case that in Figure 1, the central hexagon spans V1 , the nodes 2k 1, 12k 1 span Vk,1 , and 2k , 12k span Vk,2 . We now define the N CA (n, d)-action: 2984 APOORVA KHARE n 11m w0 1m V1 12m n w0 2m n n ··· n 1d0 m ··· V2,m w0 d0 m n Vd0 ,m Figure 2. Regular representation for N CA (n, d), with d0 = d − 1 • Let V1,n+1 denote the k-span of {B(w) : w ∈ Sn }. Then for 1 6 m 6 n + 1, each V1,m has a distinguished basis in bijection with Sn ; the same holds for each Vk,m with k ∈ [2, d − 1] and m ∈ [1, n]. Now equip all of the above spaces Vk,m with the corresponding module structure over the usual nil-Coxeter algebra of type An−1 . Such a structure is uniquely determined, if given w = si1 · · · sil ∈ Sn with all ij < n, we set Tw · B(1, k, m) := B(w, k, m) and Tw · B(1) := B(w). • We next define the action of Tn on M . Via Lemma 4.3, write w ∈ Sn as w0 sn−1 · · · sm0 with w0 , m0 unique. Now using the previous paragraph, it follows that B(w, k, m) = Tw0 Tn−1 · · · Tm0 · B(1, k, m). Correspondingly, if w ∈ Sn−1 (i.e., m0 = n), define Tn · B(w, k, m) := 1(k 6 d − 2)B(w, k + 1, m), Tn · B(w) := B(w, 1, n). 0 • On the other hand, suppose m 6 n − 1. If k > 2, then define Tn · B(w, k, m) := 0. Otherwise define Tn · B(w) := B(w0 , 1, m0 ) with w0 , m0 as in Lemma 4.3, and (see Equation 4.4): ( B(w0 sn−1 · · · sm−1 , 1, m0 ), if m0 < m, (4.9) Tn · B(w, 1, m) := 0 otherwise. It remains to show that the above graph structure indeed defines an N CA (n, d)module structure on M ; then a similar argument as above (in the n = 2 case) completes the proof. In the following argument, we will occasionally use Lemma 4.3 (as well as Remark 4.5) without reference. First notice that the algebra relations involving only T1 , . . . , Tn−1 are clearly satisfied on M as it is a Rn−1 -free module by construction. To verify that the relations involving Tn hold on M , notice (e.g. via Figure 2) that the k-basis B of M can be partitioned into three subsets: B1 := {B(w, k, m) : k > 1, m ∈ [1, n], w ∈ Sn−1 } t {B(w) : w ∈ Sn−1 }, (4.10) B2 := {B(w, k, m) : k > 2, m ∈ [1, n], w ∈ Sn \ Sn−1 }, B3 := {B(w, 1, m) : m ∈ [1, n], w ∈ Sn \ Sn−1 } t {B(w) : w ∈ Sn \ Sn−1 }. Recall by the opening remarks in Section 4 that n > 2 and d > 3. We first show that the relation Tnd = 0 holds as an equality of linear operators on each vector b ∈ B, and hence on the k-module M . We separately consider the cases b ∈ Bi for i = 1, 2, 3, as in (4.10). GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2985 (1) If b = B(w, k, m) ∈ B1 , then b lies in the “top rows” of Figure 2. It is easily verified that Tnd−k−1 · B(w, k, m) = B(w, d − 1, m), and this is killed by Tn , as desired. The same reasoning shows that Tnd kills b = B(w) for w ∈ Sn−1 . (2) Let b = B(w, k, m) ∈ B2 . Then the relation holds on b since Tn · B(w, k, m) = 0. (These correspond to vectors in Vk,m for k > 2, which do not lie in the “top rows” in Figure 2.) (3) Finally, let b ∈ B3 ; thus w ∈ Sn \ Sn−1 , and we write w = w0 sn−1 · · · sm0 by Lemma 4.3. It follows from Remark 4.5 that Tn2 · B(w, 1, m) = 0 and Tnd · B(w) = Tnd−1 · B(w0 , 1, m0 ) = 0. We next show that the relation Ti Tn = Tn Ti holds on B for all i 6 n − 2. We consider the same three cases as in (4.10). (1) Fix w ∈ Sn−1 . If b = B(w, k, m) with k > 1, then verify using the aforementioned action that both Ti Tn · B(w, k, m) and Tn Ti · B(w, k, m) equal B(si w, k + 1, m) if `(si w) > `(w) and k 6 d − 2, and 0 otherwise. Similarly, Ti Tn · B(w) = 1(`(si w) > `(w))B(si w, 1, n) = Tn Ti · B(w). (2) Let b = B(w, k, m) with w ∈ Sn \Sn−1 and k > 2. Then Ti Tn ·B(w, k, m) = 0. To compute Tn Ti · B(w, k, m), since i 6 n − 2, it follows that si w ∈ Sn \ Sn−1 . If Ti Tw = 0 then we are done since B(w, k, m) = Tw ·B(1, k, m). Else note that si w ∈ Sn \ Sn−1 , whence Tn Ti · B(w, k, m) = Tn · B(si w, k, m) = 0 from above. (3) Finally, let w ∈ Sn \ Sn−1 and write w = w0 sn−1 · · · sm0 by Lemma 4.3. First suppose b = B(w, 1, m). If `(si w) < `(w), then it is not hard to show that both Ti Tn · B(w, 1, m) and Tn Ti · B(w, 1, m) vanish. Otherwise both terms are equal to B(si w0 sn−1 · · · sm−1 , 1, m0 ). A similar analysis shows that if `(si w) < `(w), then Ti Tn · B(w) = Tn Ti · B(w) = 0, otherwise Ti Tn · B(w) = Tn Ti · B(w) = B(si w0 , 1, m0 ). Next, we show that the braid relation Tn−1 Tn Tn−1 = Tn Tn−1 Tn holds on B. This is the most involved computation to carry out. We consider the same three cases as above. (1) Fix w ∈ Sn−1 . If b = B(w, k, m) with k > 2, then it is easily verified that both sides of the braid relation kill B(w, k, m). If instead k = 1, then Tn Tn−1 Tn · B(w, 1, m) = Tn Tn−1 · B(w, 2, m) = 0. To compute the other side, first notice B(w, 1, m) = Tn · B(wsn−1 · · · sm ). Hence, Tn Tn−1 · B(w, 1, m) = Tn Tn−1 Tn · B(wsn−1 · · · sm ). Now if the braid relation holds on B(w) for all w ∈ Sn , then Tn−1 Tn Tn−1 · B(w, 1, m) = Tn−1 · Tn Tn−1 Tn · B(wsn−1 · · · sm ) 2 = Tn−1 Tn−1 Tn Tn−1 · B(wsn−1 · · · sm ) ∈ Tn−1 · M = 0, where the last equality follows from the definition of M as an Rn−1 -module. It thus suffices for this case to verify that the braid relation holds on B(w) for w ∈ Sn . This is done by considering the following four sub-cases. (a) If w ∈ Sn−2 commutes with sn−1 , sn , then both Tn Tn−1 Tn · B(w) and Tn−1 Tn Tn−1 · B(w) are easily seen to equal B(sn−1 w, 1, n − 1). 2986 APOORVA KHARE (b) Suppose w = w0 sn−2 · · · sm0 ∈ Sn−1 \ Sn−2 , with w0 ∈ Sn−2 and m0 ∈ [1, n−2]. Then using Remark 4.5 and the Rn−1 -module structure of M , we compute: Tn Tn−1 Tn · B(w) = Tn Tn−1 · B(w, 1, n) = Tn · B(w0 sn−1 · · · sm0 , 1, n) = B(w0 sn−1 , 1, m0 ), Tn−1 Tn Tn−1 · B(w) = Tn−1 Tn · B(w0 sn−1 · · · sm0 ) = Tn−1 · B(w0 , 1, m0 ) = B(sn−1 w0 , 1, m0 ), whence we are done since sn−1 commutes with w0 ∈ Sn−2 . (c) In the last two sub-cases, w = w0 sn−1 · · · sm0 ∈ Sn \ Sn−1 with m0 ∈ [1, n − 1]. As in the previous two sub-cases, first suppose w0 ∈ Sn−2 . Similar to the above computations, one verifies that both Tn Tn−1 Tn · B(w) and Tn−1 Tn Tn−1 · B(w) vanish. (d) Finally, suppose w = w00 sn−2 · · · sm · sn−1 · · · sm0 with m ∈ [1, n − 2], m0 ∈ [1, n − 1], and w00 ∈ Sn−2 . Then one verifies that Tn Tn−1 Tn · B(w) = 1(m < m0 )B(w00 sn−2 · · · sm0 −1 , 1, m) = Tn−1 Tn Tn−1 · B(w). (2) Next suppose b ∈ B2 is of the form B(w, k, m) with k > 2 and w = w0 sn−1 · · · sm0 ∈ Sn \ Sn−1 . Then Tn · B(w, k, m) = 0 by definition, so Tn Tn−1 Tn · B(w, k, m) = 0. To show that Tn−1 Tn Tn−1 kills B(w, k, m), we consider two sub-cases. If w0 ∈ Sn−2 , then Tn−1 · B(w, k, m) = 0 and we are done. Otherwise suppose w0 = w00 sn−2 · · · sm00 ∈ Sn−1 \ Sn−2 , with m00 ∈ [1, n − 2]. Now compute using Remark 4.5 and the relations verified above: Tn−1 Tn Tn−1 · B(w, k, m) = Tn−1 Tn · B(w00 sn−1 · · · sm00 sn−1 · · · sm0 , k, m) = 1(m00 < m0 )Tn−1 Tn · B(w00 sn−2 · · · sm0 −1 sn−1 · · · sm00 , k, m) = 1(m00 < m0 )Tn−1 · 0 = 0, where the penultimate equality uses that k = 2. (3) Finally, suppose b ∈ B3 . By the analysis in the first case above, we only need to consider b = B(w, 1, m) with w = w0 sn−1 · · · sm0 ∈ Sn \ Sn−1 . It is now not hard to show that both Tn Tn−1 Tn · B(w, 1, m) and Tn−1 Tn Tn−1 · B(w, 1, m) vanish if w0 ∈ Sn−2 . On the other hand, if w0 = w00 sn−2 · · · sm00 ∈ Sn−1 \ Sn−2 , then repeated use of Remark 4.5 (and equation (4.4)) shows that Tn−1 Tn Tn−1 · B(w, 1, m) = Tn−1 Tn Tn−1 · B(w00 · sn−2 · · · sm00 · sn−1 · · · sm0 , 1, m) = 1(m00 < m0 )Tn−1 Tn · B(w00 · sn−2 · · · sm0 −1 · sn−1 · · · sm00 , 1, m) = 1(m00 < m0 )1(m00 < m)Tn−1 · B(w00 · sn−2 · · · sm0 −1 · sn−1 · · · sm−1 , 1, m00 ) = 1(m00 < m0 )1(m00 < m)1(m0 < m)B(w00 sn−2 · · · sm−2 sn−1 · · · sm0 −1 , 1, m) = 1(m00 < m0 < m)B(w00 sn−2 · · · sm−2 sn−1 · · · sm0 −1 , 1, m). GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2987 Notice this calculation shows the “braid-like” action of Tn , Tn−1 on strings of the type Tn−2 · · · Tm00 , Similarly, one shows that Tn−1 · · · Tm0 , Tn · · · Tm . Tn Tn−1 Tn · B(w, 1, m) = 1(m00 < m0 < m)B(w00 sn−2 · · · sm−2 sn−1 · · · sm0 −1 , 1, m), which verifies that the last braid relation holds in the last case. Thus the algebra relations hold on all of M , making it an N CA (n, d)-module generated by B(1), as claimed. In particular, N CA (n, d) ' M as k-modules, by the analysis in the first part of this proof. This completes the proof of all but the last assertion in Theorem B. Finally, the nil-Coxeter algebra N CAl ((2, . . . , 2)) surjects onto Rl , and Rl ' Rl · B(∅) ⊂ V1 is free of k-rank (l + 1)! from above. Hence Rl ' N CAl ((2, . . . , 2)), as desired. 5. Proof of Theorem C, primitive elements, and categorification In this section we continue our study of the algebras N CA (n, d), starting with Theorem C. Proof of Theorem C. We retain the notation of Theorem B. Via the k-module isomorphism M ' N CA (n, d), we identify the basis element B(w, k, m) with Tw Tnk Tn−1 · · · Tm and B(w) with Tw , where w ∈ Sn , k ∈ [1, d − 1], and m ∈ [1, n]. Let ` : B → Z>0 be as in equation (1.10). We now claim that if T = Ti1 · · · Til is any nonzero word in N CA (n, d), then l is precisely the length of T when expressed (uniquely) in standard form (4.7). The proof is by induction on l. For l = 1, Ti is already in standard form (and nonzero). Now given a word T = Ti T 0 of length l + 1 (so T 0 has length l and satisfies the claim), write T 0 via the induction hypothesis as a word in standard form of length l. Now the proof of Theorem B shows that applying any Ti to this standard form for T 0 either yields zero or has length precisely l + 1. This proves the claim. The above analysis shows (1) and (2). Now suppose k is a field. Then the algebra N CA (n, d) has a maximal ideal m = h{Ti : i ∈ I}i; in fact, m has k-corank 1 by the proof of Theorem B. Moreover, m is local because any element of A \ m is invertible. (In particular, one understands representations of the algebra N CA (n, d), e.g. by [Kha, §6.1].) The aforementioned claim also proves that ml+1 = 0, where l := `An−1 (w◦0 ) + d + n − 2. This is because any nonzero word can be expressed in standard form without changing the length. As an immediate consequence, we have: Corollary 5.1. If k is a field and T1 , . . . , Tn all have graded degree 1, the Hilbert– Poincaré series of N CA (n, d) is the polynomial n Y qn − 1 , [n]q ! := [j]q . [n]q ! (1 + [n]q [d − 1]q ), where [n]q := q−1 j=1 The proof also uses the standard result that the Hilbert–Poincaré series of the usual nil-Coxeter algebra N CA (n, 2) is [n]q ! (see e.g. [Hum, §3.12, 3.15]). Next, we discuss a property that was explored in [Kho] for the usual nil-Coxeter algebras N CA (n, 2): these algebras are always Frobenius. We now study when the 2988 APOORVA KHARE algebras N CA (n, d) are also Frobenius for d > 3. As the following result shows, this only happens in the degenerate case of n = 1, i.e., k[T1 ]/(T1d ). Theorem 5.2. Suppose k is a field. Given n > 1 and d > 2, the algebra N CA (n, d) is Frobenius if and only if n = 1 or d = 2. One checks via equation (2.1) that these conditions are further equivalent to (the group algebra of) the “generalized Coxeter group” W (Mn,d ) being a flat deformation of N CA (n, d). The proof of Theorem 5.2 crucially uses the knowledge of “maximal”, i.e., primitive words in the algebra N CA (n, d). Formally, given a generalized Coxeter matrix M , say that an element x ∈ N C(M ) is left (respectively, right) primitive if mx = 0 (respectively, xm = 0), cf. Theorem C(3). Now x is primitive if it is both left- and right-primitive. Denote these sets of elements respectively by PrimL (N C(M )), PrimR (N C(M )), Prim(N C(M )). Proposition 5.3. Every generalized nil-Coxeter algebra N C(M ) is equipped with an anti-involution θ that fixes each generator Ti . Now θ is an isomorphism : PrimL (N C(M )) ←→ PrimR (N C(M )). Moreover, the following hold. (1) If W (M ) is a finite Coxeter group with unique longest word w◦ , then PrimL (N C(M )) = PrimR (N C(M )) = Prim(N C(M )) = kTw◦ . (2) If N C(M ) = N CA (1, d), then PrimL (N C(M )) = PrimR (N C(M )) = Prim(N C(M )) = k · T1d−1 . (3) If N C(M ) = N CA (n, d) with n > 2 and d > 3, then: (a) PrimL (N C(M )) is spanned by Tw◦ := Tw◦0 Tn Tn−1 · · · T1 and the n(d − 2) words {Tw◦0 Tnk Tn−1 · · · Tm : k ∈ [2, d − 1], m ∈ [1, n]}. (b) Prim(N C(M )) is spanned by the d − 1 words Tw◦0 Tnk Tn−1 · · · T1 , where 1 6 k 6 d − 1. In all cases, the map θ fixes both Prim(N C(M )) as well as the lengths of all nonzero words. Proof. The first two statements are obvious since θ preserves the defining relations in kh{Ti : i ∈ I}i. The assertion in (1) is standard – see e.g. [Hum, Chapter 7] – and (2) is easily verified. We next classify the left-primitive elements as in (3)(a). Suppose for some k > 2 and 1 6 m 6 n that T = Tw◦0 Tnk Tn−1 · · · Tm . Then clearly Ti T = 0 for all i < n, and Tn T = 0 since k > 2, as discussed in the proof of Theorem B. Similarly, if T = Tw◦ then Ti T = 0 for i < n, and we also computed in the proof of Theorem B that Tw◦ = T1 · · · Tn−1 Tn Tw◦0 . Hence, Tn Tw◦ = T1 · · · Tn−2 (Tn Tn−1 Tn )Tw◦0 = T1 · · · Tn−1 Tn (Tn−1 Tw◦0 ) = 0. To complete the proof of (3)(a), it suffices to show that no nonzero linear combination of the remaining words of the form Tw Tnk Tn−1 · · · Tm is left-primitive. Suppose first that there is a word w ∈ WAn−1 such that the coefficient of Tw is nonzero. In that case, choose such an element w of smallest length, and left-multiply the linear combination by Tnd−1 Tw◦0 w−1 . As discussed in the proof of Theorem B, this kills all terms Tw0 Tnk Tw00 with w0 , w00 ∈ WAn−1 and k > 1. Moreover, by [Hum, Chapter 7], GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2989 left-multiplication by Tw◦0 w−1 also kills all terms of the same length that are not Tw . Thus we are left with Tnd−1 Tw◦0 w−1 Tw = Tnd−1 Tw◦0 6= 0, so the linear combination was not left-primitive. The other case is that all words in the linear combination are of the form Tw Tnk Tn−1 · · · Tm with k > 1. Once again, choose w ∈ WAn−1 of smallest length for which the corresponding word has nonzero coefficient, and left-multiply by Tw◦0 w−1 . This yields a nonzero linear combination by the analysis in Theorem B, which proves the assertion about left-primitivity. We next identify the primitive elements in N C(M ) = N CA (n, d). The first claim is that Tk := Tw◦0 Tnk Tn−1 · · · T1 is fixed by θ. Indeed, we compute using the braid relations in type A that θ fixes Tw◦0 ∈ Rn−1 and Tw◦00 ∈ Rn−2 . Hence, (5.4) θ(Tk ) = T1 · · · Tn−1 Tnk Tw◦00 Tn−1 · · · T1 = T1 · · · Tn−1 Tw◦00 Tnk Tn−1 · · · T1 = Tk . Using this we claim that Tk is right-primitive. Indeed, if i < n, then Tk Ti = T1 · · · Tn−1 Tnk Tw◦0 Ti = 0, while for i = n, we compute: Tk Tn = Tw◦0 Tnk Tn−1 Tn · Tn−2 · · · T1 = Tnk−1 Tw◦0 Tn−1 · Tn Tn−1 Tn−2 · · · T1 = 0. We now claim that no linear combination of the remaining left-primitive elements listed in (3)(a) is right-primitive. Indeed, let m0 denote the minimum of the mvalues in words with nonzero coefficients; then m0 > 1 by the above analysis. Now right-multiply by Tm0 −1 · · · T1 . This kills all elements with m-value > m0 + 1, since Tm0 −1 commutes with Tn Tn−1 · · · Tm , hence can be taken past them to multiply against Tw◦0 and be killed. The terms with m-value equal to m0 are not killed, by the analysis in Theorem B. It follows that such a linear combination is not rightprimitive, which completes the classification of the primitive elements in (3)(b). Next, that Prim(N C(M )) is fixed by θ was shown in equation (5.4). Moreover, if N C(M ) equals N CA (n, d) or kW (M ) with W (M ) finite, then it is equipped with a suitable length function `. Now θ preserves the length because the algebra relations are `-homogeneous and preserved by θ. Remark 5.5. In light of Proposition 5.3, it is natural to ask how to write rightprimitive words in standard form. More generally, given w = w0 sn−1 · · · sm0 for unique w0 ∈ Sn−1 and m0 ∈ [1, n] (via Lemma 4.3), we have: Tm · · · Tn−1 Tnk Tw = e = sm · · · sn−1 w0 . Twe Tnk Tn−1 · · · Tm0 , where w With Proposition 5.3 in hand, we turn to the Frobenius property of N CA (n, d). The following proof reveals that N CA (n, d) is Frobenius if and only if Prim(N C(M )) is one-dimensional. Proof of Theorem 5.2. For finite Coxeter groups W (M ), the corresponding nilCoxeter algebras N C(M ) are indeed Frobenius; see e.g. [Kho, §2.2]. It is also easy to verify that N CA (1, d) = k[T1 ]/(T1d ) is Frobenius, by using the symmetric bilinear form uniquely specified by: σ(T1i , T1j ) = 1(i + j = d − 1). Thus, it remains to show that for n > 2 and d > 3, the algebra N CA (n, d) is not Frobenius. Indeed, if N CA (n, d) is Frobenius with nondegenerate invariant bilinear form σ, then for each nonzero primitive p there exists a vector ap such that 0 6= σ(p, ap ) = σ(pap , 1). It follows that we may take ap = 1 for all p. Now the linear functional σ(−, 1) : Prim(N CA (n, d)) → k is nonsingular, whence dimk Prim(N CA (n, d)) = 1. Thus n = 1 or d = 2 by Proposition 5.3. 2990 APOORVA KHARE We conclude this section by discussing the connection of N CA (n, d) to the categorification by Khovanov [Kho] of the Weyl algebra Wn := Zhx, ∂i/(∂x = 1 + x∂). Namely, the usual type A nil-Coxeter algebra An := N CA (n, 2) is a bimodule over An−1 , and this structure was studied in loc. cit., leading to the construction of tensor functors categorifying the operators x, ∂. We now explain how the algebra N CA (n, d) fits into this framework. Proposition 5.6. For all n > 1 and d > 2, we have an isomorphism of An−1 bimodules: d−1 M N CA (n, d) ' An−1 ⊕ An−1 ⊗An−2 An−1 . k=1 When d = 2, this result was shown in [Kho, Proposition 5]. For general d > 2, using the notation of [Kho], this result implies in the category of An−1 -bimodules that the algebra N CA (n, d) corresponds to 1 + (d − 1)x∂ (including the previously known case of d = 2). In particular, Proposition 5.6 strengthens Theorems B and C, which explained a left An−1 -module structure on N CA (n, d) (namely, that N CA (n, d) is free of rank 1 + n(d − 1)). Proof of Proposition 5.6. From the proof of Theorem B, the algebra N CA (n, d) ∼ has a ‘regular representation’ ϕ : M −→ N CA (n, d), sending B(w) 7→ Tw and k B(w, k, m) 7→ Tw Tn Tn−1 · · · Tm for w ∈ Sn , k ∈ [1, d − 1], and m ∈ [1, n]. Also recall the subspaces Vk,m defined in the discussion following equation (4.7): Vk,m = L w∈Sn kB(w, k, m). Ln By Theorem B, Mk := m=1 ϕ(Vk,m ) is a free left An−1 -module of rank one. It is also a free right An−1 -module of rank one, using the anti-involution θ from Proposition 5.3 and Remark 5.5. In fact, the uniqueness of the standard form (4.7) shown in the proof of Theorem B, implies that for all 1 6 k 6 d − 1, the map ϕk : An−1 ⊗An−2 An−1 → Mk , a ⊗ a0 7→ aTnk a0 , is an isomorphism of An−1 -bimodules. Now the result follows from (the proof of) Theorem B. Remark 5.7. Notice that the proof of Proposition 5.6 also categorifies Corollary 5.1. 6. Proof of Theorem D: Finite-dimensional generalized nil-Coxeter algebras We now prove Theorem D, which classifies the generalized nil-Coxeter algebras of finite k-rank. The bulk of the proof involves showing (1) =⇒ (2). We again employ the diagrammatic calculus used to show Theorem B, now applied to the five diagrams in Figure 3 below. We begin by assuming that W = W (M ) is a generalized Coxeter group, and classify the algebras N C(M ) that have finite k-rank. Following this classification, we address the remaining finite complex reflection groups W (and all d), followed by the infinite discrete complex reflection groups with their Coxeter-type presentations. Case 1. Suppose mii = 2 for all i ∈ I. In this case W (M ) is a Coxeter group, so by e.g. [Hum, Chapter 7], N C(M ) has a k-basis in bijection with W (M ), which must therefore be finite. GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2991 α B1 β1 B2 ··· Bm 0 βm0 Bm γ A + C α α A γ B10 B + β1 ··· 0 Bm 0 βm 0 0 Bm Fig. 3.2 (m0 = m − 1) γ α B20 γ D C Fig. 3.1 α B1 β1 B2 ··· Bm 0 βm 0 D C δ α B10 s t β1 B + δ γ A + A Bm B20 ··· 0 Bm 0 βm 0 γ 0 Bm Fig. 3.4 (m0 = m − 1) u C Fig. 3.3 α B1 β1 B2 ··· Bm 0 βm0 Bm γ A + α B10 β1 B20 ··· 0 Bm 0 βm0 0 Bm Fig. 3.5 (m0 = m − 1) Figure 3. Modules for the infinite-dimensional generalized nilCoxeter algebras Case 2. Suppose mαα , mγγ > 3 for some α, γ ∈ I with mαγ > 3. In this case we appeal to Figure 3.1 and work as in the proof of [Mar, Proposition 3.2]. Thus, fix a free k-module M with basis given by the countable set {Ar , Br , Cr , Dr : r > 1}, and define an N C(M )-action via Figure 3.1. Namely, Ti kills all basis vectors for all i ∈ I, with the following exceptions: Tα (Ar ) := Br , Tγ (Br ) := Cr , Tγ (Cr ) := Dr , Tα (Dr ) := Ar+1 , ∀r > 1. 2992 APOORVA KHARE (The “+” at the head of an arrow refers precisely to the index increasing by 1.) It is easy to verify that the defining relations of N C(M ) hold in Endk (M ), as they hold on each Ar , Br , Cr , Dr . Therefore M is a module over N C(M ) that is generated by A1 , but is not finitely generated as a k-module. As N C(M ) M , N C(M ) is also not a finitely generated k-module. This approach is used in the remainder of the proof, to obtain a k-basis and the N C(M )-action on it, from the diagrams in Figure 3. Thus we only mention the figure corresponding to each of the cases below. Case 3. Figure 3.1 is actually a special case of Figure 3.2, and was included to demonstrate a simpler case. Now suppose more generally that there are two nodes α, γ ∈ I such that mαα , mγγ > 3. Since the Coxeter graph is connected, there exist nodes β1 , . . . , βm−1 for some m > 1 (in the figure we write m0 := m − 1), such that α ←→ β1 ←→ ··· ←→ βm−1 ←→ γ is a path (so each successive pair of nodes is connected by at least a single edge). Now appeal to Figure 3.2, i.e., define an N C(M )-module structure on the free k-module 0 0 M := spank {Ar , B1r , . . . , Bmr , Cr , B1r , . . . , Bmr : r > 1}, where each Ti kills all basis vectors above, except for the actions obtained from Figure 3.2. Once again, M is generated by A1 , so proceed as above to show that N C(M ) is not finitely generated. Case 4. The previous cases reduce the situation to a unique vertex α in the Coxeter graph of M for which mαα > 3. The next two steps show that α is adjacent to a unique node γ, and that mαγ = 3. First suppose α is adjacent to γ with mαγ > 4. Now appeal to Figure 3.3, setting (s, t, u) (α, α, γ), and define an N C(M )module structure on M := spank {Ar , Br , Cr : r > 1}. Then proceed as above. Case 5. Next suppose α is adjacent in the Coxeter graph to two nodes γ, δ. By the previous case, mαγ = mαδ = 3. Now appeal to Figure 3.4 with m = 1, to define 0 an N C(M )-module structure on M := spank {Ar , B1r , B1r , Cr , Dr : r > 1}, and proceed as in the previous cases. We now observe that if N C(M ) is finitely generated, then so is N CM ((2, . . . , 2)), which corresponds to the Coxeter group WM ((2, . . . , 2)). Hence the Coxeter graph of M is of finite type. These graphs were classified by Coxeter [Cox1, Cox2]. We now rule out all cases other than type A, in which case the above analysis shows that d = (2, . . . , 2, d) or (d, 2, . . . , 2). Case 6. First notice that dihedral types (i.e., types G2 , H2 , I) are already ruled out by the above cases. The same cases also rule out one possibility in types B, C, H, where we may now set n > 3. For the remaining cases of types B, C, H, assume that the Coxeter graph is labelled α ←→ β1 ←→ ··· ←→ βm−1 ←→ γ, with mαα > 3, mγγ = 2, mβm−1 γ > 4. In this case we construct the N C(M )-module M by appealing to Figure 3.5; now proceed as above. Case 7. The next case is of type Dn , with n > 4. Notice that α is an extremal (i.e., pendant) vertex by the above analysis. First assume α is the extremal node GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2993 on the “long arm” of the Coxeter graph. Now appeal to Figure 3.4 with m = n − 2, to construct an N C(M )-module M . The other sub-case is when α is one of the other two extremal nodes in the Dn -graph. Define the quotient algebra N C 0 (M ) whose Coxeter graph is of type D4 (i.e., where we kill the n − 4 generators Ti in the long arm that are the furthest away from α). Now repeat the construction in the previous paragraph, using Figure 3.4 with m = 2. It is easy to verify that the space M is a module for N C 0 (M ), hence for the algebra N CD4 ((2, 2, 2, mαα )). This allows us to proceed as in the previous sub-case and show that N C 0 (M ) is not finitely generated, whence neither is N C(M ). Case 8. If the Coxeter graph is of type E, then we may reduce to the Dn -case by the analysis in the previous case. Hence it follows using Figure 3.4 that N C(M ) is not finitely generated. Case 9. If the Coxeter graph is of type F4 , then we may reduce to the Bn -case by the analysis in Case 7. It now follows from Case 6 that N C(M ) is not finitely generated. This completes the classification for generalized Coxeter groups W (M ). We now appeal to the classification and presentation of all finite complex reflection groups, whose Coxeter graph is connected. These groups and their presentations are listed in [BMR2, Tables 1–4]. In what follows, we adopt the following notation: if W = Gm for 4 6 m 6 37, then the corresponding generalized nil-Coxeter algebras will be denoted by N Cm (d). Similarly if W = G(de, e, r), then we work with N C(de,e,r) (d). In what follows, we will often claim that N CW (d) is not finitely generated (over k), omitting the phrase “unless it is the usual nil-Coxeter algebra over a finite Coxeter group”. Case 10 (Exceptional types with finite Coxeter graph). If W = Gm for m = 4, 8, 16, 25, 32, then its Coxeter graph is of type A. This case has been addressed above; thus the only possibility that N CW (d) has finite rank is that it equals N CAn ((2, . . . , 2, d)) for d > 2, as desired. Next if W = Gm for m = 5, 10, 18, 26, then its Coxeter graph is of type B, which was also addressed above and never yields an algebra of finite k-rank. Now suppose W = G29 . Then s, t, u form a sub-diagram of type B3 , whence the quotient algebra 0 N C29 generated by Ts , Tt , Tu is not finitely generated, by arguments as in Case 7 above. It follows that N C29 (d) is also not finitely generated. The next case is if W = Gm for m = 6, 9, 14, 17, 20, 21. In this case the Coxeter graph is of dihedral type, which was also addressed above. Case 11 (All other exceptional types). For the remaining exceptional values of m ∈ [4, 37], with W not a finite Coxeter group, we will appeal to Figure 3.3. There are three cases: first, suppose m = 31. In this case, set (s, t, u) (s, u, t) in Figure 3.3 and define an N Cm (d)-module M that is k-free with basis {Ar , Br , Cr : r > 1}. Now proceed as above. Next if m = 33, 34, then set (s, t, u) (w, t, u) in Figure 3.3 to define an N Cm (d)-module M , and proceed as above. Finally, fix any other m, i.e., m = 7, 11, 12, 13, 15, 19, 22, 24, 27. In this case, use Figure 3.3 to define an N Cm (d)-module M , and proceed as above. 2994 APOORVA KHARE Case 12 (The infinite families). It remains to consider the six infinite families enumerated in [BMR2], which make up the family G(de, e, r). Three of the families consist of finite Coxeter groups of types A, B, I, which were considered above. We now consider the other three families. (a) Suppose W = G(de, e, r) with e > 3. Then by [BMR2, Table 1], consider 0 the quotient algebra N C(de,e,r) of N C(de,e,r) (d) which is generated by s, t := 0 0 t2 , u := t2 , by killing all other generators Ti . The generators of N C(de,e,r) now satisfy the relations Tsds = Ttdt = Tudu = 0, Ts Tt Tu = Tt Tu Ts , T T T T · · · = T T Tu Ts · · · . } \| u s {zt u } \| s t {z (e+1) times (e+1) times 0 Thus, use Figure 3.3 to define an N C(de,e,r) -module structure on M , and proceed as above to show that N C(de,e,r) (d) is not finitely generated. (b) Suppose W = G(2d, 2, r) with d > 2; see [BMR2, Table 2]. Apply a similar argument as in the previous sub-case, using the same generators and the same figure. (c) Suppose W = G(e, e, r) with e > 2 and r > 2. If e = 2, then G(2, 2, r) is a finite Coxeter group, hence was addressed above. Next, if r > 3 then killing Ts reduces to (a quotient of) the Dn -case, which was once again addressed above. Finally, suppose r = 3 6 e. Setting s := t3 , t := t02 , u := t2 , the generators of N C(e,e,3) satisfy Tsds = Ttdt = Tudu = 0, 2 2 (Ts Tt Tu ) = (Tt Tu Ts ) , Ts Tt Ts = Tt Ts Tt , Ts Tu Ts = Tu Ts Tu , T T Tu Tt · · · = Tt Tu Tt Tu · · · . \| u t {z } \| {z } e times e times Once again, use Figure 3.3 to define an N C(e,e,3) -module structure on M , and proceed as above. This completes the proof of (1) =⇒ (2) for finite complex reflection groups. Next, by e.g. [Hum, Chapter 7], for no infinite Coxeter group W is N CW ((2, . . . , 2)) a finitely generated k-module, whence the same result holds for N CW (d) when all di > 2. We now use the classification of the (remaining) infinite complex reflection groups W associated to a connected braid diagram. These groups were described in [Pop1] and subsequently in [Mal]. Thus, there exists a complex affine space E with group of translations V ; choosing a basepoint v0 ∈ E, we can identify the semidirect product GL(V ) n V with the group A(E) of affine transformations of E. Moreover, W ⊂ A(E). Define Lin(W ) to be the image of W in the factor group GL(V ), and Tran(W ) to be the subset of W in V , i.e., (6.1) Tran(W ) := W ∩ V, Lin(W ) := W/ Tran(W ). It remains to consider three cases for irreducible infinite complex reflection groups W. Case 13. The group W is noncrystallographic, i.e., E/W is not compact. Then by [Pop1, Theorem 2.2], there exists a real form ER ⊂ E whose complexification is E, i.e., ER ⊗R C = E. Moreover, by the same theorem, restricting the elements of W to ER yields an affine Weyl group WR . Hence if N CW (d) is a finitely generated k-module, then so is N CWR ((2, . . . , 2)), which is impossible. GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2995 Case 14. The group W is a genuine crystallographic group, i.e., E/W is compact and Lin(W ) is not the complexification of a real reflection group. Such groups were studied by Malle in [Mal], and Coxeter-type presentations for these groups were provided in Tables I, II in loc. cit. Specifically, Malle showed that these groups are quotients of a free monoid by a set of braid relations and order relations, together with one additional order relation R0m0 = 1. We now show that for none of these groups W is the algebra N CW (d) (defined in Definition 1.11) a finitely generated k-module. To do so, we proceed as above, by specifying the sub-figure in Figure 3 that corresponds to each of these groups. There are three sub-cases: (1) Suppose W is the group [G(3, 1, 1)] in [Mal, Table I], or [Km ] in [Mal, Table II] for m = 4, 8, 25, 32. For these groups we appeal to Figure 3.1 and proceed as in Case 2 above. (2) For W = [K33 ], [K34 ], notice that it suffices to show the claim that given f3 Coxeter graph (i.e., a 4-cycle) with nodes labelled α1 , . . . , α4 in the A clockwise fashion, the corresponding algebra N CA f3 (d) is not a finitely generated k-module. For this we construct a module M , using Figure 4 below with n = 4. Now proceeding as above shows the claim, and hence the result for [K33 ], [K34 ]. Bn αn B1 + α1 .. . ··· B2 Figure 4. Module M for N CAfn (d) (3) For the remaining cases in [Mal, Tables I, II], we appeal to Figure 3.3 as in Case 11 above, with three suitably chosen generators in each case. Case 15. Finally, we consider the remaining “nongenuine, crystallographic” cases as in [Pop1, Table 2]. Thus, E/W is compact and Lin(W ) is the complexification of a real reflection group. In these cases, verify by inspection from [Pop1, Table 2] that the cocycle c is always trivial. Thus W = Lin(W )nTran(W ), with W 0 := Lin(W ) a finite Weyl group, and Tran(W ) a lattice of rank 2\|I 0 \|, where I 0 indexes the simple reflections in the Weyl group W 0 . We now claim that the corresponding family of generalized nil-Coxeter algebras N CW (d) are not finitely generated as k-modules. To show the claim requires a presentation of W in terms of generating reflections. The following recipe for such a presentation was communicated to us by Popov [Pop2]. Notice from e.g. [Pop1, Table 2] that Tran(W ) is a direct sum of two Lin(W )-stable lattices Λ1 and Λ2 = αΛ1 (with α 6∈ R), each of rank \|I 0 \|. Thus, Λ1 ∼ = Λ2 as ZW 0 -modules, 0 with W = Lin(W ) a finite real reflection group as above. Moreover, for j = 1, 2, the semidirect product Sj := W 0 n Λj is a real crystallographic reflection group whose fundamental domain is a simplex; this yields a presentation of Sj via \|I 0 \| + 1 generating reflections in the codimension-one faces of this simplex. One now combines these presentations for S1 , S2 to obtain a system of \|I 0 \| + 2 generators for 2996 APOORVA KHARE W ; see in this context the remarks following [Mal, Theorem 3.1]. In this setting, it follows by [Pop1, Theorem 4.5] that each Sj is isomorphic, as a real reflection group, f0 over W 0 , since the Coxeter type of Sj is determined to the affine Weyl group W by the Coxeter types of W 0 and Λj . Thus W is in some sense a ‘double affine Weyl group’. (For simply-laced W 0 , it is also easy to verify by inspection from [Pop1, Table 2] that Λj is isomorphic as a ZW 0 -module to the root lattice for W 0 , whence f0 for j = 1, 2.) Sj ∼ =W Equipped with this presentation of W from [Pop2], we analyze N CW (d) as follows. Fix a ZW 0 -module isomorphism ϕ : Λ1 → Λ2 , and choose affine reflections s0j ∈ Sj , corresponding to µ1 and µ2 = ϕ(µ1 ) respectively, which together with f0 . Then W W f0 upon the simple reflections {si : i ∈ I 0 } ⊂ W generate Sj ∼ =W quotienting by the relation: s01 = s02 . Using the presentation of N CW (d) via the corresponding \|I 0 \| + 2 generators {Ti : i ∈ I 0 } t {T01 , T02 }, N CW (d) N CW ((2, . . . , 2)) N CW ((2, . . . , 2))/(T01 −T02 ) ∼ = N C f0 ((2, . . . , 2)), W and this last term is an affine Weyl nil-Coxeter algebra, hence is not finitely generated as a k-module. Therefore neither is N CW (d), as desired. This shows that (1) =⇒ (2); the converse follows by [Hum, Chapter 7] and Theorem B. We now show that (2) and (3) are equivalent. Note from the above case-bycase analysis that if N CW (d) is not finitely generated, then either it surjects onto an affine Weyl nil-Coxeter algebra N CW f0 ((2, . . . , 2)), or one can define a module M as above, and for each r > 1 there exists a word Twr ∈ N CW (d), expressed using O(r) generators, which sends the k-basis vector A1 ∈ M to Ar . It follows in both cases that m is not nilpotent. Next, if W = W (M ) is a finite Coxeter group, then it is well-known (see e.g. [Hum, Chapter 7]) that m is nilpotent. Finally, if N C(M ) = N CA (n, d), then m is nilpotent by Theorem C. This shows (2) ⇐⇒ (3). The final statement on the length function ` and the longest element also follows from [Hum, Chapter 7] and Theorem C. Remark 6.2. If M is a generalized Coxeter matrix with some mij = ∞, then we can similarly work with M the k-span of {ar , br : r > 1}, where Ti : ar 7→ br , br 7→ 0, Tj : br 7→ ar+1 , ar 7→ 0, and all other Tk kill M . It follows that N C(M ) once again has infinite k-rank. Acknowledgments The author is grateful to Ivan Marin, Vladimir Popov, and Victor Reiner for informative and stimulating correspondences; and to James Humphreys for carefully going through an earlier draft and providing valuable feedback. The author also thanks Daniel Bump, Gunter Malle, Eric Rowell, Travis Scrimshaw, and Bruce Westbury for useful references and discussions. The author is partially supported by a Young Investigator Award from the Infosys Foundation. References [AnSc] [As] Nicolás Andruskiewitsch and Hans-Jürgen Schneider, On the classification of finitedimensional pointed Hopf algebras, Annals of Mathematics 171(1):375–417, 2010. Joachim Assion, A proof of a theorem of Coxeter, C. R. Math. Rep. Acad. Sci. Canada 1(1):41–44, 1978/79. GENERALIZED NIL-COXETER ALGEBRAS OVER COMPLEX REFLECTION GROUPS2997 [Ba] Tathagata Basak, On Coxeter diagrams of complex reflection groups, Transactions of the American Mathematical Society 364(9):4909–4936, 2012. [BSS] Chris Berg, Franco Saliola, and Luis Serrano, Pieri operators on the affine nilCoxeter algebra, Transactions of the American Mathematical Society 366(1):531–546, 2014. [BGG] Joseph N. Bernstein, Israel M. Gelfand, and Sergei I. Gelfand, Schubert cells and cohomology of the spaces G/P , Russian Mathematical Surveys 28(2):1–26, 1973. [BS] Joseph N. Bernstein and Ossip Schwarzman, Complex crystallographic Coxeter groups and affine root systems, Journal of Nonlinear Mathematical Physics 13(2):163–182, 2006. [Be] David Bessis, Zariski theorems and diagrams for braid groups, Inventiones mathematicae 145(3):487–507, 2001. [BMR1] Michel Broué, Gunter Malle, and Raphaël Rouquier, On complex reflection groups and their associated braid groups, Representations of groups (Banff, 1994), CMS Conference Proceedings 16:1–13, American Mathematical Society, Providence, 1995. [BMR2] Michel Broué, Gunter Malle, and Raphaël Rouquier, Complex reflection groups, braid groups, Hecke algebras, Journal für die reine und angewandte Mathematik 500:127–190, 1998. [Ca] Roger W. Carter, Representation theory of the 0-Hecke algebra, Journal of Algebra 104(1):89–103, 1986. [Ch] Eirini Chavli, The Broué–Malle–Rouquier conjecture for the exceptional groups of rank 2, Ph.D. thesis, 2016. [Coh] Arjeh M. Cohen, Finite complex reflection groups, Annales scientifiques de l’École Normale Supérieure, Sér. 4, 9(3):379–436, 1976. [Cox1] Harold S.M. Coxeter, Discrete groups generated by reflections, Annals of Mathematics 35(3):588–621, 1934. [Cox2] Harold S.M. Coxeter, The complete enumeration of finite groups of the form Ri2 = (Ri Rj )kij = 1, Journal of the London Mathematical Society s1-10(1):21–25, 1935. [Cox3] Harold S.M. Coxeter, Factor groups of the braid group, Proceedings of the 4th Canadian Mathematical Congress (Banff, 1957), University of Toronto Press, 95–122, 1959. [Dr] Vladimir G. Drinfeld, Degenerate affine Hecke algebras and Yangians, Functional Analysis and its Applications 20(1):58–60, 1986. [Et] Pavel Etingof, Proof of the Broué–Malle–Rouquier conjecture in characteristic zero (after I. Losev and I. Marin–G. Pfeiffer), Arnold Mathematical Journal 3(3):445–449, 2017. [EG] Pavel Etingof and Victor Ginzburg, Symplectic reflection algebras, Calogero-Moser space, and deformed Harish-Chandra homomorphism, Inventiones mathematicae 147(2):243– 348, 2002. [ES] Pavel Etingof and Olivier Schiffmann, Lectures on Quantum Groups, Lectures in Mathematical Physics, International Press, Somerville, MA, 1998. [FS] Sergey Fomin and Richard Stanley, Schubert polynomials and the nil-Coxeter algebra, Advances in Mathematics 103(2):196–207, 1994. [GHV] Matı́as Graña, István Heckenberger, and Leandro Vendramin, Nichols algebras of group type with many quadratic relations, Advances in Mathematics 227(5):1956–1989, 2011. [HV1] István Heckenberger and Leandro Vendramin, A classification of Nichols algebras of semi-simple Yetter-Drinfeld modules over non-abelian groups, Journal of the European Mathematical Society 19(2):299–356, 2017. [HV2] István Heckenberger and Leandro Vendramin, The classification of Nichols algebras over groups with finite root system of rank two, Journal of the European Mathematical Society 19(7):1977–2017, 2017. [Hug1] Mervyn C. Hughes, Complex reflection groups, Communications in Algebra 18(12):3999– 4029, 1990. [Hug2] Mervyn C. Hughes, Extended root graphs for complex reflection groups, Communications in Algebra 27(1):119–148, 1999. [Hum] James E. Humphreys, Reflection groups and Coxeter groups, Cambridge studies in advanced mathematics no. 29, Cambridge University Press, Cambridge-New YorkMelbourne, 1990. [Kha] Apoorva Khare, Generalized nil-Coxeter algebras, cocommutative algebras, and the PBW property, AMS Contemporary Mathematics 688:139–168, 2017. [Kho] Mikhail Khovanov, NilCoxeter algebras categorify the Weyl algebra, Communications in Algebra 29(11):5033–5052, 2001. 2998 [KL] [KK] [Ko] [LS] [LT] [Lo] [Lu] [Mal] [Mar] [MP] [No] [ORS] [Pop1] [Pop2] [ReS] [ST] [SW1] [SW2] APOORVA KHARE Mikhail Khovanov and Aaron D. Lauda, A diagrammatic approach to categorification of quantum groups I, Representation Theory 13:309–347, 2009. Bertram Kostant and Shrawan Kumar, The nil Hecke ring and cohomology of G/P for a Kac-Moody group G, Advances in Mathematics 62(3):187–237, 1986. David W. Koster, Complex reflection groups, Ph.D. thesis, University of Wisconsin, 1975. Alain Lascoux and Marcel-Paul Schützenberger, Fonctorialité des polynômes de Schubert, Invariant Theory (R. Fossum, W. Haboush, M. Hochster and V. Lakshmibai, Eds.), Contemporary Mathematics Volume 88, pages 585–598, American Mathematical Society, Providence, 1989. Gustav I. Lehrer and Donld E. Taylor, Unitary reflection groups, Australian Mathematical Society Lecture Series Volume 20, Cambridge University Press, Cambridge, 2009. Ivan Losev, Finite-dimensional quotients of Hecke algebras, Algebra & Number Theory 9(2):493–502, 2015. George Lusztig, Affine Hecke algebras and their graded version, Journal of the American Mathematical Society 2:599–635, 1989. Gunter Malle, Presentations for crystallographic complex reflection groups, Transformation Groups 1(3):259–277, 1996. Ivan Marin, The freeness conjecture for Hecke algebras of complex reflection groups, and the case of the Hessian group G26 , Journal of Pure and Applied Algebra 218(4):704–720, 2014. Ivan Marin and Götz Pfeiffer, The BMR freeness conjecture for the 2-reflection groups, Mathematics of Computation, published online, 2016. P.N. Norton, 0-Hecke algebras, Journal of the Australian Mathematical Society A, 27(3):337–357, 1979. Peter Orlik, Victor Reiner, and Anne V. Shepler, The sign representation for Shephard groups, Mathematische Annalen 322(3):477–492, 2002. Vladimir L. Popov, Discrete complex reflection groups, Communications of the Mathematical Institute, vol. 15, Rijksuniversiteit Utrecht, 89 pp., 1982. Vladimir L. Popov, personal communication, Feb. 2016. Victor Reiner and Anne V. Shepler, Invariant derivations and differential forms for reflection groups, preprint, arXiv:1612.01031. Geoffrey C. Shephard and John A. Todd, Finite unitary reflection groups, Canadian Journal of Mathematics 6:274–304, 1954. Anne V. Shepler and Sarah Witherspoon, A Poincaré–Birkhoff–Witt theorem for quadratic algebras with group actions, Transactions of the American Mathematical Society 366(12):6483–6506, 2014. Anne V. Shepler and Sarah Witherspoon, Poincaré–Birkhoff–Witt Theorems, Commutative Algebra and Noncommutative Algebraic Geometry, Volume I: Expository Articles (D. Eisenbud, S.B. Iyengar, A.K. Singh, J.T. Stafford, and M. Van den Bergh, Eds.), Mathematical Sciences Research Institute Proceedings Volume 67, pages 259–290, Cambridge University Press, Cambridge, 2015. Department of Mathematics, Indian Institute of Science, Bangalore – 560012, India E-mail address: khare@iisc.ac.in	4math.GR
Feudal Reinforcement Learning for Dialogue Management in Large Domains Iñigo Casanueva1 , Paweł Budzianowski1 , Pei-Hao Su2 , Stefan Ultes1 , Lina Rojas-Barahona1 , Bo-Hsiang Tseng1 and Milica Gašić1 1 Department of Engineering, University of Cambridge, UK 2 PolyAI Limited, London, UK ic340@cam.ac.uk arXiv:1803.03232v1 [cs.CL] 8 Mar 2018 Abstract Reinforcement learning (RL) is a promising approach to solve dialogue policy optimisation. Traditional RL algorithms, however, fail to scale to large domains due to the curse of dimensionality. We propose a novel Dialogue Management architecture, based on Feudal RL, which decomposes the decision into two steps; a first step where a master policy selects a subset of primitive actions, and a second step where a primitive action is chosen from the selected subset. The structural information included in the domain ontology is used to abstract the dialogue state space, taking the decisions at each step using different parts of the abstracted state. This, combined with an information sharing mechanism between slots, increases the scalability to large domains. We show that an implementation of this approach, based on Deep-Q Networks, significantly outperforms previous state of the art in several dialogue domains and environments, without the need of any additional reward signal. 1 Introduction Task-oriented Spoken Dialogue Systems (SDS), in the form of personal assistants, have recently gained much attention in both academia and industry. One of the most important modules of a SDS is the Dialogue Manager (DM) (or policy), the module in charge of deciding the next action in each dialogue turn. Reinforcement Learning (RL) (Sutton and Barto, 1999) has been studied for several years as a promising approach to model dialogue management (Levin et al., 1998; Henderson et al., 2008; Pietquin et al., 2011; Young et al., 2013; Casanueva et al., 2015; Su et al., 2016). However, as the dialogue state space increases, the number of possible trajectories needed to be explored grows exponentially, making traditional RL methods not scalable to large domains. Hierarchical RL (HRL), in the form of temporal abstraction, has been proposed in order to mitigate this problem (Cuayáhuitl et al., 2010, 2016; Budzianowski et al., 2017; Peng et al., 2017). However, proposed HRL methods require that the task is defined in a hierarchical structure, which is usually handcrafted. In addition, they usually require additional rewards for each subtask. Space abstraction, instead, has been successfully applied to dialogue tasks such as Dialogue State Tracking (DST) (Henderson et al., 2014b), and policy transfer between domains (Gašić et al., 2013, 2015; Wang et al., 2015). For DST, a set of binary classifiers can be defined for each slot, with shared parameters, learning a general way to track slots. The policy transfer method presented in (Wang et al., 2015), named Domain Independent Parametrisation (DIP), transforms the belief state into a slot-dependent fixed size representation using a handcrafted feature function. This idea could also be applied to large domains, since it can be used to learn a general way to act in any slot. In slot-filling dialogues, a HRL method that relies on space abstraction, such as Feudal RL (FRL) (Dayan and Hinton, 1993), should allow RL scale to domains with a large number of slots. FRL divides a task spatially rather than temporally, decomposing the decisions in several steps and using different abstraction levels in each sub-decision. This framework is especially useful in RL tasks with large discrete action spaces, making it very attractive for large domain dialogue management. In this paper, we introduce a Feudal Dialogue Policy which decomposes the decision in each turn into two steps. In a first step, the policy decides if it takes a slot independent or slot dependent action. Then, the state of each slot sub-policy is abstracted to account for features related to that slot, and a primitive action is chosen from the previously selected subset. Our model does not require any modification of the reward function and the hierarchical architecture is fully specified by the structured database representation of the system (i.e. the ontology), requiring no additional design. 2 Main belief state (𝑏) 𝜙1 (𝑏) 𝜙* (𝑏) 𝑄,10 Qπ (b, a) = E{R\|bt = b, at = a} (1) PT −1 (τ −t) rτ and γ is a discount where R = τ =t γ factor, with 0 ≤ γ ≤ 1. The objective of RL is to find an optimal policy π ∗ , i.e. a policy that maximizes the expected return in each belief state. In Value-based algorithms, the optimal policy can be found by greedily taking the action which maximises Qπ (b, a). In slot-filling SDSs the belief state space B is defined by the ontology, a structured representation of a database of entities that the user can retrieve by talking to the system. Each entity has a set of properties, refereed to as slots S, where each of the slots can take a value from the set Vs . The belief state b is then defined as the concatenation of the probability distribution of each slot, plus a set of general features (e.g. the communication function used by the user, the database search method...) (Henderson et al., 2014a). The set A is defined as a set of summary actions, where the actions can be either slot dependent (e.g. request(food), confirm(area)...) or slot independent1 (e.g. hello(), inform()...). The belief space B is defined by the ontology, therefore belief states of different domains will have different shapes. In order to transfer knowledge between domains, Domain Independent Parametrization (DIP) (Wang et al., 2015) proposes to abstract the belief state b into a fixed 1 We include the summary actions dependent on all the slots, such as inform(), in this group. - 𝜋* Argmax over master actions 𝑄,* - … 𝑄,* ( ) Argmax over slot independent primitives 𝑄,10 / 𝑏") 𝜋2 " … 𝜋"# 𝜋"( " … 𝜙") (𝑏) 𝑏"# 𝑏"( 𝜋1 𝑏* Background Dialogue management can be cast as a continuous MDP (Young et al., 2013) composed of a continuous multivariate belief state space B, a finite set of actions A and a reward function R(bt , at ). At a given time t, the agent observes the belief state bt ∈ B, executes an action at ∈ A and receives a reward rt ∈ R drawn from R(bt , at ). The action taken, a, is decided by the policy, defined as the function π(b) = a. For any policy π and b ∈ B, the Q-value function can be defined as the expected (discounted) return R, starting from state b, taking action a, and then following policy π until the end of the dialogue at time step T : 𝜙"( (𝑏) 𝜙"# (𝑏) 𝑏1 " 𝜋") " " " 𝑄,(/ … 𝑄,(/ 𝑄,#/ … 𝑄,#/ … 𝑄,)/ … 𝑄,)/ ( ) ( ) ( ) Argmax over slots and slot dependent primitives Figure 1: Feudal dialogue architecture used in this work. The sub-policies surrounded by the dashed line have shared parameters. The simple lines show the data flow and the double lines the sub-policy decisions. size representation. As each action is either slot independent or dependent on a slot s, a feature function φdip (b, s) can be defined, where s ∈ S∪si and si stands for slot independent actions. Therefore, in order to compute the policy, Q(b, a) can be approximated as Q(φdip (b, s), a), where s is the slot associated to action a. Wang et al. (2015) presents a handcrafted feature function φdip (b, s). It includes the slot independent features of the belief state, a summarised representation of the joint belief state, and a summarised representation of the belief state of the slot s. Section 4 gives a more detailed description of the φdip (b, s) function used in this work. 3 Feudal dialogue management FRL decomposes the policy decision π(b) = a in each turn into several sub-decisions, using different abstracted parts of the belief state in each subdecision. The objective of a task oriented SDS is to fulfill the users goal, but as the goal is not observable for the SDS, the SDS needs to gather enough information to correctly fulfill it. Therefore, in each turn, the DM can decompose its decision in two steps: first, decide between taking an action in order to gather information about the user goal (information gathering actions) or taking an action to fulfill the user goal or a part of it (information providing actions) and second, select a (primitive) action to execute from the previously selected subset. In a slot-filling dialogue, the set of information gathering actions can be defined as the set of slot dependent actions, while the set of information providing actions can be defined as the remaining actions. The architecture of the feudal policy proposed by this work is represented schematically in Figure 1. The (primitive) actions are divided between two subsets; slot independent actions Ai (e.g. hello(), inform()); and slot dependent actions Ad (e.g. request(), confirm())2 . In addition, a set of master m m actions Am = (am i , ad ) is defined, where ai corresponds to taking an action from Ai and am d to taking an action from Ad . Then, a feature function φs (b) = bs is defined for each slot s ∈ S, as well as a slot independent feature function φi (b) = bi and a master feature function φm (b) = bm . These feature functions can be handcrafted (e.g. the DIP feature function introduced in section 2) or any function approximator can be used (e.g. neural networks trained jointly with the policy). Finally, a master policy πm (bm ) = am , a slot independent policy πi (bi ) = ai and a set of slot specific policies πs (bs ) = ad , one for each s ∈ S, are defined, where am ∈ Am , ai ∈ Ai and ad ∈ Ad . Contrary to other feudal policies, the slot specific sub-policies have shared parameters, in order to generalise between slots (following the idea used by Henderson et al. (2014b) for DST). The differences between the slots (size, value distribution...) are accounted by the feature function φs (b). Therefore πm (bm ) is defined as: πm (bm ) = argmax Qm (bm , am ) (2) am ∈Am If πm (bm ) = am i , the sub-policy run is πi : πi (bi ) = argmax Qi (bi , ai ) (3) ai ∈Ai Else, if πm (bm ) = am d , πd is selected. This policy runs each slot specific policy, πs , for all s ∈ S, choosing the action-slot pair that maximises the Q function over all the slot sub-policies. πd (bs \|∀s ∈ S) = argmax Qs (bs , ad ) (4) ad ∈Ad ,s∈S Then, the summary action a is constructed by joining ad and s (e.g. if ad =request() and s=food, then the summary action will be request(food)). A pseudo-code of the Feudal Dialogue Policy algorithm is given in Appendix A. Domain Cambridge Restaurants San Francisco Restaurants Laptops SER Masks User Env. 1 0% on Std. Code CR SFR LAP Env. 2 0% off Std. # constraint slots 3 6 11 Env. 3 15% on Std. Env. 4 15% off Std. # requests 9 11 21 # values 268 636 257 Env. 5 15% on Unf. Env. 6 30% on Std. Table 1: Sumarised description of the domains and environments used in the experiments. Refer to (Casanueva et al., 2017) for a detailed description. 4 Experimental setup The models used in the experiments have been implemented using the PyDial toolkit (Ultes et al., 2017)3 and evaluated on the PyDial benchmarking environment (Casanueva et al., 2017). This environment presents a set of tasks which span different size domains, different Semantic Error Rates (SER), and different configurations of action masks and user model parameters (Standard (Std.) or Unfriendly (Unf.)). Table 1 shows a summarised description of the tasks. The models developed in this paper are compared to the state-ofthe-art RL algorithms and to the handcrafted policy presented in the benchmarks. 4.1 DIP-DQN baseline An implementation of DIP based on Deep-Q Networks (DQN) (Mnih et al., 2013) is implemented as an additional baseline (Papangelis and Stylianou, 2017). This policy, named DIP-DQN, uses the same hyperparameters as the DQN implementation released in the PyDial benchmarks. A DIP feature function based in the description in (Wang et al., 2015) is used, φdip (b, s) = ψ0 (b) ⊕ ψj (b) ⊕ ψd (b, s), where: • ψ0 (b) accounts for general features of the belief state, such as the database search method. • ψj (b) accounts for features of the joint belief state, such as the entropy of the joint belief. • ψd (b, s) accounts for features of the marginal distribution of slot s, such as the entropy of s. Appendix B shows a detailed description of the DIP features used in this work. 4.2 Feudal DQN policy A Feudal policy based on the architecture described in sec. 3 is implemented, named FDQN. Each sub-policy is constructed by a DQN policy 2 Note that the actions of this set are composed just by the communication function of the slot dependent actions, thus reducing the number of actions compared to A. 3 The implementation of the models can be obtained in www.pydial.org Env. 1 Env. 2 Env. 3 Env. 4 Env. 5 Env. 6 Task CR SFR LAP CR SFR LAP CR SFR LAP CR SFR LAP CR SFR LAP CR SFR LAP Feudal-DQN Suc. Rew. 89.3% 11.7 71.1% 7.1 65.5% 5.7 97.8% 13.1 95.4% 12.4 94.1% 12.0 92.6% 11.7 90.0% 9.7 89.6% 9.4 91.4% 11.2 90.3% 10.2 88.7% 9.8 96.3% 11.5 88.9% 7.9 78.8% 5.2 90.6% 10.4 83.0% 7.1 78.5% 6.0 DIP-DQN Suc. Rew. 48.8% -2.8 25.8% -7.4 26.6% -8.8 85.5% 9.6 85.7% 8.4 89.5% 9.7 86.1% 8.9 59.3% 0.2 71.5% 3.1 82.6% 8.7 86.1% 9.2 74.8% 6.0 74.4% 2.9 75.5% 3.2 64.4% -0.4 83.4% 8.1 71.9% 3.9 66.5% 2.7 Bnch. Rew. 13.5 11.7 10.5 12.2 9.6 7.3 11.9 8.6 6.7 10.7 7.7 5.5 10.5 4.5 4.1 10.0 3.9 3.6 Hdc. Rew. 14.0 12.4 11.7 14.0 12.4 11.7 11.0 9.0 8.7 11.0 9.0 8.7 9.3 6.0 5.3 9.7 6.4 5.5 Table 2: Success rate and reward for Feudal-DQN and DIP-DQN in the 18 benchmarking tasks, compared with the reward of the best performing algorithm in each task (Bnch.) and the handcrafted policy (Hdc.) presented in (Casanueva et al., 2017). (Su et al., 2017). These policies have the same hyperparameters as the baseline DQN implementation, except for the two hidden layer sizes, which are reduced to 130 and 50 respectively. As feature functions, subsets of the DIP features are used: φm (b) = φi (b) = ψ0 (b) ⊕ ψj (b) φs (b) = ψ0 (b) ⊕ ψj (b) ⊕ ψd (b, s) ∀s ∈ S The original set of summary actions of the benchmarking environment, A, has a size of 5 + 3 ∗ \|S\|, where \|S\| is the number of slots. This set is divided in two subsets4 : Ai of size 6 and Ad of size 4. Each sub-policy (including πm ) is trained with the same sparse reward signal used in the baselines, getting a reward of 20 if the dialogue is successful or 0 otherwise, minus the dialogue length. 5 Results The results in the 18 tasks of the benchmarking environment after 4000 training dialogues are presented in Table 2. The same evaluation procedure of the benchmarks is used, presenting the mean over 10 different random seeds and testing every seed for 500 dialogues. The FDQN policy substantially outperforms every other other policy in all the environments except Env. 1. The Figure 2: Learning curves for Feudal-DQN and DIPDQN in Env. 4, compared to the two best performing algorithms in (Casanueva et al., 2017) (DQN and GPSarsa). The shaded area depicts the mean ± the standard deviation over ten random seeds. performance increase is more considerable in the two largest domains (SFR and LAP), with gains up to 5 points in accumulated reward in the most challenging environments (e.g. Env. 4 LAP), compared to the best benchmarked RL policies (Bnch.). In addition, FDQN consistently outperforms the handcrafted policy (Hdc.) in environments 2 to 6, which traditional RL methods could not achieve. In Env. 1, however, the results for FDQN and DIP-DQN are rather low, specially for DIP-DQN. Surprisingly, the results in Env. 2, which only differs from Env. 1 in the absence of action masks (thus, in principle, is a more complex environment), outperform every other algorithm. Analysing the dialogues individually, we could observe that, in this environment, both policies are prone to “overfit” to an action 5 . The performance of FDQN and DIP-DQN in Env. 4 is also better than in Env. 3, while the difference between these environments also lies in the masks. This suggests that an specific action mask design can be helpful for some algorithms, but can harm the performance of others. This is especially severe in the DIP-DQN case, which shows good performance in some challenging environments, but it is more unstable and prone to overfit than FDQN. However, the main purpose of action masks is to reduce the number of dialogues needed to train a policy. Observing the learning curves shown in Figure 2, the FDQN model can learn a nearoptimal policy in large domains in about 1500 dialogues, even if no additional reward is used, making the action masks unnecessary. 4 An additional pass() action is added to each subset, which is taken whenever the other sub-policy is executed. This simplifies the training algorithm. 5 The model overestimates the value of an incorrect action, continuously repeating it until the user runs out of patience. 6 Conclusions and future work We have presented a novel dialogue management architecture, based on Feudal RL, which substantially outperforms the previous state of the art in several dialogue environments. By defining a set of slot dependent policies with shared parameters, the model is able to learn a general way to act in slots, increasing its scalability to large domains. Unlike other HRL methods applied to dialogue, no additional reward signals are needed and the hierarchical structure can be derived from a flat ontology, substantially reducing the design effort. A promising approach would be to substitute the handcrafted feature functions used in this work by neural feature extractors trained jointly with the policy. This would avoid the need to design the feature functions and could be potentially extended to other modules of the SDS, making textto-action learning tractable. In addition, a single model can be potentially used in different domains (Papangelis and Stylianou, 2017), and different feudal architectures could make larger action spaces tractable (e.g. adding a third subpolicy to deal with actions dependent on 2 slots). Acknowledgments This research was funded by the EPSRC grant EP/M018946/1 Open Domain Statistical Spoken Dialogue Systems References Paweł Budzianowski, Stefan Ultes, Pei-Hao Su, Nikola Mrkšić, Tsung-Hsien Wen, Inigo Casanueva, Lina M. Rojas Barahona, and Milica Gašić. 2017. Subdomain modelling for dialogue management with hierarchical reinforcement learning. In Proc of SIGDIAL. Iñigo Casanueva, Paweł Budzianowski, Pei-Hao Su, Nikola Mrkšić, Tsung-Hsien Wen, Stefan Ultes, Lina Rojas-Barahona, Steve Young, and Milica Gašić. 2017. A benchmarking environment for reinforcement learning based task oriented dialogue management. arXiv preprint arXiv:1711.11023 . Inigo Casanueva, Thomas Hain, Heidi Christensen, Ricard Marxer, and Phil Green. 2015. Knowledge transfer between speakers for personalised dialogue management. In Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue. pages 12–21. Heriberto Cuayáhuitl, Steve Renals, Oliver Lemon, and Hiroshi Shimodaira. 2010. Evaluation of a hierar- chical reinforcement learning spoken dialogue system. Computer Speech & Language 24(2):395–429. Heriberto Cuayáhuitl, Seunghak Yu, Ashley Williamson, and Jacob Carse. 2016. Deep reinforcement learning for multi-domain dialogue systems. NIPS Workkshop . Peter Dayan and Geoffrey E Hinton. 1993. Feudal reinforcement learning. In Advances in neural information processing systems. pages 271–278. Milica Gašić, Catherine Breslin, Matthew Henderson, Dongho Kim, Martin Szummer, Blaise Thomson, Pirros Tsiakoulis, and Steve Young. 2013. Pomdpbased dialogue manager adaptation to extended domains. In Proceedings of the SIGDIAL Conference. Milica Gašić, Nikola Mrkšić, Pei-hao Su, David Vandyke, Tsung-Hsien Wen, and Steve Young. 2015. Policy committee for adaptation in multidomain spoken dialogue systems. In Automatic Speech Recognition and Understanding (ASRU), 2015 IEEE Workshop on. IEEE, pages 806–812. James Henderson, Oliver Lemon, and Kallirroi Georgila. 2008. Hybrid reinforcement/supervised learning of dialogue policies from fixed data sets. Computational Linguistics 34(4):487–511. M. Henderson, B. Thomson, and J. Williams. 2014a. The Second Dialog State Tracking Challenge. In Proc of SIGdial. M. Henderson, B. Thomson, and S. J. Young. 2014b. Word-based Dialog State Tracking with Recurrent Neural Networks. In Proc of SIGdial. Esther Levin, Roberto Pieraccini, and Wieland Eckert. 1998. Using markov decision process for learning dialogue strategies. In Acoustics, Speech and Signal Processing, 1998. Proceedings of the 1998 IEEE International Conference on. IEEE, volume 1, pages 201–204. Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. 2013. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 . Alexandros Papangelis and Yannis Stylianou. 2017. Single-model multi-domain dialogue management with deep learning. In International Workshop for Spoken Dialogue Systems. B. Peng, X. Li, L. Li, J. Gao, A. Celikyilmaz, S. Lee, and K.-F. Wong. 2017. Composite Task-Completion Dialogue System via Hierarchical Deep Reinforcement Learning. ArXiv e-prints . Olivier Pietquin, Matthieu Geist, Senthilkumar Chandramohan, and Hervé Frezza-Buet. 2011. Sampleefficient batch reinforcement learning for dialogue management optimization. ACM Transactions on Speech and Language Processing (TSLP) 7(3):7. Pei-Hao Su, Pawel Budzianowski, Stefan Ultes, Milica Gasic, and Steve Young. 2017. Sample-efficient actor-critic reinforcement learning with supervised data for dialogue management. Proceedings of SigDial . Pei-Hao Su, Milica Gasic, Nikola Mrksic, Lina RojasBarahona, Stefan Ultes, David Vandyke, TsungHsien Wen, and Steve Young. 2016. Continuously learning neural dialogue management. arXiv preprint arXiv:1606.02689 . Richard S. Sutton and Andrew G. Barto. 1999. Reinforcement Learning: An Introduction. MIT Press. Stefan Ultes, Lina M. Rojas-Barahona, Pei-Hao Su, David Vandyke, Dongho Kim, Iñigo Casanueva, Paweł Budzianowski, Nikola Mrkšić, Tsung-Hsien Wen, Milica Gašić, and Steve J. Young. 2017. Pydial: A multi-domain statistical dialogue system toolkit. In ACL Demo. Association of Computational Linguistics. B DIP features This section gives a detailed description of the DIP feature functions φdip (b, s) = ψ0 (b) ⊕ ψj (b) ⊕ ψd (b, s) used in this work. The differences with the features used in (Wang et al., 2015) and (Papangelis and Stylianou, 2017) are the following: • No priority or importance features are used. • No Potential contribution to DB search features are used. • The joint belief features ψj (b) are extended to account for large-domain aspects. Feature function Feature description Feature size ψ0 (b) last user dialogue act (bin) * DB search method (bin) * # of requested slots (bin) offer happened * last action was Inform no venue * normalised # of slots (1/# of slots) normalised avg. slot length (1/avg. # of values) 7 6 5 1 1 1 1 ψj (b) prob. of the top 3 values of bj prob. of NONE value of bj entropy of bj diff. between top and 2nd value probs. (bin) # of slots with top value not NONE (bin) 3 1 1 5 5 ψd (b, s) prob. of the top 3 values of s prob. of NONE value of s diff. between top and 2nd value probs. (bin) entropy of s # of values of s with prob. > 0 (bin) normalised slot length (1/# of values) slot length (bin) entropy of the distr. of values of s in the DB 3 1 5 1 5 1 10 1 Zhuoran Wang, Tsung-Hsien Wen, Pei-Hao Su, and Yannis Stylianou. 2015. Learning domainindependent dialogue policies via ontology parameterisation. In SIGDIAL Conference. pages 412–416. Steve Young, Milica Gašić, Blaise Thomson, and Jason D Williams. 2013. Pomdp-based statistical spoken dialog systems: A review. Proceedings of the IEEE 101(5):1160–1179. A Feudal Dialogue Policy algorithm Algorithm 1 Feudal Dialogue Policy 1: for each dialogue turn do 2: observe b 3: bm = φm (b) 4: am = argmax Qm (bm , am ) total am ∈Am 5: 6: 7: if am == am i then bi = φi (b) a = argmax Qi (bi , ai ) . drop to πi ai ∈Ai 8: 9: 10: else am == am . drop to πd d then bs = φs (b) ∀s ∈ S slot, act = argmax Qs (bs , ad ) s∈S,ad ∈Ad a = join(slot, act) end if 13: execute a 14: end for 11: 12: 64 Table 3: List of features composing the DIP features. the tag (bin) denotes that a binary encoding is used for this feature. Some of the joint features ψj (b) are extracted from the joint belief bj , computed as the Cartesian product of the beliefs of the individual slots. * denotes that these features exist in the original belief state b.	2cs.AI
IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS: CONSISTENCY AND FISHER INFORMATION arXiv:1707.04371v1 [math.ST] 14 Jul 2017 JEREMIE HOUSSINEAU∗ , SUMEETPAL S. SINGH† , AND AJAY JASRA‡ Abstract. Learning the model parameters of a multi-object dynamical system from partial and perturbed observations is a challenging task. Despite recent numerical advancements in learning these parameters, theoretical guarantees are extremely scarce. In this article, we study the identifiability of these parameters and the consistency of the corresponding maximum likelihood estimate (MLE) under assumptions on the different components of the underlying multi-object system. In order to understand the impact of the various sources of observation noise on the ability to learn the model parameters, we study the asymptotic variance of the MLE through the associated Fisher information matrix. For example, we show that specific aspects of the multi-target tracking (MTT) problem such as detection failures and unknown data association lead to a loss of information which is quantified in special cases of interest. Key words. Identifiability, Consistency, Fisher Information AMS subject classifications. 62F12, 62B10 1. Introduction. A multi-object dynamical system is comprised of an unknown and randomly varying number of objects, each of which is a partially observed Markov process. Multi-target tracking refers to the problem of estimating the state of each of these objects from noisy observations that are also corrupted by detection failures and false detections (a.k.a. false alarms). This type of problem arises in many different fields such as Systems Biology [2], Robotics [12], Computer Vision [14] or Surveillance [16]. Different formulations of multi-target tracking exist, including extensions of the single-target approach to multiple targets [1] as well as formulations based on simple point processes [10]. One of the main challenges in multi-target tracking is the uncertainty in the data association, which refers to the problem of finding the right pairing between targets and recorded observations over time, a task further confounded by the corruption of these observations with false positives and detection failures. Inferentially, multitarget tracking is notoriously difficult to solve as it involves an exponentially growing numbers of possible configurations for the data association. Over the past decade there has been significant advancements towards more practical solutions to this inference problem. Some of these include solutions based on sequential Monte Carlo (SMC) [19], hierarchical SMC [15] or Gaussian mixtures [18]. In this article, both the MTT observation model and the motion model of the constituent individual targets are assumed unknown and are instead parameterised and to be inferred from the data. Although multi-target tracking has been an active research field for decades, questions concerning the identifiability and the consistency of the corresponding model parameter estimates have not received the appropriate attention. In this paper we aim to address this gap and shed some light on this issue. Building on results from the literature on Markov processes (e.g. see [9, 6]), we prove both identifiability and the consistency of the MLE of the MTT model parameters in Theorem 3. Specifically, as each constituent target of the MTT model is a partially observed Markov process, in Theorem 2 we show that identifiability transfers ∗ DSAP, National University of Singapore. Email: stahje@nus.edu.sg of Engineering, University of Cambridge and The Alan Turing Institute. Email: sss40@cam.ac.uk ‡ DSAP, National University of Singapore. Email: staja@nus.edu.sg † Department 1 2 J. HOUSSINEAU, S.S. SINGH AND A. JASRA from single to multiple targets under appropriate assumptions. The practical implications of results regarding identifiability include the understanding of the behaviour of Markov chain Monte Carlo (MCMC) techniques in multi-target tracking [13, 8], which is conditioned by the likelihood ratio between the correct parameter value and all the other possible values. The consistency of the maximum likelihood estimator raises the question of its asymptotic normality and the corresponding variance, which in turns motivates the study of the Fisher information matrix for this class of problems. It is demonstrated in Theorem 4 that there is a strict loss of information in the presence of data association uncertainty or detection failures. We characterise the Fisher information more precisely in specific illustrated cases, e.g. we show that when increasing the number of targets there is no gain in the Fisher information for the model parameters which are common to all targets if large uncertainties on the origin of the corresponding observations persist (see subsection 5.3). The Fisher information matrix is useful in applications such as sensor management [7] which aims at optimising the position of the sensor or at finding the best ratio between probability of false alarm and probability of detection. The proof of identifiability of the MTT model as well as our approach for studying the asymptotic variance of the MLE for the MTT model parameters are original and, to the best of our knowledge, the first of their kind. Consistency of the data association problem in MTT has been studied in [17] in the context of the estimation of multiple splitting and merging targets observed without noise over a fixed time interval during which n observations of the multiple targets are made at discrete times. The result in [17] is limited to the case where the number n of observation tends to infinity which effectively amounts to saying that targets are observed infinitely many times over a fixed interval which is a scenario not typically encountered in practice. In any case, our theoretical results and proof techniques are entirely different as they pertain to the MTT model parameters and not the data association. Point-process-based theoretical studies of MTT have also been conducted in [4, 5] for the stability of specific inference methods. The structure of the article is as follows: after introducing the required notations and background concepts in section 2 and section 3, the consistency of the maximum likelihood estimator is established along with its asymptotic normality for a large class of multi-object systems in section 4. Finally, in order to better understand the effect of the various parameters on the asymptotic variance, the Fisher information matrix is computed for important special cases of multi-object systems in section 5. The article concludes in section 6. 2. Notations. All random variables will be defined on the same probability space (Ω, F, P) and the expectation of a random variable X w.r.t. the probability measure P is denoted E[X]. Probability densities will be denoted by lower-case letters while probability measures will be denoted by capital letter. Similarly, random variables will be denoted with capital letters whereas their realisations will be in lower-case. The time is indexed by the set N of positive integers and for every time t ∈ N, a . finite sequence yt of Mt ∈ N0 = N ∪ {0} observation points in the observation space Y is made available. This space can be assumed to be a subset of the Euclidean space Rd with d > 0. The sequences of observations of the form (y1 , . . . , yn ) will be denoted y1:n . In the standard formulation of multi-target tracking, no more than one observation is associated with a given object at a given time step and, conversely, observations are originated from one object only. IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 3 Objects’ states are modelled as elements of a set X which is assumed to be a 0 subset of the Euclidean space Rd with d0 > 0; usually satisfying d0 ≥ d. They are propagated independently according to a Markov kernel density fθ from the state space X to itself, which depends on a parameter θ from a compact set Θ. Densities on X are defined w.r.t. a reference measure µ. The true value of the parameter θ is denoted θ∗ . The random variable Xt describing the state at time t only depends on the state xt−1 at time t − 1 as follows Xt ∼ fθ (· \| xt−1 ). This transition does not depend on time so that the associated Markov chain is said to be homogeneous. The observation process at time t given the state xt is modelled by Yt ∼ gθ (· \| xt ) where gθ is a likelihood function from X to Y, also parametrised by θ, so that the observation Yt at time t is independent from the states and observations at other times. The process (Xt , Yt )t≥1 is usually referred to as a hidden Markov model (HMM). Its law under the parameter θ ∈ Θ is denoted P̄θ when initialised with its stationary distribution assuming it exists, and Pθ (· \| x0 ) when initialised at x0 ∈ X. 3. Background. The definition of specific properties of Markov chains that will be used in the following sections is given here for completeness. Let (Xt )t≥0 be a X-valued Markov chain with transition density f and let P (· \| x) be the probability measure on (XN0 , X ⊗N0 ), where X ⊗N0 is the cylinder σ-algebra on XN0 , characterising the chain when initialised at point x ∈ X. Also, let τA be the return time to a set A ⊆ X defined as τA = inf{t ≥ 1 : Xt ∈ A}. Consider the following concepts: A set A ⊆ X is said to be accessible if τA < ∞ has positive probability under P (· \| x) for all x ∈ X. The Markov chain (Xt )t≥0 is said to be R phi-irreducible if there exists a density φ on X such that for any subset A ⊆ X, A φ(x)dx > 0 implies that A is accessible. A set A ⊆ X is said to be Harris recurrent if the event τA < ∞ happens almost surely under P (· \| x) for all x ∈ X. A phi-irreducible Markov chain is said to be Harris recurrent if any accessible set is Harris recurrent. A density q is called invariant if for all x ∈ X it holds that Z q(x) = f (x \| x0 )q(x0 )dx0 . A phi-irreducible Markov chain is called positive if it admits an invariant probability density. More details about these notions expressed in a measure-theoretic formulation can be found in [11]. These concepts will be useful when considering the long-time behaviour of the Markov chains involved in multi-target tracking problems. 4. Consistency of the maximum likelihood estimator. 4.1. The multi-target tracking model. Throughout this section, the true number of objects in the considered system will be assumed to be fixed and will be ∗ denoted by K ∗ ∈ N. We consider a Markov chain (Xt )t≥0 in XK with components independently evolving via the Markov transition fθ S from X to X. Observations at . time t are gathered into a vector yt in the space Y× = k≥0 Yk where Y0 is a notation for the set containing the empty sequence only. The observation yt is a superposition of 4 J. HOUSSINEAU, S.S. SINGH AND A. JASRA 1. the independent observation of components of Xt via the likelihood gθ from X to Y followed by a Bernoulli thinning with parameter pD corresponding to detection failure, and 2. false alarms, or clutter, generated independently of the object-originated observations and assumed to come from an i.i.d. process whose cardinality at each time is Poisson with parameter λ and common distribution Pψ which depends on the parameter ψ in a compact set Ψ and which true value is denoted ψ ∗ . The number of objects K ∗ is not assumed to be known so that it will also be considered as a parameter of the model. The parameter for the multi-target model is then . . defined as θ = [θ, K, pD , λ, ψ]t ∈ Θ = Θ × S T × (0, 1) × S C × Ψ where t is the vector T C transposition and where S and S are compact subsets of N and (0, ∞) respectively, with “T” and “C” standing for target and clutter respectively. The true parameter θ ∗ is assumed to be an interior point of Θ. Special parameter sets that are not subsets of Θ can also be introduced by fixing one or several parameters to special values, for . . . instance Θλ=0 = Θ × S T × (0, 1), ΘpD =1 = Θ × S T × S C × Ψ or Θλ=0,pD =1 = Θ × S T correspond respectively to cases where the parameters λ, pD or both have known values that are outside of their domain of definition in Θ. Alternatively, if the value of a parameter is known but inside of its domain of definition, e.g. it is known that K = 1, then the corresponding hyperplane will be expressed as Θ\|K=1 . Although the Poisson distribution is not defined for the parameter λ = 0, this parameter value is simply assumed to represent the case where there is no false alarm. The Markov transition fθ associated with the K-target process (Xt )t can simply be expressed as fθ (x \| x0 ) = K Y fθ (xi \| x0i ), i=1 0 K for any x, x ∈ X , the likelihood however takes a more sophisticated form so that additional notations are required. Let Sym(k) be the symmetric group over k letters and uk be the uniform distribution over Sym(k), also let qθ be a distribution on {0, 1}K characterised by . \|d\| qθ (d) = pD (1 − pD )K−\|d\| , for any d ∈ {0, 1}K . The variable d is such that di = 1 if and only if target i is detected for any i ∈ {1, . . . , K}. The K-target likelihood gθ (yt \| xt ) of the observations yt ∈ Y× at time t given the state x ∈ XK is characterised by X . (1) gθ (yt \| x) = Poλ (Mt − \|d\|) d∈{0,1}K \|d\|≤Mt × X Mt Y σ∈Sym(Mt ) i=\|d\|+1 \|d\| pψ yt,σ(i) Y gθ yt,σ(i) \| xr(i) uMt (σ)qθ (d) , i=1 where Poλ denotes the Poisson distribution with parameter λ, where \|d\| is the 1-norm of d, i.e. the number of detected targets and where r(i) is the ith detected target that is the integer verifying \|d1:r(i) \| = i. This choice of the likelihood gθ corresponds to a marginalisation over the observation-to-track data association. Note that \|d\| ≤ K for any d ∈ {0, 1}K so that gθ (y \| x) = 0 for any x ∈ XK if λ = 0 and if the number IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 5 of observations in y, denoted #y, is strictly greater than K. The law of the joint Markov chain (Xt , Yt )t under the parameter θ ∈ Θ is denoted P̄θ when initialised by the stationary distribution and Pθ (· \| x0 ) when assumed to start at the state x0 ∈ XK . The corresponding densities are written accordingly with lower-case letters. The objective is to study the ratio pθ (y1:n \| x0 )/pθ∗ (y1:n \| x00 ) for any x0 ∈ XK ∗ and any x00 ∈ XK . The assumptions that are considered for this purpose are detailed in the next section. 4.2. Assumptions and transferability. In order to bring a better understanding of multi-object systems as a combination of single-object systems corrupted by clutter, assumptions are primarily made on individuals systems. The properties of multi-object systems will be deduced from these whenever this is possible. A.1 The constants (2) τ− = inf inf θ∈Θ (x,x0 )∈X2 fθ (x \| x0 ) and τ+ = sup sup fθ (x \| x0 ) θ∈Θ (x,x0 )∈X2 satisfy τ− > 0 and τ+ < ∞. The condition on τ− in Assumption A.1 ensures that any point of the state space can be reached from any other point in a single time step (otherwise fθ (x \| x0 ) = 0 would hold for at least one pair (x, x0 ) ∈ X2 ) while the condition on τ+ ensures the transition is sufficiently regular when compared to the reference measure µ, i.e. the transition should be diffuse (in the sense that there should be no concentration of probability mass on a single point of the state space). Under Assumption A.1 it also holds that K K τ− ≤ fθ (x \| x0 ) ≤ τ+ (3) for any x, x0 ∈ XK , so that fθ straightforwardly satisfies the same type of conditions as fθ , since S T is compact and hence K is finite. Let Πθ be the transition kernel of the joint Markov chain (Xt , Yt )t on X × Y defined as Πθ (x, y \| x0 , y 0 ) = gθ (y \| x)fθ (x \| x0 ). The property (3) is sufficient to ensure that the joint kernel defined as Πθ (x, y \| x0 , y 0 ) = gθ (y \| x)fθ (x \| x0 ), for any x, x0 ∈ XK and any y, y 0 ∈ Y× is positive Harris-recurrent and aperiodic. In the next assumption, the expectations Ēθ∗ [·], Eψ∗ [·] and Ēθ∗ [·] are taken with respect to P̄θ∗ , Pψ∗ and P̄θ∗ respectively, also Bikp denotes the binomial distribution with success probability p and k trials. A.2 The constant . (4) b̂T sup gθ (y \| x) + = (θ,x,y)∈Θ×X×Y satisfies b̂T + < ∞, the target- and clutter-related functions Z Z T bT : y → 7 inf g (y \| x)dx and b : y → 7 sup gθ (y \| x)dx, θ − + θ∈Θ bC − θ∈Θ : y 7→ inf pψ (y) ψ and bC + : y 7→ sup pψ (y), ψ satisfy (6a) bT − (y) > 0 and bT + (y) < ∞ (6b) bC − (y) > 0 and bC + (y) < ∞ 6 J. HOUSSINEAU, S.S. SINGH AND A. JASRA for any y ∈ Y as well as (7) Ēθ∗ [\| log bT − (Y )\|] < ∞ and Eψ∗ [\| log bC − (Y )\|] < ∞, and it holds that (8) Ēθ∗ log inf BiK pD ∗ Poλ (#Y ) θ∈Θ < ∞. Assumption A.2 ensures that all points of the observation space Y can be reached from at least some states in X via (6a) although gθ (y \| x) = 0 might hold for some (x, y) ∈ X × Y. Equation (7) will ensure boundedness in the calculations related to identifiability. The upper bound (4) of the likelihood function is also assumed to be finite so that no concentration of probability mass is allowed at any point of X × Y. It is demonstrated in the following lemma that the upper and lower bounds considered in Assumption A.2 for a single target and for the clutter common distribution are sufficient to guarantee the same type of result for multiple targets. The proof is in Appendix B. Lemma 1 (Transfer of boundedness). Under Assumption A.2, it holds that the constant . b̂+ = sup sup gθ (y \| x) θ∈Θ (x,y)∈XK ×Y× is finite and that the functions Z b− : y 7→ inf gθ (y\|x)dx θ∈Θ Z and b+ : y 7→ sup gθ (y\|x)dx, θ∈Θ verify b− (y) > 0 and b+ (y) < ∞ for any y ∈ Y× as well as Ēθ∗ [\| log b− (Y )\|] < ∞. An important result that follows from the assumptions introduced so far is the uniform forgetting of the conditional Markov chain: it can be proved under Assumptions A.1 and A.2 that for any k, l ∈ N0 such that k ≤ l and any parameter θ ∈ Θ, it holds that Z Z Z p̄θ (xt \| xk , yk+1:l )p(xk )dxk − p̄θ (xt \| xk , yk+1:l )p0 (xk )dxk dxt ≤ ρt−k θ , for all t ≥ k, all probability densities p, p0 on XK and all sequences of observations . yk+1:l , where ρθ = 1 − (τ− /τ+ )K . The K-target forgetting rate ρθ will generally be smaller than the single-target rate 1−τ− /τ+ , although mixing is still guaranteed since K is finite and hence ρθ ∈ [0, 1). It is also possible to conclude about the pointwise convergence of the log-likelihood function to the function ` defined on Θ as follows ` : θ 7→ Ēθ∗ `Y−∞:0 (θ) , where `y−∞:0 is defined on Θ for any realisation y−∞:0 of the observation process as `y−∞:0 : θ 7→ lim log p̄θ (y0 \| y−m:−1 , x−m−1 ), m→∞ and this limit does not depend on x−m−1 . Indeed, under Assumptions A.1 and A.2, it holds for all K ∈ S T and all x0 ∈ XK that (9) 1 log pθ (Y1:n \| x0 ) = `(θ), n→∞ n lim P̄θ∗ -a.s. IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 7 This result shows that for any realisation y1:∞ of the observation process, the empirical average n−1 log pθ (y1:n \| x0 ) will converge to `(θ) irrespectively of the assumed initial state x0 . A continuity assumption is required in order to turn the pointwise convergence result of (9) into a uniform convergence result. A.3 For all x, x0 ∈ X and all y ∈ Y, the mappings θ 7→ fθ (x \| x0 ), θ 7→ gθ (y \| x) and ψ 7→ pψ (y) are continuous. It follows directly from Assumption A.3 that for all K ∈ S T , all x, x0 ∈ XK and all y ∈ Y× , the mappings θ 7→ fθ (x \| x0 ) and θ 7→ gθ (y \| x) are continuous on the hyperplane of Θ made of parameters with a number of targets equal to K, since these mappings are sums and products of continuous functions. Although the continuity for the multi-target Markov kernel and likelihood function is limited to hyperplanes, the result of [6, Lemma 4] can be extended to: for all θ ∈ Θ 0 lim Ēθ∗ sup `Y−∞:0 (θ ) − `Y−∞:0 (θ) = 0, δ→0 \|θ 0 −θ\|≤δ where \| · \| is the 1-norm on Θ, since θ 0 and θ will be in the same hyperplane for δ small enough. The addition of the continuity assumption enables the derivation of the following result regarding the uniform convergence of the log-likelihood function: Under Assumptions A.1 to A.3, it holds that (10) lim sup sup n→∞ θ∈Θ x ∈XK 0 1 log pθ (Y1:n \| x0 ) − `(θ) = 0, n P̄θ∗ -a.s. Since the conditional log-likelihood function log pθ (y1:n \| x0 ) is continuous and uniformly bounded, it follows from (10) that ` is also continuous on the hyperplanes of Θ of constant target number. The following identifiability assumption is considered in order to show the consistency of the maximum likelihood estimator A.4 P̄θ = P̄θ∗ if and only if θ = θ∗ and Pψ = Pψ∗ if and only if ψ = ψ ∗ Assumption A.4 is fundamental since there would be no chance to discriminate the true value θ∗ among all the other possible θ ∈ Θ \ {θ∗ } if some of these parameters did yield the same law for the observations. For instance, if the colour of the target is considered as a parameter but if the likelihood of the observations does not depend on this characteristics of the target, e.g. if the observations come from a radar, then any θ obtained by changing the colour in θ∗ would induce a law P̄θ that is equal to P̄θ∗ and Assumption A.4 would not be verified. It is shown in the next theorem that identifiability of the multi-target problem can be deduced from the identifiability of the single-target one under important special cases. The proof is in Appendix C. Theorem 2 (Transfer of identifiability). Under Assumption A.4 it holds that a) if the true parameter θ ∗ is in Θλ=0 , then it holds that P̄θ = P̄θ∗ if and only if θ = θ ∗ for any θ ∈ Θλ=0 , b) if the true parameter θ ∗ is in the subset Θ\|K=1 of Θ made of parameters of the form (θ, K, pD , λ, ψ) with K = 1, then it holds that P̄θ = P̄θ∗ if and only if θ = θ ∗ for any θ ∈ Θ\|K=1 . It is more challenging to prove that identifiability transfers to the whole parameter set Θ and this property is assumed to hold rather than demonstrated. A.5 P̄θ = P̄θ∗ if and only if θ = θ ∗ . Assumption A.5 is not a stringent condition since Theorem 2 shows that the single-target identifiability is sufficient to ensure multi-target identifiability in some 8 J. HOUSSINEAU, S.S. SINGH AND A. JASRA important special cases. Moreover, if there exists a θ 6= θ ∗ such that P̄θ = P̄θ∗ then θ has to satisfy very specific equations including ∗ K ∗ Poλ∗ . BiK pD ∗ Poλ = Bip∗ D Assumption A.5 would not hold for p∗D = 0 since identifiability w.r.t. θ∗ and K ∗ would clearly be lost in this case because of the absence of observations from the targets. The same remark can be made about K ∗ = 0 for the identifiability w.r.t. θ∗ since there is obviously no way to learn about the dynamics and observation of the targets if none of them is present. The different assumptions considered here are combined in the next section in order to prove the consistency of the maximum likelihood estimator. 4.3. Consistency and asymptotic normality. As a consequence of (9) and by the dominated convergence theorem it holds that for any θ ∈ Θ, any infinite ∗ observation sequence y1:∞ and any initial states x0 ∈ XK and x00 ∈ XK (11a) lim n→∞ (11b) pθ (y1:n \| x0 ) 1 log = `(θ) − `(θ ∗ ) n pθ∗ (y1:n \| x00 ) p̄θ (Y0 \| Y−m:−1 ) Y−m:−1 ≤ 0, = lim Ēθ∗ Ēθ∗ log m→∞ p̄θ∗ (Y0 \| Y−m:−1 ) where the inequality holds since the conditional expectations are Kullback-Leibler divergences. Yet, it could happen that some θ ∈ Θ would verify p̄θ (Y0 \| y−m:−1 ) = p̄θ∗ (Y0 \| y−m:−1 ) P̄θ∗ -almost surely for all m ∈ N0 and for all y−m:−1 , which would compromise identifiability. However, Assumption A.5 is equivalent to θ = θ ∗ if and only if p̄θ (Y1:n ) Ēθ∗ log = 0, ∀n ≥ 1. p̄θ∗ (Y1:n ) The objective is to show that this, in turn, is equivalent to θ = θ ∗ if and only if `(θ) − `(θ ∗ ) = 0 since this is the term that appears in (11). Following the same line of arguments as [6, Proposition 3], we find that under Assumptions A.1 to A.3 and A.5, it holds that `(θ) = `(θ ∗ ) if and only if θ = θ ∗ , from which we conclude that the considered approach allows for studying the identifiability of θ ∗ . Applying the strict Jensen inequality to the conditional expectation in the r.h.s. of (11), it indeed follows that lim n→∞ pθ (y1:n \| x0 ) 1 < 0, log n pθ∗ (y1:n \| x00 ) for any θ 6= θ ∗ , which implies that the likelihood of the observation sequence y1:n under the parameter θ decreases exponentially fast when compared to the likelihood under θ ∗ , irrespectively of the assumed initial states x0 and x00 . Denoting θ̂n,x0 the argument of the maximum of log pθ (y1:n \| x0 ), the consistency of the maximum likelihood estimator can be expressed as in Theorem 3 below. This theorem also states the asymptotic normality of the estimator which makes use of the Fisher information. The latter involves differentiation with respect to the parameter θ, however since the number of target K is a natural number, differentiations has to be performed for a IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 9 fixed K. This is what is understood by default when writing ∇θ . Under assumptions, the Fisher information matrix can be expressed as h i 1 Ēθ∗ ∇θ log p̄θ∗ (Y1:n ) · ∇θ log p̄θ∗ (Y1:n )t , n→∞ n I(θ ∗ ) = lim where ·t is the matrix transposition. Theorem 3. Under Assumptions A.1 to A.3 and A.5, it holds that lim θ̂n,x0 = θ ∗ n→∞ for any x0 ∈ XK with K ∈ N. Considering additionally Assumptions A.6 to A.8 (see Appendix A) and assuming that I(θ ∗ ) is positive definite, it holds that √ n(θ̂n,x0 − θ ∗ ) → N 0, I(θ ∗ )−1 , for any x0 ∈ XK and any K ∈ N, where → denotes the convergence in distribution as n tends to infinity and where N (0, V ) is the normal distribution with mean 0 and variance V . The proof of Theorem 3 follows from Lemma 1 combined with [6, Theorems 1 and 4]. It can be demonstrated that the result of Theorem 3 also holds for the special parameter sets Θλ=0 , ΘpD =1 and Θλ=0,pD =1 . These special-parameter sets will be used to understand the behaviour of the Fisher information matrix in simple cases in the next section. 5. Analysis of the Fisher information. Theorem 3 guarantees the convergence of the maximum likelihood estimator under certain conditions and proves the asymptotic normality of the estimator, the variance of the latter being the inverse of the Fisher information matrix. It is therefore of interest to understand how the Fisher information behaves in different multi-target configurations. This section is structured as follows: an equivalent observation model for which the Fisher information matrix is easier to study is introduced in subsection 5.1 and yields a characterisation of the configurations in which the information loss induced by data association uncertainty and detection failures is strictly positive. Qualitative estimates of the information loss are then obtained when isolating the different sources of loss from subsection 5.2 to subsection 5.4. Each of these qualitative estimates are confirmed by numerical results on simulated data obtained by direct Monte Carlo integration of the original expression of the Fisher information, so as to confirm the validity of the derived alternative expressions. Henceforth, if A and B are two square matrices of the same dimensions then A ≥ B is understood as A − B ≥ 0, i.e. A − B is positive semi-definite, and A > B stand for A − B > 0, i.e. A − B is positive definite. Example 1. Assuming that θ ∗ is in ΘpD =1 and that the data association is known, the joint probability of the observations becomes p̄θ∗ (y1:n ) = n Y Poλ∗ (Mt − K ∗ ) Mt Y pψ∗ yt,i i=K ∗ +1 t=1 Z × ∗ πθ×K (x0 ) ∗ n Y K∗ h i Y gθ∗ (yt,i \| xt,i )fθ∗ (xt,i \| xt−1,i ) dx0:n . t=1 i=1 10 J. HOUSSINEAU, S.S. SINGH AND A. JASRA The score is then found to be ∗ ∇θ log p̄θ∗ (y1:n ) = K X i=1 ∇θ log p̄θ∗ (y1:n,i )+ n X Mt − K ∗ t=1 λ∗ −1+ Mt X ∇θ log pψ∗ (yt,i ) i=K ∗ +1 so that, because of the independence between the targets and clutter, I(θ ∗ ) = K ∗ I(θ ∗ ) + 1 + λ∗ I C (θ ∗ ), λ∗ with I(θ ∗ ) and I C (θ ∗ ) the Fisher information for the distribution of one target and one clutter point respectively, where the gradient is taken w.r.t. θ, that is h i 1 I(θ ∗ ) = lim Ēθ∗ ∇θ log p̄θ∗ (Y1:n ) · ∇θ log p̄θ∗ (Y1:n )t n→∞ n h i C ∗ I (θ ) = Eψ∗ ∇θ log pψ∗ (Y ) · ∇θ log pψ∗ (Y )t . In spite of its simplicity, Example 1 yields important remarks: unsurprisingly, if there is no missing information and no data association uncertainty, the information increases with the number of targets. Similarly, if the Fisher information of the clutter distribution pψ∗ increases, then the overall information increases too. The interpretation for the Poisson parameter λ∗ is less straightforward, the main objective is however to study the Fisher information w.r.t. the targets rather than the false alarms so that it is of interest to compute the score without differentiating with respect to ψ or λ. Although the Fisher information becomes more difficult to compute when p∗D ∈ (0, 1), some conclusions can be drawn by focusing on the cardinality. Since the parameter θ does not affect the cardinality, only the term E[∇pD log qθ∗ (D) · ∇pD log qθ∗ (D)t ] = K ∗ /(p∗D (1 − p∗D )) remains when computing the Fisher information matrix, with D the random variable ∗ induced by Y on {0, 1}K . This term is minimal when p∗D = 0.5 and increases when p∗D goes toward 0 or 1. This is not sufficient to conclude since the fact that information is lost when detection failures happen is not taken into account in the cardinality and the information is the same for, e.g. p∗D equal to 0.99 or 0.01. Indeed, it is equally easy to estimate p∗D when an observation is always or never received. For this reason, it is useful to consider the information w.r.t. θ∗ only. The objective will therefore be to characterise how the Fisher information h i 1 (13) I(θ∗ ) = lim Ēθ∗ ∇θ log p̄θ∗ (Y1:n ) · ∇θ log p̄θ∗ (Y1:n )t , n→∞ n of a multi-object dynamical system behaves when compared to the information of the unperturbed system that excludes false alarms, detection failures and for which data association is known. We refer to the difference between (13) and the latter as the information loss. Since the Fisher information of the unperturbed system is a quantity that depends on the number of objects in the system, the aim is to express the information loss as a function of the single-object Fisher information matrix I(θ∗ ). The Fisher information matrix of the unperturbed multi-object system is clearly equal to K ∗ I(θ∗ ) because of the independence between the targets’ observation in the absence of data association uncertainty. In order to compute I(θ∗ ), we have to take the IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 11 logarithm of the probability density function ∗ p̄θ∗ (y1:n , x0:n ) = πθ×K (x0 ) ∗ n Y gθ∗ (yt \| xt )fθ∗ (xt \| xt−1 ) . t=1 However, the presence of a sum in the term gθ∗ (yt \| xt ) prevents from further analysing the Fisher information in a general setting. To avoid directly dealing with these sums, an equivalent observation model which depends explicitly on the assignment is introduced in the next section. This observation model is an important contribution since it allows us to understand the behaviour of the Fisher information for multitarget tracking. 5.1. Alternative observation model. Let dH be the Hamming metric on the symmetric group Sym(k) characterised by letting dH (σ, σ 0 ) be the number of points moved by σ 0 ◦ σ −1 for any given σ, σ 0 ∈ Sym(k). For instance, if k = 5 and if σ and σ 0 are given in Cauchy’s two-line notation as 1 2 3 4 5 1 2 3 4 5 σ= and σ0 = 1 5 2 4 3 1 3 5 4 2 then dH (σ, σ 0 ) = 3 since σ 0 ◦ σ −1 (i) 6= i if and only if i is in the set {2, 3, 5}. Let ⊕ be the vector concatenation operator such that if y = [y1 , . . . , yn ]t ∈ Yn and 0 t ] ∈ Ym then y 0 = [y10 , . . . , ym . 0 t y ⊕ y 0 = [y1 , . . . , yn , y10 , . . . , ym ] ∈ Yn+m . ∗ Let Rd be the matrix of size \|d\| × K ∗ such that (Rd )i,j = δj,r(i) for any d ∈ {0, 1}K , i.e. Rd has as many lines as there are detected targets and can be seen as a mask matrix that removes the observations of non-detected ones. Let Sσ be the permutation matrix corresponding to σ ∈ Sym(k) for any k ≥ 1, i.e. the matrix defined as   eσ(1) .   Sσ =  ...  , eσ(K ∗ ) with ei the row vector with 1 at the ith position and 0 elsewhere. The observation model with known data association is written as (14) Yt = h(Xt ) + η with h and η the multi-target observation function and the observation noise respectively, where η is i.i.d. across its K ∗ components. The false alarms are defined as a random variable Ŷ in Y× , independent of Yt , such that Ŷi ∼ pψ and Ŷi is independent of Ŷj for any 1 ≤ i, j ≤ #Ŷ . The observation model of interest can then be defined for given integers α > 0 and 0 ≤ β ≤ K ∗ as (15) Ytα,β = Sς (RD Yt ) ⊕ Ŷ , ∗ . where D is a random element of Bβ = {d0 ∈ {0, 1}K : \|1 − d0 \| ≤ β} having as a distribution the restriction qθβ of qθ to Bβ and where ς is a random permutation α drawn from the uniform law uα k with k = #Ŷ + \|d\| on the set Ak defined by . Aα k = {σ ∈ Sym(k) : dH (id, σ) ≤ α} 12 J. HOUSSINEAU, S.S. SINGH AND A. JASRA with id denoting the identity function. Henceforth, the letter ς will be used for a random permutation and σ for a realisation. The case α = 0 is not considered to avoid redundancy: it holds that A0k = A1k = {id} for any k ≥ 1 since permutations that are different from the identity move at least two points. The case of Example 1 is recovered by considering α = 1 and β = 0, i.e. ς = id and D = 1 almost surely, whereas the full data-association problem corresponds to the choice α = ∞ and β = ∞. The cardinality of Aα k is found to be Nkα = α X k i=0 i !i, with !i the subfactorial of i which is equal to the number of derangements of i letters, where a derangement refers to a permutation that moves all the elements of its domain. The subfactorial !i is defined via the same recurrence relation as the factorial i!, i.e. as !i = (i − 1)(!(i − 1)+!(i − 2)), but with the initialisation !0 = 1 and !1 = 0. The expression of Nkα can be justified as follows: the number of permutations moving a number of points less or equal to α is also the number of permutations moving exactly i points, i.e. the number of derangements of i points, multiplied by the number of ways of picking i points among k, for all 0 ≤ i ≤ α. It holds that Nkk = k! since n X n !n = n! − !(n − i). i i=1 The alternative observation model (15) brings insight about the Fisher information matrix I α,β (θ∗ ) corresponding to the observation model (15), when compared to the unperturbed case. The corresponding information loss is defined as α,β ∗ . Iloss (θ ) = K ∗ I(θ∗ ) − I α,β (θ∗ ). α,β ∗ In some cases, the relative information loss Iloss (θ )/(K ∗ I(θ∗ )) will be used instead. The next theorem is the central result of this section, its proof can be found in Appendix D. Theorem 4. Under Assumptions A.1, A.2 and A.6 to A.8, the information α,β ∗ α,β ∗ loss Iloss (θ ) verifies Iloss (θ ) ≥ 0 for any α ≥ 1 and any β ≥ 0, the inequality being strict if either α > 1 or β > 0 and if I(θ∗ ) 6= 0. Notice that the condition α > 1 would not be sufficient to make the inequality in Theorem 4 strict if λ∗ were equal to 0 since data association might have no influence in some specific configurations, e.g. when the individual likelihood does not depend on the objects’ state. Theorem 4 does not provide a quantitative characterisation of the information loss. Doing so is challenging in the general case, yet, the behaviour of the information loss can be analysed for special cases, and such will be the objective in the remainder of this section. One of the advantages with the modified observation model (15) is that the Fisher identity can be utilised as an alternative way of computing the score function based on the unobserved random variables in this model: (16) α,β α,β ∇θ log p̄θ (y1:n ) = Ēθ ∇θ log p̄θ (Y1:n , ς1:n , D1:n , X0:n ) \| Y1:n = y1:n , 13 IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS where πθ×K (x0 ) p̄θ (y1:n , σ1:n , d1:n , x0:n ) = n Y Poλ (Mt − \|dt \|) t=1 × \|d\| Mt Y pψ (yt,σt (i) ) i=\|dt \|+1 Y i=1 gθ (yt,σt (i) \| xt,r(i) ) K Y . β fθ (xt,i \| xt−1,i )uα Mt (σt )qθ (dt ) i=1 The simplification of the expression of ∇θ log p̄θ (y1:n ) is only notational. The random α,β = y1:n in (16), so variables ς1:n , D1:n and X0:n are conditioned on the event Y1:n that their respective distributions are now the conditional distributions given the observations, which are more complex than their priors. Yet, the Fisher identity enabled to move the sums and integrals outside of the logarithm, hence making easier the analysis of the Fisher information matrix. 5.2. Single static target with false alarm. Consider the case of one almostsurely detected static target with state x ∈ X which observation is corrupted by false alarms and unknown data association. The corresponding θ ∗ is in the hyperplane ΘpD =1 \|K=1 of the special parameter set ΘpD =1 composed of parameters for which K = 1. It is sufficient to study one time step since the observations at different times become independent in this case and it holds that h i 1 I(θ∗ ) = lim Ēθ∗ ∇θ log p̄θ∗ (Y1:n ) · ∇θ log p̄θ∗ (Y1:n )t n→∞ n h i = Ēθ∗ ∇θ log p̄θ∗ (Y1 ) · ∇θ log p̄θ∗ (Y1 )t . Making use of the Fisher identity (16), the Fisher information matrix I ∞,0 (θ∗ ) can be expressed as I ∞,0 (θ∗ ) = Eθ∗ X M ci (Y )cj (Y )∇θ log gθ∗ (Yi \| x) · ∇θ log gθ∗ (Yj \| x)t , i,j=1 where M = #Y and where ci (y) = gθ∗ (yi \| x)/pψ∗ (yi ) . uM (σ \| y) = PM ∗ ∗ j=1 gθ (yj \| x)/pψ (yj ) σ∈Sym(M ) X σ(1)=i Identifying the parameter θ∗ is most challenging when the distribution of the false alarm is equal to the one of the target-originated observation at θ∗ , i.e. pψ∗ = gθ∗ (· \| x), since all the observations will look alike for θ close to θ∗ . In this case it holds that ci (Y ) = 1/M for any 1 ≤ i ≤ M so that m X Poλ∗ (m − 1) X Eθ∗ ∇θ log gθ∗ (Yi \| x) · ∇θ log gθ∗ (Yi \| x)t \| m 2 m i=1 m≥1 h 1 i =E I(θ∗ ) N +1 I ∞,0 (θ∗ ) = where the expectation is taken w.r.t. the random variable N ∼ Poλ∗ . It follows that the relative information loss is equal to E[N/(N + 1)] so that it is strictly increasing 14 J. HOUSSINEAU, S.S. SINGH AND A. JASRA 1 0.9 0.8 relative loss 0.7 0.6 0.5 0.4 E[N/(N+1)] Gaussian U([-5,5]) U([-10,10]) U([-25,25]) U([-50,50]) U([-100,100]) 0.3 0.2 0.1 0 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 log(6+1) Fig. 1: Information loss as a function of the Poisson parameter λ in log-scale, calculated with 5 × 105 samples (Gaussian: worst-case scenario; U ([−a, a]): uniform distribution over [−a, a]). with λ∗ and tends to 1 when λ∗ tends to infinity. This result is supported by the experiments displayed in Figure 1 where the observation of one static target in X = R at x = 0 is corrupted by false alarms. The observation model is assumed to be linear and Gaussian with variance θ such that θ∗ = 1. Cases where the false alarm is uniform over the subset [−a, a] with a ∈ {5, 10, 25, 50, 100} are also considered. The scenario where the false alarm is distributed in the same way as the target-originated observation at θ∗ , i.e. pψ∗ = gθ∗ (· \| x), is also confirmed to be the worst-case scenario. In the next two sections, the focus will be on understanding the role played specifically by unknown data association and detection failures. 5.3. Unknown data association. In order to set the focus on data association, it is assumed that θ ∗ belongs to the special parameter set Θλ=0,pD =1 . In these conditions, the joint probability of the observations and states becomes ∗ (19) p̄θ∗ (y1:n , x0:n ) = πθ×K (x0 ) ∗ × n Y Y K∗ X t=1 σ∈Sym(K ∗ ) gθ∗ (yt,σ(i) \| xt,i )fθ∗ (xt,i \| xt−1,i ) uK ∗ (σ) . i=1 The sum in the previous expression makes it difficult to directly compute the Fisher information matrix. Some insight about it can however be obtained by considering static objects as in the following example. Example 2. Let x1 , . . . , xK ∗ be the known position of K ∗ static objects. The joint distribution of the observations is then found to be pθ∗ (y1:n ) = n Y X t=1 σ∈Sym(K ∗ ) Y K∗ gθ∗ (yt,σ(i) \| xi )uK ∗ (σ) . i=1 In this simplified setting, we can assume that gθ∗ has finite support so that the objects’ QK ∗ state can be chosen far enough from each other for i=1 gθ∗ (Yt,σ(i) \| xi ) = 0 to hold P-a.s. whenever σ 6= id. In this case, and as expected, there is no loss of information IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 15 when compared to the case with known data association. A less intuitive result can be found when all the objects’ state are equal to a given x ∈ X. In this situation, it holds that all permutations are equally probable so that ∗ pθ∗ (y1:n ) = n Y K Y gθ∗ (yt,i \| x), t=1 i=1 and once again, there is no loss of information. These two cases correspond to extreme configurations where the uncertainty on the data association is either resolvable or irrelevant. The Fisher identity can be used to provide an expression of the Fisher information for static objects as follows. For any fixed x1 , . . . , xK ∗ , the Fisher information for α = ∞ (fully unknown association) and β = 0 can be deduced from ∇θ log pθ (y) = Eθ ∇θ log pθ (Y , ς) \| Y = y X = K X ∇θ log gθ (yσ(i) \| xi )uK (σ\|y). σ∈Sym(K) i=1 The Fisher information matrix I α,β (θ∗ ) with α = ∞, β = 0 and without false alarm is found to be I ∞,0 (θ∗ ) = Eθ∗ ∇θ log pθ∗ (Y ∞,0 ) · ∇θ log pθ∗ (Y ∞,0 )t ∗ X = K X Eθ∗ uK ∗ (σ\|Y )uK ∗ (σ 0 \|Y ) Scoi (Yσ(i) ) · Scoj (Yσ0 (j) )t σ,σ 0 ∈Sym(K ∗ ) i,j=1 ∗ = K X Eθ∗ ci,k (Y )cj,l (Y ) Scoi (Yk ) · Scoj (Yl )t i,j,k,l=1 with Scoi (y) = ∇θ log gθ∗ (y \| xi ) for any y ∈ Y and with X ci,k (y) = uK ∗ (σ\|y) σ∈Sym(K ∗ ) σ(i)=k ∗ = gθ∗ (yk \| xi ) X Y gθ∗ (yσ(j) \| xj ) σ∈Sym(K ∗ ) j6=i σ(i)=k X K Y −1 gθ∗ (yσ(j) \| xj ) σ∈Sym(K ∗ ) j=1 ∗ for any y ∈ YK and any i, k ∈ {1, . . . , K ∗ }. The term ci,k (y) is the conditional probability for the object with state xi to have generated observation k given all observations y. In order to obtain a quantitative characterisation of the information loss, a special likelihood has to be introduced. We consider an observation model of the same form as the one displayed in Figure 2, i.e. such that Y is compact and there exists a collection ∗ of disjoint subsets {Bi }K i=1 of Y such that gθ (· \| xi ) uniformly distributes a probability mass > 0 outside of Bi . An example of such a distribution is given in Figure 2 for two objects. Then, for K objects, (23) I ∞,0 (θ∗ ) = K X i,j,k,l=1 k,l ∗ Ei,j (θ ) J. HOUSSINEAU, S.S. SINGH AND A. JASRA probability density 16 0.4 g3 (.\|x 1 ) g (.\|x ) 3 0.3 2 0.2 0.1 0 -10 -8 -6 -4 -2 0 2 4 6 8 10 y Fig. 2: Example of likelihood with two objects at states x1 and x2 . k,l ∗ . with Ei,j (θ ) = Eθ∗ ci,k (Y )cj,l (Y ) Scoi (Yk ) · Scoj (Yl )t for any i, j, k, l ∈ {1, . . . , K}. The objective is now to understand the behaviour of I ∞,0 (θ∗ ) when K is large. The order of the term ci,k (y) is in O(1) when i = k and in O(K −1 ) when i 6= k. The order of the summand in (23) can then be determined for the different values of i, j, k, l. For instance: • If i 6= k 6= l 6= j then Z ∇θ gθ∗ (yk \| xi ) · ∇θ gθ∗ (yl \| xj )t 2 k,l ∗ c (y)c (y) dy, (24) Ei,j (θ ) = i,k j,l k,l \|Y \ Bk \|2 Ci,j gθ∗ (yk \| xi )gθ∗ (yl \| xj ) k,l . where Ci,j = {y ∈ YK : yk ∈ Bi , yl ∈ Bj }, because gθ∗ (y \| xk ) = /\|Y \ Bk \| for all y ∈ / Bk and because Bi ∩ Bk = ∅ since i 6= k. When K increases, Y needs to be augmented at least linearly to ensure that the family {Bi }K i=1 is disjoint and (24) shows inverse proportionality with \|Y\|2 , so that it is of order O(K −4 ) at most. There are O(K 4 ) terms of this form in the sum in the r.h.s. of (23) so that the sum of these terms is of order O(1) at most. k,l ∗ • If k = l and i 6= j then Ei,j (θ ) = 0 since in this case it holds that Scoi (yk ) · Scoj (yk )t = 0 for any y ∈ YK which follows from the facts that Scoi (y) 6= 0 when y ∈ Bi only and that Bi ∩ Bj = ∅. • If i = j = k = l then Z ∇θ gθ∗ (yi \| xi ) · ∇θ gθ∗ (yi \| xi )t k,l ∗ dy, Ei,j (θ ) = 1Bi (yi )ci,i (y)2 gθ∗ (yi \| xi ) which does not depend on K or \|Y\| and is therefore of order O(1). Following the same principles for the other values of i, j, k, l, we find that I ∞,0 (θ∗ ) is of order O(K). Since the information in the idealised observation model, i.e. when data association is known, is equal to KI(θ∗ ), it follows that the relative loss is constant. In other words, for a large number of targets, adding more targets increases the information at the same rate as in the idealised model. Validation via simulations. The special likelihood is taken of the form ( . N (y; xk + m, 1) if y ∈ Bk = (xk + m − r, xk + m + r) gθ (y \| xk ) = /\|Y \ Bk \| otherwise, R with = 0.1 and with r characterised by Bk N (y; xk , 1)dy = 1 − via Bk . In this case, the displacement m is considered as the parameter θ and the true value is θ∗ = 0. The relative information loss associated with this likelihood is displayed in Figure 3a under two different configurations. The first one (Constant observation space in the 17 IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 0.2 5 0.18 4.5 0.16 4 3.5 0.12 , relative loss 0.14 3 0.1 2.5 Constant observation space Adaptive observation space 0.08 2 0.06 1.5 0.04 1 0.02 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 0 7 0.5 1 1.5 2 2.5 3 3.5 = Number of targets (a) For a varying number of objects for the (b) For varying association uncertainty α and special likelihood (105 MC runs). spatial separation τ (104 MC runs). 1 0.7 experimental 1-p 0.9 D 0.6 0.8 0.7 relative loss relative loss 0.5 0.4 0.3 0.6 0.5 0.4 0.3 0.2 0.2 0.1 0.1 0 0 1 2 3 4 5 6 number of targets 7 8 9 10 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 pD (c) For a varying number of objects with sep- (d) For a varying probability of detection pD , aration τ = 1 and α = ∞ (104 MC runs). compared to 1 − pD . Fig. 3: Information loss with association uncertainty (3a-3c) or detection failures (3d). figure) corresponds to the case where the observation space is large enough to meet the requirements associated with (23); the relative loss can be seen to increase linearly with the number of targets. The second case (Adaptive observation space in the figure) corresponds to the case where the observation space has to be augmented to fit new targets and shows a constant relative information loss. This last result is consistent with the conclusion above that the information loss is of the same order as the number of targets when the observation space has to be augmented. Further simulations. Five static objects on X = R at positions xi = τ (i − 3) with i ∈ {1, . . . , 5} are observed via a linear Gaussian model with variance equal to 1. The objective is to understand how the Fisher information matrix I α,0 (θ∗ ) evolves with α and with the position of the objects. It is assumed that θ parametrises the variance of the Gaussian observation model only, so that I α,0 (θ∗ ) is a scalar. The relative information loss is displayed in Figure 3b and confirms the intuition that the information loss increases with α, except in the case α = 1 where there is no loss by 18 J. HOUSSINEAU, S.S. SINGH AND A. JASRA definition since A1k = {id} for any k ≥ 1 so that the data association is known in this case. Also, the loss is increased when the individual likelihoods overlap while being increasingly different and then decreases when the overlap becomes negligible. The maximum is reached when τ = 1, that is when the distance \|xi − xi−1 \| between two consecutive objects is 1 for any i ∈ {2, . . . , 5}. The fact that there is no loss when τ = 0 follows from the irrelevance of data association uncertainty when all objects are at the same position, as explained in Example 2. To better understand the behaviour w.r.t. the number of targets, Figure 3c displays the relative information loss for 1 to 10 targets in the case of full data association uncertainty with τ = 1. The results for the two sets of simulations are consistent and show the same trend: the relative information loss increases with the number of targets but tend to stabilise. To sum up, there is no loss for 1 target by construction, the loss is linear in the number of targets when there are sufficiently many, and it increases the fastest during the transition between these two modes. 5.4. Detection failures. In this section, the case of detection failures is analysed when assuming that there are no false alarms, that is when θ ∗ is in the specialparameter set Θλ=0 . To establish our main result in this section (Theorem 5), we will use the concept of missing information (see for instance [3] in the context of Approximate Bayesian Computation). α,β ∗ Theorem 5. Assuming θ ∗ ∈ Θλ=0 , the information loss Iloss (θ ) for known data association with unconstrained detection failures, i.e for α = 1, β = ∞, is found to be 1,∞ ∗ . Iloss (θ ) = (1 − p∗D )K ∗ I(θ∗ ). The proof can be found in Appendix E. It follows from Theorem 5 that in the considered configuration the Fisher information matrix I 1,∞ (θ∗ ) can be made arbitrary close to 0 by making p∗D tend to 0. Also, there is no loss at all when p∗D = 1, as expected. In order to verify the result of Theorem 5 in practice, a single-object scenario with detection failures and without false alarms is considered. The object starts at time t = 0 from the position x0 = 0 and evolves in X = R according to a random walk with standard deviation 0.1 until time n = 50. The observation is linear and Gaussian with variance equal to 1. The integral over the state space in the expression of the score is computed by Monte Carlo simulation with 103 samples while the expectation in the Fisher information utilises 104 samples. The relative information loss is displayed in Figure 3d and confirms the coefficient 1 − pD found analytically in Theorem 5. The next example shows how the Fisher information evolves in general when adding new objects without involving them in data association uncertainty. Example 3. The Fisher information I α,β (θ∗ , K) of a K-object problem can be related to the information I α,β (θ∗ , K + N ) where the N new objects are not perturbed by data-association uncertainties, i.e. when the random variable ς in the observation . model (15) verifies ς\|D = id almost surely with D = {\|d1:K+1 \|, . . . , \|d1:N \|}. It then follows from Theorem 5 that I α,β (θ∗ , K + N ) = I α,β (θ∗ , K) + p∗D N I(θ∗ ) for any α > 0 and any β ≥ 0. This example gives an upper bound for the increase of the Fisher information when the number of objects is increased, since it depicts the case where there is no data association uncertainty for these objects. This would correspond in practice to a case where the added objects are in an area where there is IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 19 no false alarm and where these objects are “far” from the existing objects as well as “far” from each other, where “far” depends on the likelihood. 6. Conclusion. The first important result in this article is the proof of consistency of the maximum likelihood estimator for multi-target tracking under weak conditions, where weak means that these conditions are as often as possible applying to the single-target dynamics and observation. Asymptotic normality holds under additional assumptions and the second part of the article brings understanding to the asymptotic variance of the maximum likelihood estimate by analysing the Fisher information matrix corresponding to multi-target tracking. Qualitative results are obtained in the general case, that is, the Fisher information decreases with data association uncertainty, detection failures and in the presence of false alarms. Quantitative results are also derived in important special cases: a) one static target with false alarm and unknown data association, b) multiple static targets with unknown data association under a particular observation model, and c) multiple targets with detection failures. Future works include the study of identifiability of specific observation-to-track associations, instead of marginalising over all possibilities as considered in this article. Such an approach involves additional challenges since the parameters to be learned increase in dimensionality with time, so that it is not a special case of the results presented here. Acknowledgement. All authors were supported by the Singapore ministry of education tier 1 grant number R-155-000-182-114. Appendix A. Assumptions for Theorem 3. The following assumptions are required for the proof of the asymptotic normality of the maximum P likelihood estimator in multi-target tracking. The norm k · k is defined as kM k = i,j \|Mi,j \| for any matrix M . A.6 For all K ∈ S T , all x, x0 ∈ XK and all y ∈ Y× , the mappings θ 7→ fθ (x \| x0 ) and θ 7→ gθ (y \| x) are twice continuously differentiable on the hyperplane of Θ made of parameters with a number of target equal to K. A.7 It holds that sup sup k∇θ log fθ (x \| x0 )k < ∞ and θ∈Θ x,x0 ∈XK sup sup θ∈Θ x,x0 ∈XK k∇2θ log fθ (x \| x0 )k < ∞ and that h i Ēθ∗ sup sup k∇θ log gθ (Y \| x)k < ∞ θ∈Θ x∈XK h i Ēθ∗ sup sup k∇2θ log gθ (Y \| x)k < ∞. θ∈Θ x∈XK S A.8 For all y ∈ Y× , there exists an integrable function hy : k≥0 Xk → R+ such that supθ∈Θ gθ (y \| x) ≤ hy (x). For all θ ∈ Θ and for all x ∈ XK , there exist integrable functions h1x , h2x : Y× → R+ such that k∇θ gθ (y \| x)k ≤ h1x (y) and k∇2θ gθ (y \| x)k ≤ h2x (y). Appendix B. Proof of Lemma 1. It follows from Assumption A.2 that the supremum bC + of the clutter density p(λ,ψ) characterised for any k ∈ N0 and any 20 J. HOUSSINEAU, S.S. SINGH AND A. JASRA y ∈ Yk by p(λ,ψ) (y) = Poλ (k) sup λ∈S C X Qk i=1 pψ (yi ) verifies bC + < ∞ since sup p(λ,ψ) (y) = sup k −λ X (λbC +) e λ∈S C k≥0 k k≥0 (ψ,y)∈Ψ×Y k! C = sup eλ(b+ −1) < ∞, λ∈S C and since all the terms in the sum are positive. It then holds that b̂+ ≤ sup (pD ,K,y)∈(0,1)×S T ×Y× ≤ sup (pD ,K)∈(0,1)×S T bC + K∧#y X i=0 K i K−i (pD bT + ) (1 − pD ) i T K bC + (1 − pD + pD b+ ) < ∞, which concludes the first part of the proof. For any k ∈ N0 and any y ∈ Yk Z X gθ (y \| x)dx ≥ k Y T b− (yi ) ∧ bC − (yi ) Poλ (k − \|d\|)qθ (d) d∈{0,1}K i=1 \|d\|≤k = K∧k k XY T K b− (yi ) ∧ bC − (yi ) Poλ (k − d)BipD (d) d=0 i=1 = BiK pD ∗ Poλ (k) k Y T b− (yi ) ∧ bC − (yi ) i=1 ≥ inf θ∈Θ BiK pD k Y T ∗ Poλ (k) b− (yi ) ∧ bC − (yi ) . i=1 It also holds that inf θ∈Θ BiK pD ∗ Poλ (k) > 0 for any k ∈ N0 since the support of Poλ is N0 for any λ ∈ S C which guarantees that the convolution has also N0 as a support so that the infimum is strictly greater than zero. It follows that b− (y) > 0 and, considering (8), that Ēθ∗ [\| log b− (Y )\|] < ∞. Similarly, for any k ∈ N0 and any y ∈ Yk b+ (y) ≤ sup BiK pD ∗ Poλ (k) θ∈Θ k Y T b+ (yi ) ∨ bC + (yi ) , i=1 which is finite when k is finite. In the infinite case, noticing that Poλ (k − K)pK D is the leading term in the convolution, we find that h i k T C lim sup b+ (y) ≤ sup lim KpK D Poλ (k − K) sup b+ (y) ∨ b+ (y) k→∞ y∈Yk θ∈Θ k→∞ y∈Y −λ ik−K e C sup bT (y) ∨ b (y) <∞ + + k→∞ (k − K)! y∈Y ≤ sup cθ lim θ∈Θ h where cθ is a finite constant, which concludes the proof of the lemma. Appendix C. Proof of Theorem 2. The two cases of Theorem 2 are proved separately as follows: IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 21 a) When θ ∈ Θλ=0 , the joint probability of the observations when the system is initialised with its stationary distribution is characterised by Z P̄θ (B) = 1B (y1:n ) n Y K n Y Y gθ (yt \| xt ) πθ (x0,i ) fθ (xt,i \| xt−1,i ) dy1:n dx0:n , t=1 t=1 i=1 for any measurable subset B = B1 × · · · × Bn of (Y× )n with . gθ (y \| x) = X d∈{0,1}K \|d\|=m X m Y gθ yσ(i) \| xr(i) um (σ)qθ (d) . σ∈Sym(m) i=1 for any m ∈ N0 and any (x, y) ∈ XK × Ym . Assuming that Bt is a measurable subset of YK of the form At × · · · × At for any 1 ≤ t ≤ n, then the sum over d collapses to a single term where all targets are detected and all the terms in the sum over σ are equal, so that . wθ1 = P̄θ (B) Z Y n h K i Y Kn = pD πθ (x0 ) 1At (yt,i )gθ yt,i \| xt,i fθ (xt,i \| xt−1,i ) dy1:n dx0:n . i=1 t=1 A second case that can be considered is when Bt represents the configuration where there are m ≤ K observations without considering their locations for all 1 ≤ t ≤ n, i.e. Bt = Y × · · · × Y, in which case it holds that n . wθ2,m = P̄θ (B) = BiK pD (m) If (K, pD ) 6= (K ∗ , p∗D ) then we can show that wθ2,K = wθ2,K and wθ2,K−1 = wθ2,K−1 ∗ ∗ cannot hold at the same time for any θ ∈ Θλ=0 . Alternatively, if (K, pD ) = (K ∗ , p∗D ) then wθ1 6= wθ1 ∗ follows easily from the identifiability of θ∗ . These two cases considered together show that the distributions associated to θ and θ ∗ differ in some subset of the multi-target observation space so that P̄θ 6= P̄θ∗ . b) When θ ∈ Θ\|K=1 , the multi-target likelihood becomes gθ (y \| x) = (1 − pD )Poλ (m) m Y i=1 pψ∗ (yi ) + m Y pD X gθ∗ yi \| x Poλ (m − 1) pψ∗ (yj ), m i=1 1≤j≤m j6=i for any m ∈ N0 and any (x, y) ∈ X × Ym . Marginalising over the location of the observations at each time step and considering the case where there are m observations, i.e. Bt = Y × · · · × Y, gives . wθm = P̄θ (B) = (1 − pD )Poλ (m) + pD Poλ (m − 1), m≥1 . and wθ0 = (1 − pD )e−λ . Assuming that θ 6= θ ∗ and considering that (pD , λ) 6= (p∗D , λ∗ ) it follows that wθ0 = wθ0 ∗ , wθ1 = wθ1 ∗ and wθ2 = wθ2 ∗ cannot all hold at the same time, which concludes the proof. Appendix D. Proof of Theorem 4. 22 J. HOUSSINEAU, S.S. SINGH AND A. JASRA Lemma 6. For given integers m and K, let Pθ be a family of probability measures on YmK indexed by θ ∈ Θ and let pθ denote the corresponding probability density w.r.t. a common reference measure, for all θ, on YmK . Assume that pθ (y1 , . . . , ym ) > 0 for any θ and (y1 , . . . , ym ). For any integers α ≥ 1 and β ≥ 0, let the random vectors (Y10 , . . . , Ym0 ) be conα,β α,β ditionally independent given (Y1 , . . . , Ym ), with law PYα,β 0 \|Y1:m = PY 0 \|Y1 . . . PY 0 \|Ym 1:m 1 m is defined as in (15) via a process of thinning, augmentation with and each PYα,β 0 i \|Yi clutter with density pψ on Y and random permutation. Assume pψ > 0. 1. Consider any θ and (α, β) such that α > 1 or β > 0. If f (Y1 , . . . , Ym ) = E[f (Y1 , . . . , Ym )\|Y10 , . . . , Ym0 ] then f (Y1 , . . . , Ym ) is constant almost surely. 2. Let the probability measure of (Y10 , . . . , Ym0 ) be Pθα,β and its corresponding α,β probability density be pα,β are differθ . Assume that the densities pθ and pθ entiable w.r.t. θ then (32) E[∇θ log pθ (Y1 , . . . , Ym ) · ∇θ log pθ (Y1 , . . . , Ym )t ] α,β 0 0 0 0 t ≥ E[∇θ log pα,β θ (Y1 , . . . , Ym ) · ∇θ log pθ (Y1 , . . . , Ym ) ], with the inequality being strict if and only if α > 1 or β > 0 and if the l.h.s. is strictly greater than 0. Proof of Part 1 of Lemma 6. When α = 1 and β > 0, which corresponds to no random permutation (or no observation association uncertainty) but only random thinning, the result follows from Lemma 11 and Remark 4. When α > 1 and β = 0, which corresponds to no random thinning but only random permutation, the result follows from Corollary 9. When α ≥ 1 and β > 0, i.e. both random thinning and random permutation are present, the result of Lemma 11 based on YD , i.e. for (α = 1, β > 0), can be extended to the random variable YD0 = Sς (YD ⊕ Ŷ ) as follows: 1. By noticing that σ(YD0 ) ⊆ σ(ς, YD , Ŷ ) (i.e. the σ-algebra generated by YD0 is . coarser than the one generated by ȲD = (ς, YD , Ŷ )). Indeed, for any C = B1 × B2 × · · · ∈ B(Y× ) with all but finitely many Bi equal to Y, it holds that [ YD0−1 (C) = ȲD−1 ({σ} × Bσ(1) × Bσ(2) × . . . ) ∈ σ(ȲD ) σ where the inclusion follows by countable union. The family of subsets of the same form as C is a generating family of B(Y× ) from which the result follows. The result extends straightforwardly to any collection of m thinned observation vectors being independently perturbed by false alarm and permutation since the corresponding σalgebra is generated by the set of rectangles of the form C1 ×· · ·×Cm with Ci ∈ B(Y× ). 2. As a consequence of 1, the fact that f (Y ) = E[f (Y ) \| YD0 ] a.s. implies that E[f (Y ) \| ȲD ] = E E[f (Y ) \| YD0 ] \| ȲD = E[f (Y ) \| YD0 ] a.s., so that, since Ŷ , ς and YD are independent, it holds that f (Y ) = E[f (Y ) \| ȲD ] = E[f (Y ) \| YD ] a.s., which under the assumption of Lemma 11 implies that f (Y ) is constant almost surely. As in 1, the same result holds for collection of m observation vectors. IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 23 Proof of Part 2 of Lemma 6. Let Y be the K measurements of K targets, Ŷ be the clutter, ς the random permutation, D the K dimensional vector of deletions. Let Y 0 = Sς ((RD Y ) ⊕ Ŷ ). The missing target generated observations are Ym = RD0 Y where D 0 = 1 − D. Let pY 0 ,ς,D,Ym (y 0 , σ, d, ym ) denote the joint pdf/pmf of (Y 0 , Ym , D, ς) that depends implicitly on θ. Using the change of variable formula, noting that (Y , Ŷ ) = F (Y 0 , Ym , D, ς) where the mapping F is a permutation of (Y 0 , Ym ) and hence the Jacobian of the transformation has determinant 1, it follows that pY 0 ,Ym ,D,ς (y 0 , ym , d, σ) = pY ,Ŷ ,D,ς (F (y 0 , ym , d, σ), d, σ) for y 0 = Sσ ((Rd y) ⊕ ŷ), ym = Rd0 y where d0 = 1 − d. ∇θ log pY ,Ŷ ,D,ς (y, ŷ, d, σ) = ∇θ log pY (y), we deduce that Since it holds that ∇θ log pY 0 ,Ym ,D,ς (y 0 , ym , d, σ) = ∇θ log pY (FT (y 0 , ym , d, σ)) where FT (y 0 , ym , d, σ) is the projection of F (y 0 , ym , d, σ) on the coordinates describing Y . Now it follows that, almost surely, ∇θ log pY 0 ,Ym ,D,ς (Y 0 , Ym , D, ς) = ∇θ log pY (Y ). Let pY 0 denote the density of Y 0 . Then ∇θ log pY 0 (Y 0 ) = EPθ [∇θ log pY (Y ) \| Y 0 ]. (33) 0 Applying (33) to the joint random variables Y1:m and Y1:m defined in the lemma, it follows that 0 0 ∇θ log pα,β θ (Y1:m ) = EPθ ∇θ log pθ (Y1:m ) \| Y1:m . Let v ∈ RdΘ , then Jensen’s inequality applied to the function x 7→ x2 and to the random variable v t ∇θ log pθ (Y1:m ) yields 2 2 0 0 EPθ v t ∇θ log pθ (Y1:m ) \| Y1:m ≤ EPθ v t ∇θ log pθ (Y1:m ) \| Y1:m almost surely, so that α,β t v t EP α,β [∇θ log pα,β θ (Y1:m ) · ∇θ log pθ (Y1:m ) ]v θ 2 0 = EP α,β EPθ v t ∇θ log pθ (Y1:m ) \| Y1:m θ t ≤ v EPθ [∇θ log pθ (Y1:m ) · ∇θ log pθ (Y1:m )t ]v which proves (32). Since Jensen’s inequality has been applied to a strictly convex function, the case of equality: α,β t v t EP α,β [∇θ log pα,β θ (Y1:m ) · ∇θ log pθ (Y1:m ) ]v θ = v t EPθ [∇θ log pθ (Y1:m ) · ∇θ log pθ (Y1:m )t ]v, 0 holds if and only if, for all v ∈ RdΘ , v t ∇θ log pθ (Y1:m ) = E[v t ∇θ log pθ (Y1:m ) \| Y1:m ] 0 0 is σ(Y1:m )-measurable. Part 1 of the lemma shows that ∇θ log pθ (Y1:m ) is σ(Y1:m )measurable if and only if it is constant a.s.. Given that EPθ [∇θ log pθ (Y1:m )] = 0 it follows that the function itself is equal to 0 since it is constant, hence proving the lemma. 24 J. HOUSSINEAU, S.S. SINGH AND A. JASRA The observation model (15) does not imply the equality of the gradients of log pθ (Y 0 , Y ) and log pθ (Y ) w.r.t. θ since Pθ (dY 0 \| Y ) depends on pD , λ and ψ which are parameters included in θ. The interest is however in the information loss w.r.t. θ so that the result of Lemma 6 is satisfying. Proof of Theorem 4. The considered perturbed observation model has the same α,β properties as the one studied in [3], i.e. that ∇θ log pα,β ) = ∇θ log pθ (Y ) θ (Y , Y holds almost surely. The result of [3, Lemma 3] and [3, Remark 9] can therefore be used directly in the context of interest to give h i (m) α,β ∗ ∗ Iloss (θ ) = Ēθ∗ I α,β (θ ) , Y−∞:−1 ,Ym:∞ α,β where, for any Y−∞:−1 , any Ym:∞ and any integer m ≥ 1, I (m) ∗ α,β (θ ) Y∞:−1 ,Ym:∞ = h 1 α,β Ēθ∗ ∇θ log pθ∗ Y0:m−1 \| Y−∞:−1 , Ym:∞ · m i α,β t α,β ∇θ log pθ∗ Y0:m−1 \| Y−∞:−1 , Ym:∞ Y−∞:−1 , Ym:∞ h 1 α,β − Ēθ∗ ∇θ log pθ∗ Y0:m−1 \| Y−∞:−1 , Ym:∞ · m i α,β t α,β ∇θ log pθ∗ Y0:m−1 \| Y−∞:−1 , Ym:∞ Y−∞:−1 , Ym:∞ . The objective is to prove that (36) h (m) Ēθ∗ I α,β Y−∞:−1 ,Ym:∞ i (θ∗ ) = 0 for all m ≥ 1 implies that I(θ∗ ) = 0. From Lemma 6 applied to the involved conα,β ) = 0 almost surely ditional laws, (36) implies that ∇θ log pθ∗ (Y0:m−1 \| Y−∞:−1 , Ym:∞ α,β for almost all Y−∞:−1 and almost all Ym:∞ . Following the same principle as in [3, Lemma 4], it follows that if (36) holds for all m ≥ 1 then ∇θ log pθ∗ (Y0 \| Y−∞:−1 ) = 0 almost surely, which in turn implies that I(θ∗ ) = 0. D.1. Supporting results for the proof of Part 1 of Lemma 6. D.1.1. Supporting results for the proof of Lemma 6 for permutation uncertainty but no deletion, i.e. (α > 1, β = 0). Lemma 7 (Randomly permuting a random vector.). Let Y = (Y1 , . . . , Yn ) and Ŷ = (Yn+1 , . . . , Yn+m ). Let ς denote the randomised permutation which is independent of (Y , Ŷ ) and let Z = (Yς(1) , . . . , Yς(m+n) ). Assume ς permits, at the least, the exchange of any two indices, i.e. P(ς(i) = j, ς(j) = i, {ς(k) = k : k 6= i, j}) > 0 for all i, j. Furthermore, P(ς = (1, . . . , n)) > 0. Assume the law of (Y , Ŷ ) satisfies ν n+m PY ,Ŷ ν n+m where ν is some probability measure and ν n+m the product probability measure on Yn+m . If f (Y ) = E[f (Y )\|Z], then f (Y ) is a constant almost surely. Proof of Lemma 7. The proof is completed for the case m = 1 and easily generalised to m > 1. Likewise, to present the arguments we employ in the clearest way, we consider the case (n = 2, m = 1). (The case n + m = 2 sheds too little light to make the generalisation apparent.) IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 25 Let g(Z) = E[f (Y )\|Z]. For any σ such that P(ς = σ) > 0, E[\|f (Y ) − g(Yσ(1) , . . . , Yσ(n+1) )\|I[ς=σ] ] = 0. Since ς is independent of (Y , Ŷ ), we have E[\|f (Y ) − g(Yσ(1) , . . . , Yσ(n+1) )\|] = 0 and thus g(Yσ(1) , . . . , Yσ(n+1) ) = g(Yσ0 (1) , . . . , Yσ0 (n+1) ) almost surely for any other σ 0 such that P(ς = σ 0 ) > 0. Now consider the case (n = 3, m = 1). The preceding statements imply, PY ,Y almost everywhere, (37) g(Y1 , Y2 , Y3 ) = f (Y1 , Y2 ) (38) g(Y1 , Y2 , Y3 ) = g(Y3 , Y2 , Y1 ) (39) g(Y1 , Y2 , Y3 ) = g(Y1 , Y3 , Y2 ). Indeed these also hold ν 3 almost every. (Due to the assumption of mutual absolutely continuity, statements holds PY ,Y almost everywhere if and only if they hold ν 3 almost every.) We will show that the further implication (40) f (Y1 , Y2 ) = f (Y3 , Y2 ) = f (Y1 , Y3 ) holds ν 3 almost everywhere may be derived. Once this is done, to complete the proof, we will further manipulate (40) under the assumption that the random variables Yi are independently and identically distributed with respect to measure ν to show that f = c, for some constant c, ν 3 almost everywhere. From the first equality of (40), f (Y1 , Y2 ) = Eν 3 (f (Y3 , Y2 )\|Y1 , Y2 ) = Eν 3 (f (Y3 , Y2 )\|Y2 ) = h(Y2 ) for some function h. That is f (Y1 , Y2 ) collapses to a function of variable Y2 only, which is denoted by h(Y2 ). Using the second equality of (40), h(Y2 ) = f (Y1 , Y3 ) and thus it must be that h is a constant as Yi are independent. We now verify (37)-(38) implies f (Y1 , Y2 ) = f (Y3 , Y2 ) of (40). We have Eν 3 [\|f (Y1 , Y2 ) − g(Y3 , Y2 , Y1 )\|] = 0 and a change of variable gives Eν 3 [\|f (Y3 , Y2 )−g(Y1 , Y2 , Y3 )\|] = 0. The same procedure applied to (37)-(39) shows the second equality of (40). Corollary 8 extends Lemma 7 to the situation when Ŷ therein follows the law of a clutter process as defined in Section 5.1 (see 15.) Corollary 8. Let (Ŷ1 , Ŷ2 , . . .) be an infinite sequence of independent Y-valued random variables with Ŷi ∼ Pψ . Let Y = (Y1 , . . . , YK ) be a vector of Y-valued random variables which is independent of (Ŷ1 , Ŷ2 , . . .). Let M̂ ∈ N0 be non-negative random variable independent of (Y , Ŷ1:∞ ). Let Z = Sς (Y ⊕ Ŷ ) where Ŷ = Ŷ1:M̂ and Sς is the random permutation matrix defined as in (15) of Section 5.1. Assume (Pψ )K PY (Pψ )K . If f (Y ) = E[f (Y )\|Z], then f (Y ) is a constant almost surely. Proof. Let g(Z) = E[f (Y )\|Z] then E[\|f (Y ) − g(Z)\|\|M̂ = m] = 0 for all m such that P(M̂ = m) > 0. Since M̂ is independent of (Y , Ŷ1:∞ ) and the random permutation matrix is itself independent of (Y , Ŷ1:∞ ) given M̂ = m, the law of (Y , Z) 26 J. HOUSSINEAU, S.S. SINGH AND A. JASRA conditioned on M̂ = m satisfies the assumptions of Lemma 7. Thus, by Lemma 7, E[\|f (Y ) − g(Z)\|\|M̂ = m] = 0 implies f (Y ) = cm almost surely for some constant cm . (It is clear that cm is independent of m.) The next results extends Corollary 8 to the setting where a sequence of vectors (Y1 , . . . , Ym ), with Yi ∈ YK , is observed indirectly through the sequence of vectors (Z1 , . . . , Zm ) where each Zi is generated as in Corollary 8 by augmenting Yi with clutter and then randomly permuting it. Corollary 9. Let (Y1 , . . . , Ym ) be a sequence of random vectors with Yi ∈ YK . Let (Z1 , . . . , Zm ) be conditionally independent given (Y1 , . . . , Ym ), i.e. PZ1:m \|Y1:m = PZ1 \|Y1 . . . PZm \|Ym and each PZi \|Yi is defined as in Corollary 8. Assume (41) (Pψ )K+m PY1:m (Pψ )K+m If f (Y1:m ) = E[f (Y1:m )\|Z1:m ], then f (Y1:m ) is a constant almost surely. In particular, (41) holds if both (Y1 , . . . , Ym ) and Pψ have positively valued probability densities, i.e. p(y1 , . . . , ym ) > 0 and pψ (y) > 0, w.r.t. the dominating measures ν K+m and ν for some measure ν on Y. Proof. Due to the conditional independence assumption, E[f (Y1:m )\|Z1 , Y2:m ] = E[f (Y1:m )\|Z1 , Y2:m , Z2:m ] and since f (Y1:m ) = E[f (Y1:m )\|Z1:m ], we have f (Y1:m ) = E[f (Y1:m )\|Z1 , Y2:m ] or f (Y1 , y2:m ) = E[f (Y1:m )\|Z1 , y2:m ] PY1 ,Z1 \|y2:m − a.s. for PY2:m almost all y2:m . Invoking Corollary 8 we thus have f (Y1 , Y2:m ) = g1 (Y2:m ) a.s. for some function g1 . Proceeding in this way we can similarly show f (Y1:i−1 , Yi , Yi+1:m ) = gi (Y1:i−1 , Yi+1:m ) for i = 1, . . . , m. We can now invoke assumption (41) to show that f (Y1:m ) must be a constant almost surely in a manner similar to the proof of Lemma 7; details omitted. The final statement of the corollary is straightforward. D.1.2. Supporting results for the proof of Lemma 6 for deletion but no permutation uncertainty, i.e. (α = 1, β > 0). The following result is an intermediate result intended to convey the main idea of the analysis while the actual full blown version is contained in Lemma 11 below. Lemma 10. Let Y = (Y1 , Y2 ), D ∈ {1, 2}, Y and D are independent and 0 < P(D = 1) < 1. Assume for any measurable set C that P(Y1 ∈ C\|Y2 ∈ C) < 1 whenever 0 < P(Y1 ∈ C) < 1. Then f (Y ) = E[f (Y )\|YD ] implies f (Y ) = c almost surely for some constant c. Remark 1. The random variables (Y1 , Y2 ) can always be relabelled if the assumption is met by P(Y2 ∈ ·\|Y1 ∈ ·). The assumption is to be interpreted to mean that if Y1 is not concentrated in C then the event Y2 ∈ C does not imply Y1 ∈ C. This assures some independence between the random variables. Remark 2. The assumption is satisfied without the need of densities. For example the assumption is satisfied when Y1 and Y2 are independent. If Y = (Y1 , Y2 ) has a density, then the assumption is satisfied if p(y1 , y2 ) > 0. Indeed in this case P(Y1 ∈ C c , Y2 ∈ C) > 0 for all C ⊂ Y. IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 27 Proof of Lemma 10. The conditional expectation can always be expressed as follows: there exists a function g such that E[f (Y )\|YD ] = g(YD ) or g(YD ) is a version of the conditional expectation. Thus we have 0 = E[\|f (Y ) − g(YD )\|] = P(D = 1)E[\|f (Y ) − g(Y1 )\|] + P(D = 2)E[\|f (Y ) − g(Y2 )\|] which follows from the independence of D and Y . Since 0 < P(D = 1) < 1, both expectation terms must be zero and hence g(Y1 ) = g(Y2 ) almost surely: E[\|g(Y1 ) − g(Y2 )\|] ≤ E[\|f (Y ) − g(Y1 )\|] + E[\|f (Y ) − g(Y2 )\|] = 0. For any measurable set B, let A = g −1 (B). Since g(Y1 ) = g(Y2 ) almost surely, we have P(g(Y1 ) ∈ B) = P(g(Y1 ) ∈ B, g(Y2 ) ∈ B) = P(g(Y2 ) ∈ B) or that (42) P(Y1 ∈ A) = P(Y1 ∈ A, Y2 ∈ A) = P(Y2 ∈ A). We can now show that g(Y1 ) = c, for some constant c, almost surely. (Assuming the converse is true we will arrive at a contradiction.) If g(Y1 ) = c does not hold almost surely then there exist a set B such that 0 < P(g(Y1 ) ∈ B) < 1 and let A = g −1 (B) and thus 0 < P(Y1 ∈ A) < 1. But (42) implies P(Y1 ∈ A\|Y2 ∈ A) = 1 which violates the assumption. Lemma 11 (Multiple deletions for K > 2.). Let Y = (Y1 , . . . , YK ) be a random vector, D ⊆ {1, . . . , K} and YD denote the thinned version where components not in D have been removed. Assume 0 < P(D = σ) < 1 for all subsets σ ⊂ {1, . . . , K} such that \|σ\| = K − 1. Furthermore, assume the following: • For each i, j ∈ {1, . . . , K}, i 6= j, let Z ⊆ Y1:K\{i,j} . If P((Yi , Z) ∈ A) < 1 and IA (Yi , Z) 6= E[IA (Yi , Z)\|Z] then P((Yi , Z) ∈ A\|(Yj , Z) ∈ A) < 1. (Here f 6= g means P(f 6= g) > 0.) Then f (Y ) = E[f (Y )\|YD ] implies f (Y ) = c almost surely for some constant c. Remark 3. The assumption of Lemma 11 is more easily understood as follows: for each i, j ∈ {1, . . . , K}, i 6= j, let Z ⊆ Y1:K\{1,j} . If P(Yi ∈ A\|Z ∈ B) < 1 then P(Yi ∈ A\|Z ∈ B, Yj ∈ A) < 1 Unfortunately stating the assumption via (measurable) rectangle sets is not sufficient as the product sigma algebra is much richer than just measurable rectangles. In fact, if the set A of IA (Yi , Z) in the assumption of Lemma 11 is the measurable rectangle IA (Yi )IB (Z) then the assumption reduces to that in this remark. Remark 4. The main assumption of Lemma 11 is satisfied if ν1 × · · · × νK PY ν1 × · · · × νK where νi are probability measures, i.e. PY and ν1 × · · · × νK are mutually absolutely continuous. (The proof of this fact uses Lemma 12 and the fact that trace of a set is measurable; this is omitted.) Proof of Remark 3. The condition IA (Yi , Z) 6= E[IA (Yi , Z)\|Z] translates to IA (Yi )IB (Z) 6= E[IA (Yi )IB (Z)\|Z] and from Lemma 12 this implies that for all measurable sets C in the range of Z E[IA (Yi )IB (Z)] 6= E[IC (Z)]. Taking C = B this then implies P(Yi ∈ A, Z ∈ B) < P(Z ∈ B) and so P(Yi ∈ A\|Z ∈ B) < 1. 28 J. HOUSSINEAU, S.S. SINGH AND A. JASRA Clearly the probability P((Yi , Z) ∈ A\|(Yj , Z) ∈ A) of the assumption of Lemma 11 for the rectangle is P(Yi ∈ A\|Yj ∈ A, Z ∈ B). Proof of Lemma 11. The random variable YD belongs to Y× , that is to the disk 0 joint union ∪K k=0 Y with Y ≡ ∅. Thus we can write 0 = E[\|f (Y ) − g(YD )\|] = K X X E[\|f (Y ) − gi (Yσ )\|I[D=σ] ] i=0 σ:\|σ\|=i = K X X E[\|f (Y ) − gi (Yσ )\|]P(D = σ) i=0 σ:\|σ\|=i where g0 is a constant, gi : Yi → R are measurable functions and independence of D and Y has been invoked. If P(D = ∅) > 0 then it is trivial since this implies E[\|f (Y ) − g0 \|] = 0. So assume P(D = ∅) = 0. Having assumed 0 < P(D = σ) < 1 for all subsets σ ⊂ {1, . . . , K} such that \|σ\| = K − 1, we focus on these terms only: X E[\|f (Y ) − gK−1 (Yσ )\|]P(D = σ) σ:\|σ\|=K−1 which also implies (43) gK−1 (Yσ ) = gK−1 (Yσ0 ) or IA (Yσ ) = IA (Yσ0 ) almost surely −1 for all σ, σ 0 and A = gK−1 (B) for a measurable set B in R. For example, when σ = (1, 3, . . . , K), σ = (2, 3, . . . , K) and Z = (Y3 , . . . , YK ), we get P((Y1 , Z) ∈ A) = P((Y1 , Z) ∈ A, (Y2 , Z) ∈ A) = P((Y2 , Z) ∈ A). Henceforth we refer to gK−1 simply as g. We need to show that g(Yσ ) = c, for some constant c, almost surely. If this is not the case then there exists subsets of variables Yi ∈ Yσ , Z ⊂ Yσ and Yi ∈ / Z (recall σ ⊂ {1, . . . , K} with \|σ\| = K − 1) such that (44) g(Yσ ) = E[g(Yσ )\|Yi , Z] a.s. and g(Yσ ) 6= E[g(Yσ )\|Z] a.s. The interpretation is that g(Yσ ) can potentially be a function of the reduced set of variables (Yi , Z) (as asserted by the first equality) but it must genuinely be a function of at least the variable Yi . For clarity and simplicity assume i = 1 and Z = (Y3 , . . . , YK ). Consider the terms in the sum due to σ = (1, 3, . . . , K) and σ = (2, 3, . . . , K). By Lemma 13 below, there exists a measurable set A = g −1 (B) such that 0 < P(A) < 1 and IA (Y1 , Z) 6= E[IA (Y1 , Z)\|Z]. But (43) implies P((Y1 , Z) ∈ A\|(Y2 , Z) ∈ A) = 1 which violates the main assumption of the lemma. The following two lemmas concerns the random variables declared in the statement of Lemma 11 Lemma 12. IA (Yi , Z) = E[IA (Yi , Z)\|Z] a.s. if and only if there exits a set C in σ(Z) such that IA (Yi , Z) = IC almost surely. IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 29 Proof. Let F denote E[IA (Yi , Z)\|Z] and let C = {ω : F (ω) = 1}. Clearly C ∈ σ(Z) and we show IA (Yi , Z) = IC almost surely. The set IA (Yi , Z) 6= F has measure zero and is comprised of the disjoint sets {ω : IA (Yi , Z) = 1, F 6= 1}, {ω : IA (Yi , Z) = 0, F = 1} and {ω : IA (Yi , Z) = 0, F 6= 0, F 6= 1}. Thus the result follows since IC and IA (Yi , Z) differ precisely on the first two sets. (The reverse implication is obvious.) Lemma 13. If g(Yi , Z) 6= E[g(Yi , Z)\|Z] then there exists a measurable function IA (Yi , Z) such that 0 < P((Yi , Z) ∈ A) < 1 and IA (Yi , Z) 6= E[IA (Yi , Z)\|Z]. Proof. The proof is complete once we show that g(Yi , Z) 6= E[g(Yi , Z)\|Z] implies there exists a set A0 ∈ σ(g) such that IA0 6= E[IA0 \|Z], which in turn implies 0 < P(A0 ) < 1. (Here we write σ(g), rather than σ(g(Yi , Z)), for the smallest σ-algebra with respect to which the random variable g(Yi , Z) is measurable.) Note that for every A0 ∈ σ(g) there exists a measurable set B in R such that A0 = {ω : (Yi (ω), Z(ω)) ∈ g −1 (B)}. Assume that IA0 = E[IA0 \|Z] almost surely for all A0 ∈ σ(g). Then by Lemma 12, for each A0 ∈ σ(g) there is a set C in σ(Z) such that IA0 = IC almost surely. This will violate g(Yi , Z) 6= E[g(Yi , Z)\|Z] if g(Yi , Z) = IA0 and hence also if g was a simple function; if we replace each indicator function (measurable set) of the simple function with its almost sure counterpart from σ(Z) then it becomes σ(Z) measurable. In the general case, the result may be established since g may be approximated by a sequence of simple functions {gn } tending to g and each simple function gn itself has a σ(Z) measurable version, say g n , and g n = gn . Thus E[gn \|Z] = E[g n \|Z] = gn . Letting g = E[g(Yi , Z)\|Z], we see that E[\|g − g\|] = lim E \|gn − E[gn \|Z]\| = 0. Appendix E. Proof of Theorem 5. The case where K ∗ = 1 is first considered so that the number of observations Mt at time t can only be equal to zero or one. The joint probability of the observations and states becomes p̄θ (y1:n , x0:n ) = πθ (x0 ) n Y (1 − pD )1−Mt (pD gθ (yt \| xt ))Mt fθ (xt \| xt−1 ) , t=1 where yt is the empty sequence when Mt = 0. The size of yt at any time t can be made explicit in this expression for the sake of clarity as follows p̄θ (y1:n , x0:n , m1:n ) = πθ (x0 ) n Y (1 − pD )1−mt (pD gθ (yt \| xt ))mt fθ (xt \| xt−1 ) . t=1 Yt Let be a noisy version of the original observation Yt for any t ≥ 1 so that the HMM (Xt , Yt )t is equal in law to the HMM (Xt , Yt + Zt )t where (Zt )t is an i.i.d. sequence of random variables which common law is the uniform distribution over the ball of radius 1 and centre 0. A switching process (st )t is also introduced as follows: st = 1 when the target is detected and st = 0 otherwise. In order to study the Fisher information more easily, we introduce an alternative observation model where a detection failure at time t is replaced by an observation Yt from the target. The law of this observation model is n Y st p̄θ (ỹ1:n , x0:n , s1:n ) = πθ (x0 ) pD (1 − pD )1−st gθ (ỹt \| xt )fθ (xt \| xt−1 ) = πθ (x0 ) t=1 n h Y t=1 i [pD gθ (yt \| xt )]st [(1 − pD )gθ (yt \| xt )]1−st fθ (xt \| xt−1 ) 30 J. HOUSSINEAU, S.S. SINGH AND A. JASRA where ỹt = yt if st = 1 and ỹt = yt if st = 0. The quantity of interest is p̄θ (ỹ0 , s0 \| y−∞:−1 , y1:∞ )= ), [pD gθ (y0 \| x0 )]s0 [(1 − pD )gθ (y0 \| x0 )]1−s0 p̄θ (x0 \| y−∞:−1 , y1:∞ which we compare with the full-detection case p̄θ (y0 \| y−∞:−1 , y1:∞ ) = gθ (y0 \| x0 )p̄θ (x0 \| y−∞:−1 , y1:∞ ). To justify the equivalence of the two observation model for the considered purpose, we can verify that the score ∇θ log p̄θ (y0 , m0 \| y−∞:−1 ) is equal to the score ∇θ log p̄θ (ỹ0 , s0 \| y−∞:−1 , y1:∞ ) when → ∞. With the required modifications and after [3, Theorem 5], it follows that the loss of information Iloss (θ∗ ) when replacing the original observations by the -perturbed ones can be expressed as h i Iloss (θ∗ ) = Ēθ∗ ∇θ log p̄θ∗ (Y0 \| Y−∞:−1 , Y1:∞ ) · ∇θ log p̄θ∗ (Y0 \| Y−∞:−1 , Y1:∞ )t h i − pD Ēθ∗ ∇θ log p̄θ∗ (Y0 \| Y−∞:−1 , Y1:∞ ) · ∇θ log p̄θ∗ (Y0 \| Y−∞:−1 , Y1:∞ )t h i − (1 − pD )Ēθ∗ ∇θ log p̄θ∗ (Y0 \| Y−∞:−1 , Y1:∞ ) · ∇θ log p̄θ∗ (Y0 \| Y−∞:−1 , Y1:∞ )t . Considering the limit → ∞, it follows that . Iloss (θ∗ ) = lim Iloss (θ∗ ) →∞ h i = (1 − pD )Ēθ∗ ∇θ log p̄θ∗ (Y0 \| Y−∞:−1 ) · ∇θ log p̄θ∗ (Y0 \| Y−∞:−1 )t . In the multi-target case, it simply holds that the information loss is equal to (1 − p∗D )K ∗ I(θ∗ ) since targets’ detection are independent when the data association is known, which terminates the proof of the proposition. REFERENCES [1] S. S. Blackman, Multiple-target tracking with radar applications, Artech House, 1986. [2] N. Chenouard et al., Objective comparison of particle tracking methods, Nature methods, 11 (2014). [3] T. A. Dean, S. S. Singh, A. Jasra, and G. W. Peters, Parameter estimation for hidden Markov models with intractable likelihoods, Scandinavian Journal of Statistics, 41 (2014). [4] P. Del Moral, Mean field simulation for Monte Carlo integration, 2013. [5] P. Del Moral and J. Houssineau, Particle association measures and multiple target tracking, in Theoretical Aspects of Spatial-Temporal Modeling, Springer, 2015, pp. 1–30. [6] R. Douc, E. Moulines, and T. Ryden, Asymptotic properties of the maximum likelihood estimator in autoregressive models with Markov regime, The Annals of Statistics, 32 (2004). [7] A. Doucet, B.-N. Vo, C. Andrieu, and M. Davy, Particle filtering for multi-target tracking and sensor management, in Proceedings of the Fifth International Conference on Information Fusion, vol. 1, IEEE, 2002, pp. 474–481. [8] L. Jiang, S. S. Singh, and S. Yıldırım, Bayesian tracking and parameter learning for nonlinear multiple target tracking models, IEEE Transactions on Signal Processing, 63 (2015). [9] B. G. Leroux, Maximum-likelihood estimation for hidden Markov models, Stochastic processes and their applications, 40 (1992). [10] R. P. S. Mahler, Multitarget Bayes filtering via first-order multitarget moments, IEEE Transactions on Aerospace and Electronic Systems, 39 (2003). [11] S. P. Meyn and R. L. Tweedie, Markov chains and stochastic stability, Springer Science & Business Media, 2012. IDENTIFICATION OF MULTI-OBJECT DYNAMICAL SYSTEMS 31 [12] J. Mullane, B.-N. Vo, M. D. Adams, and B.-T. Vo, A random-finite-set approach to Bayesian SLAM, IEEE Transactions on Robotics, 27 (2011). [13] S. Oh, S. Russell, and S. Sastry, Markov chain Monte Carlo data association for multi-target tracking, IEEE Transactions on Automatic Control, 54 (2009). [14] K. Okuma, A. Taleghani, N. De Freitas, J. J. Little, and D. G. Lowe, A boosted particle filter: Multitarget detection and tracking, in European Conference on Computer Vision, Springer, 2004, pp. 28–39. [15] M. Pace and P. Del Moral, Mean-field PHD filters based on generalized Feynman-Kac flow, Journal of selected topics in signal processing, special issue on multi-target tracking, (2013). [16] Y. Pailhas, J. Houssineau, Y. Petillot, and D. E. Clark, Tracking with MIMO sonar systems: applications to harbour surveillance, IET Radar, Sonar & Navigation, (2016). [17] C. B. Storlie, J. Hannig, T. C. Lee, et al., Statistical consistency of the data association problem in multiple target tracking, Electronic Journal of Statistics, 5 (2011). [18] B.-N. Vo and W.-K. Ma, The Gaussian mixture probability hypothesis density filter, IEEE Transactions on Signal Processing, 54 (2006). [19] B.-N. Vo, S. Singh, and A. Doucet, Sequential Monte Carlo methods for multitarget filtering with random finite sets, IEEE Transactions on Aerospace and Electronic Systems, 41 (2005).	10math.ST
1 Advanced Quantizer Designs for FDD-based FD-MIMO Systems Using Uniform Planar Arrays arXiv:1703.10739v3 [cs.IT] 3 Nov 2017 Jiho Song, Member, IEEE, Junil Choi, Member, IEEE, Taeyoung Kim, Member, IEEE, and David J. Love, Fellow, IEEE UPA structure is known as full-dimension (FD) MIMO because of its ability to exploit both vertical and horizontal domain beamforming [5]–[10]. To fully exploit FD-MIMO, accurate channel state information (CSI) for both domains is critical. Channel estimation techniques relying upon time division duplexing (TDD) can leverage channel reciprocity if the transmit and receive arrays are calibrated [11], [12]. Most current cellular systems, however, exploit frequency division duplexing (FDD) where the receiver should estimate, quantize, and feed back the downlink CSI to the transmitter. The high dimensionality of a massive MIMO channel could cause large overheads for downlink channel training and quantization processes [13]–[19]. We focus on CSI quantization for FD-MIMO in this paper and refer to [18]–[20] and the references therein for the massive MIMO downlink training problem. The majority of CSI quantization codebooks have been designed under the assumption of spatially uncorrelated Rayleigh fading channels, which are uniformly distributed on the unit hypersphere when normalized. To quantize these channels, the codewords in a codebook should cover the unit sphere Index Terms—Massive MIMO, full-dimension MIMO, uniform as uniformly as possible [21]–[23]. For spatially correlated channels, the codebooks have been carefully shaped based on planar arrays, Kronecker product codebooks. the prior knowledge of channel statistics [24]–[28]. Although most previous codebooks have been designed based I. I NTRODUCTION on analytical channel models, it is difficult to represent the ASSIVE multiple-input multiple-output (MIMO) sys- properties of true three-dimensional (3D) channels for FDtems are a strong candidate to fulfill the throughput MIMO. The 3D spatial channel model (SCM) in [4], [29], requirements for fifth generation (5G) cellular networks [2]. which is an extension of the 2D SCM [30], has been extensively To maximize the number of antennas in a limited area, two- used to mimic the measured channel variations for 3GPP dimensional antenna arrays, e.g., uniform planar arrays (UPAs) standardization. Although the 3D SCM is a stochastic channel and cylindrical arrays, that host antennas in both vertical and model, it provides limited insights into practical CSI quantizer horizontal domains are being prominently considered in practice designs. Therefore, it is necessary to develop a simple channel [3], [4]. Among various 2D array solutions, UPAs are of great model that accurately represents the properties of the 3D-SCM interest to simplify signal processing for three-dimensional channel with UPA antennas. In this paper, we define a simple 3D channel model using the (3D) channels for FD-MIMO. Massive MIMO employing a sum of a finite number of scaled array response vectors. Based J. Song and D. J. Love are with the School of Electrical and Computer on this simplified channel model, we develop CSI quantizers Engineering, Purdue University, West Lafayette, IN 47907 (e-mail: {jihosong, for UPA scenarios. We first carry out performance analysis of djlove}@purdue.edu). J. Choi is with the Department of Electrical Engineering, POSTECH, Pohang, Kronecker product (KP) CSI quantizers [1], [9]. Our analytical Gyeongbuk 37673, Korea (e-mail:junil@postech.ac.kr). studies on the KP CSI quantizers provide design guidelines on T. Kim, is with Samsung Electronics Co., Ltd., Suwon, Korea how to develop a quantizer for a narrowband, single frequency (email:ty33@samsung.com). Parts of this paper were presented at the Globecom, Washington, DC USA, tone CSI using limited feedback resources. In the proposed December 4-8, 2016 [1]. quantizer, we concentrate on detecting/quantizing dominant This work was supported in part by Communications Research Team (CRT), radio paths in true channels. To maximize quantization quality, Samsung Electronics Co. Ltd., and the ICT R&D program of MSIP/IITP [2017(B0717-17-0002), Development of Integer-Forcing MIMO Transceivers we also develop a codebook for combiners that cophases and for 5G & Beyond Mobile Communication Systems]. scales the quantized beams. Both vertical and horizontal doAbstract—Massive multiple-input multiple-output (MIMO) systems, which utilize a large number of antennas at the base station, are expected to enhance network throughput by enabling improved multiuser MIMO techniques. To deploy many antennas in reasonable form factors, base stations are expected to employ antenna arrays in both horizontal and vertical dimensions, which is known as full-dimension (FD) MIMO. The most popular two-dimensional array is the uniform planar array (UPA), where antennas are placed in a grid pattern. To exploit the full benefit of massive MIMO in frequency division duplexing (FDD), the downlink channel state information (CSI) should be estimated, quantized, and fed back from the receiver to the transmitter. However, it is difficult to accurately quantize the channel in a computationally efficient manner due to the high dimensionality of the massive MIMO channel. In this paper, we develop both narrowband and wideband CSI quantizers for FDMIMO taking the properties of realistic channels and the UPA into consideration. To improve quantization quality, we focus on not only quantizing dominant radio paths in the channel, but also combining the quantized beams. We also develop a hierarchical beam search approach, which scans both vertical and horizontal domains jointly with moderate computational complexity. Numerical simulations verify that the performance of the proposed quantizers is better than that of previous CSI quantization techniques. M 2 mains are searched during beam quantization, which involves a heavy computational complexity. We thus develop a hierarchical beam search approach to reduce the complexity. We also develop a wideband quantizer for broadband communication by evolving the dual codebook structure in LTEAdvanced [31], [32]. In the dual codebook structure, a first layer quantizer is used to search correlated CSI between multiple frequency tones. Unless dominant paths are gathered in a single cluster, the LTE-Advanced codebook is not effective because only adjacent radio paths are selected and quantized using the same resolution codebook. We thus concentrate on detecting adjacent and/or sperate paths within each wideband resource block (RB) based on the proposed hierarchical beam search approach. In addition, a second layer quantizer is designed to refine the beam direction of the quantized wideband CSI according to the channel vectors in a narrowband RB. When comparing our approach to the LTE-Advanced codebook, the refined beams are cophased and scaled in our approach, while the LTE-Advanced codebook only cophase adjacent beams without considering beam refinement. In Section II, we describe FD-MIMO systems employing UPAs and discuss a simple channel model that mimics true 3D SCM channels. In Section III, we review previously reported KP codebooks. In Section IV, we develop a narrowband CSI quantizer that takes multiple radio paths into account and conduct performance analysis to develop a design guideline for CSI quantizers. In Section V, we also propose a wideband CSI quantizer assuming a multi-carrier framework. In Section VI, we present simulation results, and the conclusion follows in Section VII. Throughout this paper, C denotes the field of complex numbers, N denotes the semiring of natural numbers, CN (µ, σ 2 ) denotes the complex normal distribution with mean µ and variance σ 2 , [a, b] is the closed interval between a and b, U(a, b) denotes the uniform distribution in the closed interval [a, b], d e is the ceiling function, Γ(n) denotes the complete gamma function, Ea [·] is the expectation of independent random variable a, k · kp is the p-norm, is the Hadamard product, ⊗ is the Kronecker product, 0a is the a × 1 all zeros vector, IN is the N × N identity matrix, 1a,b is the a × b all ones matrix, Aa,b , vn {A}, and eign {A} denote (a, b)th entry, the n-th dominant eigenvector, and the principal eigenvalue of the matrix A. Also, aH , a∗ , (a)` , a[a:b] denote the conjugate transpose, element-wise complex conjugate, `-th entry, and subvector including entries between [a, b] of the column vector a, respectively. antenna structure [3]. Assuming a multi-carrier framework, an input-output expression is defined as √ y[w] = ρh[w]H f [w]s[w] + n[w], (1) where y[w] is the received baseband symbol, ρ is the signalto-noise ratio (SNR), h[w] ∈ CM is the block fading MISO channel, f [w] ∈ CM is the unit-norm transmit beamformer, s[w] ∈ C is the data symbol with the power constraint E[\|s[w]\|2 ] ≤ 1, and n[w] ∼ CN (0, 1) is the additive white Gaussian noise. Note that w ∈ {1, · · · , W } denotes the frequency tone in the multi-carrier framework. To facilitate quantizer designs, we mimic the 3D-SCM and define a simplified channel model with a few radio paths according to P . X −j2π h[w] = e w− W2−1 4tp αp dM (ψpv , ψph , w), (2) p=1 while the true 3D-SCM is used to present numerical results in Section VI. In (2), P is the number of dominant paths, 4 is the subcarrier spacing, tp is the excess tap delay of the p-th radio path, αp is the channel gain of the p-th radio path, and dM (ψpv , ψph , w) is the p-th radio path at given angles (ψpv , ψph ). In the UPA scenario, a radio path for the w-th frequency tone is represented as . dM (ψ v , ψ h , w) = dMv (ψ v , w) ⊗ dMh (ψ h , w) (3) where the array response vector dMa (ψa , w) is defined as 2πda 2πda a a T 1 . dMa (ψ a , w) = √ 1, ej λ[w] ψ , · · · , ej λ[w] (Ma −1)ψ , Ma (4) for a ∈ {v, h}. In (4), ψ v = sin φv and ψ h = sin φh cos φv [33]. Note that da is the antennas spacing, φa is the angle for the array vector, and λ[w] is the wavelength for the w-th frequency tone CSI . λ[w] = 1+ 4 fc λc w− W +1 2 (5) where fc is the center frequency satisfying c = fc λc with the speed of light c. Without loss of generality, a narrowband representation of channels is defined by plugging2 w = W2+1 into (2) h= P X dM (ψpv , ψph )αp (6) p=1 II. S YSTEM M ODEL We consider multiple-input single-output (MISO) systems1 . employing M = Mv Mh transmit antennas at the base station and a single receive antenna at the user, where Mv is the number of rows and Mh is the number of columns of the UPA 1 Although we mainly discuss MISO channel quantization to simplify presentation, the proposed channel quantizer can be easily extended to multipleinput multiple-output (MIMO) systems. The extension of the MISO channel quantizer will be discussed in Section IV-C. = Da . where D = dM (ψ1v , ψ1h ), · · · , dM (ψpv , ψph ) ∈ CM ×P is the . set of radio paths and a = [α1 , · · · , αP ]T ∈ CP is the set of complex channel gains. In the narrowband assumption, w is dropped for simplicity. We assume that the beam directions are uniformly distributed in both vertical and horizontal domains such as ψpv , ψph ∼ U(−1, 1) and are independent of channel gains αp ∼ CN (0, 1). 2 We consider the W +1 -th 2 subcarrier to ignore beam squinting effects [34]. 3 In the limited feedback beamforming approach, each user chooses a transmit beamformer among codewords in the . codebook F = {f1 , · · · , f2BT } such that vectors, the channel vector h in (6) is decomposed into both vertical and horizontal domains based on the singular value decomposition [9] yielding f = arg max \|hH f̃ \|2 , rank (H̄) h= X (uk ⊗ vk∗ )σk , (8) k=1 where BT denotes the total feedback overhead. Based on the assumption that both transmitter and receiver know the prede- where the reshaped channel is given in a matrix form T fined codebook, the BT -bit index of the selected beamformer . H̄ = h[1:Mh ] , · · · , h[(Mv −1)Mh +1:Mv Mh ] ∈ CMv ×Mh . is fed back to the transmitter over the feedback link. The majority of channel quantization codebooks have been In (8), σk denotes the k-th dominant singular value, uk ∈ CMv designed for spatially correlated and uncorrelated Rayleigh denotes the k-th dominant left singular vector, and vk ∈ CMh fading channels [21]–[28]. These analytical channel models rely denotes the k-th dominant right singular vector of H̄. The final upon rich scattering environments so that each radio path has a codeword is then obtained as f = cv ⊗ (ch )∗ , where limited effect on channel characterization. Thus, most previous v 2 cv = arg max \|uH ch = arg max \|v1H c̃h \|2 . 1 c̃ \| , beamformer codebooks focus on covering the unit hypersphere v v c̃ ∈AB /2 c̃h ∈Ah BT /2 T as uniformly as possible without considering each radio path individually. However, the analytical channel models are much Despite the advantage of a KP codebook, it has some issues. different than realistic channel models that assume only a Even with a line-of-sight (LOS) channel, a dominant radio path few dominant scatterers. Therefore, this line-packing codebook may not be accurately quantized by searching each domain design approach may not be effective when the number of separately. Also, it is not always effective to quantize only antennas is large. To accurately quantize high-dimensional a single radio path because even u1 ⊗ v1∗ may consist of massive MIMO channels, it is important to tailor the codebook multiple paths. Although the quantizer in [9] considers adding to the realistic channels consisting of a limited number of radio two beams in u1 ⊗ v1∗ , the performance improvement is limited paths. because the beams are not combined properly. f̃ ∈F III. K RONECKER P RODUCT C ODEBOOK R EVIEW - S INGLE B EAM C ASE IV. P ROPOSED NARROWBAND Q UANTIZER - M ULTIPLE B EAMS C ASE It is critical in FDD massive MIMO systems to quantize and feedback information about the high-dimensional channels to the transmitter [13]–[19]. Thus, CSI quantization codebooks have been developed to tailor the feedback link with limited overhead to the massive MIMO channels [15]–[17]. Among various CSI quantization techniques, KP codebooks are of great interest to quantize the channels in a computationally efficient manner by considering the 2D antenna structure [9], [10]. Based on the channel model in (6), KP codebooks are designed to quantize a radio path with UPA structure [6]–[8]. Most KP codebooks are based on the assumption that the covariance matrix of the channels is approximated by the KP of covariance matrices of vertical and horizontal domains such that [35] . Rh = Eh hhH Prior work has verified that most 3D SCM channel realizations are well modeled with only a few resolvable 2D radio paths [1], [9]. Thus, we assume that the channel vector on a single frequency tone can be represented by a combination of a set of multiple radio paths and its corresponding channel gain vector [1]. In (6), the array response vectors in D and the channel gain vector a contain different types of channel information. Thus, we focus on quantizing D and a using different codebooks in this paper. In our narrowband quantizer, we aim to find ' Rv ⊗ Rh . Thus, a KP codebook is of the form F = f ∈ CM : f = cv ⊗ (ch )∗ , cv ∈ AvBT /2 , ch ∈ AhBT /2 , to quantizes a single dominant path with the discrete Fourier transform (DFT) codebook AaB = aMa (1/Q), · · · , aMa (Q/Q) (7) consisting of the Q = 2B codewords iT 2q 2q 1 h . 1, ejπ( Q −1) , · · · , ejπ(Ma −1)( Q −1) . aMa (q/Q) = √ Ma Most previous KP codebooks quantize the first dominant vector in each domain separately [6]–[9]. To find singular h Cz ' (9) khk2 kCzk2 . by constructing a set of N beams C = [c1 , · · · , cN ] ∈ CM ×N . and a unit-norm weight vector z = [z1 , · · · , zN ]T ∈ CN . The radio paths constituting the channels are represented by the Kronecker product of array response vectors as in (3). Thus, each 2D beam in C can be defined by a combination of quantized array response vectors in vertical and horizontal domains such that . cn = cvn ⊗ chn , n = 1, · · · , N. We assume that Bn -bit is reserved to quantize the n-th array response vector can in the domain a and Bc -bit is reserved to quantize the weight vector z. To construct C and z under the condition of a limited PN feedback overhead of BT = Bc + n=1 2Bn -bits, the following questions should be properly addressed. 1) Beam quantization: How should the radio paths in the channel be chosen and quantized? 4 Algorithm 1 Beam quantization Initialization 1: Create an initial empty matrix C0 Beam quantization 2: for 1 ≤ n ≤ N 3: Given DFT codebooks (c̃v , c̃h ) ∈ AvBn × AhBn 4: Construct beam set Cn\|c̃v ,c̃h = [Cn−1 , c̃v ⊗ c̃h ] be defined to effectively quantize and combine the limited 5: Quantize radio path cn = cvn ⊗ chn , where number of radio paths? (cvn , chn ) = arg maxc̃v ,c̃h ∈AvB ×AhB In the following subsections, we address our channel quantin n −1 H eig1 (CH Cn\|c̃v ,c̃h hhH Cn\|c̃v ,c̃h zation procedure across two separate quantization phases and n\|c̃v ,c̃h Cn\|c̃v ,c̃h ) 6: Update beam set Cn = [Cn−1 , cvn ⊗ chn ] ∈ CM ×n evaluate quantization loss at each phase. 7: end for Remark 1: The quantized channel vector can be viewed as Final update a representation of the channel using an analog beamsteering 8: Quantized radio paths C = [c1 , · · · , cN ] ∈ CM ×N matrix C, which is realized by a set of radio frequency phase 9: Unquantized vector shifters, and a baseband beamformer z. Therefore, the proposed Hweight −1 H z̄ = v (C C) C hhH C ∈ CN 1 approach follows the hybrid beamforming architecture in [36]. 2) Beam combining: How should the quantized beams be cophased and scaled? 3) Feedback resource allocation: For a given total feedback overhead BT , how should the feedback-bit allocation scenario . rN = [B1 , · · · , BN , Bc ]T ∈ NN +1 (10) A. Phase I: Beam Quantization In the beam quantization phase, we aim to construct a selected set of quantized 2D beams C. It is well known that the DFT codebook is an effective solution to quantize array response vectors so that we quantize the n-th array response vector can with the DFT codebook AaBn . In our beam quantization approach, we select and quantize the 2D DFT beams sequentially. In the n-th update, the quantized DFT vectors and the unquantized weight vector3 are obtained by solving the maximization problem (cvn , chn , z̄n ) = arg max h n (c̃v ,c̃h ,z̃)∈Av Bn ×ABn ×C hH Cn\|c̃v ,c̃h z̃ Cn\|c̃v ,c̃h z̃ 2 2 2 , (11) . where Cn\|c̃v ,c̃h = [c1 , · · · , cn−1 , c̃v ⊗ c̃h ] ∈ CM ×n , and {c1 , · · · , cn−1 } includes previously selected DFT beams. In (11), we do not quantize the weight vector z̄n ∈ Cn in each update because it is not practical to construct a codebook for weight vectors that constantly change its dimension. The problem of choosing the n-th DFT beam is then simplified to 2! hH Cn\|c̃v ,c̃h z̃ v h (cn , cn ) = arg max max 2 z̃∈Cn h Cn\|c̃v ,c̃h z̃ 2 c̃v ,c̃h ∈Av Bn ×ABn (a) −1 = arg max eig1 (CH n\|c̃v ,c̃h Cn\|c̃v ,c̃h ) h c̃v ,c̃h ∈Av Bn ×ABn H CH n\|c̃v ,c̃h hh Cn\|c̃v ,c̃h , (12) where (a) is derived from the unquantized weight vector, which is computed based on the generalized Rayleigh quotient [37], such that z̄n\|c̃v ,c̃h = arg max = H z̃H (CH n\|c̃v ,c̃h hh Cn\|c̃v ,c̃h )z̃ z̃H (CH C v h )z̃ z̃∈Cn n\|c̃v ,c̃h n\|c̃ ,c̃ H −1 H v1 (Cn\|c̃v ,c̃h Cn\|c̃v ,c̃h ) Cn\|c̃v ,c̃h hhH Cn\|c̃v ,c̃h . 3 The bar on the top of weight vectors denotes the unquantized weight vectors. The beam quantization approach gives a set of N quantized beams C = [c1 , · · · , cN ] ∈ CM ×N (13) and the unquantized weight vector z̄ = v1 (CH C)−1 CH hhH C ∈ CN . (14) A separate beam quantization should be performed to compute each codeword candidate based on Algorithm 1. In the following Section IV-E, a practical beam search technique will be proposed to quantize a single dominant beam with moderate computational complexity as well as compute multiple codeword candidates in a hierarchical fashion. We also evaluate quantization loss due to the beam quantization as a function of the number of beams N and the feedback overhead Bn for DFT codebooks. Assuming the unquantized weight vector z̄, the beamforming gain between the channel vector h and the set of DFT beams C . Gbq = Eh \|hH Cz̄\|2 /kCz̄k22 \|hH Cz̃\|2 (15) = Eh max 2 z̃∈CN kCz̃k2 is averaged over channel realizations h in Lemma 1. Before presenting the lemma, we make the following assumption. Assumption 1: Assuming a channel vector h has already been decomposed into a set of radio paths D and channel gains a as in (6), the column vectors for each domain a ∈ {h, v} . in C are separately selected as can = aMa (qn /2Bn ) ∈ AaBn , where qn = arg max q∈{1,··· ,2Bn } 2 a Bn dH ) . Ma (ψn )aMa (q/2 (16) Considering half-wavelength antenna spacing da = λ2 , the beamforming gain between the n-th array response vector and . a a 2 the selected DFT vector Γ2na = E \|dH is derived Ma (ψn )cn \| in Appendix A as M a −1 X 2(Ma − q) sin πq/2Bn 1 2 Γna = 2 Ma + . Ma πq/2Bn q=1 5 Eh [(cv1 ⊗ ch1 )H hhH (cv1 ⊗ ch1 )]  .. R= . Eh [(cvN ⊗ chN )H hhH (cv1 ⊗ ch1 )]  ··· .. . ··· Lemma 1: A lower bound of the beamforming gain Gbq in (15) is approximated as N X P X P M Γ2nv Γ2nh − 1 Gbq ' N+ . M +N −1 qP n=1 q=n Please check Appendix B for the proof. B. Phase II: Beam Combining In the beam combining phase, we aim to compute a weight vector z, which is used to combine beams in C. To quantize weight vector H z = arg max z̃∈ZBc 2 \|h Cz̃\| , kCz̃k22 (17) we design the codebook including U = 2Bc unit-norm . combiners ZBc = {z1 , · · · , zU }. To study a codebook design framework, we model the effective channel vector based on the Kronecker correlation model as 1 CH h ' R 2 w ∈ CN , (18) . where the covariance matrix R = Eh CH hhH C ∈ CN ×N in (19) is analytically computed in Appendix C and random variables w = [w1 , · · · , wN ]T ∈ W denotes the weight vector that is subject to the equal gain subset √ . W = w ∈ CN : wn = ejϑn / N , ϑn ∼ U(0, 2π) . In our codebook design approach, we pick a set of codewords {z1 · · · zU } that maximize where the inequality in (a) is based on kCzu k22 ≤ kCk22 kzu k22 and kzu k22 = 1, and (b) is approximated by plugging the Kronecker correlation model in (18) into the maximizer. Based on the correlated Grassmannian beamforming algorithm in [24], codewords are then obtained by setting Note that I ∈ N denotes the phase quantization level. We also evaluate the quantization loss due to the beam combining as a function of the number of codewords U = 2Bc in the codebook ZBc that quantizes the baseband combiner z̄ in (14). To analyze quantization performance of ZBc , the normalized beamforming gain between the normalized effective channel and the selected unit-norm combiner z . Gbc = EhH C \|hH Cz\|2 /khH Ck22 2 \|eH u Rw\| ' Ew max (20) u∈{1,··· ,U } (wH Rw)(eH u Reu ) is averaged over the effective channel CH h. Because it is not easy to compute in a closed form, we only derive the normalized beamforming gain in the special case of N = 2 based on the following assumption. Assumption 2: For simple analysis, we assume that the . combiner is selected as z = zû , where . û = arg max cos2 θu . u∈{1,··· ,U } . Note that θu = arccos \|wH eu \|. Lemma 2: In the special case of N = 2, the normalized beamforming gain Gbc in (20) is approximated as 1 U π Gbc ' 1 + sin . 2 π U Please check Appendix D for the proof. C. MIMO Channel Quantizer In our MIMO channel quantization approach, DFT beams are chosen to maximize the sum of the channel gains of each receive antenna. The DFT vectors are quantized by solving the rewritten problem −1 (cvn , chn ) = arg max eig1 (CH n\|c̃v ,c̃h Cn\|c̃v ,c̃h ) h c̃v ,c̃h ∈Av Bn ×ABn H CH n\|c̃v ,c̃h HMIMO HMIMO Cn\|c̃v ,c̃h . 1 zu = R 2 eu /kR 2 eu k2 and picking a set of equal gain vectors {e1 · · · eU } maximizing q 2 %min = min 1 − \|eH a eb \| , 1≤a<b≤U where equal gain vectors eu ∈ EN are restricted to √ . EN = e ∈ CN : (e)n = ej2πϕn / N , ϕn ∈ 1/I · · · I/I (19) Before investigating feedback allocation solutions, we extend the proposed quantizer to MIMO channel scenarios. Assuming a MIMO system employing V receive antennas at each user, the channel matrix is defined as HMIMO = h1 , · · · , hV ∈ CM ×V . \|hH Czu \|2 1≤u≤U kCzu k2 2 (a) \|hH Czu \|2 ≥ min 1≤u≤U kCk22 1 H 2 (b) w H R 2 zu ' min , 1≤u≤U kCk22 σmin = min 1  Eh [(cv1 ⊗ ch1 )H hhH (cvN ⊗ chN )]  .. . . v h H H v h Eh [(cN ⊗ cN ) hh (cN ⊗ cN ) . After constructing a selected set of DFT beams, we next compute T orthogonal weight vectors to compute a precoding matrix for spatial multiplexing. For a given set of beams C, we compute the beamformer for the t-th layer transmission ft = Cz̄t /kCz̄t k2 ∈ CM , where the unquantized weight vector N z̄t = vt (CH C)−1 CH HMIMO HH MIMO C ∈ C 6 3.5 3 2.5 Beamforming gain G(M,P,N,rN) N Beamforming gain G(M,P,N,r ) 3 2 1.5 r ={8,0} 1 1 r2={4,3,2} r ={4,3,2} 0.5 2 r2={3,3,4} 2.5 2 1.5 r2={5,4,2} 1 r2={4,4,4} 6x6 8x8 r2={6,3,2} r3={5,2,2,2} 0.5 r3={3,2,2,2} 0 4x4 r1={10,0} r3={4,3,2,2} r3={3,3,3,2} 10x10 12x12 14x14 16x16 18x18 20x20 0 4x4 6x6 8x8 10x10 12x12 14x14 16x16 18x18 20x20 Mv × Mh (a) BT = 16-bits Mv × Mh (b) BT = 20-bits Fig. 1. Cross correlation G(M, P, N, rN ) over different feedback-bit allocation scenarios with P ∈ {3, 4, 5}. is computed based on the generalized Rayleigh quotient [37]. where M is the number of antennas, P is the number of beams The precoding matrix is then constructed as in h, N is the number of dominant beams to be quantized, and rN is the feedback-bit allocation scenario in (10). M ×T F = f1 , · · · , fT ∈ C , In the proposed quantizer, the feedback scenario are chosen where T ≤ V denotes the maximum transmission rank.4 as N, rN = arg max EP G(M, P, Ñ , r̃Ñ ) (23) D. Feedback Resource Allocation In our KP codebook procedure, quantizing more beams with a high resolution codebook increases the beamforming gain at the cost of increased feedback overhead. To effectively allocate limited feedback overhead resources, we must derive the beamforming gain between the randomly generated channel vectors h and the selected codeword Cz/kCzk2 using H \|h Cz\|2 . G = Eh , (21) kCzk22 as a function of the feedback-bit allocation scenario5 rN in (10). However, inter-dependencies across both quantization phases in Section IV-A and Section IV-B make it hard to compute the beamforming gain in a closed form. To simplify analysis, we make the following assumption. Assumption 3: Assuming the quantization phases in Section IV-A and Section IV-B work independently, the channel quantization quality in the proposed KP codebook procedure is evaluated by the combination of the quantization losses in both phases. Based on Assumption 3 that both quantization phases are independent of each other, the beamforming gain in the proposed quantizer is defined by the mixture of Gbq and Gbc such as, . G(M, P, N, rN ) = Gbq Gbc (22) 2 2 P P M Γ Γ −1 N P nv nh Gbc P N + n=1 q=n q = , M +N −1 4 Note that designing a beam combining codebook and a feedback resource allocation algorithm that support multi layer MIMO transmission are interesting topics for future research. 5 In r , B for n ∈ {1, · · · , N } denotes the size in bits of the DFT n N codebooks Aa Bn in the domain a, and Bc denotes the size in bits of the codebook for combiners ZBc . (Ñ ,r̃Ñ )∈N ×NÑ +1 by evaluating all possible scenarios r̃Ñ = [B̃1 , · · · , B̃Ñ , B̃c ]T . that considers Ñ beams in N = {1, 2, 3}. Note that the possible feedback scenarios are subject to the total feedback PÑ overhead BT = B̃c + n=1 2B̃n -bits. In (23), the expectation is taken over the number of dominant paths P since P varies depending on the channel environments. By assuming P is equally probable from 3 to 5, we plot the arithmetic mean of G(M, P, Ñ , rÑ ) in Fig. 1 with different numbers of antennas and feedback bits. As shown in the figure, quantizing one or two beams give the best performance under practical UPA scenarios and feedback overheads. Therefore, we construct the codebook F1 for quantizing a single 2D DFT beam and the codebook F2 combining two quantized 2D DFT beams based on the predefined feedback-bit allocation scenarios6 r1 = [B1 + B̌1 , 0] ∈ N2 , r2 = [B1 , B2 , Bc ] ∈ N3 , (24) respectively. The final codebook is then defined such that . F = F1 ∪ F 2 . Remark 2: In most of channel realizations, the inter-user interference due to the remained paths is negligible because most of channel gains are contained in the first and second dominant beams [1], [9]. Based on the codebook subset restriction algorithm in [?], [?], severe inter-user interference could be mitigated by reporting the remained paths having a considerable amount of channel gains. 6 The total feedback overhead for feedback scenarios are B = 2B + 2B̌ 1 1 T bits and BT = 2B1 + 2B2 + Bc -bits, where 2B̌1 = 2B2 + Bc . 7 Normalized cross correlation 0.8 (a) Wideband resource blocks (b) Narrowband resource blocks 0.7 0.6 0.5 γc 1 0.4 4x4 Fig. 2. An overview of wideband model having multiple tones. γh 6x6 8x8 10 x 10 12 x 12 14 x 14 16 x 16 M v x Mh E. Beam Search Approach It is necessary to search both vertical and horizontal domains jointly to scan for the dominant beams in a channel vector. However, this joint approach increases a computational complexity. For example, it is required to carry out 22(B1 +B̌1 ) vector computations to scan a single 2D DFT beam under the feedback-bit allocation scenario r1 in (24). To reduce the heavy computational complexity that comes with detecting the single dominant beam, we propose a multi-round beam search technique as follows. Round 1: For the channel vector h, the first dominant beam is chosen using DFT codebooks AvB1 and AhB1 in (7), which have low-resolution DFT vectors. The DFT beam is given by c1 = cv1 ⊗ ch1 , (cv1 , ch1 ) = (25) H h 2 \|h (c̃ ⊗ c̃ )\| . arg max h (c̃v ,c̃h )∈Av B ×AB 1 v 1 Later, the selected 2D DFT beam in (25) will be a baseline that guides the generation of two codeword candidates. Round 2: In this round, 2B̌1 -bits are assigned for constructing the two codeword candidates. 1) To support channel realizations having a single dominant beam, a codeword is computed based on the feedback-bit allocation scenario r1 in (24) by scanning beam directions near c1 from Round 1. The first codeword is given by f1 = ďM (θ1v , θ1h ) 2) To support channel realizations having multiple dominant beams, a codeword is computed based on the feedback-bit allocation scenario r2 in (24) by choosing an additional DFT beam to combine with c1 . The second codeword is given by f2 = (cv2 , ch2 , z) = [c1 , cv2 ⊗ ch2 ]z , k[c1 , cv2 ⊗ ch2 ]zk2 arg max h (c̃v ,c̃h ,z̃)∈Av B ×AB ×ZBc 2 hH [c1 , c̃v ⊗ c̃h ]z̃ k[c1 , c̃v ⊗ c̃h ]z̃k2 2 2 using B2 -bits size DFT codebooks AvB2 and AhB2 and Bc -bits size codebook ZBc , which is developed to combine the two 2D DFT beams as explained in the Section IV-B. Considering our CSI quantization technique, the second codebook is then defined such that v c̃1 ⊗ c̃h1 , c̃v2 ⊗ c̃h2 z̃ . M F2 = f2 ∈ C : f2 = v c̃1 ⊗ c̃h1 , c̃v2 ⊗ c̃h2 z̃ 2 over c̃v1 ∈ AvB1 , c̃h1 ∈ AhB1 , c̃v2 ∈ AvB2 , c̃h2 ∈ AhB2 , and z̃ ∈ ZBc . Round 3: Using the two codeword candidates f1 ∈ F1 and f2 ∈ F2 , the final codeword is selected with an additional bit 2 f = arg max hH f̃ . c1 , (θ1v , θ1h ) = arg max hH ďM (θ̃v , θ̃h ) Fig. 3. Normalized beamforming gains between subcarrier channel vectors. (B = 5, dv = 0.8λ, dh = 0.5λ, w ∈ {1, · · · , 150}) (27) f̃ ∈{f1 ,f2 } c1 2 B̌ θ̃ v ,θ̃ h ∈TB 1 1 . √ where ďM (θv , θh ) = M dM (θv , θh ) is defined to shift the beam directions7 and the B̌1 -bit size codebook −B̌1 1 − 2−B̌1 B̌1 −(B1 +B̌1 ) 1 − 2 T B1 = − :2 : (26) 2B1 +1 2B1 +1 is designed for refining beam directions of any DFT beams. In our CSI quantization approach, the first codebook is then defined such that . F1 = f1 ∈ CM : f1 = (c̃v1 ⊗ c̃h1 ) ďM (θ̃v , θ̃h ) V. P ROPOSED W IDEBAND Q UANTIZER We develop a wideband quantizer that takes multiple frequency tones into account. Before developing practical quantizers, we overview a broadband system model adopted in 3GPP LTE-Advanced. As shown in Fig. 2(a), W total frequency tones are divided into L wideband RBs where each wideband RB includes W L channels. Each wideband RB is written in a matrix form as W H` = h[1 + (` − 1)W /L], · · · , h[`W/L] ∈ CM × L . over c̃v1 ∈ AvB1 , c̃h1 ∈ AhB1 , and θ̃v , θ̃h ∈ TBB̌11 . As depicted in Fig. 2(b), each wideband RB is divided into R narrowband RBs H` = H`,1 , · · · , H`,R , 7 √ The Hadamard product formulation satisfies the following formulation M dM (θ1 , θ3 ) dM (θ2 , θ4 ) = dM (θ1 + θ2 , θ3 + θ4 ). where H`,r ∈ CM × LR denotes the narrowband RB that is written in a matrix form. W 8 3.5 3.6 Numerical (N=2, P=5) Numerical (N=2, P=4) Numerical (N=2, P=3) Analytical (N=2, P=5) Analytical (N=2, P=4) Analytical (N=2, P=3) 3.2 Beamforming gain Beamforming gain 3 Numerical (N=2, P=5) Numerical (N=2, P=4) Numerical (N=2, P=3) Analytical (N=2, P=5) Analytical (N=2, P=4) Analytical (N=2, P=3) 3.4 2.5 2 3 2.8 2.6 2.4 2.2 1.5 2 1.8 1 8x8 10 x 10 12 x 12 14 x 14 16 x 16 18 x 18 20 x 20 1.6 8x8 10 x 10 12 x 12 Mv × Mh 14 x 14 16 x 16 18 x 18 20 x 20 Mv × Mh (a) B1 = 4, B2 = 4, Bc = 2 (BT = 18) (b) B1 = 5, B2 = 4, Bc = 2 (BT = 20) Fig. 4. Beamforming gain comparison between the numerical results G in (21) and the analytical results G(M, P, N, rN ) in (22). Next, the correlation between the channel vectors is studied numerically based on the cross correlations over w1 6= w2 2 h[w1 ]H h[w2 ] . γh = Eh[w] , kh[w1 ]k22 kh[w2 ]k22 2 . γc1 = Eh[w] c1 [w1 ]H c1 [w2 ] , codeword quantizes channel vectors H`,r in the r = dR( wL W − e-th wideband RB ` − 1)e-th narrowband RB of the ` = d wL W by refining the beam direction of c1\|` according to f1\|`,r = ďM (θ1v , θ1h ) (θ1v , θ1h ) = arg max B θ̃ v ,θ̃ h ∈TB N1 W1 c1\|` , v h HH `,r ďM (θ̃ , θ̃ ) c1\|` 2 2 where h[w] denotes 3D-SCM channel vectors and c1 [w] with BN1 -bit codebooks in (26). denotes the dominant 2D DFT beam that is chosen from the Round 2: The second codeword is computed to support the v h DFT codebooks AB and AB in (7) for subcarrier w. As shown channel scenario having two dominant beams. The proposed in Fig. 3, it is verified that the dominant 2D DFT beams in quantizer only refines the direction of c1\|` as well as combines the different subcarriers’ channel vectors are highly correlated. with the second 2D DFT beam c2\|` . The second codeword is Based on empirical studies, the wideband quantizer is designed in such a way that the correlated information, i.e., the dominant ďM (θ2v , θ2h ) c1\|` , c2\|` z f2\|`,r = , 2D DFT beam, is shared between neighboring subcarriers. ďM (θ2v , θ2h ) c1\|` , c2\|` z 2 Level 1 (Wideband resource block): We choose two 2D 2 v h HH c1\|` , c2\|` z̃ DFT beams that are close to the channel vectors in each `,r ďM (θ̃ , θ̃ ) v h (θ2 , θ2 , z2 ) = arg max wideband RB. For supporting the `-th wideband RB, the first ďM (θ̃v , θ̃h ) c1\|` , c2\|` z̃ 2 2 θ̃ v ,θ̃ h ,z̃ 2D DFT beam is chosen as over θ̃v , θ̃h ∈ TBBN2 in (26) and z̃ ∈ ZBc . Out of 2BN1 -bits c1\|` = cv1 ⊗ ch1 , (28) allocated for eachW1narrowband RB, 2B bits are assigned for N2 v h 2 refining the first 2D DFT beam and B c bits are assigned for (cv1 , ch1 ) = arg max HH (c̃ ⊗ c̃ ) ` 2 h (c̃v ,c̃h )∈Av combining the two 2D DFT beams. BW1 ×ABW1 Round 3: Among the two codeword candidates, the final with BW1 -bit DFT codebooks. Next, the second 2D DFT beam codeword is selected with an additional bit according to is chosen using BW2 -bit DFT codebooks as 2 f`,r = arg max HH (30) `,r f̃ 2 . c2\|` = cv2 ⊗ ch2 , (29) f̃ ∈{f1\|`,r ,f2\|`,r } (cv2 , ch2 ) = arg max max c̃v ,c̃h z̃∈ZBc v h HH ` [c1\|` , c̃ ⊗ c̃ ]z̃ k[c1\|` , c̃v ⊗ c̃h ]z̃k2 2 VI. S IMULATION R ESULTS 2 over (c̃v , c̃h ) ∈ AvBW2 × AhBW2 . Level 2 (Narrowband resource block): Within the `-th wideband RB, the set of 2D DFT beams c1\|` and c2\|` will be a baseline that guide the quantization of channel vectors in each narrowband RB. To construct each set of two codeword candidates, 2BN1 -bits are allocated for each narrowband RB. Round 1: The first codeword is computed to support the channel scenario having a single dominant beam. The first We verify the performance of our CSI quantizers. Before evaluating the proposed quantizers, we pause to validate the accuracy of the approximated beamforming gain in (22). The beamforming gain between the simplified channel h and the quantized channel Cz/kCzk2 is computed as in (21). For numerical simulations, 10, 000 channels in (6) are generated by assuming a fixed number of the ray-like beams P ∈ {3, 4, 5}. In Fig. 4, it is shown that the approximated formula in (22) gives a tight lower bound on the numerical results in (21). 9 0.8 0.9 Prop. N-i (BN=21) Prop. N-ii (B N=19) 0.8 Prop. N-iii (B N=17) 0.75 Enh. KP-i (BN=22) Enh. KP-ii (B N=20) 0.7 KP (BN=22) 0.65 0.6 0.55 0.5 Prop. N-ii (B N=19) 0.7 Prop. N-iii (B N=17) 0.65 Enh. KP-i (BN=22) 0.6 Enh. KP-ii (B N=20) KP (BN=22) 0.55 0.5 0.45 0.4 0.35 0.45 4x4 Prop. N-i (BN=21) 0.75 Normalized beamforming gain GN Normalized beamforming gain GN 0.85 0.3 6x6 8x8 10 x 10 12 x 12 M v x Mh 14 x 14 16 x 16 4x4 6x6 (a) dv = 0.8λ, dh = 0.5λ 8x8 10 x 10 12 x 12 M v x Mh 14 x 14 16 x 16 (b) dv = 1.6λ, dh = 1.0λ Fig. 5. Normalized beamforming gain of narrowband quantizers. TABLE I 3D-SCM SIMULATION PARAMETERS . Tx antennas Rx antennas Scenario Carrier frequency fc Subcarrier spacing 4 Vertical antenna spacing dv Horizontal antenna spacing dh 4 × 4 to 20 × 20 Co-polarized UPA 1 Co-polarized UPA UMi NLOS 2 GHz 15 KHz 0.8λc , 1.6λc 0.5λc , 1.0λc We now evaluate the narrowband quantizers using simulations. Numerical results are obtained through Monte Carlo simulations with 10, 000 channel realizations. For generating 3D-SCM channels, we use the parameters in Table I. We evaluate the normalized beamforming gain of narrowband quantizer H 2 \|f h\| . GN = Eh , khk22 where f is the final codeword8 that is chosen in (27). We compare the beamforming gains of the proposed quantizer with that of the KP codebooks in [6], [7] and the enhanced KP codebook in [9]. Note that the feedback-bit allocation9 of each quantizer is listed in Table II and the computational complexity10 VN and feedback overhead11 BN are summarized in Table III. In Figs. 5(a) and 5(b), the normalized beamforming gains of the three quantizers are plotted with different antenna spacing 8 The perfect CSI beamformer gives the normalized beamforming gain of one. 9 The feedback-bit allocation scenarios for the proposed narrowband quantizers are predefined in (24). 10 We count the number of vector computations to evaluate the complexity. 11 The feedback overhead (per each frequency tone CSI) is assessed by the combination of the overheads for both the first and second rounds in Section IV-E. TABLE II F EEDBACK CONFIGURATIONS OF NARROWBAND QUANTIZER Prop. N-i Prop. N-ii Prop. N-iii Enh. KP-i Enh. KP-ii KP codebook 1st round B1 = 5 B1 = 5 B1 = 4 B1 + B2 B1 + B2 B1 = 11 2nd B̌1 B̌1 B̌1 B1 B1 round = 5, B2 = 4, B2 = 4, B2 = 5, B2 = 5, B2 = 4, Bc = 2 = 3, Bc = 2 = 3, Bc = 2 =5 =4 BN 21 19 17 22 20 22 VN 3,072 1,536 768 2,176 1,120 4,096 TABLE III F EEDBACK OVERHEADS AND COMPLEXITY COMPARISONS Prop. quantizer Enhanced KP KP codebook Feedback overhead BN 2(B1 + B2 ) + Bc + 1 2(B1 + B2 + 1) 2B1 Vector computations VN 22B1 + 22B2 +Bc +1 2 2B1 +B2 + 2B1 + 2B2 2B1 +1 scenarios. The proposed quantizer searches both vertical and horizontal domains jointly, while other KP codebooks search beams lying in each domain independently and integrate the results later. The 2D DFT beams, which are quantized in the proposed quantizer, are aligned by cophasing and scaling each beam. On the contrary, the quantized beams in the enhanced KP codebook [9] are simply added up together without considering phase alignment. For these reasons, the proposed quantizer generates higher beamforming gains than those of other KP codebooks. We next evaluate the normalized beamforming gain of the wideband quantizer according to H 2 . GW = Eh[w] f`,r h[w]/kh[w]k2 , where f`,r is the chosen codeword in (30). In Figs. 6(a) and 6(b), the normalized beamforming gains of wideband quantizer are compared with those of the narrowband quantizers. In the legend, the first alphabet denotes the type of quantizer, the 10 0.7 0.6 W-I-1 (BW =136) W-II-2 (BW =65) 0.6 N-ii-1 (BN =152) N-ii-2 (BN =19) 0.55 W-I-2 (BW =83) 0.55 W-II-1 (BW =120) 0.5 0.45 Normalized beamforming gain G W 0.65 Normalized beamforming gain G W W-I-1 (BW =136) W-I-2 (BW =83) W-II-1 (BW =120) W-II-2 (BW =65) 0.5 N-ii-1 (BN =152) N-ii-2 (BN =19) 0.45 0.4 0.35 0.3 0.4 8x8 10 x 10 12 x 12 14 x 14 M v x Mh 16 x 16 18 x 18 20 x 20 0.25 8x8 10 x 10 (a) dv = 0.8λ, dh = 0.5λ 12 x 12 14 x 14 M v x Mh 16 x 16 18 x 18 20 x 20 (b) dv = 1.6λ, dh = 1.0λ Fig. 6. Normalized beamforming gain comparison between the narrowband and wideband quantizers. TABLE IV F EEDBACK OVERHEADS FOR EACH WIDEBAND AND NARROWBAND RB W-I W-II Level 1: Wideband RB BW1 = 5, BW2 = 5 BW1 = 5, BW2 = 5 Level 2: Narrowband RB BN1 = 3, BN2 = 2, Bc = 2 BN1 = 2, BN2 = 1, Bc = 2 TABLE V W IDEBAND CONFIGURATIONS FOR NARROWBAND AND WIDEBAND QUANTIZERS N-1 N-2 Narrowband quantizer 1 codeword / 75 tones 1 codeword / 600 tones W-1 W-2 Wideband quantizer L = 4, R = 2 L = 1, R = 9 second alphabet denotes the feedback-bit allocation scenario in Table IV, and the final digit represents the wideband configuration12 in Table V. The total feedback overhead of the proposed wideband quantizer is defined as BW = 2(BW1 + BW2 )L + (2BN1 + 1)RL. Numerical results verify that wideband quantizers outperforms narrowband quantizers because it exploits correlation between frequency tone CSIs. The wideband quantizers also reduce feedback overhead because they can maintain quantization performance with less overhead compared to narrowband quantizers. VII. C ONCLUSION In this paper, advanced CSI quantizers based on the KP codebook structure are proposed for FD-MIMO systems using UPAs. In the proposed quantizer designs, we focused on detecting and quantizing a limited number of dominant 2D beams in 3D channel vectors by exploiting DFT codebooks. 12 In the LTE setup of scheme W-3, the first 8 narrowband RBs have 72 tone CSIs and the ninth narrowband RB has 24 tone CSIs. The codebook for combiners was designed to cophase and scale the quantized 2D DFT beams. Furthermore, we analytically derived a design guideline for practical quantizers, which is based on FD-MIMO systems with predefined feedback-bit allocation scenarios. We then developed CSI quantizers by taking the predefined feedback scenarios into account. First, a narrowband quantizer was proposed to quantize and/or combine one or two dominant 2D DFT beams. To detect and quantize beams properly, we also developed a multi-round beam search approach that scans both vertical and horizontal domains jointly under the moderate computational complexity. To reduce total feedback overhead, we also proposed a wideband quantizer that utilizes the correlated information between multiple frequency tones. Numerical simulations verified that the proposed narrowband quantizer gives better quantization performance than previous CSI quantization techniques, and the proposed wideband quantizer further improves the quantization performance with less feedback overhead compared to the narrowband quantizer in wideband settings. A PPENDIX A C ORRELATION BETWEEN ARRAY RESPONSE VECTOR AND DFT CODEWORD We discuss the correlation between the array response vector in the domain a and the selected DFT codeword to quantify the quantization performance of DFT codebooks by evaluating . a a 2 Γ2na = E dH Ma (ψn )cn h i 2 a =E max dH Ma (ψn )aMa (q/Qn ) q∈{1,··· ,Qn } = 1 E max Ma2 q∈{1,··· ,Qn } M a −1 X m=0 a e−jπm(ψn −2q/Qn +1) 2 11 (a) = MX 2 a −1 a 1 −jπm(ψn −2qn /Qn +1) e , E Ma2 m=0 where (a) is rewritten with the index of selected codeword qn = arg min q={1,··· ,Qn } ψna − 2q/Qn + 1 . To complete the lower bound in (31), we first compute the expectation of two-norm squared of the effective channel vector h i 2 E khH Ck22 = E hH [cv1 ⊗ ch1 , · · · , cvN ⊗ chN ] 2 N h i X 2 = E hH (cvn ⊗ chn ) n=1 N X P X M Γ2nv Γ2nh − 1 (a) 1 The expectation over ψna ∼ U(−1, 1) is rewritten by defining = PN + , (32) . M q the new random variable as ψ = ψna − 2qn /Qn + 1, which n=1 q=n follows U(−1/Qn , 1/Qn ) because \|ψna −2qn /Qn +1\| ≤ 1/Qn . where (a) is derived by using the correlation between The correlation formula13 is then computed over ψ, as the channel vector and the n-th selected DFT codeword MX MX E \|hH (cvn ⊗ chn )\|2 that will be discussed in Appendix E. 2 a −1 a −1 M a −1 X 1 1 = 2E e−jπmψ e−jπ(m−`)ψ E We next consider the set of N DFT vectors C in (13). The 2 Ma M a m=0 `=0 m=0 expectation of two-norm squared of C is approximated as 1 M a −1 M a −1 Z X X Qn 2 cos π(m − `)ψ 1 2 H H = 2 Ma + dψ E kCk2 = E max x̃ (C C)x̃ 1 Ma 2/Qn x̃∈CN `=0 m>` − Qn M ' max x̃H E CH C x̃ (33) a −1 M a −1 X X 2 sin π(m − `)/Qn N 1 x̃∈C = 2 Ma + Ma π(m − `)/Qn subject to kx̃k22 = 1. Note that E CH C is computed as `=0 m>`  v H v  M a −1 X 2(Ma − q) sin πq/Qn 1 (c1 ) c1 ⊗ (ch1 )H ch1 · · · (cv1 )H cvn ⊗ (ch1 )H chn = 2 Ma + .   .. .. .. Ma πq/Qn E  q=1 . . . (cvn )H cv1 ⊗ (chn )H ch1 · · · (cvn )H cvn ⊗ (chn )H chn   1 · · · Mv1Mh A PPENDIX B   .. .. .. L OWER BOUND OF NORMALIZED BEAMFORMING GAIN (34) =  . . . 1 ··· 1 Remark 3: To simplify analysis, we consider the first order Mv Mh Taylor expansion of the bivariate variables, which is derived where ca = a (q /Q ) with Q = 2Bn is chosen as in Ma n n n n as in [38], (16). Because we assume that beam directions are uniformly distributed ψna ∼ U(−1, 1), can can be chosen as one of E[A] E[A] ∂ E[B] A E[A] ∂ E[B] Qn = 2Bn codewords in the DFT codebook AaBn with equal ' + (A − E[A]) + (B − E[B]). B E[B] ∂A ∂B probabilities. For this reason, we can obtain E (cac )H cad in (34) by computing its arithmetic mean of the beamforming Expectation of the bivariate variables is then approximated as gain between two different codewords (c 6= d) as A E[A] E[A] Qc Qd 1 X X H E . 'E = E (cac )H cad = a (qc /Qc )aMa (qd /Qd ) B E[B] E[B] Qc Qd q =1 q =1 Ma c d " Q Q # Qc X Qd The beamforming gain Gbq in Lemma 1 is lower bounded M c d a −1 X X X X j2πmqd −j2πmqc 1 as = 1+ e Qd e Qc Ma Qc Qd q =1 q =1 q =1 m=1 q =1 c c d d " \|hH Cz̃\|2 # M a −1 Gbq = E max X 1 1 − ej2πm 1 − e−j2πm 2 2 z̃∈CN kCz̃k2 = Q + j2πm j2πm Ma Q2 (a) 1 − e − Qc 1 − e Qd maxz̃∈CN \|hH Cz̃\|2 m=1 ≥ E 1 kCk2 . (35) = H 2 2 M a kh Ck2 (b) =E Based on E CH C , the expectation of two-norm squared kCk22 H 2 of C in (33) is approximated as (c) E kh Ck2 ' , (31) E kCk22 E kCk22 ' eig1 E CH C eig1 {1N,N } M − 1 = + where the inequality in (a) is based on kCz̃k22 ≤ kCk22 kz̃k22 . M M and kz̃k22 = 1, (b) holds when z̃ = CH h/kCH hk2 , and (c) (a) M + N − 1 = , (36) is derived based on Remark 3. M where (a) is derived because eig1 {1N,N } = N , which holds 13 Assuming Q > M /2, the correlation formula is always positive, i.e., a a Γ2na > 0, because \|ψ\| < 2/Ma . when x̃ = √1N [1, · · · , 1]T ∈ CN . 12 E[(cvc ⊗ chc )H hhH (cvd ⊗ chd )] = X P ∗ X P v v H h h αp∗ (dH Mv (ψp )cc )(dMv (ψp )cc ) p=1 v v H h h αq∗ (dH Mv (ψq )cd )(dMv (ψq )cd ) q=1 P X h H v v h H h h = E \|αp \|2 E (cvc )H dMv (ψpv )dH Mv (ψp )cd E (cc ) dMv (ψp )dMv (ψp )cd . (37) p=1 Finally, the lower bound of Gbq in (31) is approximated by plugging the derived formulas in (32) and (36) into (31) as Gbq N X P X P M Γ2nv Γ2nh − 1 ' N+ . M +N −1 qP n=1 q=n A PPENDIX C Case 4: p 6= c, p 6= d, c 6= d. a a E (cac )H dMa (ψpa )dH Ma (ψp )cd MX M a −1 a −1 X 2q 2qc a 1 −jπm(ψpa − Q d +1) d = 2E ejπ`(ψp − Qc +1) e Ma m=0 = 1 Ma2 `=0 M a −1 M a −1 X X `=0 C OVARIANCE MATRIX OF EFFECTIVE CHANNEL VECTOR is Each entry of R in (19) rewritten in (37). Note that a a E (cac )H dMa (ψpa )dH (ψ )c p Ma d in (37) is computed depending on the different cases as follows: Case 1: p = c = d. a a 2 E (cac )H dMa (ψpa )dH Ma (ψp )cd = Γac , ejπ(`−m) m=0 h iZ mq c j2π( Q d − `q Qc ) d E e 1 −1 = 1 Ma2 M a −1 X a ejπ(`−m)ψp a dψp 2 h i q qc j2πm( Qd − Q ) c d E e m=0 1 , = Ma2 (a) where (a) is derived by computing the arithmetic mean. where Γ2ac is derived in Appendix A. Case 2: p = 6 c, c = d. i h 2 a a H a E (cac )H dMa (ψpa )dH (ψ )c = E d (ψ )a (q /Q ) M c c Ma p d Ma p a (a) = 1 , Ma where qc is chosen as in (16). Note that (a) is derived by computing the arithmetic mean. Case 3: p = c, c 6= d. H a a a a E (cac )H dMa (ψpa )dH Ma (ψp )cd = Γac E dMa (ψc )cd (a) Γac = , Ma where (a) is derived because H a a a E dH Ma (ψc )cd = E dMa (ψc )aMa (qd /Qd ) Ma −1 h i 2q 1 X −jπm(ψca − Q d +1) d E e = Ma m=0 M a −1 Z 1 X e−jπmφ (b) 1 = dφ Ma m=0 −1 2 = 1 . Ma Note that (b) is derived because φ, ψca ∼ U(−1, 1) with the . d definition of φ = ψca − 2q Qd + 1 for any qd ∈ {1, · · · , Qd } in (16). A PPENDIX D Q UANTIZATION PERFORMANCE OF ZBc The normalized beamforming gain between the effective channel vector and the selected combiner is lower bounded as 2 \|eH bc û Rw\| G =E (wH Rw)(eH û Reû ) H \|eû R(eû cos θû + kû sin θû )\|2 (a) = E (wH Rw)(eH û Reû ) " # 2 \|eH eH j∠eH Rk 2 û Reû û Rkû \| û 1+ H tan θû e û cos θû =E wH Rw \|eû Reû \| H (b) eû Reû cos2 θû 'E wH Rw H (c) E[e Reû cos2 θû ] û ' (38) E[wH Rw] . In (38), (a) is based on w = eû cos θû + kû sin θû with eû ⊥ H kû , (b) is approximated because \|eH û Rkû \| \|eû Reû \| and tan θû 1, and (c) is approximated based on Remark 3 in Appendix B. Although Gbc is simplified in (38), it is still difficult to solve in most cases. In the special case of N = 2, the equal gain vectors can be defined as . ejν w = √ [1, ejυ ]T , 2 2πu . 1 eu = √ [1, ej U ]T 2 13 using ν, υ ∼ U(−π, π), and the beamforming gain in (38) is then derived such as H 2 E[eH û Reû cos θû ] (a) E[eû Reû (1 + cos ν̂)] = H E[w Rw] 2E[wH Rw] H PU (b) `=1 E eû Reû (1 + cos ν̂) \| û = ` = 2U E[wH Rw] PU U R Uπ H π e` Re` (1 + cos τ )dτ `=1 2π − U = 2U E[wH Rw] PU 1 H e Re` U π 1 + sin = U `=1H ` 2E[w Rw] π U U π (c) 1 = 1 + sin . 2 π U . Note that (a) is derived based on the definition υ̂ = υ− 2πUû that π π 2π û π follows U(− U , U ) because \|υ− U \| ≤ U , υ ∼ U(−π, π), and û = arg min \|υ − 2πu U \| based on Assumption 2. In addition, u∈{1,··· ,U } (b) is derived by computing its arithmetic mean because û is equally probable from 1 to U , and (c) is derived because E[wH Rw] = (R1,1 + R2,2 )/2 + \|R1,2 \|E[cos(υ + ∠R1,2 )] = (R1,1 + R2,2 )/2, and the arithmetic mean of eH ` Re` is derived as U X eH Re` ` `=1 U U R1,1 + R2,2 X \|R1,2 \| cos 2π` U + ∠R1,2 = + 2 U `=1 = (R1,1 + R2,2 )/2. A PPENDIX E C ORRELATION BETWEEN CHANNEL VECTOR AND DFT CODEWORD We derive a correlation between the channel in (6) and the n-th selected 2D DFT beam h i 2 E hH (cvn ⊗ chn ) X P H 2 ∗ H v v h h =E αp dMv (ψp )cn dMh (ψp )cn p=1 P i h H (a) X 2 v v 2 = E \|αp \|2 E dH E dMh (ψph )chn Mv (ψp )cn p=1 P X 2 2 E \|αp \|2 (b) 2 = E \|αn \| Γnv Γnh + M v Mh p6=n 2 2 P − E \|αn \|2 2 = E \|αn \| Γnv Γnh + , M (b) is derived because h i h i 2 a a 2 a E dH = E dH Ma (ψp )cn Ma (ψp )aMa (qn /Qn ) MX M a −1 a −1 X 2qn a 1 n +1) jπ`(ψpa − 2q −jπm(ψ − +1) p Q Q n n e e = 2E Ma m=0 `=0 a Z Ma −1 M a −1 X 1 ejπ(`−m)ψp a 2qn 1 X E ejπ(m−`)( Qn −1) = 2 dψp Ma 2 −1 m=0 = 1 Ma2 `=0 M a −1 M a −1 X X m=0 m=` 1 = . Ma To complete the formula in (39), we compute the power of the n-th largest channel gain E[\|αn \|2 ]. Without loss of generality, we assume that the magnitude of channel gains are in descending order, i.e., \|α1 \|2 ≥ · · · ≥ \|αP \|2 . The channel gain . A = \|αp \|2 follows χ22 that is characterized by the cumulative distribution function (cdf) of FA (a) = 1 − e−a because αp ∼ CN (0, 1). We now consider the k-th order statistic (k-th smallest order statistic) of P i.i.d exponentially . distributed random variables Ak = \|αP −k+1 \|2 . Then, we refer to [39] for defining the pdf of Ak , yielding P −1 fAk (a) = P (1 − e−a )k−1 e−a(P −k+1) k−1 k−1 P −1 X k−1 (a) = P (−1)q e−a(q+P −k+1) , k − 1 q=0 q where (a) is derived based on the binomial expansion formula. Then, the expectation of k-th order statistic is defined as q Z ∞ X P ! k−1 k−1 q (−1) E Ak = ae−a(q+(P −k+1)) da (P − k)!(k − 1)! 0 q=0 P X P !(−1)q−(P −k+1) (q − 1) · · · (q − (P − k)) = (P − q)!q! q(P − k)! q=P −k+1 P P q−(P −k+1) q−1q−2 q − (P − k) (a) X q (−1) = ··· q 1 2 P −k q=1 " P # P −k P P q+1 X 1 X (b) X P q (−1) = (−1)q + q ` q=1 q q=1 `=1 " # P −k P P X P 1 (c) X 1 X P q 0 (−1) − (−1) + = ` q=0 q 0 q q=1 `=1 (d) = − P −k X `=1 (39) . a a 2 where Γ2na = E dH is derived in Appendix A. Ma (ψn )cn Note that (a) is derived because E[αp∗ αq ] = 0 when p 6= q and sin(π(` − m)) 2qn E ejπ(m−`)( Qn −1) π(` − m) = P 1 X1 + ` q=1 q P X q=P −k+1 1 . q PP Notice that (a) is derived because q=b f (q)(q − b) = PP q=b+1 f (q)(q − b) + f (b)(b − b) for any function f (·), (b) 14 (Pq )(−1)q+1 PP = PP (Pq )(−1)q−1 , (c) q=1 q q Pn 1 (nk)(−1)k+1 is derived based on = k=1 k , and (d) is k Pn k=1 derived based on k=0 (−1)k nk = 0. We now compute the n-th largest channel gain as is derived because q=1 Pn P X 1 E \|αn \|2 = E AP −n+1 = . q q=n Finally, the correlation coefficient is rewritten by plugging \|αn \|2 into (39). P X 1 M Γ2nv Γ2nh − 1 2 E hH (cvn ⊗ chn ) = P+ . M q q=n R EFERENCES [1] J. Song, J. Choi, K. Lee, T. Kim, J. Y. Seol, and D. J. Love, “Advanced quantizer designs for FD-MIMO systems using uniform planar arrays,” in Proceedings of IEEE Global Telecommunications Conference, Dec. 2016. [2] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,” IEEE Transactions on Wireless Communications, vol. 9, no. 11, pp. 3590–3600, Nov. 2010. [3] R. C. Hansen, Phased array antennas, 2nd ed. Hoboken: WileyInterscience, 2009. [4] Y. H. Nam, B. L. Ng, K. Sayana, Y. Li, J. Zhang, Y. Kim, and J. Lee, “Full-dimension MIMO for next generation cellular technology,” IEEE Communications Magazine, vol. 51, no. 6, pp. 172–179, Jun. 2013. [5] R1-150381, Discussions on FD-MIMO codebook enhancements, 3GPP TSG RAN WG1 #80 Std., Feb. 2015. [6] R1-150057, Codebook enhancements for EBF/FD-MIMO, 3GPP TSG RAN WG1 #80 Std., Feb. 2015. [7] R1-150560, Codebook for 2D antenna arrays, 3GPP TSG RAN WG1 #80 Std., Feb. 2015. [8] J. Li, X. Su, J. Zeng, Y. Zhao, S. Yu, L. Xiao, and X. Xu, “Codebook design for uniform rectangular arrays of massive antennas,” in Proceedings of IEEE Vehicular Technology Conference, Jun. 2013. [9] J. Choi, K. Lee, D. J. Love, T. Kim, and R. W. Heath, “Advanced limited feedback designs for FD-MIMO using uniform planar arrays,” in Proceedings of IEEE Global Telecommunications Conference, Dec. 2015. [10] J. Choi, T. Kim, D. J. Love, and J. Y. Seol, “Exploiting the preferred domain of FDD massive MIMO systems with uniform planar arrays,” in Proceedings of IEEE International Conference on Communications, Jun. 2015. [11] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Massive MU-MIMO downlink TDD systems with linear precoding and downlink pilots,” in Proceedings of Allerton Conference on Communication, Control, and Computing, Oct. 2013. [12] R. Rogalin, O. Y. Bursalioglu, H. Papadopoulos, G. Caire, A. F. Molisch, A. Michaloliakos, V. Balan, and K. Psounis, “Scalable synchronization and reciprocity calibration for distributed multiuser MIMO,” IEEE Transactions on Wireless Communications, vol. 13, no. 4, pp. 1815– 1831, Mar. 2014. [13] B. Hassibi and B. M. Hochwald, “How much training is needed in multiple-antenna wireless links?” IEEE Transactions on Information Theory, vol. 49, no. 4, pp. 951–963, Apr. 2003. [14] C. K. Au-Yeung and D. J. Love, “On the performance of random vector quantization limited feedback beamforming in a MISO system,” IEEE Transactions on Wireless Communications, vol. 6, no. 2, pp. 458–462, Feb. 2007. [15] C. K. Au-Yeung, D. J. Love, and S. Sanayei, “Trellis coded line packing: large dimensional beamforming vector quantization and feedback transmission,” IEEE Transactions on Wireless Communications, vol. 10, no. 8, pp. 1844–1853, Apr. 2011. [16] J. Choi, A. Chance, D. J. Love, and U. Madhow, “Noncoherent trellis coded quantization: a practical limited feedback technique for massive MIMO systems,” IEEE Transactions on Communications, vol. 61, no. 12, pp. 5016–5029, Dec. 2013. [17] J. Choi, D. J. Love, and T. Kim, “Trellis-extended codebooks and successive phase adjustment: a path from LTE-advanced to FDD massive MIMO systems,” IEEE Transactions on Wireless Communications, vol. 14, no. 4, pp. 2007–2016, Apr. 2015. [18] J. Choi, D. J. Love, and P. Bidigare, “Downlink training techniques for FDD massive MIMO systems: open-loop and closed-loop training with memory,” IEEE Journal of Selected Topics in Signal Processing, vol. 8, no. 5, pp. 802–814, Dec. 2014. [19] S. Noh, M. D. Zoltowski, Y. Sung, and D. J. Love, “Pilot beam pattern design for channel estimation in massive MIMO systems,” IEEE Journal of Selected Topics in Signal Processing, vol. 8, no. 5, pp. 787–801, Oct. 2014. [20] Y. Han, J. Lee, and D. J. Love, “Compressed sensing-aided downlink channel training for FDD massive MIMO systems,” to appear in IEEE Transactions on Communications, 2017. [21] K. K. Mukkavilli, A. Sabharwal, E. Erkip, and B. Aazhang, “On beamforming with finite rate feedback in multiple-antenna systems,” IEEE Transactions on Information Theory, vol. 49, no. 10, pp. 2562– 2579, Jan. 2003. [22] D. J. Love, R. W. Heath, and T. Strohmer, “Grassmannian beamforming for multiple-input multiple-output wireless systems,” IEEE Transactions on Information Theory, vol. 49, no. 10, pp. 2735–2747, Oct. 2003. [23] D. J. Love, R. W. Heath, V. K. Lau, D. Gesbert, B. D. Rao, and M. Andrews, “An overview of limited feedback in wireless communication systems,” IEEE Journal on Selected Areas in Communications, vol. 26, no. 8, pp. 1341–1365, Oct. 2008. [24] D. J. Love and R. W. Heath, “Grassmannian beamforming on correlated MIMO channels,” in Proceedings of IEEE Global Telecommunications Conference, Nov. 2004. [25] ——, “Limited feedback diversity techniques for correlated channels,” IEEE Transactions on Vehicular Technology, vol. 55, no. 2, pp. 718–722, Mar. 2006. [26] P. Xia and B. G. Georgios, “Design and analysis of transmit-beamforming based on limited-rate feedback,” IEEE Transactions on Signal Processing, vol. 54, no. 5, pp. 1853–1863, May 2006. [27] V. Raghavan, A. M. Sayeed, and V. V. Veeravalli, “Limited feedback precoder design for spatially correlated MIMO channels,” in Proceedings of Conference on Information Sciences and Systems, Mar. 2007. [28] V. Raghavan, J. Choi, and D. J. Love, “Design guidelines for limited feedback in the spatially correlated broadcast channel,” IEEE Transactions on Communications, vol. 63, no. 7, pp. 2524–2540, Jul. 2015. [29] Study on 3D channel model for LTE, 3GPP TR 36.873 V12.0.0 Std., Sep. 2014. [30] Spatial channel model for mutiple input multiple output simulations, 3GPP TR 25.996 V6.1.0 Std., Sep. 2003. [31] R1-103378, Performance evaluations of Rel.10 feedback framework, 3GPP TSG RAN WG1 #61 Std., May 2010. [32] R1-105011, Way forward on 8Tx codebook for Rel.10 DL MIMO, 3GPP TSG RAN WG1 #61 Std., Aug. 2010. [33] Y. Han, S. Jin, X. Li, Y. Huang, L. Jiang, and G. Wang, “Design of double codebook based on 3D dual-polarized channel for multiuser MIMO system,” EURASIP Journal on Advances in Signal Processing, vol. 2014, no. 1, pp. 1–13, Jul. 2014. [34] R. J. Mailloux, Phased array antenna handbook, 2nd ed. Artech House, 2005. [35] D. Ying, F. W. Vook, T. Thomas, D. J. Love, and A. Ghosh, “Kronecker product correlation model and limited feedback codebook design in a 3D channel model,” in Proceedings of IEEE International Conference on Communications, Jun. 2014. [36] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE Transactions on Wireless Communications, vol. 13, no. 3, pp. 1499–1513, Mar. 2014. [37] M. Borga, “Learning multidimensional signal processing,” Ph.D. dissertation, Linköping University, Linköping, Sweden, 1998, SE-581 83. [38] A. Stuart and K. Ord, Kendall’s advanced theory of statistics: distribution theory, 6th ed. London: Arnold, 1998. [39] H. David, Order statistics, 2nd ed. New York: John Wiley and Sons, 1980.	7cs.IT
Fast Distributed Algorithms for Testing Graph Properties Keren Censor-Hillel∗ Eldar Fischer∗ Gregory Schwartzman∗ Yadu Vasudev∗ arXiv:1602.03718v3 [cs.DC] 2 May 2016 May 3, 2016 Abstract We initiate a thorough study of distributed property testing – producing algorithms for the approximation problems of property testing in the CONGEST model. In particular, for the so-called dense graph testing model we emulate sequential tests for nearly all graph properties having 1-sided tests, while in the general and sparse models we obtain faster tests for trianglefreeness, cycle-freeness and bipartiteness, respectively. In addition, we show a logarithmic lower bound for testing bipartiteness and cycle-freeness, which holds even in the stronger LOCAL model. In most cases, aided by parallelism, the distributed algorithms have a much shorter running time as compared to their counterparts from the sequential querying model of traditional property testing. The simplest property testing algorithms allow a relatively smooth transitioning to the distributed model. For the more complex tasks we develop new machinery that may be of independent interest. ∗ Technion – Israel Institute of Technology, Department of Computer Science. ckeren@cs.technion.ac.il, eldar@cs.technion.ac.il, gregorys@cs.technion.ac.il, yaduvasudev@gmail.com. Supported in part by the Israel Science Foundation (grant 1696/14). 1 Introduction The performance of many distributed algorithms naturally depends on properties of the underlying network graph. Therefore, an inherent goal is to check whether the graph, or some given subgraph, has certain properties. However, in some cases this is known to be hard, such as in the CONGEST model [31]. In this model, computation proceeds in synchronous rounds, in each of which every vertex can send an O(log n)-bit message to each of its neighbors. Lower bounds for the number of √ rounds of type Ω̃( n + D) are known for verifying many global graph properties, where n is the number of vertices in the network and D is its diameter (see, e.g. Das-Sarma et al. [36])1 . To overcome such difficulties, we adopt the relaxation used in graph property testing, as first defined in [18, 20], to the distributed setting. That is, rather than aiming for an exact answer to the question of whether the graph G satisfies a certain property P , we settle for distinguishing the case of satisfying P from the case of being ǫ-far from it, for an appropriate measure of being far. Apart from its theoretical interest, this relaxation is motivated by the common scenario of having distributed algorithms for some tasks that perform better given a certain property of the network topology, or given that the graph almost satisfies that property. For example, Hirvonen et al. [24] show an algorithm for finding a large cut in triangle-free graphs (with additional constraints), and for finding an (1 − ǫ)-approximation if at most an ǫ fraction of all edges are part of a triangle. Similarly, Pettie and Su [32] provide fast algorithms for coloring triangle-free graphs. We construct fast distributed algorithms for testing various graph properties. An important byproduct of this study is a toolbox that we believe will be useful in other settings as well. 1.1 Our contributions We provide a rigorous study of property testing methods in the realm of distributed computing under the CONGEST model. We construct 1-sided error distributed ǫ-tests, in which if the graph satisfies the property then all vertices output accept, and if it is ǫ-far from satisfying the property then at least one vertex outputs reject with probability at least 2/3. Using the standard amplification method of invoking such a test O(log n) times and having a vertex output reject if there is at least one invocation in which it should output reject, gives rejection with higher probability at the price of a multiplicative O(log n) factor for the number of rounds. The definition of a graph being ǫ-far from satisfying a property is roughly one of the following (see Section 2 for precise definitions): (1) Changing any ǫn2 entries in the adjacency matrix does not give a graph that satisfies the property (dense model), or (2) changing any ǫ · max{n, m} entries in the adjacency matrix does not give a graph that satisfies the property, where m is the number of edges (general model). A particular case here is when the degrees are bounded by some constant d, and any resulting graph must comply with this restriction as well (sparse model). In a sequential ǫ-test, access to the input is provided by queries, whose type depends on the model. In the dense model these are asking whether two vertices v, u are neighbors, and in the general and sparse models these can be either asking what the degree of a vertex v is, or asking what the i-th neighbor of v is (the ordering of neighbors is arbitrary). While a sequential ǫ-test can touch only a small handful of vertices with its queries, in a distributed test the lack of ability to communicate over large distances is offset by having all n vertices operating in parallel. Our first contribution is a general scheme for a near-complete emulation in the distributed 1 Here Ω̃ hides factors that are polylogarithmic in n. 1 context of ǫ-tests originating from the dense graph model (Section 3). This makes use of the fact that in the dense model all (sequential) testing algorithms can be made non-adaptive, which roughly means that queries do not depend on responses to previous queries (see Section 2 for definition). In fact, such tests can be made to have a very simple structure, allowing the vertices in the distributed model to “band together” for an emulation of the test. There is only one additional technical condition (which we define below), since in the distributed model we cannot handle properties whose counter-examples can be “split” to disjoint graphs. For example, the distributed model cannot hope to handle the property of the graph having no disjoint union of two triangles, a property for which there exists a test in the dense model. Theorem 3.4 Any ǫ-test in the dense graph model for a non-disjointed property that makes q queries can be converted to a distributed ǫ-test that takes O(q 2 ) communication rounds. We next move away from the dense graph model to the sparse and general models, that are sometimes considered to be more realistic. In the general model, there exists no test for the property of containing no triangle that makes a number of queries independent of the number of graph vertices [2]. Here the distributed model can do better, because the reason for this deficiency is addressed by having all vertices operate concurrently. In Section 4 we adapt the interim lemmas used in the best testing algorithm constructed in [2], and construct a distributed algorithm whose number of rounds is independent of n. Theorem 4.1 Algorithm 2 is a distributed ǫ-test in the general graph model for the property of containing no triangles, that requires O(ǫ−2 ) rounds. The sparse and general models inherently require adaptive property testing algorithms, since there is no other way to trace a path from a given vertex forward, or follow its neighborhood. Testing triangle freeness sequentially uses adaptivity only to a small degree. However, other problems in the sparse and general models, such as the one we explore next, have a high degree of adaptivity built into their sequential algorithms, and we need to take special care for emulating it in the distributed setting. In the sparse model (degrees bounded by a constant d), we adapt ideas from the bipartiteness testing algorithm of [19], in which we search for odd-length cycles. Here again the performance of a distributed algorithm surpasses that of the test (a number of rounds polylogarithmic in n vs. a √ number of queries which is Ω( n) – a lower bound that is given in [20]). The following is proved in Section 5. Theorem 5.2 Algorithm 4 is a distributed ǫ-test in the bounded degree graph model for the property of being bipartite, that requires O(poly(ǫ−1 log(nǫ−1 ))) rounds. In the course of proving Theorem 5.2 we develop a method that we consider to be of independent interest2 . The algorithm works by performing 2n random walks concurrently (two starting from each vertex). The parallel execution of random walks despite the congestion restriction is achieved by making sure that the walks have a uniform stationary distribution, and then showing that congestion is “close to average”, which for the uniform stationary distribution is constant. In Section 6 we show a fast test for cycle-freeness. This makes use of a combinatorial lemma that we prove, about cycles that remain in the graph after removing edges independently with probability ǫ/2. The following summarizes our result for testing cycle-freeness. 2 This technique was recently independently and concurrently devised in [17] for a different use. 2 Theorem 6.3 Algorithm 6 is a distributed ǫ-test in the general graph model for the property of being cycle-free, that requires O(log n/ǫ) rounds. We also prove lower bounds for testing bipartiteness and cycle-freeness (matching the upper bound for the latter). Roughly speaking, these are obtained by using the probabilistic method with alterations to construct graphs which are far from being bipartite or cycle-free, but all of their cycles are of length that is at least logarithmic. This technique bears some similarity to the classic result by Erdös [13], which showed the existence of graphs with large girth and large chromatic number. The following are given in Section 7. Theorem 7.1 Any distributed 1/100-test for the property of being bipartite requires Ω(log n) rounds of communication. Theorem 7.3 Any distributed 1/100-test for the property of being cycle-free requires Ω(log n) rounds of communication. Roadmap: The paper is organized as follows. The remainder of this section consists of related work and historical background on property testing. Section 2 contains formal definitions and some mathematical tools. The emulation of sequential tests for the dense model is given in Section 3. In Section 4 we give our distributed test for triangle-freeness. In Section 5 we provide a distributed test for bipartiteness, along with our new method of executing many random walks, and in Section 6 we give our test for cycle-freeness. Section 7 gives our logarithmic lower bounds for testing bipartiteness and cycle-freeness. We conclude with a short discussion in Section 8. 1.2 Related work The only previous work that directly relates to our distributed setting is due to Brakerski and Patt-Shamir [8]. They show a tolerant property testing algorithm for finding large (linear in size) near-cliques in the graph. An ǫ-near clique is a set of vertices for which all but an ǫ-fraction of the pairs of vertices have an edge between them. The algorithm is tolerant, in the sense that it finds a linear near-clique if there exists a linear ǫ3 -near clique. That is, the testing algorithm considers two thresholds of being close to having the property (in this case – containing a linear size clique). We are unaware of any other work on property testing in this distributed setting. Testing in a different distributed setting was considered in Arfaoui et al. [5]. They study testing for cycle-freeness, in a setting where each vertex may collect information of its entire neighborhood up to some distance, and send a short string of bits to a central authority who then has to decide whether the graph is cycle-free or not. Related to having information being sent to, or received by, a central authority, is the concept of proof-labelling schemes, introduced by Korman et al. [27] (for extensions see, e.g., Baruch et al. [6]). In this setting, each vertex is given some external label, and by exchanging labels the vertices need to decide whether a given property of the graph holds. This is different from our setting in which no information other than vertex IDs is available. Another setting that is related to proof-labelling schemes, but differs from our model, is the prover-verifier model of Foerster et al. [15]. Sequential property testing has the goal of computing without processing the entire input. The wider family of local computation algorithms (LCA) is known to have connections with distributed computing, as shown by Parnas and Ron [30] and later used by others. A recent study by Göös et 3 al. [23] proves that under some conditions, the fact that a centralized algorithm can query distant vertices does not help with speeding up computation. However, they consider the LOCAL model, and their results apply to certain properties that are not influenced by distances. Finding induced subgraphs is a crucial task and has been studied in several different distributed models (see, e.g., [9, 11, 12, 26]). Notice that for finding subgraphs, having many instances of the desired subgraph can help speedup the computation, as in [11]. This is in contrast to algorithms that perform faster if there are no or only few instances, as explained above, which is why we test for, e.g., the property of being triangle-free, rather for the property of containing triangles. (Notice that these are not the same, and in fact every graph with 3/ǫ or more vertices is ǫ-close to having a triangle.) Parallelizing many random walks was addressed in [1], where the question of graph covering via random walks is discussed. It is shown there that for certain families of graphs there is a substantial speedup in the time it takes for k walks starting from the same vertex to cover the graph, as compared to a single walk. No edge congestion constraints are taken into account. √ In [37], it is shown how to perform,√ under congestion, a single random walk of length L in Õ( LD) rounds, and k random walks in Õ( kLD + k) rounds, where D is the diameter of the graph. Our method has no dependence on the diameter, allowing us to perform a multitude of short walks much faster. 1.3 Historical overview The first papers to consider the question of property testing were [7] and [35]. The original motivations for defining property testing were its connection to some Computerized Learning models, and the ability to leverage some properties to construct Probabilistically Checkable Proofs (PCPs – this is related to property testing through the areas of Locally Testable Codes and Locally Decodable Codes, LTCs and LDCs). Other motivations since then have entered the fray, and foremost among them are sublinear-time algorithms, and other big-data considerations. Since virtually no property can be decidable without reading the entire input, property testing introduces a notion of the allowable approximation to the original problem. In general, the algorithm has to distinguish inputs satisfying the property, from inputs that are ǫ-far from it. For more information on the general scheme of “classical” property testing, consult the surveys [14, 21, 33]. The older of the graph testing models discussed here is the dense model, as defined in the seminal work of Goldreich, Goldwasser and Ron [18]. The dense graph model has historically kickstarted combinatorial property testing in earnest, but it has some shortcomings. Its main one is the distance function, which makes sense only if we consider graphs having many edges (hence the name “dense model”) – any graph with o(n2 ) edges is indistinguishable in this model from an empty graph. The stricter and at times more plausible distance function is one which is relative to the actual n number of edges, rather than the maximum 2 . The general model was defined in [2], while the sparse model was defined already in [20]. The main difference between the sparse and the general graph models is that in the former there is also a guaranteed upper bound d on the degrees of the vertices, which is given to the algorithm in advance (the query complexity may then depend on d, either explicitly, or more commonly implicitly by considering d to be a constant). 4 2 Preliminaries 2.1 Additional background on property testing While the introduction provided rough descriptions of the different property testing models, here we provide more formal definitions. The dense model for property testing is defined as follows. Definition 2.1 (dense graph model [18]). The dense graph model considers as objects graphs that are given by their adjacency matrix. Hence it is defined by the following features. • Distance: Two graphs with n vertices each are considered to be ǫ-close if one can be obtained from the other by deleting and inserting at most ǫn2 edges (this is, up to a constant factor, the same as the normalized Hamming distance). • Querying scheme: A single query of the algorithm consists of asking whether two vertices u, v ∈ V form a graph edge in E or not. • Allowable properties: All properties have to be invariant under permutations of the input that pertain to graph isomorphisms (a prerequisite for them being graph properties). The number of vertices n is given to the algorithm in advance. As discussed earlier, the sparse and general models for property testing relate the distance function to the actual number of edges in the graph. They are formally defined as follows. Definition 2.2 (sparse [20] and general [2] graph models). These two models consider as objects graphs given by their adjacency lists. They are defined by the following features. • Distance: Two graphs with n vertices and m edges (e.g. as defined by the denser of the two) are considered to be ǫ-close if one can be obtained from the other by deleting and inserting at most ǫ max{n, m} edges3 . • Querying scheme: A single query consists of either asking what is the degree of a vertex v, or asking what is the i’th neighbor of v (the ordering of neighbors is arbitrary). • Allowable properties: All properties have to be invariant under graph isomorphisms (which here translate to a relabeling that affects both the vertex order and the neighbor ids obtained in neighbor queries), and reordering of the individual neighbor lists (as these orderings are considered arbitrary). In this paper, we mainly refer to the distance functions of these models, and less so to the querying scheme, since the latter will be replaced by the processing scheme provided by the distributed computation model. Note that most property testing models get one bit in response to a query, e.g., “yes/no” in response to “is uv an edge” in the dense graph model. However, the sparse and general models may receive log n bits of information for one query, e.g., an id of a neighbor of a vertex. Also, the degree of a vertex, which can be given as an answer to a query in the general model, takes log n bits. Since the distributed CONGEST model allows passing a vertex id or a vertex degree along an edge in O(1) rounds, we can equally relate to all three graph models. Another important point is the difference between 1-sided and 2-sided testing algorithms, and the difference between non-adaptive and adaptive algorithms. 3 Sometimes in the sparse graph model the allowed number of changes is ǫdn, as relates to the maximum possible number of edges; when d is held constant the difference is not essential. 5 Definition 2.3 (types of algorithms). A property testing algorithm is said to have 1-sided error if there is no possibility of error on accepting satisfying inputs. That is, an input that satisfies the property will be accepted with probability 1, while an input ǫ-far from the property will be rejected with a probability that is high enough (traditionally this means a probability of at least 2/3). A 2-sided error algorithm is also allowed to reject satisfying inputs, as long as the probability for a correct answer is high enough (traditionally at least 2/3). A property testing algorithm is said to be non-adaptive if it decides all its queries in advance (i.e. based only on its internal coin tosses and before receiving the results of any query), while only its accept/reject output may depend on the actual input. An adaptive algorithm may make each query in turn based on the results of its previous queries (and, as before, possible internal coin tosses). In the following we address both adaptive and non-adaptive algorithms. However, we restrict ourselves to 1-sided error algorithms, since the notion of 2-sided error is not a good match for our distributed computation model. 2.2 Mathematical background An important role in our analyses is played by the Multiplicative Chernoff Bound (see, e.g., [29]), hence we state it here for completeness. Fact 2.4. Suppose that X1 , ..., Xn are independent random variables taking values in {0, 1}. Let X denote their sum and let µ = E[X] denote its expected value. Then, for any δ > 0, e−δ )µ , (1 − δ)(1−δ) eδ P r[X > (1 + δ)µ] < ( )µ . (1 + δ)(1+δ) P r[X < (1 − δ)µ] < ( Some convenient variations of the bounds above are: P r[X ≥ (1 + δ)µ] < e−δµ/3 , −δ2 µ/3 P r[X ≥ (1 + δ)µ] < e , δ ∈ (0, 1) P r[X ≤ (1 − δ)µ] < e , δ ∈ (0, 1). −δ2 µ/2 3 δ≥1 Distributed emulation of sequential tests in the dense model We begin by showing that under a certain assumption of being non-disjointed, which we define below, a property P that has a sequential test in the dense model that requires q queries can be tested in the distributed setting within O(q 2 ) rounds. We prove this by constructing an emulation that translates sequential tests to distributed ones. For this we first introduce a definition of a witness graph and then adapt [22, Theorem 2.2], restricted to 1-sided error tests, to our terminology. Definition 3.1. Let P be a property of graphs with n vertices. Let G′ be a graph with k < n vertices. We say that G′ is a witness against P , if it is not an induced subgraph of any graph that satisfies P . 6 Notice that if G′ has an induced subgraph H that is a witness against P , then by the above definition G′ is also a witness against P . The work of [22] transforms tests of graphs in the dense model to a canonical form where the query scheme is based on vertex selection. This is useful in particular for the distributed model, where the computational work is essentially based in the vertices. We require the following special case for 1-sided error tests. Lemma 3.2 ( [22, Theorem 2.2]). Let P be a property of graphs with n vertices. If there exists a 1-sided error ǫ-test for P with query complexity q(n, ǫ), then there exists a 1-sided error ǫ-test for P that uniformly selects a set of q ′ = 2q(n, ǫ) vertices, and accepts if and only if the induced subgraph is not a witness against P . Our emulation leverages Lemma 3.2 under an assumption on the property P , which we define as follows. Definition 3.3. We say that P is a non-disjointed property if for every graph G that does not satisfy P and an induced subgraph G′ of G such that G′ is a witness against P , G′ has some connected component which is also a witness against P . We call such components witness components. We are now ready to formally state our main theorem for this section. Theorem 3.4. Any ǫ-test in the dense graph model for a non-disjointed property that makes q queries can be converted to a distributed ǫ-test that takes O(q 2 ) communication rounds. The following lemma essentially says that not satisfying a non-disjointed property cannot rely on subgraphs that are not connected, which is exactly what we need to forbid in a distributed setting. Lemma 3.5. The property P is a non-disjointed property if and only if all minimal witnesses that are induced subgraphs of G are connected. Here minimal refers to the standard terminology, which means that no proper induced subgraph is a witness against P . Proof. First, if P is non-disjointed and G does not satisfy P , then for every subgraph G′ of G that is a witness against P , G′ has a witness component. If G′ is minimal then it must be connected, since otherwise it contains a connected component which is a witness against P , which contradicts the minimality of G. For the other direction, if all the minimal witnesses that are induced subgraphs of G are connected, then every induced subgraph G′ that is a witness against P is either minimal, in which case it is connected, or is not minimal, in which case there is a subgraph H of G′ which is connected and a minimal witness against P . The connected component C of G′ which contains H is a witness against P (otherwise H is not a witness against P ), and hence it follows that P is non-disjointed. Next, we give the distributed test (Algorithm 1). The test has an outer loop in which each vertex picks itself with probability 5q/n, collects its neighborhood of a certain size of edges between picked vertices in an inner loop, and rejects if it identifies a witness against P . The outer loop repeats two times because not only does the sequential test have an error probability, but also with some 7 small probability we may randomly pick too many or not enough vertices in order to emulate it. Repeating the main loop twice reduces the error probability back to below 1/3. In the inner loop, each vertex collects its neighborhood of picked vertices and checks if its connected component is a witness against P . To limit communications this is done only for components of picked vertices that are sufficiently small: if a vertex detects that it is part of a component with too many edges then it accepts and does not participate until the next iteration of the outer loop. Algorithm 1: Emulation algorithm with input q for property P Variables: Uv edges known to v, Uv′ edges to update and send (temporary variables) 1 perform 2 times 2 reset the state for all vertices 3 for each vertex v simultaneously 4 Vertex v picks itself with probability 5q/n 5 if v is picked then 6 Notify all neighbors that v is picked 7 Set Uv′ = {(v, u) ∈ E\| u is picked} and Uv = ∅ 8 perform 10q times # At each iteration Uv is a subgraph of v’s connected component 9 Uv′ = Uv′ \Uv # only need recently discovered edges 10 Uv = Uv ∪ Uv′ # add them to Uv 11 if \|Uv \| ≤ 100q 2 then # don’t operate if there are too many edges 12 Send Uv′ to all picked neighbours of v # propagate known edges 13 14 15 16 17 18 19 Wait until the time bound for all other vertices to finish this iteration Set Uv′ to the union of edge sets received from neighbors if Uv ∪ Uv′ is a witness against P then Vertex v outputs reject (ending all operations) else Wait until the time bound for all other vertices to finish this iteration of the outermost loop Every vertex v that did not reject outputs accept To analyze the algorithm, we begin by proving that there is a constant probability for the number of picked vertices to be sufficient and not too large. Lemma 3.6. The probability that the number of vertices picked by the algorithm is between q and 10q is more than 2/3 . Proof. For every v ∈ V , we denote by Xv the indicator variable for the event that vertex v is picked. P Note that these are all independent random variables. Using the notation X = v∈V Xv gives that E[X] = 5q, because each vertex is picked with probability 5q/n. Using the Chernoff Bound from Fact 2.4 with δ = 4/5 and µ = 5q, we can bound the probability of having too few picked vertices: P r[X < q] = P r[X < (1 − δ)µ] < ( 1 5 e−4/5 )5q = ( 4 )q < . e 10 (1 − (4/5))(1−(4/5)) 8 For bounding the probability that there are too many picked vertices, we use the other direction of the Chernoff Bound with δ = 1 and µ = 5q, giving: P r[X > 10q] = P r[X > (1 + δ)µ] < ( e5 q 2 e 5q ) = ( ) < . 22 210 10 Thus, with probability at least 2/3 it holds that q ≤ X ≤ 10q. Now, we can use the guarantees of the sequential test to obtain the guarantees of our algorithm. Lemma 3.7. Let P be a non-disjointed graph property. If G satisfies P then all vertices output accept in Algorithm 1. If G is ǫ-far from satisfying P , then with probability at least 2/3 there exists a vertex that outputs reject. Proof. First, assume that G satisfies P . Vertex v outputs reject only if it is part of a witness against P , which is, by definition, a component that cannot be extended to some H that satisfies P . However, every component is an induced subgraph of G itself, which does satisfy P , and thus every component can be extended to G. This implies that no vertex v outputs reject. Now, assume that G is ǫ-far from satisfying P . Since the sequential test rejects with probability at least 2/3, the probability that a sample of at least q vertices induces a graph that cannot be extended to a graph that satisfies P is at least 2/3 . Because P is non-disjointed, the induced subgraph must have a connected witness against P . We note that a sample of more than q vertices does not reduce the rejection probability. Hence, if we denote by A the event that the subgraph induced by the picked vertices has a connected witness against P , then P r[A] ≥ 2/3, conditioned on that at least q vertices were picked. However, a sample that is too large may cause a vertex to output accept because it cannot collect its neighborhood. We denote by B the event that the number of vertices sampled is between q and 10q, and by Lemma 3.6 its probability is at least 2/3. We bound P r[A ∩ B] using Bayes’ Theorem, obtaining P r[A ∩ B] = P r[A\|B]P r[B] ≥ (2/3)2 . Since the outer loop consists of 2 independent iterations, this gives a probability of at least 1 − (1 − 4/9)2 ≥ 2/3 for having a vertex that outputs reject. We now address the round complexity. Each vertex only sends and receives information from its q-neighborhood about edges between the chosen vertices. If too many vertices are chosen we detect this and accept. Otherwise we only communicate the chosen vertices and their edges, which requires O(q 2 ) communication rounds using standard pipelining 4 . Together with Lemma 3.7, this proves Theorem 3.4. 3.1 Applications: k-colorability and perfect graphs Next, we provide some examples of usage of Theorem 3.4. A result by Alon and Shapira [4] states that all graph properties closed under induced subgraphs are testable in a number of queries that depends only on ǫ−1 . We note that, except for certain specific properties for which there are ad-hoc proofs, the dependence is usually a tower function in ǫ−1 or worse (asymptotically larger). From this, together with Lemma 3.2 and Theorem 3.4, we deduce that if P is a non-disjointed property closed under induced subgraphs, then it is testable, for every fixed ǫ, in a constant number of communication rounds. 4 Pipelining means that each vertex has a buffer for each edge, which holds the information (edges between chosen vertices, in our case) it needs to send over that edge. The vertex sends the pieces of information one after the other. 9 Example – k-colorability: The property of being k-colorable is testable in a distributed manner by our algorithm. All minimal graphs that are witnesses against P (not k-colorable) are connected, and therefore according to Lemma 3.5 it is a non-disjointed property. It is closed under induced subgraphs, and by [3] there exists a 1-sided error ǫ-test for k-colorability that uniformly picks O(k log(k)/ǫ2 ) vertices, and its number of queries is the square of this expression (note that the polynomial dependency was already known by [18]). Our emulation implies a distributed 1-sided error ǫ-test for k-colorability that requires O(poly(kǫ−1 )) rounds. Example – perfect graphs: A graph G is said to be perfect if for every induced subgraph G′ of G, the chromatic number of G′ equals the size of the largest clique in G′ . Another characterization of a perfect graph is via forbidden subgraphs: a graph is perfect if and only if it does not have odd holes (induced cycles of odd length at least 5) or odd anti-holes (the complement graph of an odd hole) [10]. Both odd holes and odd anti-holes are connected graphs. Since these are all minimal witnesses against the property, according to Lemma 3.5 it is a non-disjointed property. Using the result of Alon-Shapira [4] we know that the property of a graph being perfect is testable. Our emulation implies a distributed 1-sided error ǫ-test for being a perfect graph that requires a number of rounds that depends only on ǫ. 4 Distributed test for triangle-freeness In this section we show a distributed ǫ-test for triangle-freeness. Notice that since triangle-freeness is a non-disjointed property, Theorem 3.4 gives a distributed ǫ-test for triangle-freeness under the dense model with a number of rounds that is O(q 2 ), where q is the number of queries required for a sequential ǫ-test for triangle-freeness. However, for triangle-freeness, the known number of queries is a tower function in log(1/ǫ) [16]. Here we leverage the inherent parallelism that we can obtain when checking the neighbors of a vertex, and show a test for triangle-freeness that requires only O(ǫ−2 ) rounds (Algorithm 2). Importantly, our algorithm works not only for the dense graph model, but for the general graph model (where distances are relative to the actual number of edges), which subsumes it. In the sequential setting, a test for triangle-freeness in the general model requires a number of queries that is some constant power of n by [2]. Our proof actually follows the groundwork laid in [2] for the general graph model – their algorithm picks a vertex and checks two of its neighbors for being connected, while we perform the check for all vertices in parallel. Theorem 4.1. Algorithm 2 is a distributed ǫ-test in the general graph model for the property of containing no triangles, that requires O(ǫ−2 ) rounds. Our line of proof follows that of [2], by distinguishing edges that connect two high-degree vertices √ −1 from those that do not. Formally, let b = 2 ǫ m, where m is the number of edges in the graph, and denote B = {v ∈ V \| deg(v) ≥ b}. We say that an edge e = (u, v) is light if v 6∈ B or u 6∈ B, and otherwise, we say that it is heavy. That is, the set of heavy edges is H = {(u, v) ∈ E \| u ∈ B, v ∈ B}. We begin with the following simple claim about the number of heavy edges. Claim 4.2. The number of heavy edges, \|H\|, is at most ǫm/2. Proof. The number of heavy edges is \|H\| ≤ \|B\|(\|B\| − 1)/2 < \|B\|2 /2. Since \|B\|b ≤ 2m, we get that √ √2m = ǫm. This gives that \|H\| ≤ 21 \|B\|2 ≤ ǫm/2. \|B\| ≤ 2m −1 b = 2 ǫ m 10 Algorithm 2: Triangle freeness test 1 2 3 4 5 6 7 8 9 10 11 12 for each vertex v simultaneously perform 32ǫ−2 times Pick w1 , w2 ∈ N (v), w1 6= w2 uniformly at random Send w2 to w1 # Ask w1 if it is a neighbor of w2 foreach wu sent by u ∈ N (v) do # Asked by u if v is a neighbor of w if wu ∈ N (v) then Send “yes” to u else Send “no” to u if received “yes” from w1 then reject (ending all operations) accept (for vertices that did not reject) Next, we fix an iteration i of the algorithm. Every vertex v chooses two neighbors w1 , w2 . Let A = {(v, w1 ) ∈ E \| v ∈ V \ B}, where w1 is the first of the two vertices chosen by the low-degree vertex v. Let T = {e ∈ E \| e is a light edge in a triangle}, and let AT = T ∩ A. We say that an edge (v, w1 ) ∈ AT is matched if (v, w2 ) is in the same triangle as (v, w1 ). If (v, w1 ) ∈ AT is matched then {v, w1 , w2 } is a triangle that is detected by v. We begin with the following lemma that states that if G is ǫ-far from being triangle-free, then in any iteration i we can bound the expected number of matched edges from below by ǫ2 /8. Let Y be the number of matched edges. Lemma 4.3. The expected number of matched edges by a single iteration of the algorithm, E[Y ], is greater than ǫ2 /8. Proof. For every e ∈ AT , let Ye be a random variable indicating whether e is matched. Then P Y = e∈AT Ye , giving the following bound: E[Y \|AT ] = E[ X e∈AT Ye \|AT ] = X e∈AT P r[e is matched] ≥ \|AT \|/b, (1) where the last inequality follows because a light edge in AT is chosen by a vertex with degree at most b, hence the third triangle vertex gets picked with probability at least 1/b. Next, we argue that E[\|AT \|] ≥ \|T \|/b. To see why, for every edge e, let Xe be a random variable P indicating whether e ∈ A. Let X = e∈T Xe = \|AT \|. Then, E[\|AT \|] = E[X] = E[ X e∈T Xe ] = X e∈T E[Xe ] = X e∈T P r[e ∈ A] ≥ \|T \|/b, (2) where the last inequality follows because a light edge has at least one endpoint with degree at most b. Hence, this edge gets picked by it with probability at least 1/b. It remains to bound \|T \| from below, for which we claim that \|T \| ≥ ǫm/2. To prove this, first notice that, since G is ǫ-far from being triangle free, it has at least ǫm triangle edges, since otherwise 11 we can just remove all of them and make the graph triangle free with less than ǫm edge changes. By Claim 4.2, the number of heavy edges satisfies \|H\| ≤ ǫ/2m. Subtracting this from the number of triangle edges gives that at least ǫm/2 edges are light triangle edges, i.e., \|T \| ≥ ǫm/2. (3) Finally, by Inequalities (1), (2) and (3), using iterated expectation we get: E[Y ] = EAT [E[Y \|AT ]] ≥ E[ \|T \| ǫm 1 \|AT \| ]≥ 2 ≥ = ǫ2 /8. b b 2 4ǫ−1 m We can now prove the correctness of our algorithm, as follows. Lemma 4.4. If G is triangle-free then all vertices output accept in Algorithm 2. If G is ǫ-far from being triangle-free, then with probability at least 2/3 there exists a vertex that outputs reject. Proof. If G is triangle free then in each iteration v receives “no” from w1 and after all iterations it returns accept. Assume that G is ǫ-far from being triangle-free. Let Zi,v be an indicator variable for the event that vertex v detects a triangle at iteration i. First, we note that the indicators are independent, since a vertex detecting a triangle does not affect the chance of another vertex detecting a triangle (note that the graph is fixed), and the iterations are done independently. Now, let P −2 P number of detections over all iterations. Z = 32ǫ i=1 v∈V Zi,v , and notice that Z is the total P Lemma 4.3 implies that for a fixed i, it holds that E[ v∈V Zi,v ] = E[Y ] ≥ ǫ2 /8, which sums to:  −2 32ǫ X E[Z] = E  X i=1 v∈V  Zi,v  = −2 32ǫ X i=1 E " X v # Zi,v ≥ −2 32ǫ X ǫ2 /8 = 4. i=1 Using the Chernoff Bound from Fact 2.4 with δ = 3/4 and µ ≥ 4 gives P r[Z < 1] ≤ P r[Z < (1 − δ)µ] < ( e−3/4 )4 = 4/e3 < 2/3, (1 − (3/4))(1−(3/4)) and hence with probability at least 2/3 at least one triangle is detected and the associated vertex outputs reject, which completes the proof. In every iteration, each vertex initiates only two messages of size O(log n) bits, one sent to w1 and one sent back by w1 . Since there are O(ǫ−2 ) iterations, this implies that the number of rounds is O(ǫ−2 ) as well. This, together with Lemma 4.4, completes the proof of Theorem 4.1. 5 Distributed bipartiteness test for bounded degree graphs In this section we show a distributed ǫ-test for being bipartite for graphs with degrees bounded by d. Our test builds upon the sequential test of [19] and, as in the case of triangle freeness, takes advantage of the ability to parallelize queries. While the number of queries of the sequential test is 12 √ Ω( n) [20], the number of rounds in the distributed test is only polylogarithmic in n and polynomial in ǫ−1 . As in [19], we assume that d is a constant, and omit it from our expressions (it is implicit in the O notation for L below). Let us first outline the algorithm of [19], since our distributed test borrows from its framework and our analysis is in part derived from it. The sequential test basically tries to detect odd cycles. It consists of T iterations, in each of which a vertex s is selected uniformly at random and K random walks of length L are performed starting from the source s. If, in any iteration with a chosen source s, there is a vertex v which is reached by an even prefix of a random walk and an odd prefix of a random walk (possibly the same walk), then the algorithm rejects, as this indicates the existence of an odd cycle. Otherwise, the algorithm accepts. To obtain an ǫ-test the parameters √ are chosen to be T = O(ǫ−1 ), K = O(ǫ−4 n log1/2 (nǫ−1 )), and L = O(ǫ−8 log6 n). The main approach of our distributed test is similar, except that a key ingredient is that we can afford to perform much fewer random walks from every vertex, namely O(poly(ǫ−1 log nǫ−1 )). This is because we can run random walks in parallel originating from all vertices at once. However, a crucial challenge that we need to address is that several random walks may collide on an edge, violating its congestion bound. To address this issue, our central observation is that lazy random walks (chosen to have a uniform stationary distribution) provide for a very low probability of having too many of these collisions at once. The main part of the analysis is in showing that with high probability there will never be too many walks concurrently in the same vertex, so we can comply with the congestion bound. We begin by formally defining the lazy random walks that we use. Definition 5.1. A lazy random walk over a graph G with degree bound d is a random walk, that is, a (memory-less) sequence of random variables Y1 , Y2 , . . . taking values from the vertex set V , where the transition probability P r[Yk = v\|Yk−1 = u] is 1/2d if uv is an edge of G, 1 − deg(u)/2d if u = v, and 0 in all other cases. The stationary distribution for the lazy random walk of Definition 5.1 is uniform [34, Section 8]. Next, we describe a procedure to handle one iteration of moving the random walks (Algorithm 3), followed by our distributed test for bipartiteness using lazy random walks from every vertex concurrently (Algorithm 4). It is quite immediate that Algorithm 3 takes O(ξ) communication rounds. Our main result here is that Algorithm 4 is indeed a distributed ǫ-test for bipartiteness. Theorem 5.2. Algorithm 4 is a distributed ǫ-test in the bounded degree graph model for the property of being bipartite, that requires O(poly(ǫ−1 log(nǫ−1 ))) rounds. The number of communication rounds is immediate from the algorithm – it is dominated by the L calls to Algorithm 3, making a total of O(ξL) rounds, which is indeed O(poly(ǫ−1 log(nǫ−1 ))). To prove the rest of Theorem 5.2 we need some notation, and a lemma from [19] that bounds from below the probabilities for detecting odd cycles if G is ǫ-far from being bipartite. Given a source s, if there is a vertex v which is reached by an even prefix of a random walk wi from s and an odd prefix of a random walk wj from s, we say that walks wi and wj detect a violation. Let ps (k, ℓ) be the probability that, out of k random walks of length ℓ starting from s, there are two that detect a violation. Using this notation, ps (K, L) is the probability that the sequential algorithm outlined in the beginning rejects in an iteration in which s is chosen. Since we are only interested in walks of length L, we denote ps (k) = ps (k, L). A good vertex is a vertex for which this probability is bounded as follows. 13 Algorithm 3: Move random walks once with input ξ Variables: Wv walks residing in v (multiset), Hv history of walks through v Input: ξ, the maximum congestion per vertex allowed # each walk is characterized by (i, u) where i is the number of actual moves and u is the origin vertex 1 for each vertex v simultaneously 2 if \|Wv \| ≤ ξ then # give up if exceeded the maximum allowed 3 for every (i, u) in Wv do 4 draw next destination w (according to the lazy walk scheme) 5 if w 6= v then # walk exits v 6 send (i + 1, u) to w 7 remove (i, u) from Wv 8 9 wait until the maximum time for all other vertices to process up to ξ walks add the walks received by v to Wv and Hv # walks entering v Algorithm 4: Distributed bipartiteness test Variables: Wv walks residing in v (multiset), Hv history of walks through v 1 perform η = O(ǫ−9 log(nǫ−1 )) times 2 for each vertex v simultaneously 3 initialize Hv and Wv with two copies of the walk (0, v) 4 5 6 7 8 9 perform L = O(ǫ−8 log6 n) times move walks using Algorithm 3 with input ξ = γ + 2 = 3(2 ln n + ln L) + 2 for each vertex v simultaneously if Hv contains (i, u) and (j, u) for some u, even i and odd j then reject (ending all operations) # odd cycle found accept (for vertices that did not reject) Definition 5.3. A vertex s is called good if ps (K) ≥ 1/10. In [19] it was proved that being far from bipartite implies having many good vertices. Lemma 5.4 ( [19]). If G is ǫ-far from being bipartite then at least an ǫ/16-fraction of the vertices are good. In contrast to [19], we do not perform K random walks from every vertex in each iteration, but rather only 2. Hence, what we need for our analysis is a bound on ps (2). To this end, we use K as a parameter, and express ps (2) in terms of K and ps (K). Lemma 5.5. For every vertex s, ps (2) ≥ 2ps (K)/K(K − 1). Proof. Fix a source vertex s. For every i, j ∈ [K], let qi,j be the probability of walks wi , wj from s detecting a violation. Because different walks are independent, we conclude that for every i 6= j it 14 holds that qi,j = ps (2). Let Ai,j be the event of walks wi , wj detecting a violation. We have ps (K) = P r[∪i,j Ai,j ] ≤ X i,j P r[Ai,j ] = ps (2)K(K − 1)/2, which implies that ps (2) ≥ 2ps (K)/K(K − 1). Using this relationship between ps (2) and K and ps (K), we prove that our algorithm is an ǫ-test. First we prove this for the random walks themselves, ignoring the possibility that Algorithm 3 will skip moving random walks due to its condition in Line 2. Lemma 5.6. If G is ǫ-far from being bipartite, and we perform η iterations of starting 2 random walks of length L from every vertex, then the probability that no violation is detected is bounded by 1/4. Proof. Assume that G is ǫ-far from being bipartite. By Lemma 5.4, at least nǫ/16 vertices are P good, which means that for each of these vertices s, ps (K) ≥ 1/10. This implies that s∈V ps (K) ≥ nǫ/160. Now, let Xi,s be a random variable indicating whether there are two random walks starting P P at s that detect a violation. Let X = ηi=0 s∈V Xi,s . We prove that P r[X < 1] < 1/4. First, we P bound E[ s∈V Xi,s ] for some fixed i: E[X] = E " η XX i=0 s∈V = η X X i=0 s∈V = = # Xi,s = ps (2) ≥ η X X E[Xi,s ] i=0 s∈V η X X i=0 s∈V η X X 2ps (K) K(K − 1) η X nǫ 2 2 ps (K) ≥ K(K − 1) i=0 s∈V K(K − 1) i=0 160 ηnǫ ηnǫ ≥ . 80K(K − 1) 80K 2 For η = 320K 2 /nǫ = O(ǫ−9 log(nǫ−1 )) it holds that E[X] ≥ 4. Using the Chernoff Bound of Fact 2.4 with δ = 3/4 and µ ≥ 4 gives: P r[X < 1] ≤ P r[X < (1 − δ)µ] < ( e−3/4 4 )4 = 3 < 1/4, (1−(3/4)) e (1 − (3/4)) which completes the proof. As explained earlier, the main hurdle on the road to prove Theorem 5.2 is in proving that the allowed congestion will not be exceeded. We prove the following general claim about the probability for k lazy random walks of length ℓ from each vertex to exceed a maximum congestion factor of ξ walks allowed in each vertex at the beginning of each iteration. Here, an iteration is a sequence of rounds in which all walks are advanced by one step (whether or not they actually switch vertices). Lemma 5.7. With probability at least 1 − 1/n, running k lazy random walks of length ℓ originating from every vertex will not exceed the maximum congestion factor of ξ = γ + k = 3(2 ln n + ln ℓ) + k walks allowed in each vertex at the beginning of each iteration, if γ > k. 15 We show below that plugging k = 2, ℓ = L and γ = 3(2 ln n + ln L) in Lemma 5.7, together with Lemma 5.6, gives the correctness of Algorithm 4. To prove Lemma 5.7, we argue that it is unlikely for any vertex to have more than k + γ walks in any iteration. Given that this is indeed the case in every iteration, the lemma follows by a union bound. We denote by Xv,i the random variable whose value is the number of random walks at vertex v at the beginning of the i-th iteration. That is, it is equal to the size of the set Wv in the description of the algorithm. Lemma 5.8. For every vertex v ∈ V and every iteration i it holds that E[Xv,i ] = k. Proof. Let us first define random variables for our walks. Enumerating our kn walks (k from each of the n vertices) arbitrarily, let Y1r , Y2r , . . . denote the sequence corresponding to the r’th walk, that is, Yir is the vertex where the r’th walk is stationed at the beginning of the i’th iteration. In particular, Xv,i = \|{r : Yir = v}\|. Now let us define new random variables Zit in the following manner: First, we choose uniformly σ(t) for all 1 ≤ i ≤ ℓ and 1 ≤ t ≤ kn. at random a permutation σ : [rk] → [rk]. Then we set Zit = Yi t t The main thing to note is that for any fixed t, Z1 , Z2 , . . . is a random walk (as it is equal to one of the random walks Y1r , Y2r , . . .). But also, for every t, Z1t is uniformly distributed over the vertex set of G, because we started with exactly k random walks from every vertex. Additionally, since the uniform distribution is stationary for our lazy walks, this means that the unconditional distribution of each Zit is also uniform. σ(t) = v}\| = \|{t : Now, since σ is a permutation, it holds that Xv,i = \|{r : Yir = v}\| = \|{t : Yi P kn t t Zi = v}\|. The expectation (by linearity of expectation) is thus E[Xv,i ] = t=1 P r[Zi = v] = k. We can now prove Lemma 5.7. Proof of Lemma 5.7. We first claim that for every iteration i ∈ [ℓ] and every vertex v ∈ V , with probability at least 1 − 1/ℓn it holds that Xv,i ≤ k + γ. To show this, first fix some v ∈ V . Let Zj,i be the indicator variable for the event of walk wj residing at vertex v at the beginning of iteration i, P where j ∈ [kn]. Then Xv,i = kn j=1 Zj,i , and the variables Zj,i , where j ∈ [kn], are all independent. We use the Chernoff Bound of Fact 2.4 with δ = γ/k ≥ 1 and µ = k as proven in Lemma 5.8, obtaining: P r[Xv,i > k + γ] = P r[Xv,i > (γ/k + 1)k] < e−δµ/3 = e−γ/3 = e−(2 ln n+ln ℓ) = 1/ℓn2 . Applying the union bound over all vertices v ∈ V and all iterations i ∈ [ℓ], we obtain that with probability at least 1 − 1/n it holds that Xv,i ≤ k + γ for all v and i. Lemma 5.9. If G is bipartite then all vertices output accept in Algorithm 4. If G is ǫ-far from being bipartite, then with probability at least 2/3 there exists a vertex that outputs reject. Proof. If G is bipartite then all vertices output accept in Algorithm 4, because there are no odd cycles and thus no violation detecting walks. If G is ǫ-far from bipartite, we use Lemma 5.6, in conjunction with Lemma 5.7 with parameters k = 2, ℓ = L and γ = 3(2 ln n + ln L) as used by Algorithm 4. By a union bound the probability to accept G will be bounded by 1/4 + 1/n < 1/3 (assuming n > 12), providing for the required bound on the rejection probability. Lemma 5.9, with the communication complexity analysis of Algorithm 4, gives Theorem 5.2. 16 6 Distributed test for cycle-freeness In this section, we give a distributed algorithm to test if a graph G with m edges is cycle-free or if at least ǫm edges have to be removed to make it so. Intuitively, in order to search for cycles, one can run a breadth-first search (BFS) and have a vertex output reject if two different paths reach it. The downside of this exact solution is that its running time depends on the diameter of the graph. To overcome this, a basic approach would be to run a BFS from each vertex of the graph, but for shorter distances. However, running multiple BFSs simultaneously is expensive, due to the congestion on the edges. Instead, we use a simple prioritization rule that drops BFS constructions with lower priority, which makes sure that one BFS remains alive.5 Instead, our technique consists of three parts. First, we make the graph G sparser, by removing each of its edges independently with probability ǫ/2. We denote the sampled graph by G′ and prove that if G is far from being cycle-free then so is G′ , and in particular, G′ contains a cycle. Then, we run a partial BFS over G′ from each vertex, while prioritizing by ids: each vertex keeps only the BFS that originates in the vertex with the largest id and drops the rest of the BFSs. The length of this procedure is according to a threshold T = 20 log n/ǫ. This gives detection of a cycle that is contained in a component of G′ with a low diameter of up to T , if such a cycle exists, since a surviving BFS covers the component. Such a cycle is also a cycle in G. If no such cycle exists in G′ , then G′ has a some component with diameter larger than T . For large components, we take each surviving BFS that reached some vertex v at a certain distance ℓ, and from v we run a new partial BFS in the original graph G. These BFSs are again prioritized, this time according to the distance ℓ. Our main tool here is proving a claim that says that with high probability, if there is a shortest path in G′ of length T /2 between two vertices, then there is a cycle in G between them of length at most T . This allows our BFSs on G to find such a cycle. We start with the following combinatorial lemma that shows the above claim. Lemma 6.1. Given a graph G, let G′ be obtained by deleting each edge in G with probability ǫ/2, independently of other edges. Then, with probability at least 1 − 1/n3 , every vertex v ∈ G′ that has a vertex w ∈ G′ at a distance 10 log n/ǫ, has a closed path passing through it in G, that contains a simple cycle, of length at most 20 log n/ǫ. Proof. First, we show that for every pair u, v of vertices in G that are at a distance of ℓ = 10 log n/ǫ, one of the shortest paths between u and v is removed in the graph G′ with high probability. For a pair of vertices u and v at a distance ℓ in G, the probability that a shortest path is not removed is (1 − ǫ/2)ℓ , which is at most 1/n5 . Therefore, by a union bound over all pairs of vertices, with probability at least 1 − 1/n3 , at least one edge is removed from at least one shortest path between every pair of vertices that are at a distance of 10 log n/ǫ. Conditioned on this, we prove the lemma. Now, suppose that v and w are two vertices in G′ at a distance of 10 log n/ǫ. Let P ′ be this shortest path in G′ . Suppose P is the shortest path between v and w in G. If \|P \| < 10 log n/ǫ, then this path is no longer present in G′ (and thus distinct from P ′ ) and P ∪ P ′ is a closed path in G, passing through v that has a simple cycle of length at most 20 log n/ǫ. If \|P \| = 10 log n/ǫ, then there are at least two shortest paths between v and w in G of length 10 log n/ǫ, the one in G′ and 5 A more involved analysis of multiple prioritized BFS executions was used in [25], allowing all BFS executions to fully finish in a short time without too much delay due to congestion. Since we require a much weaker guarantee, we can avoid the strong full-fledged prioritization algorithm of [25] and settle for a simple rule that keeps one BFS tree alive. Also, the multiple BFS construction of [28] does not fit our demands as it may not reach all desired vertices within the required distance, in case there are many vertices that are closer. 17 one that was removed, which we choose for P . Therefore, P ∪ P ′ is a closed path passing through v of length at most 20 log n/ǫ, and hence contains a simple cycle of length at most 20 log n/ǫ in it. Next, we prove that indeed there is a high probability that G′ contains a cycle if G is far from being cycle-free. Claim 6.2. If G is ǫ-far from being cycle-free, then with probability at least 1 − e−ǫ ǫ/4-far from being cycle-free. 2 m/32 , G′ is Proof. The graph G′ is obtained from G by deleting each edge with probability ǫ/2 independently of other edges. The expected number of edges that are deleted is ǫm/2. Therefore, by the Chernoff Bound from Fact 2.4, the probability that at least 3ǫm/4 edges are deleted is at most exp(−ǫ2 m/32), and the claim follows. We now describe a multiple-BFS algorithm that takes as input a length t and a priority condition P over vertices, and starts performing a BFS from each vertex of the graph. This is done for t steps, in each of which a vertex keeps only the BFS with the highest priority while dropping the rest. Each vertex also maintains a list Lv of BFSs that have passed through it. The list Lv is a list of 3-tuples (idu , ℓ, idp ), where idu is the id of the root of the BFS, ℓ is the depth of v in this BFS tree and idp is the id of the parent of v in the BFS tree. Initially, each vertex v sets Lv to include a BFS starting from itself, and then continues this BFS by sending the tuple (idv , 1, idv ) to all its neighbors, where idv is the identifier of the vertex v. In an intermediate step, each vertex v may receive a BFS tuple from each of its neighbors. The vertex v then adds these BFS tuples to the list Lv and chooses one among Lv according to the priority condition P, proceeding with the respective BFS and discontinuing the rest. Even when a BFS is discontinued, the information that the BFS reached v is stored in the list Lv . Algorithm 5 gives a formal description of the breadth-first search that we use in the testing algorithm for cycle-freeness. Algorithm 5: BFS with a priority condition Input: Length L, Priority condition P Variables: Lv list of BFS tuples passing through v 1 for each vertex v simultaneously 2 Initialize Lv to {(idv , 0, idv )}. 3 Send (idv , 1, idv ) to all neighbors of v. 4 5 6 7 8 9 perform L times times for each vertex v simultaneously if v receives (idu1 , ℓ1 , idp1 ), . . . , (idur , ℓr , idpr ) from its neighbors then Add (idu1 , ℓ1 , idp1 ), . . . , (idur , ℓr , idpr ) to Lv . Select (iduj , ℓj , idpj ) from Lv according to P over idui Send (iduj , ℓj + 1, idv ) to all neighbors of v except pj . We now give more informal details of the test for cycle-freeness. By Lemma 6.1, we know that if there is a vertex v in G′ that has a vertex w at a distance of T /2 = 10 log n/ǫ, then there is a closed 18 path in G starting from v that contains a cycle of length 20 log n/ǫ. In the first part, each vertex gets its name as its vertex id, and performs a BFS on the graph G′ in the hope of finding a cycle. The BFS is performed using Algorithm 5, where the priority condition in the intermediate steps is selecting the BFS with the lowest origin id. If the cycle is present in a component of diameter at most 20 log n/ǫ in G′ , then it is discovered during this BFS. To check if there is a cycle, one needs to find if there are appropriate tuples (idu , ℓ, idp ) and (idu , ℓ′ , idp′ ) in Lv , for some vertex v. If no cycle is discovered in this step, then we change the ids of the vertices in the following way: The id of each vertex v is now a tuple (ℓ, v) where ℓ is the largest depth at which v occurs in a BFS tree among all the breadth-first searches that reached v. We perform a BFS in G using Algorithm 5, where the priority condition is to pick the BFS whose root has the lexicographically highest id. If there is some vertex with ℓ ≥ 10 log n/ǫ, then the highest priority vertex is such a vertex, and by Lemma 6.1, the BFS starting from that vertex will detect a cycle in G. Algorithm 6 gives a formal description of the tester for cycle-freeness. Algorithm 6: Cycle-freeness test Variables: Lv list of BFS tuples passing through v, vertex identifier idv # Construct G′ by deleting edges with probability ǫ/2. 1 for each vertex v simultaneously 2 For each neighbor u < v, mark the edge e = (u, v) ∈ G with probability ǫ/2 for deletion. 3 Send each marked edge e = (u, v) to its corresponding u. 4 Set idv = v. 5 6 7 8 9 10 11 12 13 14 15 16 17 for each vertex v simultaneously Delete all edges incident on v that have been marked for deletion. # Search for cycles in small diameter components. use Algorithm 5 to perform BFS on G′ for 20 log n/ǫ steps, with the priority condition being choosing the BFS with the lowest root id. for each vertex v simultaneously If Lv contains two tuples (idu , ℓ, idp ) and (idu , ℓ′ , idp′ ), output reject. Set idv = (ℓj , v) where ℓj is the highest among all tuples (idui , ℓi , idpi ) in Lv . use Algorithm 5 to perform BFS on G for 10 log n/ǫ steps, with the priority condition being choosing the BFS with the lexicographically highest root id. for each vertex v ∈ G simultaneously If Lv contains two tuples (idu , ℓj , idp ) and (idu , ℓ′ , idp′ ), output reject. for each vertex v ∈ G simultaneously output accept, if v did not output reject yet. We now prove the correctness of the algorithm. Theorem 6.3. Algorithm 6 is a distributed ǫ-test in the general graph model for the property of being cycle-free, that requires O(log n/ǫ) rounds. Proof. Notice that a vertex in Algorithm 6 outputs reject only when it detects a cycle. Therefore, if G is cycle-free, then every vertex outputs accept with probability 1. 19 Suppose that G is ǫ-far from being cycle-free. Notice that, with probability at least 1 − 1/n3 , the assertion of Lemma 6.1 holds. Furthermore, from Claim 6.2, we know that G′ is ǫ/4-far from 2 being cycle-free, with probability 1 − e−ǫ m/32 , and hence contains at least one cycle. This cycle could be in a component of diameter less than 20 log n/ǫ or it could be in a component of diameter at least 20 log n/ǫ in G′ . We analyse the two cases separately. Suppose there is a cycle in a component C of G′ of diameter at most 20 log n/ǫ. Let u be the vertex with the smallest id in C. In Algorithm 6, the BFS starting at u is always propagated at any intermediate vertex due to the priority condition. Furthermore, since the diameter of C is at most 20 log n/ǫ, this BFS reaches all vertices of C. Hence, this BFS detects the cycle and at least one vertex in C outputs reject. On the other hand, if the cycle is present in a component in G of diameter at least 20 log n/ǫ, then after Step 11 of the algorithm, each vertex v gets the length of the longest path from the origin, among all the BFSs that reached v, as the first component of its id. The vertex v that gets the lexicographically highest id in the component has a vertex w that is at least 10 log n/ǫ away in G′ , since the radius of the component is at least half the diameter. Therefore, by Lemma 6.1, it is part of a cycle of length at most 20 log n/ǫ in G. Hence, the vertex with the highest priority in the BFS on G is a vertex u that has a vertex at a distance of at least 10 log n/ǫ in G′ , and there is a walk through u that contain a simple cycle of length at most 20 log n/ǫ. At least one vertex on this simple cycle will output reject when Algorithm 6 is run on G. The number of rounds is O(log n/ǫ) since Algorithm 6 performs two breadth-first searches in the graph with this number of rounds. 7 Lower bounds for testing bipartiteness and cycle-freeness In this section, we prove that any distributed algorithm for ǫ-testing bipartiteness or cycle-freeness in bounded-degree graphs requires Ω(log n) rounds of communication6 . We construct boundeddegree graphs that are ǫ-far from being bipartite, such that all cycles are of length Ω(log n). We argue that any distributed algorithm that runs in O(log n) rounds does not detect a witness for nonbipartiteness. We also show that the same construction proves that every distributed algorithm for ǫ-testing cycle-freeness requires Ω(log n) rounds of communication. Formally, we prove the following theorem. Theorem 7.1. Any distributed 1/100-test for the property of being bipartite requires Ω(log n) rounds of communication. To prove Theorem 7.1, we show the existence of a graph G′ that is far from being bipartite, but all of its cycles are at least of logarithmic length. Since in T rounds of a distributed algorithm, the output of every vertex cannot depend on vertices that are at distance greater than T from it, no vertex can detect a cycle in G′ in less than O(log n) rounds, which proves Theorem 7.1. To prove the existence of G′ we use the probabilistic method with alterations, and prove the following. Lemma 7.2. Let G be a random graph on n vertices where each edge is present with probability 1000/n. Let G′ be obtained by removing all edges incident with vertices of degree greater than 2000, and one edge from each cycle of length at most log n/ log 1000. Then with probability at least 1/2 − e−100 − e−n , G′ is 1/100-far from being bipartite. 6 Our lower bound applies even to the less restricted LOCAL model of communication, which does not limit the size of the messages. 20 Since a graph that is ǫ-far from being bipartite is also ǫ-far from being cycle-free, we immediately obtain the same lower bound for testing cycle-freeness, as follows. Theorem 7.3. Any distributed 1/100-test for the property of being cycle-free requires Ω(log n) rounds of communication. The rest of this section is devoted to proving Lemma 7.2. We need to show three properties of (a) that it is far from being bipartite, (b) that it does not have small cycles, and (c) that its maximum degree is bounded. We begin with the following definition, which is similar in spirit to being far from satisfying a property and which will assist us in our proof. G′ : Definition 7.4. A graph G is k-removed from being bipartite if at least k edges have to be removed from G to make it bipartite. Note that a graph G with maximum degree d, is ε-far from being bipartite if it is εdn-removed from being bipartite. Let G be a random graph on n vertices where for each pair of vertices, an edge is present with probability 1000/n. The expected number of edges in the graph is 500(n − 1). Since the edges are sampled independently with probability 1000/n, by the Chernoff Bound from Fact 2.4, with probability at least 1 − e−10n the graph has at least 400n edges. We now show that G is far from being bipartite, with high probability. Lemma 7.5 (far from being bipartite). With probability at least 1 − e−199n , G is 20n-far from being bipartite. Proof. Fix a bipartition (L, R) of the vertex set of G such that \|L\| ≥ n/2. For each pair of vertices u, v ∈ L, let Xu,v be a random variable which is 1 if the edge (u, v) is present in G and 0 P otherwise. Its expected value is E[Xu,v ] = 1000/n. The random variable X = u,v∈L Xu,v counts the number of edges within L. By the linearity of expectation, E[X] ≥ n/2 2 1000/n ≥ 30n. Since the random variables Xu,v are independent, by the Chernoff Bound from Fact 2.4, we have that Pr[X < 20n] ≤ exp(−200n). Therefore, with probability at least 1 − exp(−200n), there are at least 20n edges within L. The total number of such bipartitions of G is at most 2n−1 . Taking a union bound over all such bipartitions, the probability that at least one of the bipartitions contains less than 20n edges within its L side is at most exp(−199n), and the lemma follows. The expected degree of a vertex v in G is 1000(1− 1/n). Therefore, by the Chernoff Bound from Fact 2.4, the probability that the degree of v is greater than 2000 is at most exp(−300(1 − 1/n)). We now show that, with sufficiently high probability, the number of edges that are incident with high degree vertices is small. We can remove all such edges to obtain a bounded-degree graph that is still far from being bipartite. Lemma 7.6 (mostly bounded degrees). With probability at least 1 − e−100 , there are at most n edges that are incident with vertices of degree greater than 2000 in G. Proof. For a pair u, v of vertices, the probability that there is an edge between them and that one of u or v is of degree greater than 2000 is Pr[(u, v) ∈ E] · Pr[u or v has degree ≥ 2000\|(u, v) ∈ E]. This is at most (1000/n) · 2 · exp(−300(1 − 1/n)). Therefore, the expected number of edges that are incident with a vertex of degree greater than 2000 is at most 1000n · exp(−300(1 − 1/n)). By Markov’s inequality, the probability that there are at least n edges that are incident with vertices of degree greater than 2000 is at most 1000 · exp(−300(1 − 1/n)). This completes the proof of the lemma. 21 We now bound the number of cycles of length at most O(log n) in the graph G. Lemma 7.7 (few small cycles). With probability at least 1/2, there are at most 2n cycles of length at most log n/ log 1000 in G. Proof. For any k fixed vertices, the probability that there is a cycle among the k vertices is at most k!(1000/n)k . Therefore the expected number of cycles in G of length at most k is at most 1000k . For k = log n/ log 1000, this means that the expected number of cycles in G of length at most log n/ log 1000 is n. Therefore, with probability at least 1/2 there are at most 2n cycles of length at most log n/ log 1000 in G. We are now ready to prove Lemma 7.2, completing our lower bounds. Intuitively, since G does not contain many high degree vertices and many small cycles, removing them to obtain G′ only changes the distance from being bipartite by a small term. Proof. With probability 1 − e−n , there are at least 400n edges in G and by Lemma 7.5 G is 20nremoved from being bipartite. By Lemma 7.6, with probability at least 1 − e−100 , there are at most n edges incident with vertices of degree greater than 2000 and by Lemma 7.7 with probability at least 1/2 there are at most 2n cycles of length at most log n/ log 1000. Hence, with probability at least 1/2 − e−100 − e−n , G′ is a graph with degree at most 2000 that is 17n-removed from being bipartite. Therefore, G′ is 1/100-far from being bipartite. 8 Discussion This paper initiates a thorough study of distributed property testing. It provides an emulation technique for the dense graph model and constructs fast distributed algorithms for testing trianglefreeness, cycle-freeness and bipartiteness. We also present lower bounds for both bipartiteness and triangle freeness. This work raises many important open questions, the immediate of which is to devise fast distributed testing algorithms for additional problems. One example is testing freeness of other small subgraphs. More ambitious goals are to handle dynamic graphs, and to find more general connections between testability in the sequential model and the distributed model. Finally, there is fertile ground for obtaining additional lower bounds in this setting, in order to fully understand the complexity of distributed property testing. References [1] Noga Alon, Chen Avin, Michal Koucký, Gady Kozma, Zvi Lotker, and Mark R. Tuttle. Many random walks are faster than one. Combinatorics, Probability & Computing, 20(4):481–502, 2011. [2] Noga Alon, Tali Kaufman, Michael Krivelevich, and Dana Ron. Testing triangle-freeness in general graphs. SIAM J. Discrete Math., 22(2):786–819, 2008. [3] Noga Alon and Michael Krivelevich. Testing k-colorability. SIAM J. Discrete Math., 15(2):211– 227, 2002. 22 [4] Noga Alon and Asaf Shapira. A characterization of the (natural) graph properties testable with one-sided error. SIAM J. Comput., 37(6):1703–1727, 2008. [5] Heger Arfaoui, Pierre Fraigniaud, David Ilcinkas, and Fabien Mathieu. Distributedly testing cycle-freeness. In Graph-Theoretic Concepts in Computer Science - 40th International Workshop, WG 2014, Nouan-le-Fuzelier, France, June 25-27, 2014. Revised Selected Papers, pages 15–28, 2014. [6] Mor Baruch, Pierre Fraigniaud, and Boaz Patt-Shamir. Randomized proof-labeling schemes. In Proceedings of the 2015 ACM Symposium on Principles of Distributed Computing, PODC 2015, Donostia-San Sebastián, Spain, July 21 - 23, 2015, pages 315–324, 2015. [7] Manuel Blum, Michael Luby, and Ronitt Rubinfeld. Self-testing/correcting with applications to numerical problems. J. Comput. Syst. Sci., 47(3):549–595, 1993. [8] Zvika Brakerski and Boaz Patt-Shamir. Distributed discovery of large near-cliques. Distributed Computing, 24(2):79–89, 2011. [9] Keren Censor-Hillel, Petteri Kaski, Janne H. Korhonen, Christoph Lenzen, Ami Paz, and Jukka Suomela. Algebraic methods in the congested clique. In Proceedings of the 2015 ACM Symposium on Principles of Distributed Computing, PODC 2015, Donostia-San Sebastián, Spain, July 21 - 23, 2015, pages 143–152, 2015. [10] Maria Chudnovsky, Neil Robertson, Paul Seymour, and Robin Thomas. The strong perfect graph theorem. ANNALS OF MATHEMATICS, 164:51–229, 2006. [11] Danny Dolev, Christoph Lenzen, and Shir Peled. "tri, tri again": Finding triangles and small subgraphs in a distributed setting - (extended abstract). In Distributed Computing - 26th International Symposium, DISC 2012, Salvador, Brazil, October 16-18, 2012. Proceedings, pages 195–209, 2012. [12] Andrew Drucker, Fabian Kuhn, and Rotem Oshman. The communication complexity of distributed task allocation. In ACM Symposium on Principles of Distributed Computing, PODC ’12, Funchal, Madeira, Portugal, July 16-18, 2012, pages 67–76, 2012. [13] Paul Erdös. Graph theory and probability. canad. J. Math, 11:34G38, 1959. [14] Eldar Fischer. The art of uninformed decisions: A primer to property testing. Current Trends in Theoretical Computer Science: The Challenge of the New Century, I:229–264, 2004. [15] Klaus-Tycho Foerster, Thomas Luedi, Jochen Seidel, and Roger Wattenhofer. Local checkability, no strings attached. In Proceedings of the 17th International Conference on Distributed Computing and Networking, ICDCN 2016, Singapore, January 4-7, 2016., 2016. [16] Jacob Fox. A new proof of the graph removal lemma. CoRR, abs/1006.1300, 2010. [17] Mohsen Ghaffari, Fabian Kuhn, and Hsin-Hao Su. Manuscript. 2016. [18] Oded Goldreich, Shafi Goldwasser, and Dana Ron. Property testing and its connection to learning and approximation. J. ACM, 45(4):653–750, 1998. 23 [19] Oded Goldreich and Dana Ron. A sublinear bipartiteness tester for bounded degree graphs. Combinatorica, 19(3):335–373, 1999. [20] Oded Goldreich and Dana Ron. Property testing in bounded degree graphs. Algorithmica, 32(2):302–343, 2002. [21] Oded Goldreich and Dana Ron. Algorithmic aspects of property testing in the dense graphs model. In Property Testing - Current Research and Surveys [outgrow of a workshop at the Institute for Computer Science (ITCS) at Tsinghua University, January 2010], pages 295– 305, 2010. [22] Oded Goldreich and Luca Trevisan. Three theorems regarding testing graph properties. Random Struct. Algorithms, 23(1):23–57, 2003. [23] Mika Göös, Juho Hirvonen, Reut Levi, Moti Medina, and Jukka Suomela. Non-local probes do not help with graph problems. CoRR, abs/1512.05411, 2015. [24] Juho Hirvonen, Joel Rybicki, Stefan Schmid, and Jukka Suomela. Large cuts with local algorithms on triangle-free graphs. CoRR, abs/1402.2543, 2014. [25] Stephan Holzer and Roger Wattenhofer. Optimal distributed all pairs shortest paths and applications. In Proceedings of the 2012 ACM Symposium on Principles of Distributed Computing, PODC ’12, pages 355–364, New York, NY, USA, 2012. ACM. [26] Jarkko Kari, Martín Matamala, Ivan Rapaport, and Ville Salo. Solving the induced subgraph problem in the randomized multiparty simultaneous messages model. In Structural Information and Communication Complexity - 22nd International Colloquium, SIROCCO 2015, Montserrat, Spain, July 14-16, 2015, Post-Proceedings, pages 370–384, 2015. [27] Amos Korman, Shay Kutten, and David Peleg. Proof labeling schemes. Distributed Computing, 22(4):215–233, 2010. [28] Christoph Lenzen and David Peleg. Efficient distributed source detection with limited bandwidth. In ACM Symposium on Principles of Distributed Computing, PODC ’13, Montreal, QC, Canada, July 22-24, 2013, pages 375–382, 2013. [29] Michael Mitzenmacher and Eli Upfal. Probability and computing - randomized algorithms and probabilistic analysis. Cambridge University Press, 2005. [30] Michal Parnas and Dana Ron. Approximating the minimum vertex cover in sublinear time and a connection to distributed algorithms. Theor. Comput. Sci., 381(1-3):183–196, 2007. [31] David Peleg. Distributed Computing: A Locality-Sensitive Approach. Society for Industrial and Applied Mathematics, 2000. [32] Seth Pettie and Hsin-Hao Su. Distributed coloring algorithms for triangle-free graphs. Inf. Comput., 243:263–280, 2015. [33] Dana Ron. Property testing: A learning theory perspective. Foundations and Trends in Machine Learning, 1(3):307–402, 2008. 24 [34] Dana Ron. Algorithmic and analysis techniques in property testing. Foundations and Trends in Theoretical Computer Science, 5(2):73–205, 2009. [35] Ronitt Rubinfeld and Madhu Sudan. Robust characterizations of polynomials with applications to program testing. SIAM J. Comput., 25(2):252–271, 1996. [36] Atish Das Sarma, Stephan Holzer, Liah Kor, Amos Korman, Danupon Nanongkai, Gopal Pandurangan, David Peleg, and Roger Wattenhofer. Distributed verification and hardness of distributed approximation. SIAM J. Comput., 41(5):1235–1265, 2012. [37] Atish Das Sarma, Danupon Nanongkai, Gopal Pandurangan, and Prasad Tetali. Distributed random walks. J. ACM, 60(1):2, 2013. 25	8cs.DS
arXiv:1802.07995v2 [math.PR] 23 Feb 2018 Multidimensional multiscale scanning in Exponential Families: Limit theory and statistical consequences Claudia König claudia.koenig@stud.uni-goettingen.de Institute for Mathematical Stochastics, University of Göttingen Axel Munk munk@math.uni-goettingen.de Institute for Mathematical Stochastics, University of Göttingen and Felix Bernstein Institute for Mathematical Statistics in the Bioscience, University of Göttingen and Max Planck Institute for Biophysical Chemistry, Göttingen, Germany Frank Werner1 frank.werner@mpibpc.mpg.de Institute for Mathematical Stochastics, University of Göttingen and Max Planck Institute for Biophysical Chemistry, Göttingen, Germany In this paper we consider the problem of finding anomalies in a d-dimensional field of independent random variables tYi uiPt1,...,nud , each distributed according to a onedimensional natural exponential family F “ tFθ uθPΘ . Given some baseline parameter θ0 P Θ, the field is scanned using local likelihood ratio tests to detect from a (large) d given system of regions R those regions R Ă t1, ..., nu with θi ‰ θ0 for some i P R. We provide a unified methodology which controls the overall family wise error (FWER) to make a wrong detection at a given error rate. Fundamental to our method is a Gaussian approximation of the asymptotic distribution of the underlying multiscale scanning test statistic with explicit rate of convergence. From this, we obtain a weak limit theorem which can be seen as a generalized weak invariance principle to non identically distributed data and is of independent interest. Furthermore, we give an asymptotic expansion of the procedures power, which yields minimax optimality in case of Gaussian observations. Keywords: exponential families, multiscale testing, invariance principle, scan statistic, weak limit, family wise error rate AMS classification numbers: Primary 60F17, 62H10, Secondary 60G50, 62F03. 1 Introduction Suppose we observe an independent, d-dimensional field Y of random variables Yi „ Fθi , 1 Corresponding i P Ind :“ t1, ..., nud , author 1 (1) where each observation is drawn from the same given one-dimensional natural exponential family model F “ tFθ uθPΘ , but with potentially different parameters θi . Prominent examples include Yi with varying normal means µi or a Poisson field with varying intensities λi . Given some baseline parameter θ0 P Θ (e.g. all µi “ 0 for a Gaussian field), we consider the problem of finding anomalies (hot spots) in the field Y , i.e. we aim to identify those regions R Ă Ind where θi ‰ θ0 d for some i P R. Here R runs through a given family of candidate regions R P Rn Ă 2In with the power set 2X of some set X. For simplicity, we will suppress the subindex n whenever it is clear from the context, i.e. write R “ Rn in what follows. Such problems occur in numerous areas of application ranging from astronomy and biophysics to genetics engineering, specific examples include detection in radiographic images (Kazantsev et al., 2002), genome screening (Jiang et al., 2016) and object detection in astrophysical image analysis (Friedenberg and Genovese, 2013), to mention a few. Our setting includes the important special cases of Gaussian (Arias-Castro et al., 2005; Sharpnack and Arias-Castro, 2016), Bernoulli (Walther, 2010), and Poisson random fields (Zhang et al., 2016). Extensions to models without exponential family structure as well as replacing the baseline parameter θ0 by a varying field of known baseline intensities can be treated as well (cf. Remark 2.1 below), but to keep the presentation simple, we restrict ourselves to the afore mentioned setting. Inline with the above mentioned references (see also Section 1.2), the problem of finding hot spots is regarded as a multiple testing problem, i.e. many ’local’ tests on the regions R are performed simultaneously, while keeping the overall error of wrong detections controllable. For a fixed region R P R the likelihood ratio test (LRT) for the testing problem @ i P R : θi “ θ0 (HR,n ) D i P R s.t. θi ‰ θ0 , (KR,n ) vs. is a powerful test in general, and often known to have certain optimality properties (depending on the structure of R, see e.g. Lehmann and Romano (2005)). Therefore, the LRT will always be considered throughout this paper as the ’local’ test. We stress, however, that our methodology could also be used for other systems of local tests, provided they obey a sufficiently well behaving asymptotic expansion. The LRT is based on the test statistic d ś ˙ ˆ supθą0 iPR fθ pYi q ś , (2) TR pY, θ0 q :“ 2 log iPR fθ0 pYi q where fθ denotes the density of Fθ , and HR,n is rejected when TR pY, θ0 q is too large. As it is not known a priori which regions R might contain anomalies, i.e. for which R P R the alternative (KR,n ) might hold true, it is of great importance to control the family wise error arising from the multiple test decisions of the local tests based on TR pY, θ0 q, R P R. Obviously, without any further restriction on the complexity of R this error cannot be controlled. To this end, we will assume that the regions R can be represented as a sequence of discretized regions in ˇ ( R “ Rn :“ R Ă Ind ˇ R “ Ind X nR˚ for some R˚ P R˚ (3) d for some system of subsets (e.g. all hypercubes) of the unit cube R˚ Ă 2r0,1s , to be specified later. This gives rise to the sequence of multiple testing problems HR,n vs. KR,n simultaneously over Rn . (4) The aim of this paper is to provide methodology to control (asymptotically) the family wise error rate (FWER) α P p0, 1q when (4) is considered as a multiple testing problem, i.e. to provide a multiple test φ for (4) (see e.g. Dickhaus, 2014) such that sup PHR,n r“any false rejection by φ”s ď α ` op1q RPRn 2 as n Ñ 8. (5) In words, this ensures that the probability of making any wrong detection is controlled by a given error level α, as n Ñ 8. This task has been the focus of several papers during the last decades, for a detailed discussion see Section 1.2. We contribute to this field by providing a general theory for a unifying method in the model (1), which includes not only Gaussian (Arias-Castro et al., 2005; Sharpnack and Arias-Castro, 2016; Kou, 2017; Cheng and Schwartzman, 2017), but also Bernoulli (Walther, 2010) or Poissonian observations (Kulldorff et al., 2005; Rivera and Walther, 2013; Tu, 2013; Zhang et al., 2016). In view of (Arias-Castro et al., 2011), where also observations from exponential families as in (1) are discussed, but the local tests are always as in the Gaussian case, we emphasize that our local test are of type (2), hence exploiting the likelihood in the exponential family, which will result in improved power (see Frick et al. (2014) for d “ 1). Our main technical contribution is to prove a weak limit theorem for the asymptotic distribution of our test statistic for general exponential family models as in (1) and arbitrary dimension d. This can be viewed as a ”multiscale” weak invariance principle for independent but not necessarily identically distributed r.v.’s. Further, we will provide an asymptotic expansion of the test’s power which leads to minimax optimal detection of the test in specific models. 1.1 Multiscale testing Our test will be of scanning-type, controlling the FWER by the maximum over the local LRT statistics in (2), i.e. Tn ” Tn pY, θ0 , Rn , vq :“ max rTR pY, θ0 q ´ penv p\|R\|qs . RPRn Here \|R\| denotes the number of points in R. The values b penv prq :“ 2v plog pnd {rq ` 1q (6) (7) where log denotes the natural logarithm, act as a scale penalization, which is necessary to guarantee optimal detection power on all scales simultaneously as it prevents smaller regions from dominating the overall test statistic, as noticed by Dümbgen and Spokoiny (2001) and others (see e.g. Dümbgen and Walther, 2008; Walther, 2010; Frick et al., 2014). The constant v in (6) can be any upper bound of VR˚ which is given as the complexity of R˚ , measured in terms of the packing number (see Remark 2.2 below). Whenever R˚ has finite VC-dimension ν pR˚ q, we can choose VR˚ “ ν pR˚ q, however, the test will have better detection properties if v is as small as possible with this property (see Section 2.3). Hence, from this point of view it is advantageous to know exactly the complexity VR˚ of R˚ , a topic which has received less attention than computing VC-dimensions. Therefore, we compute the packing numbers for three important examples of R˚ , namely hyperrectangles, hypercubes and halfspaces explicitly in the appendix. To construct a test which controls the FWER (5), we have to find a sequence of universal global thresholds q1´α,n such that P0 rTn ą q1´α,n s ď α ` op1q, (8) where P0 :“ PHI d ,n corresponds to the case that no anomaly is present. Such a threshold suffices, n as it can be readily seen that “ ‰ sup PHR,n r“any false rejection in R”s ď P0 “any false rejection in Ind ” . RPRn Given q1´α,n , the multiple test then will reject whenever Tn ě q1´α,n , and each local test rejects if TR pY, θ0 q ě q1´α,n ` penv p\|R\|q. Due to (6) and (8), any of these rejections is correct with probability ě 1 ´ α, asymptotically. To obtain the thresholds q1´α,n we provide a Gaussian approximation of the scan statistic (6) under P0 given by ˇ ˇ ﬀ « ˇÿ ˇ ˇ ˇ ´1{2 Mn ” Mn pRn , vq :“ max \|R\| (9) ˇ Xi ˇ ´ penv p\|R\|q RPRn ˇiPR ˇ 3 with i.i.d. standard normal r.v.’s Xi , i P Ind (Thm. 2.4). We also give a rate of convergence of this approximation (Thm. 2.5), which is determined by the smallest scale in Rn . Based on these results, we obtain the P0 -limiting distribution of Tn as that of « ﬀ ˚ ` ˘ \|W pR q\| a M ” M pR˚ , vq :“ sup ´ penv nd \|R˚ \| ă 8 a.s., (10) \|R˚ \| R˚ PR˚ d where W is white noise on r0, 1s and (with a slight abuse of notation) \|R˚ \| denotes the Lebesgue measure of R˚ P R˚ . This holds true as soon as R˚ and Rn have a finite complexity, R˚ consists of sets with a sufficiently regular boundary (see Assumption 1 below), and the smallest scales \|Rn \| of the system Rn are restricted suitably, see (12) and the discussion there. In case of R˚ being the subset of all hypercubes, we will also give an asymptotic expansion of the above test’s power, which allows to determine the necessary average strength of an anomaly such that it will be detected with asymptotic probability 1. This is only possible due to the penalization in (6), as otherwise the asymptotic distribution is not a.s. finite. If the anomaly is sufficiently small, we show that the anomalies which can be detected with asymptotic power one by the described multiscale testing procedure are the same as those of the oracle single scale test, which knows the size (scale) of the anomaly in advance. This generalizes findings of Sharpnack and Arias-Castro (2016) to situations where not only the mean of the signal is allowed to change, but its whole distribution. Furthermore, if the observations are Gaussian, our test with properly chosen v achieves the asymptotic optimal detection boundary, i.e. no test can have larger power in a minimax sense, asymptotically. Note finally, that weak convergence of Tn to M as in (10) can be viewed as a generalized weak invariance principle as the r.h.s. does not depend on any unknown quantity, and hence can be e.g. simulated generically in advance for any given system R as soon as a bound for the complexity of R˚ can be determined. 1.2 Literature review and connections to existing work Scan statistics and scanning-type procedures based on the maximum over an ensemble of local tests have received much attention in the literature over the past decades. To determine the quantile, a common option is to approximate the tails of the asymptotic distribution suitably, as done e.g. by Siegmund and Venkatraman (1995); Siegmund and Yakir (2000); Naus and Wallenstein (2004); Pozdnyakov et al. (2005); Fang and Siegmund (2016) for d “ 1, by Haiman and Preda (2006) for d “ 2, and by Jiang (2002) in arbitrary dimensions. If the random field is sufficiently smooth (in contrast to the setting here) the Gaussian kinematic formula can be employed, see e.g. Taylor and Worsley (2007), Adler (2000). We also mention Alm (1998), who considers the situation of a fixed rectangular scanning set in two and three dimensions. In all these papers, no penalization has been used, which automatically leads to a preference for small scales of order logpnq (see e.g. Kabluchko and Munk, 2009) and to an extreme value limit, in contrast to (10). Arias-Castro et al. (2017) study the case of an unknown null distribution and propose a permutation based approximation, which is shown to perform well in the natural exponential family setting (1), however, only for d “ 1. Conceptually most related to our work are weak limit theorems for scale penalized scan statistics, which have e.g. been obtained by Frick et al. (2014) and Sharpnack and Arias-Castro (2016). However, these results are either limited to special situations such as Gaussian observations, or to d “ 1. If a limit exists, the quantiles of the finite sample statistic can be bounded the quantiles by limiting ones as e.g. done by Dümbgen and Spokoiny (2001); Rivera and Walther (2013). Our results can be interpreted in both ways as we provide a Gaussian approximation of the scan statistic in (6) by (9) and that we obtain (10) as a weak limit. Weak limit theorems for Tn as in (6) are immediately connected to those for partial sum processes. Classical KMT-like approximations (see e.g. Komlós et al., 1976; Rio, 1993; Massart, 1989) provide in fact a strong coupling of the whole process pTR pY, θ0 qqRPRn to a Gaussian version. Results of this form have been employed for d “ 1 previously in Schmidt-Hieber et al. (2013); 4 Frick et al. (2014). Proceeding like this for general d will restrict the system Rn to scales rn s.t. \|R\| ě rn where nd´1 logpnq “ o prn q (11) as n Ñ 8, which is unfeasibly large for d ě 2. Therefore, we take a different route and employ a coupling of the maxima in (6) and (9), which relies on recent results by Chernozhukov et al. (2014), see also Proksch et al. (2017). However, in contrast to the present paper, they do not consider the local LR statistic and require that the largest scale has to go to zero. This leads to an extreme value type limit in contrast to (10) which incorporates all (larger) scales. To make use of Chernozhukov et al.’s (2014) coupling results in our general setting, we provide a symmetrizationlike upper bound for the expectation of the maximum of a partial sum process by a corresponding Gaussian version, cf. Proposition 3.2. Doing so we are able to approximate the distribution of Tn in (6) by (9) as soon as we restrict ourselves to R P Rn with \|R\| ě rn where the smallest scales only need to satisfy log12 pnq “ o prn q as n Ñ 8, (12) which compared to (11) allows for considerably smaller scales whenever d ě 2. Note that the lower scale restriction (12) does not to depend on d. However, as we consider sets in Ind here, d the corresponding lower an bound for sets in R˚ Ă 2r0,1s is n´d log12 pnq “ o pan q, which in fact depends on d as now the volume of the largest possible set has been standardized to one (see (3) and Theorem 2.9 below) and coincides with the sampling rate n´d up to a poly-log-factor. In contrast, (11) gives n´1 logpnq “ o pan q, independent of d, which only for d “ 1 achieves the ´` ˘1{10 ¯ 12 ´d sampling rate n . Under (12) we also obtain OP log pnq{rn as rate of convergence of this approximation (see (18) below). Also the asymptotic power of scanning-type procedures has been discussed in the literature. We mention Walther (2010) who studies the detection of spacial clusters in a two dimensional Bernoulli field, or Kabluchko (2011) who gives exact asymptotic expansions in a Gaussian setting. To obtain optimal power on all scales, proper penalization is necessary, as e.g. stressed by Arias-Castro et al. (2005) and Sharpnack and Arias-Castro (2016), who provide optimality results in one- and twodimensional Gaussian fields, respectively. Butucea and Ingster (2013) for d “ 2 and Kou (2017) for general d provide optimality of scanning procedures for Gaussian fields. We are able to generalize these results in case of R˚ being the set of all hypercubes to our exponential family model (1), despite the fact that under the alternative the whole distribution in (1) might change, whereas for Gaussian fields typically only the mean changes. Doing so we obtain sharp detection thresholds, which are known to be minimax in the Gaussian situation, if the parameter v in the penalization (41) is chosen to be equal to the packing number of the system of hypercubes. In contrast, if v is chosen to be the VC-dimension, the detection power turns out to be suboptimal. This emphasizes the importance of knowledge of the packing number explicitly, and an illustrative example for this will also be given in Example 2.8. As a potential alternative, weaker error measures which do not control the FWER, such as the false discovery rate (FDR) could be controlled (see e.g. Benjamini and Hochberg, 1995; Benjamini and Yekutieli, 2001; Li et al., 2016), but this is a different task and beyond the scope of our paper. 2 Theory In this section we will summarize our theoretical findings. In Section 2.1 we give an overview and details on our precise setting and present our assumptions on the set of candidate regions R˚ . Section 2.2 provides the validity of the Gaussian approximation in (9) and determines the P0 -limiting distribution of Tn . In Section 2.3 we derive an asymptotic expansion of the detection power. 5 2.1 Setting and Assumptions In the following we assume that F “ tFθ uθPΘ in (1) is a one-dimensional exponential family, which is regular and minimal, i.e. the Lebesgue densities of Fθ are of the form fθ pxq “ hpxq exp pxθ, xy ´ ψpθqq, the natural parameter space * " ż exppθxq dx ă 8 N “ θ P Rd : Rd is open and the cumulant transform ψ is strictly convex on N . Then, the moment generating function exists and the random variables Yi have sub-exponential tails, see Casella and Berger (2002) and Brown (1986) for details. Let φ pxq :“ supθPΘ rθ ¨ x ´ ψ pθqs be the Legendre-Fenchel conjugate of ψ and J px, θq :“ φ pxq ´ rθ ¨ x ´ ψ pθqs , then the LRT statistic TR pY, θ0 q in (2) can be written as g ˜ ¸ f ÿ ÿ f e pθ ¨ Yi ´ ψpθqq ´ pθ0 ¨ Yi ´ ψpθ0 qq TR pY, θ0 q “ 2 sup θ iPR iPR b ˘ ` “ 2 \|R\| J ȲR , θ0 (13) ˘ ` ř with ȲR “ \|R\|´1 iPR Yi . Note that by definition it holds J ȲR , θ0 ě 0. Remark 2.1. If observations are not drawn from an exponential family as in (1), or if θ0 P Θ is replaced by a field pθi qiPInd of known baseline intensities, then the representation of the LRT statistic TR as in (13) is not valid anymore. Our proofs rely on a third-order Taylor expansions of TR and on the sub-exponential tails of the random variables Yi (see Theorem 2.5 below), but not explicitly on the exponential family structure. Therefore, if in more general models corresponding assumptions are posed (see also Arias-Castro et al., 2017, Sec. 2.2), our results do immediately generalize to this situation. To control the supremum in (10), we have to restrict the system of regions R˚ suitably. To this end, we have to introduce some notation. For a set R˚ P R˚ and x P r0, 1sd we define d px, BR˚ q :“ inf yPBR˚ }x ´ y}2 where BR˚ denotes the topological boundary of R˚ , i.e. BR˚ “ ˝ R˚ z pR˚ q . Furthermore we define the ǫ´annulus R˚ pǫq around the boundary of R˚ for some ǫ ą 0 as ! ) d ˇ R˚ pǫq :“ x P r0, 1s ˇ d px, BR˚ q ă ǫ . Assumption 1 (Complexity and regularity of R˚ ). (a) The VC-Dimension of the set R˚ is bounded by ν pR˚ q ă 8. (b) There exists some constant C ą 0 such that \|R˚ pǫq\| ď Cǫ for all ǫ ą 0 and all R˚ P R˚ with the Lebesgue measure \|¨\|. Let us briefly comment on the above assumption. Remark 2.2. • Assumption 1(a) is a standard assumption to control the complexity of the set indexed process, see e.g. Massart (1989); van der Vaart and Wellner (1996); Dümbgen and Spokoiny (2001). • Assumption 1(a) immediately implies an upper bound for the cardinality # pRn q of Rn in (3), this is there exist constants c1 , c2 ą 0 such that # pRn q ď c1 nc2 . (14) This will allow us to apply recent results by Chernozhukov et al. (2014) to couple the process in (6) with a Gaussian version as in (9). 6 • In the following, we will also need a bound for the complexity of R˚ in terms of the packing number. The packing number Kpǫ, ρ, Wq of a subset W of R˚ w.r.t. a metric ρ is given by the maximum number m of points W1 , . . . , Wm P W s.t. ρpWi , Wj q ą ǫ for all i ‰ j, i.e. by the largest number of ǫ-balls w.r.t. ρ which can be packed inside W, see e.g. van der Vaart and Wellner (1996, Def. 2.2.3). In the following we will consider the symmetric difference R1˚ △ R2˚ :“ pR1˚ Y R2˚ q z pR1˚ X R2˚ q , and the corresponding metric ρ˚ pR1˚ , R2˚ q :“ b \|R1˚ △ R2˚ \|, for R1˚ , R2˚ P R˚ R1˚ , R2˚ P R˚ . Suppose that there exists a positive number VR˚ and constants k1 , k2 ą 0 such that ´ ¯ K pδuq1{2 , ρ˚ , tR P R˚ : \|R\| ď δu ď k1 u´k2 δ ´VR˚ (15) (16) for all u, δ P p0, 1s. If Assumption 1(a) is satisfied, then (16) holds true with VR˚ “ ν pR˚ q, which basically follows from the relationship between capacity and covering numbers and van der Vaart and Wellner (1996, Thm. 2.6.4). However, (16) might also be satisfied for considerably smaller numbers VR˚ (see the examples below). • We stress that that the assumption on the boundary smoothness (b) is satisfied whenever R˚ consists of regular Borel sets R˚ only, i.e. R˚ is a Borel set and \|BR˚ \| “ 0 for all R˚ P R˚ . Example 2.3. d 1. Consider!the set S ˚ of all hyperrectangles in r0,)1s , i.e. each S ˚ P S ˚ is of the form S ˚ “ d ˇ rs, ts :“ x P r0, 1s ˇ si ď xi ď ti for 1 ď i ď d . According to van der Vaart and Wellner (1996, Ex. 2.6.1) we have ν pS ˚ q “ 2d, and more refined computations in the appendix (cf. Lemma 5.1) show that VS ˚ “ 2d ´ 1 ` ǫ with arbitrary ǫ ą 0. Obviously, the corresponding discretization Sn consists of hyperrectangles in Ind , which are determined by their upper left and lower right corners, i.e. # pSn q ď n2d . As S ˚ consists only of regular Borel sets, the assumption on the boundary smoothness is also satisfied. d 2. We may also consider the (smaller) set Q˚ of all hypercubes in r0, 1s , i.e. each Q˚ P Q˚ is d of the form rt, t ` hs with t P r0, 1s and 0 ă h ď 1 ´ max1ďiďd ti . As Q˚ Ă S ˚ , Assumption 1 is satisfied. More precisely, according to (Despres, 2014), ν pQ˚ q “ t 3d`1 2 u, and refined computations in the appendix (cf. Lemma 5.2) show VQ˚ “ 1 independent of d (as opposed to the VC-dimension). d 3. Let H˚ be the set of all half-spaces in r0, 1s , i.e. ! ) ( d ˇ H˚ :“ Ha,α \| α P R, a P Sd´1 , Ha,α :“ x P r0, 1s ˇ xx, ay ě α . The VC-dimension of H˚ is ď d ` 1 (see e.g. Devroye and Lugosi, 2001, Cor. 4.2), which proves that Assumption 1 is satisfied. On the other hand, we prove that VH˚ “ 2 (cf. Lemma 5.3 the appendix). 2.2 Limit theory Now we are in position to show that the quantiles of the multiscale statistic in (6) can be approximated uniformly by those of the Gaussian version in (9), and furthermore that Mn pRn , vq in (9) converges to a non-degenerate limit for v ě VR˚ . For the former we require a lower bound on the d smallest scale as given in (12). Given a discretized set of candidate regions Rn Ă 2In and c ą 0 we introduce ˇ ( Rn\|c :“ R P Rn ˇ \|R\| ě c . With this notation we can formulate our main theorem: 7 Theorem 2.4 (Weak P0 limit). Let Yi , i P Ind be a field of random variables as in (1), R˚ satisfy Assumption 1 and let prn qn be a sequence such that (12) holds true. Then it holds under P0 that ˘ D ` Tn Y, θ0 , Rn\|rn , v Ñ M pR˚ , vq n Ñ 8, as (17) with M pR˚ , vq as in (10) for any fixed v P R. If furthermore v ě VR˚ in (16), then M pR˚ , vq is almost surely finite and non-degenerate. Note that M does not depend on any unknown quantities and can e.g. be simulated. However, for practical purposes it is advantageous to use the finite sample Gaussian approximation in (9) to simulate quantiles for Tn as in (6). This is justified by the following theorem: Theorem 2.5 (Gaussian approximation). Let Yi , i P Ind be a field of random variables as in (1), and let R˚ be a set of candidate regions satisfying Assumption 1(a) and let prn qn Ă p0, 8q be a sequence such that (12) holds true. Let v P R be fixed. (a) Then it holds ˘ ˘ ` ` Tn Y, θ0 , Rn\|rn , v ´ Mn Rn\|rn , v “ OP ˜ˆ log12 pnq rn ˙1{10 ¸ (18) as n Ñ 8 with Mn as in (9). (b) For all q P R we have ˇ “ ` ˘ ‰ˇ ˘ ‰ “ ` lim ˇP0 Tn Y, θ0 , Rn\|rn , v ą q ´ P Mn Rn\|rn , v ą q ˇ “ 0. nÑ8 (19) Remark 2.6. Theorems 2.4 and 2.5 are compatible in the sense that for any set of candidate regions satisfying Assumption 1, v P R and any rn Ñ 0 satisfying (12) it holds ˘ D ` Mn Rn\|rn , v Ñ M pR˚ , vq as n Ñ 8, (20) with M pR˚ , vq as in (10). Example 2.7. Suppose that R˚ is a set of candidate regions satisfying Assumption 1. Let us discuss three important examples of the model (1). ` ˘ 1. Gaussian fields: Let Yi „ N θ, σ 2 where the variance σ 2 ą 0 is fixed. In this case, ψpθq “ 12 θ2 , and ˇ ˇ a ˇY R ´ θ0 ˇ TR pY, θ0 q “ \|R\| σ 2. Bernoulli fields: Let Yi „ Bin p1, pq with p P p0, 1q. The cases p “ 0 and p “ 1 have to be excluded to obtain a natural exponential family. However, in these cases one would screen the field correctly, anyway. The natural parameter of this exponential family is θ “ log pp{p1 ´ pqq, and using ψ pθq “ log p1 ` exp pθqq we compute g ¨ » ˛ f ˜ ¸ﬁ f Y 1 ´ Y f R R ﬂ. TR pY, θ0 q “ e2\|R\| –Y R log ˝ exppθ q ‚´ p1 ´ Y R q log 1 0 1`exppθ0 q exppθ0 q`1 3. Poisson fields: Let Yi „ Poipλq with λ P R. Again, λ “ 0 has to be excluded, but this case is again trivial. The natural parameter of the exponential family is θ “ log pλq, and using ψ pθq “ exp pθq we compute d „ ˆ  ˙ YR ´ pY R ´ exppθ0 qq . TR pY, θ0 q “ 2\|R\| Y R log exppθ0 q 8 Example 2.8 (P0 -limiting distribution in the hyperrectangle / hypercube case). Recall Example d 2.3 and let S ˚ be the set of all hyperrectangles and Q˚ be the set of all hypercubes in r0, 1s . Then for prn qn as in (12) it holds under P0 that for any ǫ ą 0 max rTS pY, λ0 q ´ penv p\|S\|qs SPSn \|S\|ěrn D Ñ sup 0ďsătď1 « and ﬀ ` d ˘ \|W prs, tsq\| a ´ penv n \|rs, ts\| \|rs ´ ts\| max rTQ pY, λ0 q ´ penv p\|Q\|qs QPQn \|Q\|ěrn D Ñ sup tPr0,1sd hPp0,1´max1ďiďd ti s „ ´ ¯ \|W prt, t ` hsq\| d ´ pen phnq v hd{2  d as n Ñ 8, where W is white noise on r0, 1s . Monte-Carlo simulations (by means of (9) with n “ 96 and d “ 2) of the densities of the right-hand sides with different values of v are shown in Figure 1. The smallest possible values of v which can be chosen according to Example 2.3 are given by the packing number, i.e. v “ 3 ` ǫ and v “ 1 respectively. The corresponding results are depicted in the top row of Figure 1 with ǫ “ 0 for simplicity. Alternatively, we can use the VC-dimension v “ 4 and v “ 3 respectively, which lead to the simulated distributions shown in the bottom row of Figure 1. This nicely illustrates that using a larger value of v will lead to larger quantiles and hence a loss of detection power: As the distributions of Mn pS n , 4q and Mn pQn , 3q are extremely close, detecting in the system of squares is not easier than detecting in the system of rectangles, even though the latter is by far bigger and more complex. The explanation for this is the penalization (41), which by appropriate choice of the parameter v can be tailored to the system R˚ . 2.3 Asymptotic power In this section we will analyze the power of our multiscale testing approach in the hypercube-case. The detection power clearly depends on the size and strength of the anomaly. To describe the latter, we will frequently employ the functions mpθq :“ ψ 1 pθq “ E rY s , vpθq :“ ψ 2 pθq “ V rY s for Y „ Fθ . Heuristics The key point for the following power considerations is that the observations in (1) can be approximated as a v pθi q Fθi ´ m pθi q Yi ´ m pθ0 q m pθi q ´ m pθ0 q a a a „ `a , (21) v pθ0 q v pθ0 q v pθ0 q v pθi q ´1{2 i.e. as ’signal’ v pθ0 q pm pθi q ´ m pθa 0 qq, which is non zero on the anomaly only, aplus a standardized noise component pFθi ´ m pθi qq { v pθi q which is scaled by a factor vi :“ v pθi q {v pθ0 q. In case of Gaussian observations with variance 1, one has vi ” 1 and recovers the situation considered by Sharpnack and Arias-Castro (2016). Whenever the ’signal’ part in (21) is strong enough, the anomaly should be detected. In the following, we will make this statement mathematically precise and also give a comparison of the multiscale testing procedure with an oracle procedure. 9 1 600 600 0.8 400 400 200 200 0.6 0.4 0.2 0 0 0 2 0 (a) Mn pS n , 3q 0 2 (b) Mn pQn , 1q 0 2 (c) Comparison 1 600 600 0.8 400 400 200 200 0.6 0.4 0.2 0 0 0 2 0 (d) Mn pS n , 4q 0 2 (e) Mn pQn , 3q 0 2 (f) Comparison Figure 1: Simulated densities of the weak limits. The histograms are obtained from 104 runs of the test statistic (9). For the comparison, the corresponding densities have been estimated ), Mn pQn , vq ( )). Top row: optimal by a standard kernel estimator (Mn pS n , vq ( calibration with the covering number, bottom row: alternative calibration using the VC-dimension. d Considered alternatives Consider a given family pQ˚n qnPN of hypercube anomalies Q˚n Ă r0, 1s with Lebesgue measure \|Q˚n \| “ an P p0, 1q. The corresponding discretized anomalies Qn :“ d Ind X nQ˚n Ă Ind have size \|Qn \| „ nd an . We will consider alternatives Ki,n in (4) where θn P Θn s.t. (22) θin “ θ1n IQn . The parameters θ1n determine the total strength of the anomaly, which is given by µn pQn q :“ a m pθ1n q ´ m pθ0 q a \|Qn \| . v pθ0 q Clearly, any anomaly with fixed size or strength can be detected with asymptotic probability 1. Therefore, we will consider vanishing anomalies in the sense that an Œ 0, µn pQn q Õ 8, as n Ñ 8. (23) Furthermore, we will restrict to parameters θ1n in (22) which yield uniformly bounded variances and uniform sub-exponential tails, this is ď E rexp psY qs ď C for all 0 ď s ď t and θ P tθ0 u Y tθ1n u , (24) v ď vi “ d nPN v pθ1n q ď v̄ v pθ0 q for all i P Ind , n P N (25) for Y „ Fθ with constants t ą 0, C ą 0 and 0 ă v ă v̄ ă 8. In case of Gaussian observations with variance σ 2 , (24) and (25) are obviously satisfied, for a Poisson field this means that the intensities are bounded away from zero and infinity. 10 d Oracle and multiscale procedure Recall that by Q˚ is the set of all hypercubes in r0, 1s (cf. Example 2.3), and Qn its discretization (cf. (3)). If the size an of the anomaly is known, but its position is then one would naturally ˇ still unknown, ( restrict the set of candidate regions to R˚O :“ Q˚ P Q˚ ˇ \|S ˚ \| “ an , and consequently scan only over (cf. (3)) ˇ ( d ˇ d ˚ ˚ ˚ RO n :“ Q Ă In Q “ In X nQ for some Q P RO . As for the true anomaly Q˚ P R˚O , its discretized RO n . This gives ` versionO Qn˘ alsoOsatisfies Qn P O rise to an oracle test, which rejects whenever Tn Y, θ0 , Rn , v ą q1´α,n where q1´α,n is the 1 ´ α ` ˘ quantile of Mn RnO , v as in (9). Similar as in Theorem 2.5 one can show that this quantile sequence ensures the oracle test to have asymptotic level α. The asymptotic power of this oracle test can be seen as a benchmark for any multiscale test. To obtain a competitive multiscale ` ˘ procedure, let us choose some rn satisfying (12), and furthermore assume that rn “ o nd an , as otherwise the multiscale procedure will never be able to detect the true anomaly (as it is not contained in the set of candidate regions which we scan over). As now position and size of the anomaly are unknown, we consider all such sets in R˚MS “ Q˚ as candidate regions and consequently scan over ˇ ( d ˇ d ˚ ˚ ˚ RMS n\|rn :“ Q Ă In Q “ In X nQ for some Q P Q and \|Q\| ě rn . ` ˘ Clearly the true anomaly Q˚ satisfies Q˚ P R˚MS , and by rn “ o nd an its discretized version Qn also satisfies Qn P RMS n\|rn . This gives rise to a multiscale test, which rejects whenever ¯ ´ ¯ ´ Mn pRMS n\|rn ,v q MS MS MS Tn Y, θ0 , RMS is the 1 ´ α quantile of M :“ q where q , v ą q , v R n 1´α,n 1´α,n 1´α n\|rn n\|rn as in (9). Theorem 2.5 ensures that the multiscale test has asymptotic level α. M pQ˚ ,v q ˚ ă 8 whenever v ě VQ˚ “ 1, and it holds Now, due to Theorem 2.4 q1´α :“ q1´α O MS ˚ q1´α,n ď q1´α,n ď q1´α ă8 for all n P N and v ě 1. Asymptotic power We will now show that the multiscale procedure described above (which requires no a priori knowledge on the scale of the anomaly) asymptotically detects the same anomalies with power 1 as the oracle benchmark procedure for a known scale. Hence, the penalty choice to calibrate all scales as in (6) (where R˚ “ Q˚ ), renders the adaptation to all scales for free, at least asymptotically. This can be seen as a structural generalization of (Sharpnack and Arias-Castro, 2016, Thms. 2 and 4), as under the alternative the whole distribution in (1) and not just its mean might change. Also the power considerations in Proksch et al. (2017) restrict to this simpler case. Theorem 2.9. In the setting described above, let an Œ 0 be a sequence of scales such that 12 plog nq {nd “ o pan q as n Ñ 8. Denote by ˙ ˆ ˙ ˆ ` ˘ µ´x x`µ 2 `Φ , xě0 F x, µ, σ :“ Φ ´ σ σ the survival function of a folded normal distribution with parameters µ P R and σ 2 ą 0, where Φ is the cumulative distribution function of N p0, 1q. Let furthermore v ě VQ˚ “ 1. If (23) is satisfied, then the following holds true: (a) The single scale procedure has asymptotic power ‰ “ ` ˘ O Pθn Tn Y, θ0 , RO n , v ą q1´α,n d ˜ “ α ` p1 ´ αqF O q1´α,n ` 2v log ˆ ¸ ˙ n n ? m pθ q ´ m pθ q 1 v pθ q 0 1 1 a ` op1q. , nd{2 an , an v pθ0 q v pθ0 q 11 ` ˘ (b) If an “ o nα´d with α ą 0 sufficiently small, then the multiscale procedure has asymptotic power ” ´ ¯ ı MS Pθn Tn Y, θ0 , RMS , v ą q 1´α,n n\|rn d ˜ ě α ` p1 ´ αqF MS q1´α,n ` 2v log ˆ ¸ ˙ m pθ1n q ´ m pθ0 q v pθ1n q 1 d{2 ? a , ` op1q. ,n an an v pθ0 q v pθ0 q Remark 2.10. In (Sharpnack and Arias-Castro, 2016) a` similar ˘ result in case of Gaussian observations is shown. Note that the condition that an “ o nα´d with α ą 0 sufficiently small is missing there, but is however necessary for the proof to work. In (Proksch et al., 2017) it suffices to assume an Œ 0, as the maximum can be explicitly controlled due to its Gumbel-limit. The above Theorem allows us to explicitly describe those anomalies which will be detected with asymptotic power 1: Corollary 2.11. Under the setting in this section and the Assumptions of Theorem 2.9 any such anomaly is detected with asymptotic power 1 either by the single scale or the multiscale testing procedure if and only if c ´ ¯ ? 2v pθ0 q log a1n ´ nd{2 an \|m pθ1n q ´ m pθ0 q\| a Ñ ´8 (26) v pθ1n q as n Ñ 8. ˘ ` Example 2.12. 1. In case of Gaussian observations Yi „ N ∆n IQn , σ 2 with variance σ 2 , where the baseline mean is 0 and ∆n the size of the anomaly, this yields detection if and only if d ˆ ˙ 1 d{2 ? as n Ñ 8. \|∆n \| n an Á σ 2v log an If we calibrate the statistic with the packing number v “ VQ˚ “ 1 (cf. Example 2.3), then this coincides with the well known asymptotic detection boundary for hypercubes, see e.g. Arias-Castro et al. (2005); Frick et al. (2014) for d “ 1, Butucea and Ingster (2013) for d “ 2, or Kou (2017) for general d. ˘ ` 2. For Bernoulli r.v.’s Yi „ Ber p0 IQcn ` pn IQn with p0 , pn P p0, 1q s.t. p0 ` pn ď 1, the condition (26) reads as follows: c ´ ¯ ? 2p0 p1 ´ p0 q log a1n ´ nd{2 an \|pn ´ p0 \| a Ñ ´8. pn p1 ´ pn q Note, that the minimax detection rate is unknown in this case to best of our knowledge. ˘ ` 3. For a Poisson field Yi „ Poi λ0 IQcn ` λn IQn with λ0 , λn ą 0, Theorem 2.9 and Corollary 2.11 can only be applied if λn is a bounded sequence. In this case, (26) reduces to c ´ ¯ ? 2λ0 log a1n ´ nd{2 an \|λn ´ λ0 \| ? Ñ ´8. λn Again, the minimax detection rate is unknown in this case to best of our knowledge. 12 3 Auxiliary results Our results rely heavily on a coupling result which allows us to replace the maximum over partial sums of standardized NEF r.v.’s by a maximum over a corresponding Gaussian version. This can be obtained from recent results by Chernozhukov et al. (2014) as soon as certain moments can be controlled, which is the purpose of the following two Lemmata, proofs can be found in Section 4. In what follows, the letter C ą 0 denotes some constant, which might change from line to line. We start with controlling the maximum of powers of uniformly sub-exponential random variables: Lemma 3.1. Let Wi , i “ 1, 2, . . . be independent sub-exponential random variables s.t. there exist k1 ą 1 and k2 ą 0 s.t. P r\|Wi \| ą ts ď k1 expp´k2 tq (27) for all i. Then for all m P N there exists a constant C, s.t. for all N ě 2  „ E max \|Wi \|m ď C plog N qm . 1ďiďN It is well-known that the above bound can be improved for sub-Gaussian random variables to „  a (28) E max \|Xi \| ď C log N . 1ďiďN Now we will show that the maximum over the partial sum process of independent random variables can be bounded by the maximum over the corresponding Gaussian version. The latter can be controlled as in (28) by exploiting the fact that a maximum over dependent Gaussian random variables is always bounded by a maximum over corresponding independent Gaussian random variables (see e.g. Šidák, 1967) ﬀ « a \|XI \| (29) ď C log p# pIqq E max a IPI \|I\| i.i.d. with Xi „ N p0, 1q and XI :“ ř Xi . This allows us to prove the following: iPI Lemma 3.2. Let pZi qi“1,...,N be independent random variables with E rZi s “ 0 and denote ZI :“ ř Zi . If I is an arbitrary index set of sets tIuIPI , then there exists a constant C ą 0 independent iPI of I s.t. « ﬀ „  a \|ZI \| E max a ď C log p# pIqq E max \|Zi \| . IPI 1ďiďN \|I\| With the help of these two lemmata, the following coupling result can then be shown: Theorem 3.3 (Coupling). Let Zi , i P Ind independent, E rZi s “ 0, V rZi s “ 1, such that (27) is satisfied for all i with uniform constants k1 ą 1 and k2 ą 0. Let furthermore vi P rv, vs, i P Ind with i.i.d. 0 ă v ď v ă 8 independent of i and n, and Xi „ N p0, 1q, i “ 1, . . . , nd , and Rn , s.t. inequality (14) holds. Then ˜ˆ ˙1{6 ¸ ÿ log10 pnq 1 ÿ ´1{2 vi Xi “ OP vi Zi ´ max \|R\| . max a RPRn : RPRn : rn \|R\| iPR iPR \|R\|ěrn \|R\|ěrn 13 4 Proofs In this section we will give all proofs. In the following we will denote by pn the cardinality of Rn , i.e. pn :“ #pRn q, which by (14) satisfies logppn q „ log n. Recall that C denotes a generic constant which might differ from line to line. 4.1 Proof of the auxiliary results We start with proving the auxiliary statements from section 3. Proof of Lemma 3.1. Let hptq :“ k1 expp´k2 tq, then „  „  N P max \|Wi \| ą t “ 1 ´ P max \|Wi \| ď t ď 1 ´ p1 ´ hptqq ď N hptq. 1ďiďN 1ďiďN Let t̄ “ h´1 p1{N q ď C logpN q, then „  „  ż8 m m´1 E max \|Wi \| “m t P max \|Wi \| ą t dt 1ďiďN 1ďiďN 0 ż t̄ ż8 tm´1 N hptq dt t̄ 0 ż8 m tm´1 expp´k2 tq dt ď pC logpN qq ` k1 mN ďm tm´1 dt ` m t̄ m ď C plog N q , where the last inequality follows from integration by parts. i.i.d. Proof of Lemma 3.2. Let Xi „ N p0, 1q and ri be i.i.d. Rademacher random variables, i.e. they take the values ˘1 with probability 1{2. Step (i): Since the Xi are symmetric ˇ ˇ ˇﬀ ˇﬀﬀ « « « ˇ 1 ˇˇÿ 1 ˇˇÿ ˇˇ ˇ (30) E max a ˇ Xi ˇ “ Er E max a ˇ ri Xi ˇ . I ˇ I \|I\| ˇiPI ˇ \|I\| ˇiPI By Lemma 4.5 of (Ledoux and Talagrand, 1991) and choosing F ptq “ t, ηi “ Xi and xi :“ pci,I qI , where ci,I :“ ?1 ItiPIu (a scaled indicator function) and as norm the max ´norm, we obtain \|I\| Er ˇ ˇ ˇﬀ c ˇﬀ « π 1 ˇˇÿ ˇˇ 1 ˇˇÿ ˇˇ E max a ˇ Xi ˇ . max a ˇ ri ˇ ď I I 2 \|I\| ˇiPI ˇ \|I\| ˇiPI ˇ « (31) Step (ii): Let pZi1 q1ďiďN be a sequence of independent copies of pZi q1ďiďN and define the symĂi :“ Zi ´ Z 1 and equally the symmetrized version of ZI by Z ĂI :“ metrized version of Zi by Z i ř 1 iPI pZi ´ Zi q. Then by using the same argument as in (30) and Fubini’s theorem, we derive ﬀ « ﬀ « 1 ˇˇ Ă ˇˇ 1 E max a \|ZI \| ď 2E max a ˇZI ˇ I I \|I\| \|I\| ˇ ˇﬀﬀ « « 1 ˇˇÿ Ă ˇˇ “ 2Er E max a ˇ Zi ri ˇ I ˇ \|I\| ˇiPI ˇﬀﬀ ˇ « « 1 ˇˇÿ Ă ˇˇ “ 2E Er max a ˇ Zi ri ˇ I ˇ \|I\| ˇiPI ˇ ˇﬀﬀ « « 1 ˇˇÿ ˇˇĂˇˇ ˇˇ “ 2E Er max a ˇ ˇZi ˇ ri ˇ , I ˇ \|I\| ˇiPI 14 where the last equality holds in view of the symmetry of ri . Now we will use the contraction principle, i.e. Theorem 4.4 of (Ledoux and Talagrand, 1991) with F ptq “ t conditionally on \|ZĂi pωq\| , which is independent of pri q. By choosing xi as in Step (i) we get (after αi :“ Ă max \|Zj pωq\| j Ăj pωq\|) multiplying both sides with max \|Z j Er Therefore ˇﬀ ˇ « „ ˇ ˇ 1 ˇˇÿ ˇˇĂ ˇˇ ˇˇ 1 ˇĂ ˇ max a ˇ ˇZi pωqˇ ri ˇ ď E max ˇZi pωqˇ Er max ? I 1ďiďN I ˇ ˇ I \|I\| iPI « ˇﬀ ˇ ˇÿ ˇ ˇ ˇ ˇ ri ˇ . ˇiPI ˇ ˇﬀ ˇ ﬀ « „ ˇ ˇ 1 1 ˇˇÿ ˇˇ ˇĂ ˇ E max a \|ZI \| ď 2E max ˇZi pωqˇ Er max a ˇ ri ˇ I 1ďiďN I \|I\| \|I\| ˇiPI ˇ ˇ ˇﬀ « „ c 1 ˇˇÿ ˇˇ π E max a ˇ Xi ˇ ď 4E max \|Zi \| I 1ďiďN 2 \|I\| ˇiPI ˇ « ﬀ „  ? \|XI \| , “ 8πE max \|Zi \| E max a 1ďiďN I \|I\| « where we used (31) in the second inequality. Now the statement follows from (29). Proof of Theorem 3.3. Enumerate each region in Rn by j, 1 ď j ď pn and define vi Xij :“ a Zi ItiPRj u It\|Rj \|ěrn u , \|Rj \| Xi :“ pXij qj“1,...,pn , for some sequence rn . Then Z :“ max 1ďjďpn řN i“1 (32) d i “ 1, . . . , N “ n , Xij satisfies 1 ÿ D vi Zi . Z “ max a RPRn : \|R\| iPR \|R\|ěrn Recall that logppn q À logpnq. According to (Chernozhukov et al., 2014, Cor. 4.1) we find that řN for every δ ą 0 there exists a Gaussian version Z̃ :“ max i“1 vi Nij with independent random 1ďjďpn vectors N1 , . . . , Nn in Rpn , Ni „ N p0, E rXi Xit sq, 1 ď i ď N , such that ˇ ı ”ˇ ( logpnq ˇ ˇ P ˇZ ´ Z̃ ˇ ą 16δ À δ ´2 B1 ` δ ´1 pB2 ` B4 q logpnq logpnq ` nd where ˇ ˇﬀ N ˇÿ ˇ ˇ ˇ B1 :“ E max ˇ pXij Xik ´ E rXij Xik sqˇ 1ďj,kďpn ˇ ˇ i“1 « ﬀ N ÿ 3 B2 :“ E max \|Xij \| « 1ďjďpn B4 :“ N ÿ i“1 E „ i“1 3 max \|Xij \| It 1ďjďpn max \|Xij \|ąδ{ logppn _nqu 1ďjďpn B1 can be controlled as follows. With Xij from (32) we derive ˇﬁ ˇ » ˇ ˇ ÿ 2 2 ˇ ˇ v pZ ´ 1q i i ˇﬂ ˇ – a max B1 “ E ˇ ˇ 1ďj,kďpn : ˇ \|R \|\|R \| j k ˇ iPR XR j k \|R \|,\|R \|ěr j k n 15  . » “ E– max 1ďj,kďpn : \|Rj \|,\|Rk \|ěrn ˇﬁ ˇ a ˇ ˇ ÿ ˇ \|Rj X Rk \| ˇˇ 1 2 2 a a vi pZi ´ 1qˇˇﬂ . ˇ \|Rj \|\|Rk \| ˇ \|Rj X Rk \| iPRj XRk ˇ Using the restriction on the size of the rectangles we find: d d \|Rj X Rk \| \| mint\|Rj \|, \|Rk \|u\| 1 ď ď ? . \|Rj \|\|Rk \| \|Rj \|\|Rk \| rn ř Denote Vi :“ vi2 pZi2 ´ 1q, I :“ Rj X Rk P I Ă Ind and SI :“ iPI Vi . Now « ﬀ \|SI \| 1 . B1 ď ? E max a IPI rn \|I\| Using Lemma 3.2 we obtain C a log p#pIqq E B1 ď ? loooooomoooooon rn ? „ It remains to estimate E „ \|vi2 pZi2 max 1ďiďN logpnq „ max 1ďiďN \|vi2 pZi2  ´ 1q\|  „  2 2 ´ 1q\| “ E max vi \|Zi ´ 1\| 1ďiďN  „ 2 2 ď E max v \|Zi ´ 1\| 1ďiďN „  2 2 ď v E max \|Zi \| ` v 2 . 1ďiďN So in total we get by Lemma 3.1 a ˆ 5 ˙1{2 ˘ logpnq ` 2 log pnq B1 À ? v C logpN q2 ` v 2 À . rn rn For B2 we compute B2 ď 1 1{2 prn q E „ 3 max \|vi Zi \| 1ďiďN ď v3  E 1{2 prn q „  ˆ 6 ˙1{2 log pnq max \|Zi \|3 À , 1ďiďN rn where we again used Lemma 3.1. Now let δ ą 0 be fixed. Then ﬁ » N 3 ÿ \|v Z \| i i ﬂ B4 ď E – max I \|vi Zi \| δ u ą plog 1ďjďpn : \|Rj \|3{2 tmax pq \|Rj \|1{2 j i“1 \|Rj \|ěrn ﬀ « N 3 ď 3{2 max E \|vi Zi \| I 1{2 δr u t\|Zi \|ą \|v \| n rn 1ďiďN i log p Now let rn ą ´ 2d\|vi \| δ ¯2 plog nq2`2γ for some γ ą 1 for n ě no pδ, dq and hence B4 ď N “ ‰ max E \|vi \|3 \|Zi \|3 It\|Zi \|ąplog nqγ u 3{2 rn 1ďiďN 16 ď N v3 3{2 rn 3 max 1ďiďN N v3 ż8 ż8 plog nqγ t2 P r\|Zi \| ą ts dt t2 expp´k2 tq dt 3{2 plog nqγ rn ż N v3 1 8 u2 expp´uq du. “ 3k1 3{2 3 k γ rn 2 k2 plog nq ď 3k1 For u large enough s.t. u2 ď exppu{2q, i.e. for n ě n1 pδ, d, γq, it holds ˆ ˙ ż8 ´ u¯ 3k1 v 3 N 3k1 v 3 N k2 γ B4 ď du “ exp ´ plog nq , exp ´ k23 rn3{2 k2 plog nqγ 2 k23 rn3{2 2 and then furthermore k2 2 γ plog nq ě pd log nq which implies B4 ď 3k1 v 3 nd 1 3k1 v 3 ´d n “ . 3 3 3{2 k2 prn q k2 prn q3{2 In conclusion we obtain ˆ 10 ˙1{2 ˙1{2 ˆ 7 ” ı log pnq log pnq ` δ ´3 P \|Z ´ Z̃\| ą 16δ Àδ ´2 rn rn ˆ 4 ˙1{2 log pnq logpnq ` δ ´3 , ` rn3 nd which yields the claim. 4.2 Proofs of Section 2.2 Let us now prove the results from Section 2.2, including Theorems 2.4 and 2.5. We start with a Taylor expansion of Tn , which will allows us to apply Theorem 3.3. Lemma 4.1. Let Rn be a collection of sets s.t. (14) holds, ǫ ą 0 and prn qn Ă p0, 8q be a sequence, s.t. plog nq10`ǫ {rn Ñ 0. Suppose Yi „ Fθ0 P F , i P Ind , are i.i.d. random variables, and recall that ´1 ř for R P Rn we denote Y R :“ \|R\| iPR Yi . Then it holds ˇ ˇ ˜ˆ ˙1{4 ¸ ˇ ˇ 1 \|Y R ´ mpθ0 q\| ˇ log3 pnq ˇ 2 a max ˇTR pY, θ0 q ´ \|R\| ˇ “ OP RPRn : ˇ rn vpθ0 q ˇ \|R\|ěrn as n Ñ 8. Proof. For independent Gaussian random variables Xi „ N p0, 1q it follows from (28) and(14) that ˇﬀ «ˇ ˇ ÿ ˇˇ a ˇ ´1{2 Xi ˇ ď C log n, E ˇ max \|R\| ˇ ˇRPRn iPR hence ÿ 1 a max \|R\|´1{2 Xi “ oP p1q. logpnq RPRn : iPR \|R\|ěrn Combining this result with Theorem 3.3 (with vi “ 1 for all i) we obtain ˇ ˇ 1 ˇˇ ÿ Yi ´ mpθ0 q ˇˇ 1 ? a max a ˇ “ oP p1q. ˇ log n RPRn : \|R\| ˇiPR vpθ0 q ˇ \|R\|ěrn 17 (33) Together with (14) it follows ˇ ˇ max ˇY R ´ mpθ0 qˇ ď C RPRn : \|R\|ěrn ˆ logpnqvpθ0 q rn ˙ 12 p1 ` oP p1qq Ñ 0, n Ñ 8. ? Therefore, Y R ą mpθ0 q{ 2 in probability if n is` large uniformly ˘ enough “ ‰ over R, s.t. \|R\| ě rn . Now we are in position to analyze JpY R , θ0 q “ φ Y R ´ θ0 Y R ´ ψ pθ0 q in the definition of TR , ś see (13). As the sup pθ pYi q is attained at that θ for which ψ 1 pθq “ Y R we derive θPΘ iPR ` ˘ φpY R q “ xm´1 pY R q, Y R y ´ ψ m´1 pY R q , and therefore ` ˘ ` ˘ JpY R , θ0 q “ xm´1 pY R q, Y R y ´ ψ m´1 pY R q ´ xθ0 , Y R y ´ ψpθ0 q ` ` ˘ ˘ “ xm´1 pY R q ´ θ0 , Y R y ´ ψ m´1 pY R q ´ ψpθ0 q . Note that ȲR P Dpm´1 q for large enough n, as the latter is an open set. A Taylor expansion of ψ around θ0 and one of second order of m´1 around mpθ0 q, yields ¨ TR pY, θ0 q “ ˝2\|R\| ˜ Y R ´ mpθ0 q a vpθ0 q ¸2 ` 2\|R\|sn ˜ ¸˛1{2 Y R ´ mpθ0 q ‚ a vpθ0 q (34) with sn s.t. \|sn pxq\| ď cx3 ` oP p1q for some c ą 0. Consequently ˇ ˇ ˇ1 2 pY R ´ mpθ0 qq2 ˇˇ ˇ TR pY, θ0 q ´ \|R\| max ˇ RPRn : ˇ 2 vpθ0 q \|R\|ěrn À max \|R\| RPRn : \|R\|ěrn ´ 12 “ max \|R\| RPRn : \|R\|ěrn \|Y R ´ mpθ0 q\|3 ` oP p1q vpθ0 q3{2 ˇř ˇ3 ˇ ˇ ˇ iPR pYi ´ mpθ0 qq ˇ ˇ a a ˇ ` oP p1q ˇ \|R\| vpθ0 q ˇ ďplog3 pnqrn´1 q1{2 ` oP p1q, 1 where we again used (33). Now \|a ´ b\| ď \|a2 ´ b2 \| 2 yields the claim. Now we are in position to prove Theorem 2.5. So far we have only shown that the maximum over the local likelihood ratio statistics can be approximated by Gaussian versions, but we did not include the scale penalization penv p\|R\|q in (7). To include this in the approximation result, we will slice the maximum into scales, where the penalty-term is almost constant. Then, we show that we may bound the maximum over all scales by the sum of the maximum over theses families. The price to pay is an additional logpnq factor on the smallest scale. Proof of Theorem 2.5. (a) It follows from the triangle inequality \|}x}8 ´ }y}8 \| ď }x ´ y}8 , Lemma 4.1 and (12) that ˇ ˇ ˇ ˇ max pTR pY, θ0 q ´ penv p\|R\|qq ´ ˇ RPRn : ˇ\|R\|ěr n 18 max RPRn : \|R\|ěrn ˇ ˇ ¸ˇˇ ˜ˆ ˙1{4 ¸ ˇ Y ´ mpθ q ˇ ˇ log3 pnq ˇ ˇ R 0 1{2 ˇ \|R\| ˇ a . ˇ ´ penv p\|R\|q ˇ “ OP ˇ rn vpθ0 q ˇ ˇ ˜ Define Y R :“ \|R\|´1{2 X R :“ \|R\|´1{2 ÿ iPR ÿ ˜ Yi ´ mpθ0 q a vpθ0 q Xi , iPR ¸ i.i.d. Xi „ N p0, 1q. With this notation and a symmetry argument we find from the proof of Theorem 3.3 with vi ” 1 that ˇ ﬁ »ˇ ˇ ˇ ˆ 10 ˙1{2 ˇ R ˇˇ ˇ Rˇ ˇ log pnq P –ˇˇ max ˇY ˇ ´ max ˇX ˇˇˇ ą δ ﬂ À δ ´3 . RPRn : RPRn : rn ˇ ˇ\|R\|ěr \|R\|ěr n n 12 1{10 Let δn :“ pplog pnq{rn q Œ 0. Now define ǫj :“ jδn , j P N and Rn,j :“ tR P Rn \| exppǫj q ă \|R\| ă exppǫj`1 qu . Then the set of candidate regions Rn can be written as * " ğ ` ˘ 1 1 log log12 pnq , . . . , logpnd q Rn\|rn “ Rn,j , J :“ δn δn jPJ with \|J\| ď logpnd q δn . If we abbreviate penj :“ penv pexp pǫj qq “ d ˆ ˆ 2v log nd exppǫj q ˙ ˙ `1 , then the slicing above implies penj`1 ď penv p\|R\|q ď penj , ´? ? ? ¯ ? Using a ´ b “ pa ´ bq{ a ` b , we get for all R P Rn,j . 0 ď penj ´ penj`1 2vpǫj`1 ´ ǫj q a “a . 2v rlogpnd q ` 1 ´ ǫj s ` 2v rlogpnd q ` 1 ´ ǫj`1 s The largest index in J is δ1n logpnd q and therefore the maximal value of ǫi is given by ǭ “ logpnd q and logpnd q ` 1 ´ ǭ “ 1. Therefore, c 2vpǫj`1 ´ ǫj q v ? “ δn 0 ď penj ´ penj`1 ď . 2 2 2v This means that for n Ñ 8 the penalty terms penv p\|R\|q, R P Rn,j can be considered as constant. Now straight forward computations show that ˇ ˇ ˇ ˇ ˇ ˘ `ˇ R ˇ ˘ˇ `ˇ R ˇ ˇ max ˇY ˇ ´ penv p\|R\|q ´ max ˇX ˇ ´ penv p\|R\|q ˇ ˇ RPRn : ˇ RPRn : ˇ\|R\|ěr ˇ \|R\|ěrn n c ˇ ˇ ˇ ˇˇ ˇ ˇ ˇ v ď max ˇˇ max ˇY R ˇ ´ max ˇX R ˇˇˇ ` δn . RPRn,j jPJ RPRn,j 2 d q Now the claim follows from \|J\| ď logpn δn . (b) This is a direct consequence of (a). 19 We will now continue with the proof of Theorem 2.4. Taking into account the result of Theorem 2.5, we only have to prove Remark 2.6, which will be done in the following. Lemma 4.2. Let R˚ satisfy Assumption 1 and be equipped with the canonical metric ρ˚ as in (15) i.i.d. d and define Rn as in (3). Furthermore let W denote white noise on r0, 1s . For Xi „ N p0, 1q, d i P In define Zn pR˚ q :“ n´d{2 ÿ i{nPR˚ ÿ D Xi “ n´d{2 iPt1,...,nud \|nR˚ X Ai \| Xi , R˚ P R˚ where Ai “ pi1 ´ 1, i1 s ˆ . . . ˆ pid ´ 1, id s is the unit cube with upper corner i. Then it holds D Zn Ñ W, n Ñ 8. Proof. Note that R˚ is totally bounded w.r.t. ρ˚ . We will show the assumptions of (Kosorok, 2008, Thm. 2.1): 1. Tightness: The white noise W is tight. 2. Totally boundedness: By Markov’s inequality and standard bounds on the modulus of continuity, we obtain using Assumption 1(a) that ﬁ » — P˚ — – sup ˚ ˚ R˚ 1 ,R2 PR ˚ ρpR˚ 1 ,R2 qďδ » ﬃ \|Zn pR1˚ q ´ Zn pR2˚ q\| ą ǫﬃ ﬂ ﬁ ﬃ 1 — sup \|Zn pR1˚ q ´ Zn pR2˚ q\|ﬃ ď E— ﬂ – ǫ R˚ ,R˚ PR˚ 1 2 ˚ ρpR˚ 1 ,R2 qďδ żδa À u´2νpR˚ q ´ 1 du 0 which tends to 0 as δ Œ 0. 3. Finite dimensional convergence: The convergence of the finite-dimensional laws is an application of the central limit theorem for random fields (Dedecker, 1998, Thm 2.2) and (Dedecker, 2001, Lemma 2), which shows that ˇ ˚ ˇ ˇnR X Zd ˇ Ñ \|R˚ \| nd for regular Borel sets R˚ Ă r0, 1sd with \|R˚ \|q ą 0. Consequently, the central limit theorem shows for any fixed R˚ P R˚ that D Zn pR˚ q Ñ N p0, \|R˚ \|q as nÑ8 A similar computation shows that Cov pZn pR1˚ q, Zn pR2˚ qq Ñ \|R1˚ X R2˚ \| for all R1˚ , R2˚ P R. This shows finite dimensional convergence. 20 Now we want to apply the generalized version of the continuous mapping theorem (see e.g. Billingsley, 2013, Thm. 5.5). For c ě 0 and x P CpBpr0, 1sdq, Rq, where Bpr0, 1sdq denote the Borel sets of r0, 1sd define ¸ ˜ ˚ ` ˘ \|xpR q\| a ´ penv nd \|R˚ \| hc pxq :“ sup \|R˚ \| R˚ PR˚ : \|R˚ \|ącd hcn pxq :“ max RPRn : \|R\|ąpcnqd ˜ ¸ \|xpR{nq\| a ´ penv p\|R\|q . \|R\|{nd The necessary conditions to apply the continuous mapping theorem are given by the following Lemma: ˘ ` Lemma 4.3. Consider hc , hcn as functions CpBpr0, 1sdq, Rq, } ¨ }8 Ñ R. i) hc is uniformly continuous and phcn qnPN is a sequence of equi-continuous functions, (uniformly in n). ii) For pxn qn P CpBpr0, 1sdq, Rq, s.t. xn Ñ x it holds hcn pxn q Ñ hc pxq, n Ñ 8. Proof. i) Let ǫ ą 0, choose δ “ ǫcd{2 . Consider two functions x, y P CpBpr0, 1sdq, Rq s.t. dpx, yq “ sup \|\|xpR˚ q\| ´ \|ypR˚ q\|\| ă δ. By using \|max ai ´ max bi \| ď max \|ai ´ bi \| we find R˚ Ăr0,1sd \|hcn pxq ´ hcn pyq\| ˇ ˇ ˇ \|xpR{nq\| ´ \|ypR{nq\| ˇ δ ˇ ˇ a ď max ˇ ˇ ď d{2 “ ǫ. d RPRn ˇ ˇ c \|R\|{n d \|R\|ąpcnq Similar arguments yield the uniform continuity of hc . ii) Let pxn qn , x P CpBpr0, 1sdq, Rq, s.t. xn Ñ x. Since the functions phcm qmPN are equicontinuous, for any ǫ ą 0 we can find an N1 P N s.t. @n ą N1 @m : \|hcm pxn q ´ hcm pxq\| ă ǫ . 2 Given ǫ and N1 and n ą N1 with \|hcm pxn q ´ hcm pxq\| ă 2ǫ , choose m “ n. Then \|hcn pxn q ´ hcn pxq\| ă ǫ{2. (35) Now let us define ( A :“ R˚ P R˚ : \|R˚ \| ě cd , ( Bn :“ R{n P R˚ : R P Rn , \|R\| ě pcnqd . The set A is a compact set w.r.t. the metric ρ˚ defined in (15), w.r.t. which R˚ is totally d bounded. Furthermore Bn is a finite subset of A. If we fix x P Bpr0, 1s q and introduce g : A Ñ R by ¸ ˜ \|xpR˚ q\| ˚ ˚ ´ penv p\|R \|q , R˚ P R˚ , gpR q :“ a \|R˚ \| then it holds gpR˚ q. hc pxq “ sup gpR˚ q ď hcn pxq “ max ˚ R PBn R˚ PA 21 (36) since Bn is a subset of A. Straight forward computations show that g is continuous w.r.t. r P A s.t. hc pxq “ gpRq. r Now ρ˚ , which implies by compactness of A that there exists an R r r let Rn P Bn be a sequence s.t. Rn Ñ R, n Ñ 8 w.r.t. ρ. Then gpRn q Ñ gpRq as n Ñ 8 and hence (36) hc pxq ě hcn pxq ě gpRn q Ñ gpR̃q “ hc pxq. Consequently there exists a N2 P N s.t. @n ą N2 it holds \|hc pxq ´ hcn pxq\| ă ǫ{2, which together with (35) implies \|hcn pxn q ´ hc pxq\| ď ǫ for all n ą maxtN1 , N2 u. Now we are in position to prove Remark 2.6: Proof of Remark 2.6. By Proposition 4.2 and the generalized version of the continuous mapping theorem (see e.g. Billingsley, 2013, Thm. 5.5) we get D hcn pZn q Ñ hc pW q, n Ñ 8. The functions hcn and hc have been defined such that ´ ¯ and hc pW q “ M R˚\|cd , v , ` ˘ hcn pZn q “ Mn Rn\|pcnqd , v , i.e. for all c ą 0 holds ´ ¯ ` ˘ D Mn Rn\| pcnqd , v Ñ M R˚\|cd , v n Ñ 8. as ¯ ´ D Since M R˚\|cd , v Ñ M pR˚ , vq , c Ñ 0, we get ` ˘ lim lim Mn Rn\| pcnqd , v “ M pR˚ , vq . cÑ0 nÑ8 It can also readily be seen from the definition that ˘ ‰ “ ` lim inf P Mn Rn\| rn , v ď t ě P rM pR˚ , vq ď ts . nÑ8 Now let c ą 0 be fixed and assume rn ă pcnqd for all n P N. Then we obtain altogether that ˘ ‰ “ ` P rM pR˚ , vq ď ts ď lim inf P Mn Rn\| rn , v ď t nÑ8 “ ` ˘ ‰ ď lim sup P Mn Rn\| pcnqd , v ď t nÑ8 Ñ P rM pR˚ , vq ď ts as c Œ 0, which yields the claim. Proof of Theorem 2.4. The main statement follows from Theorem 2.5 together with Remark 2.6. It remains to show a.s. boundedness and non-degenerateness of M pR˚ , vq. We apply (Dümbgen and Spokoiny, 2001, Thm. 6.1) with ρ˚ as in (15) and σ 2 pR˚ q :“ \|R˚ \| , XpR˚ q :“ W pR˚ q. 22 Let us check the three conditions from their theorem: i) σ 2 pR1˚ q ď σ 2 pR2˚ q ` ρ˚ pR1˚ , R2˚ q2 for all R1˚ , R2˚ P R˚ is obviously fulfilled since R1˚ X R2˚ Ă R2˚ and R1˚ zR2˚ Ă R1˚ △ R2˚ . Since V rW pR˚ qs “ \|R˚ \|, ˆ 2˙ ” ı 1 η . (37) P rXpR˚ q ą σpR˚ qηs “ P W pR˚ q ą η p\|R˚ \|q1{2 ď exp ´ 2 2 ii) For ” ı P r\|XpR1˚ q ´ XpR2˚ q\| ą ρpR1˚ , R2˚ qηs “ P \|W pR1˚ q ´ W pR2˚ q\| ą \|R1˚ △ R2˚ \|q1{2 η 2 2 we compute that W pR1˚ q´W pR2˚ q „ N p0, σR “ \|R1˚ \|`\|R2˚ \|´2 CovpW pR1˚ q, W pR2˚ qq ˚ ˚ q, σ ˚ R1 ,R˚ 1 ,R2 2 and \|R1˚ △ R2˚ \| “ \|R1˚ \| ` \|R2˚ \| ´ 2 \|R1˚ X R2˚ \|. With CovpW pR1˚ q, W pR2˚ qq “ \|R1˚ X R2˚ \| we consequently find ˜ ¸ ˆ 2˙ η 2 \|R1˚ △ R2˚ \| η ˚ ˚ ˚ ˚ P r\|XpR1 q ´ XpR2 q\| ą ρpR1 , R2 qηs ď exp ´ “ exp ´ . 2 2 σR 2 ˚ ˚ ,R 1 2 iii) Is fulfilled by Assumption 1(a) (cf. Remark 2.2). (37) holds with ´W pR˚ q as well, hence we get that the statistic M pR˚ , vq ă 8 a.s. Nondegenerateness is obvious, as M is always larger than the value of the local statistic on one fixed scale, which is non-degenerate. 4.3 Proofs of Section 2.3 Let us now prove the results from Section 2.3, namely Theorem 2.9 and Corollary 2.11. First we introduce some abbreviations to ease notation. Let O q ˚ :“ q1´α,n , MS q :“ q1´α,n and denote the total signal on Q P Qn by ´1{2 µn pQq :“ \|Q\| ÿ mpθn q ´ mpθ0 q \|Q X Qn \| m pθ1n q ´ m pθ0 q i a a “ a . vpθ0 q \|Q\| v pθ0 q iPQ For brevity introduce the Gaussian process ˇ ˇ ˇ ˇ ÿ 1 ˇ ˇ n vi Xi ˇ ´ penv p\|Q\|q , γ pQq :“ ˇµ pQq ` \|Q\|´ 2 ˇ ˇ iPQ (38) Q P Qn a i.i.d. with Xi „ N p0, 1q and vi “ v pθi q {v pθ0 q. Let us now start with the analysis of the oracle procedure. As a preparation we require to leave out a suitable subset of hypercubes close to the true anomaly Qn . Therefore, choose a sequence εn such that εn Œ 0 but εn µn pQn q Ñ 8 and denote the set of all hypercubes which are close to the anomaly by ˇ ( Un :“ Q P Qn pan q ˇ µn pQq ě µn pQn q p1 ´ εn q . Furthermore define the extended neighborhood of the anomaly by ˇ ( U :“ Q P Qn pan q ˇ Q X Q1 ‰ H for some Q1 P Un , its complement by T :“ Qn pan q zU. By definition, tγ pQquQPT and tγ pQquQPUn are independent, which will allow us to compute the asymptotic power of the single-scale procedure. For a sketch of Un and U see Figure 2. 23 Figure 2: Exemplary elements of the sets Un , U and T in d “ 2: The anomaly is shown in red, the hatched cubes belong to Un , the dotted cubes to U, and all black cubes belong to T . By definition, for all Q P Un and Q1 P T it holds Q X Q1 “ H, which implies independence of tγ pQquQPT and tγ pQquQPUn . Lemma 4.4. Consider the setting from Section 2.3 and recall that q ˚ is the p1 ´ αq´quantile of Mn pQn pan qq as in (9). Then ˇ ˇ 1 ˇř ˇ (a) max \|Q\|´ 2 ˇ iPQ vi Xi ˇ “ OP p1q as n Ñ 8 QPU „  (b) lim P max γ pQq ď q ˚ “ 1 ´ α nÑ8 QPT 2 Proof. We starta with bounding the covering number N pU, ρ, ǫq w.r.t. the canonical metric ρ pQ, Q1 q “ 1 2 ´ 2 \|Q X Q \| { \|Q\| \|Q1 \|. First note that all sets in U have fixed size \|Qn \|. Hence, given some ǫ ą 0, one has ¯ ˇ ˇ ´ ` ˘2 ˇQ X Q1 ˇ ě 1 ´ ǫ \|Qn \| , ρ Q, Q1 ď ǫ ô 2 ˘ T P` ǫ 1 2 1 i.e. ρ pQ, Q q ď ǫ if and only if Q X Q contains at least 1 ´ 2 \|Q ? n \| many voxels. Consequently, as we onlyP`deal with it is straight forward to construct a ǫ-covering of U with cardinality ˘ cubes, T ď 5 \|Qn \| { 1 ´ 2ǫ \|Qn \| , i.e. 5 \|Qn \| T. N pU, ρ, ǫq ď P` 2˘ 1 ´ ǫ2 \|Qn \| (a) It follows from Dudley’s entropy integral (see e.g. Marcus and Rosen, 2006, Thm. 6.1.2) with any fixed Q1 P U that ˇ ˇﬀ « ˇÿ ˇ ˇ ˇ E max \|Q\|´1{2 ˇ vi Xi ˇ QPU ˇ ˇ iPQ ˇˇﬀ ˇﬀ ˇ ˇ ˇ ˇ ˇ « « ˇˇ ˇ ˇ ˇÿ ˇÿ ˇ ˇÿ ˇˇ ˇ ˇ ˇ ˇ ˇ ˇ 1 ´1{2 ´1{2 ´ \|Q \| v X ďE \|Q1 \|´1{2 ˇ vi Xi ˇ ` E max v X \|Q\| ˇˇ ˇ ˇ ˇ ˇ i i i i ˇˇ ˇ ˇ ˇiPQ1 ˇiPQ1 Q,Q1 PU ˇ ˇiPQ d c c ˙ ˆ ż2a ż2 2v 2 2v 2 ǫ2 `C `C dǫ ă 8 ď log N pU, ρ, ǫq dǫ ď ´ log 1 ´ π π 2 0 0 which by Markov’s inequality proves the claim. (b) A direct consequence of (a) is that ˇ ˇ « ﬀ ˇÿ ˇ ˇ ´ 12 ˇ max \|Q\| ˇ vi Xi ˇ ´ penv p\|Q\|q “ oP p1q. QPU ˇiPQ ˇ 24 Furthermore note that µn pQq “ 0 and vi ” 1, i P Q for Q P T . Consequently ˇ ˇ ﬀ ﬀ « « „  ˇÿ ˇ 1 ˇ ˇ Xi ˇ ´ penv p\|Q\|q ď q ˚ P max γ pQq ď q ˚ “P max \|Q\|´ 2 ˇ ˇiPQ ˇ QPT QPT ˇ ˇ ﬀ ﬀ « « ˇÿ ˇ ˇ ˚ ´ 12 ˇ Xi ˇ ´ penv p\|Q\|q ď q ` op1q “P max \|Q\| ˇ ˇiPQ ˇ QPQn pan q “P rMn pQn pan qq ď q ˚ s ` op1q which yields the claim. With this Lemma at hand, we are now in position to derive the asymptotic power of the oracle procedure: Proof of Theorem 2.9(a). To analyze Pθn rTn pY, θ0 , Qn pan qq ą q ˚ s, we start with showing a ě in the statement of Theorem 2.9(a). By Lemma 4.1 and the triangle inequality we can replace Tn pY, θ0 , Qn pan qq by ˇ ˇ ﬀ « ˇ ÿ Y ´ mpθ q ˇ i 0 ˇ ´ 12 ˇ a max \|Q\| ˇ ˇ ´ penv p\|Q\|q ˇiPQ QPQn pan q vpθ0 q ˇ up to oP p1q. Furthermore (21) and Theorem 3.3 allow us to approximate the latter sum by a Gaussian version, i.e. „  n ˚ ˚ max γ pQq ą q ` op1q. Pθn rTn pY, θ0 , Q pan qq ą q s “ P n QPQ pan q Now we derive P „ max n QPQ pan q „" γ pQq ą q ˚  * "  ˚ ˚ max γ pQq ą q X max γ pQq ď q QPT QPQn pan q „" "  ˚ ˚ `P max γ pQq ą q X max γ pQq ą q QPT QPQn pan q „" "  „  “P max γ pQq ď q ˚ X max γ pQq ą q ˚ ` P max γ pQq ą q ˚ QPT QPU QPT „"  „  ěP max γ pQq ď q ˚ X tγ pQn q ą q ˚ u ` P max γ pQq ą q ˚ QPT QPT „  „  “P max γ pQq ď q ˚ P rγ pQn q ą q ˚ s ` P max γ pQq ą q ˚ “P QPT QPT where we exploited Qn P U and independence of tγ pQquQPT and γ pQn q. Lemma 4.4(b) states that P rmaxQPT γ pQq ď q ˚ s “ 1 ´ α ` op1q and hence Pθn rTn pY, θ0 , Qn pan qq ą q ˚ s ě α ` p1 ´ αq P rγ pQn q ą q ˚ s ` op1q. Furthermore note that γ pQn q ` penv p\|Qn \|q follows a folded normal distribution with parameters ´1 ř 2 µ “ µn pQn q and σ 2 “ \|Qn \| iPQn vi , this is ˇ ` ˘ˇ γ pQn q „ ˇN µ, σ 2 ˇ ´ penv p\|Qn \|q . We compute µn pQn q “ a m pθ1n q ´ m pθ0 q a nd an 2, 5C p1 ` o p1qq , v pθ0 q 25 (39) ÿ v pθ1n q , v pθ0 q iPQn b ` ˘ penv p\|Q\|q “ 2v log a´1 ` op1q n \|Qn \|´1 vi2 “ (40) for all Q P Qn pan q , (41) which yields by continuity of F and Qn P Qn pan q the proposed lower bound. For the upper bound (i.e. ď in the statement of Theorem 2.9(a)) we proceed as before and obtain Pθn rTn pY, θ0 , Qn pan qq ą q ˚ s „" * "  ˚ ˚ “α ` P max γ pQq ď q n X max γ pQq ą q ` op1q QPT QPU „" " * "  ˚ ˚ ˚ “α ` P max γ pQq ď q X max γ pQq ą q X max γ pQq ą q QPT QPU QPUn „" " * "  ˚ ˚ ˚ ` P max γ pQq ď q X max γ pQq ą q X max γ pQq ď q ` op1q QPT QPU QPUn „" "  „  ˚ ˚ ˚ ďα ` P max γ pQq ď q X max γ pQq ą q ` P max γ pQq ą q ` op1q QPT QPUn QPU zUn „  „  ˚ ˚ “α ` p1 ´ αq P max γ pQq ą q ` P max γ pQq ą q ` op1q QPUn QPU zUn where we used independence of tγ pQquQPT and tγ pQquQPUn . From Lemma 4.4(a) we obtain ˇ ˇ ˇ ˇ ˇ ˇ ˇ ˇ ˇ ´ 21 ÿ ˇ ˇ ˇ ´ 12 ÿ vi Xi ˇ ď max ˇ\|Q\| vi Xi ˇ “ OP p1q max ˇ\|Q\| QPU ˇ ˇ ˇ ˇ QPU zUn iPQ iPQ and further by definition of Un that µn pQq ď p1 ´ εn q µn pQn q for all Q P UzUn . Exploiting (41) this implies „  ˚ P max γ pQq ą q QPU zUn „  b ` ´1 ˘ n ˚ ďP p1 ´ εn q µ pQn q ` OP p1q ´ 2 log an ą q „  b ` ´1 ˘ n n ˚ “P µ pQn q ´ 2 log an ´ εn µ pQn q ` OP p1q ą q “ op1q b ` ˘ if µn pQn q ´ 2 log a´1 Ñ C P r´8, 8q (as εn µn pQn q Ñ 8 by construction), and if µn pQn q ´ n b ` ˘ 2 log a´1 Ñ 8, then nothing has to be shown. Altogether this gives n „  Pθn rTn pY, θ0 , Qn pan qq ą q ˚ s ď α ` p1 ´ αq P max γ pQq ą q ˚ ` op1q. QPUn With similar arguments as in Lemma 4.4 we obtain from εn Œ 0 that „  ˚ P max γ pQq ą q “ P rγ pQn q ` oP p1q ą q ˚ s QPUn and hence the claim is proven. Now we turn to the multiscale procedure. As here different scales are considered, the set U is not large enough any more. Especially, we cannot construct a subset V such that tγ pQquV c and γ pQn q are independent and maxQPV γ pQq is still negligible. Due to this, the corresponding proof in Sharpnack and Arias-Castro (2016) is incomplete. To overcome this difficulty, we follow the 26 idea to distinguish if the anomaly Qn has asymptotically an effect on γ pQq or not. Whenever Q is sufficiently large compared to Qn , the impact will ´ ¯ asymptotically be negligible. ´γ For some sequence ǫn Œ 0 with ǫn “ O \|Qn \| with some γ ą 0 we introduce # d n δn :“ ǫn max µ pQn q , logpnq \|Qn \| rn +´1 , (42) ˇ ( V :“ Q P Qn \|rn ˇ µn pQq ě δn µn pQn q and its complement T 1 :“ Qn \|rn zV. For a sketch see Figure 3. Figure 3: Exemplary elements of the sets V and T 1 in d “ 2: The anomaly is shown in red, the hatched cubes belong to V and the dotted cubes to T 1 . However, the intersections marked in black are small enough such that they have asymptotically no influence on γ pQq. Opposed to the oracle procedure, we do not have independence of tγ pQquQPT 1 and γ pQn q. However, asymptotically a similar property holds true as shown in the following Lemma: Lemma 4.5. Consider the setting from Section 2.3 and recall that q is the p1 ´ αq´quantile of Mn pQn \|rn q as in (9). Then the following statements hold true as n Ñ 8: ˇ ˇ ´a ¯ a 1 ř ˇ ˇ ln p\|Qn \|q ` ln p´ ln pm pθ1n q ´ m pθ0 qqq (a) max ˇ\|Q\|´ 2 iPQ vi Xi ˇ “ OP QPV ˇ ˇ ˇ ˇ ˇ ´ 12 ř ˇ (b) max1 ˇ\|Q\| vi Xi ˇ “ oP p1q QPT ˇ ˇ iPQXQn „  (c) P max1 γ pQq ď q “ 1 ´ α ` op1q QPT Proof. Let us again start with the canonical metric abounding the covering number NnpV, ρ, ǫq w.r.t. n 1 2 1 1 ρ pQ, Q q “ 2 ´ 2 \|Q X Q \| { \|Q\| \|Q \|. For all Q P V it holds µ pQq ě δn µ pQn q, which implies δn a \|Q X Qn \| \|Qn \| \|Qn \| ď a ďa . \|Q\| \|Q\| Consequently, V contains only cubes Q with rn ď \|Q\| ď δn´2 \|Qn \|.?For a fixed scale k, V contains at most pC \|Qn \|q Q’s with \|Q\| “ k, and for the set ofP`such Q’s ˘ Tan ǫ-covering can be constructed as in the proof of Lemma 4.4 with at most C \|Qn \| { 1 ´ 2ǫ k elements, which gives N pV, ρ, ǫq ď C ´2 tδn \|Qn \|u ÿ k“rn ˙´1 ˆ ˙ ˆ ǫ2 \|Qn \| \|Qn \| P` ˘ T ďC 1´ . \|Qn \| ln 2 δn2 rn 1 ´ 2ǫ k 27 (a) Again with the help of Dudley’s entropy integral we find ˇ ˇﬀ « ˇÿ ˇ ˇ ´1{2 ˇ E max \|Q\| ˇ vi Xi ˇ QPV ˇiPQ ˇ c ż 2a 2v 2 `C log N pV, ρ, ǫq dǫ ď π 0 d ˆ ˆ ˜ ˙˙¸ a \|Qn \| ln p\|Qn \|q ` ln ln ďC δn2 rn ˙ ˆ b a ln p\|Qn \|q ` ln p´ ln pm pθ1n q ´ m pθ0 qqq ďC ´ ¯ ´γ where we used ǫn “ O \|Qn \| with some γ ą 0. Now Markov’s inequality gives the claim. a 1 n (b) For Q P T it holds µ pQq ă δn µn pQn q and hence \|Q X Qn \| ď δn \|Q\| \|Qn \|. Consequently ˇﬀ ˇ « ˇ ˇ ÿ ˇ ´1{2 ˇ vi Xi ˇ E max1 \|Q\| ˇ ˇ QPT ˇiPQXQ n ˇ ˇﬀ « 1 ˇ ÿ ˇ a ˆ \|Qn \| ˙ 4 ˇ ´1{2 ˇ E max \|Q X Qn \| ď δn vi Xi ˇ ˇ ˇ ˇ rn QPQn \|rn iPQXQn ďC v̄ a δn ˆ \|Qn \| rn ˙ 14 a logpnq where we used (29). As the right-hand side converges to 0 by (42), this proves the claim. (c) This can now be deduced from (a) and (b) as follows. For all Q P T 1 it holds µn pQq ď δn µn pQn q and hence ˇ »ˇ ﬁ ˇ ˇ ˇ ´1 ÿ ˇ max γ pQq ´ max1 –ˇˇ\|Q\| 2 vi Xi ˇˇ ´ penv p\|Q\|qﬂ QPT 1 QPT ˇ ˇ iPQzQn ˇﬀ ˇ « ˇ ˇ ˇ ˇ ´1 ÿ vi Xi ˇ ď max1 µn pQq ` ˇ\|Q\| 2 ˇ QPT ˇ iPQXQn ˇ ˇ ˇ ˇ ˇ ´ 21 ÿ ˇ pbq n ďδn µ pQn q ` max1 ˇ\|Q\| vi Xi ˇ “ oP p1q (43) QPT ˇ ˇ iPQXQ n n where the last estimate follows from δn µ pQn q Œ 0. Furthermore, as V contains only scales ď δn´2 \|Qn \| we obtain that ˇ «ˇ ﬀ ˇ ˇ ˇ ´ 12 ÿ ˇ max ˇ\|Q\| Xi ˇ ´ penv p\|Q\|q QPV ˇ ˇ iPQ ˇ ﬀ «ˇ ˇ ˇ ` ´2 ˘ ˇ ´ 12 ÿ ˇ Xi ˇ ´ penv δn \|Qn \| ď max ˇ\|Q\| QPV ˇ ˇ iPQ ˙ ˆ b a ` ˘ paq n ln p\|Qn \|q ` ln p´ ln pm pθ1 q ´ m pθ0 qqq ´ penv δn´2 \|Qn \| “ OP “oP p1q, (44) where we used \|Qn \| “ o pnα q with α ą 0 sufficiently small. Consequently „  P max1 γ pQq ď q QPT 28 (43) “ pbq “ (44) “ “ ˇ »ˇ ﬁ ﬁ ˇ ˇ ˇ ´1 ÿ ˇ P – max1 –ˇˇ\|Q\| 2 vi Xi ˇˇ ´ penv p\|Q\|qﬂ ď q ﬂ ` op1q QPT ˇ ˇ iPQzQn ˇ ﬀ ﬀ «ˇ « ˇ ˇ ˇ ´ 12 ÿ ˇ Xi ˇ ´ penv p\|Q\|q ď q ` op1q P max1 ˇ\|Q\| ˇ QPT ˇ iPQ ˇ ﬀ ﬀ «ˇ « ˇ ˇ ˇ ´ 12 ÿ ˇ Xi ˇ ´ penv p\|Q\|q ď q ` op1q P max ˇ\|Q\| ˇ QPQn \|rn ˇ » iPQ P rMn pQn \|rn q ď qs ` op1q which yields the claim. Proof of Theorem 2.9(b). For the multiscale procedure we have to compute a lower bound for Pθn rTn pY, θ0 , Qn \|rn q ą qs. Similar to the Proof of Theorem 2.9(a) we obtain Pθn rTn pY, θ0 , Qn \|rn q ą qs „"  „  ěP max1 γ pQq ď q X tγ pQn q ą qu ` P max1 γ pQq ą q ` op1q. QPT QPT By Lemma 4.5(b) we furthermore get „  P max1 γ pQq ď q QPT ˇ « ﬀ ﬀ «ˇ ˇ ˇ 1 ÿ ˇ n ˇ vi Xi ˇ ´ penv p\|Q\|q ď q “P max1 ˇµ pQq ` a QPT ˇ ˇ \|Q\| iPQ ˇ ﬁ ﬁ » »ˇ ˇ ˇ ÿ ˇ ˇ n 1 vi Xi ˇˇ ´ penv p\|Q\|qﬂ ď q ﬂ ` op1q, “P – max1 –ˇˇµ pQq ` a QPT \|Q\| iPQzQn ˇ ˇ which shows by independence that „" *  P max1 γ pQq ď q X tγ pQn q ą qu QPT „  “P max1 γ pQq ď q P rγ pQn q ą qs ` op1q. QPT Now the proof can be concluded as the one of Theorem 2.9(a). Proof of Corollary 2.11. The procedures have asymptotic power 1 if and only if ¸ ˜ n n a ? m pθ q ´ m pθ q v pθ q 0 1 1 a , Ñ1 F q̄ ` ´2 log pan q, nd{2 an v pθ0 q v pθ0 q as n Ñ 8 with q̄ P tq ˚ , qu respectively. The straight-forward estimate ˙ ˆ ˙* " ˆ ` ˘ µ´x ´x ´ µ 2 ,Φ F x, µ, σ ě max Φ σ σ shows that this is the case if and only if x`µ Ñ ´8 σ or x´µ Ñ ´8. σ Inserting the values for x, µ and σ and noting that q ˚ , q are uniformly bounded by the p1 ´ αq´quantile of M pQ˚ , vq gives the claim. 29 5 Appendix Recall the packing number K pǫ, ρ, Wq from Remark 2.2. In this appendix we compute the packing numbers given in Example 2.3, which will be done by means of the covering number. The covering number N pǫ, ρ, Wq of a subset W Ă R˚ w.r.t. a metric ρ is given by the minimal number of balls of radius ǫ ą 0 needed to cover W (cf. van der Vaart and Wellner, 1996, Def. 2.2.3). It is immediately clear, that ¯ ´ǫ , ρ, W , N pǫ, ρ, Wq ď K pǫ, ρ, Wq ď N 2 ´ ˇ (¯ 1{2 and hence it sufficed to compute N pδuq , ρ˚ , R P R˚ ˇ \|R\| ď δ with ρ˚ as in (15) to show (16). In the following, we will use the notation from Example 2.3. Lemma 5.1. For any ǫ ą 0 there exists a constant C depending only on the dimension d and ǫ such that for all u, δ P p0, 1s it holds ´ ¯ K pδuq1{2 , ρ˚ , tS P S ˚ : \|S\| ď δu ď Cu´p2d`ǫq δ ´p2d´1`ǫq , i.e. (16) holds true with k1 “ C, k2 “ 2d ` ǫ and VS ˚ “ 2d ´ 1 ` ǫ. Proof. We approximate the hyper-rectangles in W “ tS P S ˚ : \|S\| ď δu by hyper-rectangles with ˇ (d vertices in the lattice Lm :“ mi ˇ i “ 0, . . . , m where m has to be specified later. The set of 1 all hyper-rectangles with such vertices and size ď δ will be denoted by Wm . For S P W denote śd by k1 , . . . , kd the edge lengths. Then j“1 kj ď δ and ki ď 1. It is immediately clear that there 1 exists an approximating hyper-rectangle S 1 P Wm such that ˇ ` ˚` ˘˘2 ˇ “ ˇS △ S 1 ˇ ρ S, S 1 ď 2 pk2 ¨ ... ¨ kd ` k1 ¨ k3 ¨ ... ¨ kd ` . . . ` k1 ¨ . . . ¨ kd´1 q ¨ ď d . m 1 2m (45) 1{2 d Hence, we obtain ρ˚ pS, S 1 q ď pδuq if we choose m :“ δu . Now we have to compute the 1 cardinality of Wd{pδuq . First note that the number of possible left bottom vertices is bounded from 1 above by md “ #Lm . If we denote the edge lengths of S 1 P Wm by l1 , . . . , ld , we can find integers śd ij d i1 , ..., id such that lj “ m and j“1 ij ď δm “: N . Therefore, we obtain ´ ¯ 1 N pδuq1{2 , ρ˚ , tS P S ˚ : \|S\| ď δu ď #Wm ď md ¨ #PN (46) ! ) ˇ śd with PN :“ pi1 , . . . , id q P Nd ˇ j“1 ij ď N . To compute #PN , we employ Minkowski’s theo- ˇ rem (cf. Cassels, 1997, Sec. III.2.2), which ensures that the Lebesgue volume ∆d pN q of px1 , . . . , xd q P r1, N sd ˇ x1 ¨ . . . is comparable with #PN up to a factor of 2d . It can readily be shown by induction that ∆d pN q “ 1 d´1 N plog N q . pd ´ 1q! Inserting this into (46), we obtain ´ ¯ N pδuq1{2 , ρ˚ , tS P S ˚ : \|S\| ď δu ďmd ˚ #PN ` ˘ ď2d md ∆d δmd ` ˘d´1 Àδm2d log δmd 30 ” ´ ¯ıd´1 “δ ´p2d´1q u´2d log dd δ ´pd´1q u´d d´1 d´1 Àδ ´p2d´1q plogp1{δqq u´2d plogp1{uqq where we used px ` yqd´1 ď cxd´1 y d´1 for x, y ě 1. This proves the claim. Lemma 5.2. There exists a constant C depending only on the dimension d such that for all u, δ P p0, 1s it holds ´ ¯ K pδuq1{2 , ρ˚ , tQ P Q˚ : \|Q\| ď δu ď Cδ ´1 u´pd`1q , i.e. (16) holds true with k1 “ C, k2 “ d ` 1 and VQ˚ “ 1. Proof. We proceed as in the Proof of Lemma 5.1. In contrast to hyper-rectangles, we obtain here instead of (45) the better estimate d´1 ˇ dδ d ` ˚` ˘˘2 ˇ ρ Q, Q1 “ ˇQ △ Q1 ˇ ď . m d . Furthermore, the cardinality as all edges have the same length, i.e. we can choose m :“ δ1{d u 1 of Wm is bounded by the number of lower left vertices times the number of possibilities for an adjacent vertex, which gives ´ ¯ 1 #Wm ď md ¨ δ 1{d m “ md`1 δ 1{d . Therefore we finally obtain ´ ¯ 1 N pδuq1{2 , ρ˚ , tS P S ˚ : \|S\| ď δu ď #Wd{pδ 1{d uq ˆ ˙d`1 d ď δ 1{d “ dd`1 u´pd`1q δ ´1 , δ 1{d u which proves the claim. Lemma 5.3. There exists a constant C depending only on the dimension d such that for all u, δ P p0, 1s it holds ´ ¯ K pδuq1{2 , ρ˚ , tH P H˚ : \|H\| ď δu ď Cδ ´2 u´2 , i.e. (16) holds true with k1 “ C, k2 “ 2 and VH˚ “ 2. ˇ ( ˇ i “ 1, ..., N, j “ 1, ..., m with numbers a1 , ..., aN P Sd´1 and α1 , ..., αm P Proof. Let W “ H N,m a ,α i j ” ? ı ? 0, d . Note that Ha,α “ H for α ą d by definition and Pythagoras’ theorem. It is convenient to choose α1 , ..., αm as equidistant, e.g. αi :“ i ´ 12 ? dn, m i “ 1, . . . , m. Furthermore we choose a1 , ..., aN as a maximal system of points in Sd´1 such that ?paj , ak q ě 1 ` 1 ˘ d´1 for all j ‰ k. This implies that m ˜ˆ ˙ 1 ¸ N ď 1 d´1 d´1 S aj S Ă m j“1 ˇ ( with the spherical cap Sa pθ0 q “ e P Sd´1 ˇ ?pa, eq ď θ0 . Note that θş0 \|Sa pθ0 q\| „ π0ş 0 d´2 psin tq psin tq 31 d´2 dt dt „ θ0d´1 ” ? ı for small values of θ0 . Now, for any given a P Sd´1 and α P 0, d , we can find 1 ď i ď N and 1 ď j ď m such that \|a ´ ai \| ď ˆ 1 m 1 ˙ d´1 , ? d \|α ´ αj \| ď . m Now we split ˇ ˇ ˘˘2 ` ˚` ď \|Ha,α △ Hai ,α \| ` ˇHai ,α △ Hai ,αj ˇ ρ Ha,α , Hai ,αj and since Hai ,α △ Hai ,αj is a d ´ 1-dimensional space of width ď 1 ` 1 ˘ d´1 of hyperpyramids with opening angle ď m , we obtain ? d m and Ha,α △ Hai ,α is a union ˘˘2 ` ˚` C ď ρ Ha,α , Hai ,αj m where C is some generic constant depending only on d. Hence if we choose m “ C ´1 δ ´1 u´1 , then for each H P tH P H˚ : \|H\| ď δu there exists H 1 P WN,m such that ρ˚ pS, S 1 q ď pδuq1{2 . Now we have to estimate N . By elementary geometry it follows that ˜ ˆ ˙ 1 ¸ ˜ˆ ˙ 1 ¸ N N ď ď 1 1 d´1 1 d´1 d´1 S aj ĂS Ă , S aj 2 m m j“1 j“1 and furthermore up to boundary points, the sets on the left-hand side are disjoint. Therefore we obtain for the volumes that ˇ ˇ ˜ ˆ ˙ 1 ¸ˇ ˜ˆ ˙ 1 ¸ˇ ˇ ˇ 1 1 d´1 ˇˇ ˇˇ d´1 ˇˇ 1 d´1 ˇˇ ˇ ˇ N ˇSaj ď N ˇSaj ˇď S ˇ ˇ ˇ ˇ ˇ 2 m m which implies N „ m. Consequently, #WN,m „ m2 which proves the claim. Acknowledgements The authors gratefully acknowledge financial support by the German Research Foundation DFG through subproject A04 of CRC 755. We also thank Katharina Proksch for many helpful comments on the proof of Theorem 2.5. References Adler, R. J. (2000). On excursion sets, tube formulas and maxima of random fields. Annals of Applied Probability, pages 1–74. Alm, S. E. (1998). Approximation and simulation of the distributions of scan statistics for Poisson processes in higher dimensions. Extremes, 1(1):111–126. Arias-Castro, E., Candès, E. J., and Durand, A. (2011). Detection of an anomalous cluster in a network. Ann. Statist., 39(1):278–304. Arias-Castro, E., Castro, R. M., Tánczos, E., and Wang, M. (2017). Distribution-free detection of structured anomalies: permutation and rank-based scans. Journal of the American Statistical Association, (to appear). Arias-Castro, E., Donoho, D. L., and Huo, X. (2005). Near-optimal detection of geometric objects by fast multiscale methods. IEEE Trans. Inform. Theory, 51(7):2402–2425. 32 Benjamini, Y. and Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. J. Roy. Statist. Soc. Ser. B, 57(1):289–300. Benjamini, Y. and Yekutieli, D. (2001). The control of the false discovery rate in multiple testing under dependency. Ann. Statist., 29(4):1165–1188. Billingsley, P. (2013). Convergence of Probability Measures. John Wiley & Sons. Brown, L. D. (1986). Fundamentals of statistical exponential families with applications in statistical decision theory. lecture notes. 9:x+283. Butucea, C. and Ingster, Y. I. (2013). Detection of a sparse submatrix of a high-dimensional noisy matrix. Bernoulli, 19(5B):2652–2688. Casella, G. and Berger, R. L. (2002). Statistical inference, volume 2. Duxbury Pacific Grove, CA. Cassels, J. W. S. (1997). An introduction to the geometry of numbers. Classics in Mathematics. Springer-Verlag, Berlin. Corrected reprint of the 1971 edition. Cheng, D. and Schwartzman, A. (2017). Multiple testing of local maxima for detection of peaks in random fields. Ann. Statist., 45(2):529–556. Chernozhukov, V., Chetverikov, D., and Kato, K. (2014). Gaussian approximation of suprema of empirical processes. Ann. Statist., 42(4):1564–1597. Dedecker, J. (1998). A central limit theorem for stationary random fields. Probab. Theory Related Fields, 110(3):397–426. Dedecker, J. (2001). Exponential inequalities and functional central limit theorems for random fields. ESAIM Probab. Statist., 5:77–104. Despres, C. J. (2014). The vapnik-chervonenkis dimension of norms on Rd . arXiv preprint arXiv:1412.6612. Devroye, L. and Lugosi, G. (2001). Combinatorial Methods in Density Estimation. Springer Series in Statistics. Springer-Verlag, New York. Dickhaus, T. (2014). Simultaneous Statistical Inference. Springer, Heidelberg. With applications in the life sciences. Dümbgen, L. and Spokoiny, V. G. (2001). Multiscale testing of qualitative hypotheses. Ann. Statist., 29(1):124–152. Dümbgen, L. and Walther, G. (2008). Multiscale inference about a density. 36(4):1758–1785. Ann. Statist., Fang, X. and Siegmund, D. (2016). Poisson approximation for two scan statistics with rates of convergence. Ann. Appl. Probab., 26(4):2384–2418. Frick, K., Munk, A., and Sieling, H. (2014). Multiscale change point inference. J. R. Stat. Soc. Ser. B. Stat. Methodol., 76(3):495–580. With 32 discussions by 47 authors and a rejoinder by the authors. Friedenberg, D. A. and Genovese, C. R. (2013). Straight to the source: detecting aggregate objects in astronomical images with proper error control. J. Amer. Statist. Assoc., 108(502):456–468. Haiman, G. and Preda, C. (2006). Estimation for the distribution of two-dimensional discrete scan statistics. Methodol. Comput. Appl. Probab., 8(3):373–381. Jiang, T. (2002). Maxima of partial sums indexed by geometrical structures. Ann. Probab., 30(4):1854–1892. 33 Jiang, Y., Qiu, Y., Minn, A. J., and Zhang, N. R. (2016). Assessing intratumor heterogeneity and tracking longitudinal and spatial clonal evolutionary history by next-generation sequencing. PNAS, 113(37):E5528–E5537. Kabluchko, Z. (2011). Extremes of the standardized Gaussian noise. Stochastic Process. Appl., 121(3):515–533. Kabluchko, Z. and Munk, A. (2009). Shao’s theorem on the maximum of standardized random walk increments for multidimensional arrays. ESAIM Probab. Stat., 13:409–416. Kazantsev, I., Lemahieu, I., Salov, G., and Denys, R. (2002). Statistical detection of defects in radiographic images in nondestructive testing. Signal Processing, 82(5):791 – 801. Komlós, J., Major, P., and Tusnády, G. (1976). An approximation of partial sums of independent RV’s, and the sample DF. II. Z. Wahrscheinlichkeitstheorie und Verw. Gebiete, 34(1):33–58. Kosorok, M. R. (2008). Introduction to Empirical Processes and Semiparametric Inference. Springer Series in Statistics. Springer, New York. Kou, J. (2017). Identifying the support of rectangular signals in gaussian noise. arXiv preprint arXiv:1703.06226. Kulldorff, M., Heffernan, R., Hartman, J., Assuno, R., and Mostashari, F. (2005). A spacetime permutation scan statistic for disease outbreak detection. PLOS Medicine, 2(3). Ledoux, M. and Talagrand, M. (1991). Probability in Banach Spaces, volume 23 of Ergebnisse der Mathematik und ihrer Grenzgebiete (3) [Results in Mathematics and Related Areas (3)]. Springer-Verlag, Berlin. Isoperimetry and processes. Lehmann, E. L. and Romano, J. P. (2005). Testing Statistical Hypotheses. Springer Texts in Statistics. Springer, New York, third edition. Li, H., Munk, A., and Sieling, H. (2016). FDR-control in multiscale change-point segmentation. Electron. J. Stat., 10(1):918–959. Marcus, M. B. and Rosen, J. (2006). Markov processes, Gaussian processes, and local times, volume 100 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge. Massart, P. (1989). Strong approximation for multivariate empirical and related processes, via KMT constructions. Ann. Probab., 17(1):266–291. Naus, J. I. and Wallenstein, S. (2004). Multiple window and cluster size scan procedures. Methodol. Comput. Appl. Probab., 6(4):389–400. Pozdnyakov, V., Glaz, J., Kulldorff, M., and Steele, J. M. (2005). A martingale approach to scan statistics. Ann. Inst. Statist. Math., 57(1):21–37. Proksch, K., Werner, F., and Munk, A. (2017+). Multiscale scanning in inverse problems. arXiv preprint arXiv:1611.04537, (to appear in Ann. Statist.). Rio, E. (1993). Strong approximation for set-indexed partial-sum processes, via KMT constructions. II. Ann. Probab., 21(3):1706–1727. Rivera, C. and Walther, G. (2013). Optimal detection of a jump in the intensity of a Poisson process or in a density with likelihood ratio statistics. Scand. J. Stat., 40(4):752–769. Schmidt-Hieber, J., Munk, A., and Dümbgen, L. (2013). Multiscale methods for shape constraints in deconvolution: confidence statements for qualitative features. Ann. Statist., 41(3):1299–1328. Sharpnack, J. and Arias-Castro, E. (2016). Exact asymptotics for the scan statistic and fast alternatives. Electron. J. Stat., 10(2):2641–2684. 34 Siegmund, D. and Venkatraman, E. S. (1995). Using the generalized likelihood ratio statistic for sequential detection of a change-point. Ann. Statist., 23(1):255–271. Siegmund, D. and Yakir, B. (2000). Tail probabilities for the null distribution of scanning statistics. Bernoulli, 6(2):191–213. Taylor, J. E. and Worsley, K. J. (2007). Detecting sparse signals in random fields, with an application to brain mapping. Journal of the American Statistical Association, 102(479):913– 928. Tu, I.-P. (2013). The maximum of a ratchet scanning process over a Poisson random field. Statist. Sinica, 23(4):1541–1551. van der Vaart, A. W. and Wellner, J. A. (1996). Weak Convergence and Empirical Processes. Springer Series in Statistics. Springer-Verlag, New York. With applications to statistics. Šidák, Z. (1967). Rectangular confidence regions for the means of multivariate normal distributions. J. Amer. Statist. Assoc., 62:626–633. Walther, G. (2010). Optimal and fast detection of spatial clusters with scan statistics. Ann. Statist., 38(2):1010–1033. Zhang, N. R., Yakir, B., Xia, L. C., and Siegmund, D. (2016). Scan statistics on Poisson random fields with applications in genomics. Ann. Appl. Stat., 10(2):726–755. 35	10math.ST
Space-Efficient Latent Contracts Michael Greenberg arXiv:1604.02474v4 [cs.PL] 6 Jun 2017 Pomona College michael@cs.pomona.edu Abstract. Standard higher-order contract monitoring breaks tail recursion and leads to space leaks that can change a program’s asymptotic complexity; space-efficiency restores tail recursion and bounds the amount of space used by contracts. Space-efficient contract monitoring for contracts enforcing simple type disciplines (a/k/a gradual typing) is well studied. Prior work establishes a space-efficient semantics for manifest contracts without dependency [11]; we adapt that work to a latent calculus with dependency. We guarantee space efficiency when no dependency is used; we cannot generally guarantee space efficiency when dependency is used, but instead offer a framework for making such programs space efficient on a case-by-case basis. 1 Introduction Findler and Felleisen [6] brought design-by-contract [17] into the higher-order world, allowing programmers to write pre- and post-conditions on functions to be checked at runtime. Pre- and post-conditions are easy in first-order languages, where it’s very clear who is to blame when a contract is violated: if the precondition fails, blame the caller; if the post-condition fails, blame the callee. In higher-order languages, however, it’s harder to tell who calls whom! Who should be to blame when a pre-condition on a higher-order function fails? For example, consider the following contract: (pred(λx :Int. x > 0) 7→ pred(λy:Int. y ≥ 0)) 7→ pred(λz :Int. z mod 2 = 0) This contract applies to a function (call it f , with type (Int→Int)→Int) that takes another function (call it g, with type Int→Int) as input. The contract says that g will only be called with positives and only return naturals; f must return an even number. If f returns an odd number, f is to blame; if g returns a negative number, then it, too is to blame. But what if g is called with a non-positive number, say, −1? Who is to blame then? Findler and Felleisen’s insight was that even in a higher-order setting, there are only two parties to blame. Here, g was given to f , so any bad values given to g here are due to some nefarious action on f ’s part—blame f ! That is, the higher-order case generalizes pre- and post-conditions so that the negative positions of a contract all blame the caller while the positive positions all blame the callee. Dependent contracts—where the codomain contract can refer to the function’s argument—are particularly useful. For example, the square root function, 2 Greenberg sqrt, satisfies the contract: x :pred(λy:Real. y ≥ 0) 7→ pred(λz :Real. abs (x − z ∗ z ) < ) That is, sqrt takes a non-negative real, x , and returns a non-negative real z that’s within of the square root of x . (The dependent variable x is bound in the codomain; the variable y is local to the domain predicate.) 1.1 Contracts leak space While implementations of contracts have proven quite successful (particularly so in Racket [8,19]), there is a problem: contracts leak space. Why? The default implementation of contracts works by wrapping a function in a function proxy. For example, to check that f = λx :Int. x + 1 satisfies the contract C = pred(λz :Int. z mod 2 = 0) 7→ pred(λz :Int. z mod 2 = 0), we monitor the function by wrapping it in a function proxy monl (C , f ). When this proxy is called with an input v , we first check that v satisfies C ’s domain contract (i.e., that v is even), then we run f on v to get some result v 0 , and then check that v 0 satisfies C ’s codomain contract (that the result is even). Here the contract will always fail blaming l : one of v and v 0 will always be odd. Contracts leak space in two ways. First, there is no bound on the number of function proxies that can appear on a given function. More grievously, contracts break tail recursion. To demonstrate the issue with with tail calls, we’ll use the simplest example of mutual recursion: detecting parity. let odd = λx :Int. if (x = 0) false (even (x − 1)) and even = λx :Int. if (x = 0) true (odd (x − 1)) Functional programmers will expect this program to run in constant space, because it is tail recursive. Adding a contract breaks the tail recursion. If we add a contract to odd and call odd 5, what contract checks accumulate (Fig. 1)?1 Notice how the checks accumulate in the codomain? Even though the mutually recursive calls to even and odd are syntactically tail calls, we can’t bound the number of codomain checks that occur. That is, we can’t bound the size of the stack, and tail recursion is broken! Even though there’s only one function proxy on odd, our contracts create a space leak. 1.2 Overview and contributions Space efficiency for gradual types [24] (a/k/a contracts constrained to type tests) is well studied [14,15,25,9,23]; Greenberg [11] developed a space-efficient semantics for general, non-dependent contracts. He used a manifest calculus, conflating contracts and types; however, contracts are typically implemented in latent calculi, where contracts are distinct from whatever types may exist. Greenberg “believe[s] it would be easy to design a latent version of eidetic λH , following the translations in Greenberg, Pierce, and Weirich (GPW)” [12]; in this paper, 1 Readers may observe that the contract betrays a deeper knowledge of numbers than the functions themselves. We offer this example as minimal, not naturally occurring. Space-Efficient Latent Contracts 3 = monlodd (x :pred(λx :Int. x ≥ 0) 7→ pred(λb:Bool. b or (x mod 2 = 0)), λx :Int. if (x = 0) false (even (x − 1))) and even = λx :Int. if (x = 0) true (odd (x − 1)) let odd odd 5 −→∗C monlodd (pred[x 7→5] (. . . ), even 4) −→∗C monlodd (pred[x 7→5] (. . . ), monlodd (pred[x 7→3] (. . . ), odd monlodd (pred(λx :Int. x ≥ 0), 3))) ∗ −→C monlodd (pred[x 7→5] (. . . ), monlodd (pred[x 7→3] (. . . ), even 2)) −→∗C monlodd (pred[x 7→5] (. . . ), monlodd (pred[x 7→3] (. . . ), monlodd (pred[x 7→1] (. . . ), odd monlodd (pred(λx :Int. x ≥ 0), 1)))) ∗ −→C monlodd (pred[x 7→5] (. . . ), monlodd (pred[x 7→3] (. . . ), monlodd (pred[x 7→1] (. . . ), even 0))) Fig. 1. Contracts break tail recursion we show that belief to be well founded by giving a space-efficient semantics for a (dependent!) variant of contract PCF (CPCF) [3,4]. The rest of this paper discusses a formulation of contracts that enjoys sound space efficiency; that is, where we slightly change the implementation of contracts so that (a) programs are observationally equivalent to the standard semantics, but (b) contracts consume a bounded amount of space. In this paper, we’ve omitted some of the more detailed examples and motivation—we refer curious readers to Greenberg [11], though we intend the paper to be self-contained. We follow Greenberg’s general structure, defining two forms of dependent CPCF: CPCFC is the classic semantics; CPCFE follows the space-efficient eidetic semantics. We are able to prove space efficiency without dependency, bounding the amount of space consumed by contracts; we are unable to prove space efficiency in general with dependency, but instead offer a framework that allows for dependent contracts to be made space efficient. We offer two primary contributions: adapting Greenberg’s work to a latent calculus and extending the possibility of space efficiency to dependent contracts. There are some other, smaller, contributions as well. First, adding in nontermination moves beyond Greenberg’s strongly normalizing calculi, showing that the POPL 2015 paper’s result isn’t an artifact of strong normalization (where we can, in theory, bound the size of the any term’s evaluation in advance, not just contracts). Second, the simpler type system here makes it clear which type system invariants are necessary for space-efficiency and which are bookkeeping for proving that the more complicated manifest type system is sound. Third, by separating contracts and types, we can give tighter space bounds—the types function from Greenberg collects types that are never used in a contract, while we collect exactly contracts. Finally, we explore how space efficiency can be attained in dependent contracts. While we can’t give a guarantee for dependent 4 Greenberg Types Terms B T e ::= ::= ::= op k w v C c r ::= ::= ::= ::= ::= ::= ::= Bool \| Int \| . . . B \| T1 →T2 x \| k \| e1 op e2 \| e1 e2 \| λx :T . e \| µ(x :T ). e \| if e1 e2 e3 \| errl \| monl (C , e) \| mon(c, e) add1 \| sub1 \| . . . true \| false \| 0 \| 1 \| . . . v \| errl k \| λx :T . e \| monl (x :C1 7→ C2 , v ) \| mon(x :c1 7→ c2 , λx :T . e) predσ (e) \| x :C1 7→ C2 r \| x :c1 7→ c2 nil \| predlσ (e); r Fig. 2. Syntax of classic and space-efficient CPCF contracts, we show that it’s possible to achieve and discuss different ways to do so. 2 Classic and space-efficient Contract PCF We present classic and space-efficient CPCF as separate calculi sharing syntax and some typing rules (Fig. 2 and Fig. 3), and a single, parameterized operational semantics with some rules held completely in common (Fig. 4) and others specialized to each system (Fig. 5). The formal presentation is modal, with two modes: C for classic and E for space-efficient. While much is shared between the two modes—types, T ; the core syntax of expressions, e; most of the typing rules—we use colors to highlight parts that belong to only one system. Classic CPCF is typeset in salmon while space-efficient CPCF is in periwinkle. 2.1 Contract PCF (CPCF) Plain CPCF is an extension of Plotkin’s 1977 PCF [18], developed first by Dimoulas and Felleisen [3,4] (our syntax is in Fig. 2). It is a simply typed language with recursion. The typing rules are straightforward (Fig. 3). The operational semantics for the generic fragment also uses conventional rules (Fig. 4). Dimoulas and Felleisen use evaluation contexts to offer a concise description of their system; we write out our relation in full, giving congruence rules (EL, ER, EIf) and error propagating rules (ERaise) explicitly—we will need to restrict congruence for casts, and our methods are more transparent written with explicit congruence rules than using the subtly nested evaluation contexts of Herman et al. [14,15], which are error prone [10]. Contracts are CPCF’s distinguishing feature. Contracts, C , are installed via monitors, written monl (C , e); such a monitor says “ensure that e satisfies the contract C ; if not, the blame lies with label l ”. Monitors only apply to appropriate types (TMon). There are two kinds of contracts in CPCF: predicate contracts over base type, written predσ (e), and function contracts, written x :C1 7→ C2 . Space-Efficient Latent Contracts 5 Predicate contracts predσ (e) have two parts: a predicate on base types, e, which identifies which values satisfy the contract; and a closing substitution σ which keeps track of values substituted into the contract. For example, if ι is the identity substitution mapping variables to themselves: – predι (λx :Int. x > 0) identifies the positives; – predι (λx :Int. x > y) identifies numbers greater than an unspecified number y; and, – pred[y7→47] (λx :Int. x > y) identifies numbers greater than 47. When the closing substitution σ is the identity mapping ι, we write pred(e) instead of predι (e). In CPCFC , closing substitutions will map each variable to either (a) itself or (b) a value. Substitution into contracts is a non-issue without dependency: each contract is just closed. Having introduced dependency, we use explicit closing substitutions rather than direct substitution for three reasons: first, it simplifies our space efficiency proof for simple contracts (Sec. 4.1); second, explicitness lets us distinguish the contract pred[x 7→0] (λx :Int. x = 0) from pred[x 7→0] (λx :Int. 0 = 0); third, it emphasizes that contracts are just another form of closure. Predicates are solely over base types, not functions. Function contracts x :C1 7→ C2 are satisfied by functions satisfying their parts: functions whose inputs all satisfy C1 and whose outputs all satisfy C2 . Function contracts are dependent: the codomain contract C2 can refer back to the input to the function. For example, the contract x :pred(λz :Int. z > 0) 7→ pred(λy:Int. y > x ) is satisfied by increasing functions on the positives. Note that x is bound in the codomain, but z is not.2 When function contracts aren’t dependent, we omit the binder at the front, e.g., pred(λx :Int. x > 0) 7→ pred(λx :Int. x > 0) means operators on positives. We check that contracts are satisfied at runtime. We use explicit, delayed substitutions to keep track of which values are substituted into predicate contracts. To help with our proof of space efficiency, we don’t track variables that don’t appear in the predicate: ( predσ (e)[v /x ] = predσ[x 7→v ] (e) x ∈ fv(σ(e)) predσ (e) otherwise Alpha equivalence allows us to give fresh names to variables in the domain of σ by consistently renaming those variables inside of the predicate e. Only holding on to substitutions that close up free variables in e is a way of modeling closures. A dependent predicate closes over some finite number of variables; a compiled representation would generate a closure with a corresponding number of slots in the closing environment. Restricting substitutions to exactly those variables appearing free in the predicate serves another purpose: we can easily recover space-efficiency bounds for programs without dependent contracts (Sec. 4.1). 6 Greenberg Γ`e:T Typing rules x :T ∈ Γ Γ`x :T TVar Γ ` k : ty(k ) Γ, x :T1 ` e12 : T2 Γ ` λx :T1 . e12 : T1 →T2 ty(op) = T1 →T2 →T Γ ` e1 : T 1 Γ ` e 2 : T 2 Γ ` e1 op e2 : T TConst TAbs Γ, x :T ` e : T Γ ` µ(x :T ). e : T TOp Γ ` e1 : T1 →T2 Γ ` e2 : T1 Γ ` e1 e2 : T2 Γ ` e1 : Bool Γ ` e2 : T Γ ` e3 : T Γ ` if e1 e2 e3 : T Γ`e:T Γ`C :T Γ ` monl (C , e) : T TMon 0 0 Γ, Γ ` e : B →Bool Γ ` σ Γ ` predσ (e) : B TPred TCNil Γ ` c1 : T1 Γ, x :T1 ` c2 : T2 Γ ` x :c1 7→ c2 : T1 →T2 ∅`ι TId TMonC Γ ` C1 : T1 Γ, x :T1 ` C2 : T2 Γ ` x :C1 7→ C2 : T1 →T2 TFun Γ`r :B Γ ` predlσ (e); r : B Closing substitutions TApp TIf Γ`e:T Γ`c:T Γ ` mon(c, e) : T Γ ` predσ (e) : B Γ ` nil : B TRec Γ`c:T Γ`C :T Contract typing TBlame Γ ` errl : T TCPred TCFun Γ`σ Γ ` σ x :T ` v : T Γ, x :T ` σ[x 7→ v ] Fig. 3. Typing rules of classic and space-efficient CPCF TMap Space-Efficient Latent Contracts e1 −→m e2 Operational semantics (λx :T . e) v −→m e[v /x ] EBeta k1 op k2 −→m [[op]] (k1 , k2 ) µ(x :T ). e −→m e[µ(x :T ). e/x ] EIfTrue if true e2 e3 −→m e2 e1 −→m e10 e1 e2 −→m e10 e2 EOpL e1 −→m e10 e3 −→m if e10 e2 e3 errl e2 −→m errl errl op e2 −→m errl EIf EAppLRaise EOpLRaise EDelta EFix EIfFalse if false e2 e3 −→m e3 e2 −→m e20 v1 e2 −→m v1 e20 EAppL e1 −→m e10 e1 op e2 −→m e10 op e2 if e1 e2 7 EAppR e2 −→m e20 v1 op e2 −→m v1 op e20 if errl e2 e3 −→m errl v1 errl −→m errl v1 op errl −→m errl EOpR EIfRaise EAppRRaise EOpRRaise Fig. 4. Shared operational semantics of CPCF 2.2 Classic Contract PCF (CPCFC ) Classic CPCF gives a straightforward semantics to contracts (Figs. 4 and 5), largely following the seminal work by Findler and Felleisen [6]. To check a predicate contract, we simply test it (EMonPred), returning either the value or an appropriately labeled error. Function contracts are deferred: monl (x :C1 7→ C2 , v ) is a value, called a function proxy. When a function proxy is applied, it unwraps the proxy, monitoring the argument with the domain contract, running the function, and then monitoring the return value with the codomain contract (EMonApp). Our semantics may seem to be lax, where no monitor is applied to dependent uses of the argument in the codomain monitor [12]. In fact, it is agnostic: we could be picky by requiring that function contract monitors monl (x :C1 7→ C2 , e) have 2 Concrete syntax for such predicates can be written much more nicely, but we ignore such concerns here. 8 Greenberg EMonPred monl (predσ (e1 ), v2 ) −→C if (σ(e1 ) v2 ) v2 errl monl (x :C1 7→ C2 , v1 ) v2 −→C monl (C2 [v2 /x ], v1 monl (C1 , v2 )) e −→C e 0 −→C monl (C , e 0 ) monl (C , e) EMon monl (C , errl 0 ) −→C errl 0 monl (C , e) −→E mon(labell (C ), e) mon(nil, v1 ) −→E v1 EMonCNil mon(x :c1 7→ c2 , v1 ) v2 −→E mon(c2 [v2 /x ], v1 mon(c1 , v2 )) EMonC EMonRaise EMonLabel mon(predlσ (e); r , v1 ) −→E if (σ(e) v1 ) mon(r , v1 ) errl e 6= mon(c 0 , e 00 ) e −→E e 0 mon(c, e) −→E mon(c, e 0 ) EMonApp EMonCPred EMonCApp mon(c, errl ) −→E errl mon(c2 , mon(c1 , e)) −→E mon(join(c1 , c2 ), e) EMonCRaise EMonCJoin Fig. 5. Operational semantics of classic and space-efficient CPCF the substitution [x 7→ monl (C1 , x )] throughout C2 ; we could be indy by having 0 [x 7→ monl (C1 , x )] throughout C2 [4]. We default to a lax rule to make our proof of soundness easier, but we’ll have as a corollary that classic and space-efficient semantics yield the same result regardless of what the closing substitutions do in the codomain (Sec. 3). Standard congruence rules allow for evaluation inside of monitors (EMon) and the propagation of errors (EMonRaise). Metatheory We prove CPCFC ’s type system sound with a minimum fuss, using the usual syntactic methods. The only subtlety is that we must be careful when proving the substitution property, since we have slightly changed the definition. In this section, we only consider the typing rules for CPCFC , i.e., those typeset on white and salmon. Lemma 1 (Weakening). – If Γ ` e : T and x 6∈ fv(e) then Γ, x :T 0 ` e : T . – If Γ ` C : T and x 6∈ fv(C ) then Γ, x :T 0 ` C : T . Space-Efficient Latent Contracts 9 Proof. By mutual induction on the terms and contracts. Lemma 2 (Substitution). If ∅ ` v : T1 , then – Γ1 , x :T1 , Γ2 ` e : T2 implies Γ1 , Γ2 ` e[v /x ] : T2 , and – Γ1 , x :T1 , Γ2 ` C : T2 implies Γ1 , Γ2 ` C [v /x ] : T2 . Proof. By mutual induction the terms and contracts. (e = y) By case analysis on x = y, using weakening (Lemma 1) as appropriate. (e = k ) Immediate. (e = e1 op e2 ) By the IHs. (e = e1 e2 ) By the IHs. (e = λy:T . e1 ) By case analysis on x = y, narrowing when the two are equal. (e = µ(y:T ). e1 ) By case analysis on x = y, narrowing when the two are equal. (e = if e1 e2 e3 ) By the IHs. (e = errl ) Immediate. (e = monl (C , e)) By the IHs. (C = predσ (e)) We have Γ1 , x :T1 , Γ2 ` predσ (e) : T2 and we must show that Γ1 , Γ2 ` predσ (e)[v /x ] : T2 . By inversion, we know that: T2 = B for some base type B ; Γ1 , x :T1 , Γ2 , Γ0 ` e : B →Bool; and Γ0 ` σ. There are two cases. In both cases, we use TPred. (x ∈ fv(e)) The substitution is actually stored in σ. We find Γ1 , Γ2 , Γ0 , x :T1 ` e : B by the IH on e and Γ0 , x :T1 ` σ[x 7→ v ] by TMap and our assumption. (x 6∈ fv(e)) The substitution is ignored, and e[v /x ] = e. We find Γ1 , Γ2 , Γ0 ` e : B by the IH and Γ0 ` σ by assumption. (C = x :C1 7→ C2 ) By the IHs. As a corollary, closing substitutions σ close up exactly their context Γ0 . Lemma 3 (Closing substitutions). If Γ0 ` σ and Γ, Γ0 ` e : T then Γ ` σ(e) : T . Proof. By induction on Γ0 , using substitution (Lemma 2). Lemma 4 (Progress). If ∅ ` e : T then either (a) e is some value, v ; (b) e is some error, errl ; or, (c) e −→C e 0 . Proof. By induction on the typing derivation. ( TVar) Contradictory—variables aren’t well typed in the empty context. ( TConst) k is a value. ( TBlame) errl is an error. ( TAbs) Function abstractions are values. ( TRec) Takes a step by EFix. ( TOp) By the IH, e1 is a value, error, or steps. In the latter two cases, we step by EOpLRaise or EOpL, respectively. By the IH, e2 is a value, error, or steps. In the latter two cases we step by EOpRRaise or EOpR. If both are values, then we step by EDelta. 10 Greenberg ( TApp) By the IH, e1 is a value, error, or steps. In the latter two cases, we step by EAppLRaise or EAppL, respectively. By the IH, e2 is a value, error, or steps. In the latter two cases we step by EAppRRaise or EAppR. If both are values, we know that e1 is either a function or a function proxy; we step by either EBeta or EMonApp, depending on the shape of e1 . ( TIf) By the IH, e1 is a value, error, or steps. In the latter two cases, we step by EIfRaise or EIf, respectively./ If e1 is a value, it must be either true or false; we step by EIfTrue or EIfFalse, depending on the shape of e1 . ( TMon) By the IH, e is either a value, error, or steps. In the latter two cases, we step by EMonRaise or EMon, respectively. If e is a value, we step by EMonPred or EMonApp, depending on the shape of the contract C . Lemma 5 (Preservation). If ∅ ` e : T and e −→C e 0 then ∅ ` e 0 : T . Proof. By induction on the typing derivation, with cases on the step taken. ( TVar) Contradictory—variables aren’t well typed in the empty context. ( TConst) Contradictory—constants are normal forms. ( TBlame) Contradictory—errors are normal forms. ( TAbs) Contradictory—functions are normal forms. ( TRec) We must have stepped by EFix; by substitution (Lemma 2) and TFix. ( TOp) If we stepped by an EOpRaise rule, we are done by TBlame. If we stepped by a congruence rule, EOp, we are done by the IH and TOp. If we stepped by EDelta, by the assumption that built-in operations have sound denotations. ( TApp) If we stepped by an EAppRaise rule, we are done by TBlame.If we stepped by a congruence rule, EApp, we are done by the IH and TApp. If we stepped by EBeta, by substitution. (Lemma 2) If we stepped by EMonApp, then: the domain is well typed by TMon; we can apply the proxied function by TApp; the codomain check is well formed by substitution (Lemma 2); and the entire resulting term is well typed by TMon again. ( TIf) If we stepped by EIfRaise, by TBlame. If we stepped by EIf, by the IH and TIf. If we stepped by EIfTrue or EIfFalse, by assumption. ( TMon) If we stepped by EMonRaise, the resulting error is well typed by TBlame. If we stepped by EMon, by the IH and TMon. If we stepped by EMonPred, then: we have ∅ ` σ(e) : B →Bool by closing substitution (Lemma 3); we can type the condition by TApp; the then branch is well typed by assumption; the false branch is well typed by TBlame; and the whole lot is well typed by TIf. 2.3 Space-efficient Contract PCF (CPCFE ) How can we recover tail calls in CPCF? CPCFC will happily wrap arbitrarily many function proxies around a value, and there’s no bound on the number of Space-Efficient Latent Contracts 11 labell (predσ (e1 )) = predlσ (e1 ) labell (x :C1 7→ C2 ) = x :labell (C1 ) 7→ labell (C2 ) join(nil, r2 ) = r2 join(predlσ (e); r1 , r2 ) = predlσ (e); drop(join(r1 , r2 ), predσ (e)) join(x :c11 7→ c12 , x :c21 7→ c22 ) = x :join(c21 , c11 ) 7→ join(wrap(c12 , x , c22 ), c22 ) drop(nil, predσ (e)) = nil drop(predlσ2 (e2 ); r , predσ1 (e1 )) = ( drop(r , predσ2 (e2 )) predlσ2 (e2 ); drop(r , predσ1 (e1 )) wrap(predlσ (e), x , c) = wrap(nil, x , c) = wrap(predlσ (e); r , x , c) = wrap(y:c1 7→ c2 , x , c) = predσ1 (e1 ) ⊃ predσ2 (e2 ) predσ1 (e1 ) 6⊃ predσ2 (e2 )  l  x 6∈ fv(σ(e)) predσ (e) predlσ[x 7→mon(join(c 0 ,c),e)] (e) σ(x ) = mon(c 0 , e)   predlσ[x 7→mon(c,σ(x ))] (e) otherwise nil wrap(predlσ (e), x , c); wrap(r , x , c) y:wrap(c1 , x , c) 7→ wrap(c2 , x , c) Fig. 6. Contract labeling and predicate stack management codomain contract checks that can accumulate. The key idea is joining contracts. We’ll make two changes to the language: we’ll bound function proxies so each function has at most one, and we’ll bound stacks to avoid redundant checking. We ultimately show that contracts without dependency use constant space, but that the story for dependent functions is more complex (Sec. 4). Fortuitously, our notion of join solves both of our problems, working identically for both simple and dependent contracts. To ensure a function value can have only one proxy, we change the semantics of monitoring: when monitoring a proxied value, we join the new monitor and the old one. To bound the size of stacks contract checks, we join pending contracts to avoid redundancy. The join operation works on labeled contracts, which (a) move the label from the monitor into the contract and (b) allow us to keep track of many predicates at once (Fig. 6). Concretely, labeled contracts use the metavariable c (as opposed to C ), comprising function contracts as usual (x :c1 7→ c2 ) and predicate stacks, r (Fig. 2). A predicate stack r is a list of labeled predicates predl (e), where nil is the empty stack. The join operation takes two labeled contracts and combines them, eliminating redundant contracts as it goes. To join a new and an old predicate stack, we keep new contracts and eliminate redundant old ones; only more “recent” contracts are kept. Joining functions works contravariantly, being careful to maintain correct substitution behavior using wrap. 12 Greenberg Finally, we establish what we mean by “redundant” using predicate implication: when does one contract imply another? Definition 6 (Predicate implication). Let predσ1 (e1 ) ⊃ predσ2 (e2 ) be a relation on predicates such that: (Reflexivity) If ∅ ` predσ (e) : B then predσ (e) ⊃ predσ (e). (Transitivity) If predσ1 (e1 ) ⊃ predσ2 (e2 ) and predσ2 (e2 ) ⊃ predσ3 (e3 ), then predσ1 (e1 ) ⊃ predσ3 (e3 ). (Substitutivity) When Γi 1 , x :T 0 , Γi 2 ` predσi (ei ) : T and ∅ ` v : T 0 , if predσ1 (e1 ) ⊃ predσ2 (e2 ) then predσ1 (e1 )[v /x ] ⊃ predσ2 (e2 )[v /x ]. (Adequacy) If ∅ ` predσi (ei ) : T and predσ1 (e1 ) ⊃ predσ2 (e2 ) then ∀k ∈ KB . σ1 (e1 ) k −→m true implies σ2 (e2 ) k −→m true. (Decidability) For all ∅ ` predσ1 (e1 ) : B and ∅ ` predσ2 (e2 ) : B , it is decidable whether predσ1 (e1 ) ⊃ predσ2 (e2 ) or predσ1 (e1 ) 6⊃ predσ2 (e2 ). The entire development of space-efficiency is parameterized over this implication relation, ⊃, characterizing when one first-order contract subsumes another. We write 6⊃ for the negation of ⊃. The ⊃ relation is a total pre-order (a/k/a a preference relation)—it would be a total order, but it may not necessarily enjoy anti-symmetry. For example, we could have predι (λx :Int. x ≥ 0) ⊃ predι (λx :Int. x + 1 > 0) and vice versa, even though the two predicates aren’t equal. You can also view ⊃ as a total order up-to contextual equivalence. There is at least one workable implication relation: syntactic equality. We say predσ1 (e1 ) ⊃ predσ2 (e2 ) iff e1 = e2 and σ1 = σ2 . Since we’ve been careful to store only those values that are actually referenced in the closure of the predicate, the steps to determine these equalities are finite and computable at runtime. For example, suppose we wish to show that pred[y7→47] (λx :Int. x > y) ⊃ pred[y7→47] (λx :Int. x > y). The code part—the predicate λx :Int. x > y—is the same; an implementation might observe that the function pointers are equal. The environment has only one slot, for y, with the value 47 in it; an implementation might compare the two environments slot-by-slot. We could have given an operational semantics which behaves more like an implementation, explicitly generating conditionals and merge operations in the terms themselves, but we believe our slightly more abstract presentation is more digestible. Substitution in the codomain: lax, picky, and indy We extend Greenberg’s notion of join to account for dependency with a new function, wrap. Greenberg, Pierce, and Weirich identified two variants of latent contracts in the literature, differing in their treatment of the dependent substitution of arguments in the codomain: picky, where we monitor the value substituted in the codomain with the domain contract; and lax, where the actual parameter value substituted into the codomain is unmonitored [12]. There is a third variant, indy, which applies a monitor to the argument value but uses a different blame label [4]. These different models of substitution all exhibit different behavior for abusive contracts, where the codomain contract violates the domain contract. Space-Efficient Latent Contracts 13 There is another source of substitutions in the codomain: multiple function proxies. How do the monitors unfold when we have two function proxies? In the classic lax semantics, we find (leaving the domain check unevaluated): mon(x :c11 7→ c12 , mon(x :c21 7→ c22 , f )) v −→C mon(c12 [v /x ], mon(x :c21 7→ c22 , f ) mon(c11 , v )) −→C mon(c12 [v /x ], mon(c22 [mon(c11 , v )/x ], f mon(c21 , mon(c11 , v )))) Even though we’re using the lax semantics, we substitute contracts into the codomain. For the space-efficient semantics to be sound, it must behave exactly like the classic semantics: no matter what joins happen, CPCFE must replicate the contract substitutions done in CPCFC . We can construct an abusive contract in CPCFC —even though it has lax semantics—by having the inner function proxy abuse the outer one (Fig. 7). Why was blame raised? Because c2 ’s codomain contract abused c1 ’s domain contract. Even though CPCFC has a lax semantics, wrapping multiple function proxies leads to monitoring domains from one contract in the codomain of another—a situation ripe for abuse. Space-efficiency means joining contracts, so how can we emulate this classicsemantics substitution behavior? We use the wrap function, forcing a substitution when two function contracts are joined. By keeping track of these substitutions at every join, any joins that happen in the future will be working on contracts which already have appropriate substitutions. CPCFE uses labeled contracts (Fig. 2); substitution for labeled predicate contracts is explicit and delayed, as for ordinary contracts: ( predlσ[x 7→v ] (e) x ∈ fv(σ(e)) l predσ (e)[v /x ] = predlσ (e) otherwise nil[v /x ] = nil (predlσ (e); r )[v /x ] = predlσ (e)[v /x ]; r [v /x ] We do not do any joining when a substitution occurs (but see Sec. 6). In CPCFE , closing substitutions map each variable to (a) itself ([x 7→ x ]), (b) a monitor on itself ([x 7→ mon(c, x )]), or (c) a value. We add an evaluation rule taking ordinary contract monitors monl (C , e) to labeled-contract monitors mon(c, e) by means of the labeling function label (EMonLabel). Space-efficiency comes by restricting congruence to only apply when there are abutting monitors (cf. EMonC here in CPCFE to EMon in CPCFC ). When two monitors collide, we join them (EMonCJoin). Checking function contracts is as usual (EMonCApp is the same as EMonApp, only the latter works over labeled contracts); checking predicate stacks proceeds straightforwardly predicate-bypredicate (EMonCNil and EMonCPred). Metatheory We prove CPCFE ’s type system sound; the proof follows that for CPCFC , though there are more evaluation rules to consider here. We only consider the typing rules for CPCFE , i.e., those typeset on white and periwinkle. 14 Greenberg C1 = f :(pred(λx :Int. x > 0) 7→ pred(λx :Int. x > 0)) 7→ pred(λx :Int. x > 0) C2 = f :(pred(λx :Int. true) 7→ pred(λx :Int. true)) 7→ pred(λx :Int. f 0 = 0) Referring to the domains as Ci 1 and codomains as Ci 2 −→C −→C −→C −→C −→∗C −→C −→∗C −→C −→∗C −→∗C −→C −→C −→C −→C −→C −→C −→C −→∗C monl1 (C1 , monl2 (C2 , λf :(Int→Int). f 5)) (λx :Int. x ) monl1 (C12 [(λx :Int. x )/f ], monl2 (C2 , λf :(Int→Int). f 5) monl1 (C11 , (λx :Int. x ))) monl1 (C12 [(λx :Int. x )/f ], monl2 (C22 [monl1 (C11 , λx :Int. x )/f ], (λf :(Int→Int). f 5) monl2 (C21 , monl1 (C11 , (λx :Int. x ))))) monl1 (C12 [(λx :Int. x )/f ], monl2 (C22 [monl1 (C11 , λx :Int. x )/f ], monl2 (C21 , monl1 (C11 , (λx :Int. x ))) 5)) monl1 (C12 [(λx :Int. x )/f ], monl2 (C22 [monl1 (C11 , λx :Int. x )/f ], monl2 (pred(λx :Int. true), monl1 (C11 , (λx :Int. x )) monl2 (pred(λx :Int. true), 5)))) monl1 (C12 [(λx :Int. x )/f ], monl2 (C22 [monl1 (C11 , λx :Int. x )/f ], monl2 (pred(λx :Int. true), monl1 (C11 , (λx :Int. x )) 5))) monl1 (C12 [(λx :Int. x )/f ], monl2 (C22 [monl1 (C11 , λx :Int. x )/f ], monl2 (pred(λx :Int. true), monl1 (pred(λx :Int. x > 0), (λx :Int. x ) monl1 (pred(λx :Int. x > 0), 5))))) monl1 (C12 [(λx :Int. x )/f ], monl2 (C22 [monl1 (C11 , λx :Int. x )/f ], monl2 (pred(λx :Int. true), monl1 (pred(λx :Int. x > 0), (λx :Int. x ) 5)))) monl1 (C12 [(λx :Int. x )/f ], monl2 (C22 [monl1 (C11 , λx :Int. x )/f ], monl2 (pred(λx :Int. true), monl1 (pred(λx :Int. x > 0), 5)))) monl1 (C12 [(λx :Int. x )/f ], monl2 (C22 [monl1 (C11 , λx :Int. x )/f ], monl2 (pred(λx :Int. true), 5))) monl1 (C12 [(λx :Int. x )/f ], monl2 (C22 [monl1 (C11 , λx :Int. x )/f ], 5)) monl1 (C12 [(λx :Int. x )/f ], if ((λx :Int. monl1 (C11 , λx :Int. x ) 0 = 0) 5) 5 errl2 ) monl1 (C12 [(λx :Int. x )/f ], if (monl1 (C11 , λx :Int. x ) 0 = 0) 5 errl2 ) monl1 (C12 [(λx :Int. x )/f ], if (monl1 (pred(λx :Int. x > 0), (λx :Int. x ) monl1 (pred(λx :Int. x > 0), 0)) = 0) l2 5 err ) monl1 (C12 [(λx :Int. x )/f ], if (monl1 (pred(λx :Int. x > 0), (λx :Int. x ) (if ((λx :Int. x > 0) 0) 0 errl2 )) = 0) l2 5 err ) monl1 (C12 [(λx :Int. x )/f ], if (monl1 (pred(λx :Int. x > 0), (λx :Int. x ) (if (0 > 0) 0 errl2 )) = 0) l2 5 err ) monl1 (C12 [(λx :Int. x )/f ], if (monl1 (pred(λx :Int. x > 0), (λx :Int. x ) (if false 0 errl2 )) = 0) 5 errl2 ) monl1 (C12 [(λx :Int. x )/f ], if (monl1 (pred(λx :Int. x > 0), (λx :Int. x ) errl2 ) = 0) 5 errl2 ) errl2 Fig. 7. Abusive function proxies in CPCFC Space-Efficient Latent Contracts 15 Lemma 7 (Weakening). – If Γ ` e : T and x 6∈ fv(e) then Γ, x :T 0 ` e : T . – If Γ ` C : T and x 6∈ fv(C ) then Γ, x :T 0 ` C : T . – If Γ ` c : T and x 6∈ fv(c) then Γ, x :T 0 ` c : T . Proof. By mutual induction on the terms and contracts. Lemma 8 (Substitution). If ∅ ` v : T1 , then – Γ1 , x :T1 , Γ2 ` e : T2 implies Γ1 , Γ2 ` e[v /x ] : T2 , and – Γ1 , x :T1 , Γ2 ` C : T2 implies Γ1 , Γ2 ` C [v /x ] : T2 . – Γ1 , x :T1 , Γ2 ` c : T2 implies Γ1 , Γ2 ` c[v /x ] : T2 . Proof. By mutual induction the terms, contracts, and labeled contracts. (e = y) By case analysis on x = y, using weakening (Lemma 7) as appropriate. (e = k ) Immediate. (e = e1 op e2 ) By the IHs. (e = e1 e2 ) By the IHs. (e = λy:T . e1 ) By case analysis on x = y, narrowing when the two are equal. (e = µ(y:T ). e1 ) By case analysis on x = y, narrowing when the two are equal. (e = if e1 e2 e3 ) By the IHs. (e = errl ) Immediate. (e = monl (C , e)) By the IHs. (C = predσ (e)) We have Γ1 , x :T1 , Γ2 ` predσ (e) : T2 and we must show that Γ1 , Γ2 ` predσ (e)[v /x ] : T2 . By inversion, we know that: T2 = B for some base type B ; Γ1 , x :T1 , Γ2 , Γ0 ` e : B →Bool; and Γ0 ` σ. There are two cases. In both cases, we use TPred. (x ∈ fv(e)) The substitution is actually stored in σ. We find Γ1 , Γ2 , Γ0 , x :T1 ` e : B by the IH on e and Γ0 , x :T1 ` σ[x 7→ v ] by TMap and our assumption. (x 6∈ fv(e)) The substitution is ignored, and e[v /x ] = e. We find Γ1 , Γ2 , Γ0 ` e : B by the IH and Γ0 ` σ by assumption. (C = x :C1 7→ C2 ) By the IHs. (c = nil) Immediate. (c = predlσ (e); r ) As for the predicate contracts above, using the IH to show that r is still well typed. (c = x :c1 7→ c2 ) By the IHS. As a corollary, closing substitutions σ close up exactly their context Γ0 . Lemma 9 (Closing substitutions). If Γ0 ` σ and Γ, Γ0 ` e : T then Γ ` σ(e) : T . Proof. By induction on Γ0 , using substitution (Lemma 8). Lemma 10 (Progress). If ∅ ` e : T then either (a) e is some value, v ; (b) e is some error, errl ; or, (c) e −→E e 0 . Proof. By induction on the typing derivation. 16 Greenberg ( TVar) Contradictory—variables aren’t well typed in the empty context. ( TConst) k is a value. ( TBlame) errl is an error. ( TAbs) Function abstractions are values. ( TRec) Takes a step by EFix. ( TOp) By the IH, e1 is a value, error, or steps. In the latter two cases, we step by EOpLRaise or EOpL, respectively. By the IH, e2 is a value, error, or steps. In the latter two cases we step by EOpRRaise or EOpR. If both are values, then we step by EDelta. ( TApp) By the IH, e1 is a value, error, or steps. In the latter two cases, we step by EAppLRaise or EAppL, respectively. By the IH, e2 is a value, error, or steps. In the latter two cases we step by EAppRRaise or EAppR. If both are values, we know that e1 is either a function or a function proxy; we step by either EBeta or EMonCApp, depending on the shape of e1 . ( TIf) By the IH, e1 is a value, error, or steps. In the latter two cases, we step by EIfRaise or EIf, respectively./ If e1 is a value, it must be either true or false; we step by EIfTrue or EIfFalse, depending on the shape of e1 . ( TMon) We step by EMonLabel. ( TMonC) If the inner term e is a monitor, we step by EMonCJoin. Otherwise, by the IH, the inner term is a value, an error, or steps. If it’s an error, we step by EMonCRaise. If it steps, we step by EMonC (knowing already that that e isn’t a monitor). If e is a value, we either step by EMonCNil, step by EMonCPred, or have a value already (when e is a function). Lemma 11. If Γ ` C : T then Γ ` labell (C ) : T . Proof. By induction on the typing derivation, using TCPred and TCNil in the predicate case and TCFun in the function case. Lemma 12. If Γ ` r1 : T and Γ ` r2 : T then Γ ` join(r1 , r2 ) : T . Proof. By induction on the typing derivation of r1 . Lemma 13. If Γ ` c1 : T and Γ ` c2 : T then Γ ` join(c1 , c2 ) : T . Proof. By induction on the typing derivation of c1 , using Lemma 12 in the base case and substitution (Lemma 8). Lemma 14 (Preservation). If ∅ ` e : T and e −→E e 0 then ∅ ` e 0 : T . Proof. By induction on the typing derivation, with cases on the step taken. ( TVar) Contradictory—variables aren’t well typed in the empty context. ( TConst) Contradictory—constants are normal forms. ( TBlame) Contradictory—errors are normal forms. ( TAbs) Contradictory—functions are normal forms. Space-Efficient Latent Contracts 17 ( TRec) We must have stepped by EFix; by substitution (Lemma 8) and TFix. ( TOp) If we stepped by an EOpRaise rule, we are done by TBlame. If we stepped by a congruence rule, EOp, we are done by the IH and TOp. If we stepped by EDelta, by the assumption that built-in operations have sound denotations. ( TApp) If we stepped by an EAppRaise rule, we are done by TBlame.If we stepped by a congruence rule, EApp, we are done by the IH and TApp. If we stepped by EBeta, by substitution. (Lemma 8) If we stepped by EMonCApp, then: the domain is well typed by TMon; we can apply the proxied function by TApp; the unmonitored codomain well formed by substitution (Lemma 8); and the entire resulting term is well typed by TMon again. ( TIf) If we stepped by EIfRaise, by TBlame. If we stepped by EIf, by the IH and TIf. If we stepped by EIfTrue or EIfFalse, by assumption. ( TMon) Immediate from the assumptions, since labeling contracts preserves typing (Lemma 11). ( TMonC) If we stepped by EMonCRaise, the resulting error is well typed by TBlame. If we stepped by EMonC, by the IH and TMon. If we stepped by EMonCJoin, then by Lemma 13; if by EMonCNil, then by the assumptions . If we stepped by EMonCPred, then: we have ∅ ` σ(e) : B →Bool by closing substitution (Lemma 9); we can type the condition by TApp; the then branch is well typed by assumption; the false branch is well typed by TBlame; and the whole lot is well typed by TIf. Lemma 15 (Determinism). If e1 −→E e2 and e1 −→E e20 , then e2 = e20 . Proof. By induction on the first derivation. Recall that we exclude the CPCFC rules, so only EMonLabel fires on monitors. EMonC and EMonCJoin carefully avoid overlapping. 3 Soundness for space efficiency CPCFC and CPCFE are operationally equivalent, even though their cast semantics differ. We can make this connection formal by proving that every CPCF term either: (a) diverges in both CPCFC and CPCFE or (b) reduces to equivalent terms in both CPCFC and CPCFE . One minor technicality: some of the forms in our language are necessary only for runtime or only appear in one of the two calculi. We characterize source programs as those which omit runtime terms. Definition 16 (Source program). A well typed source program does not use TBlame or TMonC (and so TCNil, TCPred, and TCFun cannot be used). Lemma 17 (Associativity of join on predicate stacks). join(r1 , join(r2 , r3 )) = join(join(r1 , r2 ), r3 ). Proof. By induction on r1 . (r1 = nil) Immediate. 18 Greenberg (r1 = predl (e); r10 ) The IH is join(r10 , join(r2 , r3 )) = join(join(r10 , r2 ), r3 ). We calculate: join(r1 , join(r2 , r3 )) = join((predlσ (e); r10 ), join(r2 , r3 )) = predlσ (e); drop(join(r10 , join(r2 , r3 )), predσ (e)) = predlσ (e); drop(join(join(r10 , r2 ), r3 ), predσ (e)) = join(predlσ (e); drop(join(r10 , r2 ), predσ (e)), r3 ) = join(join((predlσ (e); r10 ), r2 ), r3 ) = join(join(r1 , r2 ), r3 ) (IH) Lemma 18 (Associativity of join). join(c1 , join(c2 , c3 )) = join(join(c1 , c2 ), c3 ). Proof. By induction on c1 . In the predicate stack case, by Lemma 17. In the function case, we calculate: join(x :c11 7→ c12 , join(x :c21 7→ c22 , x :c31 7→ c32 )) = join(x :c11 7→ c12 , x :join(c31 , c21 ) 7→ join(wrap(c22 , x , c31 ), c32 )) = x :join(join(c31 , c21 ), c11 ) 7→ join(wrap(c12 , x , join(c31 , c21 )), join(wrap(c22 , x , c31 ), c32 )) (IH) = x :join(c31 , join(c21 , c11 )) 7→ join(wrap(join(wrap(c12 , x , c21 ), c22 ), x , c31 ), c32 ) = join(x :c31 7→ c32 , x :join(c21 , c11 ) 7→ join(wrap(c12 , x , c21 ), c22 )) = join(x :c31 7→ c32 , join(x :c11 7→ c12 , x :c21 7→ c22 )) Lemma 19 (Idempotence of predicate stacks). If: – ∅ ` k : B, – e k −→∗E true, and – ∅ ` join(r1 , r2 ) : B , then mon(join(r1 , r2 ), k ) −→∗E w iff mon(join(r1 , drop(r2 , predσ (e))), k ) −→∗E w . Proof. By induction on length of r1 , observing that σ(e) k −→∗E true is redundantly (but successfully) checked on the left more than on the right. Greenberg identified the key property for proving soundness of a space efficient semantics: to be sound, the space-efficient semantics must recover a notion of congruence for checking. In his manifest setting, he calls it cast congruence; since CPCF uses contract monitors, we call it monitor congruence. Lemma 20 (Monitor congruence (single step)). If ∅ ` e1 : T and ∅ ` c : T and e1 −→E e2 , then mon(c, e1 ) −→∗E w iff mon(c, e2 ) −→∗E w . Proof. By cases on the step taken to find e1 −→E e2 . In the easy case, there’s no joining of coercions and the same rule can apply in both derivations. In the more interesting case, two contract monitors join. In either case, it suffices to show that the terms are ultimately confluent, since determinism (Lemma 15)will do the rest. Space-Efficient Latent Contracts 19 Join-free reductions ( EFix) We give this case explicitly, as the original proof in Greenberg [11] doesn’t talk about fixpoints. We must show mon(c, µ(x :T ). e) −→∗E w iff mon(c, e[µ(x :T ). e/x ]). The former steps to the latter immediately by EMonC and EFix, so we are done by determinism (Lemma 15). The following cases are all similar: EDelta, EBeta, EIfTrue, EIfFalse, EAppL, EAppR, EOpL, EOpR, EIf, EMonLabel, EMonCApp, and all of the ERaise rules. A word about EMonCApp: one might think that this is a joining reduction, but in fact it is not: if we can unwrap in the inner step, then the function proxy isn’t able to join with the outer monitor. Joining reductions ( EMonCNil) We must show that mon(c, mon(nil, v1 )) −→∗E w ⇔ mon(c, v1 ) −→∗E w . The terms are confluent: mon(c, mon(nil, v1 )) −→E mon(join(nil, c), v1 ) = mon(c, v1 ). ( EMonCPred) We must show that: mon(r2 , mon((predlσ (e); r1 ), v1 )) −→∗E w ⇔ mon(r2 , if (σ(e) v1 ) mon(r1 , v1 ) errl ) −→∗E w The left-hand side steps: mon(join((predlσ (e); r1 ), r2 ), v1 ) −→C mon(predlσ (e); drop(join(r1 , r2 ), predσ (e)), v1 ) by EMonCJoin, which in turn steps to if (σ(e) v1 ) mon(drop(join(r1 , r2 ), predσ (e)), v1 ) errl . By cases on the behavior of σ(e) v1 . • If σ(e) v1 −→∗E true, then both sides reduce by EIfTrue (using EMonC on the right). The left hand side reduces to mon(drop(join(r1 , r2 ), predσ (e)), v1 ) while the right hand side is mon(join(r1 , r2 ), v1 ). By idempotence of contracts (Lemma 19). • If σ(e) v1 −→∗E false, then both sides reduce by EIfFalse (using EMonC on the right); the right reduces by EMonRaise, and both sides are errl . • Finally, if σ(e) v1 diverges, then both terms diverge. 20 Greenberg ( EMonC) We must show that: mon(c, mon(c 0 , e)) −→∗E w ⇔ mon(c, mon(c 0 , e 0 )) −→∗E w given that e −→E e 0 and knowing that e isn’t another monitor. Both sides step by EMonCJoin, and then the left steps by EMon, to find mon(join(c 0 , c), e 0 ). ( EMonCJoin) We must show that: mon(c, mon(c2 , mon(c1 , e))) −→∗E w ⇔ mon(c, mon(join(c1 , c2 ), e)) −→∗E w The left-hand side joins to mon(join(c1 , join(c, c2 )), e) while the right-hand side joins to mon(join(join(c1 , c2 ), c), e). We have confluence, by associativity of join (Lemma 18). ( EMonCRaise) We must show that: mon(c, mon(c 0 , errl )) −→∗E w ⇔ mon(c, errl ) −→∗E w The left-hand steps by EMonCJoin, and then both sides can step by EMonCRaise to errl . Lemma 21 (Monitor congruence). If – ∅ ` e1 : T , – ∅ ` c : T , and – e1 −→∗E e2 , then mon(c, e1 ) −→∗E w iff mon(c, e2 ) −→∗E e2 . Diagrammatically: ∗ E e2 e1 ⇓ mon(c, e2 ) E w E ∗ ∗ mon(c, e1 ) Proof. By induction on the derivation of e1 −→∗E e2 , using single-step monitor congruence (Lemma 20). It is particularly satisfying that the key property for showing soundness of space efficiency can be proved independently of the inefficient semantics. Implementors can work entirely in the context of the space-efficient semantics, knowing that congruence ensures soundness. We show the observational equivalence of CPCFC and CPCFE by logical relations (Fig. 8), which gives us contextual equivalence—the strongest equivalence we could ask for. Lemma 22 (Similar contracts are logically related). If Γ ` C1 ∼ C2 : T and Γ ` v1 ' v2 : T then Γ ` monl (C1 , v1 ) ' monl (C2 , v2 ) : T . Space-Efficient Latent Contracts Result rules 21 e1 ∼ e2 : T k ∼ k : B ⇐⇒ ty(k ) = B v11 ∼ v21 : T1 →T2 ⇐⇒ ∀e12 ∼ e22 : T1 . v11 e12 ' v21 e22 : T2 errl ∼ errl : T Term rules e1 ' e2 : T e1 ' e2 : T ⇐⇒ (e1 diverges ∧ e2 diverges)∨(e1 −→∗C w1 ∧e2 −→∗E w2 ∧w1 ∼ w2 : T ) Contract rules (invariant relation) Γ ` C1 ∼ C2 : T Γ ` predσ1 (e1 ) ∼ predσ2 (e2 ) : B ⇐⇒ Γ ` σ1 (e1 ) ' σ2 (e2 ) : B →Bool Γ ` x :C11 7→ C12 ∼ x :C21 7→ C22 : T1 →T2 ⇐⇒ Γ ` C11 ∼ C21 : T1 ∧ Γ, x :T1 ` C12 ∼ C22 : T2 Closing substitutions and open terms Γ \|= δ Γ ` e1 ' e 2 : T Γ \|= δ ⇐⇒ ∀x ∈ dom(Γ). δ1 (x ) ∼ δ2 (x ) : Γ(x ) Γ ` e1 ' e2 : T ⇐⇒ ∀Γ \|= δ. δ1 (e1 ) ' δ2 (e2 ) : T Fig. 8. Logical relation between classic and space-efficient CPCF Proof. By induction on the type index of the invariant relation Γ ` C1 ∼ C2 : T . Let Γ \|= δ be given; we must show that δ1 (monl (C1 , v1 )) ' δ2 (monl (C2 , v2 )) : T2 . The right-hand side steps by EMonLabel to mon(labell (δ2 (C2 )), δ2 (e2 )). If T = B , then Ci = predσi (ei ), and each side reduces to if (δi (σi (ei )) δi (vi )) δi (vi ) errl (by EMonPred on the left and EMonCPred on the right). By the assumption in the invariant relation, the conditions of the if co-terminate. The two sides behave exactly the same when the conditions diverge, raise blame, or return false. When they return true, they return related values. The more interesting case is when T = T1 →T2 ; we must show that: δ1 (monl (x :C11 7→ C12 , v1 )) ∼ δ2 (mon(x :labell (C21 ) 7→ labell (C22 ), v2 )) : T2 The left-hand side is a value; the right-hand side may or may not be a value, depending on whether v2 is proxied or not. (v2 = λx :T1 . e2 ) Both sides are values. To show the logical relation, we let arguments v12 ∼ v22 : T1 be given and must show that δ1 (monl (x :C11 7→ C12 , v1 )) v12 ' δ2 (mon(x :labell (C21 ) 7→ labell (C22 ), v2 )) v22 : T2 . 22 Greenberg Each side unwraps ( EMonApp and EMonCApp). Since the LR is closed under evaluation, we can use the IH on the domain and monitor (reducing on the right to convert to a monitor of labeled contracts). Note that there is no codomain monitor unless it’s in Ci 2 ’s closing substitution already. (v2 = mon(x :c31 7→ c32 , v20 )) For brevity, let c2 i = labell (C2 i ). The right-hand side steps by EMonCJoin: δ2 (mon(x :c21 7→ c22 , mon(x :c31 7→ c32 , v20 ))) −→E δ2 (mon(join(x :c31 7→ c32 , x :c21 7→ c22 ), v20 )) which is mon(x :δ2 (join(c21 , c31 )) 7→ δ2 (join(c32 , c22 )), δ2 (v20 )). Let arguments v12 ∼ v22 : T1 be given. Now we can unwrap each side (by EMonApp and EMonCApp, respectively); we must show: monl (δ1 (C12 )[v12 /x ], δ1 (v1 ) monl (δ1 (C11 ), v12 )) ' mon(δ2 (join(wrap(c32 , x , c21 ), c22 ))[v22 /x ], δ2 (v20 ) mon(δ2 (join(c21 , c31 )), v22 )) : T2 We will use monitor congruence to resolve the extra contracts on the spaceefficient side. Since the LR is closed under evaluation, we can step the righthand side back to separate out the inner monitors, letting us recover something that behaves like v1 . That is, to show the goal above, we will show that: monl (δ1 (C12 )[v12 /x ], δ1 (v1 ) monl (δ1 (C11 ), v12 )) ' mon(δ2 (c22 )[v22 /x ], δ2 (mon(x :c31 7→ c32 , v20 )) mon(δ2 (c21 ), v22 )) : T2 . We break the check apart in order to apply the IH. Before we explain how to find this relation, why is proving that these two terms related sufficient to show that our original terms are related? It’s sufficient because monitor congruence (Lemma 21) allows us to consider each of the domain and codomain monitors separately, knowing that we’ll get identical behavior. We’ll see that the wrap in our original term is accounted for by codomain monitor that appears when we go through step-by-step. By the IH, we know that monl (δ1 (C11 ), v12 ) ' mon(δ2 (c21 ), v22 ) : T1 . If these terms terminate or diverge, than so will our original right-hand side domain monitor of join(c21 , c31 ), by Lemma 21. In that case we’re done. If the terms yield values v13 ' v23 : T1 , we can continue. Using monitor congruence to step inside of the codomain check (Lemma 21), we have: −→∗E mon(δ2 (c22 )[v22 /x ], δ2 (mon(x :c31 → 7 c32 , v20 )) mon(δ2 (c21 ), v22 )) mon(δ2 (c22 )[v22 /x ], δ2 (mon(x :c31 → 7 c32 , v20 )) v23 ) And we must show: monl (δ1 (C12 )[v12 /x ], δ1 (v1 ) v13 ) ' mon(δ2 (c22 )[v22 /x ], δ2 (mon(x :c31 7→ c32 , v20 )) v23 ) : T2 . Space-Efficient Latent Contracts 23 We can step the right hand side to: mon(δ2 (c22 )[v22 /x ], mon(δ2 (c32 )[v23 /x ], δ2 (v20 ) mon(c31 , v23 ))) By monitor congruence, we know mon(join(c21 , c31 ), v22 ) and mon(c31 , v23 ) coterminate: either they both diverge, both blame the same label, or reduce to a common value, v24 . We’re done in all but the last case, where we must consider: mon(δ2 (c22 )[v22 /x ], mon(δ2 (c32 )[v23 /x ], δ2 (v20 ) v24 )) Recall that mon(δ2 (c21 ), v22 ) −→∗E v23 ; so mon(δ2 (c32 )[v23 /x ], δ2 (v20 ) v24 ) is the same as mon(δ2 (wrap(c32 , x , c21 ))[v22 /x ], δ2 (v20 ) v24 )—that is, the wrapping done in the join captures exactly the substitution that would occur if we evaluated each layer of function proxying step by step. To patch up the derivation, we use substitutivity (Definition 6, part (6)) to see that if the codomains merged before the substitution, they would also do so after the substitution. So our modified right-hand term is still in lock step with our original, where c32 sees a value monitored by c21 but c22 doesn’t. We’ve assumed that δ1 (v1 ) ' δ2 (v2 ) : T1 →T2 , so we know that δ1 (v1 ) v13 ' mon(δ(c32 )[v23 /x ], δ(v20 ) v24 ) : T2 by back-evaluating the right-hand side a few steps. If these terms diverge or go to errors together, we are done. Otherwise, they evaluate to values v10 ' v20 : T2 . It remains to see that: monl (δ1 (C12 )[v12 /x ], v10 ) ' mon(δ2 (c22 )[v22 /x ], v20 ) : T2 which we have by the IH. We can return to our original term and see that v20 is what will be given to what’s left of c22 after joining with wrap(c32 , x , c21 )— monitor congruence let’s us know that these monitors will evaluate the same merged and unmerged. Lemma 23 (Unwinding). If ∅ ` µ(x :T ). e : T , then µ(x :T ). e −→∗m w iff there exists an n such that unrolling the fixpoint only n times converges to the same value, i.e., e[µ(x :T ). . . . e[µ(x :T ). e/x ] . . . /x ] −→∗m w . Proof. From left-to-right, by induction on the evaluation derivation, observing that there must be a finite number of unrollings. From right-to-left, by induction on n, observing that we can replace the substitution by its finite unrolling. Theorem 24 (CPCFC and CPCFE terms are logically related). 1. If Γ ` e : T as a source program then Γ ` e ' e : T . 2. If Γ ` C : T as a source program then Γ ` C ∼ C : T . Proof. By mutual induction on the typing relations. 24 Greenberg Terms Γ ` e : T Each case proceeds by letting Γ \|= δ be given, and then showing that both sides co-terminate, i.e., both diverge, raise blame, or return related values. ( TVar) By the definition of Γ \|= δ. ( TConst) By the definition of the LR. ( TAbs) By the IHs, letting some argument be given to show that the bodies behave similarly. ( TOp) By the IHs. ( TApp) By the IHs. ( TMu) We give this case in full, as it is new to the proof. Let Γ \|= δ; we must show that δ1 (µ(x :T ). e1 ) ' δ2 (µ(x :T ). e2 ) : T . Each side reduces in a single step ( EFix) to δi (ei )[δi (µ(x :T ). ei )/x ], which can be rearranged into δi (ei [µ(x :T ). ei /x ]); we can rearrange this to something that looks like a dual substitution: δi [µ(x :T ). e1 , µ(x :T ). e2 /x ](ei ). By unwinding (Lemma 23), we can see that the fixpoint on each side either (a) diverges, or (b) converge to some related values in a finite unrolling and can be replaced by a finite unrolling. In the former case, we are done because both terms diverge; in the latter case, we have Γ, x :T \|= δ[µ(x :T ). e1 , µ(x :T ). e2 /x ] and can apply the IH. ( TMon) By the IHs and Lemma 22. ( TMonC) Contradictory—doesn’t appear in source programs. ( TBlame) Contradictory—doesn’t appear in source programs. Contracts Γ ` C : T ( TPred) By the IH. ( TFun) By the IHs. 4 Bounds for space efficiency We have seen that CPCFE behaves the same as CPCFC (Theorem 24), but is CPCFE actually space efficient? For programs that don’t use dependency, yes! With dependency, the story is more complicated. 4.1 The simple case Greenberg showed that for simple contracts—without dependency—we can put a bounds on space [11]. We can recover his result in our more general framework. Observe that a given source program e starts with a finite number of predicate contracts in it. As e runs, no new predicates appear (because dependent substitutions have no effect), but predicates may accumulate in stacks. In the worst case, a predicate stack could contain every predicate contract from the original program e exactly once... but no more than that, because joins remove redundancy! Function contracts are also bounded: e starts out with function contracts Space-Efficient Latent Contracts Predicate extraction 25 preds(e), preds(C ), preds(c) : P(e × (Var * e)) preds(predσ (e)) = preds(x ) = ∅ S {(e, σ)} ∪ preds(e) ∪ [x 7→v ]∈σ preds(v ) preds(k ) = ∅ preds(λx :T . e) = preds(e) preds(x :C1 7→ C2 ) = preds(C1 ) ∪ preds(C2 ) preds(monl (C , e)) = preds(C ) ∪ preds(e) preds(mon(c, e)) = preds(c) ∪ preds(e) preds(nil) = ∅ preds(e1 e2 ) = preds(e1 ) ∪ preds(e2 ) preds(predlσ (e); r ) = {(e, σ)} ∪ preds(e) ∪ S 0 preds(e1 op e2 ) = preds(e1 ) ∪ preds(e2 ) [x 7→e 0 ]∈σ preds(e ) ∪ preds(r ) preds(if e1 e2 e3 ) = preds(x :c1 7→ c2 ) = preds(c1 ) ∪ preds(c2 ) preds(e1 ) ∪ preds(e2 ) ∪ preds(e3 ) preds(errl ) = ∅ Contract size PB : N SB : N size(C ) : N PB = \|{e ∈ preds(e) \| Γ ` predσ (e) : B }\| size(predσ (e)) = SB when ∅ ` predσ (e) : B SB = L · PB · log2 PB size(x :C1 7→ C2 ) = size(C1 ) + size(C2 ) Fig. 9. Predicate extraction and contract size of a certain height, and evaluation can only shrink that height. The leaves of function contracts are labeled with predicate stacks, so the largest contract we could ever see is of maximum height with maximal predicate stacks at every leaf. As the program runs, abutting monitors are joined, giving us a bound on the total number of monitors in a program (one per non-monitor AST node). We can make these ideas formal by first defining what we mean by “all the predicates in a program”, and then showing that evaluation doesn’t introduce predicates (Lemma 29). We let preds(e) be the set of predicates in a term, where a predicate is represented as a pair of term and a closing substitution. Lemma 25. preds(drop(r , pred(e))) ⊆ preds(r ) Proof. By induction on the predicate stack r . When r = nil, the proof is trivial; otherwise, we observe that we only lose predicates from r . Lemma 26. preds(C ) = preds(labell (C )) Proof. By induction on C ; the predicate case is immediate, while the arrow case uses the IHs. We say program e uses simple contracts when all predicates in e are closed and every predicate stack has no redundancies. As such a program reduces, no new contracts appear (and contracts may disappear). Concretely, these requirements are that if predσ (e 0 ) ∈ e or predlσ (e 0 ) ∈ e, we have fv(e 0 ) = ∅ and σ = ι. Restricting programs to simple contracts is what lets us prove that substitution and joining don’t increase the set of predicates. The “no redundancy” requirement 26 Greenberg 0 means that if predlσ (e1 ); r ∈ e then ∀predlσ0 (e2 ) ∈ r . predσ (e1 ) 6⊃ predσ0 (e2 ) (and that r itself has no redundancy, etc.). The following theorems only hold for programs that use simple contracts. Lemma 27. preds(e[e 0 /x ]) ⊆ preds(e) ∪ preds(e 0 ) Proof. By induction on e, using the absorptive property of set union. If e is a predicate contract, it has no free variables (by assumption), so the substitution doesn’t hold on to anything. Lemma 28. If ∅ ` c1 : T and ∅ ` c2 : T then preds(join(c1 , c2 )) ⊆ preds(c1 ) ∪ preds(c2 ). Proof. By induction on c1 . If c1 is a predicate stack, then so must be c2 . If c1 = nil, the proof is immediate. If c1 is a non-empty predicate stack, then by Lemma 25. In the function contract case, by the IHs. Since we’re using simple contracts, we know that wrap never does anything, because predicates have no free variables. Lemma 29 (Reduction is non-increasing in simple predicates). If ∅ ` e : T and e −→m e 0 then preds(e 0 ) ⊆ preds(e). Proof. By induction on the step taken. ( EDelta) Immediate—no predicates. ( EBeta) By substitution (Lemma 27). ( EFix) Observe that the fixpoint operator introduces no preds other than its body; then by substitution (Lemma 27). ( EIfTrue) By definition: preds(e2 ) ⊆ preds(true) ∪ preds(e2 ) ∪ preds(e3 ). ( EIfFalse) By definition, as above. ( EMonPred) By definition: the right-hand side has no new predicates and possibly one fewer. ( EMonApp) The substitution in the codomain is ignored (Lemma 27), so the right-hand side is just a rearrangement of the left. ( EMonLabel) By definition and the fact that labeling leaves the predicate set alone (Lemma 26). ( EMonCNil) By definition. ( EMonCPred) As for EMonPred above: the right-hand side has no new predicates and possibly one fewer. ( EMonCApp) As for EMonApp: we can rearrange our constituent parts— because substitutions are ignored (Lemma 27). ( EMonCJoin) By the monotonicity of join (Lemma 28), knowing that wrap does nothing. ( EL, ER, EIf, EMon, EMonC) By the IH. ( ERaise) Immediate—there are no predicates on the right-hand side. To compute the concrete bounds, we define PB as the number of distinct predicates at the base type B . We can represent a predicate stack at type B in SB bits, where L is the number of bits needed to represent a blame label. A Space-Efficient Latent Contracts 27 given well typed contract ∅ ` C : T can then be represented in size(C ) bits, where each predicate stacks are represented is SB bits and function types are represented as trees of predicate stacks. Finally, since reduction is non-increasing (Lemma 29), we can bound the amount of space used by any contract by looking at the source program, e: we can represent all contracts in our program in at most s = maxC ∈e size(C ) space—constant for a fixed source program. Readers familiar with Greenberg’s paper (and earlier work, like Herman et al. [14]) will notice that our bounds are more precise, tracking the number of holes in contracts per type (size(C )) rather than simply computing the largest conceivable type (2height(T ) ). 4.2 The dependent case In the dependent case, we can’t generally bound the number of contracts by the size of contracts used in the program. Consider the following term, where n ∈ N: let downTo = µ(f :Int→Int). monl (x :pred(λx :Int. x ≥ 0) 7→ pred(λy:Int. x ≥ y), λx :Int. if (x = 0) 0 (f (x − 1))) in downTo n How many different contracts will appear in a run of this program? As downTo runs, we’ll see n different forms of the predicate predlσi (λy:Int. x ≥ y). We’ll have one σn = [x 7→ n] on the first call, σn−1 = [x 7→ n − 1] on the second call, and so on. But n’s magnitude doesn’t affect our measure of the size of source program’s contracts. The number of distinct contracts that appear will be effectively unbounded. In the simple case, we get bounds automatically, using the smallest possible implication relation—syntactic equality. In the dependent case, it’s up to the programmer to identify implications that recover space efficiency. We can recover space efficiency for downTo by saying predσ1 (λy:Int. x ≥ y) ⊃ predσ2 (λy:Int. x ≥ y) iff σ1 (x ) ≤ σ2 (x ). Then the codomain checks from recursive calls will be able to join: downTo n −→∗E monl (pred[x 7→n] (. . . ), . . . ) −→∗E monl (pred[x 7→n] (. . . ), monl (pred[x 7→n − 1] (. . . ), . . . )) −→∗E monl (pred[x 7→n − 1] (. . . ), . . . ) Why are we able to recover space efficiency in this case? Because we can come up with an easily decidable implication rule for our specific predicates matching how our function checks narrower and narrower properties as it recurses. Recall the mutually recursive even/odd example (Fig. 1). = monlodd (x :pred(λx :Int. x ≥ 0) 7→ pred(λb:Bool. b or (x mod 2 = 0)), λx :Int. if (x = 0) false (even (x − 1))) and even = λx :Int. if (x = 0) true (odd (x − 1)) let odd 28 Greenberg We can make this example space-efficient by adding the implication that: predσ1 (λb:Bool. b or (x mod 2 = 0)) ⊃ predσ2 (λb:Bool. b or (x mod 2 = 0)) iff σ1 (x ) + 2 = σ2 (x ). Suppose we put contracts on both even and odd: let odd = monlodd (x :pred(λx :Int. x ≥ 0) 7→ pred(λb:Bool. b or (x mod 2 λx :Int. if (x = 0) false (even (x and even = monleven (x :pred(λx :Int. x ≥ 0) 7→ pred(λb:Bool. b or ((x + 1) mod 2 λx :Int. if (x = 0) true (odd (x = 0)), − 1))) = 0)), − 1))) Now our trace of contracts won’t be homogeneous; eliding domain contracts: odd 5 −→∗C monlodd (pred[x 7→5] (. . . ), even 4) −→∗C monlodd (pred[x 7→5] (. . . ), monleven (pred[x 7→4] (. . . ), odd 3)) −→∗C monlodd (pred[x 7→5] (. . . ), monleven (pred[x 7→4] (. . . ), monlodd (pred[x 7→3] (. . . ), even 2))) ∗ −→C monlodd (pred[x 7→5] (. . . ), monleven (pred[x 7→4] (. . . ), monlodd (pred[x 7→3] (. . . ), monleven (pred[x 7→2] (. . . ), odd 1)))) ∗ −→C monlodd (pred[x 7→5] (. . . ), monleven (pred[x 7→4] (. . . ), monlodd (pred[x 7→3] (. . . ), monleven (pred[x 7→2] (. . . ), monlodd (pred[x 7→1] (. . . ), even 0))))) To make these checks space efficient, we’d need several implications; we write oddp for λb:Bool. b or (x mod 2 = 0) and evenp for λb:Bool. b or ((x + 1) mod 2 = 0). The following table gives conditions on the implication relation for the row predicate to imply the column predicate: ⊃ predσ1 (oddp ) predσ1 (evenp ) predσ2 (oddp ) σ1 (x ) + 2 = σ2 (x ) σ1 (x ) + 1 = σ2 (x ) predσ2 (evenp ) σ1 (x ) + 1 = σ2 (x ) σ1 (x ) + 2 = σ2 (x ) Having all four of these implications allows us to eliminate any pair of checks generated by the recursive calls in odd and even, reducing the codomain checking to constant space—here, just one check. We could define a different implication relation, where, say, predσ1 (oddp ) ⊃ predσ2 (oddp ) iff σ1 (x ) mod 2 = σ2 (x ) mod 2. Such an implication would apply more generally than those in the table above— it isn’t always obvious how to define the implication relation. As usual, there is a trade-off between time and space. It’s possible to write contracts where the necessary implication relation for space efficiency amounts to checking both contracts. Consider the following tail-recursive factorial function: let any = λz :Int. true let fact = µ(f :Int→Int→Int). monl (x :pred(any) 7→ acc:pred(any) 7→ pred(λy:Int. x ≥ 0), λx :Int. λacc:Int. if (x = 0) acc (f (x − 1) (x ∗ acc))) Space-Efficient Latent Contracts 29 This contract isn’t wrong, just strange: if you call fact with a negative number, the program diverges and you indeed won’t get a value back out (contracts enforce partial correctness). A call of fact 3 yields monitors that check, from outside to inside, that 3 ≥ 0 and 2 ≥ 0 and 1 ≥ 0 and 0 ≥ 0. When should we say that predσ1 (λy:Int. x ≥ 0) ⊃ predσ1 (λy:Int. x ≥ 0)? We could check that σ1 (x ) ≥ σ2 (x )... but the time cost is just like checking the original contract. More starkly, consider the following function with a strange contract (where prime is a predicate identifying prime numbers): let absurd = monl (x :pred(λx :Int. true) 7→ pred(λy:Int. prime x ), λx :Int. if (x ⊂ 13) 1 (absurd 11)) How does absurd 47 run? absurd 47 = monl (pred[x 7→47] (λy:Int. prime x ), absurd 11) = monl (pred[x 7→47] (λy:Int. prime x ), monl (pred[x 7→11] (λy:Int. prime x ), 1)) What implication relation could resolve these two unrelated checks into one? We could say that predσ1 (λy:Int. prime x ) ⊃ predσ2 (λy:Int. prime x ) iff σ1 (x ) and σ2 (x ) are prime. There’s no time-wise savings, but it would allow us to collapse tall stacks of monitors. 5 Where should the implication relation come from? The simplest option is to punt: derive the implication relation as the reflexive transitive closure of a programmer’s rules. A programmer might specify how several different predicates interrelate as follows: 1 2 3 y : Int { x1 >= y } implies y : Int { x2 >= y } when x1 <= x2 y : Int { x1 > y } implies y : Int { x2 >= y } when x1 <= x2 + 1 y : Int { x1 > y } implies y : Int { x2 > y } when x1 <= x2 A default collection of such implications might come with the language; library programmers should be able to write their own, as well. But it is probably unwise to allow programmers to write arbitrary implications: what if they’re wrong? A good implementation would only accept verified implications, using a theorem prover or an SMT solver to avoid bogus implications. Rather than having programmers write their own implications, we could try to automatically derive the implications. Given a program, a fixed number of predicates occur, even if an unbounded number of predicate/closing substitution pairings might occur at runtime. Collect all possible predicates from the source program, and consider each pair of predicates over the same base type, pred(e1 ) and pred(e2 ) such that Γ ` ei : B →Bool. We can derive from the typing derivation the shapes of the respective closing substitutions, σ1 and σ2 . What are the conditions on σ1 and σ2 such that predσ1 (e1 ) ⊃ predσ2 (e2 )? We are looking for a property P (σ1 , σ2 ) such that: ∀k ∈ KB , P (σ1 , σ2 ) ∧ σ1 (e1 ) k −→∗E true ⇒ σ2 (e2 ) k −→∗E true 30 Greenberg Ideally, P is also efficiently decidable—at least more efficiently than deciding both predicates. The problem of finding P can be reduced to that of finding the weakest precondition for the safety of the following function: 1 2 3 4 fun x : B = > let y0 = v 10 , ... , yn = v 1n ( σ1 ’s bindings ) z0 = v 20 , ... , zn = v 2m ( σ2 ’s bindings *) in if e1 x then ( if e2 x then () else error ) else () Since P would be the weakest precondition, we would know that we had found the most general condition for the implication relation. Whether or not the most general condition is the best condition depends on context. We should also consider a cost model for P ; programmers may want to occasionally trade space for time, not bothering to join predicates that would be expensive to test. Finding implication conditions resembles liquid type inference [20,28,16]: programmers get a small control knob (which expressions can go in P ) and then an SMT solver does the rest. The settings are different, though: liquid types are about verifying programs, while we’re executing checks at runtime. 5.1 Implementation Implementation issues abound. How should the free variables in terms be represented? What kind of refactorings and optimizations can the compiler do, and how might they interfere with the set of contracts that appear in a program? When is the right moment in compilation to fix the implication relation? More generally, what’s the design space of closure representations and calling conventions for languages with contracts? 6 Extensions Generalizing our space-efficient semantics to sums and products does not seem particularly hard: we’d need contracts with corresponding shapes, and the join operation would push through such shapes. Recursive types and datatypes are more interesting [22]. Findler et al.’s lazy contract checking keeps contracts from changing the asymptotic time complexity of the program [7]; we may be able to adapt their work to avoid changes in asymptotic space complexity, as well. The predicates here range over base types, but we could also allow predicates over other types. If we allow predicates over higher types, how should the adequacy constraint on predicate implication (Definition 6) change? Impredicative polymorphism in the style of System F would require even more technical changes. The introduction of type variables would make our reasoning about names and binders trickier. In order to support predicates over type variables, we’d need to allow predicates over higher types—and so our notion of adequacy of ⊃ would change. In order to support predicates over quantified types, we’d need to change adequacy again. Adequacy would end up looking like the logical relation used to show relational parametricity: when would we Space-Efficient Latent Contracts 31 have ∀α.T1 ⊃ ∀α.T2 ? If we substitute T10 for α on the left and T20 for α on the right (and T10 and T20 are somehow related), then T1 [T10 /α] ⊃ T2 [T20 /α]. Not only would the technicalities be tricky, implementations would need to be careful to manage closure representations correctly (e.g., what happens if polymorphic code differs for boxed and unboxed types?). We don’t treat blame as an interesting algebraic structure—it’s enough for our proofs to show that we always produce the same answer. Changing our calculus to have a more interesting notion of blame, like indy semantics [4] or involutive blame labels [30,29], would be a matter of pushing a shallow change in the semantics through the proofs. Finally, it would make sense to have substitution on predicate stacks perform joins, saying (predlσ (e); r )[v /x ] = join(predlσ (e)[v /x ]; nil, r [v /x ]), so that substituting a value into a predicate stack checks for newly revealed redundancies. We haven’t proved that this change would be sound, which would require changes to both type and space-efficiency soundness. In particular, we’d need to (a) revise Lemma 22 and (b) simultaneously prove that substitution and joins preserve types (Lemmas 8) and 13). 7 Related work For the technique of space efficiency itself, we refer the reader to Greenberg [11] for a full description of related work. We have striven to use Greenberg’s notation exactly, but we made some changes in adapting to dependent contracts: the cons operator for predicate stacks is a semi-colon, to avoid ambiguity; there were formerly two things named join, but one has been folded into the other; and our predicates have closing substitutions to account for dependency. We place one more requirement on the implication relation than Greenberg did: monotonicity under substitution, which we call substitutivity. Substitutions weren’t an issue in his non-dependent system, but we must require that if a join can happen without having a value for x , then the same join happens when we know x ’s value. CPCF was first introduced in several papers by Dimoulas et al. in 2011 [3,4], and has later been the subject of studies of blame for dependent function contracts [5] and static analysis [27]. Our exact behavioral equivalence means we could use results from Tobin-Hochstadt et al.’s static analysis in terms of CPCFC to optimize space efficient programs in CPCFE . More interestingly, the predicate implication relation ⊃ seems to be doing some of the work that their static analysis does, so there may be a deeper relationship. Thiemann introduces a manifest calculus where the compiler optimizes casts for time efficiency: a theorem prover uses the “delta” between types to synthesize more efficient checks [26]. His deltas and our predicate implication relation are similar. He uses a separate logical language for predicates and restricts dependency (codomains can only depend on base values, avoiding abusive contracts). Sekiyama et al. [21] also use delayed substitutions in their polymorphic manifest contract calculus, but for different technical reasons. While delayed sub- 32 Greenberg stitutions resemble explicit substitutions [1] or explicit bindings [13,2], we use delayed substitutions more selectively and to resolve issues with dependency. The manifest type system in Greenberg’s work is somewhat disappointing compared to the type system given here. Greenberg works much harder than we do to prove a stronger type soundness theorem... but that theorem isn’t enough to help materially in proving the soundness of space efficiency. Developing the approach to dependency used here was much easier in a latent calculus, though several bugs along the way would have been caught early by a stronger type system. Type system complexity trade-offs of this sort are an old story. 7.1 Racket’s implementation If contracts leak space, how is it that they are used so effectively throughout PLT Racket? Racket is designed to avoid using contracts in leaky ways. In Racket, contracts tend to go on module boundaries. Calls inside of a module then don’t trigger contract checks—including recursive calls, like in the even/odd example. Racket will monitor recursive calls across module boundaries, and these checks can indeed lead to space leaks. In our terms, Racket tends to implement contract checks on recursive functions as follows: downTo = monl (x :pred(λx :Int. x ≥ 0) 7→ pred(λy:Int. x ≥ y), µ(f :Int→Int). λx :Int. if (x = 0) 0 (f (x − 1))) Note that calling downTo n will merely check that the final result is less than n— none of the intermediate values. Our version of downTo above puts the contract inside the recursive knot, forcing checks every time (Sec. 4.2). Racket also offers a less thorough form of space efficiencyvia the tail-marks-match? predicate.3 We can construct a program where Racket will avoid redundant checks, but but wrapping the underlying function with the same contract twice leads to a space leak (Figure 10). We use low-level contract mechanisms here, to highlight the compromises Racket makes for space efficiency, “strik[ing] a balance between common ways that tail recursion is broken and checking that would not be too expensive in the case that there wouldn’t have been any pile-up of redundant wrappers”.4 Finally, contracts are first-class in Racket: their monitors take two expressions, one for the contract and one to be monitored. Computing new contracts at runtime breaks our framing of space-efficiency: if we can’t predetermine which contracts arise at runtime, we can’t fix an implication relation in advance. We hope that CPCFE is close enough to Racket’s internal model to provide insight into how to achieve space efficiency for at least some contracts in Racket. 8 Conclusion We have translated Greenberg’s original result [11] from a manifest calculus [12] to a latent one [3,4]. In so doing, we have: offered a simpler explanation of the 3 4 From racket/collects/racket/contract/private/arrow.rkt. Robby Findler, personal correspondence, 2016-05-19. Space-Efficient Latent Contracts 1 2 3 4 5 6 7 8 33 ( define ( count-em-integer ? x ) ( printf " checking ~ s \ n " x ) ( integer ? x )) ( letrec ([ f ( contract (- > any / c count-em-integer ?) ( lambda ( x ) ( if ( zero ? x ) x ( f (- x 1)))) ’ pos ’ neg )]) ( f 3)) (a) Space-efficient program 1 2 3 4 5 6 7 ( letrec ([ f ( contract (- > any / c count-em-integer ?) ( contract (- > any / c count-em-integer ?) ( lambda ( x ) ( if ( zero ? x ) x ( f (- x 1)))) ’ pos ’ neg ) ’ pos ’ neg )]) ( f 3)) (a) Space-leaking program Fig. 10. Space-efficiency in Racket original result; isolated the parts of the type system required for space bounds, which were intermingled with complexities from conflating contracts and types; and, extended the original result, by covering more features (dependency and nontermination) and more precisely bounding non-dependent programs. Acknowledgments. The existence of this paper is due to comments from Sam Tobin-Hochstadt and David Van Horn that I chose to interpret as encouragement. Robby Findler provided the Racket example and helped correct and clarify a draft; Sam Tobin-Hochstadt also offered corrections and suggestions. The reviews offered helpful comments, too. References 1. Abadi, M., Cardelli, L., Curien, P.L., Lévy, J.J.: Explicit substitutions. Journal of Functional Programming (JFP) 1(4), 375–416 (1991) 2. Ahmed, A., Findler, R.B., Siek, J., Wadler, P.: Blame for all. In: Principles of Programming Languages (POPL) (2011) 3. Dimoulas, C., Felleisen, M.: On contract satisfaction in a higher-order world. TOPLAS 33(5), 16:1–16:29 (Nov 2011) 4. Dimoulas, C., Findler, R.B., Flanagan, C., Felleisen, M.: Correct blame for contracts: no more scapegoating. In: Principles of Programming Languages (POPL) (2011) 5. Dimoulas, C., Tobin-Hochstadt, S., Felleisen, M.: Complete monitors for behavioral contracts. In: Seidl, H. (ed.) Programming Languages and Systems, LNCS, vol. 7211, pp. 214–233. Springer Berlin Heidelberg (2012) 6. Findler, R.B., Felleisen, M.: Contracts for higher-order functions. In: International Conference on Functional Programming (ICFP) (2002) 34 Greenberg 7. Findler, R.B., Guo, S.Y., Rogers, A.: Lazy contract checking for immutable data structures. In: Chitil, O., Horváth, Z., Zsók, V. (eds.) Implementation and Application of Functional Languages, pp. 111–128. Springer-Verlag, Berlin, Heidelberg (2008) 8. Flatt, M., PLT: Reference: Racket. Tech. Rep. PLT-TR-2010-1, PLT Design Inc. (2010), http://racket-lang.org/tr1/ 9. Garcia, R.: Calculating threesomes, with blame. In: International Conference on Functional Programming (ICFP) (2013) 10. Greenberg, M.: Manifest Contracts. Ph.D. thesis, University of Pennsylvania (November 2013) 11. Greenberg, M.: Space-efficient manifest contracts. In: Principles of Programming Languages (POPL) (2015) 12. Greenberg, M., Pierce, B.C., Weirich, S.: Contracts made manifest. In: Principles of Programming Languages (POPL) (2010) 13. Grossman, D., Morrisett, G., Zdancewic, S.: Syntactic type abstraction. TOPLAS 22(6), 1037–1080 (Nov 2000) 14. Herman, D., Tomb, A., Flanagan, C.: Space-efficient gradual typing. In: Trends in Functional Programming (TFP). pp. 404–419 (Apr 2007) 15. Herman, D., Tomb, A., Flanagan, C.: Space-efficient gradual typing. Higher Order Symbol. Comput. 23(2), 167–189 (Jun 2010) 16. Jhala, R.: Refinement types for haskell. In: Programming Languages Meets Program Verification (PLPV). pp. 27–27. ACM, New York, NY, USA (2014) 17. Meyer, B.: Eiffel: the language. Prentice-Hall, Inc. (1992) 18. Plotkin, G.: LCF considered as a programming language. Theoretical Computer Science 5(3), 223 – 255 (1977) 19. Racket contract system (2013) 20. Rondon, P.M., Kawaguchi, M., Jhala, R.: Liquid types. In: Programming Language Design and Implementation (PLDI) (2008) 21. Sekiyama, T., Igarashi, A., Greenberg, M.: Polymorphic manifest contracts, revised and resolved. TOPLAS (2016), accepted in September; to appear 22. Sekiyama, T., Nishida, Y., Igarashi, A.: Manifest contracts for datatypes. In: Principles of Programming Languages (POPL). pp. 195–207. ACM, New York, NY, USA (2015) 23. Siek, J., Thiemann, P., Wadler, P.: Blame, coercion, and threesomes: Together again for the first time. In: Programming Language Design and Implementation (PLDI) (2015) 24. Siek, J.G., Taha, W.: Gradual typing for functional languages. In: Scheme and Functional Programming Workshop (September 2006) 25. Siek, J.G., Wadler, P.: Threesomes, with and without blame. In: Principles of Programming Languages (POPL). pp. 365–376. ACM, New York, NY, USA (2010) 26. Thiemann, P.: A delta for hybrid type checking. In: Wadler Festschrift. pp. 411– 432. LNCS 9600, Springer Switzerland 27. Tobin-Hochstadt, S., Van Horn, D.: Higher-order symbolic execution via contracts. In: OOPSLA. pp. 537–554. ACM, New York, NY, USA (2012) 28. Vazou, N., Rondon, P.M., Jhala, R.: Abstract refinement types. In: Felleisen, M., Gardner, P. (eds.) European Symposium on Programming (ESOP). pp. 209–228. Springer Berlin Heidelberg, Berlin, Heidelberg (2013) 29. Wadler, P.: A Complement to Blame. In: Ball, T., Bodik, R., Krishnamurthi, S., Lerner, B.S., Morrisett, G. (eds.) SNAPL. LIPIcs, vol. 32, pp. 309–320. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik (2015) Space-Efficient Latent Contracts 35 30. Wadler, P., Findler, R.B.: Well-typed programs can’t be blamed. In: European Symposium on Programming (ESOP) (2009)	6cs.PL
On Expressing and Monitoring Oscillatory Dynamics Petr Dluhoš, Luboš Brim, and David Šafránek ∗ Faculty of Informatics Masaryk University Botanická 68a, Brno, Czech Republic safranek@fi.muni.cz To express temporal properties of dense-time real-valued signals, the Signal Temporal Logic (STL) has been defined by Maler et al. The work presented a monitoring algorithm deciding the satisfiability of STL formulae on finite discrete samples of continuous signals. The logic has been used to express and analyse biological systems, but it is not expressive enough to sufficiently distinguish oscillatory properties important in biology. In this paper we define the extended logic STL∗ in which STL is augmented with a signal-value freezing operator allowing us to express (and distinguish) detailed properties of biological oscillations. The logic is supported by a monitoring algorithm prototyped in Matlab. The monitoring procedure of STL∗ is evaluated on a biologically-relevant case study. 1 Introduction In this paper we deal with automatic decision of the question if a given continuous signal satisfies a given temporal property. This question originally arose in the domain of analogous circuits verification [17]. The procedure deciding this question has been called monitoring [17]. A monitor is constructed for a given property allowing off-line or even on-line decision over any continuous signal arising from a technical device (real case) or a numerical simulation procedure (model case) [18]. An important fact is that the monitoring procedure is always time bounded and that validity of the given logic property (discrete nature) is decided only approximately on a signal (continuous nature). These restrictions are necessary but not limiting to significantly help in avoiding errors in systems construction [21]. In systems biology, continuous signals most typically represent the model case – theoretical behaviour of mathematical models mimicking the dynamics of biological processes. Temporal properties are employed to express biological hypothesis [19, 9, 5, 4] and the monitoring procedure provides a promising analysis tool [7, 22, 11]. Many dynamical phenomena that arise in biological systems have the form of oscillations. In physics the term oscillation represents an infinite behaviour periodically alternating certain quantities/qualities. The identifying aspect of oscillations is the fact that certain states in the phase space of the dynamical system are being repeatedly re-visited. A biological example of such phenomena are circadian rhythms. In [23] an interesting wet-lab experiment has been achieved on the cyanobacterium Cyanothece sp. that showed a relation between metabolic cycles and circadian rhythms. A mathematical model of oscillations in gene regulation of cyanobacteria has been provided in [20]. Other models targeting oscillatory behaviour can be found e.g. in [15, 14]. In contrast to physics, the notion of oscillation in biology is usually understood informally and in a wider scope. Many wet-lab experiments show oscillatory behaviour with decreasing amplitude (socalled damped oscillation). Dual notion, oscillation with increasing amplitude, is also relevant: the question targeting how many oscillations and how strongly they increase until a permanent oscillation is ∗ The work has been supported by the Grant Agency of Czech Republic grant GAP202/11/0312. Ezio Bartocci and Luca Bortolussi (Eds.): HSB 2012 EPTCS 92, 2012, pp. 73–87, doi:10.4204/EPTCS.92.6 c P. Dluhos et al. This work is licensed under the Creative Commons Attribution License. 74 On Expressing and Monitoring Oscillations achieved comes to a significant interest when tuning biological processes via mathematical models [15, 14]. When considering the population behaviour as a real-valued signal in dense-time domain, we need to quantitatively express and compactly encode the mentioned types of oscillatory phenomena. In [9], a linear-time logic with constraints over real values has been defined. It can be practically used to express many dynamical phenomena including oscillations of species concentration. The logic is interpreted over finite time-series of real-valued data obtained as discretely sampled solutions of differential equations (ODEs) (or a series of physical measurements). Formally, a technical problem may arise with the dense-time domain of considered signals. In general, in a particular interval on the time axis of the signal there may exist infinitely many points where a considered property (a predicate over real-values) changes its truth value. A rigorous semantics which treats this problem is given in [17] in terms of continuous (real-valued) signals which are required to be piece-wise affine. In such a setting, it is assumed that there exists a finite interval covering of the signal time-domain where the end points can be sampled and on which the signal behaves “reasonably” in each individual time interval. The logic with such a semantics is called Signal Temporal Logic (STL) [17] and is based on the (bounded) Metric Interval Temporal Logic (MITL) [2]. Neither STL, nor other temporal logic, as far as we know, provide any possibility to express (and distinguish) the classes of oscillations such as damped oscillations or oscillations with increasing amplitude. The reason is impossibility of globally referencing (and relatively comparing) concrete signal values occurring in time points in which some local property is satisfied. Of course, in STL we can express directly any single concrete signal, but not a universal property without references to concrete values. E.g., the damped oscillation in Figure 1c can be expressed in STL as a sequence exactly reaching the 15 local extremes in the given order. In general, there can appear any number of local extremes in the observed time interval. Such a general property cannot be expressed in STL. (a) x(t) = sin( 2πt 10 ) (b) x(t) = t 2πt 20 sin( 10 ) (c) x(t) = 20 2πt t sin( 10 ) Figure 1: Various types of oscillations. In this paper, we propose an extension of STL, denoted STL∗ , based on enriching the logic with a signal-value freeze operator ∗[ϕ] which allows referencing to a signal value in a time point when ϕ is determined true. By means of atomic propositions enriched by variables of the form x∗ referring to “frozen” values, we can express damped oscillations by the following formula: G[0,60] (( F[0,10] ∗ G[1,10] x∗ ≥ x + c ) ∧ ( F[0,10] ∗ G[1,10] x∗ ≤ x − c )). It says that ”there is always a time instant in near future which is a local maximum for some future interval and there is also another time instant in near future which is a local minimum for some future interval”. In these intervals, another local maxima and minima for more distant intervals occur, and so on. This implies that the signal values can be surrounded by a decreasing function from top and an increasing function from bottom. The constant c sets the extent of damping. P. Dluhos et al. 75 The concept of freeze quantification has been introduced in [3] to increase real-time temporal logic expressiveness by allowing every temporal operator to bind a time variable to the time it refers to. In the context of systems biology this concept has been used in Biological Oscillator Synchronisation Logic (BOSL) [6] to reason on oscillators synchronisation. In our work we shift the notion of temporal quantification to bind the real-valued signal variables to the temporal reference of time and we do it at the level of STL logic. Freezing of variable values has been used before in higher-level formalisms employing local variables [13, 1]. However, these formalisms do not support dense-time binding. It is necessary to note that when values of derivatives of signal values are known in all sampled points, some oscillatory properties can be expressed by using plain STL with predicates over derivatives. However, in such a case all needed derivatives (of all required orders) must be included within the signal. Getting these data can be computationally hard or even impossible. STL∗ allows to express (and monitor) oscillatory properties without the need of any additional information supporting the signals. The contribution of this paper is an extension of STL logic with signal value freeze quantification (Section 3) motivated by the need to express oscillations appearing in biology. The paper is primarily focused on a practical aspect, we give an algorithm for monitoring STL∗ formulae over (bounded) continuous signals (Section 4). The result is supported by a prototype implementation evaluated on a biological case study of E. coli repressilator (Section 5). Please note that a detailed discussion of STL∗ expressiveness is provided in a master’s thesis [10] that makes a preliminary version of this paper. 2 Background First, we briefly recall the real-valued signals over continuous time from [17]. For the purposes of this paper, we focus only on signals over Rn . Definition 2.1 Let T = [0, r], r ∈ Q≥0 , T ⊆ R≥0 , be a finite interval (time domain of the signal). A finite length signal s is a function s : T → Rn where \|s\| = r denotes the length of the signal s. Definition 2.2 Given a signal s : T → Rn , n denotes the order of the signal s. We use si to denote the i-th component of the signal, si : T → R (i.e., the canonical projection to the i-th element). A signal of order n represents one finite run of a continuous-time system with n variables. A variable can represent any quantity of the system, e.g., concentration of species, its derivative or any other attribute of the signal which can be measured and quantified. To translate the properties of a signal into terms of logic, we use a function transforming real values of the signal components in every time instant to the Boolean set {0, 1}, representing satisfaction or dissatisfaction of a property over the signal. These truth values are used as propositional variables in formulae of the logic. Unlike in [17], we restrict the allowed operations on signal variables to linear combinations and simple comparisons. Such a restriction is needed for our monitoring algorithm (Section 4) and does not limit expression of the properties of our interest. Definition 2.3 Linear predicate ν of order m is a function ν : Rm → B of the form ∑m i=1 ai xi ∼ b where ai , b ∈ R are real coefficients, xi are input variables, ∼∈ {<, ≤, >, ≥, =} and B is a Boolean set B = {0, 1}. The predicate ν(x1 , x2 , . . . xm ) returns 1 iff the expression ∑m i=1 ai xi ∼ b is true and 0 otherwise. However, in the logic defined in this paper, we need not only to work with the values of variables acquired in a single time instant, but also with a second set of values from another time instant. For this reason, we will define an additional function extracting truth values from the signal. 76 On Expressing and Monitoring Oscillations + Definition 2.4 Atomic predicate µ is a function µ : Sn × R+ 0 × R0 → B where Sn is a class of signals ∗ of order n. For each signal s ∈ Sn and t,t ∈ [0, \|s\|], µ is defined by the formula: df µ(s,t,t ∗ ) = ν(s1 (t), . . . , sn (t), s1 (t ∗ ), . . . , sn (t ∗ )), where ν is a linear predicate of order 2n and si are components of the signal s. Values of the atomic predicate µ for t > \|s\| or t ∗ > \|s\| are considered undefined. An atomic predicate over a signal s describes a Boolean property of the signal s, i.e., if a constraint specified by the linear predicate ν is satisfied with respect to two time instants t and t ∗ . An example of such a property is ”under condition t ∗ < t, the difference between values of variables s1 and s2 was higher in the time t ∗ than in the time t”: µ(s,t,t ∗ ) = ν(s1 (t), s2 (t), s1 (t ∗ ), s2 (t ∗ )) = s1 (t) − s2 (t) < s1 (t ∗ ) − s2 (t ∗ ), which can be rearranged to the standard form of a linear predicate according to Definition (2.3): df ν(s1 (t), s2 (t), s1 (t ∗ ), s2 (t ∗ )) = s1 (t) − s2 (t) − s1 (t ∗ ) + s2 (t ∗ ) < 0. Instead of µ(s,t,t ∗ ) = s1 (t) − s2 (t) < s1 (t ∗ ) − s2 (t ∗ ) we usually write µ(s,t,t ∗ ) = x − y < x∗ − y∗ , denoting the variables of a signal s in time t by letters x, y, z, . . . and the same variables in time t ∗ as x∗ , y∗ , z∗ , . . .. As in [17], we will avoid the undesired cases of atomic predicates where the output values of the predicate are varying infinitely often. We assume that we deal with signals that are well-behaving with respect to every µ, that is, µ(s) has a bounded variability and every change in µ(s) is detected in the sense that every point t such that µ(s[t]) 6= limt 0 −>t µ(s[t 0 ]) is included in the sampling [17]. 3 A Logic for Oscillatory Dynamics In this section we describe the syntax and semantics of STL. It is based on the Signal Temporal Logic (STL) [17] which we extend with the signal-value freeze operator ∗[ϕ]. We consider atomic predicates restricted to linear as defined in the previous section. 3.1 Syntax The formulae of the STL are inductively defined by the following grammar: ϕ ::= µi \| ¬ϕ \| ϕ1 ∨ ϕ2 \| ϕ1 U[a,b] ϕ2 \| ∗ϕ, where µi are atomic predicates from Definition 2.4, ¬ and ∨ are standard propositional logic operators, U[a,b] is a bounded until operator constrained by the closed nonsingular time interval [a, b] with rational end-points. The interpretation of this operator is similar to the one used in the Metric Interval Temporal Logic (MITL) [2]. Finally, ∗ is the newly added unary signal-value freeze operator. In the standard way, we can derive additional temporal operators such as eventually ( F[a,b] ϕ ≡ true U[a,b] ϕ) and globally ( G[a,b] ϕ ≡ ¬ F[a,b] ¬ϕ). P. Dluhos et al. 3.2 77 Semantics Formulae of STL* are interpreted over triplets (s,t,t ∗ ) where s is a real-valued signal and t,t ∗ ∈ [0, \|s\|] are two time instants. We write (s,t,t ∗ ) \|= ϕ, iff a formula ϕ is satisfied on the triplet (s,t,t ∗ ). The satisfaction of an STL* formula is defined inductively to its structure: df (s,t,t ∗ ) \|= µ ⇐⇒ µ(s,t,t ∗ ) = 1; (s,t,t ∗ ) \|= ¬ϕ ⇐⇒ (s,t,t ∗ ) 6\|= ϕ; (s,t,t ∗ ) \|= ϕ1 ∨ ϕ2 ⇐⇒ (s,t,t ∗ ) \|= ϕ1 or (s,t,t ∗ ) \|= ϕ2 ; df df df (s,t,t ∗ ) \|= ϕ1 U[a,b] ϕ2 ⇐⇒ ∃t 0 ∈ [a + t, b + t] : (s,t 0 ,t ∗ ) \|= ϕ2 ∧ ∀t 00 ∈ [t,t 0 ] : (s,t 00 ,t ∗ ) \|= ϕ1 ; (s,t,t ∗ ) \|= ∗ϕ (1) df ⇐⇒ (s,t,t) \|= ϕ. Because models for STL* are signals, we also speak about satisfaction of a formula over a signal. df Definition 3.1 The satisfaction of a formula ϕ over a signal s is defined as: s \|= ϕ ⇐⇒ (s, 0, 0) \|= ϕ. The meaning of the operator ∗ϕ is to freeze the values of the signal over which the formula is interpreted in the current time point so we can use these values inside the subformula ϕ. Formula ∗ϕ is true in time t = t0 iff the formula ϕ is true with the time t ∗ frozen in this time, i.e., t ∗ = t0 . In combination with other temporal operators, interesting properties such as ”the value of the variable can be∗ expressed x is nondecreasing on the interval I”, ϕ1 ≡ GI ∗ G[0,ε] (x ≤ x) , where ε > 0 is an arbitrary constant. Or ”at some point in the future (during interval I) the value x increases by the value of 5 within two time units”, ϕ2 ≡ FI ∗ F[0,2] (x∗ + 5 = x) . To determine the satisfaction of a formula over a signal, the signal has to be of a sufficient length. The necessary length can be computed for each formula inductively to its structure as in [17]. The freeze operator does not require any additional length: df l(µ) = 0; l(¬ϕ) = l(ϕ); l(ϕ1 ∨ ϕ2 ) = max(l(ϕ1 ), l(ϕ2 )); df df (2) df l(ϕ1 U[a,b] ϕ2 ) = max(l(ϕ1 ), l(ϕ2 )) + b; l(∗ϕ) df = l(ϕ). From Definition 3.1 we can see that if a formula ϕ contains an atomic predicate µi and there is no operator ∗ wrapping this predicate then all variables concerning the frozen time instant t ∗ are handled as if t ∗ = 0. From the definition of the semantics of STL* (1) follows that in a formula with several nested operators ∗, the meaning of a frozen variable x∗ is local. I.e., it relates to the nearest operator ∗, because the content of the variable t ∗ is overwritten by the most nested freeze operator ∗. It implies that only a single set of frozen signal values is accessible at every place in the formula. For example, in a formula: ∗ x∗ ≤ y UI (x∗ = y ∧ ∗ G[0,5] x∗ ≥ y ) , (3) the first and the second occurrence of the variable x∗ relate to the first usage of ∗ and address the time t = 0, in which the formula is interpreted. The third occurrence of x∗ , in contrast to the previous two, 78 On Expressing and Monitoring Oscillations relates to the second operator ∗ and addresses the time instant in which the subformula on the right side of UI becomes true. Semantics restricted in this way does not excessively limit the capability of expressing various biological behaviour while remaining computationally feasible. 4 Algorithm In this section we deal with monitoring of an STL* formula, which means to determine the satisfaction of a formula over a finite length signal. The monitoring procedure introduced in this paper is inspired by constrained LTL model-checking [9] and monitoring of STL formulae [17], but needed to be extended into 2D space due to freeze operator ∗. It is because the task is solved simultaneously for two time instants, actual time and frozen time. The idea of the monitoring procedure is to construct a parse tree of the formula and check the satisfaction in a bottom-up manner. First step in checking a formula ϕ over a signal s is to construct a set of time points in which the formula ϕ is satisfied. Definition 4.1 Let ϕ be a STL* formula and s be a real-valued signal. A set Sϕ,s = {(t,t ∗ ) ∈ [0, \|s\|] × [0, \|s\|] \| (s,t,t ∗ ) \|= ϕ} is called the satisfaction set of formula ϕ over signal s. We use short-form notation Sϕ whenever the signal s is obvious from the context. Now we inductively construct satisfaction sets for nodes on higher levels of the parse tree. The last step of the procedure is to decide if s \|= ϕ from satisfaction set for the formula ϕ. According to Definition 3.1, it is equivalent to checking whether the satisfaction set contains the point (0, 0). Constructing the satisfaction set for a general signal can be a very difficult task. For this reason we will consider the monitoring algorithm only for piecewise linear signals. This is a reasonable requirement because in most cases we deal with time series produced by numerical simulations of modelled systems or by some measurements. These series can be interpreted as piecewise linear signals considering the values changing linearly between two adjacent points. If required, other points can be generated in between the existing points to make the signal more precise. Definition 4.2 A real-valued signal s of order n is called piecewise linear signal iff all the projections si (t) : [0, \|s\|] → R, i = 1, . . . , n are piecewise linear functions defined on a finite set of intervals J = T S {I1 , I2 , . . . , Im } where I j ⊆ [0, \|s\|], j = 1, . . . , m and J = 0, / J = [0, \|s\|]. Theorem 4.1 (Inductive construction of the satisfaction set) Let s be a piecewise linear signal and ϕ a formula in STL. The satisfaction set Sϕ of the formula ϕ over the signal s can be constructed inductively with respect to the structure of the formula ϕ: Sµ S¬ϕ Sϕ1 ∨ϕ2 Sϕ1 U[a,b] ϕ2 S∗ϕ + ∗ {(t,t ∗ ) ∈ R+ 0 × R0 \| µ(s,t,t ) = 1}; + R+ 0 × R0 \ Sϕ ; Sϕ1 ∪ Sϕ2 ; {(t,t ∗ ) ∈ Sϕ1 \| ∃t 0 ∈ [t + a,t + b] : (t 0 ,t ∗ ) ∈ Sϕ2 ∧ ∀t 00 ∈ [t,t 0 ] : (t 00 ,t ∗ ) ∈ Sϕ1 }; + + ∗ = {(t,t ∗ ) ∈ R+ 0 × R0 \| (t,t) ∈ Sϕ ,t ∈ R0 }. = = = = Proof Each of the equations follows directly from the definition of semantics of STL (1). P. Dluhos et al. 79 The assumption of piecewise linear signals in combination with linearity of atomic predicates (Definitions 2.3, 2.4) enables us to construct satisfaction sets for atomic predicates in polynomial time and to easily compute sets belonging to higher formulae. The formalism we use for the representation of satisfaction sets are convex polytopes. Definition 4.3 Convex polytope P ⊆ Rn is a set {x = (x1 , x2 , . . . , xn ) ∈ Rn×1 \| Ax ≤ b} where A ∈ Rm×n is a matrix and b ∈ Rm×1 is a column. A convex polytope in R1 is called a line segment, in R2 a convex polygon. Each inequality ∑nj=1 ai j x j ≤ bi in Definition 4.3 identifies a subspace of Rn . The convex polytope is obtained as intersection of these subspaces. Theorem 4.2 Assume s a piecewise linear signal and J = {I1 , . . . , Im } the set of intervals on which s is defined. Let µ be an atomic predicate over the signal s. The predicate µ(s,t,t ∗ ) = ∑ni=1 ai si (t) + ∑ni=1 bi si (t ∗ ) ∼ b, ∼∈ {<, ≤, >, ≥, =} is a linear function over the signal s in variables t and t ∗ according df to Definition 2.4. Denote Si,+j = {(t,t ∗ ) ∈ Ii × I j \| µ(s,t,t ∗ ) = 1} where i, j ∈ {1, . . . m}, Ii , I j ∈ J the set of time instants at which the atomic predicate µ is satisfied over signal s on rectangular area Ii × I j and df Si,−j = (Ii × I j ) \ Si,+j its complement. For t ∈ Ii ,t ∗ ∈ I j there can arrive two situations: 1. if ∼ ≡ =, then the set Si,+j makes a line segment (possibly degenerated to a single point or an empty set); 2. if ∼ 6≡ =, then the space Ii × I j is divided by the line ∑ni=1 ai si (t) + ∑ni=1 bi si (t ∗ ) = b into two subspaces Si,+j and Si,−j (Figure 2). (Again, there might be a degenerated case when the line ∑ni=1 ai si (t) + ∑ni=1 bi si (t ∗ ) = b does not cross the rectangle {(t,t ∗ ) ∈ Ii × I j }. In this case Si,+j = 0/ and Si,−j = Ii × I j or vice versa). Proof Described properties result from the linear form of atomic predicates 2.4 and from the fact that signal s behaves linearly on each interval Ii ∈ J , i ∈ {1, . . . m}. Figure 2: An illustration of the set Si,+j for a formula x∗ < 32 x + 28 3 on the subspace Ii ×I j = [0, 4]×[10, 13]. In both cases, the set Si,+j makes a convex set easily describable by set operations over convex polytopes. In the first case, a line segment itself is a convex polytope. In the second case, either the set Si,+j is a convex polytope, or (when ∼∈ {<, >}) it is a convex polytope after subtraction of one bordering line segment, which is a polytope. S Using this algorithm we can express the satisfaction set Sµ as Sµ = i, j∈{1,...,m} Si,+j , where Si,+j are the sets described above constructed by set operations on polytopes. The set Sµ can be further simplified by 80 On Expressing and Monitoring Oscillations joining adjacent polytopes. Having satisfaction sets for atomic predicates, we could inductively construct satisfaction sets for composite formulae from Theorem 4.1 again as operations on polytopes, which could be achieved in polynomial time [8]. However, working with precise satisfaction sets as defined in Theorem 4.1 can be quite a difficult task. We have to work with polytopes of different dimensions (polygons, lines and points) and operations over sets of these objects can be unnecessarily demanding. In real-world applications, we could end up performing an expensive monitoring over noisy signals measured on imperfect devices or computed on computers with finite numerical precision. Hence think about less precise, yet faster monitoring algorithm. We can imagine a noisy signal as a signal measured or computed with finite precision. The actual values of the variables can differ from the values presented by the signal up to some error ε > 0. This inaccuracy in the domain of values implies also inaccuracy in the time domain. When investigating the time instant in which some event occurred (e.g., x > 0) we cannot be sure when exactly it happened. That is because the critical value x = 0 is burdened by an error and it is possible that x = 0, but also that x > 0 or x < 0 in that particular time instant. Imagine we are in the situation from Theorem 4.2, but working with a noisy signal. We cannot be sure about the border of the set Si,+j due to the error ε. The condition ∑ni=1 ai si (t) + ∑ni=1 bi si (t ∗ ) = b defining points on the border of Si,+j could be easily broken by changing the values of the signal s by arbitrary small values from (0, ε). For this reason, we will ignore these border points and count them as they were in Si,+j or in Si,−j depending on what would be computationally easier. As a consequence, in the first case in Theorem 4.2, Si,+j = 0, / and in the second case, it does not matter if ∼ ≡ < or ∼ ≡ ≤. As a result of this simplification, the set of allowed operators in atomic predicates is restricted to {<, >}. The remaining three operators lost their sense. With signals burden with some error ε > 0 the satisfaction of atomic predicates can be meaningfully determined only for time instants inside the satisfaction set. Hence the satisfaction of a formula using the operator =, e.g., x = b or x = y + 2, cannot be determined. Every reference to a precise value has to be replaced by a sufficiently large interval, e.g., formula x = b can be replaced by x ≥ (b − δ ) ∧ x ≤ (b + δ ) for some δ > ε/2. Based on these assumptions, monitoring can be solved approximately. For every pair of intervals Ii , I j ∈ J , the satisfaction set on this area can be described by a single convex polygon Si,+j . For the S whole satisfaction set Sµ,s we get Sµ,s = i, j∈{1,...,m} Si,+j . The problem of construction of satisfaction sets for composite formulae from Theorem 4.1 is reduced to operations with sets of convex polytopes in plane for which efficient algorithms exist [8]. Algorithm 4.1 (Approximative monitoring algorithm for piecewise linear signals) Input: A piecewise linear signal s and an STL* formula ϕ. Output: Answer to the question s \|= ϕ. Algorithm inductively constructs the satisfaction set Sϕ : All the computations are performed on parts of the signal s of sufficient length. This can be computed according to (2). If the signal does not have the sufficient length, the answer returned by the algorithm might be wrong. 1. Sµ = i, j∈{1,...,m} Si,+j . The computation of the satisfaction sets Si,+j for individual intervals is performed according to Theorem 4.2, but without determining the satisfaction of the borders (as was justified in this section). S P. Dluhos et al. 81 2. S¬ϕ = (J × J ) \ Sϕ . The result can be computed as a simple Boolean operation on polygons. It is necessary to ensure that the resulting set consists only of convex polygons, which could be achieved for example by triangulation [8]. 3. Sϕ1 ∨ϕ2 = Sϕ1 ∪ Sϕ2 . Again a Boolean polygonal operation. + + ∗ 4. S∗ϕ = {(t,t ∗ ) ∈ R+ 0 × R0 \| (t,t) ∈ Sϕ ,t ∈ R0 }. The satisfaction set for the freeze operation can be computed by: (1) finding an intersection between the line t ∗ = t and the satisfaction set Sϕ (black line segments in Figure 3a), (2) substituting values in the second component by the whole axis t ∗ , i.e., making a projection to the first component t and then a Cartesian product with the axis t ∗ (Figure 3b). (a) (b) Figure 3: Example of computing satisfaction set for the formula ∗ϕ. a) The satisfaction set represented as a set of (overlapping) convex polygons (different colors are used only for depicting the fact that the set consists of convex polygons) and the line t = t ∗ . The regions of intersection of the line and the satisfaction set are depicted in black color. b) Resulting set S∗ϕ obtained by projecting the intervals of intersection to the axis t and making a Cartesian product with the axis t ∗ . 5. Sϕ1 U[a,b] ϕ2 = {(t,t ∗ ) ∈ Sϕ1 \| ∃t 0 ∈ [t + a,t + b] : (t 0 ,t ∗ ) ∈ Sϕ2 ∧ ∀t 00 ∈ [t,t 0 ] : (t 00 ,t ∗ ) ∈ Sϕ1 }. To be more illustrative, we explain this part of the algorithm from the geometrical point of view, identifying values t with the horizontal axis and values t ∗ with the vertical axis. (a) Sϕ1 (b) Sϕ2 (c) Sϕ1 ∩ Sϕ2 Figure 4: First step of computation of Sϕ1 U[a,b] ϕ2 is to find the intersection Sϕ1 ∩ Sϕ2 . For each convex polygon P ∈ Sϕ1 ∩ Sϕ2 (Figure 4c) the following procedure is performed: 82 On Expressing and Monitoring Oscillations (a) P ∪ Sϕ1 (b) ordered vertices (c) rectangular polygons (stripes) Ri , i = 1, . . . , n − 1 Figure 5: Next, the space is divided into horizontal stripes. The task is solved in each stripe separately. (a) Vertices of polygons {P} ∪ Sϕ1 are increasingly ordered by the second component t ∗ (duplidf cates can be removed). Denote the ordered set V = (V1 ,V2 , . . .Vn ) (Figures 5a, 5b); ∗ )∈V (b) For i = 1, . . . , n − 1 every neighbouring pair of vertices Vi = (ti ,ti∗ ),Vi+1 = (ti+1 ,ti+1 specifies a rectangular polygon (stripe): df ∗ ∗ Ri = {(t,t ∗ ) ∈ R+ 0 × [ti ,ti+1 ]} (Figure 5c). Ri in a form of line segment is considered empty because we do not care about the border points. df (c) The task is solved in every stripe Ri separately. For every nonempty Ri denote Pi = Ri ∩ P and df Si = Ri ∩ Sϕ1 . Due to the way of construction of Ri , all polygons in Si and the single polygon ∗ ) or the lower (t ∗ = t ∗ ) border of the stripe Pi have all their vertices on the upper (t ∗ = ti+1 i Ri (Figure 6a). In fact, these polygons can take only the shape of a triangle or a trapezoid (Figure 7). (a) the stripe R4 → − (b) points on the left side of the line ul (c) A4 = the rightmost polygon in S40 Figure 6: The potential polygons are identified. (The solution cannot be placed further in time than the property ϕ2 . Moreover, the property ϕ1 must continuously hold at all time points until ϕ2 is satisfied.) Now we get rid of the polygons lying on the right side of Pi because their points cannot be in the solution (they represent the time past the event ϕ2 ). To do so, we isolate the upper and lower rightmost vertices of the polygon Pi . Denote them as u = (ut , ut ∗ ) and l = (lt , lt ∗ ) (Figure 6b). We get a new area: → − df R0i = {(t,t ∗ ) ∈ Ri \| (t,t ∗ ) lies on the left side of the line ul }. P. Dluhos et al. 83 df The solution can lie only in this area, so we can restrict Si to Si0 = R0i ∩ Si . Figure 7: An example of a stripe and all possible shapes of polygons inside a stripe. (d) All the adjoining polygons (sharing an edge) in Si0 are connected to form maximal seamless convex polygons (e.g., the green and orange polygons in Figure 6b). (e) Let Ai be the rightmost polygon in Si0 (Figure 6c) (Even after connecting some polygons in Si0 their shape can still be only triangular or trapezoidal (Figure 7) so there is only one rightmost polygon in Si0 ). Then A0i = Ai ∩ (Pi (b, a)) is the solution for the stripe Ri (Figure 8). df The operation is defined as A (a, b) = {(x, y) ∈ R × R \| ∃c ∈ (a, b) : (x + c, y) ∈ A} which is equal to A ⊕ (−a, −b), where ⊕ is the Minkowski sum. S 0 (f) The final satisfaction set is given by Sϕ1 U[a,b] ϕ2 = n−1 i=1 Ai . (a) P4 before application of the operator (b) the shifted polygon P4 (2.5, 2.5) (c) A04 = A4 ∩ (P4 (2.5, 2.5)), the final solution in the stripe R04 Figure 8: Last step is to consider the interval [a, b] bounding the operator U. 6. The result s \|= ϕ which is equivalent to (0, 0) ∈ Sϕ is returned. Remark 4.1 The satisfaction set for temporal operator future F[a,b] ϕ ≡ trueU[a,b] ϕ (and hence G[a,b] ϕ) can be computed more easily than according to the step 5. It can be obtained directly as S F[a,b] ϕ = Sϕ (b, a). 4.1 Time and space complexity Time and space complexity of all steps of the Algorithm 4.1 is proportional to the number of polygons they are working with. Number of polygons generated in step 1 is at most n2 for each atomic predicate, where n is the number of intervals on which the signal is defined. The number of polygons does not asymptotically change during the computation of operations described in other steps of the algorithm and all these functions do not work slower than O(n2 ) (see [8] for more details). The output polygons also keep small number of vertices during the process. The upper bound on the total computation time is therefore O(kn4 ), where n is the number of intervals on which the signal is defined and k is the size of the investigated formula. 84 5 On Expressing and Monitoring Oscillations Evaluation and Case Study A prototype of the approximative monitoring Algorithm 4.1 was implemented in Matlab (version R2011b). The Multi-Parametric Toolbox (MPT) package [16] (version 2.6.3) was used for the polygonal operations. We have not focused on efficiency of the implemented algorithm, our goal was to prove the concept of the presented algorithm. Detailed description of the implementation, results of the experiments and performance analysis can be found in [10]. We have studied a biological system of three transcriptional repressors which was designed and built into bacteria Escherichia coli [14]. Concentrations of involved proteins are periodically oscillating which causes periodic production of a green fluorescent protein. Intensity of fluorescence of this protein, which can be measured, gives evidence of the ongoing activity of the network (Figure 9). Figure 9: The fluorescent intensity of a colony of Escherichia coli containing the repressilator. The intensity is increasing overall due to the increasing number of individuals in the colony. The period of oscillations is lower than the time between cell division, but the whole repressilator is transmitted from generation to generation. Figure was taken from [14]. The network was modelled as a system of six differential equations: ṁ1 ṁ2 ṁ3 α = −m1 + 1+p n + α0 ; 3 α = −m2 + 1+pn + α0 ; 1 α = −m3 + 1+p n + α0 ; 2 p˙1 p˙2 p˙3 = −β (p1 − m1 ); = −β (p2 − m2 ); = −β (p3 − m3 ). (4) dmi df d pi , ṗi = ; m1 , m2 and m3 correspond to activity of genes lacl,tetR and cl; p1 , p2 dt dt and p3 are concentrations of corresponding proteins and α, β , α0 , n are constants (a typical behaviour of the system is depicted in Figure 10). The analysis of the system (4) and the estimation of parameters α, β , α0 , n producing the oscillatory behaviour was originally done by means of manual qualitative analysis of ODEs [14]. However, we can perform the analysis automatically using the monitoring of STL* formulae. We can specify the desired oscillatory property by the formula: df Where ṁi = df ϕ = G[10,190] F[0,50] ∗ ( F[1,50] m∗i < mi ) ∧ F[1,50] m∗i > mi ) (5) and test the satisfaction for different values of parameters on runs of the length at least 300 minutes. The expected period has to be lower than 50 minutes. We start the testing 10 minutes later after the beginning to avoid the initial swing and end the testing 50 minutes before the end of the measurement because the signal might be cut in the middle of a period. P. Dluhos et al. 85 Figure 10: A typical signal produced by the system (4). Only variables m1 , m2 and m3 are depicted. (a) α = 400, β = 0.2, α0 = 0.2, n = 2 (b) α = 400, β = 0.2, α0 = 2, n = 2 Figure 11: Sample runs of the system (4) for different sets of parameters. The initial values were m1 = 0.1, m2 = 0.3, m3 = 0.2, p1 = 0.2, p2 = 0.1, p3 = 0.3. Only the values of m1 are depicted. Formula (5) is satisfied by both runs in Figure 11. To avoid the case of damped oscillations, we can add the formula: df ψ = G[10,200] ∗ ( F[1,50] m∗i ≤ mi ) . (6) It ensures that the values reached in each period are not decreasing. By connecting formulae (5) and (6) we get the formula ϕ ∧ ψ. It does express the desired oscillatory behaviour in Figure 11a and it is not satisfied by signals of the type depicted in Figure 11b. Another property of the repressilator can be expressed by the formula: G[0,270] ∗ ( F[0,30] (m∗1 + 1 > m3 ∧ m∗1 − 1 < m3 ) , (7) which means that ”the values of m1 precede the values of m3 in such sense that for every value of m1 , m3 reaches similar value in short time after m1 did” (Figure 12). Full estimation of parameters could not be performed due to high time demands of the implemented monitoring algorithm [10]. A single run of the monitoring algorithm for the formulae (5), (6) or (7) over a signal sampled by 80 points took several hours on a regular PC. Reduction of the number of points describing the signal would lead to excessive loss of information. We have found out that the 86 On Expressing and Monitoring Oscillations Figure 12: Sample run of the system (4) satisfying the formula (7). The parameters were α = 400, β = 0.2, α0 = 0.2, n = 2 and the initial values m1 = 0.1, m2 = 0.3, m3 = 0.2, p1 = 0.2, p2 = 0.1, p3 = 0.3. Depicted variables are m1 – red and m3 – green. bottleneck in efficiency of the prototype implementation lies in polyhedral operations performed by MPT. Since we work in 2D and need only a small subset of operations, we believe that our algorithm can be significantly accelerated when an optimal implementation is employed. Here we focused on demonstrating the applicability while leaving efficiency for future work. Different task would be to ensure that the behaviour of concentrations of the fluorescent protein in the bacteria (Figure 9) is in correspondence with the model. While it could be done only manually in [14], if we had the data, we would be able to test the properties of the measurements identically as the properties of the signals produced by ODEs. 6 Conclusion We have proposed an extended Signal Temporal Logic STL∗ motivated by the need to express properties of biological dynamical systems in a detail sufficient to distinguish different shapes of oscillation. We have provided a monitoring algorithm that approximately computes the truth value of an STL∗ formula for a given continuous (piece-wise linear) signal. The method has been prototyped in Matlab and the results achieved on a case study of oscillatory behaviour of the repressilator showed that the method satisfactorily works for signals generated by numerical simulation. However, the employed library MPT for polyhedral operations appears to be not efficient enough to satisfy the needs of practical usage. For future work on the practical side, we plan to implement more efficient algorithms for the specific polyhedral operations we use in the monitoring procedure. Another straightforward direction of future development is lifting of the robustness measure [11] to the extended logic. A very inspiring work is [12] where Time-Frequency Logic (TFL) is defined by shifting STL semantics to frequency domain. TFL provides another way to express permanent oscillations. However, nonpure oscillatory behaviour such as damped oscillations require specific elaboration in the time-domain. Joining TFL semantics with the STL∗ concept of real-value freezing can be an interesting step further. References [1] Accelera Organization, Inc. (2004): SystemVerilog 3.1a Language Reference Manual. [2] R. Alur, T. Feder & T. A. Henzinger (1996): The benefits of relaxing punctuality. J ACM 43(1), pp. 116–146, doi:10.1145/227595.227602. [3] R. Alur & T. A. Henzinger (1994): A really temporal logic. J ACM 41(1), pp. 181–203, doi:10.1145/174644.174651. P. Dluhos et al. 87 [4] P. Ballarini & M. L. Guerriero (2010): Query-based verification of qualitative trends and oscillations in biochemical systems. Theor. Comput. Sci. 411(20), pp. 2019–2036, doi:10.1016/j.tcs.2010.02.010. [5] J. Barnat, L. Brim & D. Šafránek (2010): High-performance analysis of biological systems dynamics with the DiVin E model checker. Brief. in Bioinformatics 11, pp. 301–312, doi:10.1093/bib/bbp074. [6] E. Bartocci, F. Corradini, E. Merelli & L. Tesei (2010): Detecting synchronisation of biological oscillators by model checking. Theoretical Computer Science 411(20), pp. 1999 – 2018, doi:10.1016/j.tcs.2009.12.019. [7] G. Batt, D. Ropers, H. de Jong, J. Geiselmann, R. Mateescu, M. Page & D. Schneider (2005): Validation of qualitative models of genetic regulatory networks by model checking: analysis of the nutritional stress response in Escherichia coli. In: ISMB (Supplement of Bioinformatics), pp. 19–28. [8] M. de Berg, O. Cheong, M. van Kreveld & M. Overmars (2008): Computational Geometry: Algorithms and Applications, 3 edition. Springer, Berlin. [9] L. Calzone, N. Chabrier-rivier, F. Fages & S. Soliman (2006): Machine learning biochemical networks from temporal logic properties. Trans Comput Syst Biol 4220, pp. 68–94, doi:10.1007/11880646 4. [10] P. Dluhoš (2012): Specification and monitoring of oscillation properties in dynamical systems. Master’s thesis, Masaryk University. Available at http://is.muni.cz/th/269281/fi_m/thesis.pdf. [11] A. Donzé & O. Maler (2010): Robust satisfaction of temporal logic over real-valued signals. In: FORMATS’10, Springer-Verlag, Berlin, Heidelberg, pp. 92–106, doi:10.1007/978-3-642-15297-9 9. [12] A. Donzé, O. Maler, E. Bartocci, D. Nickovic, R. Grosu & S. Smolka (2012): On Temporal Logic and Signal Processing. Accepted to ATVA 2012. [13] C. Eisner & D. Fisman (2008): Augmenting a regular expression-based temporal logic with local variables. In: FMCAD ’08, IEEE Press, pp. 23:1–23:8, doi:10.1.1.140.3930. [14] M. B. Elowitz & S. Leibler (2000): A synthetic oscillatory network of transcriptional regulators. Nature 403(6767), pp. 335–338, doi:10.1038/35002125. [15] B. N. Kholodenko (2000): Negative feedback and ultrasensitivity can bring about oscillations in the mitogen-activated protein kinase cascades. Eur J Biochem 267(6), pp. 1583–1588, doi:10.1046/j.14321327.2000.01197.x. [16] M. Kvasnica, P. Grieder & M. Baotić (2004): Multi-Parametric Toolbox (MPT). [17] O. Maler & D. Nickovic (2004): Monitoring temporal properties of continuous signals. In: Proc. of FORMATS-FTRTFT, Springer, pp. 152–166, doi:10.1.1.102.2905. [18] O. Maler, D. Nickovic & A. Pnueli (2008): Checking Temporal Properties of Discrete, Timed and Continuous Behaviors. In: Pillars of Computer Science, LNCS 4800, Springer, pp. 475–505, doi:10.1007/978-3-54078127-1 26. [19] R. Mateescu, P. T. Monteiro, E. Dumas & H. de Jong (2011): CTRL: Extension of CTL with regular expressions and fairness operators to verify genetic regulatory networks. Theor. Comput. Sci. 412(26), pp. 2854–2883, doi:10.1016/j.tcs.2010.05.009. [20] F. Miyoshi, Y. Nakayama, K. Kaizu, H. Iwasaki & M. Tomita (2007): A Mathematical Model for the KaiProtein-Based Chemical Oscillator and Clock Gene Expression Rhythms in Cyanobacteria. Journal of Biological Rhythms 22(1), pp. 69–80, doi:10.1177/0748730406295749. [21] D. Nickovic & O. Maler (2007): AMT: a property-based monitoring tool for analog systems. In: Proceedings of the 5th international conference on Formal modeling and analysis of timed systems, FORMATS’07, Springer-Verlag, Berlin, Heidelberg, pp. 304–319. [22] A. Rizk, G. Batt, F. Fages & S. Soliman (2008): On a Continuous Degree of Satisfaction of Temporal Logic Formulae with Applications to Systems Biology. In: Proc. of CMSB’08, Springer, pp. 251–268, doi:10.1007/978-3-540-88562-7 19. [23] J. Červený & L. Nedbal (2009): Metabolic Rhythms of the Cyanobacterium Cyanothece sp. ATCC 51142 Correlate with Modeled Dynamics of Circadian Clock. J Biol Rhythms 24(4), pp. 295–303, doi:10.1177/0748730409338367.	5cs.CE
The Lottery Ticket Hypothesis: Training Pruned Neural Networks arXiv:1803.03635v1 [cs.LG] 9 Mar 2018 Jonathan Frankle MIT CSAIL jfrankle@mit.edu Michael Carbin MIT CSAIL mcarbin@mit.edu Abstract Recent work on neural network pruning indicates that, at training time, neural networks need to be significantly larger in size than is necessary to represent the eventual functions that they learn. This paper articulates a new hypothesis to explain this phenomenon. This conjecture, which we term the lottery ticket hypothesis, proposes that successful training depends on lucky random initialization of a smaller subcomponent of the network. Larger networks have more of these “lottery tickets,” meaning they are more likely to luck out with a subcomponent initialized in a configuration amenable to successful optimization. This paper conducts a series of experiments with XOR and MNIST that support the lottery ticket hypothesis. In particular, we identify these fortuitously-initialized subcomponents by pruning low-magnitude weights from trained networks. We then demonstrate that these subcomponents can be successfully retrained in isolation so long as the subnetworks are given the same initializations as they had at the beginning of the training process. Initialized as such, these small networks reliably converge successfully, often faster than the original network at the same level of accuracy. However, when these subcomponents are randomly reinitialized or rearranged, they perform worse than the original network. In other words, large networks that train successfully contain small subnetworks with initializations conducive to optimization. The lottery ticket hypothesis and its connection to pruning are a step toward developing architectures, initializations, and training strategies that make it possible to solve the same problems with much smaller networks. 1 Introduction Recent work on neural network pruning (e.g., [4, 3, 2]) indicates that neural networks can be dramatically simplified once trained. After training is complete, upwards of 90% of weights can be pruned without reducing accuracy. If a network can be so pruned, then the function that it learned could have been represented by a far smaller network than that used during training. However, researchers believe1 that these smaller networks cannot be trained as readily as their larger counterparts, in spite of the fact that they are demonstrably capable of representing the desired functions. In this paper, we contend that it is indeed possible to train these smaller networks directly. In fact, these small, trainable networks are embedded within the larger models that we typically train. This paper articulates a possible explanation for the disconnect between a neural network’s representation capacity and its trainability. This conjecture, which we term the lottery ticket hypothesis, 1 [4] mentions that “CNNs contain fragile co-adapted features” and that “gradient descent is able to find a good solution when the network is initially trained, but not after re-initializing some layers and retraining them...When we retrain the pruned layers, we should keep the surviving parameters instead of re-initializing them.” states that training succeeds for a given network if one of its subnetworks has been randomly initialized such that it could be trained in isolation—independent of the rest of the network—to high accuracy in at most the number of iterations necessary to train the original network. We refer to these fortuitously-initialized networks as winning tickets. Subnetworks and winning tickets. From the perspective of the lottery ticket hypothesis, a network’s initialization procedure can be thought of as drawing many samples from a distribution over initialized subnetworks. Ideally, the procedure manages to draw a subnetwork with the right architecture and weight initializations for optimization to succeed (a winning ticket). If a network of size m (where size is measured in units or weights) is being trained to solve a problem but a network of size n (where n ≤ m) is sufficient to represent the function to be learned, then the lottery ticket hypothesis views the original network as containing m n overlapping subnetworks. If the larger network is able to train successfully, it is because one of these subnetworks lucked into an initialization amenable to optimization. Metaphorically, training a network larger than is necessary to represent the function to be learned is like buying many lottery tickets. Larger networks have combinatorially more subcomponents that could facilitate successful training (i.e., more lottery tickets). The initialization strategy determines which of these subcomponents are well-situated for optimization to succeed (i.e., which tickets are winners). If a subcomponent is initialized favorably (i.e., the network picked a winning ticket), training succeeds. Identifying winning tickets. In this paper, we demonstrate that it is possible to automatically identify winning tickets by making a small but critical modification to the experiment of Han et al. [4]. We prune a trained neural network’s smallest weights (as measured by their magnitudes after training) in the same manner as Han et al.; the set of connections that survives this pruning process is the architecture of a winning ticket as anticipated by the lottery ticket hypothesis. Unique to our work, the winning ticket’s weights are the values to which these connections were initialized before training began. Where Han et al. aimed to compress networks during the training process, our goal is to find small networks that can be trained independently from the start. We show that a winning ticket extracted in this fashion, when initialized with the original weights from before training, can be trained successfully in isolation at least as fast as (and typically faster than) the full network. Methodology. To empirically assess the lottery ticket hypothesis, we use the following procedure to extract winning tickets from fully-connected networks of a variety of sizes for MNIST and, as an illustrative example, from small networks for XOR. This procedure is identical to Han et al.’s pruning process with the addition of the crucial last step: resetting the weights to their original values from before training. 1. Randomly initialize a neural network. 2. Train the network until it converges. 3. Prune a fraction of the network. 4. To extract the winning ticket, reset the weights of the remaining portion of the network to their values from (1) (i.e., the initializations they received before training began). If successful training really does rely on fortuitous initialization of a subcomponent of the network and pruning really does reveal this winning ticket, then the lottery ticket hypothesis predicts that the pruned network—when reset to the original initializations from before training—will train successfully at sizes too small for a randomly-initialized or a randomly-configured network to do so. Research questions. To test the lottery ticket hypothesis, we evaluate the following research questions: How effectively do winning tickets train in comparison to the original network and to randomly sampled networks of similar size? For a variety of network configurations and perturbations of winning tickets, we measure both convergence times and test accuracy once the network has converged. How big are winning tickets relative to the size of the original network? By training on networks of various sizes, we explore whether the size of a winning ticket remains constant for a particular learning problem or grows in proportion to the size of the larger network from which it is derived. 2 10 Units DB ZL 8 Units DB ZL 6 Units DB ZL 4 Units DB ZL 2 Units DB ZL 98.5 96.8 92.5 78.3 49.1 92.9 87.5 76.8 55.3 17.6 Figure 1: Success rates of 1000 random XOR networks, each with the specified number of hidden units. DB = percent of trials that found the correct decision boundary. ZL = percent of trials that reached zero loss. How sensitive are our results to particular pruning strategies? We test two broad classes of strategies: pruning hidden units with the lowest-magnitude incoming and/or outgoing weights (for XOR) and individually pruning the lowest-magnitude weights (for MNIST). We also study whether networks can be pruned in a single step or whether they must repeatedly be pruned, retrained, and reset in an iterative process. Results. Our experimental results support the lottery ticket hypothesis. XOR. We trained a simple network with one hidden layer to learn XOR. The minimal architecture capable of representing XOR, a hidden layer with two randomly-initialized units, reached zero loss 18% of the time. In contrast, when a network with ten hidden units that reached zero loss was iteratively pruned down to a two-unit winning ticket, the winning ticket reached zero loss 80% of the time when trained with its original initializations. MNIST. Up to a certain point, winning tickets derived by pruning converged faster than and at least as accurately as the original network; after this point, convergence times and accuracy gradually and then rapidly dropped off. In a single step, we could prune networks by 75% while still finding winning tickets that, on average, converged 38% faster than the original network and matched its accuracy. Pruning iteratively by 20% at a time, winning tickets 91% smaller than the original network converged on average 39% faster. Networks iteratively pruned by up to 98% on average still converged as fast as the original network while maintaining accuracy. When winning tickets were randomly reinitialized or their weights were randomly rearranged, convergence times increased and accuracy decreased as compared to the original network. Depending on the metric of winning ticket size, winning tickets grew either gradually or marginally with network size. Contributions and implications. • We propose the lottery ticket hypothesis as a new perspective on neural network training. • We further posit that pruning uncovers the winning tickets that the lottery ticket hypothesis predicts, leading to an algorithm for extracting winning tickets from trained networks. • We apply this algorithm to empirically evaluate these conjectures on small networks. The evidence we find supports both the lottery ticket hypothesis and our contention that pruning can extract winning tickets. Although this paper focuses mainly on measurement, it has important implications for our understanding of training. The increased representation power of large networks is not necessarily required for gradient descent to learn functions with small representations. Lurking within these large networks are small, fortuitously-initialized winning tickets that are both more efficient to train (as a product of their size) and faster to converge (as a product of their initialization). By examining the initalizations and architectures of successful winning tickets, we might find new ways of designing networks that are smaller but equally-capable (if not superior). 2 Learning the XOR Function The XOR function is among the simplest examples that distinguish neural networks from linear classifiers. Before presenting our results for MNIST, we summarize the lottery ticket hypothesis as it applies to this simple computation. The XOR function has four data points: the coordinates (0, 0), (0, 1), (1, 0), and (1, 1). The first and last points should be placed in class 0 and the middle two points in class 1. Geometrically, this problem requires a nonlinear decision boundary. In this experiment, we consider the family of fully connected networks for XOR with two input units, one hidden layer (ReLU activation), and one output unit (sigmoid activation). 3 Pruning Strategy 10 Units DB ZL 4 Units (Pruned) DB ZL 2 Units (Pruned) DB ZL One-shot Product 99.2 93.3 98.0 90.3 82.4 65.3 Input Magnitude 98.9 93.5 97.9 92.2 83.8 76.5 Output Magnitude 99.0 93.3 96.9 85.9 78.6 56.1 Product 98.5 92.9 97.6 90.3 91.5 79.4 Figure 2: Success rates of different pruning strategies on 1000 trials each. DB and ZL defined as in Figure 1. The pruned columns include only those runs for which both the original ten-unit network and the pruned winning ticket found the right decision boundary or reached zero loss. The first row of the table was obtained by pruning in one shot; all subsequent rows involved pruning iteratively Although a network of this form with two hidden units is sufficient to perfectly represent the XOR function,2 the probability that a standard training approach—one that randomly initializes the network’s weights and then applies gradient descent—correctly learns XOR for a network with two hidden units is low relative to that for a larger network. Figure 1 contains the overall success rates (percent of networks that found the right decision boundary or reached zero loss). In 1000 training runs, a network with two hidden units learned a correct decision boundary in only 49.1% of trials. Cross-entropy loss reached 0 (meaning the network learned to output a hard 0 or 1) in only 17.6% of trials. Meanwhile, an otherwise identical network outfitted with ten hidden units learned the decision boundary in 98.5% of trials and reached 0 loss in 92.9% of trials. Figure 1 charts the loss for these and other hidden layer sizes.3 To put the central question of this paper in the concrete terms of the XOR problem, why do we need to start with a neural network with ten hidden units to ensure that training succeeds when a much smaller neural network with two hidden units can represent the XOR function perfectly? We propose the lottery ticket hypothesis as an explanation for this phenomenon. The Lottery Ticket Hypothesis. Training succeeds for a given network if one of its subnetworks (a “winning ticket”) has been randomly initialized such that it can be trained in isolation to high accuracy in at most the number of iterations necessary to train the original network. According to the lottery ticket hypothesis, successful networks with a large number of parameters (e.g., the XOR network with ten hidden units) should contain winning tickets comprising a small number of fortuitously-initialized weights on which training will still succeed. 2.1 Methodology To test the lottery ticket hypothesis with the XOR function, we instantiated the experiment from Part 1 with the following details: 1. Randomly initialize a network with ten hidden units. 2. Train for 10,000 iterations on the entire training set. 3. Prune a certain number of hidden units according to a particular pruning heuristic. 4. To extract the winning ticket, reset the pruned network to the original initializations. The first three steps extract the architecture of the winning ticket; the crucial final step extracts the corresponding initializations. We ran this experiment with two different classes of pruning strategies. One-shot pruning involves pruning the network in a single pass. For example, one-shot pruning a network by 80% would involve removing 80% of its units after it has been trained. In contrast, iterative pruning involves repeating steps (2) through (4) several times, removing a small portion of n −n 1 Example satisfying weights for the first layer: −n n . Satisfying weights for the output unit: [ 1 ]. Satisfying bias for the output unit: −n/2. n ≥ 1. As n grows, the output approaches a hard 1 or 0. 3 Weights were sampled from a normal distribution centered at 0 with standard deviation 0.1; all values more than two standard deviations from the mean were discarded and resampled. Biases were initialized to 0. The network was trained for 10,000 iterations. 2 4 the units (in our case, two units) on each iteration. We find iterative pruning to be more effective for extracting smaller winning tickets; Han et al. [4] found the same for compressing large networks while maintaining accuracy. We consider three different heuristics for determining which hidden units should be pruned: • Input magnitude: remove the hidden unit with the smallest average input weight magnitudes. • Output magnitude: remove the hidden unit with the smallest output weight magnitude. • Magnitude product: remove the hidden unit with the smallest product of the magnitude of its output weight and the sum of the magnitudes of its input weights. The magnitude product heuristic achieved the best results, so we use it unless otherwise stated. 2.2 Results One-shot pruning. We generated 1000 networks with ten hidden units and pruned them down to both four and two hidden units using the magnitude product heuristic. The results of doing so appear in the first row of Figure 2. The winning tickets with two hidden units found the correct decision boundary 82.4% of the time (up from 49.1% for randomly-initialized networks with two hidden units) and reached zero loss 65.3% of the time (up from 17.6% of the time for a random network). Iterative pruning. We conducted the iterative version of the pruning experiment 1000 times, starting with networks containing ten hidden units that were eventually pruned down (in two unit increments) to networks containing a candidate winning ticket of just two hidden units. Of the 93.5% of ten hidden unit networks that reached zero loss, 84.9% 4 had a two hidden unit winning ticket that also reached zero loss (as compared to 17.6% of randomly-intialized two hidden unit networks). Likewise, of the 98.9% of ten hidden unit networks that found the correct decision boundary, 92.8% had a two-unit winning ticket that did the same (as compared to 49.1% of randomly-initialized two hidden unit networks). The four hidden unit winning tickets almost identically mirror the performance of the original ten hidden unit network. They found the correct decision boundary and reached zero loss respectively in 99% and 97% of cases where the ten hidden unit network did so. Both of these pruned trials appear in Figure 2 (under the Magnitude Product row). These experiments indicate that, although iterative pruning is more computationally demanding than one-shot pruning, it finds winning tickets at a higher rate than one-shot pruning. More importantly, they also confirm that networks with ten hidden units can be pruned down to winning tickets of two hidden units that, when initialized to the same values as they were in the original network, succeed in training far more frequently than a randomly initialized network with two hidden units. The winning tickets with four hidden units succeed nearly as frequently as the ten unit networks from which they derive. Both of these results support the lottery ticket hypothesis—that large networks contain smaller, fortuitously-initialized winning tickets amenable to successful optimization. In addition to the magnitude-product pruning heuristic, we also tested the input magnitude and output magnitude heuristics. The results of doing so appear in Figure 2. The magnitude product heuristic outperformed both. We posit that this success is due to the fact that, in the XOR case when all input values are either 0 or 1, the product of input and output weight magnitudes should mimic the activation of the unit (and therefore with its influence on the output). 3 MNIST (One-shot Pruning) In this section and those that follow, we explore the lottery ticket hypothesis as applied to the MNIST dataset. Here, we analyze the behavior of one-shot pruning; in the following section, we show the additional power that iterative pruning offers. 4 These numbers are derived from the last row of Figure 2. 93.5% of networks with ten hidden units reached zero loss. 79.4% of networks started with ten units, reached zero loss, and were pruned to into two-unit networks that also reached zero loss. 79.4% of 93.5% is 84.9%. 5 3.1 Methodology We trained and pruned a network with two fully-connected layers. We used the LeNet-300-100 architecture [6], which has 784 input units (corresponding to the pixels of the 28x28 images in MNIST), a fully-connected hidden layer with 300 units, a fully-connected hidden layer with 100 units, and ten fully-connected output units (one for each class). The hidden units have ReLU activation functions, and the output units have softmax activation functions. By default, biases were initialized to 0 and weights were randomly sampled from a normal distribution with mean 0 and standard deviation 0.1 (values more than two standard deviations from the mean were discarded and resampled). Networks were optimized using stochastic gradient descent with a learning rate of 0.05. This section follows the experimental template from Section 1: 1. Randomly initialize the network. 2. Train for 50,000 iterations on 100-example mini-batches from the training data 3. Prune a certain percentage of the weights from within each hidden layer, removing those with the lowest magnitudes. 4. To extract the winning ticket, reset the values of the weights of the pruned network to their original initializations from before training. The pruning strategy we follow for MNIST removes individual weights rather than entire units. In preliminary experiments, we found this strategy to be more effective (Srinivas and Babu explore pruning by unit in [12]). We use the simplest weight-by-weight pruning heuristic possible: remove those weights with the lowest magnitudes within each hidden layer (just as in [4]). Weights connecting to the output layer are pruned by half of the percentage at which the rest of the network is pruned to avoid severing connectivity to the output units. 3.2 Results Figure 3: The test set accuracy on MNIST as training proceeds. These charts are zoomed into the very highest levels of accuracy. Each curve shows the average progression of five trials of training at the specified pruning level. Percents are the percent of the weights in each layer that remain after pruning. The error bars show the minimum and maximum values of any one of those five trials. Dots signify the moment when the corresponding colored line has converged, with error bars showing the earliest and latest convergence times amongst the five trials. Pruning’s most substantial impact was on convergence times. When pruned to between 75% and 25% of the size of the original network, the winning tickets converged on average at least 25% faster while accuracy remained on average within 0.15% of the original network’s accuracy. The winning ticket that was pruned to 25% of the original size converged on average 38% faster than the original network. Further pruning caused convergence times to slowly rise and accuracy to drop. Figure 3 shows the test set accuracy and convergence behavior of winning tickets pruned to different levels as they were trained.5 Each curve is the average of five different runs starting from distinct, 5 We define convergence as the moment at which the 1000-iteration moving average of test accuracy changed by less than 0.002 for 1000 consecutive iterations. We measured test accuracy every 100 iterations. According to this definition of convergence, the one-shot-pruned winning tickets improved their test accuracy by an average of 0.0019 (standard deviation 0.00085) after convergence. We acknowledge that determining convergence times is an imprecise art, but this metric seems to adequately characterize the behavior of convergence for our purposes. 6 Figure 4: The test set accuracy on MNIST as training proceeds for winning tickets of various sizes and for winning tickets whose weights have been randomly reinitialized (control experiment 1). randomly initialized networks; error bars indicate the minimum and maximum value that any run took on at each point in the training process. The dots indicate the average convergence times for the curve in the corresponding color; error bars indicate the minimum and maximum convergence times. The left graph in Figure 3 shows that, for the first few pruning levels, convergence times decrease and accuracy increases. A winning ticket comprising 90% of the weights from the original network converges slightly faster than the original network but slower than a winning ticket with 75% of the original weights. This pattern continues until the network is pruned to about 55% of the original size, after which (as the right graph in Figure 3 shows) convergence times flatten and then, after about 35%, increase. When the winning ticket is between 10% and 15% of the original size of the network, it returns to the performance of the unpruned network. In the terms of the lottery ticket hypothesis, we attribute improving convergence times to the removal of unnecessary, noisy parts of the network as pruning hones in on the winning ticket. Convergence times reach a tipping point as pruning begins to remove weights that are essential to the winning ticket, after which convergence times increase and accuracy decreases. The lottery ticket hypothesis also predicts that this behavior is largely attributable to a confluence of initialization and architecture. To test this conjecture, we ran two control experiments: (1) retain the winning ticket’s architecture but randomize its weights and (2) retain the winning ticket’s weights but randomize its architecture. Control experiment 1. This experiment evaluates the extent to which initialization is a necessary component of a winning ticket. Figure 4 shows this experiment. The curves for the original network and winning tickets that are 75% and 35% of the original network’s size are the same as in Figure 3. Two curves have been added for the control experiments. Each control experiment entailed training a network that used a winning ticket’s architecture but randomly reinitialized its weights from the original initialization distribution (N (0, 0.1)). We trained three control experiments for each winning ticket, so the control curves are the average of 15 experiments. Unlike the winning tickets, the control experiments converged on average more slowly than the original network, simultaneously achieving lower levels of accuracy. These differences were substantial: the average 35% and 25% winning tickets converged 1.91 and 2.28 times as fast as the corresponding average controls. The error bars on convergence times reflect that the control trials exhibited a much wider variance in behavior. For example, the earliest-converging of the 35% control trials converged faster than the average unpruned network; however, the average 35% control trial convergence time converged 27% slower than the average original network. This experiment further supports the lottery ticket hypothesis’ emphasis on fortuitous initialization. Using the same pruned architecture, the original initialization not only withstood but benefited from pruning, while performance of the reinitialized network immediately suffered and steadily diminished as the network was further pruned. This outcome mirrors, on a larger scale, the result of the XOR experiment, in which networks with many hidden units could be pruned to smaller winning tickets that found the right decision boundary at a much higher rate than randomly-initialized small networks. Figure 5 provides a broader perspective on these patterns across all of the levels to which we pruned. The left graph shows the convergence time in relation to the percentage of the network remaining after pruning. The blue line is the average of the five winning ticket trials at each level. The convergence 7 Figure 5: The convergence times (left) and accuracies (right) when running the MNIST pruning experiment to various degrees of pruning. The blue line is the average of five trials with different starting initializations that prune and reuse the original initialization. Each of the multicolored lines represents three randomly reinitialized control trials (one for each trial with the original initialization). The error bars are the minimum and maximum value any trial takes on at each interval. Figure 6: The convergence times and accuracy for the five winning tickets at each level of pruning (blue line), the 15 trials where the winning ticket weights were reinitialized (orange line), and the 15 trials where the winning ticket weights were maintained but shuffled within each layer (green line). time initially decreases before leveling off between 55% and 35% and then slowly climbing again. In contrast, the multicolored lines, which represent the groups of control trials for each winning ticket, steadly require longer to converge as more of the network is pruned. In the control experiment, the error bars are much larger, suggesting wider variation in convergence times compared to the consistent convergence times of the winning tickets. The right graph in Figure 5 provides important context: how accurate are the networks at the moment they converge? The average trial that used the original initialization (the blue line) maintans accuracy within 0.15% of the original network when pruned down to 15%, after which accuracy drops off. In contrast, the accuracy of the average control trial drops below this level when the network has been pruned by about 30%, falling precipitously when pruned by 15% (0.6% below the original network’s accuracy). This experiment supports the lottery ticket hypothesis’ prediction that fortuitous initialization is a necessary ingredient to make a winning ticket. The winning ticket’s structure alone is insufficient to explain its success. Control experiment 2. This experiment evaluates the extent to which architecture is a necessary component of a winning ticket. For each winning ticket at each level of pruning, we randomly shuffled the locations of the weights in each hidden layer while maintaining their original initializations. The results of doing so appear in Figure 6. Just as in Figure 5, the blue line traces winning tickets pruned to various sizes. The orange line is the average of all 15 of the trials from control experiment 1 (reinitializing the winning tickets). The green line is the average of all 15 of the trials from control experiment 2 (shuffling the winning tickets without reinitializing). The convergence times for the two control experiments are similar: they start increasing immediately and increase more rapidly as the network gets smaller. The accuracy of control experiment 2 drops off slightly earlier than control experiment 1, which itself dropped off before the winning ticket did. This experiment supports the lottery ticket hypothesis’ prediction that winning tickets emerge from a combination of initialization and structure. Neither initialization (control experiment 1) nor structure (control experiment 2) alone is sufficient to explain the better performance of the winning tickets. 8 Figure 7: The convergence times and accuracy of winning tickets extracted from fully-connected networks for MNIST using one-shot pruning (orange) and iterative pruning (blue). Note that the x-axis is logarithmic. Summary. The first notable result from this set of experiments is that, even when pruned to sizes much smaller than the original network, winning tickets are still able to converge at all. This supports the core prediction of the lottery ticket hypothesis: pruning reveals smaller subcomponents that were originally initialized such that they can train successfully in isolation. Not only do these networks train successfully, but they converge faster than and maintain the accuracy of the networks from which they derive. Furthermore, winning tickets emerge from a confluence of both fortuitous initalization and structure. 4 MNIST (Iterative Pruning) In the XOR experiment in Section 2, iterative pruning [4]—repeatedly training, pruning, reinitializing, and pruning again—arrived at winning tickets that were more likely to train successfully In this section, we find that iterative pruning makes it possible to extract winning tickets from our MNIST network that are far smaller than those generated by one-shot pruning. 4.1 Methodology We use the same experimental setup (network architecture, initialization strategy, and optimization strategy) as in Section 3. We follow a similar procedure repetitively in order to iteratively prune. 1. Randomly initialize the network. 2. Train for 50,000 iterations on 100-example mini-batches from the training data 3. Prune 20% of the weights from within each hidden layer and 10% of the weights in the output layer, removing those with the lowest magnitudes. 4. Reset the weight values of the pruned network to their initializations from before training. 5. Repeat steps (2) through (4) until the network has been pruned to the desired size. The result of the last iteration of (4) is the winning ticket. We iteratively prune the incoming weights of the first and second layers of the network by 20% and the weights of the output layer by 10%.6 We start with a network with two fully-connected hidden layers with 300 and 100 hidden units and prune the network until just under 1% of the original weights remained. Comparison to one-shot pruning. Figure 7 shows the difference in convergence times and accuracy between one-shot pruning (orange) and iterative pruning (blue). (Note that the x-axis is logarithmic in Figure 7 and in most figures in this section.) The average iteratively pruned winning tickets reach initially reach lower convergence times. These convergence times flatten when the original network is pruned to between 41% (36% faster than the original network) and 8.5% (38% faster than the original network) of the original network size, as compared to between 55% (44% faster than the original network) and 40% (41% faster than the 6 As mentioned in Section 3, we prune the output layer at a lower rate to reduce the chances of severing connectivity to any of the output units. 9 Figure 8: The convergence times and accuracy of winning tickets extracted by iteratively pruning and control experiments. The blue line is the average of five winning tickets. The orange line is control experiment 1: winning tickets that have been reinitialized. The green line is control experiment 2: winning tickets whose weights were randomly shuffled. The red line is the performance of one-shot pruning. The locations where the control trials cut off are those at which, according to our metric, they no longer converged. original network) for one-shot pruning. The average iteratively pruned network returns to the original convergence time when pruned to 1.8% (as compared to between 5% and 10% for one-shot pruning). Likewise, accuracy actually increases slightly for many winning tickets, returning to the original network’s accuracy at a winning ticket size of 2.8% on average. In contrast, one-shot pruning begins to drop off when the winning ticket is 15% of the size of the original network. Although iterative pruning can extract much smaller winning tickets than one-shot pruning, it is far more costly to find those winning tickets. Extracting a winning ticket with one-shot pruning requires training the original network a single time, regardless of how much the network is pruned. In contrast, iteratively pruning a network by 20% at each iteration until it is about 5% of the original network’s size requires training the network 14 times. However, since our goal is to understand the behavior of winning tickets rather than to find them efficiently, iterative pruning’s compelling advantage is that it is able to extract smaller winning tickets that maintain convergence and accuracy performance, placing a tighter upper-bound on the size of a network’s winning ticket. 4.2 Results In this section, we re-run the control experiments from Section 3.1. Just as before, we aim to explore the extent to which architecture and initialization are responsible for a winning ticket’s ability to continue to converge at such small sizes. Figure 8 contains the average results of performing control experiment 1 (randomly reinitializing the winning ticket’s weights) in orange and control experiment 2 (randomly shuffling the winning ticket’s weights) in green. For comparison, the curve in red is the performance of one-shot pruning. Control experiment 1. Just as with one-shot pruning, average convergence times for control experiment 1 begin increasing as soon as the network is pruned and continue to grow at a steady rate. The error bars in Figure 8 reflect that convergence times vary widely for pruned networks that are reinitialized. The average control trial’s accuracy begins dropping off (more than 0.15% lower than the original networks accuracy) when the network is pruned to about 31%, whereas the average iteratively pruned network drops below this level when pruned to 1.5%. Just as with the one-shot experiment, this control trial indicates that initialization plays a critical role in making a winning ticket. Control experiment 2. Average convergence times for control trial 2 increase steadily in a pattern similar to those from control trial 1. Error bars indicate that these convergence times similarly vary widely. Accuracy begins dropping off earlier, at about 51.2% and more steeply, potentially suggesting that architecture might be more important than initialization. Summary. The control experiments for iterative pruning put the results from Section 3 in sharper relief. Iterative pruning makes it possible to extract smaller winning tickets than from one-shot pruning that reach lower convergence times than the original network while maintaining or exceeding its level of accuracy. The control experiments show that both initialization and network architecture 10 Figure 9: The distributions of the initializations of weights that survived iterative pruning across ten iterative pruning runs. The graphs contain the initializations of the network pruned to (left to rigth) 100%, 51.2%, 16.8%, and 5.50%. The blue, orange, and green lines are the distributions of initial weights for the first hidden layer, second hidden layer, and output layer, respectively. play factor into creating a winning ticket, with control trial 2 suggesting that network architecture might be slightly more important. The experiments on XOR and MNIST support the lottery ticket hypothesis. Embedded within larger networks are small subcomponents that are fortuitously initialized in a manner conducive to successful training. We extracted the winning ticket architectures by pruning and determined the corresponding initializations by resetting the winning ticket’s connections to their original from before training. These networks not only trained successfully but, in the case of iteratively-pruning with the MNIST network, converged faster and more accurately. Meanwhile, neither architecture nor initialization alone could entirely account for this result. We next investigate the architecture and initializations of these small winning tickets (Section 5) and the behavior of winning tickets when subjected to a wider variety of parameters (Section 6). 5 Examining Winning Tickets In this section, we briefly explore the internal structure of winning tickets that result from iterativelypruning our MNIST network. We have already found evidence to support the claim that winning tickets arise out of a confluence of architecture and initialization. What exactly do those architectures and initializations look like? Initializations. Figure 9 shows the initialization distributions for winning tickets at four different levels of pruning. Note that these are the values of the winning ticket’s weights from before training. The graph in the upper left contains the initial weights of the entire network (no pruning), which are initialized according to a normal distribution with mean 0 and standard deviation 0.1. The graph in the upper right contains the weights after iteratively pruning the network down to 51.2% of its original size. The blue, orange, and green lines are the distriutions of initial weights for the first hidden layer, second hidden layer, and output layer, respectively. At 51.2%, the remaining weights already show the impact of pruning. The first and second hidden layer’s distributions are bimodal, with two peaks mirrored opposite 0. Since these distributions plot the original initializations of the weights that survive the pruning process (i.e., the weights before training), these distributions were created by removing samples from a formerly normal distribution. These peaks on the 51.2% graph appear to be the left and right tails of the original normal distribution. The missing weights have been pruned. Interestingly, pruning occurs after training and these are graphs of weights before training. In other words, for these distributions to emerge, small weights from before training must have remained small after training. The second hidden layer (orange) retains more of its center than the first hidden layer, indicating that those weights likely moved more during training. The output distribution (green) more closely resembles the original normal distribution, indicating that its weights probably moved significantly during training. One other contributing factor to the output distribution is that we prune it at a slower rate, meaning that the effects of pruning make take longer to appear. 11 Figure 10: For each unit in the current layer, how many units in the previous layer connect to it? The left graph is for the first hidden layer. The middle graph is for the second hidden layer. The right graph is for the output layer. The blue, orange, and green lines are for the winning tickets when iteratively pruned to 80%, 16.8%, and 5.50%, respectively. Each point on the line represents a single unit; the units have been sorted in descending order by the number of connections they have. These data points were collected over ten trials. The pattern for 51.2% pruning plays out in more extreme form at 16.8% (lower left graph) and 5.50% (lower right graph). The middles of the first and second hidden layer distributions continue to get hollowed out, and the same happens (albeit more slowly) to the output distribution. Even with these extreme-looking input distributions, the corresponding networks converged faster than the original network and retained the same accuracy. Considering the extent to which the particular pruning strategy we pursued left its imprint on the winning tickets, it is worth considering the impact that other pruning strategies would have and, more broadly, whether the winning tickets we found are a product of the pruning strategy we pursued or whether the pruning strategy we pursued happens to exploit a deeper reality of the way neural networks behave. Architecture. As the network is pruned, it becomes sparser. Figure 10 shows the distributions of surviving connections aggregated across ten trials7 between a unit in layer n and units in layer n − 1 when the network is pruned to 80% (blue), 16.8% (orange), and 5.50% (green) of its original size. The left, middle, and right graphs show the first hidden layer, second hidden layer, and output layer. When pruned to 80%, the network remains almost fully-connected, with only slight differences between the units with the most and least connections. As the network is further pruned, units in the first hidden layer continue to have a roughly equal number of connections to units in the input layer. Even when the network is pruned to 5.50%, only a small fraction of the hidden units in the first layer have been eliminated entirely. The second hidden layer becomes less evenly connected as more weights are pruned. By the time the network is pruned to 5.50%, nearly a third of the units in the second hidden layer have been fully disconnected, and there is a steep decline from the best-connected units to the worst-connected. The output layer shows a less severe slope, likely because every output unit serves a clear function and because we prune the output layer at a slower rate. The winning tickets are quite sparse. Even when the network is pruned by nearly 95%, only a fraction of the units have been eliminated entirely. No units maintain a large number of connections after pruning; instead, nearly all units retain a proportionally small number of connections. 6 Exploring MNIST Parameters This section explores the sensitivity of the MNIST results to the parameters of the lottery ticket experiment. Namely, we explore the role that initialization and network size play in the properties of the winning tickets that emerge. 6.1 Initialization Although our default network was initialized from a normal distribution with mean 0 and standard deviation 0.1, we experimented with several other standard deviations to explore the effect of larger 7 We also removed any edges that did not have a path to an output unit. 12 Figure 11: The convergence times and accuracy for groups of five winning tickets initialized with various standard deviations ≥ 0.1. Figure 12: The convergence times and accuracy for groups of five winning tickets initialized with various standard deviations ≤ 0.1. and smaller weights on the behavior of winning tickets. One might expect that our pruning strategy would be especially vulnerable to initializing network weights too large: by selecting for the highestmagnitude weights, it might exacerbate exploding gradients. Likewise, it might be resilient to initializing network weights too small, since it will select for the largest weights after training. In this section, we present the results using the one-shot pruning strategy. The results for iterative pruning were similar. Figure 11 shows the convergence times and accuracy for winning tickets of networks initialized with standard deviations larger than 0.1 (0.2, 0.4, and 0.8). As expected, convergence times increase and accuracy decreases as the standard deviations increase. We did not explore whether the extent to which this behavior resulted from exploding gradients or weaknesses in the pruning strategy. Figure 12 contains the same information for winning tickets of networks initialized with standard deviations smaller than 0.1. A standard deviation of 0.1 produces the fastest convergence times but cedes a certain amount of accuracy in doing so. In contrast, a standard deviation of 0.025 causes winning tickets to converge more slowly but to higher-accuracy optima. This behavior suggests that there are sweet spots for both convergence times (stddev=0.1) and accuracy (stddev=0.025) and a tradeoff-space in between. 6.2 Network Size We experimented with increasing the size from the default network (layers of 300 and 100 hidden units) in order to determine whether there is a fixed winning ticket size for a particular learning problem, or whether larger networks naturally beget larger winning tickets. We consider two possible definitions of the “size” of a network’s winning ticket: • A winning ticket is the minimal network that minimizes convergence time. Since convergence times initially decrease with pruning, this heuristic looks for the winning ticket with the lowest possible convergence time. • A winning ticket is the minimal network that retains the accuracy of the original network. Accuracy remains relatively flat as smaller and smaller winning tickets are created; it then 13 Figure 13: The convergence times and accuracy for groups of five winning tickets extracted from networks of various sizes with the one-shot pruning strategy. Error bars were elided to improve readability. The legend contains the size of the network (e.g., 300-100 means a network with hidden layers of 300 and 100 units). All networks were initialized with a standard deviation of 0.05. Figure 14: The convergence times and accuracy for groups of five winning tickets extracted iteratively from networks of various sizes. Error bars were elided to improve readability. All networks were initialized with a standard deviation of 0.05. reaches a tipping point and drops off rapidly. This definition considers the winning ticket to be the last moment before this accuracy drop-off takes place. One-shot pruning. We trained networks whose sizes were multiples of the original network size. The results of doing so and applying the one-shot pruning strategy appear in Figure 13, which plots convergence times and accuracy according to the number of weights in the winning ticket. According to the convergence-based definition of a winning ticket, the winning ticket sizes increase gradually with the size of the network. The LeNet-300-100 architecture appears to reach this point at about 140,000 weights, where the LeNet-600-300 does so at about 200,000 weights. The same pattern holds for the larger architectures. Larger networks are capable of representing more sophisticated functions, so pruning larger networks may produce different network architectures that exploit this additional representation capacity to converge faster. Indeed, the larger the network, the lower the convergence times its winning tickets were able to achieve and the larger the size at which it reached them. The accuracy-based definition of a winning ticket agreed. As the bottom graph of Figure 13 illustrates, the accuracy of larger networks dropped off steeply at slightly earlier times than the accuracy of smaller networks. However, these differences were quite small—on the order of tens of thousands of weights. Although winning ticket size does seem to increase with network size by this definition, the changes were very slight and winning ticket sizes close to uniform. Iterative pruning. As Figure 14 reflects, the convergence and accuracy trends for iteratively pruning larger networks remains the same as in the one-shot case. Larger networks reach their minimum convergence times at gradually larger sizes, but accuracy plummets in unison. There are two key differences worth noting in the iterative case. 14 Figure 15: The convergence times and accuracy of winning tickets extracted by iteratively pruning by different rates on each iteration. Error bars have been elided for readability. Note that the x-axis is not logarithmic. First, both minimum convergence times and accuracy dropoffs occur at much smaller network sizes than in the one-shot experiments. This result coincides with the other iterative experiments, which demonstrate that iterative pruning creates winning tickets that can be pruned to much smaller sizes before convergence times increase and accuracy diminishes. Whereas the accuracy dropoff took place when the networks had about 150,000 weights in the one-shot experiments, it occurs when the iteratively-derived winning tickets have just tens of thousands of weights. Second, each of the accuracy graphs has a small bulge upwards just before dropping off, indicating that accuracy actually increases slightly when the winning tickets are smallest. These bulges occur at the same winning ticket size in all cases, regardless of the initial size of the network. Summary. The analysis in this subsection leaves many open questions for future research. Although we do not undertake extensive analysis of the internal structure of winning tickets in this study, comparing equally-sized winning tickets derived from networks of different sizes would shed light on the extent to which the winning tickets themselves are similar or different between various initial network sizes. 6.3 Exploring Iterative Pruning Rates Choosing the exact rate at which to prune on each iteration of iterative pruning entails balancing the performance of the resulting winning ticket with the number of iterations necessary to extract that winning ticket. Figure 15 shows the convergence times and accuracy for the LeNet-300-100 architecture iteratively pruned at different rates on each iteration. (Note that the x-axis is not logarithmic.) This experiment can be thought of as exploring middle grounds between one-shot pruning and iteratively pruning at a small rate. Although pruning by a larger percentage (e.g., 70% or 50%) on each iteration reaches smaller winning tickets faster, those winning tickets are pruned too aggressively and fail to match the convergence times or accuracy of winning tickets pruned more slowly. On the other end of the spectrum, iteratively pruning by 10% appears to achieve the best convergence times and accuracy but would require training the network 28 times to extract a winning ticket 5% of the original network’s size. For our experiments, we prune by 20%, which balances performance with the amount of training required. 6.4 Weight Resetting Before each training iteration of our iterative pruning approach, we reset the weights of the unpruned connections to their original values from before training. Doing so is part of our experiment to evaluate the lottery ticket hypothesis: exploring how well winning tickets obtained by pruning train in isolation. We conjecture that resetting before each training iteration makes it easier to find small winning tickets. In effect, each iteration is a recursive pruning problem in which a subnetwork that trains effectively when starting from the original initializations must be pruned to a slightly smaller network that does the same. In contrast, Han et al. [4] interleave training and pruning without ever resetting weights. After a round of training, low-magnitude weights are pruned and then training continues based on the trained 15 Figure 16: The convergence times and accuracy of winning tickets extracted by iteratively pruning using weight resetting between iterations (our strategy, in blue) and by continuing to use the trained weights after pruning (Han et al [4]’s strategy, in orange). weights. These differences in approaches reflect two different goals: Han et al. want to produce the smallest possible trained network, while we wish to find a pruned network that trains successfully from the start. Figure 16 shows the convergence times and accuracy achieved by the winning tickets extracted using these two pruning strategies. To simulate Han et al.’s strategy, we iteratively trained a network, pruned low-magnitude weights, and continued training using the trained weights. At each iteration, we copied the resulting network, reset its weights to the original initializations, and trained the network to obtain the results in Figure 16. As Figure 16 shows, Han et al.’s pruning strategy is quite effective at finding small networks that rain successfully, although our strategy of resetting weights at each iteration maintains lower convergence times and higher accuracy for slightly longer. However, since Figure 16 is on a logarithmic scale, these differences appear only at very small network sizes. 7 Related Work Pruning. LeCun et al. first explored pruning as a way to reduce the size of neural networks [7]; they pruned based on the second derivative of the loss function with respect to each weight. Hassibi et al. [5] build on this approach. More recently, Han et al. [4, 3, 2] showed that these techniques could be used to substantially reduce the size of modern image-recognition networks. Since then, a rich variety of neural network pruning approaches have emerged (e.g., pruning smallest weights [4], pruning units in a Bayesian fashion [9], pruning entire convolutional filters [8, 10], fusing redundant units to increase network diversity [11]). The goal of this literature on pruning is to compress trained neural networks, reducing the size of a large model such that it can run efficiently on restricted computational platform (e.g., a mobile device) without sacrificing accuracy. In contrast, we aim to make it possible to train small neural networks from the start. In [4] and follow-up work, network compression takes place in three iterative steps. First, a large network is trained. Second, weights or units are pruned according to a heuristic. Third, the network is further trained using the already-trained weights. Han et al. find that, without this third retraining step, network performance drops off much earlier in the pruning process. Han et al. also caution that the pruned network should not be re-initialized after training, but do not consider reusing the values to which the surviving weights were initialized in the original network as we do. Our work builds off of the literature on pruning by shedding light on the mechanisms that make pruning possible. The fact that networks can be pruned while maintaining accuracy indicates that the function to be learned can be represented with a much smaller network than the one used for training. We aim to understand why pruning is possible and investigate whether small networks can be trained directly (rather than pruning large networks to smaller sizes after training). The lottery ticket hypothesis posits that large networks have small, fortuitously-initialized subnetworks that facilitate successful training. From this point of view, neural network pruning finds these winning tickets. To evaluate the lottery ticket hypothesis on small, fully-connected networks, we leverage Han 16 et al.’s experimental approach, except that we make a crucial modification: after pruning we reset each weight to its original value. Our results explain or complement those of Han et al. The lottery ticket hypothesis offers insight into why Han et al. are able to prune networks. Many of the trends we see (e.g., that the accuracy of iteratively-pruned winning tickets drops off at very small winning ticket sizes or that the original initializations of pruned networks take on a bimodal distribution) parallel those that Han et al. find when continuing to train pruned networks based on their trained weights. Dropout. Dropout [13] creates a smaller subnetwork for each training iteration by randomly removing a subset of the units (or weights). At inference-time, a unit’s activation is reduced by the probability that it was dropped out. Intuitively, dropout is intended to reduce overfitting and improve generalization by forcing units to remain robust to changes in the network. Follow-up work on dropout [1] has characterized training with dropout as “perform[ing] gradient descent...with respect to...the ensemble of all possible subnetworks” and inference with dropout as approximately computing the average over this ensemble. In the terminology of dropout, our experiment aims to discover a single, particularly successful member of this ensemble of subnetworks. Our dropout heuristic is that, after training the network once without dropout, we drop out the lowest k% of weights (by magnitude after training) with probability 1 and all other weights with probability 0. In other words, we perform an extremely aggressive, coarse-grained form of dropout based on examining the results of training the network once without dropout. However, our goal is different. Dropout is designed to regularize a network during training, a process that can be used to produce sparse networks. We aim to directly find small (and, in the case of the networks we found, sparse) networks that can be trained from start to finish without removing further weights. Our broader formulation of the lottery ticket hypothesis does closely relate to dropout’s notion of ensemble learning. The lottery ticket hypothesis views a randomly-initialized large network as a collection of a combinatorial number of small networks (i.e., lottery tickets) of which one (i.e., the winning ticket) must be initialized fortuitously to enable training to succeed. From this point of view, a large network begins with the possibility of coalescing toward one of an exponential number of subnetworks, and gradient descent drives it toward the subnetwork comprising the winning ticket that we find. 8 Limitations This work is limited in several ways. We only examine fully-connected networks, and for two of the smallest possible examples (XOR and MNIST). We do not consider convolutional networks or larger networks that better reflect real-world examples.8 Our evidence for the lottery ticket hypothesis is purely experimental; we do not offer any theoretical analysis to formally support this claim. Finally, although we analyze the structure and initialization distributions of winning tickets for MNIST, we have yet to devise a way to turn these observations into useful strategies for training smaller networks. We anticipate exploring these avenues in future work and updating this paper as we do so. 9 Conclusions and Future Work This paper proposes a new hypothesis to explain why large neural networks are amenable to substantial pruning yet the pruned networks cannot be trained effectively from scratch. This conjecture, known as the lottery ticket hypothesis, holds that training succeeds when a subcomponent of the larger network is randomly initialized in a fashion that is suitable for optimization. Furthermore, it conjectures that pruning uncovers these winning tickets. To empirically evaluate this hypothesis, we devised an experiment based on the work of Han et al. [4] where, after pruning a trained network, remaining weights are reset to their original initializations. If the lottery ticket hypothesis holds and pruning 8 However preliminary experiments with CIFAR10 on a convolutional network reflect the same behavior described in this paper for MNIST. 17 uncovers these winning tickets, then these pruned networks should train successfully in isolation when reset to their original initializations. On XOR, we found that winning tickets derived from larger networks were able to learn the decision boundary and reach zero loss far more frequently than those that were randomly initialized. On MNIST, winning tickets converged more quickly and reached higher accuracy than the original network. Control experiments supported the claim that winning tickets represent a confluence of fortuitious initialization and network architecture. This paper articulates a new perspective on neural network training and supports that view empirically. Now that a foundation has been laid, there are numerous research directions to further evaluate the lottery ticket hypothesis and exploit this perspective to improve network design and training. Larger examples. The largest network that we examine is a fully-connected network for MNIST. Repeating the experiments outlined in this paper for a convolutional network, larger networks, and harder learning tasks would make it possible to understand whether the lottery ticket hypothesis holds more generally and how it manifests in these settings. Understanding winning tickets. This paper focuses mainly on the behavioral properties of the lottery ticket hypothesis, pruning, and winning tickets. One logical next step is to systematically analyze the architectures and initializations of lottery tickets. To what extent are winning tickets unique artifacts created by randomly initializing large networks and getting lucky? To what extent is there common structure between multiple winning tickets for the same task? What do winning tickets tell us about the functions that neural networks learn for particular tasks? Lottery ticket networks. The lottery ticket hypothesis and the existence of winning tickets demonstrate that small networks can be trained from start to finish. The most concrete follow-up to this work would be to exploit the lessons learned by leveraging winning tickets to develop new network architectures and initialization regimes that allow smaller networks to be trained for a wider variety of learning tasks. Doing so could reduce the amount of computation needed to train neural networks. References [1] Pierre Baldi and Peter J Sadowski. 2013. Understanding dropout. In Advances in neural information processing systems. 2814–2822. [2] Song Han, Huizi Mao, and William J. Dally. 2015. Deep Compression: Compressing Deep Neural Network with Pruning, Trained Quantization and Huffman Coding. CoRR abs/1510.00149 (2015). arXiv:1510.00149 http://arxiv.org/abs/1510.00149 [3] Song Han, Huizi Mao, and William J Dally. 2015. A deep neural network compression pipeline: Pruning, quantization, huffman encoding. arXiv preprint arXiv:1510.00149 10 (2015). [4] Song Han, Jeff Pool, John Tran, and William Dally. 2015. Learning both weights and connections for efficient neural network. In Advances in neural information processing systems. 1135–1143. [5] Babak Hassibi, David G Stork, and Gregory J Wolff. 1993. Optimal brain surgeon and general network pruning. In Neural Networks, 1993., IEEE International Conference on. IEEE, 293– 299. [6] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient-based learning applied to document recognition. Proc. IEEE 86, 11 (1998), 2278–2324. [7] Yann LeCun, John S Denker, and Sara A Solla. 1990. Optimal brain damage. In Advances in neural information processing systems. 598–605. [8] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. 2016. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710 (2016). [9] Christos Louizos, Karen Ullrich, and Max Welling. 2017. Bayesian compression for deep learning. In Advances in Neural Information Processing Systems. 3290–3300. 18 [10] Jian-Hao Luo, Jianxin Wu, and Weiyao Lin. 2017. Thinet: A filter level pruning method for deep neural network compression. arXiv preprint arXiv:1707.06342 (2017). [11] Zelda Mariet and Suvrit Sra. 2015. Diversity Networks: Neural Network Compression Using Determinantal Point Processes. arXiv preprint arXiv:1511.05077 (2015). [12] Suraj Srinivas and R Venkatesh Babu. 2015. Data-free parameter pruning for deep neural networks. arXiv preprint arXiv:1507.06149 (2015). [13] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research 15, 1 (2014), 1929–1958. 19	2cs.AI
Infinitely generated semigroups and polynomial complexity arXiv:1503.04610v2 [math.GR] 11 May 2016 J.C. Birget Abstract This paper continues the functional approach to the P-versus-NP problem, begun in [2]. Here we focus on the monoid RMP2 of right-ideal morphisms of the free monoid, that have polynomial input balance and polynomial time-complexity. We construct a machine model for the functions in RMP2 , and evaluation functions. We prove that RMP2 is not finitely generated, and use this to show separation results for time-complexity. 1 Introduction In [2] we defined the monoids of partial functions fP and RMP2 . The question whether P = NP is equivalent to the question whether these monoids are regular. The monoid fP consists of all partial functions A∗ → A∗ that are computable by deterministic Turing machines in polynomial time, and that have polynomial I/O-balance. The submonoid RMP2 consists of the elements of fP that are right-ideal morphisms of A∗ . One-way functions (according to worst-case time-complexity) are exactly the nonregular elements of fP. It is known that one-way functions (according to worst-case time-complexity) exist iff P 6= NP. Also, f ∈ RMP2 is regular in fP iff f is regular in RMP2 . Hence, P = NP iff fP is regular, iff RMP2 is regular. We refer to [9, 15] for background on P and NP. The original motivation for studying RMP2 in addition to fP was that RMP2 is reminiscent of the Thompson-Higman groups [14, 16, 12, 7, 6, 5] and the Thompson-Higman monoids [4]. It also quickly turned out that RMP2 , while having the same connection to P-vs.-NP as fP, has different properties than fP (e.g., regarding the Green relations, and actions on {0, 1}ω ; see [2, 3]). It is hard to know whether this approach will contribute to a solution of the P-vs.-NP problem, but the monoids fP and RMP2 are interesting by themselves. Above and in the rest of the paper we use the following notation and terminology. We have an alphabet A, which will be {0, 1} unless the contrary is explicitly stated, and A∗ denotes the set of all strings over A, including the empty string ε. For x ∈ A∗ , \|x\| denotes the length of the string x. For a partial function f : A∗ → A∗ , the domain is Dom(f ) = {x ∈ A∗ : f (x) is defined}, and the image is Im(f ) = f (A∗ ) = f (Dom(f )). When we say “function”, we mean partial function (except when we explicitly say “total function”). Similarly, for a deterministic input-output Turing machine with input-output alphabet A, the domain of the machine is the set of input words for which the machine produces an output; and the set of output words is the image of the machine. A function f : A∗ → A∗ is called polynomially balanced iff there exists polynomials p, q such that for all x ∈ Dom(f ): \|f (x)\| ≤ p(\|x\|) and \|x\| ≤ q(\|f (x)\|). The polynomial q is called an input balance function for f . As we said already, fP is the set of partial functions f : A∗ → A∗ that are polynomially balanced, and such that x ∈ Dom(f ) 7−→ f (x) is computable by a deterministic polynomial-time Turing machine. Hence, Dom(f ) is in P when f ∈ fP, and it is not hard to show that Im(f ) is in NP. Clearly, fP is a monoid under function composition. A function f : A∗ → A∗ is said to be one-way (with respect to worst-case complexity) iff f ∈ fP, but there exists no deterministic polynomial-time algorithm which, on every input y ∈ Im(f ), outputs some x ∈ A∗ such that f (x) = y. By “one-way” we will always mean one-way with respect to worst-case complexity; hence, these functions are not “cryptographic one-way functions” (in the sense of, e.g., 1 [8, 13, 10]). However, they are important for the P-vs.-NP problem because of the following folklore fact (see e.g., [11] p. 33): One-way functions exist iff P 6= NP. As is easy to prove (see the Introduction of [2]), f ∈ fP is not one-way iff f is regular in fP. By definition, an element f in a monoid M is regular iff there exists f ′ ∈ M such that f f ′ f = f ; in this case, f ′ is called an inverse of f .1 A monoid M is called regular iff all the elements of M are regular. In summary we have: The monoid fP is regular iff P = NP. Let us look in more detail at the monoid RMP2 . A right ideal of A∗ is a subset R ⊆ A∗ such that R A∗ = R (i.e., R is closed under right-concatenation by any string). For two strings v, w ∈ A∗ , we say that v is a prefix of w iff (∃x ∈ A∗ )[ vx = w ]. A prefix code in A∗ is a set P ⊂ A∗ such that no word in P is a prefix of another word in P . For any right ideal R there exists a unique prefix code PR such that R = PR A∗ ; we say that PR generates R as a right ideal. For details, see e.g. [6, 4]; a good reference on prefix codes, and variable-length codes in general is [1]. A right-ideal morphism is a partial function h : A∗ → A∗ such that for all x ∈ Dom(h) and all w ∈ A∗ : h(xw) = h(x) w. In that case, Dom(h) and Im(h) are right ideals. For a right-ideal morphism h, let domC(h) (called the domain code) be the prefix code that generates Dom(h) as a right ideal. Similarly, let imC(h), called the image code, be the prefix code that generates Im(h). So a right-ideal morphism h is determined by h\|domC(h) (the restriction of h to its domain code). In general, imC(h) ⊆ h(domC(h)), and it can happen that imC(h) 6= h(domC(h)). We define RMP2 = {f ∈ fP : f is a right-ideal morphism of A∗ }. By Prop. 2.6 in [2], f ∈ RMP2 is regular in RMP2 iff f is regular in fP. The monoid RMP2 is regular iff P = NP. We saw (Cor. 2.9 in [2]) that fP and RMP2 are not isomorphic, that the group of units of RMP2 is trivial (Prop. 2.12 in [2]), and that RMP2 has only one non-0 J -class (Prop. 2.7 in [2]). In [3] we will see that RMP2 has interesting actions on {0, 1}ω , and has interesting homomorphic images (some of which are regular monoids, and some of which are regular iff P = NP). Overall, RMP2 seems to have “more structure” than fP. It is proved in [2] (Section 3) that fP is isomorphic to a submonoid of RMP2 . To prove this, we use an encoding of the three-letter alphabet {0, 1, #} into words over the two-letter alphabet {0, 1}; this encoding will also be used here. First, we encode the alphabet {0, 1, #} by code(0) = 00, code(1) = 01, code(#) = 11. A word x1 . . . xn ∈ {0, 1, #}∗ is encoded to code(x1 ) . . . code(xn ). For a fixed k > 0, a k-tuple of words (u1 , . . . , uk−1 , uk ) ∈ {0, 1}∗ × . . . × {0, 1}∗ is encoded to code(u1 # . . . uk−1 #) uk = code(u1 ) 11 . . . code(uk−1 ) 11 uk ∈ {0, 1}∗ . A function f ∈ fP is encoded to f C ∈ RMP2 , defined by domC(f C ) = code(Dom(f ) #), so Dom(f C ) = code(Dom(f )) 11 {0, 1}∗ ; and f C (code(x #) v) = code(f (x) #) v, for all x ∈ Dom(f ) and v ∈ {0, 1}∗ ; equivalently, f C (code(x) 11 v) = code(f (x)) 11 v. Then for every L ⊆ {0, 1}∗ , code(L#) is a prefix code, which belongs to P iff L is in P. And f ∈ fP iff f C ∈ RMP2 . The transformation f 7→ f C is a isomorphic embedding of fP into RMP2 ; moreover, f C is regular in RMP2 iff f is regular in fP. From here on, the alphabet denoted by A will always be {0, 1}. In [2] (Section 4) we introduced a notion of polynomial program for Turing machines with built-in polynomial counter (for input balance and time-complexity). These programs form a machine model that characterizes the functions in fP. For a polynomial program w, we let φw ∈ fP denote the function computed by this program. For every polynomial q of the form q(n) = a nk + a (where a, k are positive integers), we constructed an evaluation map evC q ∈ fP such that for every polynomial program w with built-in polynomial pw (n) ≤ q(n) (for all n ≥ 0), and all x ∈ A∗ , evC = code(w) 11 φw (x) q code(w) 11 x 1 The terminology varies, depending on the field. In semigroup theory f ′ such that f f ′ f = f is called a semi-inverse or a pseudo-inverse of f , in numerical mathematics f ′ is called a generalized inverse, in ring theory and in category theory it’s called a weak inverse. In semigroup theory the term “inverse” of f is only applied to f ′ if f ′ f f ′ = f ′ holds in addition to f f ′ f = f . It is easy to see that if f f ′ f = f then f ′ f f ′ (= v) satisfies f vf = f and vf v = v. 2 C if x ∈ Dom(φw ); if x 6∈ Dom(φw ) then evC q code(w) 11 x is undefined. We used evq , with any polynomial q of degree ≥ 2 with large enough coefficient, to prove the following: First, fP is finitely generated (Theorem 4.5 in [2]). Second, evC q is complete in fP with respect to inversive polynomial reduction (Section 5 of [2]). Later in this paper (Def. 2.4 and following) we define completeness and various reductions for RMP2 , along the same lines as for fP. Note that fP and RMP2 , in their entirety, do not have evaluation maps that belong to fP, respectively RMP2 (since such maps would not have polynomially bounded complexity). That is the reason why we restrict ev and evR to complexity ≤ q(.), and why we need precise machine models for fP and RMP2 (as opposed to more intuitive “higher-level” models). In Section 2 we define a machine model that characterizes the functions in RMP2 ; and for any large CC enough polynomial q we construct evaluation maps evRC for the functions in RMP2 that q and evRq have balance and time-complexity ≤ q. We prove that evRCC is complete in RMP2 (and in fP) with q respect to inversive Turing reduction. In Section 3 we prove that RMP2 is not finitely generated, and in Section 4 we show that infinite generation has some complexity consequences, i.e., infinite generation can be used for a time-complexity lower-bound argument. 2 Machine model and evaluation maps for RMP2 The evaluation map evC q : code(w) 11 x 7−→ code(w) 11 φw (x), that we constructed for fP in [2], works in particular when φw ∈ RMP2 (provided that φw has time-complexity and input-balance ≤ q). But P C evC q is not a right-ideal morphism and, moreover, ev q can evaluate functions that are not in RM2 . We want to construct an evaluation map that belongs to RMP2 , and that evaluates exactly the elements of RMP2 that have balance and complexity ≤ q. In [2] we constructed a machine model for fP, namely a class of Turing machines with built-in polynomial counter (for controlling the time-complexity and the input-balance). We will refine these Turing machines in order to obtain a machine model for accepting the right ideals in P, and for computing the functions in RMP2 . We will consider deterministic multi-tape Turing machines with input-output alphabet A, with a read-only input tape, and a write-only output tape. Moreover we assume that on the input tape and on the output tape, the head can only move to the right, or stay in place (but cannot move left). We assume that the input tape has a left endmarker #, and a right endmarker B (the blank symbol). At the beginning of a computation of such a machine M on input z ∈ A∗ , the input tape has content # z B, with the input tape head on # ; initially, all other tapes are blank (i.e., they are filled with infinitely many copies of the letter B). The output tape does not need endmarkers (since it is write-only). We assume that M has a special output state qout , and that M only goes to state qout when the output is complete; the output state is a halting state (i.e., M has no transition from state qout ). An important convention for a Turing machine M with non-total input-output function fM is the following: If M on input x halts in a state that is not qout , then there is no output (even if the output tape contains a non-blank word). So, in that case, fM (x) is undefined. The content of the output tape is considered unreadable, or hidden, until the output state qout is reached. This kind of Turing machine can compute any partial recursive function (the restrictions on the input and output tapes do not limit the machine, because of the work-tapes). To compute a function in fP, we add a built-in polynomial (used as a bound on input balance and time-complexity); see Section 4 in [2]. In order to obtain a machine model for the functions in RMP2 the above Turing machines (with built-in polynomial) will be restricted so that they compute right-ideal morphisms of A∗ . This is done in two steps: First, sequential functions and sequential Turing machines are introduced. From this it is easy to obtain a class of Turing machines that compute right-ideal morphisms (which are a special kind of sequential functions). Recall that by “function” we mean partial function. By definition, a function f : A∗ → A∗ is sequential iff 3 for all x1 , x2 ∈ Dom(f ): if x1 is a prefix of x2 then f (x1 ) is a prefix of f (x2 ). Obviously, every right-ideal morphism is a sequential function. A sequential Turing machine is a deterministic multi-tape Turing machine M (with special input tape and special output tape and output state, according to the conventions above), with input-output function fM , such that the following holds. For every x ∈ Dom(fM ) and every word z ∈ A∗ : in the computation of M on input xz, the input-tape head does not start reading z B until fM (x) has been written on the output tape. To “read a letter ℓ” (in zB) means to make a transition whose input letter is ℓ. So, the input tape has content # xz B, with the input-tape head on the left-most letter of zB (but no transition has been made on that letter yet), and the output tape now has content fM (x). Of course, at this moment the computation of M on input xz is not necessarily finished; the state is not necessarily qout , the output might still grow, and qout might be reached eventually, or not; if qout is never reached, there is no final output. The sequential Turing machines form a machine model for the partial recursive sequential functions. If we let the machines have a built-in polynomial we obtain a machine model for the sequential functions in fP. Finally, to obtain a machine model for the functions in RMP2 we take the sequential Turing machines with built-in polynomial, with the following additional condition. For every x ∈ Dom(fM ) and every word z ∈ A∗ : in the computation of M on input xz, once fM (x) has been written on the output tape (after x was read on the input tape), the remaining input z is copied to the output tape; at this point the state qout is reached. We call such a machine an RMP2 -machine. The following shows how, from an fP-machine for a function f , an RMP2 -machine for f can be constructed, provided that f ∈ RMP2 . Let us first consider right ideals in P, rather than functions. For any polynomial program w for a Turing machine Mw that accepts a language L ∈ P, we construct a new polynomial program v describing a Turing machine Mv that behaves as follows: On any input x ∈ {0, 1}∗ , Mv successively examines prefixes of x until it finds a prefix, say p, that is accepted by Mw ; Mv does not read the letter of x that comes after p until it has decided that p 6∈ L. As soon as Mv finds a prefix p of x such that p ∈ L, Mv accepts the whole input x. If Mw accepts no prefix of x, Mv rejects x. Thus, Mv accepts LA∗ (the right ideal generated by L); if L is a right ideal then LA∗ = L. If Mw has time-complexity ≤ T (.) (a polynomial) then Mv has time-complexity ≤ T (.)2 . Let us now consider functions in RMP2 . Given any polynomial program w for a function φw ∈ fP, we construct a new polynomial program v such that Mv , on input x, successively examines all prefixes of x until it finds a prefix p in Dom(φw ); let φw (p) = y. Then, on input x, the machine Mv outputs y z, where z is such that x = p z. Note that since p is the shortest prefix of x such that p ∈ Dom(φw ), we actually have p ∈ domC(φw ) (if Dom(φw ) is a right ideal). The machine Mv does not read the letter of x that comes after a prefix p until it has decided that p 6∈ Dom(φw ) or p ∈ domC(φw ). Hence, the function computed by Mv is in RMP2 . This construction describes a transformation f ∈ fP 7−→ fpref ∈ RMP2 , where fpref is defined as follows: fpref (x) = f (p) z, where x = p z, and p is the shortest prefix of x that belongs to Dom(f ); so, p ∈ domC(fpref ). Thus for every f ∈ fP we have: f ∈ RMP2 iff fpref = f . Based on RMP2 -machines we can construct evaluation maps for RMP2 . Let q be a polynomial where q(n) = a nk + a for some integers a, k ≥ 1. We define evRC q , as follows: evRC = code(w) 11 φw (x), q code(w) 11 x for all RMP2 -programs w with built-in polynomial pw ≤ q, and for all x ∈ Dom(φw ). The details of C P the construction are the same as for evC q ; see Section 4 in [2]. Although evRq belongs to RM2 and 4 evaluates all RMP2 -programs w with built-in polynomial ≤ q, we will prove in Theorem 4.3 that the complexity of evRC q is higher than q. The following doubly coded evaluation function is usually more useful for RMP2 -programs. It is defined by evRCC code(w) 11 code(u) 11 v = code(w) 11 code(φw (u)) 11 v, q when u ∈ domC(φw ), v ∈ A∗ , and w is as before. CC To give a relation between evRC we will use the following partial recursive right-ideal q and evRq P morphism γ, defined for very RM2 -program w ∈ A∗ and every x ∈ Dom(φw ) by γ code(w) 11 x = code(w) 11 code(u) 11 v, where x = uv, and u is the shortest prefix of x such that u ∈ Dom(φw ); equivalently, u ∈ domC(φw ). When x 6∈ Dom(φw ), γ code(w) 11 x is undefined. Essentially, γ finds the shortest prefix of x that belongs to Dom(φw ) (or equivalently, to domC(φw )). The function γ can be evaluated by examining successively longer prefixes of x until a prefix u ∈ Dom(φw ) is fund. So γ is computable with recursive domain, when w ranges over RMP2 -programs. For any fixed RMP2 -program w, let γw be γ restricted to this w, i.e., γw = γ\|code(w) 11 A∗ . In other words, Dom(γw ) = code(w) 11 Dom(φw ), and γw (code(w) 11 uv) = code(w) 11 code(u) 11 v o by Dom(γ o ) = Dom(φ ) (as opposed to when u ∈ domC(φw ), v ∈ A∗ . Similarly we define γw w w code(w) 11 Dom(φw )), and o (uv) = code(w) 11 code(u) 11 v γw o ) = Im(γ ) = code(w) 11 domC(φ ) 11 A∗ . when u ∈ domC(φw ), v ∈ A∗ . So, Im(γw w w P o Then γw and γw belong to RM2 for every fixed w. But γ itself is not polynomial-time computable, since it has to work for all possible RMP2 -programs w. Another restricted form of γ that belongs to RMP2 is obtained by choosing a fixed polynomial q, and defining γq as the restriction of γ to the set {code(w) 11 x : w is a RMP2 -program with built-in polynomial ≤ q, and x ∈ Dom(φw )}. Hence, γq ∈ RMP2 . We also define the functions π0 , π1 , ρ0 , ρ1 ∈ RMP2 by πa (x) = ax, ρa (ax) = x, for all x ∈ {0, 1}∗ and a ∈ {0, 1}. For a word w = an . . . a1 with ai ∈ {0, 1} we denote πan ◦ . . . ◦ πa1 by πw , and ρan ◦ . . . ◦ ρa1 by ρw . o =γ ◦π o Then we have: γw w code(w) 11 , and γw = γw ◦ ρcode(w) 11 . Another important function in RMP2 is the decoding function, defined for any u, v ∈ A∗ by decode(code(u) 11 v) = uv, so domC(decode) = {00, 01}∗ 11, and imC(decode) = {ε}. We also define a second-coordinate decoding function, for all u1 , u2 , v ∈ A∗ , by decode2 code(u1 ) 11 code(u2 ) 11 v = code(u1 ) 11 u2 v. So, decode2 ∈ RMP2 , domC(decode2 ) = {00, 01}∗ 11 {00, 01}∗ 11, and imC(decode2 ) = {00, 01}∗ 11. CC Now we can formulate a relation between evRC q and evRq : CC evRC ◦ γq . q = decode2 ◦ evRq In order to show that evRCC is complete with respect to inversive reduction in RMP2 , we will q adapt the padding and unpadding functions (defined for fP in [2], Section 4) to RMP2 . Although for 5 RMP2 we keep the same names as for the corresponding (un)padding functions in fP, the functions are slightly different. The padding procedure begins with the function expand(.), defined by expand code(w) 11 code(u) 11 v = code(ex(w)) 11 04 \|code(u)\| 2 +8 \|code(u)\|+2 01 code(u) 11 v, 2 for all u ∈ domC(φw ), v ∈ A∗ , and RMP2 -programs w. The word 04 \|code(u)\| +8 \|code(u)\|+2 01 is of the 2 form code(s) for a word s ∈ 0∗ 1; and 04 \|code(u)\| +8 \|code(u)\|+2 01 code(u) is also a code word, namely 2 code(su). Since 0∗ 1 and its subset (00)∗ 01 are prefix codes, code(s) = 04 \|code(u)\| +8 \|code(u)\|+2 01 is uniquely determined as a prefix of code(su). Here, ex(w) is an RMP2 -program obtained from w so that φex(w) (00)h 01 code(u) 11 v = (00)h 01 code(φw (u)) 11 v, for all u ∈ domC(φw ), v ∈ A∗ , and h > 0. Moreover, if n 7→ a nk + a is the built-in polynomial of the program w then the built-in polynomial of ex(w) is pe (n) = ae n⌈k/2⌉ + ae , with ae = max{12, ⌈a/2k ⌉ + 1}. The detailed justification of the numbers used in the definition of expand and ex (as well as reexpand, recontr, and contr below) is given in [2], Section 4. It is important that expand uses the prefix u of x for padding (in the format code(u) 11, where u ∈ domC(φw )). If the whole input x were used for computing the amount of padding, expand would o , in order to isolate the not be a right-ideal morphism. This is the reason why we introduce γw or γw prefix u ∈ domC(φw ) of x. We iterate expansion (padding) by applying the following function, where ex(.) is as above: reexpand code(ex(z)) 11 0k 01 code(u) 11 v = code(ex(z)) 11 04k 2 +8k+2 01 code(u) 11 v, where k > 0, u, v ∈ A∗ , and z is any RMP2 -program; k is even in the context where reexpand will be used. Repeated contraction (unpadding) is carried out by applying the following function, for k > 0: recontr code(ex(z)) 11 (00)k 01 code(y) 11 v √ = code(ex(z)) 11 (00)max{1, ⌊ k/2⌋−1} 01 code(y) 11 v; √ note that max{1, ⌊ k/2⌋ − 1} ≥ 1. The unpadding procedure ends with the application of the function contr code(ex(z)) 11 (00)k 01 code(y) 11 v = code(z) 11 code(y) 11 v, if 2 ≤ \|(00)k \| = 2k ≤ 4 \|code(y)\|2 + 8 \|code(y)\| + 2. The functions expand(.), ex(.), reexpand(.), recontr(.), and contr(.), are undefined in the cases where no output has been specified above. Lemma 2.1 Let q2 be the polynomial defined by q2 (n) = 12 n2 + 12. For any φw ∈ RMP2 , where w is a RMP2 -program with built-in polynomial q (of the form q(n) = a nk + a for positive integers a, k), we have for all u ∈ domC(φw ), v ∈ A∗ : (⋆) φw (uv) m o = decode ◦ ρcode(w) 11 ◦ contr ◦ recontr 2m ◦ evRCC q2 ◦ reexpand ◦ expand ◦ γw (uv) m o = ρcode(w) 11 ◦ decode2 ◦ contr ◦ recontr2m ◦ evRCC q2 ◦ reexpand ◦ expand ◦ γw (uv), where m = ⌈log2 (a + k)⌉. 6 Proof. This is similar to the proof of Prop. 4.5 in [2], with a few modifications. For u ∈ domC(φw ), v ∈ A∗ , γo w uv 7−→ expand code(w) 11 code(u) 11 v code(ex(w)) 11 04 \|code(u)\| 7−→ reexpand m 7−→ 2 +8 \|code(u)\|+2 01 code(u) 11 v code(ex(w)) 11 0Nm+1 01 code(u) 11 v, where N1 = 4 \|code(u)\|2 + 8 \|code(u)\| + 2, so \|0N1 01\| = (2 (\|code(u)\| + 1))2 ; by induction, Ni = 2 + 8N Ni 01\| = (2 (N 2 4 Ni−1 i−1 + 2 for 1 < i ≤ 2m + 1, and \|0 i−1 + 1)) . The above string, which will CC now be the argument of evq2 , has length > Nm+1 + 2 + \|code(u)\|, which is much larger than the time it takes to simulate the machine with program w on input u. So evRCC q2 can now be applied correctly. Continuing the calculation, evRCC q2 7−→ code(ex(w)) 11 0Nm+1 01 code(φw (u)) 11 v recontr 2m 7−→ code(w) 11 00 01 code(φw (u)) 11 v. We use 2m in recontr 2m because φw (u) could be much shorter than u; but because of polynomial input balance, \|u\| ≤ pw (\|φw (u)\|). Note that doing more input padding than necessary does not do any harm; and recontracting (unpadding) more than needed has no effect (by the definition of recontr). Hence contr can now be applied correctly. We complete the calculation: contr 7−→ code(w) 11 code(φw (u)) 11 v decode2 7−→ code(w) 11 φw (u) v ρcode(w) 11 7−→ φw (u) v. ✷ Lemma 2.2. RMP2 has the following infinite generating set: {decode, ρ0 , ρ1 , π0 , π1 , contr, recontr, evRCC q2 , reexpand, expand} ∪ {γw : w is any RMP2 -program}. Here, decode can be replaced by decode2 . Yet another infinite generating set of RMP2 is k {ρ0 , ρ1 , π0 , π1 } ∪ {evRC q : q is any polynomial of the form q(n) = a n + a with a, k ∈ N≥1 }. o = γ ◦π Proof. The first infinite generating set follows from Lemma 2.1. Recall that γw w code(w) 11 . The second generating set follows in a straightforward way from the proof of Prop. 4.5 in [2]. ✷ Proposition 2.3 RMP2 is generated by a set of regular elements of RMP2 . Proof. The generators ρ0 , ρ1 , π0 , π1 are easily seen to be regular. Thus, using the second infinite generating set in Lemma 2.2, it is enough to factor evRC q into regular elements. We have: evRC q = ρ2,q ◦ Eq , where Eq and ρ2,q are defined as follows: For every RMP2 -program w with built-in polynomial ≤ q, and every u ∈ domC(φw ) and v ∈ A∗ , Eq code(w) 11 code(u) 11 v = code(w) 11 code(u) 11 code(φw (u)) 11 v; and for all z, y, x, v ∈ A∗ such that \|x\| ≤ q(\|y\|), ρ2,q code(z) 11 code(x) 11 code(y) 11 v = code(z) 11 code(y) 11 v. The functions are undefined otherwise. It is easy to see that Eq and ρ2,q have polynomial-time inversion algorithms (i.e., they are regular), and belong to RMP2 . ✷ P We will show now that evRCC q2 is complete in RM2 and in fP, with respect to a certain “inversive reduction”. We need to recall some definitions from [2] concerning reductions between functions in fP or RMP2 , and in particular, reductions that “preserve one-wayness” (inversive reductions). 7 Definition 2.4 Let f1 , f2 : A∗ → A∗ be two polynomially balanced right-ideal morphisms. (1) We say that f2 simulates f1 (denoted by f1 4 f2 ) iff there exist α, β ∈ RMP2 such that f1 = β◦f2 ◦α. (2) We have a polynomial-time Turing simulation of f1 by f2 (denoted by f1 4T f2 ) iff f1 can be computed by an oracle RMP2 -machine that can make oracle calls to f2 ; such oracle calls can, in particular, be calls on the membership problem of Dom(f2 ). In the above definition, f1 , f2 need not be polynomial-time computable. Since RMP2 is J 0 -simple (Prop. 2.7 in [2]), every f1 ∈ RMP2 is simulated by every f2 ∈ RMP2 −{0} (for each of the above simulations). Definition 2.5 (Inversive reduction). If 4X is a simulation between right-ideal morphisms (e.g., as in the previous definition) then the corresponding inversive reduction is defined as follows. We say that f1 inversively X-reduces to f2 (denoted by f1 6inv,X f2 ) iff (1) f1 4X f2 , and (2) for every inverse f2′ of f2 there exists an inverse f1′ of f1 such that f1′ 4X f2′ ; here, f2′ and f1′ range over all polynomially balanced right-ideal morphisms A∗ → A∗ . Note that J 0 -simplicity (Prop. 2.7 in [2]) does not apply for inversive reduction since f2′ , f1′ do not range over just RMP2 . One easily proves the following about polynomially balanced right-ideal morphisms f1 , f2 (see [2], Section 5): If f1 6inv,T f2 and f2 ∈ RMP2 , then f1 ∈ RMP2 ; if, in addition, f2 is regular, then f1 is regular (equivalently, if, in addition, f1 is one-way, then f2 is one-way). Definition 2.6 A polynomially balanced right-ideal morphism f0 is complete in a set S (of right-ideal morphisms) with respect to an (inversive) reduction 6inv,X iff f0 ∈ S, and for all φ ∈ S: φ 6inv,X f0 . See Section 5 of [2] for more details and properties of these simulations and reductions; in [2] the focus was on fP, whereas here we concentrate on RMP2 . The simulations in Def. 2.4 are similar to the standard notions of reductions between decision problems. The concept of inversive reduction was first introduced in [2]; it is the appropriate notion of reduction between functions when one-wayness is to be preserved under upward reduction (and regularity is to be preserved under downward reduction). In the above definitions we only refer to polynomially balanced inverses; this is justified by the following Proposition, according to which “balanced functions have balanced inverses”. Proposition 2.7 Suppose f is a right-ideal morphism with balance ≤ q(.) (where q(.) is a polynomial), and f has an inverse f1′ with time-complexity ≤ T (.). Then f has an inverse f ′ with balance ≤ q and time-complexity ≤ T (.) + c q(.) (for some constant c > 1). The inverse f ′ can be chosen as a restriction of f1′ . Proof. Let f ′ be the restriction of f1′ to the set {y ∈ Dom(f1′ ) : \|y\| ≤ q(\|f1′ (y)\|) and \|f1′ (y)\| ≤ q(\|y\|)}. Then f ′ obviously has balance ≤ q. Note that since f1′ is an inverse of f we have Im(f ) ⊆ Dom(f1′ ). To show that f ′ is an inverse of f it is sufficient to check that the domain of f ′ contains Im(f ). Let y = f (x) ∈ Im(f ) for some x ∈ Dom(f ). Then f (f1′ (y)) = y, since f1′ is an inverse. Checking \|y\| ≤ q(\|f1′ (y)\|): \|y\| = \|f (f1′ (y))\| ≤ q(\|f1′ (y)\|; the inequality holds since q is a balance for f on input f1′ (y). Checking \|f1′ (y)\| ≤ q(\|y\|): \|f1′ (y)\| ≤ q(\|f (f1′ (y))\|) since q is a balance for f on input f1′ (y); and q(\|f (f1′ (y))\|) = q(\|y\|) since f (f1′ (y)) = y. To find a time-complexity bound for f ′ , we first compute f1′ (y) in time ≤ T (\|y\|); thereby we also verify that y ∈ Dom(f1′ ). To check whether y is in the domain of f ′ we first compare \|y\| and \|f1′ (y)\| in time ≤ \|y\| + 1. 8 Checking \|y\| ≤ q(\|f1′ (y)\|): If \|y\| ≤ \|f1′ (y)\| then we automatically have \|y\| ≤ q(\|f1′ (y)\|). If \|y\| ≥ we compute q(\|f1′ (y)\|) in time O(q(\|f1′ (y)\|)) (≤ O(q(\|y\|))), by writing the number \|f1′ (y)\| in binary, and then evaluating q (see Section 4 of [2] for a similar computation). Then we check \|y\| ≤ q(\|f1′ (y)\|) in time ≤ \|y\| + 1. Checking \|f1′ (y)\| ≤ q(\|y\|) is done in a similar way, in time ≤ O(q(\|y\|)) + \|y\| + 1. ✷ \|f1′ (y)\| P Theorem 2.8 The map evRCC q2 is complete for RM2 with respect to inversive Turing reduction. Proof. Lemma 2.1 provides the following simulation of φw by evRCC q2 : m o φw = decode ◦ ρcode(w′ ) 11 ◦ contr ◦ recontr2m ◦ evRCC q2 ◦ reexpand ◦ expand ◦ γw . To obtain an inversive Turing simulation, let e′ be any inverse of evRCC q2 . Slightly modifying the proof of Prop. 5.6 in [2], we apply e′ to any string of the form code(ex(w)) 11 0Nm+1 11 code(p) 11 z, where p ∈ φw (domC(φw )), and z ∈ A∗ ; then for any p ∈ imC(φw ) (⊆ φw (domC(φw ))), and z ∈ A∗ : e′ code(ex(w)) 11 0Nm+1 11 code(p) 11 z = code(ex(w)) 11 0Nm+1 11 code(t) 11 z, ′ ′ ′ ′ for some t ∈ φ−1 w (p) ⊆ Dom(φw ). Based on e we now construct an inverse φw of φw such that φw 4T e ; for any y ∈ Im(φw ) we define o (y). φ′w (y) = decode ◦ ρcode(w′ ) 11 ◦ contr ◦ recontr 2m ◦ e′ ◦ reexpandm ◦ expand ◦ δw o (y) is defined by Here, δw o (y) = code(w) 11 code(p) 11 z, δw o (.) is similar to γ o (.), except that δ o (.) uses imC(φ ), when y = pz with p ∈ imC(φw ), z ∈ A∗ . So, δw w w w o (.) uses domC(φ ). We saw that γ o ∈ RMP ; but unless P = NP, δ o will not be in RMP whereas γw w w 2 w 2 in general. o (y) can be computed by an RMP -machine M that makes oracle calls to Dom(e′ ) The value δw 2 and to e′ as follows. On input y, M considers all prefixes of y of increasing lengths, p1 , . . . , pk , until pj ∈ Im(φw ) is found. Since pj is the first prefix in Im(φw ), we have pj ∈ imC(φw ) and o (y) = code(w) 11 code(p ) 11 z. To test for each p whether p ∈ Im(φ ), M pads p to produce δw j i i w i Nm+1 11 code(p ). 0Nm+1 11 code(pi ); if pi ∈ Im(φw ) then e′ code(ex(w)) 11 • is defined on input 0 i Thus, if code ex(w)) 11 0Nm+1 11 code(pi ) 6∈ Dom(e′ ), then pi 6∈ Im(φw ). On the other hand, if code ex(w)) 11 0Nm+1 11 code(pi ) ∈ Dom(e′ ), then let ti ∈ φ−1 w (pi ) be such that ′ N m+1 e code(ex(w)) 11 0 11 code(pi ) = code(ex(w)) 11 0Nm+1 11 code(ti ). One oracle call to e′ yields this, and hence ti . Then we can use φw to check whether ti ∈ Dom(φw ); and this holds iff pi ∈ Im(φw ). This way, M can check whether pi ∈ Im(φw ). Thus, if y ∈ Im(φw ), M will find pj ∈ Im(φw ). When y 6∈ Im(φw ), M produces no output; this doesn’t matter since we do not care how φ′w is defined outside of Im(φw ). o (y) is known, the remaining simulation Once δw decode ◦ ρcode(w′ ) 11 ◦ contr ◦ recontr 2m ◦ e′ ◦ reexpandm ◦ expand o (y) = code(w) 11 code(p) 11 z, yields φ′ (y). of e′ , applied to δw w The function φ′w is an inverse of φw : Indeed, for x ∈ Dom(φw ), we have φw (x) = pz for some p ∈ imC(φw ), z ∈ A∗ . Then o (pz) = code(ex(w)) 11 0Nm+1 11 code(p) 11 z; reexpandm ◦ expand ◦ δw and applying e′ then yields code(ex(w)) 11 0Nm+1 11 code(t) 11 z, 9 for some t ∈ φ−1 w (p). Applying decode ◦ ρcode(w′ ) 11 ◦ contr ◦ recontr2m ′ ′ now yields tz. Finally, φw (tz) = pz, since t ∈ φ−1 w (p). So, φw φw φw (x) = φw φw (pz) = φw (tz) = pz = φw (x). ✷ P We show next that evRCC q2 is not only complete for RM2 , but for all of fP. P Proposition 2.9 The map evRCC q2 (∈ RM2 ) is complete for fP with respect to 6inv,T . Proof. By Prop. 5.6 in [2], evC q2 is complete in fP for inversive simulation. By Prop. 5.17 in [2], P C C C C C C C maps into evq2 6inv (evq2 ) . Moreover, (evq2 )C 6inv,T evRCC q2 ; indeed, (ev q2 ) ∈ RM2 (since f 7→ f CC P P C C C RM2 ), and we just saw that evRq2 is complete in RM2 . Hence evq2 6inv (evq2 ) 6inv,T evRCC ✷ q2 . 3 Non-finite generation In [2] we proved that fP is finitely generated, and we left open the question whether RMP2 is also finitely generated. We will now answer this question negatively. We will use the following general compactness property: If a semigroup S is finitely generated, and if Γ is any infinite generating set of S, then S is generated by some finite subset of this set Γ. Theorem 3.1. RMP2 is not finitely generated. Proof. We saw that RMP2 is generated by the infinite set {ρ0 , ρ1 , π0 , π1 , decode2 , contr, recontr, evRCC q2 , reexpand, expand} ∪ {γw : w is an RMP2 -program}. Let us assume, by contradiction, that RMP2 is finitely generated. Then a finite generating set can be extracted from this infinite generating set, so RMP2 is generated by Γfin = {ρ0 , ρ1 , π0 , π1 , decode2 , contr, recontr, evRCC q2 , reexpand, expand} ∪ {γi : i ∈ F }, where F is some finite set of RMP2 -programs. So for every γw there is a word in Γ∗fin that expresses γw as a finite sequence of generators. Recall that Dom(γw ) = code(w) 11 Dom(φw ), and for any x ∈ Dom(φw ), γw (code(w) 11 x) = code(w) 11 code(u) 11 v, where x = uv and u ∈ domC(φw ). The proof strategy will consist in showing that there are infinitely many functions γw that do not have a correct representation over Γfin . More precisely, for all RMP2 -programs w and all u ∈ domC(φw ), we have γw (code(w) 11 u) = code(w) 11 code(u) 11; so γw (code(w) 11 u) ∈ {00, 01}∗ 11 {00, 01}∗ 11. On the other hand, we will show that there exist (infinitely many) RMP2 -programs w such that for every X ∈ (Γfin )∗ that represents γw , there exist (infinitely many) u ∈ domC(φw ) such that: X(code(w) 11 u) = code(w) 11 code(u1 ) 11 u2 , where u2 is non-empty; so, X(code(w) 11 u) 6∈ {00, 01}∗ 11 {00, 01}∗ 11. Thus we obtain a contradiction. We consider the RMP2 -programs w such that domC(φw ) satisfies: (1) no word in domC(φw ) contains 11 as a subsegment; (2) for all i ∈ F , domC(φi ) 6= domC(φw ); (3) for any u ∈ domC(φw ) and any integer n > 0, there exists v ∈ domC(φw ) of length \|v\| > n such that u = u0 c, u0 is a prefix of v, and \|c\| ≤ 4. Equivalently: ∀u ∈ domC(φw ) ∀n > 0 ∃v ∈ domC(φw ), \|v\| > n ∃u0 , c, z ∈ A∗ [ v = u0 z, u = u0 c, \|c\| ≤ 4 ]. 10 We can picture this as a path in the tree of A∗ , labeled by u and ending at vertex u; at vertex u0 along this path, at distance ≤ 4 from vertex u, a second path branches off and ends at vertex v (of length \|v\| > n). The following family of examples shows that there exist infinitely many RMP2 -programs w that satisfy properties (1)-(3). In each of these examples (parameterized by a ∈ {0, 1}∗ ) we have domC(φw ) = {code(an ) 0010 : n > 0}, where a ∈ {0, 1}∗ is any fixed non-empty word (depending on w), chosen so that domC(φw ) 6= domC(φi ) for all i ∈ F (thus property (2) holds). Any word a that is long enough will work; indeed, for different words a the above prefix codes are different, whereas F is finite. Property (1) follows from the definition of code (namely, code(0) = 00, code(1) = 01). Property (3) holds because for every u = code(am ) 0010 and every n > 0, we can take u0 = code(am ) and v = code(an+m ) 0010. The set {code(an ) 0010 : n > 0} is a regular language, with regular expression (code(a))+ 0010. Let X ∈ Γ∗fin be a representation of γw , where w is any RMP2 -program from the family of examples above with properties (1)-(3). We will consider certain suffixes Si of X, over Γfin . Let S0 be the shortest suffix of X such that for all u ∈ domC(φw ), S0 (code(w) 11 u) is of the form code(x) 11 code(y) 11 z ∈ {00, 01}∗ 11 {00, 01}∗ 11 {0, 1}∗ . Then S0 exists since X itself (representing γw ) maps code(w) 11 u to code(w) 11 code(u) 11 ∈ {00, 01}∗ 11 {00, 01}∗ 11 {0, 1}∗ . Inductively we define S1 , S2 , . . . , Si , . . . , where Si is the shortest suffix of X that has Si−1 as a strict suffix, and such that for all u ∈ domC(φw ) we have: Si code(w) 11 u ∈ {00, 01}∗ 11 {00, 01}∗ 11 {0, 1}∗ . So, Si code(w) 11 u is of the form code(w1 ) 11 code(u1 ) 11 u2 with w1 , u1 , u2 ∈ A∗ . Then X = SN for some N ≥ 0 (and \|X\| > N ). Theorem 3.1 now follows from the next Lemma, according to which there are (infinitely many) u ∈ domC(φw ) such that SN (code(w) 11 u) = code(w) 11 code(u1 ) 11 u2 , with u2 non-empty. On the other hand, X = SN , and X represents γw , hence by the definition of γw we have for every u ∈ domC(φw ): SN (code(w) 11 u) = code(w) 11 code(u) 11; so u2 is empty. Thus, the assumption that X (over the finite generating set Γfin ) represents γw , leads to a contradiction. Lemma 3.2 Let γw be such that domC(φw ) = {code(an ) 0010 : n > 0} for some word a ∈ {0, 1}∗ , chosen so that the program w satisfies properties (1)-(3). Let X be a word over Γfin that represents γw , and let \|X\| be the length of X over Γfin . Let S0 , . . . , SN be the suffixes of X defined above, with SN = X. Then there exist ℓ and n with ℓ > n > 0 such that for all i = 0, . . . , N and all u ∈ domC(φw ) with \|u\| ≥ ℓ: Si (code(w) 11 u) = code(w1 ) 11 code(u1 ) 11 u2 , for some w1 , u1 , u2 ∈ A∗ . Moreover, u2 has a non-empty common suffix with u, and this common suffix has length at least n. Proof. We have for all u ∈ domC(φw ): Si (code(w) 11 u) = code(w1 ) 11 code(u1 ) 11 u2 , for some w1 , u1 , u2 ∈ A∗ . We want to show that there is ℓ such that for all u ∈ domC(φw ) with \|u\| ≥ ℓ: u2 has a non-empty (sufficiently long) suffix in common with u; the number n is an auxiliary parameter. We take u of the form u = code(am ) 0010 and use induction on i = 0, . . . , N . Proof for S0 : The only generators from Γfin that can occur in S0 are π0 , π1 , ρ0 , ρ1 and γj (for j ∈ F ). Indeed, the other generators in Γfin (namely decode2 , contr, recontr, evRCC q2 , reexpand, expand) are only applicable to inputs of the form code(x) 11 code(y) 11 z; so, S0 would end before a generator in {decode2 , contr, recontr, evRCC q2 , reexpand, expand} can be applied. Moreover, S0 cannot start with a generator in {decode2 , contr, recontr, evRCC q2 , reexpand, expand}; indeed, for all inputs code(w) 11 u ∈ domC(X), u = code(am ) 0010 contains no 11, so these generators are not defined on any element of domC(X). So, S0 is over {π0 , π1 , ρ0 , ρ1 } ∪ {γj : j ∈ F }. 11 The actions of π0 , π1 , ρ0 , ρ1 can change an input in at most \|S0 \| positions at the left end of the input, so these actions preserve a common suffix u2 and u of length ≥ \|u\| − \|S0 \|. Thus, if S0 consists only of instances of π0 , π1 , ρ0 , ρ1 , the Lemma holds for S0 if \|u\| ≥ ℓ ≥ n + \|S0 \| and n > 0. Suppose now that S0 contains γj for some j ∈ F . Then (if m > \|S0 \|/\|a\|), instances of π0 , π1 , ρ0 , ρ1 will transform the input u = code(w) 11 code(am ) 0010 into a word code(x) 11 s code(ak ) 0010 (for some x, s ∈ A∗ , k > 0), such that γj can be applied. This action changes an input in < \|S0 \| positions at the left end of the input. Since γj is assumed to be applicable now, we must also have x = j and s = y0 z for some y0 ∈ domC(φj ), z ∈ A∗ . Then the output of γj is γj code(j) 11 s code(ak ) 0010 = code(j) 11 code(y0 ) 11 z code(ak ) 0010, thus the common suffix of u2 and u could decrease by length ≤ \|y0 \| under the action of γj . So we let ℓ ≥ n + \|S0 \| + \|y0 \| and n > 0. Also, at most one γj (with j ∈ F ) occurs in S0 , since after γj the output is of the form code(w1 ) 11 code(u1 ) 11 u2 , which marks the end of the action of S0 . This proves the Lemma for S0 . Inductive step “Si → Si+1 ”, for 0 ≤ i < N : By induction we assume that for all u ∈ domC(φw ) with \|u\| ≥ ℓ, we have Si (code(w) 11 u) = code(w1 ) 11 code(u1 ) 11 u2 for some w1 , u1 , u2 ∈ A∗ , where u2 and u have a common suffix of length ≥ n (> 0). Let us write Si+1 = Ti+1 Si ; then Ti+1 is non-empty (by the definition of Si+1 ). We also let T0 = S0 . Claim 1: If Ti+1 contains a generator g ∈ {contr, recontr, evRCC q2 , reexpand, expand, decode2 }, then g is the first (i.e., rightmost) letter of Ti+1 , and g occurs only once. Indeed, if g were applicable later in Ti+1 , the output of the generator preceding g would be of the form code(w1 ) 11 code(u1 ) 11 u2 , so Si+1 would have ended before g was applied. Claim 2: If Ti+1 contains a generator g ∈ {contr, recontr, evRCC q2 , reexpand, expand} ∪ {γj : j ∈ F }, then g is the last (i.e., leftmost) letter of Ti+1 , and g occurs only once. Indeed, such a generator outputs a word of the form code(w1 ) 11 code(u1 ) 11 u2 . So, Si+1 ends after such a generator. As a consequence of Claims 1 and 2, if Ti+1 contains a generator g ∈ {contr, recontr, evRCC q2 , reexpand, expand}, then Ti+1 consists of just g. A generator of this form does not change u2 . So we can assume for the remaining cases that Ti+1 is of the form ti+1 , or ti+1 · decode2 , or γj · ti+1 , or γj · ti+1 · decode2 , where j ∈ F and ti+1 is over the generators π0 , π1 , ρ0 , ρ1 . Let code(w1 ) 11 code(u1 ) 11 u2 be the input of Ti+1 (and this is also the output of Si ), where u2 and u have a common suffix of length ≥ n. • Case where Ti+1 is over the generators π0 , π1 , ρ0 , ρ1 : Then Ti+1 changes the input in at most \|Ti+1 \| positions at the left end of the input, so u2 will not be affected if ℓ − n ≥ \|Ti+1 \| (and n > 0). • Case where Ti+1 = ti+1 · decode2 , with ti+1 over π0 , π1 , ρ0 , ρ1 : The output of decode2 is of the form code(w1 ) 11 u1 u2 , so the common suffix of u2 and u is preserved by decode2 . The action of ti+1 , containing only generators from {π0 , π1 , ρ0 , ρ1 }, affects at most \|ti+1 \| positions near the left side of the input, so u2 is not changed if ℓ − n ≥ \|Ti+1 \| (and n > 0). • Case where Ti+1 = γj · ti+1 , with ti+1 over π0 , π1 , ρ0 , ρ1 : Applications of π0 , π1 , ρ0 , ρ1 change fewer than \|ti+1 \| letters of the input near the left end, so the common suffix is not affected if ℓ − n ≥ \|Ti+1 \|. When γj is applied, the output produced will be of the form code(j) 11 code(yi+1 ) 11 z code(an ) 0010, where yi+1 ∈ domC(φj ). Then u2 will not be affected if we pick ℓ ≥ n + \|Ti+1 \| + \|yi+1 \| and n > 0. • Case where Ti+1 = γj · ti+1 · decode2 , with ti+1 over π0 , π1 , ρ0 , ρ1 : This case can be handled as a combination of the previous two cases. m In all the aboveP cases the constraints are fulfilled PNfor all i = 0, . . . , N , and for all u = code(a ) 0010, if N m ≥ N + \|X\| + i=0 \|yi \| (using the fact that i=0 \|Ti \| = \|X\|). Note that the words yi do not depend on the choice of the input u = code(am ) 0010, whenever m is long enough; indeed, to determine all yi we can apply each Si to the infinite word code(a)ω ∈ {0, 1}ω . ✷ 12 Notation. For a given polynomial q (of the form q(n) = a nk + a with integers a, k ≥ 1), let (q) S2 = {f ∈ RMP2 : f is computed by an RMP2 -program with built-in polynomial ≤ q}. (q) We call w an S2 -program iff w is an RMP2 -program with built-in balance and time-complexity polynomial ≤ q. (q) (q) (q) Let RM2 = hS2 i, i.e., the submonoid of RMP2 generated by the set S2 . Obviously, we have: Proposition 3.3 For any set of polynomials {qi : i ∈ N} of the form qi (n) = ai nki + ai , such that S (qi ) sup{ai : i ∈ N} = +∞ = sup{ki : i ∈ N}, we have: = RMP2 . ✷ i∈N RM2 (q) The non-finite generation result for RMP2 also holds for RM2 , and the proof is similar. We need a few preliminary facts. (q) Lemma 3.4 For every polynomial q of the form q(n) = a nk +a with a, k ≥ 2, and every S2 -program (q) w we have: γw ∈ S2 . Proof. Recall that γw (code(w) 11 uv) = code(w) 11 code(u) 11 v, where u ∈ domC(φw ). The input balance of γw is ≤ q. Indeed, the input is shorter than the output; and the output length is 2 \|w\| + 2 + 2 \|u\| + 2 + \|v\|, which is less than q(\|code(w) 11 uv\|) = q(2 \|w\| + 2 + \|u\| + 2 + \|v\|) when q(n) ≥ 2 n2 + 2. To compute code(w) 11 code(u) 11 v from input code(w) 11 uv, an RMP2 -machine can proceed as follows: First, the machine reads and outputs code(w) 11. Then it runs the program w on input uv, i.e., it simulates the corresponding RMP2 -machine Mw (which has built-in polynomial q), with an extra tape and a few modifications. While searching for a prefix of uv in domC(φw ), the longest prefix examined so far is kept on the extra tape; the output φw (u) of Mw will not be written on the output tape. Once u (the prefix of uv in domC(φw )) has been found (and written on the extra tape), code(u) 11 v is appended on the output tape. All this takes time ≤ \|code(w) 11\| + q(\|u\|) + \|code(u) 11 v\| = 2 \|w\| + 2 + q(\|u\|) + 2 \|u\| + 2 + \|v\|; this is < q(2 \|w\| + 2 + \|u\| + 2 + \|v\|) = q(\|code(w) 11 uv\|) when q(n) ≥ 2 n2 + 2. ✷ (q) Lemma 3.5 Let q be a polynomial that is larger than a certain polynomial of degree 5. Then RM2 is generated by (q) {ρ0 , ρ1 , π0 , π1 , decode2 , contr, recontr, evRCC q2 , reexpand, expand} ∪ {γz : z is an S2 -program}. (q) Proof. When w is an S2 -program then as a consequence of Lemma 2.1, m o φw = ρcode(w′ ) 11 ◦ decode2 ◦ contr ◦ recontr2m ◦ evRCC q2 ◦ reexpand ◦ expand ◦ γw , m = ρcode(w′ ) 11 ◦ decode2 ◦ contr ◦ recontr2m ◦ evRCC q2 ◦ reexpand ◦ expand ◦ γw ◦ πcode(w) 11 , (q) where q2 is a certain polynomial of degree 2. So the above generating set does indeed generate RM2 . (q) We still need to show that these generators belong to RM2 . The functions ρ0 , ρ1 , decode2 , contr, recontr, reexpand, expand, π0 , π1 have balance and complexity (q) (q) ≤ 4 (n + 1)2 . And γw ∈ RM2 if w is an S2 -program (by Lemma 3.4). Let us verify that evRCC q2 has 5 balance ≤ q2 and complexity O(n ). By definition, evRCC q2 (code(w) 11 code(u) 11 v) = code(w) 11 code(φw (u)) 11 v. Then evRCC q2 has balance ≤ q2 , since on an output of length n = 2 \|w\| + 2 + 2 \|φw (u)\| + 2, the input length is ≤ 2 \|w\| + 2 + 2 q2 (\|φw (u)\|) + 2 ≤ q2 2 \|w\| + 2 + 2 \|φw (u)\| + 2 = q2 (n). (q) When φw can be computed by an RM2 -machine with built-in polynomial pw (≤ q2 ), then 2 2 evRCC q2 (code(w) 11 code(u) 11) can be computed in time ≤ c \|w\| pw (\|u\|) ≤ c \|w\| q2 (\|u\|) , for some 5 constant c > 0 (see the proof of Prop. 4.4 in [2]). Since q2 has degree 2, evRCC q2 has complexity O(n ). (q) Thus, there exists q of degree 5 such that the above generators belong to RM2 . 13 ✷ Theorem 3.6 For any polynomial q such that q(n) = a nk + a, with k ≥ 5 and a > a0 (for some (q) constant a0 > 1), we have: RM2 is not finitely generated. (q) Proof. The proof is very similar to the proof of Theorem 3.1. We saw in Lemma 3.5 that RM2 is generated by the infinite set (q) {ρ0 , ρ1 , π0 , π1 , decode2 , contr, recontr, evRCC q2 , reexpand, expand} ∪ {γz : z is an S2 -program}. (q) Let us assume, by contradiction, that RM2 is finitely generated. Then a finite generating set can (q) be extracted from this infinite generating set; so RM2 is generated by Γfin = {ρ0 , ρ1 , π0 , π1 , decode2 , contr, recontr, evRCC q2 , reexpand, expand} ∪ {γi : i ∈ F }, (q) (q) where F is some finite set of S2 -programs. For every S2 -program w let X be a word in Γ∗fin that expresses γw as a finite sequence of generators. From here on, the proof is identical to the proof of Theorem 3.1. We use the fact that domC(φw ) = {code(an ) 0010 : n > 0} is a finite-state language, so for such a program w, γw has linear complexity (q) (being computable by a Mealy machine) and belongs to S2 . ✷ 4 4.1 Some complexity consequences of non-finite generation Hierarchy and separation (q) Proposition 4.1 Let q be a polynomial of the form q(n) = a nk + a such that a, k ≥ 1. The set S2 , (q) and hence the monoid RM2 , are contained in a finitely generated submonoid of RMP2 . Proof. Let w be a RMP2 -program such that φw has I/O-balance and time-complexity ≤ q. Then evRC q can simulate φw directly, without any need of padding and unpadding. So we have for all u ∈ domC(φw ), v ∈ A∗ : φw (uv) = ρcode(w) 11 ◦ evRC q ◦ πcode(w) 11 (uv). (q) So S2 is contained in the submonoid generated by {π0 , π1 , ρ0 , ρ1 , evRC q }. (Compare with Lemma 2.2 and the proof of Prop. 4.5 in [2].) ✷ The proof of Prop. 4.1 yields the following chain of submonoids in which non-finitely generated and finitely generated submonoids alternate. Corollary 4.2 Let . . . < qi < qi+1 < . . . be any sequence of polynomials such that for all i ≥ 0, P P qi+1 is is large enough so that evRC qi has an RM2 -program with built-in polynomial qi+1 . Then RM2 contains a strict inclusion chain, which is infinite in the upward direction, (q ) . . . $ RM2 i (q i+1 $ hπ0 , π1 , ρ0 , ρ1 , evRC qi iRMP $ RM2 2 ) $ ... ... . Proof. The strictness of the inclusions in the chain follows from the fact that non-finite generation and finite generation alternate. ✷ Theorem 4.3 Let q be a polynomial of the form q(n) = a nk + a such that a > 1, k ≥ 1. The (q) submonoid RM2 ⊆ RMP2 has the following properties: (1) (q) RM2 6= RMP2 . (q) (2) If q(n) ≥ 2 (n + 1)2 (for all n ∈ N), then RM2 contains elements of arbitrarily high polynomial balance and time-complexity. (3) (q) (q) S2 6= RM2 , if q(n) ≥ 2 (n + 1)2 . 14 (4) (q) evRC q 6∈ RM2 , if k ≥ 5 and a ≥ a0 (where a0 is as in Theorem 3.6). Moreover, evRC q has balance ≤ q, but its time-complexity is not ≤ q. (5) Let q1 , q2 be polynomials of the above form, such that q1 (n) < q2 (n) for all n ∈ N. Suppose also that q1 (n) = a nk + a with k ≥ 5 and a ≥ a0 (as in (4)), and that q2 is large enough so that (q2 ) (q1 ) (q ) evRC $ RM2 2 . q1 ∈ RM2 . Then RM2 (q) Proof. (1) Since RM2 is contained in a finitely generated submonoid of RMP2 (Prop. 4.1), and RMP2 is not contained (by ⊆) in a finitely generated submonoid of RMP2 (itself), inequality follows. 2 (2) Consider the function s : 0n 1x 7→ 02n 1x, for all n ≥ 0, x ∈ {0, 1}∗ . Then s has time-complexity ≤ 2 (n + 1)2 . Indeed, a Turing machine on input 0n 1 can read this word n times, each time turning 2 an input 0 into some new letter a, and each time writing 0n on the output tape; this produces 0n in 2 the output; then one more copy of 0n is made, followed by 1. This takes time ≤ 2 (n + 1)2 . m Then, sm (i.e., the composition of m instances of s) has complexity ≥ 2m n2 (since the output (q) length is that high, the time must be at least that much too). Thus the functions sm ∈ RM2 (as m grows) have unbounded complexity, both in their degree and in their coefficient. (q) (3) By (2), RM2 contains functions with arbitrarily high polynomial balance and time-complexity, (q) whereas S2 only contains functions with balance and complexity ≤ q. (q) (4) By Prop. 4.1, RM2 is contained in the submonoid generated by {π0 , π1 , ρ0 , ρ1 , evRC q }; and we (q) (q) (q) easily see that π0 , π1 , ρ0 , ρ1 ∈ RM2 . Hence, if evRC q belonged to RM2 , the monoid RM2 would be finitely generated, contradicting Theorem 3.6. The input balance of evRC q is ≤ q (see Lemma 3.5, where this is proved for q2 ). It follows that the (q) C time-complexity of evRC q is not ≤ q, otherwise we would have evRq ∈ RM2 . (q ) (q ) 1 2 C (5) This follows from (4) since evRC q 6∈ RM2 , but evRq ∈ RM2 . ✷ Corollary 4.4 (Strict complexity hierarchy of submonoids in RMP2 ). There exists an infinite sequence of polynomials (qi : i ∈ N), each of the form qi (n) = ai nki + ai with ki , ai > 1, and with qi (n) < qi+1 (n) for all i, n ∈ N, such that the following holds: S (q ) (q ) (qi ) RM2 i $ RM2 i+1 for all i, and = RMP2 . i∈N RM2 Moreover, RMP2 (which is not finitely generated) is the union of a ⊂-chain of 4-generated submonoids. Proof. The first statements follow from Theorem 4.3 (1) and Prop. 3.3. The last statement follows from Cor. 4.2. ✷ (q ) Since each RM2 i contains functions of arbitrarily high polynomial complexity (by Theorem 4.3 (q ) (2)), the monoids RM2 i form a strict complexity hierarchy of a new sort, different from the usual (q) complexity hierarchies. The fact that S2 6= RMP2 could have been shown by a diagonal argument. It is not clear whether classical separation techniques from complexity theory would show the results (1), (3), (4), (5) of Theorem 4.3. Remark: The monoid fP, being finitely generated, does not contain an infinite strict complexity hierarchy of monoids (but it can contain hierarchies of sets). Indeed, we have in general: Fact. A finitely generated monoid M does not contain any infinite strict ω-chain of submonoids whose union is M . S Indeed, if we had a chain (Mi : i ∈ ω) with M0 $ . . . $ Mi $ Mi+1 $ . . . . . . $ S i∈ω Mi = M , then there would exist j such that Mj contains a finite set of generators of M (since i∈ω Mi = M ). Then Mj = M , contradicting the strict hierarchy. 15 This Fact does not hold for chains over arbitrary order types; it holds for limit ordinals. The non-finitely generated monoid RMP2 contains the encoding fPC as a submonoid (see Section 3 in [2]). And fPC is finitely generated (being an isomorphic copy of fP), and fPC contains an isomorphic copy of RMP2 . This leads to non-ω strict chains of submonoids of fP and of RMP2 . 4.2 Irreducible functions (q) Another consequence of non-finite generation is that RMP2 and RM2 have “irreducible” elements, i.e., elements that cannot be expressed by composition of lower-complexity elements. We make this precise in the next definitions. In this subsection we do not use evaluation maps, so we can use “polynomials” q(n) = a nk + a where we drop the requirement that a, k are integers, i.e., we now allow real numbers ≥ 1. Definition 4.5 The inf complexity degree of f ∈ RMP2 is (q) df = inf{k ∈ R≥1 : f ∈ S2 for some polynomial q of the form q(n) = b nk + b, for some b > 1 }. We also define the inf complexity coefficient cf of f by cf = inf{Cf (ε) : ε ∈ R>0 }, where (q) Cf (ε) = inf{a ∈ R≥1 : f ∈ S2 for some polynomial q of the form q(n) = a ndf +ε + a}. The inf complexity polynomial of f is the polynomial qf given by qf (n) = cf · (ndf + 1) (for all n ∈ N). (qf ) Since df and cf are defined by infimum, f might not be in S2 following. . By the definition of inf we have the Proposition 4.6 For any polynomial q(n) = a nk + a with k > df and a > cf : (q) f ∈ RM2 . ✷ On the other hand, for every ε1 > 0, ε2 > 0: (p1 ) for any polynomial p1 (n) = b ndf −ε1 + b with any b > 1; (p2 ) where p2 (n) = (cf − ε2 ) · (ndf + 1). f 6∈ S2 f 6∈ S2 Definition 4.7 Let us choose δ1 , δ2 ∈ R>0 . A function f ∈ RMP2 is called (δ1 , δ2 )-reducible iff (q) f ∈ RM2 for some polynomial q(n) = (cf − δ2 ) · (ndf −δ1 + 1). And f is called (δ1 , δ2 )-irreducible iff f is not (δ1 , δ2 )-reducible. (q) (q) In other words, f is (δ1 , δ2 )-reducible iff f is a composite of elements of S2 i.e., f ∈ RM2 , where q(n) = (cf − δ2 ) · (ndf −δ1 + 1). So, f can be factored into functions that “have strictly lower complexity than f ” (regarding both the degree and the coefficient). Note that in the definition of df and cf we (q) (q) used S2 , not RM2 (Def. 4.5). (q ) (q ) Proposition 4.8 For all δ1 , δ2 ∈ R>0 and all polynomials q1 , q2 such that RM2 1 $ RM2 2 , there (q ) (q ) exist (δ1 , δ2 )-irreducible functions in RM2 2 − RM2 1 . (q ) (q ) Proof. By contradiction, assume that there exist δ1 , δ2 such that every f ∈ RM2 2 − RM2 1 is (q ) (δ1 , δ2 )-reducible, i.e., f can be factored as f = fm ◦ . . . ◦ f1 , where fi ∈ RM2 2 (i = 1, . . . , m) with inf degree dfi < df − δ1 and inf coefficient cfi < cf − δ2 . By the contradiction assumption, among these (q ) (q ) factors, those that are in RM2 2 − RM2 1 can themselves be factored into elements of degree and (q ) (q ) coefficient lower by amount δ1 , respectively δ2 . I.e., a factor fi ∈ RM2 2 − RM2 1 can be factored as 16 fi = fi,mi ◦ . . . fi,1 with dfi,j < dfi − δ1 and cfi,j < cfi − δ2 ; hence, dfi,j < df − 2 δ1 and cfi,j < cf − 2 δ2 , for j = 1, . . . , mi . By repeating this process we keep reducing the degree and the coefficient by at least δ1 , respectively δ2 , in each step. After a finite number of steps we obtain a factorization of f into (q ) (q ) (q ) functions in RM2 1 , contradicting the assumption that RM2 1 $ RM2 2 . ✷ Remark: A finitely generated monoid, like fP, does not contain irreducible functions of arbitrarily large complexity. Indeed, all elements are expressible as a composite of elements of bounded complexity (namely the maximum complexity of the finitely many generators). Acknowledgement: The paper benefitted from the referee’s thoughtful reading and advice. References [1] J. Berstel, D. Perrin, Theory of Codes, Academic Press (1985). [2] J.C. Birget, “Semigroups and one-way functions”, International J. of Algebra and Computation 25.1-2 (2015) 3-36. [3] J.C. Birget, “Polynomial-time right-ideal morphisms and congruences”, http://arXiv.org/abs/1511.02056 (Nov. 2015). [4] J.C. Birget, “Monoid generalizations of the Richard Thompson groups”, J. of Pure and Applied Algebra 213(2) (2009) 264-278. [5] J.C. Birget, “Circuits, coNP-completeness, and the groups of Richard Thompson”, International J. of Algebra and Computation 16(1) (Feb. 2006) 35-90. [6] J.C. Birget, “The groups of Richard Thompson and complexity”, International J. of Algebra and Computation 14(5,6) (Dec. 2004) 569-626. [7] J.W. Cannon, W.J. Floyd, W.R. Parry, “Introductory notes on Richard Thompson’s groups”, L’Enseignement Mathématique 42 (1996) 215-256. [8] W. Diffie, M. Hellman, “New directions in cryptography”, IEEE Trans. Information Theory 22.6 (1976) 644-654. [9] D.Z. Du, K.I. Ko, Theory of Computational Complexity, Wiley (2000). [10] O. Goldreich, Foundations of Cryptography: Basic Tools, Cambridge U.P. (2001). [11] L.H. Hemaspaandra, M. Ogihara, The Complexity Theory Companion, Springer (2002). [12] G. Higman, “Finitely presented infinite simple groups”, Notes on Pure Mathematics 8, The Australian National University, Canberra (1974). [13] L. Levin, “The tale of one-way functions”, Problemy Peredatshi Informatsii 39.1 (2003) 92-103. [14] R. McKenzie, R.J. Thompson, “An elementary construction of unsolvable word problems in group theory”, in Word Problems, (W. Boone, F. Cannonito, R. Lyndon, editors), North-Holland (1973) pp. 457-478. [15] Ch. Papadimitriou, Computational Complexity, Addison-Wesley (1994). [16] Richard J. Thompson, “Embeddings into finitely generated simple groups which preserve the word problem”, in Word Problems II, (S. Adian, W. Boone, G. Higman, editors), North-Holland (1980) pp. 401-441. 17	4math.GR
Simple Problems The Simplicial Gluing Structure of Pareto Sets and Pareto Fronts arXiv:1709.10377v1 [math.OC] 18 Apr 2017 Naoki Hamada Fujitsu Laboratories Ltd. 4-1-1 Kamikodanaka, Nakahara-ku Kawasaki, Japan hamada-naoki@jp.fujitsu.com ABSTRACT 1.2 Related Studies in Other Fields Quite a few studies on real-world applications of multi-objective optimization reported that their Pareto sets and Pareto fronts form a topological simplex. Such a class of problems was recently named the simple problems, and their Pareto set and Pareto front were observed to have a gluing structure similar to the faces of a simplex. This paper gives a theoretical justiﬁcation for that observation by proving the gluing structure of the Pareto sets/fronts of subproblems of a simple problem. The simplicity of standard benchmark problems is studied. Theories concerning the easiness of covering solutions have been developed in several optimization-related ﬁelds as well as the EMO community. Most of them are studies on topological properties of solution sets. 1.2.1 Contractibility. The earliest work can be found in 1951 by Koopmans [11] Assertion 4.14 in which he applied the linear programming to economics and pointed out some conditions making the Pareto front contractible. Peleg [18] generalized this result and showed that the Pareto front is contractible if the feasible objective region is a convex set. Afterward, the study spread to operations CCS CONCEPTS research, and the closedness, the (arcwise) connectedness, and the •Applied computing → Multi-criterion optimization and decisioncontractibility of the Pareto set/front under general settings were making; •Mathematics of computing → Evolutionary algostudied on linear programming in 1970’s and on (quasi) convex prorithms; Nonconvex optimization; Geometric topology; gramming in 1980’s. These results are collected in Luc [15] Section 6. Recently, similar results were obtained under more general probKEYWORDS lem classes such as the lexicographic quasiconvexity [20] and the multi-objective optimization, continuous optimization, problem class, arcwise cone-quasiconvexity [12]. The (arcwise) connectedness of stratiﬁcation the Pareto set is a necessary condition that the homotopy method covers the solutions. ACM Reference format: Naoki Hamada. 2017. Simple Problems. In Proceedings of GECCO ’17 Companion, Berlin, Germany, July 15-19, 2017, 10 pages. DOI: http://dx.doi.org/10.1145/3067695.3076069 1 INTRODUCTION 1.1 Motivation The success of evolutionary multi-objective optimization (EMO) is widely spreading over various academic and industrial ﬁelds. Recent numerical studies showed that decomposition-based EMO algorithms such as MOEA/D [33], NSGA-III [3], and AWA [6–8, 24, 25] have an ability to approximate the entire Pareto set and Pareto front of many-objective problems. In contrast to their abundance of experimental successes, the theory shedding light on why they work is still under developing. Especially, the problem class in which decomposition-based EMO algorithms can cover the entire Pareto set/front has not been understood. This paper discusses some problem class in which solutions are well-behaved for scalarization. Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proﬁt or commercial advantage and that copies bear this notice and the full citation on the ﬁrst page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). GECCO ’17 Companion, Berlin, Germany © 2017 Copyright held by the owner/author(s). 978-1-4503-4939-0/17/07. . . $15.00 DOI: http://dx.doi.org/10.1145/3067695.3076069 1.2.2 Decomposition. The decomposition approach considers not only a given problem but also its subproblems each optimizing a subset of objective functions and studies the relation among their solutions. Lowe et al. [14] in 1984 showed that the weak Pareto set of a convex programming is the union of the Pareto sets of subproblems. Malivert et al. [16] extended the result to explicitly quasiconvex upper semicontinuous functions. Popovici [19] named this property the Pareto reducible and gave a suﬃcient condition independent of convexity. Ward [32] showed that the strictly Pareto solutions to a convex programming problem are completely surrounded by the Pareto solutions of subproblems. Recent studies [12, 20, 21] revealed that the Pareto reducibility of the lexicographic quasiconvex programming problem is closely related to the contractibility, through the simply shadiness [1] of the Pareto front. 1.2.3 Stratification. From pure mathematics, the singularity theory of diﬀerentiable maps gives a decomposition of solutions. In 1973, Smale [26] applied this theory to an economic problem and stated that the Pareto set of a pure exchange economy of m agents is homeomorphic to an (m − 1)-simplex, provided the quasiconvexity and monotonicity of the agents’ utility functions. Lovison et al. [13] pointed out that each face of this simplex corresponds to the Pareto set of each subproblem optimizing a subset of objective functions. In Smale [26] and its sequels [27–31], he discussed the stratiﬁcation of Pareto critical points of generic maps GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany N. Hamada with the transversality and rank assumption of derivatives. de Melo [2] showed that the C ∞ -maps whose Pareto critical points admit a stratiﬁcation are generic, i.e., they form a dense subset of the space of C ∞ -maps under the Whitney topology. Recently, Lovison et al. [13] collected related works to this approach and further developed de Melo’s result, and showed that local Pareto sets of suﬃciently proper maps admit a Whitney stratiﬁcation. 1.3 Our Approach These attempts are going on in two courses: the linear/convex analysis originated by Koopmans and the global analysis by Smale. The former seems much restrictive for the global optimization nature of EMO algorithms. The latter approach is suﬃciently general but currently hard to compute. We need a handy theory for understanding the behavior of EMO algorithms. Recently, Hamada et al. [8] deﬁned a class of problems called the simple problem. They pointed out, without rigorous proofs, that the Pareto set and Pareto front of a simple problem are both homeomorphic to a simplex and the faces of the simplex correspond to the Pareto sets and their images of the subproblems. They also discussed that this property is closely related to scalarization. This paper gives rigorous proofs for their arguments. 1.4 Contribution (1), and the evaluation map f , depending on the context. We call the empty set ∅ a 0-objective problem and deﬁne its evaluation map as the empty map ∅ : ∅ → R0 to a one-point set R0 = {0}. For problems f , д such that д ⊆ f as sets, we say that д is a subproblem of f and f is a superproblem of д. We call the set of all subproblems of a problem f the decomposition of f and denote by 2 f = {д \| д ⊆ f }. If solutions x, y ∈ X to a problem f satisfy the conditions ∀fi ∈ f , fi (x) ≤ fi (y) and ∃f j ∈ f , f j (x) < f j (y), then we say that x f -dominates y and denote it by x ≺ f y. We denote f (x) = f (y) by x = f y and (x ≺ f y) ∨ (x = f y) by x f y. If a solution x ∗ ∈ X to a problem f satisﬁes ∀x ∈ X , x 6≺ f x ∗ , then x ∗ is called a Pareto solution to f . The set of all Pareto solutions of a problem f = { f 1 , . . . , fm } is called the Pareto set, denoted by X ∗ (f ) or X ∗ (f 1 , . . . , fm ). The image of X ∗ (f ) by a map д is denoted by дX ∗ (f ). Especially, the image f X ∗ (f ) is called the Pareto front. Through the paper, we abbreviate the composition of mapsψ ◦ϕ toψϕ. The above notation, дX ∗ , can be considered as д ◦ X ∗ by regarding X ∗ as a map X ∗ : 2 F → 2X . In this paper, we give a proof that the boundary of the Pareto set (resp. Pareto front) of a simple problem is the union of the interior of the Pareto sets (resp. their images) of subproblems. This property enables us to numerically compute a stratiﬁcation of the Pareto set/front. Additionally, we investigate the simplicity of benchmark problems widely-used in the EMO community: all problems in ZDT suite [34] and DTLZ suite [4] are non-simple, ﬁve of problems in WFG suite [10] can be simple under a very restrictive situation, and MED problem [7] is always simple. 3 SIMPLE PROBLEM 1.5 First, we present the deﬁnition of simple problem and its graphical intuition. Contents The rest of the paper is organized as follows. Section 2 prepares basic notions and notations used in subsequent sections. Section 3 gives some properties of solutions to simple problems and their relation to scalarization. Section 4 discusses the simplicity of existing benchmark problems. Section 5 gives conclusions and remarks for future work. 2 PRELIMINARIES This paper considers the following n-variable m-objective problem: minimize f (x) = (f 1 (x), . . . , fm (x)) n x ∈X ⊆R : Rn → (1) where we call f Rm the evaluation map, fi : Rn → R (i = 1, . . . ,m) the i-th objective function, Rn the variable space, X ⊆ Rn the feasible region, x ∈ Rn a solution, Rm the objective space, and f (x) ∈ Rm an evaluation value. We will make various problems by removing some of objective functions from (1) and discuss a gluing structure of their solutions. To write such arguments clearly, we deﬁne the problem as a ﬁnite set of objective functions F = {f 1 , . . . , fm } and regard (1) as a notation for it. We abuse f to denote the problem F , the equation This section introduces the deﬁnition of simple problem and shows its solution structure. Section 3.1 presents the deﬁnition of simple problem. Section 3.2 shows some inclusion properties of solutions among subproblems. Section 3.3 shows that those solutions have the gluing structure of a topological simplex. Section 3.4 points out that this gluing structure enables decomposition-based EMO algorithms to cover the Pareto set and Pareto front. 3.1 Deﬁnition Deﬁnition 3.1 (simple problem [8]). A problem f is simple or has simplicity if every subproblem д ∈ 2 f satisﬁes the following conditions: (S1) X ∗ (д) ≈ ∆k−1 if д is k-objective, (S2) д\|X ∗ (д) : X ∗ (д) → Rk is an embedding. n o Í Here, ∆k−1 = (x 1 , . . . , xk ) ∈ [0, 1]k \| xi = 1 is the standard (k − 1)-simplex. By A ≈ B, we denote that topological spaces A and B are homeomorphic, which is deﬁned as there are continuous maps ψ : A → B and ϕ : B → A such that ϕψ = idA and ψϕ = idB . Such maps, ψ and ϕ, are called homeomorphisms. The topology of X ∗ (д) is induced from the variable space Rn . That is, any open set U in X ∗ (д) can be written as U = V ∩ X ∗ (д) with some open set V in Rn under the Euclidean topology. Similarly, all spaces discussed in this paper are implicitly equipped with induced topologies from either the variable space Rn or the objective space Rm . By ϕ\|A′ : A′ → B, we denote the restriction of a map ϕ : A → B to a set A′ ⊆ A, which is the composite ϕι : A′ → B with the inclusion map ι : A′ ֒→ A, a 7→ a. The embedding of A to Simple Problems GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany X(f1, f2, f3) 2 f1(x) f2(x) 1.5 minimize x 1 , x 2 ∈[−2, 2] 1 RGB(x ) = (y1 (x ), y2 (x ), y3 (x )) ( (f i ∈ д) f i − f i (x ) / f i − f i yi (x ) = 0 (f i < д) f i = max f i (x ) \| x ∈ X ∗ (д) + 1e-16 f i = min f i (x ) \| x ∈ X ∗ (д) f3(x) 0.5 + 3(x 2 − 1)2 x 12 x2 f 1 (x ) = x 22 f 2 (x ) = 2(x 1 − 1)2 + 0 -0.5 -1 f 3 (x ) = 3(x 1 + 1)2 + 2(x 2 + 1)2 -1.5 -2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 x1 X(f1, f2) 2 1.5 1.5 0.5 0.5 0.5 x2 1 -0.5 0 -0.5 -1 -1 -1 -1.5 -1.5 -2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -2 -1.5 -1 -0.5 x1 0 0.5 1 1.5 2 -2 1.5 1.5 0.5 0.5 -0.5 x2 0.5 x2 1 0 0 -0.5 -1 -1 -1.5 -1.5 -2 0 0.5 1 1.5 2 1 1.5 2 X(f3) f3(x) 0 -1 -0.5 0.5 -0.5 -1.5 -2 0 1.5 1 -1 -0.5 2 f2(x) 1 -1.5 -1 x1 X(f2) 2 f1(x) -2 -1.5 x1 X(f1) 2 x2 0 -0.5 -1.5 -2 f1(x) f3(x) 1.5 1 0 X(f3, f1) 2 f2(x) f3(x) 1 x2 x2 X(f2, f3) 2 f1(x) f2(x) -2 -2 -1.5 -1 x1 -0.5 0 x1 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 x1 Figure 1: A simple 3-objective problem f = { f 1 , f 2 , f 3 } and its subproblem д’s Pareto set X ∗ (д). The Pareto sets are colored by converting the f 1 , f 2 , f 3 -coordinates to RGB using the equations in the top-right cell. are homeomorphic to ∆0 , a point! A map on a one-point set always satisﬁes (S2); there is no special implication. Finally, there is a 0-objective problem (not shown in Fig. 1) that corresponds to the Let us cultivate the intuitive understanding of this deﬁnition case of no objective function. This is just required for formality. with Fig. 1. When considering a simple problem, we also deal with Note that the conditions (S1) and (S2) only impose the problem its subproblems optimizing all subsets of given objective functions. structure within the Pareto set and admit an arbitrary structure The ﬁrst subproblem is the original problem itself optimizing all out of the Pareto set. This is contrastive to conventional probthe objectives. The example in the ﬁgure is 3-objective, and the lem classes such as linear/convex/polynomial programming probcondition (S1) imposes that its Pareto set, X ∗ (f 1 , f 2 , f 3 ), is homeolems which regulate their structures over the entire domain. The morphic to ∆2 , a surface created by bending and stretching a trisimplicity is independent of those problem classes. In fact, evangle (without cutting and connecting). The condition (S2) guarery class of linear/convex/polynomial programming problems conantees that the restricted evaluation map f : X ∗ (f 1 , f 2 , f 3 ) → tains both of simple problems and non-simple problems! As a ref X ∗ (f 1 , f 2 , f 3 ) and its inverse map f −1 : f X ∗ (f 1 , f 2 , f 3 ) → X ∗ (f 1 , f 2 , f 3 ) sult, the simplicity characterizes a new aspect of “easiness to solve”. are bijective, continuous, and thus homeomorphisms. This implies 3.2 Inclusion Properties that the Pareto front, f X ∗ (f 1 , f 2 , f 3 ), is also homeomorphic to ∆2 , and every point on f X ∗ (f 1 , f 2 , f 3 ) continuously corresponds to a This section shows inclusion relations of the Pareto sets and its unique solution on X ∗ (f 1 , f 2 , f 3 ) and vice versa. Next, we remove images for subproblems of a simple problem. To isolate the conone of objectives, resulting in three 2-objective subproblems. Their sequence of assuming the simplicity from general properties of Pareto sets, X ∗ (f 1 , f 2 ), X ∗ (f 2 , f 3 ), X ∗ (f 3 , f 1 ), are homeomorphic to Pareto-optimality, we begin without the assumption. ∆1 by (S1), a curve without loops. By (S2), their Pareto fronts are Proposition 3.2. For any problem f (possibly non-simple) and also a curve in which each point continuously corresponds to a any subproblem д ⊆ f , the following relations hold: unique Pareto solution. Again removing an objective, we get three X ∗ (д) ⊆ X w (д) ⊆ X w (f ) 1-objective subproblems. Their Pareto sets, X ∗ (f 1 ), X ∗ (f 2 ), X ∗ (f 3 ), B is the composite ιϕ : A → B of a homeomorphism ϕ : A → B ′ and the inclusion map ι : B ′ ֒→ B. GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany N. Hamada where X w (д) is the weak Pareto set of д, that is, the set of points each x ∈ X satisfying ∀y ∈ X , ∃fi ∈ д : fi (x) ≤ fi (y). (2) Proof. The ﬁrst inclusion, X ∗ (д) ⊆ X w (д), is a well-known fact (see, for example, Miettinen [17] Section 2.5). The second relation, X w (д) ⊆ X w (f ), directly follows from (2). In the deﬁnition of simplicity, the conditions (S1) and (S2) are imposed on all subproblems as well as a given problem, which means that the subproblems inherit the simplicity from the superproblem. Proposition 3.7. If a problem f is simple, then any subproblem д ⊆ f is simple. Proof. If д ⊆ f , then 2д ⊆ 2 f . Thus, if all the problems in 2 f satisfy (S1) and (S2), then also do in 2д , which implies that if f is simple, then д is. However, it does not hold that X ∗ (д) ⊆ X ∗ (f ) in general. Example 3.3. Consider a 1-variable 2-objective problem minimize f 1 (x) = 0, f 2 (x) = x. x ∈[0, 1] Clearly, the Pareto sets are X ∗ (f 1 ) = [0, 1] and X ∗ (f 1 , f 2 ) = {0}, which implies X ∗ (f 1 ) X ∗ (f 1 , f 2 ). In contrast, simple problems do not have such ill-behaved solutions. Proposition 3.4. If a problem f is simple, then X w (f ) = X ∗ (f ). Proof. From Proposition 3.2, we have X w (f ) ⊇ X ∗ (f ). We will prove X w (f ) ⊆ X ∗ (f ) by contradiction. Suppose that a point x ∈ X w (f ) \ X ∗ (f ) exists. Then, since x is weakly Pareto-optimal, the condition ∀y ∈ X , ∃fi ∈ f : fi (x) ≤ fi (y) (3) holds. On the other hand, since x is not Pareto-optimal, the condition ∃y ∈ X \ {x } , ∀fi ∈ f : fi (y) ≤ fi (x) (4) holds. A point y in (4) should satisfy (3), which means ∃fi ∈ f : fi (y) = fi (x). By (3), we have x ∈ X ∗ (fi ) and thus y ∈ X ∗ (fi ). This contradicts (S2); more speciﬁcally, fi cannot be an injection on X ∗ (fi ). Therefore, the simplicity ensures the inclusion relationship of Pareto sets. Proposition 3.5. For a simple problem f and any subproblem д ⊆ f , it holds that X ∗ (д) ⊆ X ∗ (f ). Proof. Combine Proposition 3.2 with Proposition 3.4. Therefore, propositions that hold for a simple problem also hold for its subproblems. For example, the actual assertion of Proposition 3.4 is that any subproblem of f , as well as f itself, does not have a weakly Pareto-optimal point which is not Pareto-optimal. The interpretation of Proposition 3.6 is bit more complicated: given a simple problem h, the assertion holds for any pair of problems f , д such that д ⊆ f ⊆ h. Henceforth, we will not repeat this property, but it is always valid when propositions involve simple problems. Similarly to that the empty set is a subset of every set, a 0objective problem is a subproblem of every problem. Therefore, if 0-objective problems do not exist or are not simple, then there cannot exist any simple problem. Let us check them. Proposition 3.8. There exists a 0-objective problem ∅; it is unique and simple. Proof. The existence and the uniqueness follow from those of the empty set and empty map. Let us check the simplicity. Since the only subproblem of ∅ is ∅ itself, it suﬃces to show that ∅ satisﬁes (S1) and (S2). (S1): Since ∅ is the 0-objective problem, it suﬃces to show X ∗ (∅) ≈ −1 ∆ . It holds indeed as ∆−1 = ∅ and X ∗ (∅) = ∅. (S2): By X ∗ (∅) = ∅, the restricted evaluation map, ∅\|X ∗ (∅) : ∗ X (∅) → R0 , is the evaluation map ∅ : ∅ → R0 itself. Since ∅ : ∅ → R0 can be decomposed into a homeomorphism id : ∅ → ∅ and an inclusion map ι : ∅ ֒→ R0 as ∅ = ιid, the restriction ∅\|X ∗ (∅) : X ∗ (∅) → R0 is an embedding. Using this fact, we can see the topology of the image of the Pareto set. Although, the condition (S1) itself addresses only the topology of the Pareto set, combined with (S2), we show that the Pareto front and the image in the superproblem have the same topology. Now we have conﬁrmed that for each subproblem, the Pareto set and its image are well-behaved. The next section investigates that those solutions are nicely glued together. Proposition 3.6. For a simple problem f and any subproblem д ⊆ f , if д is k-objective, then The goal of this section is to give a proof that the solutions to a simple problem have a special gluing structure as shown in Figs. 2 and 3. This structure is an analogy of the faces of a simplex. A 2-simplex [v 1 ,v 2 , v 3 ] is a triangle spanned by vertices v 1 ,v 2 ,v 3 and its boundary ∂[v 1 ,v 2 ,v 3 ] is the union of three edges, or 1simplices, [v 1 ,v 2 ], [v 2 , v 3 ], [v 3 ,v 1 ]. Each edge [vi , v j ] has the boundary consisting of two points, or 0-simplices, [vi ] and [v j ]. The boundary of each vertex [vk ] is the empty set, or the (−1)-simplex. We can expand these relations using A = int A ⊔ ∂A and see that the boundary of a simplex is expressed as the disjoint union of the X ∗ (д) ≈ дX ∗ (д) ≈ f X ∗ (д) ≈ ∆k−1 . Proof. By Deﬁnition 3.1, X ∗ (д) ≈ дX ∗ (д) ≈ ∆k−1 is evident. We will show the nontrivial part: X ∗ (д) ≈ f X ∗ (д). It suﬃces to show that the restriction f \|X ∗ (д) is an embedding. Remember that the restriction f \|X ∗ (f ) is an embedding by deﬁnition and X ∗ (д) ⊆ X ∗ (f ) in Proposition 3.5. In general, any restriction of an embedding is again an embedding. 3.3 Gluing Properties Simple Problems GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany open faces: X( f3) ∂[v 1 ,v 2 ,v 3 ] = int[v 1 ,v 2 ] ⊔ int[v 2 ,v 3 ] ⊔ int[v 3 ,v 1 ] X( f2, f3) ⊔ int[v 1 ] ⊔ int[v 2 ] ⊔ int[v 3 ]. Generally, Pareto sets and their images may have a more complex topological structure. However, for a simple problem, we have seen in Proposition 3.6 that those form topological manifolds with boundary (hereafter, simply call manifolds) homeomorphic to a simplex. In a k-manifold M, a point having an open neighborhood homeomorphic to Rk is called an interior point, and the set of all interior points is called the interior of M, denoted by int M. The other points are boundary points having an open neighborhood homeomorphic to [0, ∞) × Rk−1 , and the set of all boundary points is called the boundary of M, denoted by ∂M. For a simple problem, a similar relation holds among Pareto sets as shown in Fig. 2: ∂X ∗ (f 1 , f 2 , f 3 ) = int X ∗ (f 1 , f 2 ) ⊔ int X ∗ (f 2 , f 3 ) ⊔ int X ∗ (f 3 , f 1 ) ⊔ int X ∗ (f 1 ) ⊔ int X ∗ (f 2 ) ⊔ int X ∗ (f 3 ). (5) X( f3, f1) X( f1, f2, f3) X( f2) X( f1 , f2) X( f1) Figure 2: A simple 3-objective problem f = {f 1 , f 2 , f 3 } and its gluing structure of the Pareto sets X ∗ (д) of subproblems д ⊆ f . The boundary of the Pareto set X ∗ (д) of each subproblem д ⊆ f consists of the Pareto sets X ∗ (h) of all subproblems h ⊂ д. fX( f2) fX( f1 , f2) The same relation holds for the images as shown in Fig. 3: fX( f1) ∂ f X ∗ (f 1 , f 2 , f 3 ) = int f X ∗ (f 1 , f 2 ) ⊔ int f X ∗ (f 2 , f 3 ) ⊔ int f X ∗ (f 3 , f 1 ) ⊔ int f X ∗ (f 1 ) ⊔ int f X ∗ (f 2 ) ⊔ int f X ∗ (f 3 ). fX( f1, f2, f3) (6) It is known that such a gluing structure of solutions commonly appears in facility location problems, studied for a long time (see for example Rodrı́guez-Chı́a et al. [22] and the references therein). It is also seen in other applications and exploited as a heuristic for practitioners [5]. Now we start to show (5) and (6) for an arbitrary number of objectives. First of all, let us see some basic properties of the Pareto front that hold for any class of problems. Lemma 3.9. For an m-objective (possibly non-simple) problem f whose Pareto front f X ∗ (f ) forms an (m − 1)-manifold, the projection Rm → Ri −1 × {0} × Rm−i ≃ Rm−1 , π−i : (y1 , . . . , ym ) 7→ (y1 , . . . , yi −1, 0, yi +1, . . . , ym ) restricted to int f X ∗ (f ) is an embedding. Proof. Generally, any projection and its restriction to any open set are continuous and open, and any injective continuous open map is an embedding. Thus, it suﬃces to show that π−i is injective when restricted to int f X ∗ (f ). If π−i \|int f X ∗ (f ) is not injective, then int f X ∗ (f ) contains two points having the same coordinates except for the i-th value. This means that one point f -dominates another, contradicting the deﬁnition of the Pareto front f X ∗ (f ). This lemma asserts that the interior of the Pareto front can be ﬂattened while keeping its topology and ordering. As shown in Fig. 4, the projection induces a coordinate neighborhood (π−i , U ) at any point y ∈ int f X ∗ (f ) where for any points y, y ′ ∈ U , it holds that y j ≤ y j′ ⇔ π−i (y)j ≤ π−i (y ′)j for all j , i. This property is the key to investigate the interaction between the topology and the dominance-ordering on the interior of the Pareto front. fX( f2, f3) fX( f3, f1) fX( f3) Figure 3: A simple 3-objective problem f = {f 1 , f 2 , f 3 } and its gluing structure of the Pareto set images f X ∗ (д) for all subproblems д. Although the shape is diﬀerent from Fig. 2, the topology is the same. Lemma 3.10. Consider an m-objective (possibly non-simple) problem f whose Pareto front f X ∗ (f ) forms an (m − 1)-manifold. The following statement holds for any y ∈ int f X ∗ (f ): any neighborhood U ⊆ f X ∗ (f ) of y has a point that д-dominates y and is (f \ д)dominated by y for each д such that ∅ ⊂ д ⊂ f . Proof. By Lemma 3.9, the projection π−i restricted to f X ∗ (f ) is an embedding, and thus π−i (U ) is a neighborhood of a point π−i (y) in Rm−1 . Then, we can take an (m − 1)-hyper-cube centered at π−i (y) in f X ∗ (f ) and can write its vertices as (y1 ± ε, . . . , yi −1 ± ε,yi +1 ± ε, . . . , ym ± ε). Here, ε is a small positive number and signs, ±, run over all possible combinations. Among these vertices, let π−i (v) be the one such that coordinates related to д ⊆ f −i = f \ { fi } are −ε and related to f −i \ д are +ε. Thus, it holds that y ≻д v and y ≺(f −i \д) v. Especially, д = f −i implies y ≺fi v and д = ∅ implies y ≻fi v. Otherwise, it means that y ≺ f v or y ≻ f v, contradicting y,v ∈ f X ∗ (f ). Repeating the above argument for all i, we complete the proof. GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany π-1 π-2 1 0.5 f3 fX*(f) 0 U π-3 0 y 0.2 0.4 0.6 0 0.2 0.4 f2 f1 0.8 0.6 0.8 1 1 Figure 4: A 3-objective (possibly non-simple) problem f = { f 1 , f 2 , f 3 } and its Pareto front projections. The projection π−i restricted to int f X ∗ (f ) is injective and thus an embedding. By this fact, any interior point y and its neighborhood U in f X ∗ (f ) are mapped to an interior point π−i (y) and its neighborhood π−i (U ) in π−i f X ∗ (f ). Other than the Pareto front f X ∗ (f ), do Lemmas 3.9 and 3.10 extend to the Pareto set image f X ∗ (д) of a subproblem д ⊂ f ? The answer is NO: for general problems, the projection π−i is not an embedding of int f X ∗ (д). Example 3.11. Again consider the problem in Example 3.3: minimize f 1 (x) = 0, f 2 (x) = x. x ∈[0, 1] The Pareto set of the subproblem f 1 is X ∗ (f 1 ) = [0, 1], and its image is f X ∗ (f 1 ) = {(0, y) \| 0 ≤ y ≤ 1}. The interior int f X ∗ (f 1 ) = {(0, y) \| 0 < y < 1} can be projected to π−2 int f X ∗ (f 1 ) = f 1X ∗ (f 1 ) = {0}, which implies π−2 \|int f X ∗ (f1 ) is not injective and thus not an embedding. Furthermore, since X ∗ (f 1 , f 2 ) = {0}, we see that the set X ∗ (f 1 ) \ X ∗ (f 1 , f 2 ) = (0, 1] is weakly Pareto-optimal and not Pareto-optimal in f . The existence of these weak Pareto optima disrupts the injectivity of π−2 . Contrary, if the problem is simple, then there are no solutions that are weakly Pareto-optimal and not Pareto-optimal, which enable us to extend Lemmas 3.9 and 3.10 to the image f X ∗ (д). Corollary 3.12. Consider a simple problem f and a subproblem д ⊆ f . The restriction of the projection π−i to int f X ∗ (д) is an embedding. Proof. If π−i \|int f X ∗ (д) is not injective, then int f X ∗ (д) contains two points that have the same coordinates except for the i-th value. This means that f has a weak Pareto solution which is nonPareto, contradicting Proposition 3.4. Corollary 3.13. Consider a simple problem f and a subproblem д ⊆ f . For any y ∈ int f X ∗ (д), the following statement holds: for any neighborhood U ⊆ f X ∗ (д) of y, there exists a point that h-dominates y and is (д \ h)-dominated by y for each h such that ∅ ⊂ h ⊂ д. N. Hamada Proof. Chose an objective function fi ∈ f \ д and let the remaining set be f −i = f \{ fi }. By Corollary 3.12, f X ∗ (д) is mapped to f −i X ∗ (д) by π−i , homeomorphically. Next, chose another f j ∈ f −i \д and let the remainder be f −i j = f −i \ f j . Again by Corollary 3.12, f −i X ∗ (д) is mapped to f −i j X ∗ (д) by π−j , homeomorphically. Though the repeated application of projections as long as it is an embedding, the original Pareto set image is ﬁnally mapped to дX ∗ (д). Let the composite of used projections π−i , π−j , . . . be π−∗ . Generally, the composite of embeddings is again an embedding. Thus, a point y ∈ int f X ∗ (д) and its neighborhood U ⊆ f X ∗ (д) is mapped homeomorphically to a point π−∗ (y) ∈ int дX ∗ (д) and its neighborhood π−∗ (U ) ⊆ дX ∗ (д). Proposition 3.6 asserts дX ∗ (д) ≈ ∆ \|д \|−1 , which together with Lemma 3.10 completes the proof. By this property, we can see the simplicity ensures that Pareto set images of subproblems are located on the boundary of their superproblems. Lemma 3.14. For a simple problem f and subproblems д, h such that h ⊂ д ⊆ f , the following relation holds: ∂ f X ∗ (д) ⊇ f X ∗ (h). Proof. Suppose there exists a point y ∈ f X ∗ (h) that is an interior point of f X ∗ (д). Now, y ∈ int f X ∗ (д) and h ⊂ д holds and thus by Corollary 3.13, some neighborhood U ⊆ f X ∗ (д) of y has a point h-dominating y. This contradicts y ∈ f X ∗ (h). Our next question is whether a similar relation holds for the Pareto set before mapped by f . To check this, we need the following lemma: Lemma 3.15. For a simple problem f and a subproblem д ⊆ f , the map f commutes with the boundary and interior: f ∂X ∗ (д) = ∂ f X ∗ (д), ∗ ∗ f int X (д) = int f X (д). (7) (8) Proof. Generally, an embedding maps boundary to boundary and interior to interior. Now f is an embedding on X ∗ (д) ⊆ X ∗ (f ), and thus it holds that f ∂ = ∂ f and f int = int f . Using this fact, we show that under the simplicity, Pareto sets of subproblems have the same relation as its images. Corollary 3.16. For a simple problem f and a (proper!) subproblem д ⊂ f , the following relation holds: ∂X ∗ (f ) ⊇ X ∗ (д). Proof. Since f is an embedding on X ∗ (f ), there is the inverse map f −1 : f X ∗ (f ) → X ∗ (f ). Lemma 3.15 converts Lemma 3.14 to the assertion as follows: ∂ f X ∗ (f ) ⊇ f X ∗ (д) ⇔ f ∂X ∗ (f ) ⊇ f X ∗ (д) −1 ⇔ f f ∂X ∗ (f ) ⊇ f −1 f X ∗ (д) ⇔ ∂X ∗ (f ) ⊇ X ∗ (д). The last key to the main theorem is the sphere embedding. Lemma 3.17. Every embedding f : S n → S n is surjective and thus a homeomorphism where S n is an n-sphere S n = x ∈ Rn+1 \| kx k = 1 . Simple Problems GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany Proof. Suppose f is not surjective. Then, there exists a point y ∈ S n \ f (S n ) and a stereographic projection with north pole y, denoted by π : S n \ {y} → Rn . Generally, any stereographic projection is an embedding, and the composite of embeddings is an embedding. Therefore, π f is an embedding of S n into Rn . This contradicts the well-known fact that S n cannot be embedded into Rn . Remark 1. To keep the proof elementary, here we assumed that S n ֒6 → Rn is known and derived that S n ֒→ S n is surjective. There is an alternative proof deriving both by the Mayer-Vietoris exact sequence in a uniﬁed fashion. Consult Hatcher [9], the two paragraphs after the proof of Proposition 2B.1 (pp. 169–170). Now, we show the goal of this section. Theorem 3.18. For a simple problem f and a subproblem д ⊆ f , it holds that ∂X ∗ (д) = Ä int X ∗ (h), (9) h ⊂д ∂ f X ∗ (д) = Ä int f X ∗ (h). Proof. First, we show (10). What to be proven are: (a) if h , h ′ , then int f X ∗ (h) ∩ int f X ∗ (h ′ ) = ∅, Ð (b) ∂ f X ∗ (д) = h ⊂д int f X ∗ (h). (a) When h ⊃ h ′, it holds from Lemma 3.14 that ∂ f X ∗ (h) ⊇ f X ∗ (h ′ ). The same holds for the inverse case h ⊂ h ′. We thus consider the case there is no inclusion relation between h and h ′. Assume there exists a point y ∈ int f X ∗ (h) ∩ int f X ∗ (h ′ ), and let U be a neighborhood ofy in f X ∗ (h ∪h ′ ). If there exists a point z ∈ U \ f X ∗ (h) ∪ f X ∗ (h ′ ) , then y ≺h z and y ≺h ′ z hold, implying y ≺(h∪h′ ) z. This contradicts z ∈ U ⊆ f X ∗ (h ∪ h ′ ). If z does not exist, then the dimension of U must be equal to that of f X ∗ (h) or f X ∗ (h ′ ), which contradicts dim U = h ∪ h ′ − 1. Consequently, such y cannot exist. (b) Since Lemma 3.14 states ∂ f X ∗ (д) ⊇ f X ∗ (h), we have ∂ f X ∗ (д) ⊇ f X ∗ (h) h ⊂д ⊇ Ø From the second to third lines, we used the general property of a map f (A) ∪ f (B) = f (A ∪ B). 3.4 Relation to Scalarization Equations (7)–(10) together deﬁne a gluing structure of the Pareto sets and their images of subproblems of a simple problem. This structure induces a natural stratiﬁcation1 of the Pareto set (resp. the Pareto front) where each stratum is the interior of the Pareto set (resp. its image) of a subproblem. Therefore, we can numerically compute the stratiﬁcation by solving each subproblem. Points spreading over all strata can be a good covering of the Pareto set/front. To see why this structure enables decomposition-based EMO algorithms to cover the Pareto set/front, consider the weighted Tchebyshev-norm scalarization minimize fw (x) = max wi (fi (x) − z i ) (10) h ⊂д Ø We can get (9) by converting (10) with Lamma 3.15: Ã ∂ f X ∗ (д) = h ⊂д int f X ∗ (h) Ã ∗ ⇔ f ∂X (д) = h ⊂д f int X ∗ (h) Ã ⇔ f ∂X ∗ (д) = f h ⊂д int X ∗ (h) −1 −1 Ã ∗ ⇔ f f ∂X (д) = f f h ⊂д int X ∗ (h) Ã ⇔ ∂X ∗ (д) = h ⊂д int X ∗ (h). int f X ∗ (h). h ⊂д Ð Thus, there is the inclusion map ι : h ⊂д int f X ∗ (h) ֒→ ∂ f X ∗ (д). Generally, any inclusion map is an embedding. Combining (a) with Lemma 3.14, we have f X ∗ (h) ∩ f X ∗ (h ′ ) = f X ∗ (h ∩ h ′ ). Therefore, Ð ∗ h ⊂д int f X (h) is the union of manifolds each homeomorphic to a simplex, which are glued as the faces of a simplex. This fact enÐ sures h ⊂д int f X ∗ (h) ≈ ∂∆ \|д \|−1 ≈ S \|д \|−2 . Contrary, f X ∗ (д) ≈ ∆ \|д \|−1 implies ∂ f X ∗ (д) ≈ ∂∆ \|д \|−1 ≈ S \|д \|−2 . As Lemma 3.17 ensures that S n ֒→ S n is surjective, the inclusion map ι is surjective, Ð implying that h ⊂д int f X ∗ (h) = ∂ f X ∗ (д) holds. i x ∈X (11) where the weight w = (w 1 , . . . , wm ) is chosen from ∆m−1 and the utopian point is ﬁxed to be z i = inf x ∈X fi (x). Let ei be the i-th standard base in Rm whose i-th coordinate is one and the other coordinates are zero. The standard (m − 1)-simplex is rewritten as ∆m−1 = [e 1 , . . . , em ]. Using the notation of the weight-optima correspondence Ø X ∗ (fw ), S(W ) = w ∈W a well-known fact of the optima to (11) can be written as S([ei 1 , . . . , ei k ]) = X w (fi 1 , . . . , fi k ) (12) for any choice of an arbitrary number of indices i 1 , . . . , i k ∈ {1, . . . ,m}. If the problem is simple, then we can go further: Proposition 3.4 extends (12) to S([ei 1 , . . . , ei k ]) = X ∗ (fi 1 , . . . , fi k ), and by Corollary 3.16 we have S(∂[ei 1 , . . . , ei k ]) = ∂X ∗ (fi 1 , . . . , fi k ). Therefore, a weight on each face gives a boundary point of each stratum with corresponding indices. Unfortunately, the L∞ -norm, as well as other existing scalarization methods including the weighted sum, the augmented Chebyshevnorm, PBI [33], and IPBI [23], does NOT give the correspondence between the interiors: S(int[ei 1 , . . . , ei k ]) , int X ∗ (fi 1 , . . . , fi k ). Nevertheless, once boundary points of a stratum are obtained, we can ﬁnd new weights corresponding to interior points of the stratum by interpolating the weights used for the boundary points. Thus, the grid arrangement or divide-and-conquer generation of weights over [ei 1 , . . . , ei k ] practically often hit interior points of X ∗ (fi 1 , . . . , fi k ). 1 The smoothness of the stratiﬁcation is determined from that of the evaluation map. GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany 4 SIMPLICITY OF BENCHMARKS This section investigates the simplicity of benchmark problems widely-used in the EMO community: ZDT suite [34], DTLZ suite [4], WFG suite [10], and MED problem [7]. 4.1 N. Hamada Proof. Every problem has an objective function ignoring some variables as follows: DTLZ1–6: fm (y, z) = (1 + д(z))f (y1 ) DTLZ7: f 1 (y, z) = y1 DTLZ8, 9: f 1 (y, z) = y10.1 + y20.1 + · · · + y 0.1 ⌊n/m ⌋ ZDT Suite The ZDT suite has six n-variable 2-objective problems named ZDT1– 6. The decision variables are split into the position variables y = (y1 , . . . , yk ) (0 < k < n) and the distance variables z = (z 1 , . . . , z l ) (l = n − k), deﬁning the problems in the following uniﬁed format: minimize f (y, z) = (f 1 (y, z), f 2 (y, z)), Therefore, DTLZ1–9 are all non-simple. Furthermore, Huband et al. [10] Table VII shows that DTLZ1–6 has f that is not injective on X ∗ (f ) and DTLZ7 has a disconnected Pareto front. This is another evidence for the non-simplicity of DTLZ1–7. (y,z)∈Y ×Z 4.3 WFG Suite where f 1 (y, z) = f (y1 ), f 2 (y, z) = д(z)h(f 1 (y),д(z)), The WFG suite contains nine problems, WFG1–9, having the form: minimize f (x) = (f 1 (x), . . . , fm (x)), X = Y × Z = (Y1 × · · · × Yk ) × (Z 1 × · · · × Zl ), Y1 = [0, 1], ( [−5, 5] Y2 , . . . , Yk , Z 1 , . . . , Zl = [0, 1] (y,z)∈Y ×Z where fi (x) = xm + 2i × hi (x 1 , . . . , xm−1 ), Y → X 1 × · · · × Xm−1 , tY : (y1, . . . , yk ) 7→ (x 1 , . . . , xm−1 ), Z → Xm , tZ : (z 1, . . . , z l ) 7→ xm , (ZDT4), (otherwise). Users can make diﬀerent problems by changing placeholder functions f , д, h. For the concrete speciﬁcation of f , д, h for ZDT1–6, see Zitzler et al. [34]. The above general formulas are enough to show that the problems are non-simple. X = X 1 × · · · × Xm = [0, 1]m , Y = [0, 2] × [0, 4] × · · · × [0, 2k], Z = [0, 2k + 2] × [0, 2k + 4] × · · · × [0, 2n]. Theorem 4.1. ZDT1–6 are all non-simple, independent of the choice of variable dimension n and position-variable dimension k. Additionally, this suite cannot create simple problems no matter how f , д, h are speciﬁed unless their domains are modiﬁed. Proof. First, we exclude ZDT5 from the following analysis since it is a binary-variable problem which is clearly non-simple. Then, for ZDT1–4, 6, the function f deﬁning f 1 depends on a single variable, y1 . The other variables can take an arbitrary value on the optima of f , and thus X ∗ (f 1 ) = Y1∗ (f ) × Y2 × · · · × Yk × Z . This means X ∗ (f 1 ) 6≈ ∆0 and contradicts the simplicity condition (S1). Consequently, ZDT1–6 are all non-simple. The reason why this suite cannot be simple is that f depends only on y1 . Generally, when the problem has an objective function independent of some variables, its Pareto set extends to higher dimensions than usual, contradicting the simplicity condition (S1). The existence of unused variables is a quick test for non-simplicity. 4.2 DTLZ Suite The DTLZ suite consists of nine problems named DTLZ1–9. Their decision variables are split into position variables y and distance variables z, as ZDT, but the number of objectives m can be set arbitrarily. See Deb et al. [4] for deﬁnition. Theorem 4.2. DTLZ1–9 are all non-simple, independent of the choice of variable dimension n, objective dimension m, and positionvariable dimension k. The functions h 1 , . . . , hm , tY , t Z are placeholders. In this suite, the variables y and z are mapped by the transformation functions tY and t Z to the position variables x 1 , . . . , xm−1 and the distance variable xm , then passed to the objective functions fi . For this reason, y and z are called the position-related variables and the distancerelated variables, respectively. For the concrete deﬁnition, see Huband et al. [10]. Theorem 4.3. WFG2, 4, 5, 9 are always non-simple. WFG1, 3, 6–8 are simple if and only if the dimension of the position-related variables y is k = 1. Here, one can set k = 1 only when the number of objectives is m = 2 because Huband et al. [10] Table XIV shows that these problems require k mod (m − 1) = 0. Proof. First, consider WFG2. Huband et al. [10] Table XIV shows that WFG2 has a disconnected Pareto front. Such a front cannot be homeomorphic to ∆m−1 , which contradicts the property of a simple problem shown in Proposition 3.6. The following discussion treats the rest of the problems. Let us check the properties of the Pareto set X ∗ (f ) in the transformed variable space X = [0, 1]m and the map f on X ∗ (f ). As described in Huband et al. [10], all problems2 have X ∗ (f ) = [0, 1]m−1 × {0} . By the properties of the shape functions hi shown in Huband et al. [10] Table X, f is an embedding upon X ∗ (f ) if and only if m = 2.3 For 2 As opposed to Huband et al. [10], WFG2 actually has a diﬀerent Pareto set in which x 1 is conditioned to be Pareto-optimal. To avoid a complication caused by this diﬀerence, we ﬁrst ﬁnished WFG2. 3 Except for Disconnected used only in WFG2. Simple Problems GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany 4.4 MED m > 2, it holds that f (0, x 2 , . . . , xm−1 , 0) = 0, and thus f is not injective on X ∗ (f ). Since the transformation functions, tY and t Z , are surjective as described in Huband et al. [10], the composite evaluation map f t : (tY ,tZ ) This is a single problem, MED, deﬁned as follows: minimize f (x) = (f 1 (x), . . . , fm (x)), n x ∈X =R where fi (x) = x − xi∗ xi∗ f Y × Z −−−−−−→ X − → Rm is not injective on the Pareto set (Y × Z )∗ (f t) in the untransformed variable space Y × Z for m > 2. This does not meet the simplicity condition (S2). In the case of k > m −1, the transformation function tY involves Reduction: Weighted Sum or Reduction: Non-separable to decrease the dimension of Y . By examining Huband et al. [10] Table XI, we can see that both functions are not injective. Therefore, tY maps two diﬀerent points y, y ′ ∈ Y to the same Pareto solution (x 1 , . . . , xm−1 ) ∈ [0, 1]m−1 , contradicting the simplicity condition (S2). There remains the case k = 1 and m = 2 where the problems may be simple. First, let us consider the 1-objective subproblems. From Huband et al. [10] Table X, we have X ∗ (f 1 ) = {(0, 0)} , X ∗ (f 2 ) = {(1, 0)} . This implies that the simplicity condition (S1) for these problems is equivalent to the following criterion: tY−1 (0), tY−1 (1), t Z−1 (0) are all a point. Since (S2) automatically follows from (S1) when the problem is 1objective, the above criterion is a necessary and suﬃcient condition for the simplicity. We can see from Huband et al. [10] Table XI that tY−1 (0) and t Z−1 (0) are always a point, but it depends on the case whether tY−1 (1) is a point or not. WFG4, 5, 9 introduce Shift: Deceptive or Shift: Multi-modal into tY , making tY−1 (1) not a point. Thus, these problems are non-simple. For WFG1, 3, 6–8, tY−1 (1) becomes a point, and their 1-objective subproblems are simple. Next, let us consider their 2-objective subproblems. It holds that X ∗ (f 1 , f 2 ) = [0, 1] × {0} , and it has been conﬁrmed that f = {f 1 , f 2 } is an embedding on X ∗ (f 1 , f 2 ) and t Z−1 (0) is a point. Therefore, the equivalent condition to the simplicity is as follows: tY−1 : [0, 1] → [0, 2] embedding. For k = 1 and m = 2, the transformation function tY in WFG1, 3, 6–8 can be simpliﬁed into the form tY (y) = y α . Thus, its inverse, tY−1 (x) = x 1/α , is an embedding. Now, we have checked that WFG1, 3, 6–8 are simple if and only if k = 1 and m = 2. Note that Huband et al. [10] Table XIV describes that WFG3 has a degenerate Pareto front, which seems to be an evidence that WFG3 is always non-simple. However, the degeneracy actually occurs only when m > 2. In our analysis for m = 2, the Pareto front of WFG3 forms a line segment, which does not disrupt the simplicity. pi , = (0, . . . , 0, 1 , 0, . . . , 0), \| {z } \| {z } i −1 (13) n−i 0 < pi < ∞. The front-shape parameters, pi , which determine the convexity of the Pareto front, are user-speciﬁed parameters as well as the variable dimension n and the number of objectives m. Theorem 4.4. MED is always simple independent of the choice of parameters n,m,pi . Additionally, changing individual optima, xi∗ , does not break the simplicity as long as they are aﬃnely independent. Proof. First, consider the case of pi = 1. This corresponds to a facility location problem under the L2 -norm. The Pareto set of this ∗ [32]. Thus, if problem is known as the convex hull of x 1∗ , . . . , xm ∗ ∗ m ≤ n + 1 holds and x 1 , . . . , xm are aﬃnely independent, then the ∗ . Indeed, convex hull is the (m − 1)-simplex spanned by x 1∗ , . . . , xm by the deﬁnition of xi∗ , this problem can be deﬁned only when ∗ are aﬃnely independent. Thus, the Pareto m ≤ n, and x 1∗ , . . . , xm ∗ set X (f ) is an (m − 1)-simplex, which ensures that the problem f satisﬁes the simplicity condition (S1). Analyzing the gradient of the L2 -norm, we can see that the map f is an embedding on X ∗ (f ), which satisﬁes (S2). The same argument applies to the subproblems, conﬁrming that they satisfy (S1) and (S2); the problem f is simple. The case pi , 1 can be considered as the composite of fi in the case of pi = 1 and the pi -th power. Since any positive power is an order-preserving homeomorphism [0, ∞) → [0, ∞), the composition preserves the simplicity of the facility location problem. 5 CONCLUSIONS In this paper, we have discussed the simple problem and showed that the Pareto sets of its subproblems (resp. their images) constitute a stratiﬁcation of its Pareto set (resp. its Pareto front). This topological property gives a theoretical guarantee that decompositionbased EMO algorithms can obtain an entire approximation of the Pareto set as well as the Pareto front. We have also investigated the simplicity of benchmark problems widely-used in the EMO community. All problems in the ZDT and DTLZ suites are non-simple. The WFG suite contains ﬁve simple problems under a very restrictive situation but usually does not, whereas the MED problem is always simple. We believe that the absence of simple problems in the standard benchmark suites is a considerable gap between the benchmark and the real-world since there are many evidences that a large portion of nowadays applications seems to be simple. Additionally, real-world applications involving simulations can be black-box; it would be important to develop an estimation method for the simplicity of black-box problems from a ﬁnite set of approximate solutions. GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany REFERENCES [1] J. Benoist. 2003. Contractibility of Eﬃcient Frontier of Simply Shaded Sets. Journal of Global Optimization 25, 3 (2003), 321–335. DOI: http://dx.doi.org/10.1023/A:1022448811953 [2] W. de Melo. 1976. On the Structure of the Pareto Set of Generic Mappings. Boletim da Sociedade Brasileira de Matemática – Bulletin of the Brazilian Mathematical Society 7, 2 (1976), 121–126. DOI:http://dx.doi.org/10.1007/BF02584786 [3] K. Deb and H. Jain. 2014. An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point-Based Nondominated Sorting Approach, Part I: Solving Problems With Box Constraints. IEEE Transactions on Evolutionary Computation 18, 4 (Aug 2014), 577–601. DOI: http://dx.doi.org/10.1109/TEVC.2013.2281535 [4] K. Deb, L. Thiele, M. Laumanns, and E. Zitzler. 2005. Scalable Test Problems for Evolutionary Multiobjective Optimization. In Evolutionary Multiobjective Optimization, A. Abraham, L. Jain, and R. Goldberg (Eds.). Springer London, 105–145. DOI:http://dx.doi.org/10.1007/1-84628-137-7 6 [5] R.M. Everson, D.J. Walker, and J.E. Fieldsend. 2013. Edges of Mutually Nondominating Sets. In Proceedings of the 15th Annual Conference on Genetic and Evolutionary Computation (GECCO ’13). ACM, New York, NY, USA, 607–614. DOI:http://dx.doi.org/10.1145/2463372.2463452 [6] N. Hamada, Y. Nagata, S. Kobayashi, and I. Ono. 2010. Adaptive Weighted Aggregation: A Multiobjective Function Optimization Framework Taking Account of Spread and Evenness of Approximate Solutions. In Proceedings of the 2010 IEEE Congress on Evolutionary Computation (CEC 2010). 787–794. [7] N. Hamada, Y. Nagata, S. Kobayashi, and I. Ono. 2011. Adaptive Weighted Aggregation 2: More Scalable AWA for Multiobjective Function Optimization. In Proceedings of the 2011 IEEE Congress on Evolutionary Computation (CEC 2011). 2375–2382. [8] N. Hamada, Y. Nagata, S. Kobayashi, and I. Ono. 2011. On Scalability of Adaptive Weighted Aggregation for Multiobjective Function Optimization. In Proceedings of the 2011 IEEE Congress on Evolutionary Computation (CEC 2011). 669–678. [9] A. Hatcher. 2002. Algebraic Topology. Cambridge University Press, Cambridge, New York. http://opac.inria.fr/record=b1122188 [10] S. Huband, P. Hingston, L. Barone, and L. While. 2006. A Review of Multiobjective Test Problems and a Scalable Test Problem Toolkit. IEEE Transactions on Evolutionary Computation 10, 5 (Oct. 2006), 477–506. DOI: http://dx.doi.org/10.1109/TEVC.2005.861417 [11] T.C. Koopmans. 1951. Analysis of Production as an Eﬃcient Combination of Activities. In Activity Analysis of Production and Allocation (Proceedings of a Conference) (Cowles Commission Monograph), T.C. Koopmans (Ed.). John Wiley & Sons, Inc., New York, 33–97. [12] D. La Torre and N. Popovici. 2010. Arcwise Cone-Quasiconvex Multicriteria Optimization. Operations Research Letters 38, 2 (2010), 143–146. DOI: http://dx.doi.org/10.1016/j.orl.2009.11.003 [13] A. Lovison and F. Pecci. 2014. Hierarchical Stratiﬁcation of Pareto Sets. ArXiv e-prints (July 2014). arXiv:math.OC/1407.1755 http://arxiv.org/abs/1407.1755. [14] T.J. Lowe, J.-F. Thisse, J.E. Ward, and R.E. Wendell. 1984. On Eﬃcient Solutions to Multiple Objective Mathematical Programs. Management Science 30, 11 (1984), 1346–1349. DOI:http://dx.doi.org/10.1287/mnsc.30.11.1346 [15] D.T. Luc. 1989. Theory of Vector Optimization. Lecture Notes in Economics and Mathematical Systems, Vol. 319. Springer-Verlag. [16] C. Malivert and N. Boissard. 1994. Structure of Eﬃcient Sets for Strictly Quasi Convex Objectives. Journal of Convex Analysis 1 (1994), 143–150. [17] K.M. Miettinen. 1999. Nonlinear Multiobjective Optimization. International Series in Operations Research & Management Science, Vol. 12. Springer-Verlag, GmbH. [18] B. Peleg. 1972. Topological Properties of the Eﬃcient Point Set. Proc. Amer. Math. Soc. 35, 2 (1972), 531–536. http://www.jstor.org/stable/2037641 [19] N. Popovici. 2005. Pareto Reducible Multicriteria Optimization Problems. Optimization 54, 3 (2005), 253–263. DOI: http://dx.doi.org/10.1080/02331930500096213 [20] N. Popovici. 2006. Structure of Eﬃcient Sets in Lexicographic Quasiconvex Multicriteria Optimization. Operations Research Letters 34, 2 (2006), 142–148. DOI: http://dx.doi.org/10.1016/j.orl.2005.03.003 [21] N. Popovici. 2007. Explicitly quasiconvex set-valued optimization. Journal of Global Optimization 38, 1 (2007), 103–118. DOI: http://dx.doi.org/10.1007/s10898-006-9085-1 [22] A.M. Rodrı́guez-Chı́a and J. Puerto. 2002. Geometrical Description of the Weakly Eﬃcient Solution Set for Multicriteria Location Problems. Annals of Operations Research 111 (2002), 181–196. DOI:http://dx.doi.org/10.1023/A:1020905820371 [23] H. Sato. 2014. Inverted PBI in MOEA/D and Its Impact on the Search Performance on Multi and Many-objective Optimization. In Proceedings of the 2014 Annual Conference on Genetic and Evolutionary Computation (GECCO ’14). ACM, New York, NY, USA, 645–652. DOI:http://dx.doi.org/10.1145/2576768.2598297 [24] T. Shioda and I. Ono. 2015. Adaptive Weighted Aggregation with Enhanced Relocation and Its Performance Evaluation. Transaction of the Japanese Society for Evolutionary Computation 6, 2 (2015), 104–117. DOI: N. Hamada http://dx.doi.org/10.11394/tjpnsec.6.104 [25] T. Shioda and I. Ono. 2015. Adaptive Weighted Aggregation with Step Size Control Weight Adaptation for Multiobjective Continuous Function Optimization. SICE Journal of Control, Measurement, and System Integration 8, 5 (2015), 303– 311. DOI:http://dx.doi.org/10.9746/jcmsi.8.303 [26] S. Smale. 1973. Global Analysis and Economics I: Pareto Optimum and a Generalization of Morse Theory. In Dynamical Systems, M. M. Peixoto (Ed.). Academic Press, 531–544. DOI:http://dx.doi.org/10.1016/B978-0-12-550350-1.50044-8 [27] S. Smale. 1974. Global Analysis and Economics IIA: Extension of a Theorem of Debreu. Journal of Mathematical Economics 1, 1 (1974), 1–14. DOI: http://dx.doi.org/10.1016/0304-4068(74)90031-7 [28] S. Smale. 1974. Global Analysis and Economics III: Pareto Optima and Price Equilibria. Journal of Mathematical Economics 1, 2 (1974), 107–117. DOI: http://dx.doi.org/10.1016/0304-4068(74)90002-0 [29] S. Smale. 1974. Global Analysis and Economics IV: Finiteness and Stability of Equilibria with General Consumption Sets and Production. Journal of Mathematical Economics 1, 2 (1974), 119–127. DOI: http://dx.doi.org/10.1016/0304-4068(74)90003-2 [30] S. Smale. 1974. Global Analysis and Economics V: Pareto Theory with Constraints. Journal of Mathematical Economics 1, 3 (1974), 213–221. DOI: http://dx.doi.org/10.1016/0304-4068(74)90013-5 [31] S. Smale. 1976. Global Analysis and Economics VI: Geometric Analysis of Pareto Optima and Price Equilibria under Classical Hypotheses. Journal of Mathematical Economics 3, 1 (1976), 1–14. DOI: http://dx.doi.org/10.1016/0304-4068(76)90002-1 [32] J. Ward. 1989. Structure of Eﬃcient Sets for Convex Objectives. Mathematics of Operations Research 14, 2 (1989), 249–257. [33] Q. Zhang and H. Li. 2007. MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition. IEEE Transactions on Evolutionary Computation 11, 6 (December 2007), 712–731. DOI:http://dx.doi.org/10.1109/TEVC.2007.892759 [34] E. Zitzler, K. Deb, and L. Thiele. 2000. Comparison of Multiobjective Evolutionary Algorithms: Empirical Results. IEEE Transaction on Evolutionary Computation 8, 2 (2000), 173–195. DOI:http://dx.doi.org/10.1162/106365600568202	9cs.NE
Distributed Spanner Approximation Keren Censor-Hillel∗ Michal Dory∗ arXiv:1802.03160v1 [cs.DS] 9 Feb 2018 February 12, 2018 Abstract We address the fundamental network design problem of constructing approximate minimum spanners. Our contributions are for the distributed setting, providing both algorithmic and hardness results. Our main hardness result shows that √an α-approximation for the minimum directed kspanner √ √problem for k ≥ 5 requires Ω(n/ α log n) rounds using deterministic algorithms or Ω( n/ α log n) rounds using randomized ones, in the Congest model of distributed computing. Combined with the constant-round O(n )-approximation algorithm in the Local model of [Barenboim, Elkin and Gavoille, 2016], as well as a polylog-round (1 + )-approximation algorithm in the Local model that we show here, our lower bounds for the Congest model imply a strict separation between the Local and Congest models. Notably, to the best of our knowledge, this is the first separation between these models for a local approximation problem. Similarly, a separation between the directed and undirected cases is implied. We also prove that the minimum weighted k-spanner problem for k ≥ 4 requires a near-linear number of rounds in the Congest model, for directed or undirected graphs. In addition, we show lower bounds for the minimum weighted 2-spanner problem in the Congest and Local models. On the algorithmic side, apart from the aforementioned (1 + )-approximation algorithm for minimum k-spanners, our main contribution is a new distributed construction of minimum 2-spanners that uses only polynomial local computations. Our algorithm has a guaranteed approximation ratio of O(log(m/n)) for a graph with n vertices and m edges, which matches the best known ratio for polynomial time sequential algorithms [Kortsarz and Peleg, 1994], and is tight if we restrict ourselves to polynomial local computations. An algorithm with this approximation factor was not previously known for the distributed setting. The number of rounds required for our algorithm is O(log n log ∆) w.h.p, where ∆ is the maximum degree in the graph. Our approach allows us to extend our algorithm to work also for the directed, weighted, and client-server variants of the problem. It also provides a Congest algorithm for the minimum dominating set problem, with a guaranteed O(log ∆) approximation ratio. ∗ Technion, Department of Computer Science, {ckeren,smichald}@cs.technion.ac.il. Supported in part by the Israel Science Foundation (grant 1696/14). 1 Introduction A k-spanner of a graph G is a sparse subgraph of G that preserves distances up to a multiplicative factor of k. First introduced in the late 80’s [56, 57], spanners have been central for numerous applications, such as synchronization [2, 3, 57], compact routing tables [4, 12, 58, 64], distance oracles [6, 62, 65], approximate shortest paths [25, 32], and more. Due to the prominence of spanners for many distributed applications, it is vital to have distributed algorithms for constructing them. Indeed, there are many efficient distributed algorithms for finding sparse spanners in undirected graphs, which give a global guarantee on the size of the spanner. A prime example are algorithms that construct (2k − 1)-spanners with O(n1+1/k ) edges, for a graph with n vertices [7, 17, 18, 28, 40], which is optimal in the worst case assuming Erdős’s girth conjecture [33]. As opposed to finding spanners with the best worst-case sparsity, this paper focuses on the network design problem of approximating the minimum k-spanner, which is a fundamental optimization problem. This is particularly crucial for cases in which the worst-case sparsity is Θ(n2 ) such as 2spanners (complete bipartite graphs) or directed spanners. Spanner approximation is at the heart of a rich line of recent work in the sequential setting, presenting approximation algorithms [8,14,15,21,23], as well as hardness of approximation results [19, 31, 45]. There are only few distributed spanner approximation algorithms known to date. A distributed algorithm with an expected approximation ratio of O(log n) for the minimum 2-spanner problem is e √n) for given in [21]. This was recently extended to k > 2, achieving an approximation ratio of O( directed k-spanners [22], which matches the best approximation known in the sequential setting [8]. Yet, in the distributed setting, it is possible to obtain better approximations if local computation is not polynomially bounded. A constant time O(n )-approximation algorithm for directed or undirected minimum k-spanner, which takes exp(O(1/)) + O(k) rounds for any constant > 0 and a positive integer k, is given in [5]. In addition, we show a polylogarithmic time (1+)-approximation algorithm for these problems, following the framework of a recent algorithm for covering problems [39] (see Section 6). This approximation is much better than the best approximation that can be acheived in the sequential setting, due to the hardness results of [19, 31]. All these algorithms work in the classic Local model of distributed computing [51], where vertices exchange messages of unbounded size in synchronous rounds. A natural question is whether we can obtain good approximations efficiently also in the Congest model [54], where the messages exchanged are bounded by O(log n) bits. In the undirected case, efficient constructions of (2k − 1)-spanners with O(n1+1/k ) edges in the Congest model [7, 28] imply O(n1/k )-approximations, since any spanner of a connected graph has at least n − 1 edges. However, for directed graphs there are no efficient algorithms in the Congest model. Our contribution in this paper is twofold. We provide the first hardness of approximation results for minimum k-spanners in the distributed setting. Our main hardness result shows that there are no efficient approximation algorithms for the directed k-spanner problem for k ≥ 5 in the Congest model. This explains why all the current approximation algorithms for the problem require large messages, and also creates a strict separation between the directed and undirected variants of the problem, as the latter admits efficient approximations in the Congest model. In addition, we provide new distributed algorithms for approximating the minimum k-spanner problem and several variants in the Local model. Our main algorithmic contributaion is an algorithm for minimum 2-spanners that uses only polynomial local computations and guarantees an approximation ratio of O(log m n ), which matches the best known approximation for polynomial sequential algorithms [46]. On the way to obtaining our results, we develop new techniques, both algorithmically and for obtaining our lower bounds, which can potentially find use in studying various related problems. 1 1.1 1.1.1 Our contributions Hardness of approximation We show several negative results implying hardness of approximating various spanner problems in both the Local and Congest models. While there are many recent hardness of approximation results for spanner problems in the sequential setting [14, 19, 31, 45], to the best of our knowledge ours are the first for the distributed setting. (I.) Directed k-spanner for k ≥ 5 in the CONGEST model: Perhaps our main negative result is a proof for the hardness of approximating the directed k-spanner problem for k ≥ 5 in the Congest model. Theorem 1.1. Any (perhaps randomized) distributed α-approximation algorithm in the Congest √ n n . model for the directed k-spanner problem for k ≥ 5 takes Ω( √α·log n ) rounds, for 1 ≤ α ≤ 100 When restricting attention to deterministic algorithms, we prove a stronger lower bound of e √ Ω( nα ), for any α ≤ nc for a constant c > 1. For example, this gives that a constant or a polylogarithmic approximation ratio for the directed e √n) rounds using randomized algorithms or k-spanner problem in the Congest model requires Ω( e Ω(n) rounds using deterministic algorithms. Even an approximation ratio of only n is hard and e 1/2−/2 ) rounds using randomized algorithms or Ω(n e 1−/2 ) rounds using deterministic requires Ω(n ones, for any 0 < < 1. Moreover, in the deterministic case, even an approximation ratio of nc , for e √n) rounds. This is to be contrasted with an approximation of n, appropriate values of c, requires Ω( which can be obtained without any communication by taking the entire graph, since any k-spanner has at least n − 1 edges. LOCAL vs. CONGEST. The major implication of the above is a strict separation between the Local and Congest models, since the former admits a constant-round O(n )-approximation algorithm [5]1 and a polylogarithmic (1 + )-approximation algorithm (see Section 6) for directed k-spanners. Such a separation was previously known only for global problems (problems that are subject to an Ω(D) lower bound, where D is the diameter of the graph), and for local decision problems (such as determining whether the graph contains a k-cycle). To the best of our knowledge, ours is the first separation for a local approximation problem. Directed vs. undirected. Our lower bound also separates the undirected and directed kspanner problems, since there are efficient algorithms in the Congest model for constructing (2k − 1)-spanners with O(n1+1/k ) edges [28, 40] which imply an O(n1/k )-approximation. The best randomized algorithm for the task takes k rounds [28], and the best deterministic algorithm is a recent algorithm which takes O(n1/2−1/k ) rounds for a constant even k [40]. Achieving the e 1/2−1/2k ) rounds using randomization, or same approximation for directed graphs necessitates Ω(n e 1−1/2k ) rounds using deterministic algorithms. Ω(n (II.) Weighted k-spanner for k ≥ 4 in the CONGEST model: In addition to the above main result, we consider weighted k-spanners, and show that any α-approximation for the weighted e n ) rounds, and that Ω(n) e undirected k-spanner problem for k ≥ 4 requires Ω( rounds are needed for k the weighted directed k-spanner problem. 1 In [5], a constant time randomized algorithm for directed k-spanner is presented. However, the deterministic network decomposition presented in [5] gives a polylogarithmic deterministic approximation for directed k-spanner as well, which shows the separation also for the deterministic case. 2 Weighted vs. unweighted. As these lower bounds hold also for randomized algorithms, we obtain yet another separation, between the weighted and the unweighted variants of the problem, since the aforementioned k-round (2k − 1)-spanner constructions imply an O(n1/k ) approximation for the unweighted case. LOCAL vs. CONGEST. Since both the constant-round algorithm for approximating kspanners within a factor of O(n ) [5] and the (1 + )-approximation algorithm that we give in Section 6 are suitable for the weighted case, our hardness result for the weighted case implies the separation between the Local and Congest models also when having weights. This holds also for the undirected weighted case. (III.) Weighted 2-spanner in the LOCAL and CONGEST models: Finally, we show lower bounds for the weighted 2-spanner problem, which, in a nutshell, are obtained by a reduction that captures the intuition that approximating the minimum weight 2-spanner is at least as hard as approximating the minimum vertex cover (MVC). We emphasize that the reduction from the set cover problem to the unweighted 2-spanner problem given in [45] is inherently sequential, by requiring the addition of a vertex that is connected to all other vertices in the graph, and hence is unsuitable for the distributed setting. q Our reduction implies that Ω( logloglog∆∆ ) or Ω( logloglogn n ) rounds are required for a logarithmic approximation ratio for weighted 2-spanner in the Local model, by plugging in the lower bounds e 2 ) lower bound for an exact solution for MVC given in [48]. In addition, our reduction implies an Ω(n for weighted 2-spanner in the Congest model, by using the near-quadratic lower bound for exact MVC given recently in [11]. This is tight up to logarithmic factors since O(n2 ) rounds allow learning the entire graph topology and solving essentially all natural graph problems. 1.1.2 Distributed approximation algorithms We show new distributed algorithms for approximating minimum k-spanners. Our main algorithmic contribution is a new algorithm for the minimum 2-spanner problem that uses only polynomial local computations (see Section 4). In addition, we show that if local computation is not polynomially bounded it is possible to achieve (1 + )-approximation for minimum k-spanners (see Section 6). (I.) Distributed (1 + )-approximation of minimum k-spanners: In Section 6, we present (1+ )-approximation algorithms for spanner problems, following the framework of a recent algorithm for covering problems [39]. We show the following. Theorem 1.2. There is a randomized algorithm with complexity O(poly(log n/)) in the Local model that computes a (1 + )-approximation of the minimum k-spanner w.h.p, where k is a constant. The algorithm is quite general and can be adapted similarly to additional variants. Theorem 1.2 shows that although spanner problems are hard to approximate in the sequential setting, it is possible to achieve extremely strong approximations for them efficiently in the Local model. This demonstrates the power of the Local model. However, the algorithm is based on learning neighborhoods of polylogarithmic size and solving NP-complete problems (finding optimal spanners). It is desirable to design also algorithms that work with more realistic assumptions. We next focus on the 2-spanner problem and show a new algorithm that uses only polynomial local computations and uses the power of the Local model only for learning neighborhoods of diameter 2. 3 (II.) Distributed approximation of minimum 2-spanners: If we restrict ourselves to polynomial local computations, the best algorithm for the minimum 2-spanner problem is the O(log n)-round O(log n)-approximation in expectation of Dinitz and Krauthgamer [21],2 which solves even the more general problem of finding fault-tolerant spanners. However, this still leaves several open questions regarding minimum 2-spanners. First, the best approximation to the problem in the sequential setting is O(log m n ) where m is the number of edges in the graph. Can we achieve such approximation also in the distributed setting? Second, the approximation ratio holds only in expectation. Can we design an algorithm that guarantees the approximation ratio? Third, this algorithm requires learning neighborhoods of logarithmic radius, and hence a direct implementation of it in the Congest model is not efficient. Can we design a more efficient algorithm in the Congest model? We design a new algorithm for the minimum 2-spanner problem, answering some of these questions. Our algorithm obtains an approximation ratio of O(log m n ) always, within O(log n log ∆) rounds w.h.p,3 where ∆ is the maximum vertex degree, summarized as follows. Theorem 1.3. There is a distributed algorithm for the minimum 2-spanner problem in the Local model that guarantees an approximation ratio of O(log m n ), and takes O(log n log ∆) rounds w.h.p. Our approximation ratio of O(log m n ) matches that of the best approximation in the sequential setting up to a constant factor [46], and is tight if we restrict ourselves to polynomial local computations [45]. In addition, the approximation ratio of our algorithm is guaranteed, rather than only holding in expectation. This is crucial for the distributed setting since, as opposed to the sequential setting, running the algorithm several times and choosing the best solution completely blows up the complexity because learning the cost of the solution requires collecting global information. Note that although our algorithm can be converted into an algorithm with a guaranteed polylogarithmic time complexity and an approximation ratio that holds only in expectation, the opposite does not hold. Another feature of our algorithm is that it uses the power of the Local model only for learning the 2-neighborhood of vertices. A direct implementation of our algorithm in the Congest model yields an overhead of O(∆) rounds, which is efficient for small values of ∆. We address this issue further in Section 1.3. (III.) Distributed approximation of additional 2-spanners: The techniques we develop for constructing and analyzing our spanner have the advantage of allowing us to easily extend our construction to the directed, weighted and client-server variants of the problem. We obtain the same approximation ratio for the directed case as in the undirected case, and for the weighted case we give an approximation ratio of O(log ∆), both improving upon the O(log n) approximation in expectation of [21]. For the client-server 2-spanner case, which to the best of our knowledge ours is the first distributed approximation, we obtain an approximation ratio that matches that of the sequential algorithm [29]. (IV.) Distributed approximation of MDS: Finally, our technique also gives an efficient algorithm for the minimum dominating set (MDS) problem, which obtains an approximation ratio of O(log ∆) always. Our algorithm for MDS works even in the Congest model and takes O(log n log ∆) rounds w.h.p. The MDS problem has been studied extensively by the distributed 2 In [21], a time complexity of O(log2 n) rounds is claimed. However, the algorithm is based on sampling a certain decomposition O(log n) times independently, which takes O(log n) rounds each time. From the independence of the decompositions, the computations can be parallelized in the Local model, achieving a time complexity of O(log n) rounds. See also [22]. 3 As standard in this setting, a high probability refers to a probability that is at least 1 − n1c for a constant c ≥ 1. 4 computing community, with several efficient algorithms for MDS in the Congest obtaining an approximation ratio of O(log ∆) in expectation [43, 48, 49]. To the best of our knowledge, our algorithm is the first that guarantees this approximation ratio always. 1.2 1.2.1 Technical overview Hardness of approximation We prove Theorem 1.1 by a reduction from 2-party communication problems, as has been proven fruitful for various lower bounds for the Congest model [1, 11, 24, 35, 42, 63]. In principle, a family of graphs is constructed depending on the input strings of the two players, such that the solution to the required Congest problem uniquely determines whether the input strings of the players satisfy a certain Boolean predicate. The most common usage is of set-disjointness, although other 2-party communication problems have been used as well [10, 26, 34, 55]. The two players can simulate a distributed algorithm for solving the Congest problem, and deduce their output for the 2-party communication problem accordingly. This yields a lower bound for the Congest problem, based on known lower bounds for the communication complexity of the 2-party problem, by incorporating the cost of the simulation itself. The prime caveat in using this framework for approximation problems is that in the above examples a modification of a single input bit has a slight influence on the graph. For example, when showing a lower bound for computing the diameter, any bit of the input affects the distance between one pair of vertices [1, 35, 42]. This is sufficient when computing some global property of the graph. Indeed, the distance between a single pair of vertices can change the diameter of the graph. The challenge in designing a construction for approximating k-spanners is that now any single bit needs to affect drastically the size of the minimum k-spanner. In more detail, any k-spanner has at least n − 1 edges and, hence, for a meaningful lower bound for an α-approximation, any input bit must affect at least Ω(αn) edges. We manage to overcome the above challenge by constructing a graph that captures this requirement and allows a reduction from set-disjointness. The main technical ingredient is a dense component in which many edges are affected by single input bits. This component resides in its entirety within the set of vertices that is simulated by a single player of the two, thus resulting in a non-symmetric graph construction. This is crucial for our proof, as otherwise the density of this component would imply a dense cut between the two sets of vertices simulated by the players, which in turn would nullify the achievable lower bound. For having this property, we believe that our construction may give rise to follow-up lower bound constructions for additional local approximation problems. Our graph construction is designed using several parameters, which allows us to show trade-offs between the time complexity of an algorithm and its approximation ratio, and gives lower bounds even for large values of α. Our stronger lower bounds for the deterministic case are obtained using the 2-party gapdisjointness problem rather than the more common set-disjointness problem. Since gap-disjointness allows more slack, we obtain stronger lower bounds, at the price of them holding only for deterministic algorithms. We believe that the flexibility of the gap-disjointness problem may be useful in showing additional strong lower bounds for approximation problems. Our stronger lower bounds for the weighted case are obtained by assigning weights to the edges of the graph in a manner which allows us to shave off certain edges that affect the bound. 5 1.2.2 Distributed approximation of minimum 2-spanners Our algorithm for approximating minimum 2-spanners is inspired by the sequential greedy algorithm of Kortsarz and Peleg [46], in which dense stars are added to the spanner one by one, obtaining an approximation ratio of O(log m n ). A star is a subset of edges between a vertex v and some of its neighbors. The density of a star is the ratio between the number of edges 2-spanned by the star and the size of the star, where an edge e = {u, v} is 2-spanned by a star S if S includes a path of length two between u and v. A roughly intuition for the greedy algorithm is that if S is a dense star then adding its edges to the spanner allows 2-spanning many edges by adding only a small number of edges to the spanner. A direct implementation of this greedy approach in the distributed setting is highly expensive, since deciding upon the densest star inherently requires collecting global information. Moreover, one would like to leverage the ability of the distributed setting to add multiple stars to the spanner simultaneously. To address both sources of inefficiency, rather than computing the star that is the densest in the entire graph, we compute all the stars that are the densest in their local 2neighborhood. While greatly speeding up the running time, adding all of these locally densest stars to the spanner is too extreme, and results in a poor approximation ratio. Instead, we consider these stars as candidates for being added to the spanner. The key challenge is then to break symmetry among the candidates, while balancing the need to choose many stars in parallel (for a fast running time) with the need to bound the overlap in spanned edges among the candidates (for a small approximation ratio). We tackle this conflict by constructing a voting scheme for breaking symmetry by choosing among the stars based on a random permutation. Interestingly, our approach is inspired by a parallel algorithm for set cover [60]. We let each edge vote for the first candidate that 2-spans it according to the random permutation. A candidate that receives a number of votes which is at least 18 of the edges it 2-spans is added to the spanner, and we continue this process iteratively. Since we add to the spanner only stars receiving many votes, this approach guarantees that there is not too much overlap in the edges 2-spanned by different stars, which eventually culminates in a proof of an approximation ratio of O(log m n ), which matches the one obtained by the greedy approach. A tricky obstacle lies in showing that our algorithm completes in O(log n log ∆) rounds w.h.p. This is because, as opposed to the set cover case, there may be as many as 2∆ different stars centered at each vertex, and a vertex may be required to add candidate stars multiple times during the execution of the algorithm. It turns out that an arbitrary choice for a candidate among all densest stars centered at a vertex is incapable of providing an efficient time complexity. To overcome this issue, we design a subtle mechanism for proposing a candidate star, and pair it with a proof that our algorithm indeed completes in the claimed number of rounds. 1.3 Discussion While our results in this paper significantly advance the state-of-the-art in distributed approximation of minimum k-spanners, intriguing questions remain open. First, the landscape of the trade-offs between the approximation ratio and the running time of distributed minimum k-spanner algorithms is yet to be fully mapped. For example, the O(log ∆) factor in the running time of our approximation algorithm for weighted 2-spanner is tight up to an O(log log ∆) factor, due to our reduction from MVC and the known lower bounds for it. However, it remains open whether the O(log n) factor is necessary. Additional gaps remain open for other various approximation ratios. In particular, an interesting question is to show a lower bound for approximating the undirected unweighted 6 minimum k-spanner problem. A curious question is whether our algorithm can be efficiently made to work in the Congest model. A direct implementation would yield an overhead of O(∆) for the running time, for computing the densities of stars, and for sending the candidate stars. We emphasize that knowing the density of the neighborhood of vertices is crucial for additional algorithms, such as the state-of-the-art (∆ + 1)-coloring algorithm of Harris et al. [41]. Another interesting question is to design an efficient deterministic algorithm achieving the same approximation ratio. For larger values of the stretch k, our lower bounds imply a strict separation between the Local and Congest models for the number of rounds required for approximating directed minimum k-spanners. Such a separation was previously known only for global problems (problems that are subject to an Ω(D) lower bound, where D is the diameter of the graph), and for local decision problems (such as determining whether the graph contains a k-cycle). Interestingly, ours is the first separation for a local approximation problem. It is a central open question whether such separations hold also for local symmetry breaking problems. Interestingly, our algorithm, as well as other distributed approximation algorithms for the minimum k-spanner in the Local model, work also for directed graphs, achieving the same approximation ratio and round complexity. However, our hardness results create a strict separation between the undirected and directed variants in the Congest model. It will be interesting to show such separations for other problems. 1.4 Additional related work Spanners have been studied extensively in the distributed setting, producing many efficient algorithms for finding sparse spanners in undirected graphs [7, 17, 18, 27, 28, 40]. These algorithms construct (2k − 1)-spanners with O(n1+1/k ) edges for any fixed k ≥ 2, with the fastest completing in k rounds [17, 28], which is tight [17]. Many additional works construct various non-multiplicative spanners in the distributed setting, such as [59] and the excellent overview within. Many recent studies address spanner approximations in the sequential setting. The greedy algorithm of [46] achieves an approximation ratio of O(log m n ) for the minimum 2-spanner problem. This was extended to the weighted, directed and client-server cases [29,45]. Approximation algorithms for the directed k-spanner problem for k > 2 are given in [8,9,20,23,30], with the best approximation √ e 1/3 ) for k = 3, 4 [8, 23]. These ratio of O( n log n) for k > 4, and an approximation ratio of O(n approximation ratios are matched by a recent distributed O(k log n)-round algorithm, that uses only polynomial local computations [22]. Approximation algorithms are given also for pairwise spanners and distance preservers [14], for spanners with lowest maximum degree [13, 15, 22, 47], for fault-tolerant spanners [21, 23], and more. Hardness of approximation results in the sequential setting give that for k = 2, no polynomial algorithm gives an approximation ratio better than Θ(log n) [45], which shows that the sequential greedy algorithm is optimal. For k > 2, the problem is even harder. For any constant > 0 and k ≥ 3 there are no polynomial-time algorithms that approximate the k-spanner problem within a 1− factor better than 2(log n)/k [19], or the directed k-spanner problem within a factor better than 1− 2(log n) [31]. Similar results are known for additional variants [14, 31]. Spanner problems are closely related to covering problems such as set cover, minimum dominating set (MDS), and minimum vertex cover. Indeed, some of the ingredients of our algorithms borrow ideas from distributed and parallel algorithms for such problems. Our symmetry breaking scheme is inspired by the parallel algorithm for set cover of Rajagopalan and Vazirani [60], however, the general structure of this algorithm requires global coordination and hence is not suitable for the distributed setting. There are also several ideas inspired by the distributed MDS algorithm of Jia et 7 al. [43], such as, rounding the densities and comparing densesties in 2-neighborhoods. However, [43] breaks the symmetry between the candidates in a different way which results in an approximation ratio of O(log ∆) in expectation. The connection between spanners to set cover is used also in [9] where they show that covering the edges of a graph by stars is also useful for approximating the directed k-spanner problem for k > 2. In this context, we also mention the distributed algorithm of [37] for the minimum connected dominating set problem, which also uses stars as the main component for its construction. Our work is, however, incomparable, especially since the minimum connected dominating set problem is a global problem, admitting an Ω(D) lower bound even in the Local model. 1.5 Preliminaries Let G = (V, E) be a connected undirected graph with n vertices and maximum degree ∆. Let S ⊆ E be a subset of the edges, and let k ≥ 1. We say that an edge e = {u, v} is covered by S if there is a path of length at most k between u and v in S. A k-spanner of G is a subgraph of G that covers all the edges of G. A k-spanner of a subgraph G0 ⊆ G is a subgraph of G that covers all the edges of G0 . For a directed graph, we say that a directed edge e = (u, v) is covered by a subset of edges S, if S includes a directed path of length at most k from u to v, and define a k-spanner for a directed graph accordingly. In the minimum k-spanner problem the input is a connected undirected graph G = (V, E) and the goal is to find the minimum size k-spanner of G. The directed k-spanner problem is defined accordingly, with respect to directed graphs. In the weighted k-spanner problem each edge e has a non-negative weight w(e) and the goal P is to find the k-spanner of G having minimum cost, where the cost of a spanner H is w(H) = e∈H w(e).4 In the client-server k-spanner problem, introduced in [29], the input is a connected undirected graph G = (V, E) that its edges are divided to two types: clients C and servers S (there may be edges e ∈ C ∩ S), and the goal it to find the minimum size k-spanner of C that includes only edges of S. In the distributed setting, the input for the k-spanner problem is the communication graph G itself. Each vertex initially knows only the identities of its neighbors, and needs to output a subset of its edges such that the union of all outputs is a k-spanner. The communication in the network is bidirectional, even when solving the directed k-spanner problem. Roadmap: In Section 2, we present our hardness of approximation results for directed and weighted k-spanners in the Congest model. In Section 3, we provide hardness of approximation results for weighted 2-spanners. In Section 4, we present our algorithm for the minimum 2-spanner problem and show its extensions to other variants. In Section 5, we describe our MDS algorithm. Finally, in Section 6, we show our (1 + )-approximation for minimum k-spanners. 2 Hardness of approximation in the CONGEST model In this section, we prove hardness of approximation results for approximating k-spanners in the Congest model. As explained in Section 1.2.1, we build upon the previous used framework of reducing 2-party communication problems to distributed problems for the Congest model. The key technical challenge that we overcome is how to plant a dense subgraph into the construction, 4 There is another variant of the weighted k-spanner problem, in which the weight of an edge represents a length. We emphasize that in our case all the edges have length 1. 8 without inducing a large cut between the vertices simulated by the two players, but while still having the choice of edges taken from the dense subgraph to the spanner depend on both inputs. We describe a graph construction that allows us to provide a reduction from problems of 2-party communication. In the latter setting, two players, Alice and Bob, receive input strings a = (a1 , ..., aN ) and b = (b1 , ..., bN ), respectively, of size N . Their goal is to solve a problem related to their inputs, while communicating a minimum number of bits. For example, the set disjointness problem requires the players to decide if their input strings represent disjoint subsets of [N ], that is, they need to decide if there is a bit 1 ≤ i ≤ N such that ai = bi = 1. The communication complexity of set disjointness is known to be linear in the length of the strings [50, 61]. Lemma 2.1. Solving the set disjointness problem on input strings of size N , requires exchanging Ω(N ) bits, even using randomized protocols. We start by showing that approximating the directed k-spanner problem in the Congest model is hard for k ≥ 5, and then modify our construction to provide hardness results for the weighted case. The general approach is to build a dense graph G, where some of its edges depend on the inputs of Alice and Bob, such that if the inputs of Alice and Bob are disjoint then there is a sparse 5-spanner in G (which is also a k-spanner for k ≥ 5), and otherwise any k-spanner has many edges. By simulating the distributed approximation algorithm for the k-spanner problem, Alice and Bob solve set disjointness. Hence, depending on the parameters of our graph construction, a communication lower bound for the latter would imply a lower bound on the number of rounds required for the former. In [35,42], a reduction from set disjointness is used in order to show a lower bound for computing the diameter of a graph. The main idea is that each bit of the inputs affects the distance between two vertices in the graph, and if the distance between any of these pairs of vertices is long it affects the diameter of the graph. This idea is useful also for showing lower bound for spanner problems, and indeed one of the elements in our construction is similar to the constructions in [35,42]. However, the main difference in our case is that the distance between one pair of vertices in the graph does not affect significantly the size of the minimum spanner. In order to overcome it, we suggest the following construction. Our graph consists of two subgraphs. One of them depends on the inputs, and the other one is a complete bipartite graph D that each of its sides is divided to blocks of size β. We connect the two subgraphs in such a way that each bit i of the inputs affects β 2 edges of D, which must be added to the spanner if and only if ai = bi = 1. Let `, β be positive integers. We construct a graph G = G(`, β) according to the parameters ` and β. Later we plug-in different values of ` and β in order to obtain several trade-offs. The graph G(`, β) is a directed graph, with V = X1 ∪X2 ∪Y1 ∪Y2 ∪Y3 , where X1 = {x1i \|1 ≤ i ≤ `}∪{x2i \|1 ≤ i ≤ `}, Y1 = {yi1 \|1 ≤ i ≤ `} ∪ {yi2 \|1 ≤ i ≤ `}, X2 = {xij \|1 ≤ i ≤ `, 1 ≤ j ≤ β}, Y2 = {yij \|1 ≤ i ≤ `, 1 ≤ j ≤ β}, and Y3 = {yi3 \|1 ≤ i ≤ `}. See Figure 1 for an illustration. The set of edges consists of a matching between X1 and Y1 that includes all the directed edges 1 (xi , yi1 ) and (x2i , yi2 ), for 1 ≤ i ≤ `. In addition, there is a complete bipartite graph D between the vertices of X2 and Y2 that includes all the directed edges (xij , yrs ) for 1 ≤ i, r ≤ `, 1 ≤ j, s ≤ β. For each vertex xij ∈ X2 there is an edge (xij , x1i ). For each vertex yij ∈ Y2 there is an edge (yi3 , yij ). In addition, the graph includes the edges (yi2 , yi3 ), for 1 ≤ i ≤ `. In addition, the two input strings a, b of length `2 bits, denoted by aij , bij for 1 ≤ i, j ≤ `, affect G in the following way. The edge (x1i , x2j ) is in G if and only if aij = 0, and the edge (yi1 , yj2 ) is in G if and only if bij = 0. 9 𝑋2 𝑌2 𝑥11 𝑦11 𝑥12 𝑦12 𝑋1 𝑌1 𝑥11 𝑦11 𝑥21 𝑦21 𝑥ℓ1 𝑦ℓ1 𝑥12 𝑦12 𝑥22 𝑦22 𝑥ℓ2 𝑦ℓ2 𝑌3 𝑥1𝛽 𝑦13 𝑦1𝛽 𝑦23 𝑥ℓ1 𝑦ℓ1 𝑥ℓ2 𝑦ℓ2 𝑥ℓ𝛽 𝑦ℓ𝛽 𝑦ℓ3 Figure 1: The graph G, with some of its edges omitted for clarity. The red dashed edges are examples of optional edges which depend on the input strings a and b. Note that the number of vertices in G is n = Θ(`β), and that D consists of (`β)2 = Θ(n2 ) edges, and recall the goal of constructing a sparse k-spanner for G with k ≥ 5. Since D is a dense subgraph, taking its edges to the spanner would be expensive. However, in order to avoid taking the edges of D to the spanner, the spanner must include a directed path of length at most k between every pair of vertices xij , yrs , which does not include edges of D. The existence of such a path depends on the input strings in the following way. Claim 2.2. If one of the edges (x1i , x2r ), (yi1 , yr2 ) is in G, there is a directed path of length 5 between the vertices xij , yrs that does not contain edges of D. Otherwise, the only directed path from xij to yrs is the path that consists of the edge (xij , yrs ). Proof. Note that any directed path from xij to yrs that does not include the edges of D must begin with the edge (xij , x1i ) and must end with the two edges (yr2 , yr3 ), (yr3 , yrs ). Hence, the existence of such a path depends on whether there is a directed path from x1i to yr2 . We show that there is a directed path of length 2 from x1i to yr2 if at least one of the edges (x1i , x2r ), (yi1 , yr2 ) is in G. Otherwise, there is no directed path of any length from x1i to yr2 . Let P be a directed path from x1i to yr2 . The path P must cross the cut between X1 to Y1 either by the edge (x1i , yi1 ) or by the edge (x2r , yr2 ), since any path (of any length) from x1i can only cross the cut through the edge (x1i , yi1 ) or by an edge of the form (x2j , yj2 ). However, if j 6= r, yr2 is not reachable from yj2 . If P crosses by the edge (x1i , yi1 ) the only way to reach yr2 from yi1 is by the edge (yi1 , yr2 ). In the second case, the edge (x1i , x2r ) must be the first edge in the path. In conclusion, if one of the edges (x1i , x2r ), (yi1 , yr2 ) is in G, then there is a directed path of length 5 between the vertices xij , yrs that does not contain edges of D. Otherwise, there is no directed path of any length from x1i to yr2 . In this case, the only directed path from xij to yrs is the path that consists of the edge (xij , yrs ). 10 Claim 2.2 captures the essence of why our construction is suitable for an approximation problem. Next, we use our graph construction and this claim in order to show our hardness results. 2.1 Randomized directed k-spanner In this section, we address the √directed k-spanner problem for k ≥ 5, and show that obtaining an n α-approximation requires Ω( √α log ) rounds in the Congest model, even when using randomized n algorithms. Lemma 2.3. Let G = G(`, β) for β ≥ `, let k ≥ 5, and let c = 7. If the input strings a, b are disjoint, then there is a k-spanner of size at most c`β for G. Otherwise, any k-spanner for G includes at least β 2 edges of D. Proof. If the input strings a, b are disjoint, then for every pair of indexes i, r at least one of the edges (x1i , x2r ), (yi1 , yr2 ) is in G. Hence, by Claim 2.2, there is a directed path of length at most 5 between every two vertices xij , yrs , which does not contain edges of D. This gives a 5-spanner of size at most c`β edges for G by taking all the edges not in D, since there are at most 2`β + 2`2 + 3` such edges, which is at most c`β since ` ≤ β and c = 7. This is also a k-spanner for any k ≥ 5. If the input strings are not disjoint, then there is a pair of indexes i, r such that neither of the edges (x1i , x2r ), (yi1 , yr2 ) is in G. Hence, by Claim 2.2, there is no directed path between the vertices xij , yrs except for the path that includes the edge (xij , yrs ). Therefore, we need to take all the edges (xij , yrs ) to the spanner for all values of j and s, which means adding β 2 edges of D to the spanner. Let k ≥ 5 and let A be a distributed α-approximation algorithm for the minimum k-spanner problem. Denote by T (n) the time complexity of A on a graph with n vertices. The approximation ratio α = α(n) of the algorithm A may depend on n, and we assume that it is a monotonic increasing function of n, and that if n = Θ(n0 ), then α(n) = Θ(α(n0 )). Our goal is to show that A can be used to solve set disjointness. If α · c`β < β 2 , then by Lemma 2.3, the algorithm A gives a protocol for set disjointness, in which case we show a lower bound of Ω( log` n ) on the time complexity of A, as stated in the following lemma. Lemma 2.4. Let G = G(`, β). If there is a threshold t such that if the input strings a, b are disjoint, an optimal spanner of G has at most t edges, and otherwise each spanner of G includes more than α(n) · t edges of D, then T (n) = Ω( log` n ). Proof. We use A to solve set disjointness on input strings of length N = `2 in the following way. Let a, b be two input strings of length N , given to Alice and Bob respectively. We take the graph G = G(`, β) and define VB = Y1 , VA = V \ VA . Since the input strings a and b affect only edges between vertices within VA and within VB respectively, it holds that Alice knows all the edges adjacent to vertices in VA and Bob knows all the edges adjacent to vertices in VB . The cut between VA to VB consists of Θ(`) edges: the 2` edges of the matching between X1 to Y1 , and the ` edges between Y1 to Y3 . Now Alice and Bob simulate A on G as follows. Alice simulates the vertices in VA and Bob simulates the vertices in VB . At each round, Alice and Bob exchange the messages going over the cut between VA and VB in either direction. Messages that are sent between vertices in VA or between vertices in VB are simulated locally by Alice and Bob, without any communication. Since the size of messages is O(log n) bits, and the size of the cut is Θ(`), they can simulate one round of A by exchanging at most O(` · log n) bits, and therefore they can simulate the entire execution of A by exchanging at most O(T (n) · ` · log n) bits. 11 At the end of the simulation, Alice knows which of the edges of D are taken to the spanner. If there are more than α(n) · t edges of D in the spanner, Alice concludes that the input strings are not disjoint, and otherwise she concludes that they are disjoint. To show that this produces the correct output, recall the condition of the lemma that if the input strings are disjoint then the size of an optimal spanner is at most t and otherwise it is more than α(n) · t. Therefore, if the input strings are disjoint, since A is an α(n)-approximation algorithm, it constructs a spanner with at most α(n) · t edges, in which case Alice indeed outputs that the input strings are disjoint. Otherwise, if the input strings are not disjoint, the size of any spanner is more than α(n) · t edges, in which case Alice indeed outputs that the input strings are not disjoint. Hence, Alice and Bob solve set disjointness by exchanging O(T (n) · ` · log n) bits. However, any (perhaps randomized) protocol that solves disjointness on inputs of size N = `2 requires exchanging ` `2 Ω(`2 ) bits by Lemma 2.1. This gives T (n) = Ω( `·log n ) = Ω( log n ). Using Lemma 2.3 and Lemma 2.4, we prove our following main theorem. Theorem 1.1. Any (perhaps randomized) distributed α-approximation algorithm in the Congest √ n n . model for the directed k-spanner problem for k ≥ 5 takes Ω( √α·log n ) rounds, for 1 ≤ α ≤ 100 Proof. We show that there is a threshold t that distinguishes whether the inputs are disjoint. Then, using Lemma 2.4, we get a lower bound on the round complexity of A. 0 We define G = G(`, β) with the following choice of the q parameters β, `. Let n be a positive integer, and let c = 7. Let q = dα(n0 )ce + 1. Let ` = b n 100 α(n) ≤ ensures that cq ≤ n = Θ(`β) = Θ(q`2 ) = Θ(q · n0 , n0 q ) n0 cq c, and let β = q`. The requirement which shows that ` is positive. The number of vertices in G is = Θ(n0 ). In addition, note that n ≤ c`β, since the number of 0 vertices in G is 2`β + 5`, which gives n ≤ c`β = cq`2 ≤ cq · ncq = n0 . Let t = c`β. By Lemma 2.3, If the inputs are disjoint, there is a k-spanner for G having at most t = c`β edges, and otherwise any k-spanner for G includes at least β 2 edges of D. By the definition of q, it holds that α(n0 ) · c < q, which gives α(n0 ) · c`β < q`β = β 2 . Since n ≤ n0 , it holds that α(n) ≤ α(n0 ), which gives α(n) · t = α(n) · c`β < β 2 . q Hence, t satisfies the conditions of Lemma 2.4, which gives T (n) = Ω( log` n ). Since ` = Θ( nq ) = √ q n ). Θ( α(n) ), it holds that T (n) = Ω( log` n ) = Ω( √ n α(n)·log n Theorem 1.1 shows that achieving a constant or a polylogarithmic approximation ratio for the e √n) rounds, and even achieving an directed k-spanner problem in the Congest model requires Ω( e 1/2−/2 ) rounds, for any 0 < < 1. approximation ratio of n is hard, requiring Ω(n This proves a strict separation between the Local and Congest models, since there is a constant round O(n )-approximation algorithm [5], and a polylogarithmic (1 + )-approximation algorithm (see Section 6) for directed k-spanner in the Local model. It also separates the undirected and directed k-spanner problems, since there are randomized k-round algorithms in the Congest model for constructing (2k − 1)-spanners with O(n1+1/k ) edges [28]. These algorithms obtain an approximation ratio of O(n1/k ) for the undirected minimum (2k − 1)-spanner problem in k rounds, where achieving the same approximation for the directed e 1/2−1/2k ) rounds according to Theorem 1.1. problem requires Ω(n 2.2 Deterministic directed k-spanner We next show that any deterministic algorithm solving the directed k-spanner problem for k ≥ 5, n requires Ω( √α·log ) rounds. The trick that allows a stronger lower bound is that we use a different n 12 problem from communication complexity, which we refer to as the gap disjointness problem. This problem is also mentioned in [34]. In the gap disjointness problem, Alice and Bob receive the input strings a = (a1 , ..., aN ) and b = (b1 , ..., bN ), respectively, and their goal is to distinguish whether their input strings are disjoint N or are far from being disjoint. The inputs are far from being disjoint if there are at least 12 indexes i, such that ai = bi = 1. If the inputs are neither disjoint nor far from being disjoint, any output of Alice and Bob is valid. The gap disjointness problem can be easily solved by randomized protocols exchanging O(1) bits. However, solving the problem deterministically requires exchanging Ω(N ) bits. Lemma 2.5. Solving the gap disjointness problem deterministically on input strings of size N requires exchanging Ω(N ) bits. For a proof of Lemma 2.5, see example 5.5 in [50], where it is shown that approximating the size of the intersection \|a ∩ b\| requires exchanging Ω(N ) bits. The proof relies only on showing that distinguishing between disjoint inputs and inputs with intersection of more than N6 bits is difficult N (note that any such inputs have intersection of size at least 12 ). Hence, the exact same proof shows that solving gap disjointness requires exchanging Ω(N ) bits using a deterministic protocol. In order to use set disjointness for the proof of Theorem 1.1, it was necessary to devise a construction where each bit of the input affects many edges of the spanner, in order to argue that even if there is only one index i such that ai = bi = 1, then the players can correctly decide whether the inputs are disjoint by checking the size of the spanner. However, when we use gap disjointness, the players need to distinguish only between the case that the inputs are disjoint and the case that they are far from being disjoint, which allows much more flexibility and gives stronger lower bounds for the deterministic case. Lemma 2.6. Let G = G(`, β) for 1 ≤ β ≤ `, let k ≥ 5 and let c = 7. If the input strings a, b are disjoint, then there is a k-spanner of size at most c`2 . If the input strings are far from being disjoint, 2 any k-spanner for G includes at least β12 `2 edges of D. Proof. If the input strings are disjoint, taking all the edges not in D is a 5-spanner, as shown in the proof of Lemma 2.3. These are at most 2`β + 2`2 + 3` edges not in D, which is at most c`2 since β ≤ ` and c = 7. This is also a k-spanner for any k ≥ 5. `2 If the input strings are far from being disjoint then there are at least 12 pairs (i, r) such that `2 1 2 1 2 none of the edges (xi , xr ), (yi , yr ) are in G. Hence, by Claim 2.2, there are at least 12 pairs (i, r) such that there is no directed path between the vertices xij , yrs except for the path that consists of the edge (xij , yrs ). For each such pair, we need to take all the directed edges (xij , yrs ) to the spanner for all the values of j and s, which means adding β 2 edges to the spanner. Summing over 2 `2 all the 12 pairs, we get that any k-spanner must include at least β12 `2 edges of D. Let k ≥ 5 and let A be a deterministic distributed α-approximation algorithm for the minimum k-spanner problem. Denote by T (n) the round complexity of A on a graph with n vertices. The following lemma adapts Lemma 2.4 to the gap disjointness problem. Its proof is the same as the proof of Lemma 2.4, with the difference that now Alice concludes that the input strings are far from being disjoint if and only if the constructed spanner has more than α(n) · t edges of D. Also, now the lower bound holds only for the deterministic case, since it relies on the communication complexity of gap disjointness. Lemma 2.7. Let G = G(`, β). If there is a threshold t such that if the input strings a, b are disjoint then an optimal k-spanner of G has at most t edges, and if the input strings are far from being disjoint then each k-spanner of G includes more than α(n) · t edges of D. Then, T (n) = Ω( log` n ). 13 Using Lemma 2.6 and Lemma 2.7, we show the following. Theorem 2.8. Any deterministic distributed α-approximation algorithm in the Congest model n for the directed k-spanner problem for k ≥ 5 takes Ω( √α·log ) rounds, for 1 ≤ α ≤ cn0 for a constant n c0 > 1. Proof. We construct the graph G = G(`, β) withp the following choice for the parameters `, β. Let n0 n0 be a positive integer, and let c = 7. Let β = d 12α(n0 )ce + 1, and let ` = b cβ c. 0 The number of vertices in G is n = Θ(`β) = Θ( nβ β) = Θ(n0 ). In addition, it holds that n ≤ c`β 0 n since the number of vertices in G is 2`β + 5`, which gives n ≤ c`β ≤ cβ cβ = n0 . In order to use Lemma 2.6 we need to verify that β ≤ `. Note that p n = c1 `βpfor a constant 2 ≤ c1 ≤ c. It follows that β ≤ ` if and only if c1 β 2 ≤ n. Since β =q Θ( α(n)) = c2 α(n) for a constant c2 , if we choose c0 = c1 c22 , we get that if α(n) ≤ n c0 , then β ≤ n c1 , which gives c1 β 2 ≤ n as needed. We now define t = c`2 . By Lemma 2.6, if the input strings a, b are disjoint, then there is a k-spanner of size at most t = c`2 . Otherwise, if the input strings are far from being disjoint, then any 2 k-spanner for G includes at least β12 `2 edges of D. By the choice of β and since n ≤ n0 , it holds that 2 12α(n) · c < β 2 , which gives α(n) · t = α(n) · c`2 < β12 `2 , which shows that t satisfies the conditions of Lemma 2.7. Using Lemma 2.7 we get that T (n) = Ω( log` n ). Note that now ` = Θ( βn ) = Θ( √ n ), α(n) which shows that T (n) = Ω( √ n ). α(n)·log n Theorem 2.8 shows that achieving a constant or a polylogarithmic approximation ratio for the e directed k-spanner problem in the Congest model requires Ω(n) rounds for any deterministic e 1−/2 ) rounds, for algorithm. In addition, even an approximation ratio of n is hard, requiring Ω(n n any 0 < < 1. Notably, even an approximation ratio of c for appropriate values of c is hard, e √n) rounds. This is to be contrasted with the fact that obtaining an approximation requiring Ω( ratio of n requires no communication, since any k-spanner has at least n − 1 edges. Theorem 2.8 separates the Local and the Congest models, since the deterministic network decomposition described in [5] gives a deterministic O(n )-approximation for directed k-spanner for a constant k in polylogarithmic time in the Local model. It also separates the undirected and directed k-spanner problems for deterministic algorithms. Currently the best deterministic algorithm in the Congest model for the undirected k-spanner problem, is a recent algorithm [40] which constructs (2k−1)-spanners of size O(n1+1/k ) in O(n1/2−1/k ) rounds for a constant even k (in the Local nodel there is a k-round deterministic algorithm for this problem [17]). This gives an O(n1/k )-approximation for undirected (2k − 1)-spanners. Achieving the e 1−1/2k ) rounds according to Theorem 2.8. same approximation for the directed problem requires Ω(n 2.3 Weighted k-spanner We extend our construction to the weighted case, showing that any approximation for the weighted e k-spanner in the Congest model takes Ω(n) rounds for k ≥ 4, even for randomized algorithms. A similar result holds for the weighted undirected case. In the weighted case, rather than guaranteeing that each input bit affects many edges of the spanner, we simply assign weight 0 to all the edges that are not in D and weight 1 to all the edges of D. Hence, taking even a single edge from D is very expensive if we can avoid it. This allows us to show a simpler construction, obtaining a stronger lower bound for the weighted case, as follows. We build a graph Gw (`) = Gw = (Vw , Ew ) which is the same as G, except for the following differences (see Figure 2). We define β = 1, and change the set of vertices to be Vw = V \ Y3 . 14 Since β = 1, the vertices in X2 and Y2 are only of the form xi1 , yi1 for 1 ≤ i ≤ `. We change their names from xi1 , yi1 to xi and yi , respectively. For each 1 ≤ i ≤ ` we replace the two edges (yi2 , yi3 ), (yi3 , yi1 ) ∈ E by the edge (yi2 , yi ) ∈ Ew . Since β = 1, the size of the cut between Y1 and the rest of the graph is still Θ(`). 𝑋2 𝑌2 𝑥1 𝑦1 𝑥2 𝑦2 𝑥ℓ 𝑦ℓ 𝑋1 𝑌1 𝑥11 𝑦11 𝑥21 𝑦21 𝑥ℓ1 𝑦ℓ1 𝑥12 𝑦12 𝑥22 𝑦22 𝑥ℓ2 𝑦ℓ2 Figure 2: The graph Gw , with some of its edges omitted for clarity. The red dashed edges are examples of optional edges which depend on the input strings a and b. The following theorem states our lower bound for the weighted directed case. Theorem 2.9. Any (perhaps randomized) distributed α-approximation algorithm in the Congest model for the weighted directed k-spanner problem for k ≥ 4 takes Ω( logn n ) rounds. Proof. Let n = 6` for a positive integer `, and let Gw = Gw (`). Note that the number of vertices in Gw is exactly n. There is a 4-spanner of cost 0 for Gw if and only if there is a path of length at most 4 of edges of weight 0 between every pair of vertices xi , yj . A path of length at most 4 between xi and yj that includes only edges of weight 0, must start with the edge (xi , x1i ) and must end with the edge (yj2 , yj ). Following the proof of Claim 2.2, we argue that there is such a path if and only if one of the edges (x1i , x2j ), (yi1 , yj2 ) is in Gw . Otherwise, there is no directed path of weight 0 between xi and yj . It follows that for every k ≥ 4, there is a k-spanner of cost 0 for Gw if and only if the inputs a and b are disjoint. Hence, a distributed α-approximation algorithm A for the weighted k-spanner problem can be used to solve set disjointness: we define VB = Y1 , VA = V \ VB and let Alice and Bob simulate the algorithm on Gw as before. At the end of the simulation, Alice concludes that the inputs are disjoint if and only if none of the edges of D are taken to the spanner. If the inputs are disjoint, then there is a spanner of cost 0. Hence, for any α ≥ 1, an αapproximation must return a spanner of cost 0 if such exists. Otherwise, any spanner must include at least one of the edges of D which proves that the output of Alice is indeed correct. As in the proof of Lemma 2.4, we get that T (n) = Ω( log` n ). Since n = 6`, this gives T (n) = Ω( logn n ). We prove a similar bound for the weighted undirected k-spanner problem for k ≥ 4. In the undirected case, we would like to construct a similar graph Gw , with only modifying all of its edges to be undirected. It would still hold that there is a path of length at most 4 of edges of weight 0 15 between the vertices xi , yj if and only if one of the edges {x1i , x2j }, {yi1 , yj2 } is in Gw , following the same proof. However, since the edges are undirected, there may be a path of length longer than 4 of edges of weight 0 between the vertices xi , yj , even if none of the edges {x1i , x2j }, {yi1 , yj2 } is in Gw , which requires us to modify our construction in order for our bounds to apply also for k > 4. We change the construction of Gw as follows. For each 1 ≤ i ≤ ` we replace the edge {yi2 , yi } by a path of length k − 3, by adding to the graph k − 4 vertices yi3 , ..., yik−2 and the required edges for constructing the path {yi2 , yi3 , ..., yik−2 , yi }. All of the edges of the path have weight 0. Any path of length at most k of edges of weight 0 between xi to yj must start with the edge {xi , x1i } and must end with the path of length k − 3 that we added between yj2 to yj . Hence, there is a path of length at most k between xi to yj of edges of weight 0 if and only if there is a path of length 2 between x1i and yj2 . This can only happen if one of the edges {x1i , x2j }, {yi1 , yj2 } is in the graph. The rest of the proof is exactly the same as in the directed case. However, we added Θ(k`) vertices to the graph. Hence, the number of vertices in the graph is n = Θ(k`) and not Θ(`) as e n ) to the undirected before, which gives ` = Θ( nk ). This allows us to prove a lower bound of Ω( k e problem (which is still Ω(n) for small values of k). Theorem 2.10. Any (perhaps randomized) distributed α-approximation algorithm in the Congest n model for the weighted undirected k-spanner problem for k ≥ 4 takes Ω( k·log n ) rounds. 3 Hardness of approximation of weighted 2-spanner In this section, we show that approximating the weighted 2-spanner problem is at least as hard as approximating the (unweighted) minimum vertex cover (MVC) problem. Therefore, known lower bounds for MVC translate directly to lower bounds for the weighted 2-spanner problem. In the MVC problem the input is a graph G = (V, E) and the goal is to find a minimum set of vertices C that covers all the edges. That is, it is required that for each edge e = {v, u}, at least one of v and u is in C. Let G = (V, E) be an input graph to the MVC problem. We construct a new graph GS = (VS , ES ) in the following way (see Figure 3). For each vertex v ∈ V , there are 3 vertices in VS : v1 , v2 , v3 . We connect these 3 vertices by a triangle, where the edge {v1 , v2 } has weight 1, and the edges {v1 , v3 }, {v2 , v3 } have weight 0. In addition, for each edge {v, u} ∈ E, there are 3 edges in ES : {v1 , u1 }, {v2 , u2 }, both having weight 0, and one of the edges {v1 , u2 }, {u1 , v2 }, according to the order of the IDs of v and u, having weight 2. 0 𝑣3 0 𝑣1 1 0 𝑣2 𝑢1 2 𝑢3 1 𝑢2 0 0 0 Figure 3: For each vertex v ∈ G, there is a corresponding triangle in GS between the vertices v1 , v2 , v3 . The edge {v, u} ∈ G has 3 corresponding edges in GS : {v1 , u1 }, {v2 , u2 }, {v1 , u2 }. We show that a solution for the weighted 2-spanner problem in GS gives a solution for MVC in G. 16 Claim 3.1. The cost of the minimum 2-spanner in GS is exactly the size of the minimum vertex cover in G. Proof. Let C be a minimum vertex cover in G. We construct a 2-spanner HC for GS as follows. First, HC includes all the edges having weight 0. In addition, for every v ∈ C, we add to HC the edge {v1 , v2 }. Note that these edges have weight 1, and all the other edges we add to HC have weight 0, hence, the cost of HC is exactly \|C\|. We now show that HC is a 2-spanner. All the edges having weight 0 in GS are added to the spanner, and hence they are covered. All the edges having weight 1 are covered by edges of weight 0, since an edge {v1 , v2 } is covered by the path {v1 , v3 }, {v3 , v2 }. Let {v1 , u2 } be an edge of weight 2 in GS , and let {v, u} be the corresponding edge in G. Since C is a vertex cover, at least one of the vertices v, u is in C. In the former case, we add {v1 , v2 } to HC , hence, the edge {v1 , u2 } is covered by the path {v1 , v2 }, {v2 , u2 } (note that {v2 , u2 } has weight 0 and is included in HC ). In the latter case, {v1 , u2 } is covered by the path {v1 , u1 }, {u1 , u2 }. Hence, HC is a 2-spanner having cost \|C\|. In the other direction, let H be a minimum cost 2-spanner in GS having cost w(H). We construct a vertex cover CH in G with size w(H). We start by converting H into a 2-spanner H 0 with the same cost. First, H 0 contains all the edges having weight 0 in GS and all the edges having weight 1 in H. In addition, if H includes an edge {v1 , u2 } having weight 2, we replace it in H 0 by the two edges {v1 , v2 }, {u1 , u2 }, each having weight 1. This transformation clearly cannot increase the cost. We next show that H 0 is still a 2-spanner. Since H 0 includes all the edges having weight 0 in GS , it covers all the edges of weight 0 or 1 by edges of weight 0, as explained above. In addition, any edge of weight 2 that is covered in H by a path of length 2 that includes only edges of weight 0 or 1, is covered in H 0 in the same way. Let e = {x1 , y2 } be an edge of weight 2, covered in H by a path P of length at most 2 that includes the edge e0 = {v1 , u2 } ∈ H having weight 2 (e0 may be different than e). It holds that e0 6∈ H 0 , since e0 has weight 2, hence, we added the edges {v1 , v2 } and {u1 , u2 } to H 0 . Since P has length at most 2, it follows that x1 = v1 or y2 = u2 . In the first case, e = {v1 , y2 } and the path {v1 , v2 }, {v2 , y2 } covers e (we added {v1 , v2 } to H 0 , and since {v1 , y2 } ∈ ES , then {v2 , y2 } is also in ES and has weight 0). In the second case, e = {x1 , u2 } and {x1 , u1 }, {u1 , u2 } is a path of length 2 that covers e in H 0 . Therefore, H 0 is a 2-spanner with the same cost of H. We define CH = {v\|{v1 , v2 } ∈ H 0 }. The size of CH is exactly w(H 0 ) since the edges {v1 , v2 } ∈ H 0 are exactly all the edges of H 0 having weight 1, and H 0 includes only edges of weight 0 or 1. In addition, we claim that CH is a vertex cover. Let {v, u} ∈ E, then one of the edges {v1 , u2 } or {u1 , v2 } is in ES . Assume w.l.o.g that e = {v1 , u2 } ∈ ES . Note that e 6∈ H 0 since it has weight 2. Since H 0 is a 2-spanner it includes a path of the form {v1 , x}, {x, u2 } that covers e. It must hold that x = v2 or x = u1 (if, for example, x = w1 such that w1 = 6 u1 , then the edge {w1 , u2 } has weight 2 and is not in H 0 ). Hence, at least one of the edges {v1 , v2 }, {u1 , u2 } is in H 0 , which means that at least one of v, u is in CH as needed. In conclusion, the cost of a minimum 2-spanner in GS is exactly the size of the minimum vertex cover in G. We can now relate the above to the number of rounds required for distributed algorithms that solve or approximate these two problems. Lemma 3.2. Let A be a distributed α-approximation algorithm for the weighted 2-spanner problem that takes T (n) rounds on a graph with n vertices. Then there is an α-approximation algorithm for MVC that takes 3T (3n) rounds on a graph with n vertices. Proof. We describe an algorithm AM V C that approximates MVC. Let G be an input graph for MVC. The algorithm AM V C simulates A on the graph GS , in the following way. Each vertex 17 v ∈ V simulates AM V C on the vertices v1 , v2 , v3 . Each time a message is sent on one of the 3 edges corresponding to an edge {v, u} ∈ E, we send this message over the edge {v, u} ∈ E. Since we may need to send 3 different messages on this edge, each round of A can be simulated in three rounds of AM V C .5 When A finishes, we convert the solution H to a vertex cover CH as described in the proof of Claim 3.1, without any communication. From Claim 3.1, it follows that if H is an α-approximation for the weighted 2-spanner problem in GS , then CH is an α-approximation for MVC in G. Let n be the number of vertices in G. The number of vertices in GS is 3n from the definition of GS , hence the time complexity of simulating A on GS is 3T (3n). Lemma 3.2 shows that if A works in the Congest model, then AM V C works in the Congest model as well. Hence, lower bounds for approximating MVC in both the Local and the Congest models give lower bounds for the weighted 2-spanner problem. This gives the following results. Theorem 3.3. To obtain a constant or a polylogarithmic approximation ratio for the weighted 2-spanner problem, even in the Local model, q there are graphs on which every distributed algorithm log ∆ requires at least Ω( log log ∆ ) rounds and Ω( logloglogn n ) rounds. Theorem 3.3 follows from Theorem 14 in [48] and from Lemma 3.2. Note that the number of vertices and the maximum degree in GS are equal up to a constant factor to the number of vertices and maximum degree in G. In addition, Theorem 13 in [48], allows us to show trade-offs between the time complexity of a distributed algorithm for weighted 2-spanner to the approximation ratio it gets. Theorem 3.4. For every integer k > 0, there are graphs G, such that in k communication rounds in the Local model, every distributed algorithm for the weighted 2-spanner problem on G has ! 1 1−o(1) k+1 n 4k2 approximation ratios of at least Ω and Ω ∆ k . k In the Congest model, solving MVC optimally takes Ω which carries over to exact spanners, as follows. n2 log2 n rounds (see Theorem 2 in [11]), Theorem 3.5. Any distributed in the Congest model that solves the weighted 2-spanner algorithm n2 problem optimally requires Ω log2 n rounds. All of these lower bounds hold also for randomized algorithms. Remarks: Our reduction from MVC can be adapted to obtain additional bounds. First, by changing the weights of all edges having weight 2 to have weight 1, we obtain that an α-approximation for the weighted 2-spanner problem gives a 2α-approximation for MVC. This implies the lower bounds of Theorem 3.3 and Theorem 3.4 also for graphs with 0, 1 weights. This can be viewed as lower bounds for the 2-spanner augmentation problem, in which we are given an initial set of edges and need to augment it with the minimal number of edges that induces a 2-spanner. Further, the same lower bounds hold for the directed weighted case. We modify the construction such that the edges of the triangle for vertex v are (v1 , v2 ), (v1 , v3 ), (v3 , v2 ). For an edge {v, u} ∈ E, ES includes 5 directed edges: (v1 , u1 ), (u1 , v1 ), (v2 , u2 ), (u2 , v2 ) and one of the edges (v1 , u2 ), (u1 , v2 ). The weights of all the edges remain as in the undirected case. 5 In the Local model we can send these 3 messages in one round. However, we spend three different rounds in order for the simulation to work also in the Congest model. 18 4 Distributed approximation for 2-spanner problems Here we present our distributed approximation algorithm for the minimum 2-spanner problem. We need the following terminology and notation. A v-star is a non-empty subset of edges between v and a subset of its neighbors. The density S\| of a star S with respect to a subset of edges H, denoted by ρ(S, H), equals \|C \|S\| , where CS is the set of edges of H 2-spanned by the star S, where an edge e = {u, w} is 2-spanned by the v-star S if S includes the edges {v, u}, {v, w}. Note that S covers all the edges 2-spanned by S and also the edges of S, but it 2-spans only non-star edges. The densest v-star with respect to H is the v-star having maximal density with respect to H. The density of a vertex v with respect to H, denoted by ρ(v, H), is the density of the densest v-star. If H is clear from the context, we refer to ρ(S, H) and ρ(v, H) as the density of S and the density of v, and denote them by ρ(S) and ρ(v), respectively. The rounded density of a star S with respect to H, denoted by ρ̃(S, H), is obtained by rounding ρ(S, H) to the closest power of 2 that is greater than ρ(S, H). Similarly, the rounded density of a vertex v with respect to H, denoted by ρ̃(v, H), is obtained by rounding ρ(v, H) to the closest power of 2 that is greater than ρ(v, H). The full v-star is the star that includes all the edges between v and its neighbors. The 2-neighborhood of a vertex v consists of all the vertices at distance at most 2 from v. In our algorithm, each vertex v maintains a set Hv that includes all the edges 2-spanned by the full v-star that are still not covered by the edges added to the spanner. The algorithm proceeds in iterations, where in each iteration the following is computed: 1. Each vertex v computes its rounded density ρv = ρ̃(v, Hv ), and sends it to its 2-neighborhood. 2. Each vertex v such that ρv ≥ ρu for each u in its 2-neighborhood and ρ(v, Hv ) ≥ 1 is a candidate. Let Sv be a v-star with density at least ρ4v , chosen according to Section 4.1 (a choice which is central for our analysis to carry through). Vertex v informs its neighbors about Sv . Let Cv be the edges of Hv 2-spanned by Sv . 3. Each candidate v chooses a random number rv ∈ {1, ..., n4 } and sends it to its neighbors.6 4. Each uncovered edge that is 2-spanned by at least one of the candidates, votes for the first candidate that 2-spans it according to the order of the values rv . If there is more than one candidate with the same minimum value, it votes for the one with minimum ID. 5. Each star Sv for which v receives at least spanner. \|Cv \| 8 votes from edges it 2-spans is added to the 6. Each vertex v updates the set Hv in its 2-neighborhood by removing from it edges that are now covered. 7. If the maximal density in the 2-neighborhood of v is at most 1, v adds to the spanner all the edges adjacent to it that are still not covered, and outputs the edges adjacent to it that were added to the spanner during the algorithm. At the end of the algorithm all the edges are covered by spanner edges, since we add to the spanner edges that are not 2-spanned during the algorithm. 6 Knowing the exact value of n is unnecessary, and the typical assumption that the vertices know a polynomial upper bound on n suffices. 19 Since all the candidates have maximal rounded density in their 2-neighborhood, it follows that all the candidates that cover the same edge have the same rounded density, which is crucial in the analysis. In addition, rounding the densities guarantees that there are only O(log ∆) possible values for the maximal rounded density, which allows us to show an efficient time complexity. Each iteration takes constant number of rounds in the Local model. For example, to calculate ρ̃(v, Hv ), each vertex v learns all the edges between its neighbors that are still uncovered, by having each vertex u send to its neighbors a list of its neighbors w such that the edges {u, w} are still not covered. We next show that the algorithm requires only polynomial local computations. We can compute the densest v-star in polynomial time as in the sequential algorithm (see Lemma 2.1 in [46]). This is the maximal density problem, that can be solved in polynomial time using flow techniques [36]. This allows us to compute the rounded density of a vertex. We next explain how we choose the star Sv in polynomial time. Other computations in the algorithm are clearly polynomial. 4.1 Choosing the star Sv In step 2 of each iteration, a candidate vertex v chooses a v-star with density at least ρ4v . However, there may be multiple v-stars with such density, and choosing an arbitrary star between them does not meet our claimed round complexity, and it is crucial to choose the stars in a certain way. In addition, we have to find such star using only polynomial local computations. We next describe how to choose the star Sv . Let Hvi be the subset Hv at the beginning of iteration i. It holds that Hvi+1 ⊆ Hvi for all i. Let ρ = ρ̃(v, Hvi ). The star Svi that v chooses in iteration i is defined as follows. If i is the first iteration in which v is a candidate with rounded density ρ, then Svi is chosen as follows. First, v computes the densest v-star, denote it by S. Now, if there is an edge e such that ρ(S ∪ {e}, Hvi ) ≥ ρ4 , then v adds such an edge to S. Otherwise, if there is a disjoint v-star S 0 such that ρ(S 0 , Hvi ) ≥ ρ4 , then v adds the edges of S 0 to S. Now v continues in the same manner until there is no edge or disjoint star it can add to S without decreasing the density below ρ4 . The resulting star is Svi . If v is already a candidate with rounded density ρ in iteration i − 1, then if ρ(Svi−1 , Hvi ) ≥ ρ4 , we define Svi = Svi−1 . Otherwise, if Svi−1 contains a star with density at least ρ4 with respect to Hvi , we define Svi as follows. v starts by computing the densest v-star S that is contained in Svi , and then adds to it edges or disjoint stars as before, however, it only considers adding edges or disjoint stars from Svi−1 . This guarantees that Svi ⊆ Svi−1 . If Svi−1 does not contain a star with density at least ρ4 with respect to Hvi , v chooses an arbitrary v-star with rounded density ρ. (We later show that this never happens). The computations are polynomial. v adds edges to S at most n times. Each time it adds edges it does the following computation: it checks if there is an edge such that ρ(S ∪ {e}, Hvi ) ≥ ρ4 , and since there are at most n optional edges, the computation is polynomial. It also checks if there is a disjoint star with density at least ρ4 . To compute this, it computes the densest star that is disjoint to S. 4.2 Analysis In this section, we present the analysis of our distributed approximation algorithm for the minimum 2-spanner problem, and prove the following theorem. Theorem 1.3. There is a distributed algorithm for the minimum 2-spanner problem in the Local model that guarantees an approximation ratio of O(log m n ), and takes O(log n log ∆) rounds w.h.p. Let H be the set of edges of the spanner produced by the algorithm. When the algorithm ends, ∗ all the edges are covered, hence H is a 2-spanner. We show that its size it at most O(log m n )\|H \|, 20 where H ∗ is the set of edges of a minimum 2-spanner. Afterwards, we show that the time complexity of the algorithm is O(log n log ∆) rounds, w.h.p. 4.2.1 Approximation ratio We start by showing that our algorithm guarantees an approximation ratio of O(log m n ). The analysis of the sequential algorithm of [46] that obtains the same approximation ratio strongly depends on the facts that stars are added to the spanner one by one and that the star that is added at each step has maximal rounded density. These allow dividing the edges to several subsets according to the order in which they are covered in the algorithm, and bounding the number of edges in each subset. Our analysis borrows ideas from the above analysis, but requires a more sophisticated accounting, since our algorithm adds multiple stars in each iteration, with varying densities. In addition to overcoming these uncertainties, a compelling aspect of our approach is that it easily extends to other variants of the problem, such as the client-server 2-spanner problem [29]. To show the approximation ratio, we assign each edge e ∈ E a value cost(e) such that the sum of the costs of all edges is closely related both to \|H\| and \|H ∗ \|, by satisfying X m ∗ \|H \|, \|H\| ≤ 8 cost(e) ≤ O log n e∈E which implies our claimed approximation ratio. We write H = H1 ∪ H2 , where H1 are edges added to the spanner during the algorithm, and H2 are edges added to the spanner at the end of the algorithm, when the maximal density in the 2-neighborhood of a vertex is at most 1. For an edge e ∈ H2 , we set cost(e) = 1. For an edge e ∈ H1 , let i be the iteration in which e is first covered in the algorithm. The edge e may be covered by a candidate star Sv that it votes for and is added to the spanner at iteration i. In this case, we set cost(e) = ρ1 , where ρ is the density of the star Sv that e chooses at iteration i. Another option is that e is covered as a result of adding other stars to the spanner at iteration i: it may be covered either by a different star than the one it votes for, or by a path of length 2 that is created by edges added to the spanner at iteration i together with edges added at previous iterations. In each of these cases, we set cost(e) = 0. We first show the left inequality above. P Lemma 4.1. \|H\| ≤ 8 · e∈E cost(e). P P \|H1 \| ≤ 8 P Proof. ToPprove that \|H\| ≤ 8 e∈E cost(e), it is enough to show thatP e∈E\H2 cost(e) and \|H2 \| ≤ 8 e∈H2 cost(e). The second inequality follows since \|H2 \| = e∈H2 1 = e∈H2 cost(e). We next prove the first inequality. P Let Stars be the set of stars added to H1 in the algorithm. It holds that \|H1 \| ≤ S∈Stars \|S\|, since each edge of H1 is included in at least one star. Let Sv be a star added to H1 at iteration i, having density ρ at that iteration. Recall that we add Sv to the spanner since it gets at least \|C8v \| votes from the edges it 2-spans. Denote by V otes(Sv ) the set of edges that vote for Sv at iteration i. For each e ∈ V otes(Sv ), we defined cost(e) = ρ1 , which gives, X e∈V otes(Sv ) cost(e) ≥ 1 \|Cv \| \|Sv \| \|Cv \| \|Sv \| · = · = . ρ 8 \|Cv \| 8 8 P Hence, for each S ∈ Stars, \|S\| ≤ 8 · e∈V otes(S) cost(e). For each edge e, there is at most one star S ∈ Stars such that e ∈ V otes(S), since an edge votes for at most one star at the iteration in which 21 it is covered. In addition, an edge e ∈ V otes(S) is 2-spanned during the algorithm, which means that e 6∈ H2 . Hence, we get X X X X \|H1 \| ≤ \|S\| ≤ 8 · cost(e) ≤ 8 · cost(e). S∈Stars S∈Stars e∈V otes(S) e∈E\H2 This completes the proof of Lemma 4.1. P To bound e∈E cost(e) from above, let r = m n , and f = dlog re. We divide the edges of E f +1 to f + 2 subsets {Ej }j=0 according to their costs, and show that for each j, the sum of costs ∗ of P edges in Ej is at most ∗O(\|H \|). Since there are f + 2 = O(log r) subsets, we conclude that e∈E cost(e) ≤ O(log r)\|H \|. Let E0 = {e : cost(e) = 0} and Ef +1 = H2 . Note that all the edges not in E0 and Ef +1 are edges that were 2-spanned in the algorithm by the candidate star they vote for. We divide these edges to f subsets as follows. Let E1 = {e 6∈ H2 : 0 < cost(e) ≤ 2r }, and for 2 ≤ j ≤ f , let j−1 j Ej = {e 6∈ H2 : 2 r < cost(e) ≤ 2r }. For each edge e, it holds that cost(e) ≤ 1, since the density of stars added to H1 during the algorithm is at least 1, and since we defined cost(e) = 1 for edges f +1 Ef . e ∈ H2 . Hence, for each edge e, we have cost(e) ≤ 2r , which gives E = ∪fj=0 P Lemma 4.2. For every 0 ≤ j ≤ f + 1, e∈Ej cost(e) = O(\|H ∗ \|). P Proof. For j = 0, the claim holds trivially. For j = 1, it holds that e∈E1 cost(e) ≤ 2r · \|E\| ≤ 2n = O(\|H ∗ \|), where the last equality follows from the fact that any 2-spanner for G has at least n − 1 edges, since G is connected. For 2 ≤ j ≤ f , let Hj∗ be the set of edges of a minimum 2-spanner for Ej . For each P vertex v, let ∗ Sj (v) be the full v-star in Hj∗ . We define Starsj = {Sj∗ (v)}v∈V . We next show that e∈Ej cost(e) ≤ P P P P 9 S∈Starsj \|S\|. To prove this, we write e∈Ej cost(e) = e∈Ej ∩Starsj cost(e)+ e∈Ej \Starsj cost(e). P P Since cost(e) ≤ 1, we get e∈Ej ∩Starsj cost(e) ≤ S∈Starsj \|S\|. P P We now show that e∈Ej \Starsj cost(e) ≤ 8 S∈Starsj \|S\|. Consider a specific star S ∈ Starsj , and let (e1 , ..., e` ) be the edges of Ej 2-spanned by S according to the order in which they were 2-spanned in the algorithm, breaking ties arbitrarily. Note that all the edges in Ej for 1 ≤ j ≤ f are 2-spanned in the algorithm as explained above. The density of S at the beginning of the iteration ` in which e1 is 2-spanned is at least \|S\| , since S may 2-span additional edges that are not in Ej . Since all the candidates that 2-span an edge have the same rounded density because they all have maximal rounded density in their 2-neighborhood, it holds that the density of the star Sv that ` ` 2-spans e1 is at least 4\|S\| , as v chooses a star with density at least ρ4v ≥ 4\|S\| . Hence, cost(e1 ) ≤ 4\|S\| ` . This gives, ` ≤ 4\|S\| cost(e1 ) ≤ each edge e ∈ Ej , cost(e) r 4\|S\|, where the 2j−1 2j ≤ r . Therefore, ` X i=1 cost(ei ) ≤ last inequality follows since e1 ∈ Ej . Note that for 2j 2j r `≤ · j−1 4\|S\| = 8\|S\|. r r 2 Let CS be the set of edges of Ej 2-spanned by the star S. Since Hj∗ is a 2-spanner for Ej , every edge e ∈ Ej \ Starsj is 2-spanned by at least one star S ∈ Starsj . Summing over all the stars in Starsj gives X X X X cost(e) ≤ cost(e) ≤ 8 \|S\|. e∈Ej \Starsj S∈Starsj e∈CS 22 S∈Starsj P Note that S∈Starsj \|S\| = 2\|Hj∗ \|, since each edge of Hj∗ is included in exactly two stars in Starsj . In addition \|Hj∗ \| ≤ \|H ∗ \| since H ∗ covers all the edges P of E, and in particular all the edges of Ej , and Hj∗ is the minimum 2-spanner of Ej . This gives e∈Ej cost(e) = O(\|H ∗ \|), which completes the proof for 2 ≤ j ≤ f . For j = f + 1, we define Hj∗ and Starsj as before. Let S ∈ Starsj , and let (e1 , ..., e` ) be the edges of Ej = H2 that are 2-spanned by S according to the order in which they were added to H2 in the algorithm, breaking ties arbitrarily. It must hold that ` ≤ \|S\|, as otherwise the density of S is greater than P` one at the iteration in which e1 is added to H2 , which contradicts the algorithm. This P gives i=1 cost(e∗i ) ≤ ` ≤ \|S\|. Following the same arguments for the case 2 ≤ j ≤ f , we get e∈Ej cost(e) = O(\|H \|), which completes the proof. P Lemmas 4.1 and 4.2 give \|H\| ≤ 8 e∈E cost(e) ≤ O(log r)\|H ∗ \|, which proves the following claimed approximation ratio. Lemma 4.3. The approximation ratio of the algorithm is O(log m n ). 4.2.2 Time complexity We now show that our algorithm completes in O(log n log ∆) rounds, w.h.p. In [43, 60], a potential function argument is given for analyzing the set cover and minimum dominating set problems that are addressed. We analyze our algorithm along a similar argument, but our algorithm necessitates a more intricate analysis, mainly due to the fact that each vertex may be the center of multiple stars that are added during the algorithm, rather than being chosen only once for a dominating set. The latter may contain at most n vertices, while for the spanner constructed by our algorithm there are initially n2∆ possible stars which may constitute it. Nevertheless, we show how to get a time complexity of O(log n log ∆) rounds for our minimum 2-spanner algorithm, which matches the time complexity of the above set cover and dominating set algorithms. The crucial component in proving our small time complexity is showing that as long as the rounded density of v does not change between iterations, v always chooses a star Sv that is equal to or contained in the star that it chooses in the previous iteration. As explained in Section 4.1, if the rounded density of v is the same in iterations i and i + 1, v tries to choose a star Svi+1 which is contained in Svi . We show that this is always the case. The following will be useful in our analysis. Observation 1. Let x1 , x2 , ..., xn be non-negative numbers, and let y1 , y2 , ..., yn be positive numbers, then Pn xi xi xi i=1 min ≤ Pn ≤ max . i i yi yi i=1 yi n o n o x In addition, the inequalities become equalities only if for all j, yjj = mini xyii = maxi xyii . Observation 1 follows from writing Pn i=1 xi = Pn xi i=1 yi · yi . We now prove the following. Claim 4.4. Let v be a candidate star in iteration i. If ρ̃(v, Hvi ) = ρ̃(v, Hvi+1 ) = ρ, then v chooses a star contained in Svi in iteration i + 1. Proof. Assume to the contrary that there is an iteration i such that ρ̃(v, Hvi ) = ρ̃(v, Hvi+1 ) = ρ, and there is no star contained in Svi with density at least ρ4 with respect to Hvi+1 . Let i0 be the first 0 iteration in which ρ̃(v, Hvi0 ) = ρ, and let i0 be the first iteration after i0 where ρ̃(v, Hvi ) = ρ and 0 0 ρ there is no star contained in Svi −1 with density at least 4 with respect to Hvi . Let S ∗ be the densest 23 0 0 0 v-star with respect to Hvi . Then ρ(S ∗ , Hvi ) ≥ ρ2 since ρ̃(v, Hvi ) = ρ. Let S0 be the full v-star, and let (S1 , S2 , ..., Sk ) be the sequence of stars chosen by v between iteration i0 and iteration i0 − 1 in the order in which they were chosen. For all 0 ≤ j ≤ k, it holds that Sj ⊆ Sj−1 , since i0 is the first iteration in which this does not hold. We next show by induction that S ∗ ⊆ Sj for all 0 ≤ j ≤ k. In particular this will give S ∗ ⊆ Sk . 0 Hence, at iteration i0 there is a star contained in Svi −1 = Sk with density at least ρ4 , in contradiction to the definition of i0 . For j = 0, the claim holds trivially since S0 is the full v-star. Assume that S ∗ ⊆ Sj−1 , and assume to the contrary that S ∗ * Sj . Note that both Sj and S ∗ are contained in Sj−1 by the induction hypothesis. Let j 0 be the iteration in which Sj is chosen. Since S ∗ * Sj , we can write S ∗ = S1 ∪ S2 where S1 = S ∗ ∩ Sj and S2 = S ∗ \ Sj . It holds that 0 0 0 0 2 \|+\|C12 \| ρ(S ∗ , Hvj ) ≥ ρ(S ∗ , Hvi ) ≥ ρ2 . We can write ρ(S ∗ , Hvi ) = \|C1 \|+\|C where C1 are edges of Hvi \|S1 \|+\|S2 \| 2-spanned by S1 , C2 are edges 2-spanned by S2 , and C12 are edges 2-spanned by S ∗ with one 0 endpoint in S1 and one endpoint in S2 . Since S ∗ is the densest star with respect to Hvi it follows 0 12 \| 1\| ∗ i0 that \|C2 \|+\|C ≥ ρ(S ∗ , Hvi ) ≥ ρ2 , as otherwise by Observation 1, \|C \|S2 \| \|S1 \| > ρ(S , Hv ), which shows that S1 is a denser star than S ∗ . This shows that at least one of 0 0 In the first case, ρ(S2 , Hvj ) ≥ ρ(S2 , Hvi ) = \|C2 \| \|S2 \| \|C2 \| \|S2 \| and \|C12 \| \|S2 \| is at least ρ4 . ≥ ρ4 , which shows that S2 is a disjoint star to ρ 12 \| Sj with density at least ρ4 that is contained in Sj−1 . In the second case, \|C \|S2 \| ≥ 4 . For an edge e all the edges of C e = {v, u} ∈ S2 , denote by C12 12 with endpoint u. It follows that there is an ρ e edge e ∈ S2 such that \|C12 \| ≥ 4 . By Observation 1 and since the density of Sj is at least ρ4 we \|C \|+\|C e \| j 12 get ρ(Sj ∪ {e}) ≥ \|S ≥ ρ4 , where Cj are the edges 2-spanned by Sj . Either way we get a j \|+1 contradiction to the definition of Sj . This completes the proof. The rest of the analysis is based on a potential function argument which is described in [43, 60] for the set cover and minimum dominating set problems. Let ρ = maxv∈V ρv at the beginning of P iteration i. We define the potential function φ = v:ρv =ρ \|Cv \|, where Cv is the set of edges of Hv that are 2-spanned by the star Sv which v chooses at iteration i. Note that the potential function may increase between iterations if the value of ρ changes. However, since we round the values ρv to powers of two, there may be only O(log ∆) different values for ρ. The obstacle is that φ may increase between iterations even if the value of ρ does not change, because a vertex v might change the stars Sv in different iterations. However, by Claim 4.4, as long as the rounded density of the vertex remains the same among iterations, it always chooses a star that is contained in the star that it chooses in the previous iteration. Hence, the size of the set of edges Cv can only decrease between the end of the last iteration to the beginning of the next one. It follows that as long as ρ does not change, the value of φ can only decrease between iterations. Our goal is to show that if the value of ρ does not change between iterations, the potential function φ decreases by a multiplicative factor between iterations in expectation. Having this, we get a time complexity of O(log n log ∆) rounds w.h.p. The following lemma shows that if the value of ρ does not change between iterations, the potential function φ decreases by a multiplicative factor between iterations in expectation. The proof follows the lines of the proofs in [43, 60], and is included here for completeness. We say that an iteration is legal if the random numbers rv chosen by the candidates in this iteration are different. Lemma 4.5. If φ and φ0 are the potentials at the beginning and end of a legal iteration, then E[φ0 ] ≤ c · φ for some positive constant c < 1. 24 In order to prove Lemma 4.5 we need the following definitions. Let s(e) be the number of candidates that 2-span the edge e. For a candidate v, we sort the edges in Cv according to s(e) in non-increasing order. Let T (v) and B(v) be the sets of the first d\|Cv \|/2e edges, and the last d\|Cv \|/2e edges in the sorted order, respectively. Indeed, if \|Cv \| is odd, the sets T (v) and B(v) share an edge. For a pair (Sv , e), where Sv is a candidate star that 2-spans e, we say that (Sv , e) is good if e ∈ T (v). We next show that if e ∈ T (v) chooses Sv in a legal iteration, then the star Sv is added to the spanner with constant probability. Claim 4.6. Let i be a legal iteration. If e, e0 are both 2-spanned by S in iteration i, and s(e) ≥ s(e0 ), then P r[e0 chooses S\|e chooses S] ≥ 12 . Proof. Let Ne , Ne0 , Nb be the number of candidates that 2-span e but not e0 , e0 but not e, and both e and e0 , respectively. P r[e and e0 choose S] P r[e chooses S\|e chooses S] = = P r[e chooses S] 0 1 Ne +Ne0 +Nb 1 Ne +Nb = \|Ne \| + \|Nb \| . \|Ne \| + \|Ne0 \| + \|Nb \| It holds that Ne ≥ Ne0 since s(e) ≥ s(e0 ). This gives, P r[e0 chooses S\|e chooses S] = \|Ne \| + \|Nb \| 1 \|Ne \| + \|Nb \| ≥ ≥ . \|Ne \| + \|Ne0 \| + \|Nb \| 2\|Ne \| + \|Nb \| 2 Claim 4.7. If (Sv , e) is a good pair in a legal iteration i, then P r[Sv is chosen\|e chooses Sv ] ≥ 31 . Proof. Assume that e chooses Sv . Denote by X the number of edges in B(v) that choose Sv , and let X 0 = \|B(v)\| − X. Note that e ∈ T (v) since (Sv , e) is good, therefore s(e) ≥ s(e0 ) for any edge e0 ∈ B(v). By Claim 4.6, any edge e0 ∈ B(v) chooses Sv with probability at least 12 . Hence, E[X] ≥ \|B2v \| . Equivalently, E[X 0 ] ≤ \|B2v \| . Using Markov’s inequality we get P r[X < \|Bv \| 3 3 2 ] = P r[X 0 > \|Bv \|] ≤ P r[X 0 ≥ E[X 0 ]] ≤ . 4 4 2 3 Hence, we get P r[X ≥ \|B4v \| ] ≥ 13 . Since \|Bv \| ≥ \|C2v \| , it holds that X ≥ \|C8v \| with probability at least \|Cv \| 1 3 . In this case, at least 8 edges choose Sv , and it is added to the spanner. This completes the proof. We can now bound the value of the potential function, by proving Lemma 4.5. Proof of Lemma 4.5. Let φ and φ0 be the potential function at the beginning and end Pvalues of the P P of a legal iteration i. It holds that φ = v:ρv =ρ \|Cv \| = (Sv ,e) 1 = e s(e), where we sum over all the edges 2-spanned by candidates having rounded density ρ, and over all the pairs (Sv , e) where v is a candidate having rounded density ρ that 2-spans e. Note that the rounded density of all the candidates that 2-span an edge e is the same, since they have maximal rounded density in their 2-neighborhood. If the edge e chooses the star Sv , and the star Sv is added to the spanner, φ decreases by s(e). We ascribe this decrease to the pair (Sv , e). Since e chooses only one candidate, 25 any decrease in φ is ascribed only to one pair. Hence, we get X P r[e chooses Sv , Sv is chosen] · s(e) E[φ − φ0 ] ≥ (Sv ,e) ≥ X P r[e chooses Sv ] · P r[Sv is chosen\|e chooses Sv ] · s(e) (Sv ,e) is good ≥ X (Sv ,e) is good 1 1 1 · · s(e) = s(e) 3 3 X 1. (Sv ,e) is good Since at least half of the pairs are good, we get E[φ − φ0 ] ≥ 16 φ, or equivalently E[φ0 ] ≤ 56 φ, which completes the proof. In conclusion, we get the following. Lemma 4.8. The time complexity of the algorithm is O(log n log ∆) rounds w.h.p. Proof. It holds that the maximum density of a star of size k is at most O(k 2 ) and the algorithm terminates when the maximum density is at most 1. Since densities are rounded to powers of 2, ρ = maxv∈V ρv may obtain at most O(log ∆) values, where ∆ is the maximum degree. In addition, by Lemma 4.5, if ρ has the same value at iterations j and j + 1, and j is a legal iteration, then the value of φ decreases between these iterations by a factor of at least 1/c in expectation. Since the random numbers rv are chosen from {1, ..., n4 }, they are different w.h.p, giving that if ρ has the same value in any two consecutive iterations then the value of φ decreases between these iterations by a constant factor in expectation. Since φ ≤ n3 , after O(log (n3 )) = O(log n) iterations in expectation, the value of ρ must decrease. This shows that the time complexity is O(log n log ∆) rounds in expectation. A Chernoff bound then gives that this also holds w.h.p. Lemma 4.3 and Lemma 4.8 complete the proof of Theorem 1.3. 4.3 Additional 2-spanner approximations Here we show that the algorithm extends easily to the following variants: the directed 2-spanner problem, the weighted 2-spanner problem and the client-server 2-spanner problem. We describe the differences in the algorithm and analysis in each of these cases. 4.3.1 Directed 2-spanner approximation In the directed case we consider directed stars. A v-star 2-spans a directed edge (u, w) if it includes the directed edges (u, v), (v, w). A v-star may include both ingoing and outgoing edges of v. The definition of densities follows the definition in the undirected case. In order to give an algorithm that requires only polynomial local computations for the directed variant, we show how to approximate the rounded density and the densest star in the directed case. The rest of the analysis follows the undirected case, and gives the following. Theorem 4.9. There is a distributed algorithm for the directed 2-spanner problem in the Local model that guarantees an approximation ratio of O(log m n ), and takes O(log n log ∆) rounds w.h.p. To compute an approximation for the densest (directed) v-star, we look at all the edges between neighbors of v, and remove all the directed edges (u, w) that cannot be 2-spanned by a v-star (such an edge can be 2-spanned by a v-star, only if the two directed edges (u, v), (v, w) exist in the graph). 26 Now we ignore the directions of edges and compute the densest v-star as in the undirected case. Let Sv be the star computed. Let ρU = ρU (Sv ) be the undirected density of v (when ignoring edges that cannot be 2-spanned by a directed path), and let ρD be the directed density of v. We will show that ρU ρU 2 ≤ ρD ≤ ρU , and that 2 ≤ ρD (Sv ), which shows that Sv gives a 2-approximation for the densest directed v-star. When computing ρD (Sv ) we replace each undirected edge {v, u} in Sv by the two directed edges (v, u), (u, v) if both of them exist in the graph, or by the one that exists otherwise. Claim 4.10. ρU 2 ≤ ρD (Sv ). Proof. Let Cv be the edges 2-spanned by Sv in the undirected case. Since Sv is the densest v\| undirected v-star, ρU = ρU (Sv ) = \|C \|Sv \| . The directed star Sv 2-spans all the edges in Cv (some of them may be counted twice in the directed case, which only increases the density), and it contains at most twice edges because we replaced each undirected edge by at most two edges, which gives \|Cv \| ρD (Sv ) ≥ 2\|S = ρ2U . v\| Claim 4.11. ρU 2 ≤ ρD ≤ ρU . Proof. By Claim 4.10, ρ2U ≤ ρD (Sv ) ≤ ρD . We next show ρD ≤ ρU . Let SD be the densest directed v-star, and let CD be the directed edges 2-spanned by SD . We write CD = C1 ∪ C2 where an edge (u, w) ∈ CD is in C2 if and only if the edge (w, u) is also in CD . We write SD = S1 ∪ S2 where an edge (v, u) ∈ SD is in S2 if and only if the edge (u, v) is also in SD . Now we look at the undirected density of SD (if we have two directed edges (w, u), (u, w) they are replaced by one undirected edge). ρU (SD ) = \|C1 \| + \|S1 \| + \|C2 \| 2 \|S2 \| 2 = \|C1 \| 2 \|S1 \| 2 + + \|C1 \|+\|C2 \| 2 \|S1 \|+\|S2 \| 2 ≥ min \|C1 \| \|C1 \| + \|C2 \| , \|S1 \| \|S1 \| + \|S2 \| = min \|C1 \| , ρD , \|S1 \| where the second inequality follows from Observation 1, and the last equality follows since SD is the 1 \|+\|C2 \| densest directed v-star, and its directed density equals \|C \|S1 \|+\|S2 \| . We next show that \|C1 \| \|S1 \| ≥ ρD . Since ρD = \|C1 \|+\|C2 \| \|S1 \|+\|S2 \| , if \|C1 \| \|S1 \| < ρD , we get by Observation 1 2\| that \|C \|S2 \| > ρD . Note that the directed star S2 2-spans all the directed edges in C2 (and it may 2-span additional edges), because all the edges in C2 appear in both directions, which means that the paths that 2-span them also appear in both directions in SD , which means that all the edges of these paths are in S2 . This shows that S2 is a directed star with density greater 1\| than ρD , in contradiction to the definition of ρD . In conclusion \|C \|S1 \| ≥ ρD , which shows that n o 1\| ρU ≥ ρU (SD ) ≥ min \|C \|S1 \| , ρD = ρD . Claims 4.10 and 4.11 show that we can approximate the directed density and the densest directed star using polynomial local computations. Having this, we can adapt the algorithm to the directed 2-spanner problem. We approximate the directed density of v with ρD (Sv ), which gives a 2-approximation. Then, we round the value of it to the closest power of two that is greater than ρD (Sv ), and denote the rounded value by ρv , this is a 2-approximation to the rounded density of v.7 While the value of ρv remains the same we choose stars similarly to the undirected case, the only difference is that when we look for a dense disjoint star, we do not necessarily find the densest directed disjoint star but a 2-approximation for it. For the analysis to work, we need to look for 7 Since we compute an approximation to the density, the value of ρv may increase between iterations. To avoid such cases, we always define it to be the minimum between the value in the last iteration and the value computed in the current iteration. This is always a 2-approximation for the rounded density since the density can only decrease between iterations. 27 disjoint stars with density at least ρ8v and not ρ4v as in the undirected case, and we add edges or disjoint stars to the star S we choose as long as the density of S is at least ρ8v . The rest of the analysis is similar, the constants change sightly since we choose stars that are less dense, and work with an approximation to the density. 4.3.2 Weighted 2-spanner approximation In the weighted case, the cost of a spanner is w(H), rather than \|H\| as in the unweighted case. This requires several changes in the algorithm and the analysis. Let W be the ratio between the maximum and minimum positive weights of an edge. We show the following. Theorem 4.12. There is a distributed algorithm for the weighted 2-spanner problem in the Local model that guarantees an approximation ratio of O(log ∆), and takes O(log n log (∆W )) rounds w.h.p. P We next describe the differences in the weighted case. For a star S, we define w(S) = e∈S w(e). \|CS \| If w(S) 6= 0, we define ρ(S, H) = w(S) , where CS is the set of edges of H 2-spanned by the star S. If w(S) = 0, we define ρ(S, H) = 0. We emphasize that we take the number of potentially 2-spanned edges and not the sum of their weights, since, intuitively, all edges need to be covered (as opposed to taking the sum of weights of the edges of the star, which is due to the need to optimize the cost of the spanner). In the beginning of the algorithm we add all the edges of weight 0 to the spanner. By doing this, all the edges covered by stars S of weight 0 are already covered. Hence, the algorithm should only consider stars for which w(S) > 0. When we round the densities to the closest power of two, we include also negative powers of two, since the density of a star may be smaller than 1, depending on the weights. A slight difference is that now a vertex terminates if the density in its 1 2-neighborhood is at most wmax where wmax is the maximal weight of an edge adjacent to a vertex in its 2-neighborhood. In such case, it adds to the spanner all the edges adjacent to it that are still not covered. We denote by H2 all these edges. The rest of the algorithm is the same, according to the new definition of ρ. As was observed in the sequential algorithm for the weighted case [45], we can find the densest star in the weighted case using flow techniques as well. We next describe the differences in the analysis. The cost of the solution obtained by the algorithm is now w(H), and the cost of an optimal solution is w(H ∗ ). We give to edges e 6∈ H2 a cost as in the unweighted case, but depending on the new definition of the density ρ. In addition, for edges e ∈ H2 we define cost(e) = w(e). Our goal is to show that X w(H) ≤ 8 cost(e) ≤ O(log ∆)w(H ∗ ). e∈E P The proof that w(H) ≤ 8 Pe∈E cost(e) is the same as the proof of Lemma P 4.1 with minor changes. Note that w(H2 ) = e∈H2 cost(e) by definition. Now w(H1 ) ≤ S∈Stars w(S), for the same P reason as in the unweighted case. In addition, the new definition of ρ, gives that w(S) ≤ 8 e∈V otes(S) cost(e), and the rest of the proof follows. However, the difference from the unweighted case is that we can no longer show an approximation ratio of O(log m n ). This is because the density of stars added in the algorithm may now be smaller than 1, and because the weight of an optimal 2-spanner may be smaller than n − 1. Still, we show an approximation ratio of O(log ∆). Some elements of our analysis have similar analogues in the classic analysis of the greedy set cover algorithm [16, 44, 53]. First, instead of Lemma 4.2 we show the following. 28 cost(e) ≤ O(log ∆)w(H ∗ ). P ∗ 0 Proof. First, we show that e∈E 0 cost(e) ≤ O(log ∆)w(H P ), where E = E \ (H2 ∩ E0 ) and E0 are edges with cost 0. Edges in E0 clearly do not affect e∈E cost(e). All the edges not in E0 or H2 are 2-spanned in the algorithm, as in the unweighted case. For each vertex v, let S ∗ (v) be the full v-star in H ∗ . We define Stars∗ = {S ∗ (v)}v∈V . Consider a star S ∈ Stars∗ and let (e1 , ..., e` ) be the sequence of edges 2-spanned by S according to the order in which they are 2-spanned in the algorithm. Assume first that w(S) 6= 0. The density of S at the ` beginning of the iteration in which e1 is 2-spanned is w(S) . All the candidates that 2-span e1 have the same rounded density since they all have maximal rounded density in their 2-neighborhood. In ` particular, the density of the star that 2-spans e1 is at least 4w(S) , as v chooses a star with density Lemma 4.13. at least ρv 4 ≥ P e∈E ` 4w(S) . Hence, cost(e1 ) ≤ iteration in which ej is 2-spanned is at ` X cost(ej ) ≤ 4w(S) · j=1 4w(S) ` . Similarly, the density of S at the beginning of 4w(S) least `−j+1 w(S) , which gives cost(ej ) ≤ `−j+1 . This gives, ` X j=1 the 1 = O(log `)w(S) = O(log ∆)w(S). `−j+1 The last equality is because the number of edges ` 2-spanned by a star is at most ∆2 . For a star S ∈ Stars∗ such that w(S) = 0, note that cost(e) = 0 for all the edges 2-spanned by S, since they are all covered P`at the beginning of the algorithm without voting for any candidate. Hence, we get in this case j ) = 0 = O(log ∆)w(S).P j=1 cost(eP P e∈E e∈E 0 ∩Stars∗ cost(e) P We now write P0 cost(e) = P + e∈E 0 \Stars∗ cost(e). It holds that cost(e) ≤ w(S). We next bound 0 ∗ ∗ e∈E ∩Stars S∈Stars e∈E 0 \Stars∗ cost(e). Let CS be the set of edges 2-spanned by the star S. Since H ∗ is a 2-spanner, every edge e ∈ E 0 \ Stars∗ is 2-spanned by at least one star S ∈ Stars∗ . Summing over all the stars in Stars∗ we get, X X X X cost(e) ≤ cost(e) ≤ O(log ∆) w(S). e∈E 0 \Stars∗ S∈Stars∗ e∈CS S∈Stars∗ P P In conclusion, e∈E P 0 cost(e) = O(log ∆) ∗ S∈Stars∗ w(S). It holds P that S∈Stars∗ w(S) = 2w(H ) since each edge of H ∗ is included in exactly two stars. ∗ This gives, e∈E 0 cost(e) = O(log ∆)w(H P ). To complete the proof, we bound e∈H2 cost(e). Let H2∗ be an optimal spanner for H2 . We define Stars∗ as before, with respect to H2∗ . Let S ∈ Stars∗ and let (e1 , ..., e` ) be the sequence of edges of H2 2-spanned by S according to the order in which they are added to H2 in the algorithm. ` 1 From the definition of H2 it must hold that w(S) ≤ wmax where wmax is the maximal weight in the 2-neighborhood of e1 (which in particular contains P the star S and P all the edges 2-spanned by it), as otherwise e1 was not added to H2 . This givesP `i=1 cost(e) = `i=1 w(e) ≤ ` · wmax ≤ w(S). ∗ ∗ ∗ Following as before, this 2 )). Since \|H2 \| ≤ \|H \|, Pthe same arguments P P gives e∈H2 cost(e) ≤ O(w(H ∗ we get e∈E cost(e) = e∈E 0 cost(e) + e∈H2 cost(e) = O(log ∆)w(H ). This completes the proof. P In conclusion, we get w(H) ≤ 8 e∈E cost(e) ≤ O(log ∆)w(H ∗ ), which completes the proof of the O(log ∆)-approximation ratio, giving the following lemma. Lemma 4.14. The approximation ratio of the algorithm is O(log ∆). To prove the round complexity, there are minor changes in the proof of Claim 4.4. First, we replace the size of a star \|S\| by its cost w(S) in order to work with the new definition of ρ. Note 29 that adding edges of weight 0 to a star can only increase its density, which shows that all the v-stars chosen in the algorithm, and in particular the star Sj , contain all the edges of weight 0 adjacent to v. This shows that the star S2 = S ∗ \ Sj includes only edges with positive weight. The proof carries over if w(S1 ) 6= 0. If w(S1 ) = 0, then all the edges 2-spanned by S1 are already 2-spanned at 0 \|+\|C12 \| 1 \|+\|C2 \|+\|C12 \| the beginning of the algorithm which shows ρ2 ≤ ρ(S ∗ , Hvi ) = \|Cw(S = \|C2w(S , as needed. 1 )+w(S2 ) 2) In addition, in the second case of the proof instead of showing that there is an edge e ∈ S2 with e e \| ≥ ρ , we show that \|C12 \| ≥ ρ . \|C12 4 4 w(e) The number of possible densities depends on the weights, in the following way. Let Wmax , Wmin max be the maximum and the minimum positive weights of an edge. Recall that W = W Wmin . The 2 maximum density of a star is at most W∆min since a star 2-spans at most ∆2 edges. In addition, the 1 algorithm terminates when the maximum density is Wmax . Since we round the densities to powers of two, there may at most O(log ∆W ) different non-zero values for the densities. The rest of the proof is exactly the same as in the unweighted case. 4.3.3 Client-server 2-spanner approximation Recall that in the Client-Server 2-spanner problem, the edges of the graph are divided to two types: clients and servers, and the goal is to cover all the client edges with server edges. Let C be the set of client edges, let V (C) be all the vertices that touch client edges, and let ∆S be the maximum degree in the subgraph of G that includes all the server edges. We show the following. Theorem 4.15. There is a distributed algorithm for the client-server 2-spanner problem in the Local model that guarantees an approximation ratio of O(min{log \|V\|C\| (C)\| , log ∆S }), and takes O(log n log ∆S ) rounds w.h.p. There are slight differences in the algorithm. First, throughout the algorithm and analysis, S\| we consider only stars composed of server edges, and for each such star we define ρ(S, H) = \|C \|S\| , where CS is the set of client edges of H 2-spanned by the star S. The set of edges Hv that a vertex v maintains consists only of client edges 2-spanned by the star that includes all the server edges adjacent to v. Now v terminates if the maximal density in its 2-neighborhood is below 12 and not at most 1 as before (since not all the client edges are server edges, perhaps the best way to cover a client edge is to take a path of length 2 that covers it, the density of the corresponding star is 12 ). Now cost(e) ≤ 2 which changes slightly the constants in the analysis. When v terminates, it adds an uncovered edge e to the spanner only if e is both a client and a server edge. These edges are the edges of H2 . Note that since not all the edges are server edges, there may be client edges that cannot be covered by server edges, in which case there is no solution to the problem, and our algorithm covers only all the edges that may be covered by server edges. When we analyze the algorithm, we assume that there is a solution to the problem, otherwise H ∗ is not defined. For other cases, we can restrict the client edges to be only edges that can be covered by server edges, and get a new problem that has an optimal solution H ∗ , and the approximation ratio we get is w.r.t to H ∗ . For the analysis, there are slight differences as follows. First, we give costs only to client edges, since these are the only edges we need to cover. We give the costs as in the minimum 2-spanner algorithm. In particular, cost(e) = 1 for e ∈ H2 . Our goal is to show that X \|C\| \|H\| ≤ 8 cost(e) ≤ O log \|H ∗ \|. \|V (C)\| e∈C 30 P The proof that \|H\| ≤ 8 e∈C cost(e) is exactly the same as the proof of Lemma 4.1. We next P \|C\| ∗ show that e∈C cost(e) ≤ O log \|V\|C\| (C)\| \|H \|. Let r = \|V (C)\| , and f = dlog re. We define the sets Ej according to the new definition of r. Let E1 = {e ∈ C \ H2 : 0 < cost(e) ≤ 2r }, and for j−1 j 2 ≤ j ≤ f + 1, let Ej = {e ∈ C \ H2 : 2 r < cost(e) ≤ 2r }. We define E0 as before, and Ef +2 = H2 . Since the stars added to H1 in the algorithm have density at least 12 , then cost(e) ≤ 2 for each edge +2 e ∈ C. This gives, C = ∪fj=0 Ej . We next show the following. P Lemma 4.16. For every 0 ≤ j ≤ f + 2, e∈Ej cost(e) = O(\|H ∗ \|). Proof. For j = 0, 2 ≤ j ≤ f + 1, and j = f + 2 the proof follows 0, 2 ≤ j ≤ f and P the cases j = 2 j = f + 1 in the proof of Lemma 4.2. For j = 1, it holds that e∈E1 cost(e) ≤ r · \|C\| ≤ 2\|V (C)\| = O(\|H ∗ \|). In the first inequality, we use the fact that we give costs only to edges of C. The last edges, which we prove next. equality follows from the fact that H ∗ includes at least \|V (C)\| 4 Let GC = (V (C), C), and let C1 , ..., C` be the connected components of GC . Note that each connected component of GC includes at least two vertices (since it includes at least one edge of C), which means that the number ` of connected components is at most \|V (C)\| . For a connected P` 2 component Ci , denote by ni the number of vertices in Ci , so that \|V (C)\| = i=1 ni . For a connected component Ci , denote by Hi all the edges of H ∗ that cover the edges in Ci . It holds that \|Hi \| ≥ ni −1 since Ci is connected, and the edges of Hi need to connect all the vertices in Ci , otherwise there is an edge in Ci which is not covered in H ∗ . In addition, for each edge e ∈ Hi , at least one of the vertices of e is in Ci , otherwise it cannot cover an edge in Ci . It follows that an edge e ∈ H ∗ can be in at most two different subsets Hi , Hj . This gives ` ` ` i=1 i=1 i=1 1X \|V (C)\| \|V (C)\| \|V (C)\| 1X 1X 1 \|H \| ≥ − = , \|Hi \| ≥ (ni − 1) = ni − ` ≥ 2 2 2 2 2 4 4 ∗ which completes the proof. P Pf +2 P ∗ By Lemma 4.16, we get cost(e) = e∈C e∈Ej cost(e) = O(log r)\|H \|. Since \|H\| ≤ j=0 P P 8 · e∈C cost(e), we have \|H\| ≤ 8 e∈C cost(e) ≤ O(log r)\|H ∗ \|, which shows an approximation ratio of O(log \|V\|C\| (C)\| ). P In addition, we can show that e∈C cost(e) ≤ O(log ∆S ), following the proof of Lemma 4.13, by replacing w(S) and w(H) by \|S\| and \|H\|. This shows an approximation ratio of O(log ∆S ) to the problem. Note that in the minimum 2-spanner problem, m n is half of the average degree in G, and ∆ is the maximum degree in G, hence an approximation ratio of O(log m n ) is better than O(log ∆). However, \|C\| in the client-server variant, it may be the case that ∆S ≤ \|V (C)\| depending on the client and server edges in G. The time analysis is the same as in the minimum 2-spanner problem. Note that there may be at most O(log ∆S ) different values for ρ because we consider only stars composed of server edges. This completes the proof of Theorem 4.15. 5 Distributed approximation for MDS In this section, we show that our algorithm can be modified to give an efficient algorithm for the minimum dominating set (MDS) problem, guaranteeing an approximation ratio of O(log ∆). In the MDS problem the goal is to find a minimum set of vertices D such that each vertex is either in D or has a neighbor in D. Our algorithm for MDS has the same structure of the algorithm of Jia et 31 al. [43], but it differs from it in the mechanism for symmetry breaking. Our approach guarantees an approximation ratio of O(log ∆), where in [43] the O(log ∆)-approximation ratio holds only in expectation. The following states our results for MDS. Theorem 5.1. There is a distributed algorithm for the minimum dominating set problem in the Congest model that guarantees an approximation ratio of O(log ∆), and takes O(log n log ∆) rounds w.h.p. For MDS, we define the star Sv centered at the vertex v as the set of vertices that contains v and all of its neighbors. Note that there is only one star centered at each vertex, which simplifies both the algorithm and its analysis. The density of a star S with respect to a subset of vertices U , denoted by ρ(S, U ), is defined as \|S ∩ U \|. The density of a vertex v with respect to U , denoted by ρ(v, U ), is defined as \|Sv ∩ U \|. The definition of the rounded density is the same as for our algorithm for the minimum 2-spanner problem. A vertex v maintains a set Uv that contains all the vertices in Sv that are still not covered by the vertices that have already been added to the dominating set, where a vertex is covered by a set if it is in that set or has a neighbor in that set. Our algorithm proceeds in iterations, where in each iteration the following is computed: 1. Each vertex v computes its rounded density ρv = ρ̃(v, Uv ), and sends it to its 2-neighborhood. 2. Each vertex v such that ρv ≥ ρu for each u in its 2-neighborhood is a candidate. Vertex v informs its neighbors that it is a candidate. Let Cv = Sv ∩ Uv . 3. Each candidate v chooses a random number rv ∈ {1, ..., n4 } and sends it to its neighbors. 4. Each uncovered vertex that is covered by at least one of the candidates, votes for the first candidate that covers it according to the order of the values rv . If there is more than one candidate with the same minimum value, it votes for the one with the minimum ID. 5. If v receives at least \|Cv \| 8 votes from vertices it covers then it is added to the dominating set. 6. Each vertex updates the set Uv by removing from it vertices that are now covered. If Uv = ∅, v outputs 1 if and only if it was added to the dominating set in the previous step. A crucial difference from our spanner approximation algorithm is that the densities are now based on the number of uncovered neighbors of v, and not the number of uncovered edges that can be potentially covered by a star. For this reason, all the computations in the algorithm can be implemented efficiently in the Congest model. The analysis of our MDS algorithm follows the same lines as the analysis of our minimum 2-spanner algorithm. We denote by D the dominating set produced by the algorithm, and by D∗ a minimum dominating set. We assign each vertex v with a value cost(v), which equals ρ1 if v is covered for the first timeP by a candidate having density ρ that v votes for, and otherwise, cost(v) = 0. We show that \|D\| ≤ 8 v∈V cost(v) ≤ O(log ∆)\|D∗ \|, which implies our claimed approximation ratio. P Lemma 5.2. \|D\| ≤ 8 · u∈V cost(u). Proof. The proof is similar to the proof of Lemma 4.1. For a vertex v ∈ D we denote by V otes(v) the vertices that vote for v. If v is added to D then it holds that at least \|C8v \| vertices vote 32 1 ρ, where ρ is the density of v, which is \|Cv \|, by P definition. Hence, for each vertex v ∈ D, it holds that u∈V otes(v) cost(u) ≥ ρ1 · \|C8v \| = 18 . Since each vertex otes(v), summing over all the vertices in D gives that P u is Pin at most one set V P \|D\| ≤ 8 · v∈D u∈V otes(v) cost(u) ≤ 8 · u∈V cost(u). P The proof that v∈V cost(v) ≤ O(log ∆)\|D∗ \| is similar to the proof of Lemma 4.13, where Stars∗ isPreplaced by D∗ , edges are replaced by vertices, and w(S) is replaced by 1 (note that the P ∗ equality S∈Stars∗ w(S) = 2w(H ) is replaced by v∈D∗ 1 = \|D∗ \|). Together with Lemma 5.2, this proves the approximation ratio of O(log ∆). For the time analysis, the main difference is that for each vertex v there is only one star Sv , which simplifies the proof (Claim 4.4 is no longer required). Let ρ = maxv∈V ρv at the beginning of P iteration i. We define the potential function φ = v:ρv =ρ \|Cv \|. If the value of ρ does not change between iterations, the value of φ can only decrease between iterations. By the definition of the densities, the density of a vertex v is a at most \|Sv \| which is at most ∆ + 1. Since we round the densities there may be at most O(log ∆) different values for ρ. Following the same analysis as in the analysis of our minimum 2-spanner algorithm (with the difference that edges are replaced by vertices, and a candidate is a vertex and not a star) we can show that if the value of ρ does not change between iterations, then the potential function φ decreases by a multiplicative factor between iterations in expectation. This gives a time complexity of O(log n log ∆) rounds w.h.p. Together with the approximation ratio, this proves Theorem 5.1. for it. The cost of each of these vertices is 6 Distributed (1 + )-approximation for spanner problems In this section, we show distributed (1 + )-approximation algorithms for spanner problems, following the framework of a recent algorithm for covering problems [39] (see Section 7).8 In a nutshell, the vertices invoke a network decomposition algorithm on the graph Gr , for a value of r = O(log n/) that can be computed by all vertices locally, given and a polynomial bound on n. This decomposes the graph into clusters of logarithmic diameter, colored by a logarithmic number of colors. Finally, by increasing order of colors, the vertices of each color select edges for the spanner. We show that indeed clusters of the same color can make their choices in parallel, and that the method of choosing edges to the spanner results in a (1 + ) approximation factor, giving the following. Theorem 1.2. There is a randomized algorithm with complexity O(poly(log n/)) in the Local model that computes a (1 + )-approximation of the minimum k-spanner w.h.p, where k is a constant. Proof. We start by describing a sequential (1 + )-approximation algorithm, and then explain how to implement it in the Local model using network decomposition. In the algorithm, the vertices start adding edges to the spanner H, which is initialized to be empty, while keeping track of all the edges covered by edges of H. At the beginning, all the edges are uncovered. To describe how this is done, we need the following notation. For a given integer d, denote by Bd (v) the subgraph of all the vertices within distance at most d from v and all the edges between them. For a vertex v and d ≥ 1, let g(v, d) be the size of an optimal spanner for all of the uncovered edges in Bd (v) (notice that the spanner can use both covered or uncovered edges of the whole graph G). We process the vertices according to a given order v1 , v2 , ..., vn . In step i, we look for the smallest radius ri such that g(vi , ri + 2k) ≤ (1 + )g(vi , ri ). Since an optimal spanner has size at most n2 , increasing the radius without the condition being met can only happen at most ri = O(log n/) times. We add to H an optimal spanner for all the uncovered edges in Bri +2k (vi ), and mark all 8 The presentation of the framework in [39] is slightly different and goes through an intermediate SLOCAL model. 33 the edges covered by the new edges of H as covered. In particular, all the edges of Bri +2k (vi ) are covered after this step. Note that an optimal spanner for Bri +2k (vi ) is contained in Bri +3k (vi ), which shows that step i depends only on a polylogarithmic neighborhood around vi . We next prove the approximation ratio of the algorithm. Denote by Ei all the edges of Bri (vi ) that are uncovered before step i. Since all the edges of Bri +2k (vi ) are covered after step i, it follows that Ei and Ej are at distance at least 2k + 1 for i 6= j. Let H ∗ be an optimal spanner, and let Hi∗ be the minimum set of edges in H ∗ that covers Ei . By the definition of a k-spanner, Hi∗ is contained in Bri +k (vi ), which shows that the subsets Hi∗ are disjoint. In step i, we added to H at most (1 + )g(vi , ri ) ≤ (1 + )\|Hi∗ \| edges, where the inequality follows since g(vi , ri ) is the size of an optimal spanner for Ei where Hi∗ is a spanner for Ei . Since ∪ni=1 Hi∗ ⊆ H ∗ and the subsets Hi∗ are disjoint, summing over all i gives \|H\| ≤ (1 + )\|H ∗ \|, which completes the approximation ratio proof. We now show how to implement the algorithm in the Local model (see also Proposition 3.2 in [38]). Let r = O(log n/) be such that r > ri + 4k for all i, and consider the graph Gr on the same set of vertices, where two vertices are connected if they are at distance at most r in the network graph G. Notice that in the Local model, any algorithm on Gr can be simulated by the vertices of G with an overhead of r rounds. The vertices invoke the randomized network decomposition algorithm of Linial and Saks [52] on the graph Gr . This algorithm decomposes a graph into clusters of diameter O(log n) that are colored with O(log n) colors, within O(log2 n) rounds. Invoked on Gr , this completes in poly(log n/) rounds. We assign a vertex v the label (qv , IDv ) where qv is the color of the cluster of v and IDv is the id of v. The lexicographic increasing order of the labels provides the order of the vertices. The distributed k-spanner algorithm runs in O(log n) phases, where in each phase `, the vertices of color ` are active, and collect all of the information of their cluster in Gr and its neighbors. Since the diameter of each cluster is at most O(log n), this completes in poly(log n/) rounds. Each vertex of the cluster then locally simulates the sequential algorithm for all the vertices in its cluster, according to their order. It can do so, since the sequential algorithm depends only on r-neighborhoods of vertices, and every two vertices in the same r-neighborhood are neighbors in Gr , which means they are either in the same cluster or in two clusters with different colors. This guarantees that the algorithm can indeed be executed in parallel for vertices of the same color. This completes the proof. The correctness of the algorithm relies only on the fact that the definition of k-spanners is local: an optimal spanner for Bd (v) is contained in Bd+k (v). Hence, the algorithm can be adapted similarly to the weighted, directed and client-server variants. In the weighted case the complexity is O(poly(log (nW )/)), where W is the ratio between the maximum and minimum positive weights of an edge. Acknowledgment: We would like to thank Seri Khoury for fruitful discussions. References [1] Amir Abboud, Keren Censor-Hillel, and Seri Khoury. Near-linear lower bounds for distributed distance computations, even in sparse networks. In Proceedings of the 30th International Symposium on Distributed Computing (DISC), pages 29–42, 2016. [2] Baruch Awerbuch, Boaz Patt-Shamir, David Peleg, and Michael Saks. Adapting to asynchronous dynamic networks. In Proceedings of the twenty-fourth annual ACM symposium on Theory of computing (STOC), pages 557–570, 1992. 34 [3] Baruch Awerbuch and David Peleg. Network synchronization with polylogarithmic overhead. In Proceedings of the 31st Annual Symposium on Foundations of Computer Science (FOCS), pages 514–522, 1990. [4] Baruch Awerbuch and David Peleg. Routing with polynomial communication-space trade-off. SIAM Journal on Discrete Mathematics, 5(2):151–162, 1992. [5] Leonid Barenboim, Michael Elkin, and Cyril Gavoille. A fast network-decomposition algorithm and its applications to constant-time distributed computation. Theoretical Computer Science, 2016. [6] Surender Baswana and Sandeep Sen. Approximate distance oracles for unweighted graphs in expected O(n2 ) time. ACM Transactions on Algorithms (TALG), 2(4):557–577, 2006. [7] Surender Baswana and Sandeep Sen. A simple and linear time randomized algorithm for computing sparse spanners in weighted graphs. Random Structures & Algorithms, 30(4):532– 563, 2007. [8] Piotr Berman, Arnab Bhattacharyya, Konstantin Makarychev, Sofya Raskhodnikova, and Grigory Yaroslavtsev. Approximation algorithms for spanner problems and directed steiner forest. Information and Computation, 222:93–107, 2013. [9] Piotr Berman, Sofya Raskhodnikova, and Ge Ruan. Finding sparser directed spanners. In IARCS Annual Conference on Foundations of Software Technology and Theoretical Computer Science (FSTTCS), pages 424–435, 2010. [10] Keren Censor-Hillel, Telikepalli Kavitha, Ami Paz, and Amir Yehudayoff. Distributed construction of purely additive spanners. In International Symposium on Distributed Computing (DISC), pages 129–142, 2016. [11] Keren Censor-Hillel, Seri Khoury, and Ami Paz. Quadratic and near-quadratic lower bounds for the CONGEST model. In 31st International Symposium on Distributed Computing, DISC 2017, October 16-20, 2017, Vienna, Austria, pages 10:1–10:16, 2017. [12] Shiri Chechik. Compact routing schemes with improved stretch. In Proceedings of the 2013 ACM symposium on Principles of distributed computing (PODC), pages 33–41, 2013. [13] Eden Chlamtác and Michael Dinitz. Lowest degree k-spanner: Approximation and hardness. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques (APPROX/RANDOM), pages 80–95, 2014. [14] Eden Chlamtáč, Michael Dinitz, Guy Kortsarz, and Bundit Laekhanukit. Approximating spanners and directed steiner forest: Upper and lower bounds. In Proceedings of the TwentyEighth Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 534–553, 2017. [15] Eden Chlamtac, Michael Dinitz, and Robert Krauthgamer. Everywhere-sparse spanners via dense subgraphs. In IEEE 53rd Annual Symposium on Foundations of Computer Science (FOCS), pages 758–767, 2012. [16] Vasek Chvatal. A greedy heuristic for the set-covering problem. Mathematics of operations research, 4(3):233–235, 1979. 35 [17] Bilel Derbel, Cyril Gavoille, David Peleg, and Laurent Viennot. On the locality of distributed sparse spanner construction. In Proceedings of the twenty-seventh ACM symposium on Principles of distributed computing (PODC), pages 273–282, 2008. [18] Bilel Derbel, Mohamed Mosbah, and Akka Zemmari. Sublinear fully distributed partition with applications. Theory of Computing Systems, 47(2):368–404, 2010. [19] Michael Dinitz, Guy Kortsarz, and Ran Raz. Label cover instances with large girth and the hardness of approximating basic k-spanner. ACM Transactions on Algorithms (TALG), 12(2):25, 2016. [20] Michael Dinitz and Robert Krauthgamer. Directed spanners via flow-based linear programs. In Proceedings of the forty-third annual ACM symposium on Theory of computing (STOC), pages 323–332, 2011. [21] Michael Dinitz and Robert Krauthgamer. Fault-tolerant spanners: better and simpler. In Proceedings of the 30th annual ACM SIGACT-SIGOPS symposium on Principles of distributed computing (PODC), pages 169–178, 2011. [22] Michael Dinitz and Yasamin Nazari. Distributed distance-bounded network design through distributed convex programming. In Proceedings of the The 21st International Conference on Principles of Distributed Systems (OPODIS), 2017. [23] Michael Dinitz and Zeyu Zhang. Approximating low-stretch spanners. In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 821–840, 2016. [24] Andrew Drucker, Fabian Kuhn, and Rotem Oshman. On the power of the congested clique model. In Proceedings of the 2014 ACM symposium on Principles of distributed computing (PODC), pages 367–376, 2014. [25] Michael Elkin. Computing almost shortest paths. ACM Transactions on Algorithms (TALG), 1(2):283–323, 2005. [26] Michael Elkin. An unconditional lower bound on the time-approximation trade-off for the distributed minimum spanning tree problem. SIAM J. Comput., 36(2):433–456, 2006. [27] Michael Elkin. A near-optimal distributed fully dynamic algorithm for maintaining sparse spanners. In Proceedings of the twenty-sixth annual ACM symposium on Principles of distributed computing (PODC), pages 185–194, 2007. [28] Michael Elkin and Ofer Neiman. Efficient algorithms for constructing very sparse spanners and emulators. In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 652–669, 2017. [29] Michael Elkin and David Peleg. The client-server 2-spanner problem with applications to network design. In 8th International Colloquium on Structural Information and Communication Complexity (SIROCCO), pages 117–132, 2001. [30] Michael Elkin and David Peleg. Approximating k-spanner problems for k > 2. Theoretical Computer Science, 337(1):249–277, 2005. 36 [31] Michael Elkin and David Peleg. The hardness of approximating spanner problems. Theory of Computing Systems, 41(4):691–729, 2007. [32] Michael Elkin and Jian Zhang. Efficient algorithms for constructing (1+ε, β)-spanners in the distributed and streaming models. In Proceedings of the twenty-third annual ACM symposium on Principles of distributed computing (PODC), pages 160–168, 2004. [33] Paul Erdős. Extremal problems in graph theory. In Theory Of Graphs And Its Applications, Proceedings of Symposium Smolenice, pages 29–36. Publ. House Cszechoslovak Acad. Sci., Prague, 1964. [34] Orr Fischer, Tzlil Gonen, and Rotem Oshman. Distributed property testing for subgraphfreeness revisited. CoRR, abs/1705.04033, 2017. [35] Silvio Frischknecht, Stephan Holzer, and Roger Wattenhofer. Networks cannot compute their diameter in sublinear time. In Proceedings of the twenty-third annual ACM-SIAM symposium on Discrete Algorithms (SODA), pages 1150–1162, 2012. [36] Giorgio Gallo, Michael D Grigoriadis, and Robert E Tarjan. A fast parametric maximum flow algorithm and applications. SIAM Journal on Computing, 18(1):30–55, 1989. [37] Mohsen Ghaffari. Near-optimal distributed approximation of minimum-weight connected dominating set. In Proceedings of the 41st International Colloquium on Automata, Languages, and Programming (ICALP), pages 483–494, 2014. [38] Mohsen Ghaffari, David G Harris, and Fabian Kuhn. On derandomizing local distributed algorithms. arXiv preprint arXiv:1711.02194, 2017. [39] Mohsen Ghaffari, Fabian Kuhn, and Yannic Maus. On the complexity of local distributed graph problems. In Proceedings of the 49th Annual ACM SIGACT Symposium on Theory of Computing (STOC), pages 784–797. ACM, 2017. [40] Ofer Grossman and Merav Parter. Improved deterministic distributed construction of spanners. In 31st International Symposium on Distributed Computing, DISC 2017, October 16-20, 2017, Vienna, Austria, pages 24:1–24:16, 2017. [41] David G Harris, Johannes Schneider, and Hsin-Hao Su. Distributed (∆+1)-coloring in sublogarithmic rounds. In Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing (STOC), pages 465–478, 2016. [42] Stephan Holzer and Roger Wattenhofer. Optimal distributed all pairs shortest paths and applications. In Proceedings of the 2012 ACM symposium on Principles of distributed computing (PODC), pages 355–364. ACM, 2012. [43] Lujun Jia, Rajmohan Rajaraman, and Torsten Suel. An efficient distributed algorithm for constructing small dominating sets. Distributed Computing, 15(4):193–205, 2002. [44] David S Johnson. Approximation algorithms for combinatorial problems. Journal of computer and system sciences, 9(3):256–278, 1974. [45] Guy Kortsarz. On the hardness of approximating spanners. Algorithmica, 30(3):432–450, 2001. [46] Guy Kortsarz and David Peleg. Generating sparse 2-spanners. Journal of Algorithms, 17(2):222– 236, 1994. 37 [47] Guy Kortsarz and David Peleg. Generating low-degree 2-spanners. SIAM Journal on Computing, 27(5):1438–1456, 1998. [48] Fabian Kuhn, Thomas Moscibroda, and Roger Wattenhofer. Local computation: Lower and upper bounds. Journal of the ACM (JACM), 63(2):17, 2016. [49] Fabian Kuhn and Rogert Wattenhofer. Constant-time distributed dominating set approximation. In Proceedings of the twenty-second annual symposium on Principles of distributed computing (PODC), pages 25–32, 2003. [50] Eyal Kushilevitz and Noam Nisan. Communication Complexity. Cambridge University Press, New York, NY, USA, 1997. [51] Nathan Linial. Locality in distributed graph algorithms. SIAM J. Comput., 21(1):193–201, 1992. [52] Nathan Linial and Michael Saks. Low diameter graph decompositions. Combinatorica, 13(4):441– 454, 1993. [53] László Lovász. On the ratio of optimal integral and fractional covers. Discrete mathematics, 13(4):383–390, 1975. [54] David Peleg. Distributed Computing: A Locality-Sensitive Approach. SIAM, 2000. [55] David Peleg and Vitaly Rubinovich. A near-tight lower bound on the time complexity of distributed minimum-weight spanning tree construction. SIAM J. Comput., 30(5):1427–1442, 2000. [56] David Peleg and Alejandro A Schäffer. Graph spanners. Journal of graph theory, 13(1):99–116, 1989. [57] David Peleg and Jeffrey D Ullman. An optimal synchronizer for the hypercube. SIAM Journal on computing, 18(4):740–747, 1989. [58] David Peleg and Eli Upfal. A trade-off between space and efficiency for routing tables. Journal of the ACM (JACM), 36(3):510–530, 1989. [59] Seth Pettie. Distributed algorithms for ultrasparse spanners and linear size skeletons. Distributed Computing, 22(3):147–166, 2010. [60] Sridhar Rajagopalan and Vijay V Vazirani. Primal-dual rnc approximation algorithms for set cover and covering integer programs. SIAM Journal on Computing, 28(2):525–540, 1998. [61] Alexander A. Razborov. On the distributional complexity of disjointness. Theoretical Computer Science, 106(2):385–390, 1992. [62] Liam Roditty, Mikkel Thorup, and Uri Zwick. Deterministic constructions of approximate distance oracles and spanners. In International Colloquium on Automata, Languages, and Programming (ICALP), pages 261–272, 2005. [63] Atish Das Sarma, Stephan Holzer, Liah Kor, Amos Korman, Danupon Nanongkai, Gopal Pandurangan, David Peleg, and Roger Wattenhofer. Distributed verification and hardness of distributed approximation. SIAM Journal on Computing, 41(5):1235–1265, 2012. 38 [64] Mikkel Thorup and Uri Zwick. Compact routing schemes. In Proceedings of the thirteenth annual ACM symposium on Parallel algorithms and architectures (SPAA), pages 1–10, 2001. [65] Mikkel Thorup and Uri Zwick. Approximate distance oracles. Journal of the ACM (JACM), 52(1):1–24, 2005. 39	8cs.DS
arXiv:1704.04782v1 [cs.DC] 16 Apr 2017 A Security Monitoring Framework For Virtualization Based HEP Infrastructures A. Gomez Ramirez1 , M. Martinez Pedreira2 , C. Grigoras2 , L. Betev2 , C. Lara1 and U. Kebschull1 for the ALICE Collaboration 1 2 Infrastructure and Computer Systems for Data Processing (IRI), Goethe-University Frankfurt CERN, Geneva, Switzerland E-mail: andres.gomez@cern.ch Abstract. High Energy Physics (HEP) distributed computing infrastructures require automatic tools to monitor, analyze and react to potential security incidents. These tools should collect and inspect data such as resource consumption, logs and sequence of system calls for detecting anomalies that indicate the presence of a malicious agent. They should also be able to perform automated reactions to attacks without administrator intervention. We describe a novel framework that accomplishes these requirements, with a proof of concept implementation for the ALICE experiment at CERN. We show how we achieve a fully virtualized environment that improves the security by isolating services and Jobs without a significant performance impact. We also describe a collected dataset for Machine Learning based Intrusion Prevention and Detection Systems on Grid computing. This dataset is composed of resource consumption measurements (such as CPU, RAM and network traffic), logfiles from operating system services, and system call data collected from production Jobs running in an ALICE Grid test site and a big set of malware. This malware was collected from security research sites. Based on this dataset, we will proceed to develop Machine Learning algorithms able to detect malicious Jobs. 1. Introduction Frequently in HEP computing, and also in general purpose Grid computing, user supplied code and data are deployed and executed in farms around the world, while the exact location is normally irrelevant. This allows scientists from many areas beyond physics to use huge computational power to solve complicated scientific problems, such as weather modeling, brain simulation, among others. However it also creates cyber-security challenges. Operators and administrators need tools to monitor for security incidents. User code and data should be isolated from different users, also from the physical computers and networks, in order to restrict access to sensitive elements in the organizations. We propose a novel paradigm and developed a framework that focus on protecting and monitoring user payload execution. Moreover, it enforces isolation in the environment in such a way that Jobs cannot access sensitive resources. This tool enables the Job behavior analysis in order to detect possible intrusions. This is accomplished by collecting and processing data generated by Jobs such as logs, system calls and resource consumption data. Traditional Intrusion Detection and Prevention Systems (IDPS) perform attack detection by using fixed rules based on signatures, identical to traditional monitoring systems. Therefore, we employ Machine Learning (ML) to overcome the mentioned drawbacks, achieving generalization among attack variants. Currently there is no tool that provides isolation, while monitoring security incidents by ML algorithms in Grid computing [1]. The authors have defined a threat model that guides the design and implementation of the described framework [1]. It is devised to detect attackers in the protected system trying actions like the following: • • • • • • Exploit unknown or unfixed software/hardware vulnerabilities. Listen to user network traffic to gather sensitive clear text information. Perform a ’man in the middle’ attack. Tamper other user Jobs. Escalate privileges. Access sensitive server configuration data. As a proof of concept we are implementing the described framework for the ALICE Grid at CERN. ALICE (A Large Ion Collider Experiment) is a dedicated Pb–Pb detector designed to exploit the physics potential of nucleus-nucleus interactions at the Large Hadron Collider at CERN [2, 3]. The ALICE experiment has developed the ALICE production environment (AliEn) [4], which implements many components of the Grid technologies that are needed to analyze HEP data. Through AliEn, the computing centers that participate in the ALICE Grid can be seen and used as a single entity. Any available node executes Jobs and file access is transparent to the user, wherever in the world a file might be [5]. Figure 1 shows a picture of the ALICE Grid. Figure 1. ALICE Grid computing farms aroud the world. This document is organized as follows: section 2 introduces a security by isolation strategy for a distributed system. Section 3 explains a method for collecting relevant data from the isolated infrastructure. Section 4 shows how the collected data can be used to determine the security status of the system. In section 5 we detail our Intrusion Detection and Prevention model. Section 6 summarizes the current state of the project and the challenges faced in the design and implementation of the desired methodology. Finally, section 7 gives conclusions on the work done. 2. Security by isolation Security by isolation enforces application space separation [8]. The idea is that, if one process is compromised and utilized to attack the entire system, other components can stay untouched. Several technologies provide secure isolation. Virtual Machines (VM) and Linux Containers (LC) are very popular examples. 2.1. Linux containers LC are an extension of the virtual memory concept to allow the isolation of network interfaces, the PID tree and mount points [9]. Separation of containers from the rest of the system is enforced by the Kernel, so they can not affect the host or other containers. LC technology uses namespaces and Cgroups [10] to have a private view of the system and a limited resource assignment. Containers provide a set of advantages over VM. They are lightweight and fast, boot in milliseconds and have just a few MB of intrinsic disk and memory usage. It has been shown [11], that they provide a better performance than VMs. Commonly, VM are used in Grid and Cloud computing to achieve isolation, however LC performance and comparable security features make it a suitable alternative[6, 7]. 2.2. Proposed isolation architecture We propose the usage of LC to enforce HEP Grid site user isolation, also extensible to broader scientific computing and clouds. To achieve this we require a batch Job orchestrator allowing the execution of user processes in containers on computing clusters. Section 6 gives further details about this requirement and the selected solution. As shown in Figure 2, we switch from an environment without isolation, where Jobs have access to the server and other user Jobs, to an environment where Jobs run in their one process space, without access to other Jobs or sensitive components. Pilot Job Bob's Grid Alice's Grid Job Job Bob's Grid Job Container Pilot Job Alice's Grid Job Pilot Job Container Figure 2. Desired isolation scenario. An isolated environment for Job execution is not enough. Jobs could still perform several kind of attacks or not allowed activities such as, Distributed Denegation of Service (DDoS), Bitcoin mining, malware/botnet hosting, among others. Consequently, we need to monitor the activity and detect incidents inside the LC, as described in the next sections. 3. Monitoring data mining HEP distributed computing systems use continuous automated monitoring to help administrators to find and fix situations affecting the normal operation [12]. The resulting monitoring data can be used to find or even predict software and hardware failures. It is a valuable source of security information as well. In this document we focus on the measurement of metrics related to the batch Jobs being submitted to the distributed system. There are several relevant metrics that we can collect, for instance: • Job and system logs. • System call sequence. • Resource usage data (such as CPU, RAM and network traffic). Furthermore our goal is to chose the best information about Job behavior without affecting the habitual performance. We decided to employ data mining and intelligent algorithms, given their ability to find correlations and analyze trends in big datasets [13], in order to provide a better understanding of security related events. 4. Machine learning based security monitoring Machine Learning is a set of mathematical models that simulate the human learning abilities[14, 15]. In the context of Intrusion Detection, ML helps analyzing big amounts of data by learning the expected behavior and identifying abnormal situations. Traditional industrial IDS use rather fixed rules and search for known attack signatures. However they have problems when unknown or slightly different intrusion methods are used [13]. We have selected supervised training to analyze the collected data. In supervised training, a set of already classified and tagged data (training dataset) is used to model a function (for example a Neural Network) in order to make it able to classify new unseen data (test dataset) [16]. Figure 3. Monitored data gathering and processing. 4.1. Training dataset We have collected a training and testing dataset by gathering monitoring data from production ALICE Grid Jobs (Figure 3). Additionally we have executed a big set of Linux malware samples. The data on the first part is tagged as normal data and the second as malicious. This dataset is utilized to compare several Machine Learning algorithms to find the one that gives the best accuracy. Following is the list of ML algorithms selected. They will be tested to define which one gives the best accurate results for our dataset: • Support Vector Machines. • Multilayer Neural Networks. • Recurrent Neural Networks. Figure 4 shows a scheme of the proposed architecture for the ML usage. Figure 4. Proposed architecture. 5. Intrusion prevention and detection When analyzing Job monitoring data, our goal is finding security incidents. This security incidents can be found by analyzing anomalies in the system, things that go beyond the common state, probably caused by malicious software. Besides, even if our execution environment is sandboxed, there are many possible attacks that can still affect the distributed infrastructure. If a user’s Job is misbehaving, the proposed framework should raise an alarm and perform predefined actions, for instance terminate the malicious processes. Figure 5 shows the desired implementation of the proposed system regarding Intrusion Detection. Intrusion Detection and Prevention • Alerts for Admins • Automated Actions Figure 5. Proposed Intrusion Detection in the worker nodes. 5.1. Challenges Improving the provided security should not impact the system performance. This is especially important in HEP computing. On the other hand, we also need innovative ways to analyze the trace and log data in an efficient way. Another important challenge is to reduce the amount of false positives and false negatives, since the system administrators rely on the accuracy of the security monitoring framework. 6. Proof of concept and testing environtment So far we have already deployed a testing ALICE Grid site based on AliEn [4] in a local Linux cluster, with five Ubuntu 14.04 nodes. In order to orchestrate and run the Jobs inside Linux Containers we have tested three different tools that offer such functionality: • Kubernetes [17]. • Apache Mesos [18]. • Docker Swarm [19]. At the end we have decided to work with Docker Swarm, because it allows to carry out the simplest deployment, which is an important requirement for our research environment. We use Docker [20] as LC engine, with Centos 6 [21] container images. We have developed AliEn interfaces for the mentioned batch systems. CVMFS [22] is installed on the hosts and shared as a volume inside the AliEn container to allow access to HEP libraries. Currently we execute one Job per container. This is useful to increase the traceability between different Jobs. Also, this is the natural micro service model for LC. As a monitoring infrastructure for collecting data from normal Grid Jobs we have Prometheus [23] and Sysdig [24]. Prometheus allows to take resource usage data directly from containers and collect it via a RESTful interface. Sysdig enables to capture system calls in Linux OS in a fast a reliable way. We have develop a custom Python library to integrate these tools and make them fit our needs. This infrastructure has been utilized for the execution and measurement of ALICE production Jobs, that are tagged as normal Jobs. A network isolated machine was used for malware data collection. This machine has the same setup as the Grid worker nodes. We have downloaded a set of 10000 Linux malware samples from a security research web site [25]. We ran the samples and collected the same information as for the normal Jobs (logs, sequence of system calls, resource usage data). Finally we obtained a combined dataset that allows to train and test our selected Machine Learning algorithms. A representation of the implemented components are shown in the Figure 6. 7. Conclusions Distributed computing is a fundamental component of High Energy Physics collaborations. Improving security in this kind of infrastructures requires innovative tools to automatically detect security related incidents. Security by isolation is also necessary to protect sensitive components allowing traceability on Job activities. We propose the usage of Linux Containers in order to provide isolation without highly decreasing the expected performance. We use Machine Learning techniques to provide generalization, overcoming common IDS difficulties on finding even slightly different threats. We describe the ongoing development process of a new security monitoring framework for Linux Containers based HEP infrastructures. This is being tested as a proof of concept for the ALICE experiment at CERN. We have collected a dataset of normal and malicious monitoring information from Grid Jobs and malware samples, that will be utilized to train and test ML algorithms. These algorithms should enable autonomous Intrusion Detection and Prevention as an important component of the proposed new framework. As future work we plan to explore how our approach can be used to detect anomalies that go beyond the security scope, for instance to find hardware failures or even human mistakes. Acknowledgments Authors acknowledge assistance from CERN security department specially Stefan Lueders and Romain Wartel. This work is supported by the German Federal Ministry of Education and Research. SYSDIG • System calls • Connections • Resource usage Figure 6. Proof of concept implementation. References [1] Gomez Ramirez A, Lara C and Kebschull U 2015 JPCS. Intrusion Prevention and Detection in Grid Computing - The ALICE Case. [2] The ALICE collaboration 1995 ALICE - Technical Proposal for A Large Ion Collider Experiment at the CERN LHC, CERN, Geneve, Rep. CERN-LHCC-95-71; LHCC-P-3. [3] The ALICE collaboration 2008 JINST. The ALICE Experiment at the CERN LHC, , vol. 3, no. 08, Aug. [4] Begnasco S et al. 2015 JPCS. AliEn: ALICE environment on the GRID. IOP Publishing. [5] Christoph E et al. 2005 Technical Design Report LCG. LHC computing Grid: Technical Design Report. Version 1.06. 20 Jun 2005. CERN, Geneve. Available from: https://cds.cern.ch/record/840543. [6] Abed A, Clancy C and Levy C 2015 Security and Trust Management, Lecture Notes in Computer Science. Intrusion Detection System for Applications Using Linux Containers. Springer International Publishing. [7] Modi C et al. 2013 JNCA. A survey of intrusion detection techniques in Cloud. Jan;36(1):42-57. [8] Mansfield-Devine S 2010 Computer Fraud & Security. Security through isolation. 5, 8 - 11. 1361-3723. [9] Graber S, 2014 Ubuntu Foundations Team, LXC 1.0: Blog post series. [updated 2014 Jan 17; cited 2015 May 11]. Available from: https://www.stgraber.org/2013/12/20/lxc-1-0-blog-post-series/ [10] Menage P 2007 Proceedings of the Linux Symposium. Adding Generic Process Containers to the Linux Kernel. June 27. [11] Shiseki A 2016 The 22nd International Conference on Computing in High Energy and Nuclear Physics, CHEP 2016. Cloud Computing as a Scientific and Technical Application Development and Execution Platform. October 10-14. [12] Lara C et al. 2011 JPCS. Autonomous System Management for the ALICE High-Level-Trigger Cluster using the SysMES Framework. doi:10.1088/1742-6596/331/5/052003. [13] Azad V 2013 IJITCS. Data Mining in Intrusion Detection: A Comparative Study of Methods, Types and Data Sets. 75-90. [14] Yuping L et al. 2012 IJACT. An Intrusion Detection Approach Using SVM and Multiple Kernel Method. Jan 31;4(1):463-9. [15] Gascon H et al. 2014 ACM Press; 2013. Structural detection of android malware using embedded call graphs. [cited 2014 Nov 17]. p. 45-54. Available from: http://dl.acm.org/citation.cfm?doid=2517312.2517315 [16] Bishop C 2006 Library of Congress ISBN Pattern recognition and machine learning. New York. Springer. [17] Google, Kubernetes. 2016 [updated 2016 Dec 1; cited 2016 Dec.1]. Available from: http://kubernetes.io/ [18] Apache foundation, Mesos 2016 [updated 2016 Dec 1; cited 2016 Dec.1]. Available from: https://mesos.apache.org/ [19] Docker, Docker Swarm 2016 [updated 2016 Dec 1; cited 2016 Dec.1]. Available from: https://docs.docker.com/swarm/ [20] Docker, Docker 2016 [updated 2016 Dec 1; cited 2016 Dec.1]. Available from: https://docs.docker.com [21] CentOS Project, Centos6 2016 [updated 2016 Dec 1; cited 2016 Dec.1]. Available from: https://www.centos.org/ [22] Buncic P et al 2008 Proceedings of the XII. International Workshop on Advanced Computing and Analysis Techniques in Physics Research (ACAT08). CernVM - a virtual appliance for LHC applications. Erice, PoS(ACAT08)012. [23] Prometheus 2016 [updated 2016 Dec 1; cited 2016 Dec.1]. Available from: https://prometheus.io/ [24] Draios, Sysdig 2016 [updated 2016 Dec 1; cited 2016 Dec.1]. Available from: http://www.sysdig.org/ [25] VirusShare 2016 [updated 2016 Dec 1; cited 2016 Dec.1]. Available from: https://virusshare.com/about.4n6	2cs.AI
Can Machines Think in Radio Language? arXiv:1710.02648v3 [cs.AI] 15 December 2017 Yujian Li College of Computer Science, Beijing University of Technology, Beijing 100124, China Email: liyujian@bjut.edu.cn Abstract: People can think in auditory, visual and tactile forms of language, so can machines principally. But is it possible for them to think in radio language? According to a first principle presented for general intelligence, i.e. the principle of language's relativity, the answer may give an exceptional solution for robot astronauts to talk with each other in space exploration. Keywords: The principle of language’s relativity; First principle; Radio language; Space exploration. Figure 1. Two robonauts are talking about the earth in a Morse-code radio language on the moon. One says “THE EARTH IS OUR HOME”, the other “IT IS TRULY BEAUTIFUL”. This is a potential application of Morse code to artificial intelligence in the future, beyond classical transmission of text information between people. In computer science, one of the biggest unsolved problems is to develop intelligent machines. Since a seminal paper by Turing on the topic of artificial intelligence [1], the central question “Can machines think?” began to excite interest in building systems that learn and think like people. That is a fascinating dream! Recently, the interest has been renewed again because of impressive progress with deep learning [2], in spite of great difficulties such as the Character Challenge and the Frostbite Challenge [3], to perform a variety of tasks as rapidly and flexibly as people do. What does it mean for a system to learn and think like a person? Lake et al. argued that this system should build causal models of the world, ground learning in intuitive theories of physics and psychology, and harness compositionality and learning-to-learn [3]. They claimed that these key ideas of core ingredients would play an active and important role in producing human-like learning and thought. Undoubtedly, their claim is attractive and promising for the ultimate dream of implementing machines with human-level general intelligence. However, the claim says little about a person’s ability to communicate and think in natural language, which is clearly vital for human intelligence [4]. The question is, how to develop a capacity of language for machines? Language is a basic tool in human society, playing an essential role in communication and thought. People are accustomed to thinking in sound language (sound thinking). In different countries, people generally speak different languages. There are about 50007000 languages spoken all over the world, 90% of them used by less than 100000 people. As estimated by UNESCO (The United Nations’ Educational, Scientific and Cultural Organization), the most widely spoken languages are: Mandarin Chinese, English, Spanish, Hindi, Arabic, Bengali, Russian, Portuguese, Japanese, German and French. In practice, a language usually takes forms of speech and text, but can be encoded into whistle, sign or braille. This may lead to an interesting question, can machines think in language with other forms, e.g. radio? To all appearance, a Chinese can think in Chinese, an American can think in English, a Spanish can think in Spanish, and so on. From the viewpoint of daily life, all these forms of language, even including other forms such as whistle, sign and braille, should be equivalent for people to think about the world. In my opinion, this quite common point can be generalized as a first principle to establish a theory of mind (and intelligence more broadly), termed the principle of language’s relativity, or “the principle of symbolic relativity” [5]. The principle is described as follows. All admissible forms of language are equivalent for an intelligent system to think about the world. In the principle of language’s relativity, an admissible form means that the system can use it for thinking, i.e. the formulation of thoughts about the world. Therefore, the principle can be stated in other words, All admissible forms of language are equivalent with respect to the formulation of thoughts about the world. Note that this principle is named with inspiration from the principle of relativity in physics [6], namely, All admissible frames of reference are equivalent with respect to the formulation of the fundamental laws of physics. That is, physic laws are the same in all reference frames - inertial or non-inertial. By analogy, it can be stated that, thoughts about the world are the same in all language forms – speech, text, whistle, sign or braille. Therefore, in this sense a language form can also be regarded as a reference frame to formulate thoughts. If taken as the first postulate of intelligence theory, the principle of language’s relativity implies that language is independent of modality. This explains why any human language can be encoded into a lot of different media such as using auditory, visual, and tactile stimuli. Moreover, it can give profound and original insights to guide engineering future generations of intelligent machines. For example, principally robots can think in radio language. Such robots would be tremendously useful in space exploration, where radio language is much more convenient for them to talk with each other than sound language for lack of air. Since no person have an inborn ability to receive and send radio waves, the radio form of language is not admissible for human. Thus, radio language is a novel and creative idea for robots to think, although radio itself is certainly very ordinary for information transmission and remote control. Clearly, thinking in radio language (radio thinking), is a different way to implement intelligence than people can. One may argue that, even without language, artificial intelligence (e.g. by residual network [7], deep Q-network [8] and AlphoGo [9]) could equal or even beat human intelligence in deep learning performance of some tasks such as object recognition, video games and board games. In practical realization, it is still reasonable to require that an intelligent machine is able to communicate through language. It goes without saying, language is an essential ability for general intelligence. Yet nobody is quite sure of what intelligence is. Perhaps in most general purpose, intelligence measures an agent’s ability to achieve goals in a wide range of environments [10]. Nonetheless, this informal definition, together with the mathematic description [10], plays a limited role in design of intelligent machines, albeit bringing together some key features from many expert definitions of human intelligence. To understand the nature of intelligence, not only a far-reaching definition is further expected, but also a comprehensive theory. What does this theory look like? At the very least, it should contain just a few first principles at the system level. The few principles must be fundamental and independent in all phenomena of intelligence, and cannot be deduced from any other principles in physics, chemistry and biology. Although these principles may not lead to anything like Maxwell equations or E=mc2, they should capture the essential characteristics of intelligence comprehensively in perspectives between science and philosophy. More importantly, they should be able to make a guide to engineering intelligent machines, especially with some different intelligence from human. “Human-like” implicates imitation before grasping the nature of intelligence clearly, while “human-different” implicates creation after understanding genuinely. One of such principles is the principle of language’s relativity, others to be certain. Obviously, the principle is independent of physics, chemistry and biology. Furthermore, it can account for modalityindependence of language, and give rise to a revolutionary idea of radio thinking. As a high-level intelligence envisaged for future robot astronauts, the importance of radio thinking should be emphasized again. It may overturn a public view of how robonauts talk with each other in space exploration. Imagine two robonauts are talking about the earth on the moon (see Figure 1). Traditionally, people think that they would talk in sound language, just as could be seen in some science fiction films. Nevertheless, in reality they cannot do so at all without air. Alternatively, they can talk with each other in a Morse-code radio language that lacks any neural mechanisms. This is an excellent thought experiment to show that, although human language is an ability developed on the basis of neural mechanisms in the brain, intelligent machines may have a capacity of radio language not based on these mechanisms. Thereby, high-level intelligence may not be brain-like. The brain-like intelligence tries to achieve intelligence as demonstrated by brains [11], preferably those of highly evolved creatures. But the nature of intelligence can be understood in a brain-different way, just like the secret of how to fly. Indeed, without flapping its wings, an airplane can fly in a bird-different way based on the theory of aerodynamics, not the bird’s brain. For space exploration, autonomous robonauts would be extremely helpful on the moon or the other planets. Since the environments change, there are a lot of barriers to implement these robonauts. But they should require certain kinds of human-different intelligence for high-level autonomy in actions. For example, radio thinking helps them to plan, and radio talking helps them to collaborate. Kirobo is the world’s first talking robot sent into space [12]. However, Kirobo was tasked to be a companion, not an explorer. It could talk in sound language only inside the spacecraft. Radio talking may be an exceptional solution outside. It is a terrific endeavor to engineer machines with general intelligence. A theory of intelligence requires a few first principles, for example, the principle of language’s relativity. Where are other principles in the Universe? What are their implications in science? How do they play roles in technology? These questions are the challenges set ahead for remarkable years of discovery. References [1] Turing, A. M. Computing Machinery and Intelligence. Mind 49, 433-460 (1950). [2] LeCun, Y., Bengio, Y. & Hinton, G. E. Deep learning. Nature 521, 436-444 (2015). [3] Lake, B. M., Ullman, T. D., Tenenbaum, J. B., Gershaman, S. J. Building machines that learn and think like people. Behavioral and Brain Sciences, accepted for publication, https://doi.org/10.1017/S0140525X16001837. [4] Mikolov, T., Joulin, A., & Baroni, M. A Roadmap towards Machine Intelligence. arXiv preprint: 1511.08130 (2016). [5] Yujian Li. Reveal the secrets of consciousness  also on theory of cognitive relativity. In 100 Interdisciplinary Science Puzzles of the 21st Century. Science Press, Beijing, 2005, edited by Xixian Li et al. [6] https://en.wikipedia.org/wiki/Principle_of_relativity. [7] He, K., Zhang, X., Ren, S., et al. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2016), 770-778. [8] Mnih, V., Kavukcuoglu, K., Silver D., et al. Human-level control through deep reinforcement learning. Nature 518, 529-533 (2015). [9] Silver, D., Huang, A., Maddison, C. J., et al. Mastering the game of Go with deep neural networks and tree search. Nature 529, 484-489 (2016). [10] Legg, S., Hutter, M. Universal Intelligence: A Definition of Machine Intelligence. Minds & Machines 17, 391-444 (2007). [11] Sendhoff, B., Körner, E., Sporns, O. Creating Brain-like Intelligence. LANAI 5436, 1-14 (2009). [12] http://www.telegraph.co.uk/news/science/space/10221399/Talking-robot-astronaut-blasts-into-space.html.	2cs.AI
1 Interpreting Deep Visual Representations via Network Dissection arXiv:1711.05611v1 [cs.CV] 15 Nov 2017 Bolei Zhou∗ , David Bau∗ , Aude Oliva, and Antonio Torralba Abstract—The success of recent deep convolutional neural networks (CNNs) depends on learning hidden representations that can summarize the important factors of variation behind the data. However, CNNs often criticized as being black boxes that lack interpretability, since they have millions of unexplained model parameters. In this work, we describe Network Dissection, a method that interprets networks by providing labels for the units of their deep visual representations. The proposed method quantifies the interpretability of CNN representations by evaluating the alignment between individual hidden units and a set of visual semantic concepts. By identifying the best alignments, units are given human interpretable labels across a range of objects, parts, scenes, textures, materials, and colors. The method reveals that deep representations are more transparent and interpretable than expected: we find that representations are significantly more interpretable than they would be under a random equivalently powerful basis. We apply the method to interpret and compare the latent representations of various network architectures trained to solve different supervised and self-supervised training tasks. We then examine factors affecting the network interpretability such as the number of the training iterations, regularizations, different initializations, and the network depth and width. Finally we show that the interpreted units can be used to provide explicit explanations of a prediction given by a CNN for an image. Our results highlight that interpretability is an important property of deep neural networks that provides new insights into their hierarchical structure. Index Terms—Convolutional Neural Networks, Network Interpretability, Visual Recognition, Interpretable Machine Learning. F 1 I NTRODUCTION BSERVATIONS of hidden units in large deep neural networks have revealed that human-interpretable concepts sometimes emerge as individual latent variables within those networks. For example, object detector units emerge within networks trained to recognize places [1]; part detectors emerge in object classifiers [2]; and object detectors emerge in generative video networks [3]. This internal structure has appeared in situations where the networks are not constrained to decompose problems in any interpretable way. The emergence of interpretable structure suggests that deep networks may be learning disentangled representations spontaneously. While it is commonly understood that a network can learn an efficient encoding that makes economical use of hidden variables to distinguish the input, the appearance of a disentangled representation is not well understood. A disentangled representation aligns its variables with a meaningful factorization of the underlying problem structure, and encouraging disentangled representations is a significant area of research [4]. If the internal representation of a deep network is partly disentangled, one possible path for understanding its mechanisms is to detect disentangled structure, and simply read out the human interpretable factors. We address the following three key issues about the deep visual representations in this work: • What is a disentangled representation of neural networks, and how can its factors be quantified and detected? • Do interpretable hidden units reflect a special alignment of feature space, or are interpretations a chimera? • What differences in network architectures, data sources, and training conditions lead to the internal representations with greater or lesser entanglement? O • • B. Zhou and D.Bau contributed equally to this work. B. Zhou, D. Bau, A.Oliva, and A. Torralba are with CSAIL, MIT, MA, 02139. E-mail: {bzhou, davidbau, oliva, torralba}@csail.mit.edu To examine these issues, we propose a general analytic framework, Network Dissection, for interpreting deep visual representations and quantifying their interpretability. Using Broden, a broadly and densely labeled dataset, our framework identifies hidden units’ semantics for any given CNN, then aligns them with human-interpretable concepts. Building upon the preliminary result published at [5], we begin with a detailed description of the methodology of Network Dissection. We use the method to interpret a variety of deep visual representations trained with different network architectures (AlexNet, VGG, GoogLeNet, ResNet, DenseNet) and supervisions (supervsied training on ImageNet for object recognition and on Places for scene recognition, along with various self-taught supervision tasks). We show that interpretability is an axis-aligned property of a representation that can be destroyed by rotation without affecting discriminative power. We further examine how interpretability is affected by different training datasets, training regularizations such as dropout [6] and batch normalization [7], and fine-tuning between different data sources. Our experiments reveal that units emerge as semantic detectors in the intermediate layers of most deep visual representations, while the degree of interpretability can vary widely across changes in architecture and training. We conclude that representations learned by deep networks are more interpretable than previously thought, and that measurements of interpretability provide insights about the structure of deep visual representations that that are not revealed by their classification power alone1 . 1.1 Related Work Visualizing deep visual representations. Though CNN models are notoriously known as black boxes, a growing number of 1. Code, data, and more dissection results are available at the project page http://netdissect.csail.mit.edu/. 2 techniques have been developed to visualize the internal representations of convolutional neural networks. The behavior of a CNN can be visualized by sampling image patches that maximize activation of hidden units [1], [8], [9], or by using variants of backpropagation to identify or generate salient image features [8], [10], [11]. Back-propagation together with a natural image prior can be used to invert a CNN layer activation [12], and an image generation network can be trained to invert the deep features by synthesizing the input images [13]. [14] further synthesizes the prototypical images for individual units by learning a feature code for the image generation network from [13].These visualizations reveal the image patterns that have been learned in a deep visual representation and provide a qualitative guide to the interpretation and interpretability of units. In [1], a quantitative measure of interpretability was introduced: human evaluation of visualizations to determine which individual units behave as object detectors in a network trained to classify scenes. However, human evaluation is not scalable to increasingly large networks such as ResNet [15], with more than 100 layers. Therefore the aim of the present work is to develop a scalable method to go from qualitative visualization to quantitative interpretation. Analyzing the properties of deep visual representations. Various intrinsic properties of deep visual representations have been explored. Much research has focused on studying the power of CNN layer activations to be used as generic visual features for classification [16], [17]. The transferability of activations for a variety of layers has been analyzed, and it has been found that higher layer units are more specialized to the target task [18]. Susceptibility to adversarial input reveals that discriminative CNN models are fooled by particular image patterns [19], [20]. Analysis of correlation between different random initialized networks reveal that many units converge to the same set of representations after training [21]. The question of how representations generalize has been investigated by showing that a CNN can easily fit a random labeling of training data even under explicit regularization [22]. Our work focuses on another less explored property of deep visual representations: interpretability. Unsupervised learning of deep visual representations. Recent work on unsupervised learning or self-supervised learning exploits the correspondence structure that comes for free from unlabeled images to train networks from scratch [23], [24], [25], [26], [27]. For example, CNN is trained by predicting image context [23], by colorizing gray images [28], [29], by solving image puzzle [24], and by associating the images with ambient sounds [30]. The resulting deep visual representations learned from different unsupervised learning tasks are compared by evaluating them as generic visual features on classification datasets such as Pascal VOC. Our work provides an alternative approach to compare deep visual representations in terms of their interpretability, beyond just their discriminative power. 2 F RAMEWORK OF N ETWORK D ISSECTION The notion of a disentangled representation rests on the human perception of what it means for a concept to be mixed up. Therefore we define the interpretability of deep visual representation in terms of the degree of alignment with a set of human-interpretable concepts. Our quantitative measurement of interpretability for deep visual representations proceeds in three steps: 1) Identify a broad set of human-labeled visual concepts. TABLE 1 Statistics of each label type included in the dataset. Category scene object part material texture color Classes 468 584 234 32 47 11 Sources ADE [32] ADE [32], Pascal-Context [34] ADE [32], Pascal-Part [35] OpenSurfaces [33] DTD [36] Generated Avg sample 38 491 854 1,703 140 59,250 2) Gather the response of the hidden variables to known concepts. 3) Quantify alignment of hidden variable−concept pairs. This three-step process of network dissection is reminiscent of the procedures used by neuroscientists to understand similar representation questions in biological neurons [31]. Since our purpose is to measure the level to which a representation is disentangled, we focus on quantifying the correspondence between a single latent variable and a visual concept. In a fully interpretable local coding such as a one-hot-encoding, each variable will match exactly with one human-interpretable concept. Although we expect a network to learn partially nonlocal representations in interior layers [4], and past experience shows that an emergent concept will often align with a combination of a several hidden units [2], [17], our present aim is to assess how well a representation is disentangled. Therefore we measure the alignment between single units and single interpretable concepts. This does not gauge the discriminative power of the representation; rather it quantifies its disentangled interpretability. As we will show in Sec. 3.2, it is possible for two representations of perfectly equivalent discriminative power to have very different levels of interpretability. To assess the interpretability of any given CNN, we draw concepts from a new broadly and densely labeled image dataset that unifies labeled visual concepts from a heterogeneous collection of labeled data sources, described in Sec. 2.1. We then measure the alignment of each hidden unit of the CNN with each concept by evaluating the feature activation of each individual unit as a segmentation model for each concept. To quantify the interpretability of a layer as a whole, we count the number of distinct visual concepts that are aligned with a unit in the layer, as detailed in Sec. 2.2. 2.1 Broden: Broadly and Densely Labeled Dataset To be able to ascertain alignment with both low-level concepts such as colors and higher-level concepts such as objects, we have assembled a new heterogeneous dataset. The Broadly and Densely Labeled Dataset (Broden) unifies several densely labeled image datasets: ADE [32], OpenSurfaces [33], Pascal-Context [34], Pascal-Part [35], and the Describable Textures Dataset [36]. These datasets contain examples of a broad range of objects, scenes, object parts, textures, and materials in a variety of contexts. Most examples are segmented down to the pixel level except textures and scenes, which are given for full images. In addition, every image pixel in the dataset is annotated with one of the eleven common color names according to the human perceptions classified by van de Weijer [37]. Samples of the types of labels in the Broden dataset are shown in Fig. 1. The purpose of Broden is to provide a ground truth set of exemplars for a broad set of visual concepts. The concept labels in Broden are normalized and merged from their original datasets so 3 red (color) yellow (color) wrinkled (texture) meshed (texture) wood (material) fabric (material) foot (part) door (part) airplane (object) waterfall (object) art studio (scene) beach (scene) Fig. 1. Samples from the Broden Dataset. The ground truth for each concept is a pixel-wise dense annotation. Top activated images Segmented images using the binarized unit activation map Semantic segmentation annotations Segmented annotations Fig. 2. Scoring unit interpretability by evaluating the unit for semantic segmentation. that every class corresponds to an English word. Labels are merged based on shared synonyms, disregarding positional distinctions such as ‘left’ and ‘top’ and avoiding a blacklist of 29 overly general synonyms (such as ‘machine’ for ‘car’). Multiple Broden labels can apply to the same pixel: for example, a black pixel that has the Pascal-Part label ‘left front cat leg’ has three labels in Broden: a unified ‘cat’ label representing cats across datasets; a similar unified ‘leg’ label; and the color label ‘black’. Only labels with at least 10 image samples are included. Table 1 shows the number of classes per dataset and the average number of image samples per label class. Totally there are 1197 visual concept classes included. 2.2 Scoring Unit Interpretability The proposed network dissection method evaluates every individual convolutional unit in a CNN as a solution to a binary segmentation task to every visual concept in Broden, as illustrated in Fig. 3. Our method can be applied to any CNN using a forward pass without the need for training or backpropagation. For every input image x in the Broden dataset, the activation map Ak (x) of every internal convolutional unit k is collected. Then the distribution of individual unit activations ak is computed. For each unit k , the top quantile level Tk is determined such that P (ak > Tk ) = 0.005 over every spatial location of the activation map in the dataset. To compare a low-resolution unit’s activation map to the inputresolution annotation mask Lc for some concept c, the activation map is scaled up to the mask resolution Sk (x) from Ak (x) using bilinear interpolation, anchoring interpolants at the center of each unit’s receptive field. Sk (x) is then thresholded into a binary segmentation: Mk (x) ≡ Sk (x) ≥ Tk , selecting all regions for which the activation exceeds the threshold Tk . These segmentations are evaluated against every concept c in the dataset by computing intersections Mk (x)∩Lc (x), for every (k, c) pair. The score of each unit k as segmentation for concept c is reported as a the intersection over union score across all the images in the dataset, P \|Mk (x) ∩ Lc (x)\| IoUk,c = P , (1) \|Mk (x) ∪ Lc (x)\| where \| · \| is the cardinality of a set. Because the dataset contains some types of labels which are not present on some subsets of inputs, the sums are computed only on the subset of images that have at least one labeled concept of the same category as c. The value of IoUk,c is the accuracy of unit k in detecting concept c; we consider one unit k as a detector for concept c if IoUk,c exceeds a threshold. Our qualitative results are insensitive to the IoU threshold: different thresholds denote different numbers of units as concept detectors across all the networks but relative orderings remain stable. For our comparisons we report a detector if IoUk,c > 0.04. Note that one unit might be the detector for multiple concepts; for the purpose of our analysis, we choose the top ranked label. To quantify the interpretability of a layer, we count the number unique concepts aligned with units. We call this the number of unique detectors. Figure 2 summarizes the whole process of scoring unit interpretability: By segmenting the annotation mask using the receptive field of units for the top activated images, we compute the IoU for each concept. The IoU evaluating the quality of the segmentation of a unit is an objective confidence score for interpretability that is comparable across networks. Thus this score enables us to compare interpretability of different representations and lays the basis for the experiments below. Note that network dissection works only as well as the underlying dataset: if a unit matches a human-understandable concept that is absent in Broden, then it will not score well for interpretability. Future versions of Broden will be expanded to include more kinds of visual concepts. 3 INTERPRETING DEEP VISUAL REPRESENTATIONS For testing we prepare a collection of CNN models with different network architectures and supervision of primary tasks, as listed in Table 2. The network architectures include AlexNet [38], GoogLeNet [39], VGG [40], ResNet [15], and DenseNet [41]. For supervised training, the models are trained from scratch (i.e., not pretrained) on ImageNet [42], Places205 [43], and Places365 [44]. ImageNet is an object-centric dataset, which contains 1.2 million images from 1000 classes. Places205 and Places365 are 4 Freeze trained network weights Upsample target layer Car Person Grass Door t Uni e n O tion va Acti Fabric Pixel-wise segmentation Blue Conv Conv Conv Conv Network being probed Conv Input image Evaluate on segmentation tasks Fig. 3. Illustration of network dissection for measuring semantic alignment of units in a given CNN. Here one unit of the last convolutional layer of a given CNN is probed by evaluating its performance on various segmentation tasks. Our method can probe any convolutional layer. TABLE 2 Tested CNN Models Training none Supervised Self Network AlexNet AlexNet GoogLeNet VGG-16 ResNet-152 DenseNet-161 AlexNet dataset or task random ImageNet, Places205, Places365, Hybrid. ImageNet, Places205, Places365. ImageNet, Places205, Places365, Hybrid. ImageNet, Places365. ImageNet, Places365. context, puzzle, egomotion, tracking, moving, videoorder, audio, crosschannel,colorization. objectcentric, transinv. two subsets of the Places Database, which is a scene-centric dataset with categories such as kitchen, living room, and coast. Places205 contains 2.4 million images from 205 scene categories, while Places365 contains 1.6 million images from 365 scene categories. “Hybrid” refers to a combination of ImageNet and Places365. For self-supervised training tasks, we select several recent models trained on predicting context (context) [23], solving puzzles (puzzle) [24], predicting ego-motion (egomotion) [25], learning by moving (moving) [26], predicting video frame order (videoorder) [45] or tracking (tracking) [27], detecting object-centric alignment (objectcentric) [46], colorizing images (colorization) [28], inpainting (contextencoder) [47], predicting cross-channel (crosschannel) [29], predicting ambient sound from frames (audio) [30], and tracking invariant patterns in videos (transinv) [48]. The self-supervised models we analyze are comparable to each other in that they all use AlexNet or an AlexNetderived architecture, with one exception model transinv [48], which uses VGG as the base network. In the following experiments, we begin by validating our method using human evaluation. Then, we use random unitary rotations of a learned representation to test whether interpretability of CNNs is an axis-independent property; we find that it is not, and we conclude that interpretability is not an inevitable result of the discriminative power of a representation. Next, we analyze all the convolutional layers of AlexNet as trained on ImageNet [38] and as trained on Places [43], and confirm that our method reveals detectors for higher-level concepts at higher layers and lower-level concepts at lower layers; and that more detectors for higher-level concepts emerge under scene training. Then, we show that different network architectures such as AlexNet, VGG, and ResNet yield different interpretability, while differently supervised training tasks and self-supervised training tasks also yield a variety of levels of interpretability. Additionally we show the impact of different training conditions, examine the relationship between discriminative power and interpretability, and investigate a possible way to improve the interpretability of CNNs by increasing their width. Finally we utilize the interpretable units as explanatory factors to the prediction given by a CNN. 3.1 Human Evaluation of Interpretations Using network dissection, we analyze the interpretability of units within all the convolutional layers of Places-AlexNet and ImageNetAlexNet, then compare with human interpretation. Places-AlexNet is trained for scene classification on Places205 [43], while ImageNet-AlexNet is the identical architecture trained for object classification on ImageNet [38]. Our evaluation was done by raters on Amazon Mechanical Turk (AMT). As a baseline description of unit semantics, we used human-written descriptions of each unit from [1]. These descriptions were collected by asking raters to write words or short phrases to describe the common meaning or pattern selected by each unit, based on a visualization of the top image patches. Three descriptions and a confidence were collected for each unit. As a canonical description we chose the most common description of a unit (when raters agreed), and the highest-confidence description (when raters did not agree). Some units may not be interpretable. To identify these, raters were shown the canonical descriptions of visualizations and asked whether they were descriptive. Units with validated descriptions are taken as the set of interpretable units. To compare these baseline descriptions to network-dissectionderived labels, we ran the following experiment. Raters were shown a visualization of top images patches for an interpretable unit, along with a word or short phrase description, and they were asked to vote (yes/no) whether the given phrase was descriptive of most of the image patches. The baseline human-written descriptions were randomized with the labels derived using net dissection, and the origin of the labels was not revealed to the raters. Table 3 summarizes the results. The number of interpretable units is shown for each layer, and average positive votes for descriptions of interpretable units are shown, both for humanwritten labels and network-dissection-derived labels. Human labels are most highly consistent for units of conv5, suggesting that humans have no trouble identifying high-level visual concept detectors, while lower-level detectors are more difficult to label. 5 Fig. 4. The annotation interface used by human raters on Amazon Mechanical Turk. Raters are shown descriptive text in quotes together with fifteen images, each with highlighted patches, and must evaluate whether the quoted text is a good description for the highlighted patches. TABLE 3 Human evaluation of our Network Dissection approach. Interpretable units Human consistency Network Dissection conv1 57/96 82% 37% conv2 126/256 76% 56% conv3 247/384 83% 54% conv4 258/384 82% 59% conv5 194/256 91% 71% Similarly, labels given by network dissection are best at conv5, and are found to be less descriptive for lower layers. Comparison of the human interpretation and the labels predicted by network dissection is plotted in Fig. 5. A sample of units is shown together with both automatically inferred interpretations and manually assigned interpretations taken from [1]. We can see that the predicted labels match the human annotation well, though sometimes they capture a different description of a visual concept, such as the ‘crosswalk’ predicted by the algorithm compared to ‘horizontal lines’ given by the human for the third unit in conv4 of Places-AlexNet in Fig. 5. Confirming intuition, color and texture concepts dominate at lower layers conv1 and conv2 while more object and part detectors emerge in conv5. 3.2 Measurement of Axis-aligned Interpretability We conduct an experiment to determine whether it is meaningful to assign an interpretable concept to an individual unit. Two possible hypotheses can explain the emergence of interpretability in individual hidden layer units: Hypothesis 1. Interpretability is a property of the representation as a whole, and individual interpretable units emerge because interpretability is a generic property of typical directions of representation space. Under this hypothesis, projecting to any direction would typically reveal an interpretable concept, and interpretations of single units in the natural basis would not be more meaningful than interpretations that can be found in any other direction. Hypothesis 2. Interpretable alignments are unusual, and interpretable units emerge because learning converges to a special basis that aligns explanatory factors with individual units. In this model, the natural basis represents a meaningful decomposition learned by the network. Hypothesis 1 is the default assumption: in the past it has been found [19] that with respect to interpretability “there is no distinction between individual high level units and random linear combinations of high level units.” Network dissection allows us to re-evaluate this hypothesis. We apply random changes in basis to a representation learned by AlexNet. Under hypothesis 1, the overall level of interpretability should not be affected by a change in basis, even as rotations cause the specific set of represented concepts to change. Under hypothesis 2, the overall level of interpretability is expected to drop under a change in basis. We begin with the representation of the 256 convolutional units of AlexNet conv5 trained on Places205 and examine the effect of a change in basis. To avoid any issues of conditioning or degeneracy, we change basis using a random orthogonal transformation Q. The rotation Q is drawn uniformly from SO(256) by applying 2 Gram-Schmidt on a normally-distributed QR = A ∈ R256 with positive-diagonal right-triangular R, as described by [49]. Interpretability is summarized as the number of unique visual concepts aligned with units, as defined in Sec. 2.2. Denoting AlexNet conv5 as f (x), we find that the number of unique detectors in Qf (x) is 80% fewer than the number of unique detectors in f (x). Our finding is inconsistent with hypothesis 1 and consistent with hypothesis 2. We also test smaller perturbations of basis using Qα for 0 ≤ α ≤ 1, where the fractional powers Qα ∈ SO(256) are chosen to form a minimal geodesic gradually rotating from I to Q; these intermediate rotations are computed using a Schur decomposition. Fig. 6 shows that interpretability of Qα f (x) decreases as larger rotations are applied. Fig. 7 shows some examples of the linearly combined units. Each rotated representation has exactly the same discriminative power as the original layer. Writing the original network as g(f (x)), note that g 0 (r) ≡ g(QT r) defines a neural network that processes the rotated representation r = Qf (x) exactly as the original g operates on f (x). We conclude that interpretability is neither an inevitable result of discriminative power, nor is it a prerequisite to discriminative power. Instead, we find that interpretability is a different quality that must be measured separately to be understood. We repeat the complete rotation (α = 1) on Places365 and ImageNet 10 times, the result is shown in Fig. 8. We observe the drop of interpretability for both of the network, while it drops more for the AlexNet on Places365. It is because originally the interpretability of AlexNet on Places365 is higher than AlexNet on ImageNet thus the random rotation damages more. 3.3 Network Architectures with Supervised Learning How do different network architectures affect disentangled interpretability of the learned representations? We apply network dissection to evaluate a range of network architectures trained on ImageNet and Places. For simplicity, the following experiments focus on the last convolutional layer of each CNN, where semantic detectors emerge most. Results showing the number of unique detectors that emerge from various network architectures trained on ImageNet and Places are plotted in Fig. 9. In terms of network architecture, we find that interpretability of ResNet > DenseNet > VGG > 6 h:green sky (object) Places veined (texture) orange (color) h:color yellow lacelike (texture) ImageNet red (color) h:grid pattern perforated (texture) h:yellow food (material) h:red woven (texture) h:yellow banded (texture) yellow (color) h:striped sky (object) h:blue grid (texture) sky (object) conv1 h:corrugated tree (object) h:black&white banded (texture) h:pink or red lined (texture) red (color) h:grass chequered (texture) h:sky grass (object) h:bed h:tree car (object) h:car h:pattern crosswalk (part) h:horiz. lines mountain (scene) h:orange muzzle (part) h:animal face wheel (part) h:blue sky swirly (texture) h:round cat (object) h:nosed head (part) h:face leg (part) h:mesh dotted (texture) conv2 h:windows bed (object) conv3 h:montain h:wheels h:animal faces h:leg conv5 conv4 Fig. 5. Comparison of the interpretability of all five convolutional layers of AlexNet, as trained on classification tasks for Places (top) and ImageNet (bottom).Four examples of units in each layer are shown with identified semantics. The segmentation generated by each unit is shown on the three Broden images with highest activation. Top-scoring labels are shown above to the left, and human-annotated labels are shown above to the right. Some disagreement can be seen for the dominant judgment of meaning. For example, human annotators mark the first conv4 unit on Places as a ‘windows’ detector, while the algorithm matches the ‘chequered’ texture. Rotation of Representation 20 30 20 10 0 10 rotate 0.4 rotate 0.6 rotate 0.8 rotate 1 Fig. 6. Interpretability over changes in basis of the representation of AlexNet conv5 trained on Places. The vertical axis shows the number of unique interpretable concepts that match a unit in the representation. The horizontal axis shows α, which quantifies the degree of rotation. Baseline (individual units) car (single unit 87) skyscraper (single unit 94) Rotate 1 (linear combinations) IoU 0.16 car (combination, closest to unit 173) IoU 0.16 skyscraper (combination, closest to unit 94) IoU 0.06 IoU 0.05 tree (single unit 228) IoU 0.10 tree (combination, closest to unit 228) IoU 0.02 head (single unit 3) IoU 0.09 head (combination, closest to unit 70) IoU 0.02 closet (single unit 107) IoU 0.06 closet (combination, closest to unit 34) IoU 0.02 random combination AlexNet on ImageNet object scene part material texture color 20 15 10 5 0 baseline random combination Fig. 8. Complete rotation (α = 1) repeated on AlexNet trained on Places365 and ImageNet respectively. Rotation reduces the interpretability significantly for both of the networks. 350 Number of unique detectors rotate 0.2 baseline 25 300 250 200 object scene part material texture color 150 100 50 0 R es N en et se 15 N 2-P e R t16 lac es N 1- es D en et Pla 36 se 15 ce 5 N 2-I s3 et 16 ma 65 1 g VG -Im eNe G ag t -P eN la e VG ces t 20 G V G oo G 5 g G H G LeN -Pla ybr id oo c e gL t- es G eN Pla 36 oo e ce 5 gL t-P s3 eN la 65 et ces VG Ima 205 g Al ex G-I eN N ma et et - ge A Pla Ne Al lex ces t ex Ne 3 N t- 65 Al et-P Hyb ex rid l a N e ce Al t-Im s20 ex a 5 N ge et -ra Net nd om 0 baseline object scene part material texture color Number of unique detectors 30 AlexNet on Places365 40 Number of unique detectors object scene part material texture color D Number of unique detectors 40 Fig. 9. Interpretability across different architectures trained on ImageNet and Places. Fig. 7. Visualizations of the best single-unit concept detectors of five concepts taken from individual units of AlexNet conv5 trained on Places (left), compared with the best linear-combination detectors of the same concepts taken from the same representation under a random rotation (right). For most concepts, both the IoU and the visualization of the top activating image patches confirm that individual units match concepts better than linear combinations. In other cases, (e.g. head detectors) visualization of a linear combination appears highly consistent, but the IoU reveals lower consistency when evaluated over the whole dataset. GoogLeNet > AlexNet. Deeper architectures usually appear to allow greater interpretability, though individual layer structure is different across architecture. Comparing training datasets, we find Places > ImageNet. As discussed in [1], one scene is composed of multiple objects, so it may be beneficial for more object detectors to emerge in CNNs trained to recognize scenes. AlexNet VGG-16 GoogLeNet ResNet-152 DenseNet-161 bi ca cy r cl e po ol ca ai tab t rp le la ho ne us m e ot bu or s bi ke be to d ile s t ho ky rs w e at e do r si g bo nk ttl e se ch a ai bo r o co k st w ov ro e a bo tr d ok ack bu cas i m ld e o in tv unt g m ai on n ito tra r sn in ow bi sh rd e la lf sh mp w e in g ep do ra w ss p ba an e pa thtu in b tin g s cu of sh a sk cu ion ys rta cr in a pe pe rs r o ta n ce ble ilin se g pl at a fe nt w nc si ash e de e st wa r ai lk rw a ﬁ y ﬂo eld w pl er at bo e a ch an tre t si de e gn li bo er ar d units Densenet (Imagenet) 26 13 Closet 57 objects do g ca t ca h ai ors r rp e la ne bu s bi rd b m t ed ot ra or in b bo ike tt to le i po wa let ol te ta r bl c e sh ow pe eep r b so m icy n ou cl nt e ho ain u gr se as ro s a pl d ba an th t tu b se bo a ok ch po a tte si ir n tv dpl k m an on t i si st tor w gn ov in b e do oa w rd pa n sh e sn elf ow ro c ﬁe k ta ld ce ble il ﬂo ing w e tre r e te n s t sk fe ofa ys n cr ce a w pe as r he bo r bo pl at ok at si ca e de se w be p alk dc illo lo w th e ea s ra rth ilin g cu sea sh t io n pa l sa m n co ben d m ch pu t w la er ar m dr p ob e 86 43 62 objects units Resnet152 (Imagenet) ca be r d ai bu rp s la w ne at po d er ol og ta bl e m ou ca t pa nta i in n ti gr ng as ro s a bo d o to k ce ile ilin t tra g in bi sk cy y c ho le sk h rse ys ou cr se ap po e tte so r dp fa la n si t n pl k an m bo t ot tt or le bi k bo e at pe tree rs on se a w tra at ck er fa la ll m p te n bi t rd se s at w tove as he si raili r de n w w st a g or ai lk k rw ch sur ay gr andface an e ds lier t cu and rta ch in ﬂ ai w bu ow r in i er do ld in tv wpa g m n on e i br tor id bo ge ok ba ca ll s ea e r sh th el ca f sw s iv sno e el w ch sh air ee pa p l pi m t p ch gr illow o co un un d t t er ca able bi ne ﬁe t si cra ld be gnb dle o dc a lo rd th fe es nc e co w units Densenet (Places) 36 18 73 objects la n tra e in ca bu r b s ho ed us e c m d at ou o nt g a ho in r w se at gr er a bo ss ttl e tre bo e at bu sof ild a in to g il pl et an m ot bi t o po rb rd ol ike ta bl e sk y te nt se po a b tte o dp ok la fe nt bi nce cy pa c in le tin r g pe oad rs cu on rta in si sn nk ow ta bl e w b she in at l do ht f ch wp ub an an e s de sk tai lier ys rw cr ay ap st er ov ro e c tra k ck w ﬁeld si ash de e w r a ch lk a co ir w or ﬂ ur k ow t su e w rfac r at e e br rfal id l be ge n sh ch ee tv rai p m lin on g i m tor irr bo p or ok alm ca se rp ai 37 19 61 objects units Resnet152 (Places) 7 0 0 0 0 Fig. 10. Histogram of the object detectors from the ResNet and DenseNet trained on ImageNet and Places respectively. layer161 unit 1639 IoU=0.225 layer161 unit 2035 IoU=0.199 layer161 unit 1126 IoU=0.076 layer161 unit 1492 IoU=0.282 layer161 unit 1518 IoU=0.205 layer161 unit 1788 IoU=0.201 layer161 unit 2028 IoU=0.113 layer161 unit 1356 IoU=0.067 layer161 unit 1519 IoU=0.155 layer161 unit 1512 IoU=0.125 res5c unit 2011 IoU=0.171 res5c unit 1573 IoU=0.217 res5c unit 264 IoU=0.125 res5c unit 674 IoU=0.265 res5c unit 1243 IoU=0.172 res5c unit 9 IoU=0.161 res5c unit 1718 IoU=0.195 res5c unit 766 IoU=0.094 res5c unit 74 IoU=0.256 res5c unit 963 IoU=0.156 inception_5b unit 758 IoU=0.159 inception_4e unit 750 IoU=0.203 inception_4e unit 56 IoU=0.139 inception_4e unit 824 IoU=0.168 inception_4e unit 92 IoU=0.164 inception_5b unit 235 IoU=0.136 inception_4e unit 225 IoU=0.152 inception_4e unit 714 IoU=0.105 inception_5b unit 603 IoU=0.154 inception_4e unit 759 IoU=0.144 conv5_3 unit 213 IoU=0.125 conv5_3 unit 142 IoU=0.205 conv5_3 unit 85 IoU=0.086 conv5_3 unit 191 IoU=0.153 conv5_3 unit 151 IoU=0.150 conv5_3 unit 107 IoU=0.065 conv5_3 unit 491 IoU=0.112 conv4_3 unit 336 IoU=0.068 conv5_3 unit 20 IoU=0.149 conv5_3 unit 204 IoU=0.077 conv5 unit 235 IoU=0.017 conv5 unit 180 IoU=0.090 conv5 unit 55 IoU=0.087 conv5 unit 10 IoU=0.040 conv5 unit 13 IoU=0.101 conv3 unit 255 IoU=0.015 conv5 unit 250 IoU=0.051 conv5 unit 16 IoU=0.062 conv5 unit 174 IoU=0.029 conv5 unit 28 IoU=0.049 Dog Plant Bus Airplane Fig. 11. Comparison of several visual concept detectors identified by network dissection in DenseNet, ResNet, GoogLeNet, VGG, and AlexNet. Each network is trained on Places365. The two highest-IoU matches among convolutional units of each network is shown. The segmentation generated by each unit is shown on the four maximally activating Broden images. Some units activate on concept generalizations, e.g., GoogLeNet 4e’s unit 225 on horses and dogs, and 759 on white ellipsoids and jets. 8 147 50 13 25 textures 1 color 39 textures 1 color 37 2 colors units airplane train car bus bed house cat dog mountain horse water grass bottle tree boat sofa building toilet plant bird motorbike pool table sky tent sea book pottedplant fence bicycle painting road person curtain sink snow table shelf bathtub windowpane chandelier stairway skyscraper stove rock track ﬁeld washer sidewalk chair court ﬂower work surface waterfall bridge bench sheep railing tvmonitor mirror palm bookcase highway closet nursery art gallery skyscraper corridor reception airport terminal cockpit shoe shop attic classroom lighthouse playground bakery-shop youth hostel dining room volleyball court-outdoor jacuzzi-indoor conference room sauna landing deck ofﬁce mountain snowy auditorium alley castle bowling alley laundromat gas station windmill 160 scenes game room subway interior operating room playroom sandbox staircase pantry escalator-indoor ﬁre escape forest-broadleaf parking garage-indoor bathroom carrousel gymnasium-indoor 14 parts badlands cubicle-ofﬁce cemetery butchers shop 4 materials rope bridge dinette-vehicle beauty salon amusement arcade forest-needleleaf street waiting room airplane cabin dolmen golf course bus interior corn ﬁeld ice skating rink-indoor beach poolroom-home elevator-door home theater greenhouse-indoor zen garden gazebo-exterior art studio park theater-indoor procenium martial arts gym coast ﬁeld-cultivated living room bullring slum balcony-interior television studio catwalk water tower ice cream parlor bank vault creek desert-sand 19 21 textures 139 10 textures 1 color Resnet152 (Places) 0 9 5 0 1 object 1 scene 8 parts 1 material creek desert-sand vineyard supermarket kindergarden classroom locker room warehouse-indoor building facade bridge aqueduct pasture river 29 objects ﬂorist shop-indoor waterfall-block car interior-backseat orchard boxing ring music studio crevasse ball pit amusement park archive galley cloister-indoor fountain hayﬁeld lobby topiary garden garage-indoor jewelry shop bar temple-east asia earth ﬁssure amphitheater home ofﬁce desert-vegetation batters box parlor rubble shopfront clothing store kasbah kitchen junkyard bookstore shopping mall-indoor museum-indoor cavern-indoor swimming pool-outdoor natural history museum greenhouse-outdoor dam barn doorway-outdoor apartment building-outdoor medina ﬁeld-wild hot spring mountain campsite ballroom market-outdoor arrival gate-outdoor driveway inn-indoor excavation window seat construction site library-indoor clean room utility room hangar-indoor artists loft basketball court-outdoor football ﬁeld roundabout screen crosswalk body headboard wheel head hair roof shop window coach back pillow balcony torso monitor food tile carpet paper dotted 83 scenes cracked stratiﬁed lined cobwebbed potholed 5 parts chequered 1 material spiralled ﬂecked interlaced zigzagged freckled polka-dotted veined frilly banded striped perforated grooved ﬁbrous grid lacelike scaly porous meshed swirly pleated gauzy 21 textures crystalline crosshatched matted paisley studded pitted bubbly sprinkled bumpy marbled braided red Resnet152 continued ... units 17 objects 1 scene 44 14 Alexnet (Egomotion) 8 8 textures dog cat grass tree bicycle sea sky water road car painting windowpane mountain motorbike book sidewalk bus mountain snowy wheel hair head ear muzzle arm leg screen food zigzagged dotted chequered banded cobwebbed wafﬂed perforated striped polka-dotted frilly spiralled studded honeycombed grid meshed sprinkled veined porous cracked interlaced crosshatched red yellow 4 4 objects 1 scene 1 part 92 units 0 6 0 sky ceiling grass tree forest-broadleaf head banded lined dotted perforated grid chequered crosshatched spiralled 12 units bus airplane bed train pool table dog car horse house road boat motorbike tree grass water seat cat shelf ground sea skyscraper stove fence work surface toilet bottle sofa sink building cockpit conference room street closet poolroom-home classroom nursery ball pit highway corridor airport terminal attic bathroom castle art gallery alley dining room ofﬁce kitchen bar mountain snowy living room staircase shoe shop skyscraper lighthouse water tower gazebo-exterior subway interior bowling alley 60 objects bakery-shop waiting room cathedral-indoor landing deck laundromat auditorium gas station crevasse ice skating rink-indoor windmill bus interior pantry greenhouse-indoor amphitheater supermarket casino-indoor utility room cloister-indoor building facade bookstore reception viaduct car interior-backseat rope bridge cubicle-ofﬁce airplane cabin warehouse-indoor gymnasium-indoor bridge park sauna escalator-outdoor badlands pagoda waterfall-fan playroom galley waterfall-block lake-natural apartment building-outdoor jacuzzi-indoor campsite library-indoor kindergarden classroom artists loft bullring carrousel 47 scenes junkyard bow window-outdoor creek hayﬁeld ski slope mountain crosswalk screen head shop window coach food 12 parts potholed paisley 2 materials striped cobwebbed lined sprinkled interlaced studded banded dotted frilly stratiﬁed spiralled woven grooved ﬁbrous meshed crystalline bumpy scaly braided 22 21 Alexnet (Imagenet) 0 72 13 textures Alexnet (Puzzle Solving) sky ball pit veined chequered meshed striped dotted perforated studded lacelike frilly cobwebbed red units 8 2 parts units airplane bus car road grass plant windowpane person bed train motorbike horse tree house water pool table table toilet boat mountain dog painting book curtain ﬂower building grandstand bridge railing swimming pool tent signboard chair work surface skyscraper bottle chandelier seat stove shelf track cabinet cushion 44 objects sofa snow waterfall pillow sink sea rock desk sidewalk plate palm ground sky 13 scenes swivel chair ﬂoor fence ceiling ball pit skyscraper mountain snowy bakery-shop laundromat closet cockpit shoe shop corridor 13 parts conference room 1 material castle forest-needleleaf ﬁeld-cultivated windmill staircase classroom escalator-indoor playground pantry ofﬁce bowling alley cemetery desert-sand highway street amusement park lighthouse creek nursery corn ﬁeld 20 textures wheat ﬁeld 1 colorshopping mall-indoor attic cavern-indoor youth hostel sandbox bar reception childs room casino-indoor bookstore pagoda gymnasium-indoor waterfall-fan slum art gallery butchers shop hair wheel drawer screen head crosswalk roof balcony seat cushion pot shop window hand carpet food chequered striped wafﬂed meshed paisley dotted banded swirly cracked stratiﬁed lined polka-dotted perforated grid sprinkled lacelike ﬁbrous interlaced frilly spiralled potholed woven freckled studded veined red 4 9 5 0 5 objects 1 scene units car grass airplane mountain painting tree ceiling dog bus road pool table water horse skyscraper plant motorbike cat bed sea track sink ﬁeld stairway shelf work surface building house waterfall ground sidewalk book chair 32 objects sky windowpane ﬂoor toilet person railing 6 scenes washer signboard table ﬂower chandelier bird 6 parts ball pit skyscraper 2 materials mountain snowy closet swimming pool-outdoor stadium-baseball coast corridor auditorium creek highway conference room jacuzzi-indoor hair head screen wheel roof headboard crosswalk seat cushion drawer leg arm body shop window food 25 textures chequered 1 color perforated striped grid spiralled dotted swirly lined meshed studded paisley banded polka-dotted cobwebbed grooved stratiﬁed porous honeycombed woven interlaced red 0 units car tree grass sea mountain highway head hair banded lined chequered studded zigzagged striped perforated ﬂecked grid cracked meshed gauzy dotted units 10 35 Alexnet (Ambient Sound) Googlenet (Places) 5 7 colors VGG16 (Places) 0 17 textures AlexNetGAPWide (Places) 6 5 scenes water grass tree car plant windowpane road mountain airplane skyscraper dog sea ceiling building person horse bed track book pool table cabinet chair painting waterfall sidewalk sink shelf sky house stove ﬂoor bus mountain snowy ball pit pantry building facade skyscraper street hair wheel head screen crosswalk shop window food wood lined dotted banded studded grid honeycombed paisley zigzagged wafﬂed meshed cracked stratiﬁed chequered perforated sprinkled potholed grooved pleated matted freckled swirly spiralled woven ﬁbrous cobwebbed red Alexnet (Places) 3 6 objects 0 20 grass road sky water tree dog ball pit forest-broadleaf mountain snowy building facade highway chequered lined banded porous lacelike veined grid frilly ﬂecked perforated cracked stratiﬁed potholed freckled studded woven polka-dotted purple orange red yellow blue pink green 10 units 0 Fig. 12. Comparison of unique detectors of all types on a variety of architectures. More results are at the project page. Alexnet (Video Tracking) 61 objects 279 9 20 10 60 40 20 0 80 2 nv co 3 nv co 4 nv co nv co 5 nv co 3-3 co nv 4-3 co 5-1 nv -2 v5 n co 60 40 20 0 c e a 2 b b -s2 n-4 n-4 n-5 n-5 n-3 orm x7 tio tio tio tio tio 2-n ep ep ep ep ep nv inc inc inc inc inc co -3 v5 n co c ResNet152 on Places365 200 object scene part material texture color 100 0 1 nv co GoogLeNet on Places365 120 object scene part material texture color Number of unique detectors Number of unique detectors Number of unique detectors 30 VGG16 on Places365 80 object scene part material texture color Number of unique detectors AlexNet on Places365 40 object scene part material texture color 150 100 50 0 -1 BN -7 v1 on 16 -1 ise w Elt 8 -23 ise 8 5 e-3 wis Elt w Elt 8 7 e-4 wis Elt Elt 10 e-5 wis 10 5 0 80 1 10 0 2 nv co 3 nv co 4 nv co nv co 5 nv co 3-3 co nv 4-3 co 5-1 nv - v5 n co object scene part material texture color 2 -3 40 scene part material texture color 80 60 colorization tracking 0.091 dotted (texture) 40 10 0.140 chequered (texture) 0.167 20 0.063 perforated (texture) 0.085 head (part) 0 a e c 2 2 b b 7-s n-5 n-4 n-4 n-5 n-3 orm tio tio tio tio tio -7x 2-n ep ep ep ep ep v1 nv n inc inc inc inc inc co co head (part) 0.061 sky (object) 0.056 grass (object) 0 -1 16 BN 0.069 sky (object) Elt e-1 wis 8 Elt 3 e-2 wis 0.120 58 78 10 e-3 e-4 e-5 wis wis wis Elt Elt Elt 0.048 red-c (color) 0.100 20 la c et es -Im 36 ag 5 eN e t ob rac t je kin ct ce g nt r tra ic ns in v au di o m co ovi n lo riz g at io n p cr os uz sc zle ha eg nn om el ot io co n nt f r e a Al ex me xt or N e co t-r der nt an ex do te m nc od er et -P N N ex Fig. 15. The top ranked concepts in the three top categories in four selfsupervised networks. Some object and part detectors emerge in audio. Detectors for person heads also appear in puzzle and colorization. A variety of texture concepts dominate models with self-supervised training. Al ex Al puzzle 20 car (object) v5 n co scene part material texture color chequered (texture) 0.102 head (part) 0 Fig. 14. Semantic detectors emerge across different supervision of the primary training task. All these models use the AlexNet architecture and are tested at conv5. Fig. 10 shows the histogram of object detectors identified inside ResNet and DenseNet trained on ImageNet and Places respectively. DenseNet161-Places365 has the largest number of unique object detectors among all the networks. The emergent detectors differ across both training data source and architecture. The most frequent object detectors in the two networks trained on ImageNet are dog detectors, because there are more than 100 dog categories out of the 1000 classes in the ImageNet training set. Fig. 11 shows the examples of object detectors grouped by object categories. For the same object category, the visual appearance of the unit as detector varies not only within the same network but also across different networks. DenseNet and ResNet has such good detectors for bus and airplane with IoU > 0.25. Fig. 12 showcases the unique interpretable units of all types on a variety of networks. Fig. 13 shows the unique interpretable detectors over different layers for different network architectures trained on Places365. We observe that more object and scene detectors emerge at the higher layers across all architectures: AlexNet, VGG, GoogLeNet, and ResNet. This suggests that representation ability increases over layer depth. Because of the compositional structure of the CNN layers, the deeper layers should have higher capacity to represent concepts with larger visual complexity such as objects and scene parts. Our measurements confirm this, and we conclude that higher network depth encourages the emergence of visual concepts with higher semantic complexity. 3.4 40 audio 30 20 60 nv co 30 50 Number of unique detectors 100 scene part material texture color 40 Number of unique detectors 15 Number of unique detectors scene part material texture color 20 Number of unique detectors Number of unique detectors AlexNet on ImageNet VGG16 on ImageNet GoogLeNet on ImageNet ResNet152 on ImageNet Fig. and ResNet152 trained on Places365. All five conv layers of 25 13. Comparison of interpretability of 50 the layers for AlexNet, VGG16, GoogLeNet, 100 60 object the selected layers of VGG, GoogLeNet, object object object AlexNet and and ResNet are included. CNNs to predict the neighborhoods of two image patches, while [28] trains networks by colorizing images. Totally we investigate 12 networks trained for different self-supervised learning tasks. How do different supervisions affect those internal representations? Here we compare the interpretability of the deep visual representations resulting from self-supervised learning and supervised learning. We keep the network architecture the same as AlexNet for each model (one exception is the recent model transinv which uses VGG as the base network). Results are shown in Fig. 14. We observe that training on Places365 creates the largest number of unique detectors. Self-supervised models create many texture detectors but relatively few object detectors; apparently, supervision from a self-taught primary task is much weaker at inferring interpretable concepts than supervised training on a large annotated dataset. The form of self-supervision makes a difference: for example, the colorization model is trained on colorless images, and almost no color detection units emerge. We hypothesize that emergent units represent concepts required to solve the primary task. Fig. 15 shows some typical visual detectors identified in the self-supervised CNN models. For the models audio and puzzle, some object and part detectors emerge. Those detectors may be useful for CNNs to solve the primary tasks: the audio model is trained to associate objects with a sound source, so it may be useful to recognize people and cars; while the puzzle model is trained to align the different parts of objects and scenes in an image. For colorization and tracking, recognizing textures might be good enough for the CNN to solve primary tasks such as colorizing a desaturated natural image; thus it is unsurprising that the texture detectors dominate. Representations from Self-supervised Learning Recently many work have explored a novel paradigm for unsupervised learning of CNNs without using millions of annotated images, namely self-supervised learning. For example, [23] trains deep 3.5 Representations from Captioning Images To further compare the representations from supervised learning and self-supervised learning, we train a CNN from scratch using 10 COCO captioning dataset: AlexNet on Places205 Number of unique detectors Train from Scratch 40 30 20 object scene part material texture color 10 0 10 0 10 2 10 4 10 6 10 4 10 6 Validation accuracy 0.6 Fig. 16. Example images in the COCO captioning dataset, the CNN+LSTM image captioning model, and the network dissection result. Training ResNet18 from scratch using the supervision from captioning images leads to a lot of emergent object detectors. Number of unique detectors 100 object scene part material texture color 80 60 200 Number of detectors object scene part material texture color 150 100 0.4 0.2 0 10 0 10 2 Training iteration Fig. 18. The evolution of the interpretability of conv5 of Places205AlexNet over 3,000,000 training iterations. The accuracy on the validation at each iteration is also plotted. The baseline model is trained to 300,000 iterations (marked at the red line). 40 50 20 0 ba se lin re e pe at re 1 pe at re 2 pe N oD at3 ro Ba po u tc hN t or m ba se lin e re pe at 1 re pe at 2 re pe a N oD t3 ro Ba pou t tc hN or m 0 Fig. 17. Effect of regularizations on the interpretability of CNNs. the supervision of captioning images, which generates natural language sentence to describe the image contents. Specially we use the image captioning data from COCO dataset [50], where the images in COCO dataset are annotated with five image captions. Then we train a CNN plus LSTM as the image captioning model similar to [51]. Features of ResNet18 are used as input to the LSTM for generating image captions. The CNN+LSTM architecture and the network dissection results on the last convolutional layer of the ResNet18 are shown in Fig.16. We can see that many object detectors emerge in the CNN, and conclude that supervision from natural language captions contains high-level semantics to train the CNN. 3.6 Training Conditions Training conditions such as the number of training iterations, dropout [6], batch normalization [7], and random initialization [21], are known to affect the representation learning of neural networks. To analyze the effect of training conditions on interpretability, we take the Places205-AlexNet as the baseline model and prepare several variants of it, all using the same AlexNet architecture. For the variants Repeat1, Repeat2 and Repeat3, we randomly initialize the weights and train them with the same number of iterations. For the variant NoDropout, we remove the dropout in the FC layers of the baseline model. For the variant BatchNorm, we apply batch normalization at each convolutional layer of the baseline model. Repeat1, Repeat2, Repeat3 all have nearly the same top-1 accuracy 50.0% on the validation set. The variant without dropout has top-1 accuracy 49.2%. The variant with batch norm has top-1 accuracy 50.5%. Fig. 17 shows the interpretability of units in the CNNs over different training conditions. We find several effects: 1) Comparing different random initializations, the models converge to similar levels of interpretability, both in terms of the unique detector number and the total detector number; this matches observations of convergent learning discussed in [21]. 2) For the network without dropout, more texture detectors emerge but fewer object detectors. 3) Batch normalization seems to decrease interpretability significantly. The batch normalization result serves as a caution that discriminative power is not the only property of a representation that should be measured. Our intuition for the loss of interpretability under batch normalization is that the batch normalization ‘whitens’ the activation at each layer, which smooths out scaling issues and allows a network to easily rotate axes of intermediate representations during training. While whitening apparently speeds training, it may also have an effect similar to random rotations analyzed in Sec. 3.2 which destroy interpretability. As discussed in Sec. 3.2, however, interpretability is neither a prerequisite nor an obstacle to discriminative power. Finding ways to capture the benefits of batch normalization without destroying interpretability is an important area for future work. Fig. 18 plots the interpretability of snapshots of the baseline model at different training iterations along with the accuracy on the validation set. We can see that object detectors and part detectors begin emerging at about 10,000 iterations (each iteration processes a batch of 256 images). We do not find evidence of transitions across different concept categories during training. For example, units in conv5 do not turn into texture or material detectors before becoming object or part detectors. In Fig. 19, we keep track of six units over different training iteration. We observe that some units start converging to the semantic concept at early stage. For example, unit138 starts detecting mountain snowy as early as at iteration 2446. We also observe that units evolve over time: unit74 and unit108 detect road first before they start detecting car and airplane respectively. 11 Fig. 19. The interpretations of units change over iterations. Each row shows the interpretation of one unit. Fine-tuning a pre-trained network to a target domain is commonly used in transfer learning. The deep features from the pre-trained network show good generalization across different domains. The pre-trained network also makes the training converge faster and results in better accuracy, especially if there is not enough training data at the target domain. Here we analyze what happens inside the representation and how the interpretation of the internal units evolve during the transfer learning. We run two sets of experiments: fine-tuning Places-AlexNet to ImageNet and fine-tuning ImageNet-AlexNet to Places. We want to see how individual units mutate across domains. The interpretability results of the model checkpoints at different fine-tuning iteration are plotted in Fig. 20. We can see that the training indeed converges faster compared to the network trained from scratch on Places in Fig. 18. The interpretations of the units also change over finetuning. For example, the number of unique object detectors first drop then keep increasing for the network trained on ImageNet being fine-tuned to Places365, while it is slowly dropping for the network trained on Places being fine-tuned to ImageNet. Fig. 21 shows some examples of the individual unit evolution happening in the network trained from ImageNet to Places365 and the network trained from Places365 to ImageNet. For each network, we show six units with their interpretation at the beginning of finetuning and at the end of fine-tuning. For example, in the network fine-tuned from ImageNet to Places365, unit15 which detects the white dogs first, mutates to detect waterfall; unit136 and unit144 which detect dogs first, mutate to detect horse and cow respectively, as a lot of scene categories in Places such as pasture and corral contain such animals. On the other hand, in the network fine-tuned from Places365 to ImageNet, a lot of units mutate to various kinds of dog detectors. Interestingly though those units mutate to detect different concepts, those concepts share low-level similarity such as colors and textures. In Fig. 22, we zoom into two units from each of the two finetuning processes and plot the history of concept evolution. We can see that some units switch its top ranked label several times before converging to some concept: unit15 in the fine-tuning of ImageNet to Places365 flipped to white, crystalline, before reaches the waterfall concept. On the other hand, some units switch faster, for example unit132 in the fine-tuning of Places365 to ImageNet switches from hair to dog at early stage of fine-tuning. 3.8 Layer Width vs. Interpretability From AlexNet to ResNet, CNNs for visual recognition have grown deeper in the quest for higher classification accuracy. Depth has been shown to be important to high discrimination ability, and we have seen in Sec. 3.3 that interpretability can increase with depth as well. However, the width of layers (the number of units per Number of unique detectors Transfer Learning between Places and ImageNet Validation accuracy 3.7 Places365 to ImageNet 50 ImageNet to Places365 40 object scene part material 30 texture color 40 30 20 20 10 10 0 10 0 10 2 10 4 10 6 0 10 0 0.6 0.6 0.4 0.4 0.2 0.2 0 0 10 10 2 10 4 10 6 10 2 0 0 10 10 Training iteration 2 10 4 10 4 10 6 10 6 Training iteration Fig. 20. a) Fine-tune AlexNet from ImageNet to Places365. b) Fine-tune AlexNet from Places365 to ImageNet. ImageNet to Places365 Before After Places365 to ImageNet Before After Fig. 21. Units mutate from a) the network fine-tuned from ImageNet to Places365 and b) the network fine-tuned from Places365 to ImageNet. Six units are shown with their semantics at the beginning of the finetuning and at the end of the fine-tuning. layer) has been less explored. One reason is that increasing the number of convolutional units in a layer significantly increases computational cost while yielding only marginal improvements in classification accuracy. Nevertheless, some recent work [52] shows that a carefully designed wide residual network can achieve classification accuracy superior to the commonly used thin and deep counterparts. To explore how the width of layers affects interpretability of CNNs, we do a preliminary experiment to test how width affects emergence of interpretable detectors. We remove the FC layers of the AlexNet, then triple the number of units at the conv5, i.e., from 256 units to 768 units, as AlexNet-GAP-Wide. We further triple the number of units for all the previous conv layers except conv1 for the standard AlexNet, as AlexNet-GAP-WideAll. Finally we put a global average pooling layer after conv5 and fully connect the 12 ImageNet to Places365 before before after unit15 after unit100 Places365 to ImageNet after before before after unit132 unit31 Fig. 22. The history of one unit mutation during the fine-tuning from ImageNet to Places365 (top) and Places365 to ImageNet (low). Number of unique detectors 140 object scene part material texture color 120 100 80 60 40 3.10 20 AlexNet AlexNet-GAP-Wide nv 1 nv 2 co nv 3 co nv 4 co nv 5 co co v3 nv 4 co nv 5 co v2 co n v1 co n co n co co nv 1 nv 2 co nv 3 co nv 4 co nv 5 0 AlexNet-GAP-WideAll Fig. 23. Comparison of the standard AlexNet, AlexNet-GAP-Wide, and AlexNet-GAP-WideAll. Widening the layer brings the emergence of more detectors. Networks are trained on Places365. pooled 768-feature activations to the final class prediction. After training on Places365, the AlexNet-GAP-Wide and the AlexNet-GAP-WideAll obtain similar classification accuracy on the validation set as the standard AlexNet (∼ 0.5% top1 accuracy lower and higher), but it has many more emergent unique concept detectors at conv5 for AlexNet-GAP-Wide and at all the conv layers for AlexNet-GAL-WideAll, as shown in Fig. 23. We have also increased the number of units to 1024 and 2048 at conv5, but the number of unique concepts does not significantly increase further. This may indicate a limit on the capacity of AlexNet to separate explanatory factors; or it may indicate that a limit on the number of disentangled concepts that are helpful to solve the primary task of scene classification. 3.9 the performance on the test split. We compute the classification accuracy averaged across classes. We include the event8 [53], action40 [54], indoor67 [55], sun397 [56], caltech101 [57], and caltech256 [58]. The classification accuracies on six image datasets using the deep features are plotted in Fig. 24. We can see that the deep features from supervised trained networks perform much better than the ones from the self-supervised trained networks. Networks trained on Places have better features for scene-centric datasets such as sun397 and indoor67, while networks trained on ImageNet have better features for object-centric datasets such as caltech101 and action40. Fig. 25 plots the number of the unique object detectors for each representation over that representation’s classification accuracy on three selected datasets. We can see there is positive correlation between them. Thus the supervision tasks that encourage the emergence of more concept detectors may also improve the discrimination ability of deep features. Interestingly, on some of the object centric dataset, the best discriminative representation is the representation from ResNet152-ImageNet, which has fewer unique object detectors compared to the ResNet152-Places365. We hypothesize that the accuracy on a representation when applied to a task is dependent not only on the number of concept detectors in the representation, but on how well the concept detectors captures the characteristics of the hidden factors in the transferred dataset. Discrimination vs. Interpretability Activations from the higher layers of pre-trained CNNs are often used as generic visual features (noted as deep features), generalizing very well to other image datasets [16], [43]. It is interesting to bridge the generalization of the deep visual representation as generic visual features with their interpretability. Here we first benchmark the deep features from several networks on six image classification datasets for their discriminative power. For each network, we feed in the images and extract the activation at the last convolutional layer as the visual feature, then train a linear SVM with C = 0.001 on the train split and evaluate Explaining the Predictions for the Deep Features After we interpret the units inside the deep visual representation, we show that the unit activation along with the interpreted label can be used to explain the prediction given by the deep features. Previous work [59] uses the weighted sum of the unit activation maps to highlight which image regions are most informative to the prediction, here we further decouple at individual unit level to segment the informative image regions. We first plot the Class-specific units. After the linear SVM is trained, we can rank the elements of the feature according to their SVM weights to obtain the elements of the deep features which contribute most to that class. Those elements are units that act as explanatory factors, and we call those top ranked units associated with each output class class-specific units. Fig. 27 shows the classspecific units of ResNet152-ImageNet and ResNet152-Places365 for one class from action40 and sun397 respectively. For example, for the Walking the dog class from action40, the top three classspecific units from ResNet152-ImageNet are two dog detection unit and one person detection unit; for the Picnic area class from sun397, the top three class-specific units from ResNet152-Places365 are plant detection unit, grass detection unit, and fence detection unit. The intuitive match between visual detectors and the classes they explain suggests that visual detectors from CNNs behave as the bag-of-semantic-words visual features. We further use the individual units identified as concept detectors to build an explanation of the individual image prediction given by a classifier. The procedure is as follows: Given any image, let the unit activation of the deep feature (for ResNet the GAP activation) be [x1 , x2 , ..., xN ], where each xn represents the value summed up from the activation map P of unit n. Let the top prediction’s SVM response be s = n wn xn , where [w1 , w2 , ..., wN ] is the SVM’s learned weight. We get the top ranked units in Figure 26 by ranking [w1 x1 , w2 x2 , ..., wN xN ], which are the unit activations weighted by the SVM weight for the event8 0.973 0.966 0.963 0.955 0.946 0.933 0.928 0.927 0.923 0.920 0.918 0.916 0.904 0.904 0.897 0.875 0.852 0.838 0.836 0.829 0.825 0.806 0.782 0.739 0.688 0.621 0.551 0.5 0.766 0.748 0.708 0.691 0.691 0.674 0.563 0.555 0.555 0.527 0.527 0.520 0.498 0.492 0.485 0.484 0.353 0.351 0.347 0.346 0.329 0.299 0.271 0.241 0.208 0.187 0.177 1 Accuracy Accuracy 1 13 action40 0.8 0.6 0.4 0.2 0 R es R Net es 1 N 52 e R es t50 Ima -Im g N e et 15 VG ag Ne e t 2 VG -P G-H Ne G la t o G og VG G-I ce ybr oo Le G m s3 id gL N -P ag 65 e l e Al ex eNe t-Imace Ne N V t-P a s3 t et G l ge 6 -P G ac N 5 Al la -P e e ex ce la s3 t N s3 ce 65 e G Al Al t-P 65- s20 oo ex ex la G 5 gL N Ne ce AP s e Al Al eN t-I t-H 20 ex ex et ma yb 5 N N -P g rid et et la eN -P -P ce e la la s2 t ce ce 0 s2 s3 5 05 65 -B co cr lo au N os riz d sc at io ha ion tr nn ob ack el je co ing ct nt ce ex n t eg p tric om uz ot zle Al ex fra m ion N m ov et eo in -ra rd g nd er om R es R Ne G esNt15 o 2 R og et5 -Im es Le 0 a N N -Im ge et et a N 15 -I g e 2- ma eN t Pl ge e t a VG VG ce Ne s t G VG - G-H 36 G 5 oo G Im y gL VG -P ag bri eN G la eN d G et -Places et oo gL Ale -Pl ce 20 a A A eN xN ce s3 5 Al lex lex et- et- s3 65 ex Ne N Pl Hy 65 N t-P et- ac br et l Im es id Al Pla ace ag 205 e c s e Al xN es 20 Ne ex et 36 5- t N -P 5- BN et la G -P ce A la s3 P ce 6 s 5 co co 20 lo n 5 riz te cr at xt os io sc au n ha di tra nn o c el ob je p king c u eg tce zzl n e Al ex om tric N m oti et o - o n fra ran vin m do g eo m rd er 0 indoor67 0.4 Accuracy 0.6 1 0.8 0.6 0.4 0.2 0.2 0 0 R es R Net e 1 G 15 oo sNe 52 Accuracy gL t5 -Im VG2-P eN 0-I ag l G ac G et ma eN -P es oo -Im g e R e t es G gL VG VGlac 36 oo e G G es 5 VG VG ag Ne N 0.690 et G G eN t R gL Ne -P -H 36 es e t- la y 5 Al 15 -I -H e 0.666 N P b ex 2- m y t Al ReNet et- lacces rid N Pl ag bri 0.657 ex s 15 P e 20 e a Al Al t-Im ce eN d N Ne 2- lac s3 5 0.655 ex ex a s3 et et t5 Im e 65 A -P 0 a s2 A N le le e Ne ge 65 0.649 G la -Im ge 05 xN xN t-P t-H Ne Al oog V ces ag Ne 0.629 et et la yb t ex L GG 36 e t -P -P ce ri 0.605 N eN - 5- Ne la la s3 d e I ce ce 6 0.597 Al t-P et-I ma GA t s2 s2 5 ex la m ge P 05 05 a 0.515 Al N ce g N A e ex et s2 eN t co co -BN l e 0.513 N P 05 e xN cr lo n et la - t et VG os riz tex 0.513 Al Al -P ce BN -P G sc at t ex ex la s2 la -P ha ion 0.472 N Ne ce 05 ce la n et t- s3 0.460 s3 ce ne -Im H 6 VG 65 s3 l ag ybr 5 G 0.454 -G 65 oo eN id cr G A gL 0.447 os co e eN Pla puz P sc nt t G 0.381 c ha ex es zle e oo t t n 20 P gL Number of unique nobject detectors 0.287 e eN lac au 5 co tra aud l 0.271 es di et lo ck io -P tr 36 o 0.268 r i ob lacack 5 ob izat ng 0.250 je es ing je p ion ct 2 ct u 0.250 ce 05 c z Al ego ent zle n 0.208 ex m ric eg mo tric o A N m ti om vi 0.188 l f et e r xN am o ng -ra ov on 0.160 fra n in et eo tion -ra rd m do g 0.142 eo m nd er rd 0.105 om er 0.102 R es R es G N o e R Ne G ogL t15 oo e 2 G es t15 oo Ne 2 Accuracy gL Ne -Pl gL t5 Im eN t-P ac eN 0- ag e l es I et ma eN VG t-P ace 36 R 5 l g e I m es eN t VGG- aces20 a V 0.827 V N G G g e R G Pla s365 e e t -P ce 5 R esN 0.809 Al t15 G- G-H Ne Al es e V la s ex 2- Im y t ex N t5 G ce 20 0.782 N Pl ag br N et1 0- G s3 5 et 5 Im -H 6 Al et-I ace eN id 0.766 - 2 m e e y 5 s3 t Al x a V G Pla -Im age bri 0.754 N g 65 ex G o e e N V G t- N Al og V ces ag Ne d 0.732 et G -P H e G ex L G 36 eN t t Gunique la yb object N eN -I 5- e Number detectors 0.585 Al -Plof e t e ac -P ce rid Al t-P et-I ma GA 0.571 G Al xN es lac s36 e l m g P oo ex et 36 es 5 Al xN ace ag eN Al g N -P 5 2 0.519 ex et s2 eN et e L e l - 0 N -P 05 e 0.515 G xN eN t-P ace GA 5 e la t oo et et la s P Al Al t-P ce -BN 0.494 gL -P -P ce 20 ex ex lac s2 eN lac lac s3 5 0.493 N Ne es 05 et t- 3 et es2 es3 65 -Im H 6 -P 0 6 0.489 la 5- 5 a yb 5 ce B 0.483 cr geNrid cr s2 N os co e 0.476 os co 0 sc nt t 5 co sch nte 0.433 ha ex lo an xt nn t co 0.387 riz n lo au el at el riz d 0.380 io a n a tr ti io tr u 0.377 ob ack on ob ack dio 0.343 je p ing je p ing c u c u 0.322 eg tce zzl eg tce zz om ntr e om ntr le 0.319 ic ic A Al o 0.255 le fra m otio ex fra m tio xN m o n N m ov n 0.229 et eo vin et eo in -ra rd g -ra rd g 0.221 nd er nd er om om 0.190 0.148 Accuracy 1 0.8 0.827 0.797 0.792 0.788 0.785 0.768 0.756 0.750 0.674 0.672 0.664 0.615 0.607 0.580 0.574 0.510 0.430 0.424 0.402 0.392 0.362 0.339 0.335 0.285 0.259 0.206 0.178 0.934 0.926 0.919 0.918 0.889 0.862 0.861 0.854 0.798 0.785 0.750 0.743 0.737 0.728 0.713 0.704 0.678 0.669 0.650 0.649 0.622 0.576 0.546 0.517 0.511 0.466 0.343 caltech101 sun397 caltech256 0.8 1 0.8 0.6 0.6 0.4 80 0 60 AlexNet-Places365-GAP et 20 frameorder AlexNet-random 0.6 0.7 0.8 0.9 VGG-Places365 VGG-Hybrid AlexNet-Places365 AlexNet-Places205 0 objectcentric tracking puzzle moving audio crosschannel frameorder context 0 egomotion colorization AlexNet-random 0.1 ResNet152-Places365 0.2 0.3 0.4 0.5 0.6 60 40 20 VGG-Hybrid AlexNet-Places365-GAP AlexNet-Places365 AlexNet-Hybrid VGG-Places205 VGG-Places365 VGG-ImageNet GoogLeNet-Places365 GoogLeNet-Places205 GoogLeNet-ImageNet AlexNet-Places205 AlexNet-ImageNet audio tracking moving crosschannel colorization frameorder puzzle context 0 AlexNet-random egomotion 0 0.2 0.4 0.6 objectcentric AlexNet-Places205-BN Accuracy on indoor67 0.8 1 0.7 0 0.3 0.8 ResNet50-ImageNet 60 VGG-ImageNet GoogLeNet-Places365 GoogLeNet-Places205 GoogLeNet-ImageNet AlexNet-Places205 AlexNet-Places365-GAP AlexNet-Places365 40 AlexNet-Hybrid AlexNet-ImageNet objectcentric tracking puzzle moving audio crosschannel frameorder context 0 AlexNet-random egomotion colorization 20 0 0.1 0.2 0.3 0.4 Accuracy on sun397 AlexNet-Places205-BN 0.5 0.6 VGG-Hybrid VGG-ImageNet AlexNet-Places365-GAP GoogLeNet-ImageNet AlexNet-Places365 AlexNet-Places205 AlexNet-Hybrid AlexNet-Places205-BN objectcentric tracking audio puzzle crosschannel context egomotion colorization frameorder AlexNet-ImageNet moving 0.4 0.5 0.6 0.7 0.8 Accuracy on caltech101 100 ResNet152-ImageNet VGG-Places205 VGG-Hybrid VGG-Places365 80 ResNet50-ImageNet ResNet152-ImageNet GoogLeNet-Places365 GoogLeNet-Places205 20 ResNet152-Places365 100 Number of unique object detectors Number of unique object detectors 80 ResNet152-ImageNet ResNet50-ImageNet 40 AlexNet-random Fig. 24. The classification accuracy datasets. Accuracy on event8 of deep features on the six image Accuracy on action40 100 60 AlexNet-Places205-BN AlexNet-ImageNet 20 ResNet152-Places365 VGG-Places205 80 VGG-Places365 GoogLeNet-Places365 VGG-ImageNet GoogLeNet-Places205 GoogLeNet-ImageNet AlexNet-Hybrid AlexNet-Places365-GAP 40 1 ResNet152-ImageNet VGG-Places205 VGG-ImageNet GoogLeNet-ImageNet AlexNet-Places365 AlexNet-Places205 AlexNet-Hybrid AlexNet-Places205-BN AlexNet-ImageNet objectcentric tracking moving puzzle audio crosschannel colorization egomotion context 100 ResNet152-Places365 0.2ResNet50-ImageNet 60 GoogLeNet-Places365 GoogLeNet-Places205 40 0 0.5 0.4 100 ResNet50-ImageNet VGG-Places205 ResNet152-ImageNet 80 VGG-Hybrid VGG-Places365 es N Number of unique object detectors R 0.2 Number of unique object detectors ResNet152-Places365 100 VGG-Places205 80 VGG-Places365 40 20 0.7 1 ResNet152-Places365 ResNet50-ImageNet 60 0.9 VGG-ImageNet AlexNet-Places365-GAP GoogLeNet-Places205 AlexNet-Places365 AlexNet-Places205 AlexNet-Places205-BN ResNet152-ImageNet VGG-Hybrid GoogLeNet-Places365 GoogLeNet-ImageNet AlexNet-Hybrid objectcentric AlexNet-ImageNet tracking moving audio crosschannel frameorder context puzzle 0 egomotion colorization 0 AlexNet-random 0.2 0.4 0.6 0.8 Accuracy on caltech256 1 Fig. 25. The number of unique object detectors in the last convolutional layer compared to each representations classification accuracy on three datasets. Supervised (in red) and unsupervised (in green) representations clearly form two clusters. top predicted class. Then we simply upsample the activation map of the top ranked unit to segment the image. Image segmentations using individual unit activation are plotted in Fig. 26a. The unit segmentation explain the prediction explicitly. For example, the prediction for the first image is Gardening, and the explanatory units detect plant, grass, person, flower, and pot. The prediction for the second image is Riding a horse, the explanatory units detect horse, fence and dog. We also plot some incorrectly predicted samples in Figure 26b. The segmentation gives the intuition as to why the classifier made mistakes. For example, for the first image the classifier predicts cutting vegetables rather than the true label gardening, because the second unit incorrectly considers the ground as table. 4 C ONCLUSION Network Dissection translates qualitative visualizations of representation units into quantitative interpretations and measurements of interpretability. We have found that the units of a deep representation are significantly more interpretable than expected for a basis of the representation space. We have investigated the interpretability of deep visual representations resulting from different architectures, training supervisions, and training conditions. Furthermore, we have shown that interpretability of deep visual representations is 14 relevant to the power of the representation as a generalizable visual feature. We conclude that interpretability is an important property of deep neural networks that provides new insights into their hierarchical structure. Our work motivates future work towards building more interpretable and explainable AI systems. ACKNOWLEDGMENTS This work was partially funded by DARPA XAI program to A.T.; This work was also partly supported by the National Science Foundation under Grants No. 1524817 to A.T., and No. 1532591 to A.T. and A.O.; the Vannevar Bush Faculty Fellowship program sponsored by the Basic Research Office of the Assistant Secretary of Defense for Research and Engineering and funded by the Office of Naval Research through grant N00014-16-1-3116 to A.O.; the MIT Big Data Initiative at CSAIL, the Toyota Research Institute MIT CSAIL Joint Research Center, Google and Amazon Awards, and a hardware donation from NVIDIA Corporation. B.Z. is supported by a Facebook Fellowship. R EFERENCES B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Object detectors emerge in deep scene cnns,” International Conference on Learning Representations, 2015. [2] A. Gonzalez-Garcia, D. Modolo, and V. Ferrari, “Do semantic parts emerge in convolutional neural networks?” arXiv:1607.03738, 2016. [3] C. Vondrick, H. Pirsiavash, and A. Torralba, “Generating videos with scene dynamics,” arXiv:1609.02612, 2016. [4] Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A review and new perspectives,” IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 8, pp. 1798–1828, 2013. [5] D. Bau, B. Zhou, A. Khosla, A. Oliva, and A. Torralba, “Network dissection: Quantifying interpretability of deep visual representations,” in Proc. CVPR, 2017. [6] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: a simple way to prevent neural networks from overfitting.” Journal of Machine Learning Research, vol. 15, no. 1, pp. 1929–1958, 2014. [7] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” arXiv:1502.03167, 2015. [8] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” Proc. ECCV, 2014. [9] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Region-based convolutional networks for accurate object detection and segmentation,” IEEE transactions on pattern analysis and machine intelligence, 2016. [10] A. Mahendran and A. Vedaldi, “Understanding deep image representations by inverting them,” Proc. CVPR, 2015. [1] [11] K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep inside convolutional networks: Visualising image classification models and saliency maps,” International Conference on Learning Representations Workshop, 2014. [12] A. Mahendran and A. Vedaldi, “Understanding deep image representations by inverting them,” in Proc. CVPR, 2015. [13] A. Dosovitskiy and T. Brox, “Generating images with perceptual similarity metrics based on deep networks,” in Advances in Neural Information Processing Systems, 2016, pp. 658–666. [14] A. Nguyen, A. Dosovitskiy, J. Yosinski, T. Brox, and J. Clune, “Synthesizing the preferred inputs for neurons in neural networks via deep generator networks,” In Advances in Neural Information Processing Systems, 2016. [15] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. CVPR, 2016. [16] A. S. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, “Cnn features off-the-shelf: an astounding baseline for recognition,” arXiv:1403.6382, 2014. [17] P. Agrawal, R. Girshick, and J. Malik, “Analyzing the performance of multilayer neural networks for object recognition,” Proc. ECCV, 2014. [18] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How transferable are features in deep neural networks?” In Advances in Neural Information Processing Systems, 2014. [19] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus, “Intriguing properties of neural networks,” arXiv:1312.6199, 2013. [20] A. Nguyen, J. Yosinski, and J. Clune, “Deep neural networks are easily fooled: High confidence predictions for unrecognizable images,” in Proc. CVPR, 2015. [21] Y. Li, J. Yosinski, J. Clune, H. Lipson, and J. Hopcroft, “Convergent learning: Do different neural networks learn the same representations?” arXiv:1511.07543, 2015. [22] C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals, “Understanding deep learning requires rethinking generalization,” International Conference on Learning Representations, 2017. [23] C. Doersch, A. Gupta, and A. A. Efros, “Unsupervised visual representation learning by context prediction,” in Proc. CVPR, 2015. [24] M. Noroozi and P. Favaro, “Unsupervised learning of visual representations by solving jigsaw puzzles,” in Proc. ECCV, 2016. [25] D. Jayaraman and K. Grauman, “Learning image representations tied to ego-motion,” in Proc. ICCV, 2015. [26] P. Agrawal, J. Carreira, and J. Malik, “Learning to see by moving,” in Proc. ICCV, 2015. [27] X. Wang and A. Gupta, “Unsupervised learning of visual representations using videos,” in Proc. CVPR, 2015. [28] R. Zhang, P. Isola, and A. A. Efros, “Colorful image colorization,” in Proc. ECCV. Springer, 2016. [29] ——, “Split-brain autoencoders: Unsupervised learning by cross-channel prediction,” in Proc. CVPR, 2017. [30] A. Owens, J. Wu, J. H. McDermott, W. T. Freeman, and A. Torralba, “Ambient sound provides supervision for visual learning,” in Proc. ECCV, 2016. [31] R. Q. Quiroga, L. Reddy, G. Kreiman, C. Koch, and I. Fried, “Invariant visual representation by single neurons in the human brain,” Nature, vol. 435, no. 7045, pp. 1102–1107, 2005. a) b) Correct label: gardening Correct label: brushing Fig. 26. Segmenting images using top activated units weighted by the class label from ResNet152-Places365 deep feature. a) the correctly predicted samples. b) the incorrectly predicted samples. 15 Images from Walking the dog (action40) ResNet152-ImageNet ResNet152-Places365 Images from Picnic area (sun397) ResNet152-ImageNet ResNet152-Places365 Fig. 27. Class-specific units from ResNet152-ImageNet and ResNet152Places365 on one class from action40 and sun397. For each class, we show three sample images, followed by the top 3 units from ResNet152ImageNet and ResNet152-Places365 ranked by the class weight of linear SVM to predict that class. SVM weight, detected concept name and the IoU score are shown above each unit. [32] B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba, “Scene parsing through ade20k dataset,” Proc. CVPR, 2017. [33] S. Bell, K. Bala, and N. Snavely, “Intrinsic images in the wild,” ACM Trans. on Graphics (SIGGRAPH), 2014. [34] R. Mottaghi, X. Chen, X. Liu, N.-G. Cho, S.-W. Lee, S. Fidler, R. Urtasun, and A. Yuille, “The role of context for object detection and semantic segmentation in the wild,” in Proc. CVPR, 2014. [35] X. Chen, R. Mottaghi, X. Liu, S. Fidler, R. Urtasun, and A. Yuille, “Detect what you can: Detecting and representing objects using holistic models and body parts,” in Proc. CVPR, 2014. [36] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, “Describing textures in the wild,” in Proc. CVPR, 2014. [37] J. Van De Weijer, C. Schmid, J. Verbeek, and D. Larlus, “Learning color names for real-world applications,” IEEE Transactions on Image Processing, vol. 18, no. 7, pp. 1512–1523, 2009. [38] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105. [39] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” in Proc. CVPR, 2015. [40] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv:1409.1556, 2014. [41] G. Huang, Z. Liu, K. Q. Weinberger, and L. van der Maaten, “Densely connected convolutional networks,” Proc. CVPR, 2017. [42] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet large scale visual recognition challenge,” Int’l Journal of Computer Vision, 2015. [43] B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva, “Learning deep features for scene recognition using places database,” In Advances in Neural Information Processing Systems, 2014. [44] B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba, “Places: A 10 million image database for scene recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. [45] I. Mikjjsra, C. L. Zitnick, and M. Hebert, “Shuffle and learn: unsupervised learning using temporal order verification,” in Proc. ECCV, 2016. [46] R. Gao, D. Jayaraman, and K. Grauman, “Object-centric representation learning from unlabeled videos,” arXiv:1612.00500, 2016. [47] D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros, “Context encoders: Feature learning by inpainting,” in Proc. CVPR, 2016. [48] X. Wang, K. He, and A. Gupta, “Transitive invariance for self-supervised visual representation learning,” arXiv preprint arXiv:1708.02901, 2017. [49] P. Diaconis, “What is a random matrix?” Notices of the AMS, vol. 52, no. 11, pp. 1348–1349, 2005. [50] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in European conference on computer vision. Springer, 2014, pp. 740–755. [51] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan, “Show and tell: A neural image caption generator,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 3156–3164. [52] S. Zagoruyko and N. Komodakis, “Wide residual networks,” arXiv:1605.07146, 2016. [53] L.-J. Li and L. Fei-Fei, “What, where and who? classifying events by scene and object recognition,” in Proc. ICCV, 2007. [54] B. Yao, X. Jiang, A. Khosla, A. L. Lin, L. Guibas, and L. Fei-Fei, “Human action recognition by learning bases of action attributes and parts,” in Proc. ICCV, 2011. [55] A. Quattoni and A. Torralba, “Recognizing indoor scenes,” in Proc. CVPR, 2009. [56] J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba, “Sun database: Large-scale scene recognition from abbey to zoo,” in Proc. CVPR, 2010. [57] L. Fei-Fei, R. Fergus, and P. Perona, “Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories,” Computer Vision and Image Understanding, 2007. [58] G. Griffin, A. Holub, and P. Perona, “Caltech-256 object category dataset,” 2007. [59] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” in Proc. CVPR, 2016. Bolei Zhou is a Ph.D. Candidate in Computer Science and Artificial Intelligence Lab (CSAIL) at the Massachusetts Institute of Technology. He received M.Phil. degree in Information Engineering from the Chinese University of Hong Kong and B.Eng. degree in Biomedical Engineering from Shanghai Jiao Tong University in 2010. His research interests are computer vision and machine learning. He is an award recipient of Facebook Fellowship, Microsoft Research Asia Fellowship, and MIT Greater China Fellowship. David Bau is a PhD student at the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). He received an A.B. in Mathematics from Harvard in 1992 and an M.S. in Computer Science from Cornell in 1994. He coauthored a textbook on numerical linear algebra. He was a software engineer at Microsoft and Google and developed ranking algorithms for Google Image Search. His research interest is interpretable machine learning. Aude Oliva is a Principal Research Scientist at the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). After a French baccalaureate in Physics and Mathematics, she received two M.Sc. degrees and a Ph.D in Cognitive Science from the Institut National Polytechnique of Grenoble, France. She joined the MIT faculty in the Department of Brain and Cognitive Sciences in 2004 and CSAIL in 2012. Her research on vision and memory is cross-disciplinary, spanning human perception and cognition, computer vision, and human neuroscience. She received the 2006 National Science Foundation (NSF) Career award, the 2014 Guggenheim and the 2016 Vannevar Bush fellowships. Antonio Torralba received the degree in telecommunications engineering from Telecom BCN, Spain, in 1994 and the Ph.D. degree in signal, image, and speech processing from the Institut National Polytechnique de Grenoble, France, in 2000. From 2000 to 2005, he spent postdoctoral training at the Brain and Cognitive Science Department and the Computer Science and Artificial Intelligence Laboratory, MIT. He is now a Professor of Electrical Engineering and Computer Science at the Massachusetts Institute of Technology (MIT). Prof. Torralba is an Associate Editor of the International Journal in Computer Vision, and has served as program chair for the Computer Vision and Pattern Recognition conference in 2015. He received the 2008 National Science Foundation (NSF) Career award, the best student paper award at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) in 2009, and the 2010 J. K. Aggarwal Prize from the International Association for Pattern Recognition (IAPR).	1cs.CV
Semantic Photometric Bundle Adjustment on Natural Sequences Rui Zhu, Chaoyang Wang, Chen-Hsuan Lin, Ziyan Wang, Simon Lucey The Robotics Institute, Carnegie Mellon University arXiv:1712.00110v1 [cs.CV] 30 Nov 2017 {rz1, chaoyanw, chenhsul, ziyanw1}@andrew.cmu.edu, slucey@cs.cmu.edu Abstract The problem of obtaining dense reconstruction of an object in a natural sequence of images has been long studied in computer vision. Classically this problem has been solved through the application of bundle adjustment (BA). More recently, excellent results have been attained through the application of photometric bundle adjustment (PBA) methods – which directly minimize the photometric error across frames. A fundamental drawback to BA & PBA, however, is: (i) their reliance on having to view all points on the object, and (ii) for the object surface to be well textured. To circumvent these limitations we propose semantic PBA which incorporates a 3D object prior, obtained through deep learning, within the photometric bundle adjustment problem. We demonstrate state of the art performance in comparison to leading methods for object reconstruction across numerous natural sequences. Figure 1: Overview of the proposed method for semantic object-centric photometric bundle adjustment. Given a sequence of images with small motion (top row), we recover a full 3D dense shape of the object, as well as the camera poses w.r.t. the global coordinate system (middle row). Our method enables reprojection to the image plane with the estimated cameras (bottom row) to optimize for the photometric consistency as well as silhouette and depth constraints. 1. Introduction In this paper we are primarily concerned with the goal of obtaining dense 3D object reconstructions from short natural image sequences. One obvious strategy is to employ classical bundle adjustment (BA) [18] across the sequence where we can simultaneously recover camera pose and 3D points. Although reliable, this strategy is problematic as it: (i) can only recover 3D points if they are observed in the image sequence, and (ii) the density of the reconstruction is dependent on how textured the surface of the object is across the image sequence. Recently, photometric extensions to bundle adjustment have been proposed [1, 4] that directly minimize the photometric consistency between frames with respect to pose and 3D points. Borrowing upon the terminology of [1] we shall refer to these methods collectively herein as photometric bundle adjustment (PBA). PBA has recently proved advantageous over classical BA for problems where dense reconstructions are required due to their ability directly minimize for photometric consistency. Even with these recent innovations PBA is still, however, fundamentally limited in performance by (i) and (ii). There have been numerous efforts within the computer vision community to bring semantic prior into the task of object/scene 3D reconstruction. Convolutional Neural Networks (CNN) are proving remarkably useful for this task when one is provided with scene [17, 24] or object category [25, 7] specific labels & priors. A powerful characteristic of these semantic CNN methods is their ability to circumvent the fundamental limitations of (i) and (ii). For example, [25, 7, 21, 6] offer strategies for inferring a dense 3D reconstruction of an object from a single image even when a substantial portion of the projected 3D points are self-occluded. More recently, semantic CNN strategies have been proposed that attempt to incorporate multiple frames [3, 9]. Most of these previous efforts have been 1 trying to attack the problem of 3D reconstruction as a supervised learning problem – where geometry is largely treated as a label to be predicted. Although attractive in its simplicity this strategy has some inherent drawbacks. First, these geometric labels can be problematic to obtain – hand labeling can be error prone, and rendering can lack the necessary realism. Second, the predicted labels from these network models do not adhere to geometric constraints – such as photometric consistency – making the results unreliable. Recently, the application of geometric constraints within the offline CNN learning process has been entertained including reprojected silhouette matching [22, 25], depth matching [14], and even photometric consistency [19, 13, 12]. A fundamental problem with these emerging methods, however, is that the geometric constraints are not enforced at test time – dramatically reducing their effectiveness. Given these concerns, we argue that instead of incorporating geometric constraints into semantic CNN strategies offline, one should instead incorporate object semantics within the PBA pipeline. As we demonstrate in Fig. 1 and our results section this strategy gives the best of both worlds – semantic knowledge of the object with photometric consistency. In this paper, we propose an enhanced semantic PBA method which works on natural sequences as the classic PBA does, and give extensive evaluations on both synthetic and natural sequence domains. We summarize our contributions as follows: • We provide the first approach of its kind (to our knowledge) for semantic object-centric PBA on natural sequences – which gives the global 6DoF camera poses of each frame and the dense 3D shape, with PBA-like accuracy but denser depth maps. • We systematically evaluate the local optimality of our proposed optimization pipeline, as well as our enhanced objective which takes use of classic PBA results as off-the-shelf initialization or regularizer in our method. • We collect a new dataset for the task of object-centric shape reconstruction, consisting of rendered sequences of full ground truth in cameras, depth maps, and the shape in canonical pose, as well as natural sequences annotated with ShapeNet [2] models, making the dataset feasible for evaluating both PBA methods with camera estimations [11, 8], and methods which only recover aligned shapes or depth maps [20, 13, 9] by end-to-end learning on ShapeNet . 1.1. Related Work Photometric Bundle Adjustment: Photometric bundle adjustment (PBA) which is an optimization based method sitting entirely upon the visual cue of photometric consistency across all input frames [5, 15, 1, 11]. In PBA the shape is recovered by jointly optimizing for depth maps corresponding to the visible pixels in template frames [5, 15, 1], as well as camera motion. As a result the formulation of classic PBA is solely to recover the geometry of the scene, completely agnostic to semantics of the scene/object. Some works in PBA aims for small motion videos in particular [11, 10]. Shape Reconstructing with Deep Learning: As previously mentioned, early deep learning methods [7, 21, 6, 3, 9] solve the task of object-centric (object shape only) reconstruction with supervision from shape labels. Recently, an emerging school of thought seeks to bring in geometry back to the task, including reprojected silhouette matching [22, 25], depth matching [14], and even photometric consistency [19, 13, 12]. One issue of most of these methods is they assume known cameras, in global frame. This is in fact a too strong assumption to hold for natural sequences where global camera poses are not readily availabe. While some others do not account for camera motion at all [20, 3], which creates a gap from classic PBA where camera motions are instead the direct output. Semantic PBA: Recent work of Zhu et al. [26] also proposed to apply shape priors within PBA for 3D object reconstruction. In spite of the similarity in the formulation of the problem, Zhu et al.’s approach was problematic in a number of ways. First, the performance of Zhu et al. relies heavily on the initialization point given by CNN pose/style predictors trained predominantly on rendered images, which is suspicious for its reliability on natural sequences. Instead, we utilize a more reliable source – relative camera pose from PBA for initialization. Second, due to the limitations of the method, such as weak perspective camera model assumption and unreliable initialization source, Zhu et al.’s evaluation was restricted to rendered sequence. Thus they did not conduct quantitative comparisons between actual PBA methods [11, 8] w.r.t. camera pose error or depth error on real world sequences; while we give an extensive evaluation of our method on real sequences. Third, Zhu et al. did not give a proper analysis of the characteristics of their objective function, which results in using inadequate optimization techniques for their approach; In this paper, through inspecting the property of different cost functions, we propose a more robust and efficient optimization pipeline. Notation: Vectors are represented with lower-case bold font (e.g. a). Matrices are in upper-case bold (e.g. M) while scalars are lower-cased (e.g. a). Italicized upper-cased characters represent sets (e.g. S ). To denote the lth sample in a set (e.g. images, shapes), we use subscript (e.g. Ml ). Calligraphic symbols (e.g. π) denote functions. Images are defined as sampling function over the pixel coordinates, i.e. I(u) : R2 → R3 . Generator Offline PBA Pipeline & Initializers Camera Camera Pipeline of our approach Our camera model Figure 2: (Left) Pipeline of the optimization. Given an input sequence, we first run it through offline PBA pipeline to provide rough estimation for the depth maps as future depth map constraint, and camera motion as initialization to our camera motion parameters. Our style & pose initializer is also called to initialize the global pose of the template frame, as well as the style vector. Starting from the initialization, we optimize all variables until convergence through PBA-like pipeline with combined objective from photometric consistency Lph , silhouette matching error Lcd and inverse depth error Linvd . (Right) The perspective camera model. We adopt the camera model in BlenderTM by centering the object in the world frame, and positioning the camera with identity pose (in grey) along the Z axis, looking at the origin point. We show in this figure camera transformation from the identity camera can be parameterized with exponential twist ∆ω (solid blue arrow) and translation ∆t (red arrows) w.r.t. the camera local axes. 2. Approach M 2.1. Preliminary Camera model We assume a perspective camera with known intrinsics K. The camera extrinsics are parameterized as concatenation of exponential coordinates (also known as twist) of rotation and translation vector: p = [ω; t] ∈ R6 . The camera projection function is written as π(x; p) = K(R(ω)x + t). (1) Given a short window of L frames as in PBA [11, 1], we define the first frame as the target frame (frame 0), and the subsequent L − 1 frames as source frames. The relative camera pose between the target frame and source frame is denoted as ∆pl = [∆ωl ; ∆tl ]. The global pose is thus computed as a composing of relative pose of the source frame and the global pose of the target frame: pl = [∆ωl ◦ ω0 ; ∆tl + t0 ]. (2) We define the pose composition rule as: R(∆ω ◦ ω) = R(∆ω)R(ω), ∆t ◦ t = ∆t + t. (3) The reprojection of one point x onto frame l with the corresponding global pose is framed as sampling the image with reprojected pixel location Il (π(xj ; pl )). Reprojection with pseudo-raytracing Reprojection from a point set X given camera pose p can be viewed as first reasoning the visible part of the point set with a masking p function Xp := M(X; p) where Xp = {xj }j=1 . The mask function is implemented as pseudo-raytracing [14] by projecting the points to a enlarged inverse depth plane by a factor U , and then perform max pooling in a neighbourhood of U × U to figure out the visible point with biggest inverse depth. By doing so the mask function gives both the indices of the Mp visible points, and an inverse depth map. 2.2. Overview Our method takes in a RGB sequence taken by a calibrated camera moving around an object. The category(e.g. cars, airplanes, chairs) of the object is assumed to be known, and we have a rich repository of aligned CAD models (e.g. ShapeNet [2]) for this category. We define the world coordinate system as one attached to the objects as chosen by the CAD dataset, and the calibrated perspective camera model is parametrized by full 6DoF rotation and translation (see Fig.2). The goal of our method is to recover the full 3D shape of the object in the world frame, as well as the 6DoF parameters of the camera pose of each frame. We adopt the category-specific shape prior from Zhu et.al [26] to learn a shape space from the repository of ShapeNet [2] CAD models. We use dense point cloud as the shape parameterization in our work, considering learning shape space of point clouds has been made possible by several works [6, 16, 26]. The shape prior is a learned categoryspecific point cloud generator, written as a function of a style vector s ∈ RS which represents the sub-category object style. The output of the shape prior is the set of gener- ated points defined as X := {xi }N i=1 = G(s). For the 6DoF poses of the total L frames, we break the pose parameters into two sets: the global camera pose of the target frame p0 ∈ R6 , and the relative camera pose between each source L−1 frame and the corresponding target frame {∆pl ∈ R6 }l=1 . The overall pipeline of our method is illustrated in Fig. 2. We formulate the inference of the style vector and camera pose parameters as optimization steps with geometric objectives. The parameters are initialized with an off-theshelf initialization pipeline (see 2.4), and the optimization steps are taken to minimize this objective over the parameter space. In this paper, we propose to take advantage of the cheap and rough outputs from other methods to regularize our optimization procedure. For each frame, we get cheap segmentation masks (silhouettes) from recent state-of-theart instance segmentation method FCIS [23]. Considering traditional PBA methods gives as results semi-dense inverse depth and camera motion, we also borrow the readily available although error-prone outputs from PBA pipelines (e.g. openMVS [8]) to add another inverse depth loss with the estimated depth. We also take advantage of their camera motion estimation to initialize the relative camera pose of each source frame w.r.t. its target frame. 2.3. Optimization Objective Photometric Consistency The basic objective is formulated as the photometric consistency between the corresponding pixel pairs from the target frame and each source frame. Classic PBA methods usually track a set of sparse points through all the frames in a window, while with our formulation we are able to get dense correspondence automatically derived from reprojection. Considering that visible points may differ in each frame due to camera motion and occlusion, in this work we formulate the bi-directional photometric consistency as: Lph (p0 , {∆pl }L−1 (4) l=1 , s) = " Mp L−1 X X0 Lδ1 I0 π(xj ; p0 ) − Il π(xj ; p0 ◦ ∆pl ) l=1 Mpl + X k=1 j=1 Lδ1 Il π(xk ; p0 ◦ ∆pl ) − I0 π(xk ; p0 ) # where Lδ (·) is the Huber loss. Silhouette Error and Inverse Depth Error as Extra Constraints In Zhu et al. [26] the silhouette error is utilized as an extra constraint in the objective. We adopt the same objective but with the estimated silhouette with FCIS [23] that produce instance segmentation. We will show later that this constraint is still effective although the masks are errorprone. We write the silhouette distance for frame l as the 2D Chamfer distance between the set of pixel locations Ul1 in- side the rough silhouette, and the ones projected down from our camera model Ul2 . L−1 Lcd (p0 , {∆pl }l=1 , s) = L−1 1 X X min \|\|uk − uj \|\|22 uj ∈Ul2 L uk ∈Ul1 l=0 X min \|\|uj − uk \|\|22 . + uj ∈Ul2 (5) uk ∈Ul1 Finally, apart from cheap camera motion, we are able to get semi-dense depth map for each frame from the off-theshelf PBA pipeline. In this case we further formulate the extra objective term of inverse depth error as: L−1 Linvd (p0 , {∆pl }l=1 , s) = L−1 1 X Lδ2 (d′ l − αdl ) L (6) l=0 where dl is given by our reprojection module. Note that α here should be updated on the fly. Considering we will be getting confident camera poses for all frames in a few optimization steps, α can be robustly solved by finding a scale that best aligns the estimated camera poses with the ones from the offline estimator. Specifically, we can solve for an αl for each source frame l:  R0 (ω0 )x + t0 = αl R′ 0 x + t′ 0 (7) Rl (ωl )x + tl = αl R′ l x + t′ l . L−1 The solution for α is average over all {αl }l=1 : α= L−1 1 X arg min \|\|Rl (ωl )x + tl α L−1 (8) l=1 T − R′ l RT0 (R0 (ω0 )x + t0 ) − α(t′ l − R′ l R′ 0 t′ 0 )\|\|22 . The combined objective is given by: L = Lph + λ1 Lcd + λ2 Linvd , (9) where ablative study about the weight factors λ1 and λ2 is included the Appendix. 2.4. Initialization Style and Pose Initialization We improve upon existing learning-based pipeline to provide initialization for style and template frame camera pose. For style, unlike [25, 26] where a single-image based regressor is adopted, we use a recurrent network architecture to leverage the sequential information in an effort to alleviate ambiguity in style from a single viewpoint. Details about the architecture of our regressor and the training process as well as dataset are included in the Appendix. To generate more accurate style available to give an accurate estimation of the scale factor α. Instead we seek to bring the estimation of α in the loop by aligning the estimated inverse map d′ 0 to our reprojected inverse depth d0 : arg min \|\|d′ 0 − αd0 \|\|22 . α (11) The camera motion is then initialized by solving:  R0 (ω0 )x + t0 = α R′ 0 x + t′ 0 ∆Rl (∆ωl )R0 (ω0 )x + ∆tl + t0 = α R′ l x + t′ l (12) where the solution for ∆ωl and ∆tl is: Figure 3: Local cost surface of p0 and ∆pl . The yellow dot marks the optimum while the yellow line shows the search space for p0 if ∆pl is initialized with offline camera motion. vectors, we exploit cheap silhouette from Li et al. [23] to mask off the background. To find a coarse pose initialization, we first utilize BlenderTM to render templates under varying camera poses. After that we retrieve a coarse pose by finding one template which has the maximum IoU with the target silhouette. Camera Motion Initialization One observation of ours is, the photometric consistency error is problematic in our optimization of the objective. On the one hand Lph is locally bi-linear with Gauss-Newton solvers, where in traditional PBA the template term is fixed so that the residual is linear and the problem is locally convex. The problem is worsened in the way the variables are initialized as in Zhu et al. [26], L−1 where {∆pl }l=1 are initialized to all zeros when we have a poorly-initialized p0 . To illustrate this we show in Fig. 3 by plotting the cost surface over local perturbation of p0 and ∆pl . We show that the cost surface is highly non-convex between the initialization and the optimum if both variables are initialized far from the ground truth (red arrows). However we show with the camera motion parameters initialized correctly, the search space of p0 is constrained to the yellow line with better curvature for better convergence (green arrows). Inspired by this observation, at the beginning of our pipeline we run an off-the-shelf PBA pipeline (e.g. [8]) to L−1 acquire camera poses {p′l }l=0 for every frame, as well as a semi-dense point cloud X ′ . By formulation we have the following relation between our camera model (left) and the off-the-shelf estimator (right): Rl (ωl )x + tl = α R′l (ωl′ )x′ + t′ l . (10) Unfortunately no correspondence between x and x′ is ( T ∆Rl (∆ωl ) = R′ l R′ 0 T T ∆tl = R′ l R′ 0 t0 + α t′ 1 − R′ l R′ 0 t′ 0 . (13) We use Equ. 13 for initializing the camera motion parameters before the optimization steps. 2.5. Optimization Pipeline Given reasonable initialization from the above steps, we solve our objective with gradient-descent based methods. Particularly we found off-the-shelf L-BFGS solver gives efficient solution to our problem. We summarize our optimization pipeline in Algorithm 1. Algorithm 1 Optimization of the objective 1: 2: 3: 4: 5: 6: 7: 8: 9: L−1 procedure L(p0 , {∆pl }l=1 , s) p0 , s ← Initializer(I0 ) L−1 L−1 α, {∆pl }l=1 , {d′ l }l=0 ← openMVS [8], Equ. 13 while L > δL or step < maximum iterations do p0 ← L-BFGS update on p0 L−1 L−1 {∆pl }l=1 ← L-BFGS update on {∆pl }l=1 s ← L-BFGS update on s α ← Update on α by Equ. 8 L−1 return {p0 , {∆pl }l=1 , s} 3. Evaluation 3.1. Data Preparation Rendered Data To enable evaluation of our methods against Zhu et al. [26] which is only feasible on rendered domain, we follow Zhu et al. by rendering small baseline sequences from ShapeNet [2] cars. Please refer to Appendix for the statistics which is inherently identical to Zhu et al. apart from our perspective camera versus weak-perspective by Zhu et al. GT point frame 16 cloud frame 1 frame 2 frame 1 pred seg init … … … Figure 4: (Left) Examples of our test sequences. We show on the left three examples of rendered sequence, natural sequence of model car and real car respectively. (Right) Visualization of style & pose initialization. We also show some quantitative results on style and pose initialization for some natural sequences. The three columns on the right correspond to the first frame of a sequence, foreground segmentation of that first frame, and style & pose initialization viewed in rendered image, respectively. Natural Data In view of the absence of object-centric and category-specific natural video dataset, we collect our own test set with a mixture of sequences from toy model cars and real cars. We carefully choose the models with ground truth CAD available in ShapeNet [2]. Each video is shot with iPhone 6s (30fps) with around 30 ◦ of rotational motion and moderate translation, and the images are scaled down to 512 × 512. There are 15 videos collected from 4 toy car models and 12 videos from 6 real cars for evaluation. For each sequence, we annotate the template frame and last frame with ground truth CAD models and 6DoF pose. Samples of our test data can be found in Fig. 4. We also visualize some qualitative results of our initialization on natural sequences in Fig. 4. Although the style retrieved from regressor is not very precise in color or details, our style regressor is able to yield style prediction that are close in 3D shape. 3.2. Evaluation on Rendered Sequences In this section, we give both qualitative and quantitative results of our methods against 1) classic PBA methods of open-source openMVS [8], and Ham et al. [11] which is a PBA pipeline specifically optimized for small motion videos, 2) learning based methods [6, 3] which gives shapes in canonical poses. In our experiments, we set the weights in the optimization objective as λ1 = 0.1 with δ1 = 100 (we use unnormalized RGB values in range [0, 255]), and λ1 = 1000 with δ2 = 10. Ablative study on different settings of the hyper-parameters is included in the Appendix. Since we are evaluating on synthetic sequences from BlenderTM , we have access to full ground truth for both the camera pose of every frame (in world coordinate system), as well as the ground truth dense depth map and dense shape. However given our formulation is to output full 3D shape in an object-centric manner, while openMVS and Ham et al. give semi-dense point cloud of the entire scene by best effort, we are not able to align three outputs to the ground truth shape. Instead we choose to follow Ham et al. [11] to measure the depth error of the recovered points by reprojecting the shape onto the image plane with the estimated cameras. To compare with openMVS [8] and Ham et al. [11] which give only relative camera motion of every source frames w.r.t. to the first frame, we offer ground truth camera pose of the first frame to these two methods and find a rigid-body transformation to align their camera pose of the first frame to the ground truth. The scale ambiguity is solved by Equ.8. By doing this, the shape and camera poses of all methods are ideally aligned to the world coordinate system, and we are able to measure the camera error by calculating the camera position error as the distance of the estimated camera center to its ground truth, and the camera orientation error as the acute angle between the estimated camera orientation and its ground truth. We report the average error of depth maps and camera poses in Table 1 and the statistics in Fig. 6. The results show that our method achieves comparable camera error as openMVS [8], but slightly worse than Ham et al. [11] which is specifically optimized for small motion videos for better camera tracking. For the depth map error we achieve SLAM-like performance by outperforming both openMVS [8] and Ham et al. [11] in addition to much denser results thanks to the shape prior which produces full 3D shapes. Moreover, again thanks to the shape prior, we show in Fig. 7 that we only need a few observations of the object to give confident results, even at two frames, while classic methods [8, 11] start at considerable amount of motion to perform camera tracking. Finally we experiment on all the rendered sequences by perturbing upon their ground truth p0 and calculate the average p0 at convergence. We show in Fig. 8 that our method achieves better convergence in face of large initialization error in pose. 3.3. Evaluation on Natural Sequences We evaluate our methods against others on the objectcentric dataset we collected. The dataset includes sequences of a mixture of toy cars and real cars. The sequences are annotated with aligned CAD models retrieved from the ShapeNet dataset [2]. Considering it is not possible to get the 6DoF camera poses for all frames through human annotation, we evaluate on the depth error of the annotated first frame of each sequence, as well as the density of the reprojected points against the ground truth. The quantitative frame L-1 frame 0 frame L-1 frame 0 frame L-1 – – frame 0 frame L-1 – – [3, 6] Ham et al. [11] openMVS [8] Ours Input frame 0 Figure 5: Results of our methods on natural sequences, and comparison with openMVS [8], Ham et al. [11], and deep learning based methods [3, 6]. For openMVS [8] and Ham et al. [11] we align the cameras to the world coordinate frame (note that our method automatically produces camera poses in world frame), and we draw the aligned shape, estimated camera trajectory (blue) with orientation (red), annotated camera (black) of two frames (marked with black dot). We also project the shape with the estimated camera, and color the reprojected points with their inverse depths (brighter is closer). In the last row we show for each sample results from 3D-R2N2 [3] (left) and Fan et al. [6] (right). results are reported in Table 1. We also show qualitative results on 4 natural sequences (2 model cars plus 2 real cars) in Fig. 5. Each sequence consists of L = 91 frames with roughly 30 degree of camera rotation and moderate translation. We show that we achieve comparable camera poses and denser inverse depths against openMVS [8] and Ham et al. [11]. Additionally ours recovers semantic information including full 3D shape detached from the map, and again global cameras. For sequence 3 and 4 Ham et al. [11] gives degraded solution when it fails at camera tracking or deification due to little motion in frame 3 or significant lighting change in sequence 4. We attempt to give part of its results Zhu et al. Ham et al. openMVS Ours Depth Error 0.0732 0.0682 0.0731 0.0627 Density 0.8543 0.4732 0.5862 0.9342 Rendered Sequences Cam. Location Error 0.0134 0.0045 0.0098 0.0102 Cam. Orientation Error 2.0254◦ 0.9832◦ 1.3220◦ 1.2343◦ Natural Sequences Depth Error Density 0.0804 0.6558 0.0798 0.6987 0.0756 0.9076 Table 1: Quantitative comparison on both rendered and natural sequences. 10 5 12 8 1 6 0.8 8 4 0.6 % Pixel count 10 6 4 openMVS Ham et. al Zhu et. al Ours 2 0 0 0.1 0.2 0.4 2 0.2 0 Zhu et. al Ours 0 0 0.4 0.3 Depth error / Threshold of depth error Figure 6: Depth error on the rendered test set. On the left axis shows the histogram of depth error distribution, and on the right axis gives the percentage of pixels under the threshold. 1 openMVS Ham et. al Zhu et. al Ours 0.2 0.15 0.1 0.8 Density Average depth error 0.25 0.6 0 20 40 60 Number of views 80 100 4 6 8 10 Figure 8: Convergence analysis between ours and Zhu et al. [26]. Ours is more robust against initialization error in p0 . evaluate our approach on both rendered and natural settings against both classic PBA methods and deep learning based methods, and prove that it is capable to produce dense full 3D shape in world coordinates, as well as depth maps of PBA-like quality. 0.4 openMVS Ham et. al Zhu et. al Ours 0.2 0.05 2 0 0 20 40 60 80 Number of views Figure 7: Depth error on the rendered test set. On the left axis shows the histogram of depth error distribution, and on the right axis gives the percentage of pixels under the threshold. in the figure for readers’ reference. We also notice that learning based methods [6, 3] which are mostly trained on rendered images suffer from the domain gap in our test natural sequences. 4. Conclusion In this paper we propose the method of semantic photometric bundle adjustment for object-centric shape reconstruction from natural sequence, which exerts geometric constraints over the camera pose as well as the full 3D shape generated by a learned semantic shape prior. We extensively References [1] H. Alismail, B. Browning, and S. Lucey. Photometric bundle adjustment for vision-based slam. In Asian Conference on Computer Vision, pages 324–341. Springer, 2016. 1, 2, 3 [2] A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, J. Xiao, L. Yi, and F. Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 2, 3, 5, 6 [3] C. B. Choy, D. Xu, J. Gwak, K. Chen, and S. Savarese. 3dr2n2: A unified approach for single and multi-view 3d object reconstruction. arXiv preprint arXiv:1604.00449, 2016. 1, 2, 6, 7, 8 [4] J. Engel, V. Koltun, and D. Cremers. Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 1 [5] J. Engel, T. Schöps, and D. Cremers. Lsd-slam: Large-scale direct monocular slam. In European Conference on Computer Vision, pages 834–849. Springer, 2014. 2 [6] H. Fan, H. Su, and L. Guibas. A point set generation network for 3d object reconstruction from a single image. arXiv preprint arXiv:1612.00603, 2016. 1, 2, 3, 6, 7, 8 [7] R. Girdhar, D. F. Fouhey, M. Rodriguez, and A. Gupta. Learning a predictable and generative vector representation for objects. In European Conference on Computer Vision, pages 484–499. Springer, 2016. 1, 2 [8] A. Goldberg, S. Hed, H. Kaplan, R. Tarjan, and R. Werneck. Maximum flows by incremental breadth-first search. Algorithms–ESA 2011, pages 457–468, 2011. 2, 4, 5, 6, 7 [9] J. Gwak, C. B. Choy, M. Chandraker, A. Garg, and S. Savarese. Weakly supervised 3d reconstruction with adversarial constraint. 1, 2 [10] H. Ha, S. Im, J. Park, H.-G. Jeon, and I. So Kweon. Highquality depth from uncalibrated small motion clip. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5413–5421, 2016. 2 [11] C. Ham, M.-F. Chang, S. Lucey, and S. Singh. Monocular depth from small motion video accelerated. In 3D Vision (3DV), 2017 Fifth International Conference on 3D Vision. IEEE, 2017. 2, 3, 6, 7 [12] M. Ji, J. Gall, H. Zheng, Y. Liu, and L. Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. arXiv preprint arXiv:1708.01749, 2017. 2 [13] A. Kar, C. Häne, and J. Malik. Learning a multi-view stereo machine. arXiv preprint arXiv:1708.05375, 2017. 2 [14] C.-H. Lin, C. Kong, and S. Lucey. Learning efficient point cloud generation for dense 3d object reconstruction. arXiv preprint arXiv:1706.07036, 2017. 2, 3 [15] R. A. Newcombe, S. J. Lovegrove, and A. J. Davison. Dtam: Dense tracking and mapping in real-time. In Computer Vision (ICCV), 2011 IEEE International Conference on, pages 2320–2327. IEEE, 2011. 2 [16] C. R. Qi, L. Yi, H. Su, and L. J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 3 [17] K. Tateno, F. Tombari, I. Laina, and N. Navab. Cnn-slam: Real-time dense monocular slam with learned depth prediction. arXiv preprint arXiv:1704.03489, 2017. 1 [18] B. Triggs, P. F. McLauchlan, R. I. Hartley, and A. W. Fitzgibbon. Bundle adjustmenta modern synthesis. In International workshop on vision algorithms, pages 298–372. Springer, 1999. 1 [19] S. Tulsiani, T. Zhou, A. A. Efros, and J. Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. arXiv preprint arXiv:1704.06254, 2017. 2 [20] J. Wu, Y. Wang, T. Xue, X. Sun, W. T. Freeman, and J. B. Tenenbaum. MarrNet: 3D Shape Reconstruction via 2.5D Sketches. In Advances In Neural Information Processing Systems, 2017. 2 [21] J. Wu, C. Zhang, T. Xue, B. Freeman, and J. Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems, pages 82–90, 2016. 1, 2 [22] X. Yan, J. Yang, E. Yumer, Y. Guo, and H. Lee. Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. In Advances in Neural Information Processing Systems, pages 1696–1704, 2016. 2 [23] J. D. X. J. Yi Li, Haozhi Qi and Y. Wei. Fully convolutional instance-aware semantic segmentation. 2017. 4, 5 [24] T. Zhou, M. Brown, N. Snavely, and D. G. Lowe. Unsupervised learning of depth and ego-motion from video. arXiv preprint arXiv:1704.07813, 2017. 1 [25] R. Zhu, H. Kiani Galoogahi, C. Wang, and S. Lucey. Rethinking reprojection: Closing the loop for pose-aware shape reconstruction from a single image. In The IEEE International Conference on Computer Vision (ICCV), Oct 2017. 1, 2, 4 [26] R. Zhu, C. Wang, C.-H. Lin, Z. Wang, and S. Lucey. Objectcentric photometric bundle adjustment with deep shape prior. arXiv preprint arXiv:1711.01470, 2017. 2, 3, 4, 5, 8	1cs.CV
arXiv:1711.04969v2 [stat.ML] 22 Jan 2018 Straggler Mitigation in Distributed Optimization Through Data Encoding Can Karakus UCLA Los Angeles, CA karakus@ucla.edu Yifan Sun Technicolor Research Los Altos, CA Yifan.Sun@technicolor.com Suhas Diggavi UCLA Los Angeles, CA suhasdiggavi@ucla.edu Wotao Yin UCLA Los Angeles, CA wotaoyin@math.ucla.edu Abstract Slow running or straggler tasks can significantly reduce computation speed in distributed computation. Recently, coding-theory-inspired approaches have been applied to mitigate the effect of straggling, through embedding redundancy in certain linear computational steps of the optimization algorithm, thus completing the computation without waiting for the stragglers. In this paper, we propose an alternate approach where we embed the redundancy directly in the data itself, and allow the computation to proceed completely oblivious to encoding. We propose several encoding schemes, and demonstrate that popular batch algorithms, such as gradient descent and L-BFGS, applied in a coding-oblivious manner, deterministically achieve sample path linear convergence to an approximate solution of the original problem, using an arbitrarily varying subset of the nodes at each iteration. Moreover, this approximation can be controlled by the amount of redundancy and the number of nodes used in each iteration. We provide experimental results demonstrating the advantage of the approach over uncoded and data replication strategies. 1 Introduction Solving large-scale optimization problems has become feasible through distributed implementations. However, the efficiency can be significantly hampered by slow processing nodes, network delays or node failures. In this paper we develop an optimization framework based on encoding the dataset, which mitigates the effect of straggler nodes in the distributed computing system. Our approach can be readily adapted to the existing distributed computing infrastructure and software frameworks, since the node computations are oblivious to the data encoding. In this paper, we focus on problems of the form min f (w) := w∈Rp 1 min kXw − yk2 , 2n w∈Rp (1) where X ∈ Rn×p , y ∈ Rn×1 represent the data matrix and vector respectively. The function f (w) is mapped onto a distributed computing setup depicted in Figure 1, consisting of one central server and m worker nodes, which collectively store the row-partitioned matrix X and vector y. We focus on batch, synchronous optimization methods, where the delayed or failed nodes can significantly slow down the overall computation. Note that asynchronous methods are inherently robust to delays caused 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA. by stragglers, although their convergence rates can be worse than their synchronous counterparts. Our e = SX and ye = Sy, approach consists of adding redundancy by encoding the data X and y into X respectively, where S ∈ R(βn)×n is an encoding matrix with redundancy factor β ≥ 1, and solving the effective problem 1 1 e min fe(w) := minp kS (Xw − y) k2 = minp kXw − yek2 w∈R 2βn w∈R 2βn w∈Rp (2) instead. In doing so, we proceed with the computation in each iteration without waiting for the stragglers, with the idea that the inserted redundancy will compensate for the lost data. The goal is to design the matrix S such that, when the nodes obliviously solve the problem (2) without waiting for the slowest (m − k) nodes (where k is a design parameter) the achieved solution approximates the original solution w∗ = arg minw f (w) sufficiently closely. Since in large-scale machine learning and data analysis tasks one is typically not interested in the exact optimum, but rather a “sufficiently" good solution that achieves a good generalization error, such an approximation could be acceptable in many scenarios. Note also that the use of such a technique does not preclude the use of other, non-coding straggler-mitigation strategies (e.g., [23, 21, 1] and references therein), which can still be implemented on top of the redundancy embedded in the system, to potentially further improve performance. Focusing on gradient descent and L-BFGS algorithms, we show that under a spectral condition on S, one can achieve an approximation of the solution of (1), by solving (2), without waiting for the stragglers. We show that with sufficient redundancy embedded, and with updates from a sufficiently large, yet strict subset of the nodes in each iteration, it is possible to deterministically achieve linear convergence to a neighborhood of the solution, as opposed to convergence in expectation (see Fig. 4). Further, one can adjust the approximation guarantee by increasing the redundancy and number of node updates waited for in each iteration. Another potential advantage of this strategy is privacy, since the nodes do not have access to raw data itself, but can still perform the optimization task over the jumbled data to achieve an approximate solution. Although in this paper we focus on quadratic objectives and two specific algorithms, in principle our approach can be generalized to more general, potentially non-smooth objectives and constrained optimization problems, as we discuss in Section 4 ( adding a regularization term is also a simple generalization). Our main contributions are as follows. (i) We demonstrate that gradient descent (with constant step size) and L-BFGS (with line search) applied in a coding-oblivious manner on the encoded problem, achieves (universal) sample path linear convergence to an approximate solution of the original problem, using only a fraction of the nodes at each iteration. (ii) We present three classes of coding matrices; namely, equiangular tight frames (ETF), fast transforms, and random matrices, and discuss their properties. (iii) We provide experimental results demonstrating the advantage of the approach over uncoded (S = I) and data replication strategies, for ridge regression using synthetic data on an AWS cluster, as well as matrix factorization for the Movielens 1-M recommendation task. Related work. Use of data replication to aid with the straggler problem has been proposed and studied in [21, 1], and references therein. Additionally, use of coding in distributed computing has been explored in [12, 7]. However, these works exclusively focused on using coding at the computation level, i.e., certain linear computational steps are performed in a coded manner, and explicit encoding/decoding operations are performed at each step. Specifically, [12] used MDS-coded distributed matrix multiplication and [7] focused on breaking up large dot products into shorter dot products, and perform redundant copies of the short dot products to provide resilience against stragglers. [20] considers a gradient descent method on an architecture where each data sample is replicated across nodes, and designs a code such that the exact gradient can be recovered as long as fewer than a certain number of nodes fail. However, in order to recover the exact gradient under any potential set of stragglers, the required redundancy factor is on the order of the number of straggling nodes, which could mean a large amount of overhead for a large-scale system. In contrast, we show that one can converge to an approximate solution with a redundancy factor independent of network size or problem dimensions (e.g., 2 as in Section 5). Our technique is also closely related to randomized linear algebra and sketching techniques [13, 6, 16], used for dimensionality reduction of large convex optimization problems. The main difference between this literature and the proposed coding technique is that the former focuses on reducing the 2 M N1 N2 2 kX1 w − y1 k M Nm 2 kX2 w − y2 k N1 2 N2 2 kXm w − ym k Nm 2 kS1 (Xw − y)k kS2 (Xw − y)k kSm (Xw − y)k2 Figure 1: Left: Uncoded distributed optimization with partitioning, where X and y are partitioned as > > > > X = X1> X2> . . . Xm and y = y1> y2> . . . ym . Right: Encoded distributed optimization, where node i stores (Si X, Si y), instead of (Xi , yi ). The uncoded case corresponds to S = I. problem dimensions to lighten the computational load, whereas coding increases the dimensionality of the problem to provide robustness. As a result of the increased dimensions, coding can provide a much closer approximation to the original solution compared to sketching techniques. 2 Encoded Optimization Framework Figure 1 shows a typical data-distributed computational model in large-scale optimization (left), as well as our proposed network consists of m machines, where encoded model (right). Our computing > > > > e machine i stores Xi , yei = (Si X, Si y) and S = S1 S2 . . . Sm . The optimization process is oblivious to the encoding, i.e., once the data is stored at the nodes, the optimization algorithm proceeds exactly as if the nodes contained uncoded, raw data (X, y). In each iteration t, the central server broadcasts the current estimate wt , and each worker machine computes and sends to the server e > (X ei wt − yei ). the gradient terms corresponding to its own partition gi (wt ) := X i Note that this framework of distributed optimization is typically communication-bound, where communication over a few slow links constitute a significant portion of the overall computation time. We consider a strategy where at each iteration t, the server only uses the gradient updates from the first k nodes to respond in that iteration, thereby preventing such slow links and straggler nodes from stalling the overall computation: 1 X 1 e> e X (XA wt − yeA ), get = gi (wt ) = 2βηn βηn A i∈At k where At ⊆ [m], \|At \| = k are the indices of the first k nodes to respond at iteration t, η := m e and XA = [Si X]i∈At . (Similarly, SA = [Si ]i∈At .) Given the gradient approximation, the central server then computes a descent direction dt through the history of gradients and parameter estimates. For the remaining nodes i 6∈ At , the server can either send an interrupt signal, or simply drop their updates upon arrival, depending on the implementation. Next, the central server chooses a step size αt , which can be chosen as constant, decaying, or through e t that is needed to compute the step size. We exact line search 1 by having the workers compute Xd again assume the central server only hears from the fastest k nodes, denoted by Dt ⊆ [m], where Dt 6= At in general, to compute αt = −ν d> et t g , > > e X e D dt dt X (3) D eD = [Si X] where X i∈Dt , and 0 < ν < 1 is a back-off factor of choice. Our goal is to especially focus on the case k < m, and design an encoding matrix S such that, for any sequence of sets {At }, {Dt }, f (wt ) universally converges to a neighborhood of f (w∗ ). Note that in general, this scheme with k < m is not guaranteed to converge for traditionally batch methods like L-BFGS. Additionally, although the algorithm only works with the encoded function fe, our goal is to provide a convergence guarantee in terms of the original function f . 1 Note that exact line search is not more expensive than backtracking line search for a quadratic loss, since it only requires a single matrix-vector multiplication. 3 3 Algorithms and Convergence Analysis Let the smallest and largest eigenvalues of X > X be denoted by µ > 0 and M > 0, respectively. Let η with β1 < η ≤ 1 be given. In order to prove convergence,we will consider a family of matrices (β) S where β is the aspect ratio (redundancy factor), such that for any > 0, and any A ⊆ [m] with \|A\| = ηm, > (1 − )I SA SA (1 + )I, (4) for sufficiently large β ≥ 1, where SA = [Si ]i∈A is the submatrix associated with subset A (we drop dependence on β for brevity). Note that this is similar to the restricted isometry property (RIP) used in compressed sensing [4], except that (4) is only required for submatrices of the form SA . Although this condition is needed to prove worst-case convergence results, in practice the proposed encoding scheme can work well even when it is not exactly satisfied, as long as the bulk of the eigenvalues of > SA SA lie within a small interval [1 − , 1 + ]. We will discuss several specific constructions and their relation to property (4) in Section 4. Gradient descent. We consider gradient descent with constant step size, i.e., wt+1 = wt + αdt = wt − αe gt . The following theorem characterizes the convergence of the encoded problem under this algorithm. Theorem 1. Let ft = f (wt ), where wt is computed using gradient descent with updates from a set 2ζ of (fastest) workers At , with constant step size αt ≡ α = M (1+) for some 0 < ζ ≤ 1, for all t. If S satisfies (4) with > 0, then for all sequences of {At } with cardinality \|At \| = k, κ2 (κ − γ1 ) f (w∗ ) , t = 1, 2, . . . , 1 − κγ1 1+ , and γ1 = 1 − 4µζ(1−ζ) where κ = 1− M (1+) , and f0 = f (w0 ) is the initial objective value. t ft ≤ (κγ1 ) f0 + The proof is provided in Appendix B, which relies on the fact that the solution to the effective “instantaneous" problem corresponding to the subset At lies in the set {w : f (w) ≤ κ2 f (w∗ )}, and therefore each gradient descent step attracts the estimate towards a point in this set, which must eventually converge to this set. Note that in order to guarantee linear convergence, we need κγ1 < 1, which can be ensured by property (4). Theorem 1 shows that gradient descent over the encoded problem, based on updates from only k < m nodes, results in deterministically linear convergence to a neighborhood of the true solution w∗ , for sufficiently large k, as opposed to convergence in expectation. Note that by property (4), by controlling the redundancy factor β and the number of nodes k waited for in each iteration, one can control the approximation guarantee. For k = m and S designed properly (see Section 4), then κ = 1 and the optimum value of the original function f (w∗ ) is reached. Limited-memory-BFGS. Although L-BFGS is originally a batch method, requiring updates from all nodes, its stochastic variants have also been proposed recently [14, 3]. The key modification to ensure convergence is that the Hessian estimate must be computed via gradient components that are common in two consecutive iterations, i.e., from the nodes in At ∩ At−1 . We adapt this technique to our scenario. For t > 0, define ut := wt − wt−1 , and X m rt := (gi (wt ) − gi (wt−1 )) . 2βn \|At ∩ At−1 \| i∈At ∩At−1 Then once the gradient terms {gt }i∈At are collected, the descent direction is computed by dt = −Bt get , where Bt is the inverse Hessian estimate for iteration t, which is computed by (`+1) Bt (`) = Vj> Bt Vj` + ρj` uj` u> j` , ` 4 ρk = 1 rk> uk , Vk = I − ρk rk u> k (0) r> r (e σ) t t I, and Bt := Bt with σ with j` = t − σ e + `, Bt = r> e := min {t, σ}, where σ is the L-BFGS t ut memory length. Once the descent direction dt is computed, the step size is determined through exact line search, using (3), with back-off factor ν = 1− 1+ , where is as in (4). For our convergence result for L-BFGS, we need another assumption on the matrix S, in addition to (4). Defining S̆t = [Si ]i∈At ∩At−1 for t > 0, we assume that for some δ > 0, δI S̆t> S̆t (5) for all t > 0. Note that this requires that one should wait for sufficiently many nodes to finish so that the overlap set At ∩ At−1 has more than a fraction β1 of all nodes, and thus the matrix S̆t can 1 be full rank. This is satisfied if η ≥ 21 + 2β in the worst-case, and under the assumption that node delays are i.i.d., it is satisfied in expectation if η ≥ √1β . However, this condition is only required for a worst-case analysis, and the algorithm may perform well in practice even when this condition is not satisfied. The following lemma shows the stability of the Hessian estimate. Lemma 1. If (5) is satisfied, then there exist constants c1 , c2 > 0 such that for all t, the inverse Hessian estimate Bt satisfies c1 I Bt c2 I. The proof, provided in Appendix A, is based on the well-known trace-determinant method. Using Lemma 1, we can show the following result. Theorem 2. Let ft = f (wt ), where wt is computed using L-BFGS as described above, with gradient updates from machines At , and line search updates from machines Dt . If S satisfies (4) and (5), for all sequences of {At }, {Dt } with \|At \| = \|Dt \| = k, t ft ≤ (κγ2 ) f0 + where κ = 1+ 1− , and γ2 = 1 − 4µc1 c2 M (c1 +c2 )2 κ2 (κ − γ2 ) f (w∗ ) , 1 − κγ2 , and f0 = f (w0 ) is the initial objective value. The proof is provided in Appendix B. Similar to Theorem 1, the proof is based on the observation that the solution of the effective problem at time t lies in a bounded set around the true solution w∗ . As in gradient descent, coding enables linear convergence deterministically, unlike the stochastic and multi-batch variants of L-BFGS [14, 3]. Generalizations. Although we focus on quadratic cost functions and two specific algorithms, our 2 approach can potentially be generalized for objectives of the form kXw − yk + h(w) for a simple 2 convex function h, e.g., LASSO; or constrained optimization minw∈C kXw − yk (see [11]); as well as other first-order algorithms used for such problems, e.g., FISTA [2]. In the next section we demonstrate that the codes we consider have desirable properties that readily extend to such scenarios. 4 Code Design We consider three classes of coding matrices: tight frames, fast transforms, and random matrices. nβ Tight frames. A unit-norm frame for Rn is a set of vectors F = {φi }i=1 with kφi k = 1, where β ≥ 1, such that there exist constants ξ1 ≥ ξ2 > 0 such that, for any u ∈ Rn , 2 ξ1 kuk ≤ nβ X 2 \|hu, φi i\| ≤ ξ2 kuk2 . i=1 The frame is tight if the above satisfied with ξ1 = ξ2 . In this case, it can be shown that the constants are equal to the redundancy factor of the frame, i.e., ξ1 = ξ2 = β. If we form S ∈ R(βn)×n by rows that are a tight frame, then we have S > S = βI, which ensures kXw − yk2 = β1 kSXw − Syk2 . Then for any solution w e∗ to the encoded problem (with k = m), ∇fe(w e∗ ) = X > S > S(X w e∗ − y) = β(X w e∗ − y)> X = β∇f (w e∗ ). 5 > Figure 3: Sample spectrum of SA SA for various constructions with low redundancy, and large k (normalized). > Figure 2: Sample spectrum of SA SA for various constructions with high redundancy, and relatively small k (normalized). Therefore, the solution to the encoded problem satisfies the optimality condition for the original problem as well: ∇fe(w e∗ ) = 0, ⇔ ∇f (w e∗ ) = 0, and if f is also strongly convex, then w e∗ = w∗ is the unique solution. Note that since the computation is coding-oblivious, this is not true in general for an arbitrary full rank matrix, and this is, in addition to property (4), a desired property of the encoding matrix. In fact, this equivalency extends beyond smooth unconstrained optimization, in that D E ∇fe(w e∗ ), w − w e∗ ≥ 0, ∀w ∈ C ⇔ h∇f (w e∗ ), w − w e∗ i ≥ 0, ∀w ∈ C for any convex constraint set C, as well as −∇fe(w e∗ ) ∈ ∂h(w e∗ ), ⇔ −∇f (w e∗ ) ∈ ∂h(w e∗ ), for any non-smooth convex objective term h(x), where ∂h is the subdifferential of h. This means that tight frames can be promising encoding matrix candidates for non-smooth and constrained optimization too. In [11], it was shown that when {At } is static, equiangular tight frames allow for a close approximation of the solution for constrained problems. A tight frame is equiangular if \|hφi , φj i\| is constant across all pairs (i, j) with i 6= j. nβ Proposition 1 (Welch bound [22]). Let F = {φi }i=1 be a tight frame. Then ω(F ) ≥ Moreover, equality is satisfied if and only if F is an equiangular tight frame. q β−1 2nβ−1 . Therefore, an ETF minimizes the correlation between its individual elements, making each submatrix > SA SA as close to orthogonal as possible, which is promising in light of property (4). We specifically evaluate Paley [15, 10] and Hadamard ETFs [19] (not to be confused with Hadamard matrix, which is discussed next) in our experiments. We also discuss Steiner ETFs [8] in Appendix D, which enable efficient implementation. Fast transforms. Another computationally efficient method for encoding is to use fast transforms: Fast Fourier Transform (FFT), if S is chosen as a subsampled DFT matrix, and the Fast WalshHadamard Transform (FWHT), if S is chosen as a subsampled real Hadamard matrix. In particular, one can insert rows of zeroes at random locations into the data pair (X, y), and then take the FFT or FWHT of each column of the augmented matrix. This is equivalent to a randomized Fourier or Hadamard ensemble, which is known to satisfy the RIP with high probability [5]. Random matrices. A natural choice of encoding is using i.i.d. random matrices. Although such random matrices do not have the computational advantages of fast transforms or the optimalitypreservation property of tight frames, their eigenvalue behavior can be characterized analytically. In particular, using the existing results on the eigenvalue scaling of large i.i.d. Gaussian matrices [9, 18] and union bound, it can be shown that r 2 ! 1 > 1 P max λmax SA SA > 1 + → 0, (6) βηn βη A:\|A\|=k r 2 ! 1 > 1 P min λmin S SA < 1 − → 0, (7) βηn A βη A:\|A\|=k 6 Figure 4: Left: Sample evolution of uncoded, replication, and Hadamard (FWHT)-coded cases, for k = 12, m = 32. Right: Runtimes of the schemes for different values of η, for the same number of iterations for each scheme. Note that this essentially captures the delay profile of the network, and does not reflect the relative convergence rates of different methods. as n → ∞, where σi denotes the ith singular value. Hence, for sufficiently large redundancy and problem dimension, i.i.d. random matrices are good candidates for encoding as well. However, for finite β, even if k = m, in general for this encoding scheme the optimum of the original problem is not recovered exactly. Property (4) and redundancy requirements. Using the analytical (6)–(7) on i.i.d. Gaus bounds 1 √ sian matrices, one can see that such matrices satisfy (4) with = O , independent of problem βη dimensions or number of nodes m. Although we do not have tight eigenvalue bounds for subsampled ETFs, numerical evidence (Figure 2) suggests that they may satisfy (4) with smaller than random matrices, and thus we believe that the required redundancy in practice is even smaller for ETFs. Note that our theoretical results focus on the extreme eigenvalues due to a worst-case analysis; in practice, most of the energy of the gradient will be on the eigen-space associated with the bulk of the eigenvalues, which the following proposition suggests can be mostly 1 (also see Figure 3), which means even if (4) is not satisfied, the gradient (and the solution) can be approximated closely for a modest redundancy, such as β = 2. The following result is a consequence of the Cauchy interlacing theorem, and the definition of tight frames. Proposition 2. If the rows of S are chosen to form an ETF with redundancy β, then for η ≥ 1 − β1 , 1 > β SA SA has n(1 − βη) eigenvalues equal to 1. 5 Numerical Results Ridge regression with synthetic data on AWS EC2 cluster. We generate the elements of matrix X i.i.d. ∼ N (0, 1), the elements of y i.i.d. ∼ N (0, p), for dimensions (n, p) = (4096, 6000), and 2 e − ye + λ kwk2 , for regularization parameter λ = 0.05. We solve the problem minw 1 Xw 2βn 2 evaluate column-subsampled Hadamard matrix with redundancy β = 2 (encoded using FWHT for fast encoding), data replication with β = 2, and uncoded schemes. We implement distributed L-BFGS as described in Section 3 on an Amazon EC2 cluster using the mpi4py Python package, over m = 32 m1.small worker node instances, and a single c3.8xlarge central server instance. We assume the central server encodes and sends the data variables to the worker nodes (see Appendix D for a discussion of how to implement this more efficiently). Figure 4 shows the result of our experiments, which are aggregated over 20 trials. As baselines, we consider the uncoded scheme, as well as a replication scheme, where each uncoded partition is replicated β = 2 times across nodes, and the server uses the faster copy in each iteration. It can be seen from the right figure that one can speed up computation by reducing η from 1 to, for instance, 0.375, resulting in more than 40% reduction in the runtime. Note that in this case, uncoded L-BFGS fails to converge, whereas the Hadamard-coded case stably converges. We also observe that the data replication scheme converges on average, but in the worst case, the convergence is much less smooth, since the performance may deteriorate if both copies of a partition are delayed. 7 Figure 5: Test RMSE for m = 8 (left) and m = 24 (right) nodes, where the server waits for k = m/8 (top) and k = m/2 (bottom) responses. “Perfect" refers to the case where k = m. Figure 6: Total runtime with m = 8 and m = 24 nodes for different values of k, under fixed 100 iterations for each scheme. Matrix factorization on Movielens 1-M dataset. We next apply matrix factorization on the MovieLens-1M dataset [17] for the movie recommendation task. We are given R, a sparse matrix of movie ratings 1–5, of dimension #users × #movies, where Rij is specified if user i has rated movie j. We withhold randomly 20% of these ratings to form an 80/20 train/test split. The goal is to recover user vectors xi ∈ Rp and movie vectors yi ∈ Rp (where p is the embedding dimension) such that Rij ≈ xTi yj + ui + vj + µ, where ui , vj , and µ are user, movie, and global biases, respectively. The optimization problem is given by   X X X min (Rij − ui − vj − xTi yj − µ)2 + λ  kxi k22 + kuk22 + kyj k22 + kvk22  . xi ,yj ,ui ,vj i i,j: observed j (8) We choose µ = 3, p = 15, and λ = 10, which achieves a test RMSE 0.861, close to the current best test RMSE on this dataset using matrix factorization2 . Problem (8) is often solved using alternating minimization, minimizing first over all (xi , ui ), and then all (yj , vj ), in repetition. Each such step further decomposes by row and column, made smaller by the sparsity of R. To solve for (xi , ui ), we first extract Ii = {j \| rij is observed}, and solve the > resulting sequence of regularized least squares problems in the variables wi = [x> i , ui ] distributedly > > using coded L-BFGS; and repeat for w = [yj , vj ] , for all j. As in the first experiment, distributed coded L-BFGS is solved by having the master node encoding the data locally, and distributing the encoded data to the worker nodes (Appendix D discusses how to implement this step more efficiently). The overhead associated with this initial step is included in the overall runtime in Figure 6. The Movielens experiment is run on a single 32-core machine with 256 GB RAM. In order to simulate network latency, an artificial delay of ∆ ∼ exp(10 ms) is imposed each time the worker completes a task. Small problem instances (n < 500) are solved locally at the central server, using the built-in function numpy.linalg.solve. Additionally, parallelization is only done for the ridge regression instances, in order to isolate speedup gains in the L-BFGS distribution. To reduce overhead, we create a bank of encoding matrices {Sn } for Paley ETF and Hadamard ETF, for n = 100, 200, . . . , 3500, and then given a problem instance, subsample the columns of the appropriate matrix Sn to match the dimensions. Overall, we observe that encoding overhead is amortized by the speed-up of the distributed optimization. Figure 5 gives the final performance of our distributed L-BFGS for various encoding schemes, for each of the 5 epochs, which shows that coded schemes are most robust for small k. A full table of results is given in Appendix C. 2 http://www.mymedialite.net/examples/datasets.html 8 Acknowledgments This work was supported in part by NSF grants 1314937 and 1423271. References [1] G. Ananthanarayanan, A. Ghodsi, S. Shenker, and I. Stoica. Effective straggler mitigation: Attack of the clones. In NSDI, volume 13, pages 185–198, 2013. [2] A. Beck and M. Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM Journal on Imaging Sciences, 2(1):183–202, 2009. [3] A. S. Berahas, J. Nocedal, and M. Takác. A multi-batch l-bfgs method for machine learning. In Advances in Neural Information Processing Systems, pages 1055–1063, 2016. [4] E. J. Candes and T. Tao. Decoding by linear programming. IEEE Transactions on Information Theory, 51 (12):4203–4215, 2005. [5] E. J. Candes and T. Tao. Near-optimal signal recovery from random projections: Universal encoding strategies? IEEE Transactions on Information Theory, 52(12):5406–5425, 2006. [6] P. Drineas, M. W. Mahoney, S. Muthukrishnan, and T. Sarlós. Faster least squares approximation. Numerische mathematik, 117(2):219–249, 2011. [7] S. Dutta, V. Cadambe, and P. Grover. Short-dot: Computing large linear transforms distributedly using coded short dot products. In Advances In Neural Information Processing Systems, pages 2092–2100, 2016. [8] M. Fickus, D. G. Mixon, and J. C. Tremain. Steiner equiangular tight frames. Linear Algebra and Its Applications, 436(5):1014–1027, 2012. [9] S. Geman. A limit theorem for the norm of random matrices. The Annals of Probability, pages 252–261, 1980. [10] J. Goethals and J. J. Seidel. Orthogonal matrices with zero diagonal. Canad. J. Math, 1967. [11] C. Karakus, Y. Sun, and S. Diggavi. Encoded distributed optimization. In 2017 IEEE International Symposium on Information Theory (ISIT), pages 2890–2894. IEEE, 2017. [12] K. Lee, M. Lam, R. Pedarsani, D. Papailiopoulos, and K. Ramchandran. Speeding up distributed machine learning using codes. In Information Theory (ISIT), 2016 IEEE International Symposium on, pages 1143–1147. IEEE, 2016. [13] M. W. Mahoney et al. Randomized algorithms for matrices and data. Foundations and Trends R in Machine Learning, 3(2):123–224, 2011. [14] A. Mokhtari and A. Ribeiro. Global convergence of online limited memory BFGS. Journal of Machine Learning Research, 16:3151–3181, 2015. [15] R. E. Paley. On orthogonal matrices. Studies in Applied Mathematics, 12(1-4):311–320, 1933. [16] M. Pilanci and M. J. Wainwright. Randomized sketches of convex programs with sharp guarantees. IEEE Transactions on Information Theory, 61(9):5096–5115, 2015. [17] J. Riedl and J. Konstan. Movielens dataset, 1998. [18] J. W. Silverstein. The smallest eigenvalue of a large dimensional wishart matrix. The Annals of Probability, pages 1364–1368, 1985. [19] F. Szöllősi. Complex hadamard matrices and equiangular tight frames. Linear Algebra and its Applications, 438(4):1962–1967, 2013. [20] R. Tandon, Q. Lei, A. G. Dimakis, and N. Karampatziakis. Gradient coding. ML Systems Workshop (MLSyS), NIPS, 2016. [21] D. Wang, G. Joshi, and G. Wornell. Using straggler replication to reduce latency in large-scale parallel computing. ACM SIGMETRICS Performance Evaluation Review, 43(3):7–11, 2015. [22] L. Welch. Lower bounds on the maximum cross correlation of signals (corresp.). IEEE Transactions on Information theory, 20(3):397–399, 1974. [23] N. J. Yadwadkar, B. Hariharan, J. Gonzalez, and R. H. Katz. Multi-task learning for straggler avoiding predictive job scheduling. Journal of Machine Learning Research, 17(4):1–37, 2016. 9 A Lemmas In the proofs, we will ignore the normalization constants on the objective functions for brevity. Let ftA := eA w − yeA k2 (we set A ≡ At ). Let w eA wt − yeA k2 , and fA (w) := kX kX et∗ denote the solution to the effective ∗ eA w − yeA k2 . “instantaneous" problem at iteration t, i.e., w et = arg min kX w Stronger versions of the following lemma has been proved in [11, 16], but we include a weakened version of this result here for completeness. Lemma 2. For any t and {At }, f (w et∗ ) ≤ κ2 f (w∗ ) Proof. Define e = w et∗ − w∗ and note that kX w et∗ − yk = kXw∗ − y + Xek ≤ kXw∗ − yk + kXek by triangle inequality, which implies ∗ ∗ 2 f (w et ) = kX w et − yk ≤ 1 + kXek kXw∗ − yk 2 kXw − yk = 1 + ∗ 2 kXek kXw∗ − yk 2 f (w∗ ). (9) Now, for any c > 0, consider > SA (Xw∗ − y) e> X > SA kSA Xek2 (a) ≤ −2 λmin λmin > > > e X SA SA − cI (Xw∗ − y) 2c > > = −2 − e X (Xw∗ − y) λmin λmin > SA − cI (Xw∗ − y) e> X > SA (b) = −2 λmin > (c) SA e> X > cI − SA kXw∗ − yk ≤2 λmin > (d) SA cI − SA ≤ 2 kXw∗ − ykkXek, λmin kXek2 ≤ eA w where (a) follows by expanding and re-arranging X et∗ − yeA 2 eA w∗ − yeA ≤ X 2 , which is since w et∗ is the minimizer of this function; (b) follows by the fact that ∇f (w∗ ) = X > (Xw∗ − y) = 0 by optimality of w∗ for f ; (c) follows by Cauchy-Schwarz inequality; and (d) follows by the definition of matrix norm. Since this is true for any c > 0, we choose c = λmax +λmin , 2 which gives kXek λmax − λmin ≤ = κ − 1. kX w et∗ − yk λmin Plugging this back in (9), we get f (w e∗ ) ≤ κ2 f (w∗ ), which completes the proof. Lemma 3. If A fet+1 − feA (w̄) ≤ γ fetA − feA (w̄) for all t > 0, for some w̄ ∈ Rp , and for some 0 < γ < 1, then ft ≤ (κγ)t f0 + κ−γ f (w̄) . 1 − κγ Proof. Since for any w, >e λmin f (w) = λmin kXw − yk2 ≤ (Xw − y)> SA SA (Xw − y) = feA (w), and similarly feA (w) ≤ λmax f (w), we have λmin ft+1 − λmax f (w̄) ≤ γ (λmax ft − λmin f (w̄)) , which can be re-arranged into the linear recursive inequality ft+1 ≤ κγft + (κ − γ)f (w̄) , 10 where κ = we get λmax . λmin By considering such inequalities for 0 ≤ τ ≤ t, multiplying each by (κγ)t−τ and summing, ft ≤ (κγ)t f0 + (κ − γ)f (w̄) t−1 X (κγ)τ τ =0 κ−γ f (w̄) . ≤ (κγ) f0 + 1 − κγ t Lemma 4. feA (w) is λmin µ-strongly convex. > e eA Proof. It is sufficient to show that the minimum eigenvalue of X XA is bounded away from zero. This can easily be shown by the fact that > e > eA u> X XA u = u> X > SA SA Xu ≥ λmin kXuk2 ≥ λmin µkuk2 , for any unit vector u. Lemma 5. Let M ∈ Rp×p be a positive definite matrix, with the condition number (ratio of maximum eigenvalue to the minimum eigenvalue) given by κ. Then, for any unit vector u, √ u> M u 2 κ ≥ . kM uk κ+1 Proof. Let W be the subspace spanned by {u, M u}, and let W ∈ Rp×2 be a matrix whose columns form an orthonormal basis for W. Then u and M u can be represented as u = W r1 M u = W r2 for some r1 , r2 ∈ R2 , which implies r2 = W > M W r1 = Q> DQr1 , where since W > M W is still a positive definite matrix it has the 2 × 2 eigen-decomposition Q> DQ. Defining qi = Qri for i = 1, 2, note that the quantity we are interested in can be equivalently represented as u> M u r > r2 q > q2 = 1 = 1 , kM uk kr2 k kq2 k where q2 = Dq1 . Further note that for any unit vector v, v > Dv = v > Q> DQv = v > W > M W v, and since kW vk = 1, the condition number of D (the ratio of the two non-zero elements of D) cannot be larger than that of M , which is κ (since otherwise once could find unit vectors u1 = W v1 and u2 = W v2 such that u> 1 M u1 > κ, which is a contradiction). Representing q1 = [cos φ sin φ]> for some angle φ, kqq22 k can then be u> 2 M u2 > > q1 q2 written as q2 = √ d21 2 cos φ √ d22 2 sin φ . Note that minimizing the inner product kq is equivalent 2k d1 +d2 d1 +d2 to maximizing the function d2 − 1 tan φ d1 d2 −1 tan tan tan φ − φ = d1 1 + dd21 tan2 φ 1 over φ. By setting the derivative to zero, we find that the maximizing φ is given by cos−1 dd22 −d . Therefore +d1 q √ 2 dd21 u> M u q1> q2 d1 d2 2 κ 2 2 = = p cos φ + p sin φ ≥ ≥ , kM uk kq2 k 1+κ 1 + dd2 d21 + d22 d21 + d22 1 which is the desired result. 11 Proof of Lemma 1. First note that > rt> ut = X > S̆t> S̆t [(Xwt − y) − (Xwt−1 − y)] (wt − wt−1 ) = (wt − wt−1 )> X > S̆t> S̆t X (w − wt−1 ) ≥ µkut k2 , (10) by (5) Also consider 2 (wt − wt−1 )> X > S̆t> S̆t X (wt − wt−1 ) krt k2 , = rt> ut (wt − wt−1 )> X > S̆t> S̆t X (wt − wt−1 ) which implies µ ≤ krt k2 ≤ (1 + )M, rt> ut again by (4). Now, setting j` = t − σ e + `, consider the trace tr (`+1) Bt = tr (`) Bt (`) − tr (`) + tr ≤ tr Bt (`) Bt uj` u> j` Bt ! (`) u> j` Bt uj` ! rj` rj>` rj>` uj` + tr rj` rj>` rj>` uj` ! kr k2 j (`) + >` = tr Bt rj` uj` (`) + (1 + )M, ≤ tr Bt which implies tr (Bt ) ≤ (1 + )M (e σ + d). It can also be shown (similar to [3]) that (`) (`+1) · = det Bt det Bt rj>` uj` (`) u> j` Bt uj` r > uj kuj` k2 j (`) · ` `2 · = det Bt (`) kuj` k u> j` Bt uj` µ (`) ≥ det Bt , (1 + )M (e σ + d) σe (0) µ which implies det (Bt ) ≥ det Bt . Since Bt ≥ 0, its trace is bounded above, and its (1+)M (e σ +d) determinant is bounded away from zero, there must exist 0 < c1 ≤ c2 such that c1 I Bt c2 I. B Proofs of Theorem 1 and Theorem 2 Throughout the section, we will consider a particular iteration t, and denote n o > > λmin := min λmin SA SA , λmin SD SD n o > > λmax := max λmax SA SA , λmax SD SD , where λmin (·) and λmax (·) denote the minimum and maximum eigenvalues of a matrix. We will also denote with w et∗ the solution to the effective “instantaneous" problem at iteration t, i.e., w et∗ = 2 eA w − yeA , where A ≡ At , and we ignore the normalization constants on the objective functions arg min X w eA w − yeA k2 , and fetA := kX eA wt − yeA k2 . for. Finally, we define feA (w) := kX 12 B.1 Proof of Theorem 1 Using convexity, and the choices that dt = −e gt and αt = α, we have 1 > > feA (wt − αt dt ) − feA (wt ) ≤ αt get> dt + αt2 d> t X SA SA Xdt 2 (a) 1 λmax 2 > > > = −αke gt k2 + α2 get> X > SA SA Xe gt ≤ −αke gt k2 + α get X Xe gt 2 2 (b) 2ζ (1 − ζ) λmax M ≤ −α 1 − α ke g t k2 = − ke gt k2 2 M λmax (c) 4µζ (1 − ζ) eA ≤− f (wt ) − feA (w et∗ ) , M λmax > where (a) follows by the fact that SA SA λmax I; (b) follows since X > X M I; and (c) follows by strong convexity. Re-arranging this inequality, and using the definition of γ1 , we get A fet+1 − feA (w et∗ ) ≤ γ1 fetA − feA (w et∗ ) . Then, Lemma 3 with w̄ = w et implies ft ≤ (κγ1 )t f0 + κ − γ1 f (w et∗ ) . 1 − κγ1 Finally, Lemma 2 implies f (w et∗ ) ≤ κ2 f (w∗ ), which concludes the proof. B.2 Proof of Theorem 2 Using convexity and the closed-form expression for the step size, we have 1 > > feA (wt + αt dt ) − feA (wt ) ≤ αt get> dt + αt2 d> t X SA SA Xdt 2 2 2 > > ν 2 get> dt ν get> dt 1 d> t X SA SA Xdt + · =− > > > > > > > 2 dt X > SD SD Xdt dt X > SD SD Xdt dt X SD SD Xdt ! ! > 2 > 2 > > > > > 2 (a) SD − ν2 SA dt X ν SA SA − 2νSD SD Xdt z > SD SA z dt get > = dt get = −ν 2 2 2 > > > > > kXd tk 2 dt X SD SD Xdt z SD SD z > 2 (c) > 2 (b) dt get dt get λmin − ν2 λmax λmin − ν2 λmax ν ≤ − ≤ −ν 2 2 λmin kXdt k2 M λmin kdt k2 2 > ν get Bt get λmin − 2 λmax (d) ν = − M λ2min kBt get k2 ν ν (e) (f) 4ν λmin − 2 λmax c1 c2 c1 c2 8µν λmin − 2 λmax 2 e e et∗ ) ≤− ke g k ≤ − t 2 2 f (wt ) − f (w 2 2 M λmin M λmin (c1 + c2 ) (c1 + c2 ) (h) (g) 4µc1 c2 et∗ ) . = − fe(wt ) − fe(w et∗ ) = − (1 − γ2 ) feA (wt ) − feA (w M (c1 + c2 )2 Xdt ; (b) follows by the fact that the term in parenthesis is an increasing kXdt k > S̆t S̆t z and by Assumption 1; (c) follows by the assumption that X > X M I; where (a) follows by defining z = > function of the quadratic form z e (d) follows by the definition of dt ; (e)follows by Lemmas 5 and 1; (f) follows by strong convexity of f (by min Lemma 4), which implies ke gt k2 ≥ 2µ fe(θt ) − fe(w et∗ ) ; (g) follows by choosing ν = λλmax ; and (h) follows using the definition of γ2 . Re-arranging the inequality, we obtain A fet+1 − feA (w et∗ ) ≤ γ2 fetA − feA (w et∗ ) , and hence applying first Lemma 3 with w̄ = w et , and then Lemma 2, we get the desired result. C Full results of Movielens 1-M experiment Tables 1 and 2 give the test and train RMSE for the Movielens 1-M recommendation task, with a random 80/20 train/test split. 13 uncoded replication gaussian paley hadamard m = 8, k = 1 train RMSE 0.804 0.783 0.781 0.775 0.779 test RMSE 0.898 0.889 0.877 0.873 0.874 runtime 1.60 1.76 2.24 1.82 1.82 m = 8, k = 4 train RMSE 0.770 0.766 0.765 0.763 0.765 test RMSE 0.872 0.872 0.866 0.868 0.870 runtime 2.96 3.13 3.64 3.34 3.18 m = 8, k = 6 train RMSE 0.762 0.760 0.762 0.758 0.760 test RMSE 0.866 0.871 0.864 0.860 0.864 runtime 5.11 4.59 5.70 5.50 5.33 Table 1: Full results for Movielens 1-M, distributed over m = 8 nodes total. Runtime is in hours. An uncoded scheme running full batch L-BFGS has a train/test RMSE of 0.756 / 0.861, and a runtime of 9.58 hours. uncoded replication gaussian paley hadamard m = 24, k = 3 train RMSE 0.805 0.791 0.783 0.780 0.782 test RMSE 0.902 0.893 0.880 0.879 0.882 runtime 2.60 3.22 3.98 3.49 3.49 m = 24, k = 12 train RMSE 0.770 0.764 0.767 0.764 0.765 test RMSE 0.872 0.870 0.866 0.868 0.868 runtime 4.24 4.38 4.92 4.50 4.61 Table 2: Full results for Movielens 1-M, distributed over m = 24 nodes total. Runtime is in hours. An uncoded scheme running full batch L-BFGS has a train/test RMSE of 0.757 / 0.862, and a runtime of 14.11 hours. D Efficient encoding using Steiner ETF We first describe Steiner ETF, based on the construction proposed in [8]. Steiner equiangular tight frames. Let v be a power of 2, let H ∈ Rv×v be a real Hadamard matrix, and let hi be the ith column of H, for i = 1, . . . , v. Consider the matrix V ∈ {0, 1}v×v(v−1)/2 , where each column is the incidence vector of a distinct two-element subset of {1, . . . , v}. For instance, for v = 4,   1 1 1 0 0 0  1 0 0 1 1 0  V = . 0 1 0 1 0 1  0 0 1 0 1 1 Note that each of the v rows have exactly v − 1 non-zero elements. We construct Steiner ETF S as a v 2 × v(v−1) 2 √ matrix by replacing each 1 in a row with a distinct column of H, and normalizing by v − 1. For instance, for the above example, we have   h2 h3 h4 0 0 0 1  h2 0 0 h3 h4 0  S= √  . 0 h2 0 h3 0 h4  3 0 0 h2 0 h3 h4 v In general, this procedure results in a matrix S with redundancy factor β = 2 v−1 . In full generality, Steiner ETFs can be constructed for larger redundancy levels; we refer the reader to [8] for a full discussion of these constructions. Efficient distributed encoding. Steiner ETF allows for a distributed and efficient implementation for encoding a given matrix X. Note that the encoding matrix S consists of v blocks, each corresponding to a row of V . Consider the following partition for S: > S = S1> S2> . . . Sv> , 14 where Si ∈ Rv×v(v−1)/2 is a horizontal block. Note that multiplication with a block Si X can be computed by directly finding the non-zero column indices for the block Si , which is given by B1,i ∪ B2,i , where 1 B1,i = j(2v − 1 − j) + (i − 1 − j), for j = 0, . . . , i − 2 2 1 B2,i = (i − 1)(2v − 1 − i) + j, for j = 1, . . . , v − i 2 Multiplication can then be implemented by simply taking a Hadamard transform of the corresponding rows of X, whose indices are given by B1,i ∪ B2,i . In the case of dimension mismatch, one can append zero rows to X, or remove some rows from S to make the multiplication well-defined. Therefore, one can partition the blocks {SiS } across worker nodes, and worker node k can read the corresponding rows of X from the pool of data, given by iinIk B1,i ∪ B2,i , where Ik is the blocks assigned to worker k, and then apply Fast Hadamard Transform for each block. Note that processing of blocks can be further parallelized within each node, using multiple cores. In practice, we have observed that the performance of Steiner ETF significantly improves if the rows of SX are shuffled after encoding. This can be implemented by having the nodes exchange rows of SX after encoding; however, this incurs a significant communication cost. A more practical approach could be one where each block is assigned to multiple nodes at random, i.e., each block is encoded by multiple nodes. The worker nodes can then drop a subset of their encoded rows such that each row is retained by exactly one node, based on some pre-defined row-allocation rule. Note that this has the same effect as having the nodes randomly exchange encoded rows with each other. Space complexity. One might raise the point that many large-scale datasets are sparse, and this sparsity is lost after encoding, significantly increasing the memory usage. To address this issue, first consider the case where each node has access to the entire dataset (X, y), for the sake of argument. Then a worker node would compute the gradient corresponding to block i using the order of operations represented by the following parenthesization: gi (wt ) = X > Si> (Si (Xw − y)) , since this would only require matrix-vector multiplications. Now, note that for a given block Si , there are only v − 1 non-zero rows, out of v(v − 1)/2. Therefore, in order to compute gi (wt ) in the above order of operations, one would only need to store the rows of X and y corresponding to these non-zero rows, and then apply the v transformation corresponding to Si whenever needed. If each node is assigned m blocks, this would require v(v−1) storing at most m sparse rows, which means that the memory usage would only increase by a constant factor (on the order of redundancy factor). 15	7cs.IT
An Effective AHP based Metaheuristic Approach to Solve Supplier Selection Problem Tamal Ghosh1, Tanmoy Chakraborty, Pranab K Dan Department of Industrial Engineering & Management, West Bengal University of Technology. BF 142, Sector 1, Salt Lake City, Kolkata 700064, India. Email: tamal.31@gmail.com, chakrabortytanmoy.85@gmail.com, danpk.wbut@gmail.com. Abstract The supplier selection problem is based on electing the best supplier from a group of prespecified candidates, is identified as a Multi Criteria Decision Making (MCDM), is proportionately significant in terms of qualitative and quantitative attributes. It is a fundamental issue to achieve a trade-off between such quantifiable and unquantifiable attributes with an aim to accomplish the best solution to the abovementioned problem. This article portrays a metaheuristic based optimization model to solve this NP-Complete problem. Initially the Analytic Hierarchy Process (AHP) is implemented to generate an initial feasible solution of the problem. Thereafter a Simulated Annealing (SA) algorithm is exploited to improve the quality of the obtained solution. The Taguchi robust design method is exploited to solve the critical issues on the subject of the parameter selection of the SA technique. In order to verify the proposed methodology the numerical results are demonstrated based on tangible industry data. Keywords: supplier rating; vendor section; AHP; simulated annealing; metaheuristic; Taguchi method. 1. Introduction Over the decades the supplier selection problem has been dragging the attention of the researchers and practitioners in the vicinity of Supply Chain Management study. Therefore several techniques have been proposed by the academicians and professionals to solve this Multi Criteria Decision Making (MCDM) problem. As Dickson (1966) stated that contemporary firms have the alternatives to select a suitable technique to decipher the complicated job of selecting suppliers. These techniques varies from the adoption of uncomplicated methods such as the selection of the vendor offering lowermost tender to the effective complicated methods which exploit the unquantifiable attributes of the supplier selection problem. Although many empirical studies are already contributed in the proposed area of research, yet the researchers and practitioners are involved in improving the solution methodologies due to the involvement of growing complexities of technology innovation. In manufacturing industry the procurement cost of raw materials from the outside suppliers and the other unquantifiable attributes such as quality of the materials delivered and the time of delivery are substantially crucial and nearly 70% to 80% of the total cost involved in the whole product development phase (Weber, 1 Corresponding Author: Email: tamal.31@gmail.com, Phone: +91-33-2334-1014, Fax: +91-33-2334-1031. 1 Current and Benton, 1991). Since the early 1990’s a set of hard to estimate unquantifiable attributes have turned out to be the key components of the supplier selection problem which further helped in developing novel and complicated methodologies (Ellram, 1990). Selecting an optimal set of suppliers on this basis certainly increase the sustainability of the firms in present global competition (Thompson, 1990). Traditionally the supplier selection problem is based on the purchasing cost, quality and delivery functionality of the raw materials. Consideration of the above aspects enhances the complexities of the problem in polynomial time. The selection decision becomes extensively significant for any person-in-charge of the procurement department due to the several different levels of success of the rival suppliers under these stated circumstances (Aissaoui, Haouari and Hassini, 2007). The vendor with the lowermost tender might not be the best in delivering the material of proper quality. In case of the government bodies, the practice of decision-making optimization model in vendor selection processes would be extremely helpful than the private firms. In the government framework, there is a legal need to obey the formal rules and procedures that control vendor selection processes. Therefore such decision models could be immensely supportive in maintaining transparency and doing fair business (Schooner, 2003). It is evident from the past literature that the precise and sophisticated methodologies are urgently required to solve the supplier selection problems (Scott, 1995). For that matter Kanagaraj and Jawahar (2009) have implemented a Simulated Annealing Algorithm (SAA) tool for supplier selection problems to obtain the optimal or near-optimal solutions quickly. Saen (2009) introduced a technique based on Data Envelopment Analysis (DEA) to rank the suppliers in the presence of nondiscretionary factors. An integrated model of Fuzzy-AHP and decision support system was proposed by Gnanasekaran, Velappan and Manimaran (2010) for supplier selection to reduce the cost and enhance the product quality with the help of the ReliabilityBased Total Cost of Ownership (RBTCO) model. This integrates purchasing, maintenance and stoppage costs along with the realistic constraints based on product reliability and weight restraint. Nonlinear Integer Programming (NLIP) is used to develop the mathematical formulation of the RBTCO model. In present article a multiobjective supplier selection problem has been introduced which has incorporated faulty materials and delay in delivery cost as decisive factors. Thereafter a Simulated Annealing (SA) based optimization model is implemented to solve this problem. The initial solution to the SA procedure is generated using Analytic Hierarchy Process (AHP). In order to select the optimal set of parameters to the SA, Taguchi’s Design of Experiments (DOE) technique is exploited. The rest of the article is structured in following manner, section 2 presents a brief literature survey of the proposed area of study, section 3 demonstrates the formulation of the proposed multiobjective problem model, section 4 elaborates the proposed solution methodology and numerical results are depicted in section 5 followed by the managerial implications and conclusion of this research. 2. Literature Survey Li, Fun and Hung (1997) reported that the supplier selection methodologies are practiced to facilitate the selection process which further have a substantial impact on the selection outcomes. Numerous supplier selection methods have been established and categorized over the decades. Petroni and Braglia (2000) proposed linear weighting method using supplier rating based on different attributes which is a faster and inexpensive method to instigate. Although various drawbacks and limitations are also indicated in their study. Cost proportion (Timmerman, 1986) and aggregate cost of proprietorship (Ellram, 1990) based on aggregate cost approaches assemble all the cost components of the supplier selection process in fiscal units which are very flexible methods. These are exact methods and due to the complexities and time involvement these methods are moderately expensive to implement. Mathematical programming approaches primarily exploit the quantifiable factors; these approaches 2 comprise the Principal Component Analysis (PCA) (Petroni and Braglia, 2000) and the Artificial Neural Network (ANN) (Choy, Lee and Low, 2002). Bello (2003) stated in his research that the PCA approach is advantageous in terms of its competency in managing various differing aspects. The ANN approaches are also useful in cost minimization and time reduction. Multiple Attribute Utility Theory method is practiced for global vendor selection problems, in which the surroundings are more complex and uncertain (Zhao and Bross, 2004). Chen, Lin and Huang (2006) demonstrated that the Fuzzy set theory technique controls such situations for supplier performance evaluation. The approach is helpful to the managers to purchase from the suppliers according to their own choice. a fuzzy weighted additive and mixed integer linear programming is developed. The model aggregates weighted membership functions of objectives to construct the relevant decision functions, in which objectives have different relative importance. Ng (2008) proposed a weighted linear program for the supplier selection problem. This paper demonstrates a transformation technique with the mathematical model which can solve the problem without an optimizer. Amid, Ghodsypour and O’Brien (2009) demonstrated fuzzy multiobjective and additive model to consider the imprecision of information along with the order quantities to each supplier using price breaks. The objective functions used are, minimizing the net cost, rejected items and late deliveries, while satisfying other constraints such as capacity and demand requirement. Authors further stated weighted max–min model in the similar problem environment (Amid, Ghodsypour and O’Brien, 2010) with the help of Analytic Hierarchy Process (AHP). The proposed model could be utilized to find out the appropriate order to each supplier. Saen and Zohrehbandian (2008) proposed DEA approach with quantity discount policy to select the best supplier. Wu (2009) stated a hybrid model using data envelopment analysis (DEA), decision trees (DT) and neural networks (NNs) to evaluate supplier performance. The model comprised of two parts. The first part utilized DEA and categorizes the suppliers into several clusters thereafter the second part used the performance-related data of the company to train decision trees and intelligent neuro model, the precised results are obtained. However Saen (2008) stated an approach based on super-efficiency analysis based on DEA to rank the suppliers in presence of volume discounts. Author further proposed another DEA based method to rank the suppliers in the presence of weight limitations and dual-role factors (Saen, 2010). The Analytical Hierarchical Process (AHP) is the most exploited method in supplier selection problems, which is a Multi Criteria Decision Making technique. It is implemented to rank the alternatives when several criteria are believed to be considered and it permits the managers to formulate complicated problems in the form of a hierarchical relationship (Saaty, 1980). The AHP is comparatively straightforward method to practice. This technique integrates tangible and intangible attribute of the problems. A detail survey (Tahriri et al., 2008) of the supplier selection methods portrays that the AHP is the most frequently employed method in supplier selection. The AHP hierarchy typically contains three distinct levels, which are objectives, factors, and alternatives. AHP suggests an way to rank the alternative choices based on the manager’s decisions relating the significance of the criteria. Due to this fact AHP is preferably appropriate for the abovementioned problem. The problem hierarchy provides itself to an analysis based on the impact of a given level on the next higher level (Saaty, 1980). Managerial judgments are stated in terms of the pair-wise comparisons of entries on a specified level of the hierarchy based on their influences on the next higher level. Each of the pair-wise comparisons signifies an approximation of the proportion of the weights of the two criteria being compared. Since AHP exploits a proportionate scale for personal decisions, the relative weights reflect the relative importance of the norms in attaining the objective of the hierarchy (Tam and Tummala, 2001). Yahya and Kingsman (1990) used Saaty’s AHP method to determine primacy in selecting suppliers. The authors employed vendor rating in determining how to allot business and where inadequate progress work is utilized. Akarte (2001); Handfield, Walton and 3 Sroufe (2002); Yu and Jing (2004); Liu and Hai (2005); Rajkumar, Kannan and Jayabalan (2009) also utilized AHP technique in their study as an integral part. Most of the supplier selection literature generally used traditional methods. Traditional techniques are not efficient when the solution state space is large and various constraints cause the vendor selection problem more complicated. Very few articles utilized state-of-the-art methods such as metaheuristics. Recently Arunkumar, Karunamoorthy and Makeshwaraa, (2007); Rezaei and Davoodi (2008); Kubat, and Yuce, (2006); Ding, Benyoucef and Xie (2003) have proposed Genetic Algorithm based metaheuristic approach to solve supplier selection problems in multiobjective environment. This present research introduces an efficient metaheuristic approach based on Simulated Annealing to solve the supplier selection problem. 3. Problem Formulation Supplier selection processes are primarily reliant on the specific objectives being solved with the problem and the relevant constraints related to the objectives. In this article a case of a Leading construction firm of India has been considered to derive the multiobjective optimization model. This firm constructs commercial buildings as well as residential units in large scale such as IT/ITeS SEZ, Shopping malls, hospitality and retail units, logistics and industrial squares etc. Due to the heavy construction approach, this company requires various raw materials such as steel beams, cement, light weight bricks, cast iron etc. The proposed model is articulated considering the constrictions which are managed rationally by this firm to select suppliers. Each raw material supplied by the supplier would have different constraints and characteristics such as percentage of faulty materials supplied, percentage delay in delivery and unit purchasing cost of the materials. Each supplier certainly has its own capacity to supply. The firm has specific requirements of material in certain period. If raw material j is being supplied by supplier k, then the initial procurement cost is defined as, 𝑀 𝑁 𝑃𝑐 = ∑ ∑ 𝜑𝑗𝑘 𝑞𝑗𝑘 (1) 𝑗=1 𝑘=1 where φjk is the unit cost of jth material supplied by kth supplier. qjk is the amount of material type j procured by ith supplier. The total substandard delivery is defined as, 𝑀 𝑁 𝑄𝑑 = ∑ ∑ 𝛿𝑗𝑘 𝑞𝑗𝑘 (2) 𝑗=1 𝑘=1 where δjk is the percentage of faulty items of jth material supplied by kth supplier. qjk is the amount of material type j procured by ith supplier. The total delay in delivery is defined as, 4 𝑀 𝑁 𝐷𝑑 = ∑ ∑ 𝜃𝑗𝑘 𝑞𝑗𝑘 (3) 𝑗=1 𝑘=1 where θjk is the delay percentage in delivery of jth material supplied by kth supplier. qjk is the amount of material type j, procured by ith supplier. In this supplier selection problem three objectives are simultaneously considered, (i) to minimize the total procurement cost, (ii) to minimize the total number of faulty items supplied and (iii) to minimize the total number of delay days in delivery in procuring various raw materials from various suppliers. The quality function (2) demonstrates the number of faulty items supplied by the suppliers. The faulty items are generally detected by the receiving firm while relocating the raw materials to the inventory. The firm’s strategy is to return back the substandard items to the suppliers and request them to replace those items within stipulated time which is substantially one week. Therefore in a project under fixed schedule, supply of substandard material could incur the total cost in terms of one week delay time. Thus the quality function of equation (2) and total delay in delivery of equation (3) could be transformed into a non-compliance cost component using, 𝑄𝑐 = (7 + 𝐷𝑑 ) × 𝐶𝑑 (4) Hence the proposed multiobjective model aims in minimizing the total cost function f(x), 𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑓(𝑥) = 𝑃𝑐 + 𝑄𝑐 (5) subject to 𝑞𝑗𝑘 ≤ 𝑋𝑗𝑘 (6) ∑ 𝑞𝑗𝑘 ≥ 𝑌𝑗 (7) 𝑁 𝑘=1 where Xjk is the capacity of supplier k while supplying material j Yjk is the demand of material j to the firm in certain period Constraint (6) confirms that each supplier supplies according to its capacity. Constraint (7) confirms that total raw material procured should harmonize the firm’s demand. This proposed multiobjective model is validated using simulated annealing (SA) approach by considering these abovementioned objectives and constraints which further gives the optimal selection results of suppliers. In order to define the SA approach, a quick solution generation method is believed to be identified and the generated feasible solution is assumed to be used as the initial solution to the SA algorithm. Therefore in this research Analytic Hierarchy Process (AHP) is utilized to serve the purpose. The next section would demonstrate the details of the solution methodologies involved in this approach. 5 4. Research Methodology To facilitate the present research work authors have visited an eminent construction firm operating in Kolkata, India. Currently the firm is involved in a big project based on development of an IT/ITeS Special Economic Zone (SEZ) in sector V, Rajarhat, Newtown in Salt Lake City Kolkata. Authors have prepared the questionnaires based on the information gathered from the professionals of the abovementioned firm. On the basis of the experts opinion the AHP analysis has been carried out in this study. The proposed optimization methodology is developed using Simulated Annealing (SA) algorithm to achieve the optimal solutions. The SA algorithm simulates the natural annealing process, where particles of a solid organize themselves into a thermal equilibrium. An introduction to SA can be obtained from the book by Aarts and Korst (1990). The general applications concerns combinatorial optimization problems of the following form where S is a predetermined set of feasible solutions. minx∈S g(x) (8) The algorithm exploits a pre-defined neighbourhood structure on ‘S’. A control parameter called temperature in resemblance to the natural annealing process governs the search behaviour. In each iteration, a neighbour solution y to the current solution x is figured out. If y has an improved objective function value than x, the solution y is accepted, that is, the current solution x is swapped by y. Alternatively if y does not attain a better objective function value than x, the solution y is only recognized with a specific probability depending on (i) the difference of the objective function values in x and y, and (ii) the temperature parameter. The pseudocode below exhibits the general SA method. Pseudocode (SA) initialize; repeat generate a candidate solution; evaluate the candidate; determine the current solution; reduce the temperature; until termination condition is met; In this article, factors which affect the decision makers’ strength are analysed. The process follows these steps,  Using the AHP method, the weight of the factors are obtained from qualitative expressions and each supplier’s weight is also achieved and final composite criteria weights are determined by AHP. This further rank the suppliers according to their composite scores. Therefore according to the ranks the suppliers could be selected.  Simulated Annealing (SA) takes the ranks obtained from previous step and attempts to search for optimal set of suppliers based on the objectives defined in section 3. 4.1. Initial Solution Generation In this article AHP is exploited to select initial set of suppliers. This solution is an initial feasible solution to the SA method based optimization model. 6 4.2. Analytical Hierarchy Process (AHP) The AHP is a multi-criteria decision technique that exploits hierarchical relationships to represent a problem. Primacies for substitutes are acquired based on the opinion of the experts (Saaty, 1980). The method consists of several important steps: outlining the shapeless problem into shape, obtaining the AHP hierarchical relationships, forming pairwise comparison matrices, approximating the relative weights, examining the consistency and finally attaining the overall ranking (Lee, Chen and Chang, 2008). AHP can empower managers to represent the interface of several factors in complicated and shapeless circumstances. The technique is based on the pairwise comparison of decision variables with respect to the factors or substitutes. A pairwise comparison matrix is obtained of size n×n, where n is the number of criteria to be compared. The AHP method is adopted in this article is stated as, Step 1: The hierarchical relationship of the problem is obtained and presented in Figure 1. The proposed AHP method decomposes the problem into three levels (Saaty, 1980): (i)The first level demonstrates the main objective: the selection of suppliers, the second level depicts the criteria and the last level reports the six suppliers to be compared. Figure 1. Hierarchical relationship of the supplier selection problem Step 2: Calculation of the pairwise comparison matrix for each level is required. For the pairwise comparison, a ranking scale is used for the criteria evaluation (Saaty, 1980). The scale is a crisp scale ranging from 1 to 9, as presented in Table 1. This scale values are assigned to the criteria based on the experts opinion of the AHP questionnaire sheet prepared by the authors. The pairwise comparison matrix for all the criteria is presented in Table 2. The last column of table 2 depicts the overall importance of the criteria over each other. The AHP procedure is presented as, Consider [Ax = λmaxx] where A is the comparison matrix of size n×n, for n criteria. x is the Eigenvector of size n×1 λmax is the Eigenvalue, λmax  > n. To find the ranking of priorities, namely the Eigen Vector X: 1. 1.1. 1.2. 1.3. Initialization: Take the squared power of matrix A, i.e., A2=A×A Find the row sums of A2 and normalize this array to find E0. Set A:=A2 7 2. 2.1. 2.2. 2.3. 2.4. 2.5. 2.6. Main: Take the squared power of matrix A, i.e., A2=A×A Find the row sums of A2 and normalize this array to find E1. Find D= E1 - E0. If the elements of D are close to zero, then X= E1 else set A:=A2 , set E0:=E1 and go to Step 2.1. STOP. Table 1. Saaty’s 9 point scale Importance Value 1 3 5 7 9 2, 4, 6, 8 Definition Equal Strong Moderate Strong Fairly Strong Very Strong Absolute Strong Intermediate Values Description Two factors are equally Contributing to objective One factor is marginally superior over other One factor is strongly superior over other One factor is very stongly superior over other The highest level of superiority of one factor over other According to the negotiation required Table 2. Comparison matrix of main attributes Cost Quality Delivery Cost 1 5 9 Quality 1/5 1 3 Delivery 1/9 1/3 1 Relative Weight (PV) 0.062941 0.265433 0.671625 Table 3 to Table 5 present the pairwise comparison matrices of the suppliers with each of the criteria selected in this research. The last columns of pairwise matrices present the calculated relative weights of the suppliers over each other. Table 3. Pairwise comparison matrix of suppliers with respect to cost S1 S2 S3 S4 S5 S1 1 1/3 1/2 1/7 1/6 S2 3 1 2 1/5 1/4 S3 2 1/2 1 1/3 1/3 S4 7 5 4 1 2 S5 6 4 3 1/2 1 S6 5 3 2 1/4 1/3 Relative Weight 0.40201 0.18549 0.21024 0.03831 0.05611 S6 1/5 1/3 1/2 4 3 1 0.10786 Table 4. Pairwise comparison matrix of suppliers with respect to quality S1 S2 S3 S4 S5 S6 S1 1 1/3 1/5 3 1/7 1/8 S2 3 1 1/4 4 1/5 1/7 S3 5 4 1 5 1/3 1/4 S4 1/3 1/4 1/5 1 1/8 1/9 S5 7 5 3 8 1 1/3 S6 8 7 4 9 3 1 Relative Weight 0.26949 0.15914 0.07354 0.43202 0.04087 0.02493 8 Table 5. Pairwise comparison matrix of suppliers with respect to delivery S1 S2 S3 S4 S5 S6 S1 1 1/5 1/2 3 1/6 1/7 S2 5 1 1/2 4 1/3 1/4 S3 2 2 1 5 1/3 1/5 S4 1/3 1/4 1/5 1 1/7 1/9 S5 6 3 3 7 1 1/4 S6 7 4 5 9 4 1 Relative Weight O.26324 0.11631 0.10426 0.43627 0.05268 0.02723 The overall rating of each supplier is computed by adding the product of the relative weight of each criterion and the relative weights of the suppliers considering the corresponding criteria (Table 6). Table 6 demonstrates that supplier 4 (overall ranking is 1 and composite score is 0.4099) is the best supplier followed by supplier 1, 2, 3, 5 and 6. 4.3 Computation of consistency index and consistency ratio AHP procedure requires the computation of consistency ratio to ensure the precision of the obtained solution. The consistency index (CI) of pairwise comparison matrix is calculated using, 𝐶𝐼 = 𝜆𝑚𝑎𝑥 − 𝑛 𝑛−1 (9) and the random consistency index (RI) is computed as, 𝑅𝐼 = 1.98 𝑛−2 𝑛 (10) where λmax is the maximum eigenvalue and n is the size of pairwise comparison matrix. Thus the consistency ratio (CR) is obtained using, 𝐶𝐼 𝐶𝑅 = (11) 𝑅𝐼 CI and CR are computed to understand the consistency of the solution obtained. In general CI and CR are believed to be < 0.1. Table 7 provides the computed values of the CI and CR for all the pairwise comparison matrices. All the computed values are < 0.1. Therefore the computed results are acceptable. Table 6. Composite score matrix from above four matrices S1 S2 S3 S4 S5 S6 Cost (PV=0.0629) 0.402014 0.18549 0.210246 0.0383076 0.0561086 0.107865 Quality (PV=0.2654) 0.26949 0.159145 0.0735405 0.432023 0.0408689 0.0249321 Delivery (PV=0.6716) 0.263243 0.11631 0.104265 0.436267 0.0526848 0.0272297 Composite Score (Rank) 0.2733 (2) 0.1319 (3) 0.1026 (4) 0.4099 (1) 0.0496 (5) 0.0315 (6) Table 7. CI, RI, CR values for all the pairwise comparison matrices CI RI CR Cost 0.055157 1.32 0.0417 Quality 0.0944284 1.32 0.07153 Delivery 0.095729 1.32 0.07252 Between criteria 0.0145319 0.66 0.022 9 4.4. Fitness function of SA procedure SA procedure examines the fitness score of each solution generated in the solution neighbourhood. Fitness calculation is one of the most significant steps of the metaheuristic method because it decides which solution is to be stored and which one is to be eliminated. The multiobjective function f(x) is required to compute the fitness score. In order to facilitate the computation, the f(x) is transformed into F(x) and expressed in equation (12). Two weights weight1 and weight2 are also assigned in the equation. It is considered that weight1+weight2=1. 𝐹(𝑥) = 𝑡𝑜𝑡𝑎𝑙 𝑐𝑜𝑠𝑡 𝑡𝑜𝑡𝑎𝑙 𝑐𝑜𝑠𝑡 + 𝑞𝑢𝑎𝑙𝑖𝑡𝑦 𝑐𝑜𝑠𝑡 × 𝑤𝑒𝑖𝑔ℎ𝑡1 + 𝑡𝑜𝑡𝑎𝑙 𝑐𝑜𝑠𝑡 𝑡𝑜𝑡𝑎𝑙 𝑐𝑜𝑠𝑡 + 𝑑𝑒𝑙𝑎𝑦 𝑐𝑜𝑠𝑡 × 𝑤𝑒𝑖𝑔ℎ𝑡2 (12) where total cost is computed by summing up total procurement cost and non-compliance cost. Since the firm gives more importance to the delivery issues rather than the quality or cost issues, thus Weight1 and weight2, are assigned with prefixed values of 0.3 and 0.7 respectively. The Proposed Simulated Annealing Algorithm This subsection describes the proposed SA algorithm in depth. The initial input is a solution string which is generated from AHP technique. Therefore the initial input string S0 which is obtained from Table 6 using the ranks of the suppliers, S0 = [2, 3, 4, 1, 5, 6] The size of the solution string is 6 which is the total number of suppliers to be evaluated. Each bit of the string represents the rank of the corresponding supplier (string index). Therefore S0 states supplier 1 to 6 retain the ranks 2, 3, 4, 1, 5 and 6 respectively. Thereafter this multiobjective SA procedure is set to maximize F(x) of equation (12). Some symbolizations used in the algorithm are introduced as, Scur  current solution Si neighbourhood solution Sbest best solution found so far Tinit  initial temperature Tfinal  freezing temperature T  current temperature α  temperature reducing factor M  Markov chain length iter  iteration number fi  current fitness value fbest  best fitness value The steps of the proposed algorithm can be summarized as follows. Step 1. Obtain an initial solution S0 by using Analytic Hierarchy Process method Step 2. Evaluate S0 and Calculate corresponding fitness value F0; F0 = F(S0) Step 3. Set Fbest= F0, Set Sbest = S0= Scur. Step 4. Initialize SA Heuristic and its parameters: Tinit, Tfinal, α, M, iter = 0, count=0, count1=0. Step 5. If count < M, then repeat Steps 5.1 to 5.9. Step 5.1. Generate a new supplier rank configuration neighbourhood searching by performing exchange-move (randomly selecting two suppliers and interchanging their ranks). Step 5.2. Read suppliers rank configuration from above steps and generate corresponding neighbourhood solution Si. Step 5.3. If F(Si) > Fbest, then Sbest = Si, Scur = Si, count = count + 1, go to Step 5. 10 Step 5.4. If F(Si) = Fbest, then S = Si, count1= count1 + 1, count = count + 1, go to Step 5. Step 5.5. Compute δ = F(Si)- F(Scur). Obtain a random variable r in the range of U(0,1). Step 5.6. If eδ/T > r, Step 5.7. set Scur = Si, Step 5.8. count1= 0; Step 5.9. else count1 = count1 + 1. Step 5.10. iter = iter + 1. Step 5.11. until freezing temperature (Tfinal) is reached; Step 5.12. reduce the temperature using Ti = α×Ti-1 function; The SA procedure is repetitively employed until a solution is achieved which attains the highest fitness score. All the parameters and counters are initialized in step 4. A special move, namely exchange-move, is utilized in the proposed algorithm to guide the solution searching procedure. It is spotted that exchange-move ordinarily leads to the improved solutions effortlessly and competently which is practiced as a principle component for finding better neighbourhood solution in step 5.1. The algorithm also verifies the number of instances when neighbourhood solutions become static. If this number attains a pre-fixed constant value, the fitness value of current configuration is compared to the optimal solution obtained thus far to conclude whether to prolong the iterations or stop with the best solution achieved. 5. Experiments and Verifications In order to apply the proposed SA algorithm as a solution methodology the effects of changing the values of the various parameters are studied. Determining the optimal set of parameters are crucial in this respect. Therefore in this article Taguchi’s robust design method (Taguchi, 1994) is employed to determine the optimal parameters set. 5.1 Taguchi Method for Parameters Selection The parameters are Initial temperature (Tinit), temperature reducing factor (α) and Markov chain length (M) (final temperature (Tfinal) is taken as constant value = -∞) and termed as factors, and each factor has three discrete levels (Table 8). Hence an L9 orthogonal array is used, and this recommends that 9 sets of Taguchi experiments are prerequisite and the results are evaluated by using an analysis of variance (ANOVA) technique. The parameter settings for each experiment are shown in Table 9. Table 8. Levels of parameters tested levels Parameters M α Tinit 20 0.75 10 1 30 0.85 20 2 40 0.95 30 3 Table 10 presents the results of the corresponding ANOVA analysis with Signal-to-Noise ratio (Larger-the-better). In Table 10, the variance ratios (F ratios) of the factors are determined. A test of significance at 95% confidence level is employed to spot the significance of these factors. The P values of the factors Tinit , α, M are investigated and all the values of the parameters are seen to be less than the critical level with degrees of freedom at (2, 8). This suggests that all the parameters are significant factors in the proposed approach. The response table (Table 11) depicts the average of each response characteristic for each level of each of the factors. Table 11 include the ranks based on Delta (δ) statistic, which compares the relative magnitude of effects. Ranks are assigned based on δ values. Using the level averages in the response table optimal set of levels of the factors could be determined which yields the best result. 11 Table 9. The Experimental Settings of the Taguchi Experiments Experiments 1 2 3 4 5 6 7 8 9 Tinit 10 10 10 20 20 20 30 30 30 α 0.75 0.85 0.95 0.75 0.85 0.95 0.75 0.85 0.95 M 20 30 40 30 40 20 40 20 30 responses 0.962364 0.923400 0.945652 0.956225 0.945652 0.932652 0.947348 0.962652 0.959657 Table 10. ANOVA table Factors Tinit α M Residual Error Total Degrees of Freedom 2 2 2 2 8 Factor Sum of squares 0.000301 0.000222 0.000078 0.000875 0.001475 Mean Square (Variance) F Ratio 0.000150 0.34 0.000111 0.25 0.000039 0.09 0.000438 P Value 0.744 0.798 0.918 The ranks indicate that Initial temperature (Tinit) has the greatest influence followed by temperature reducing factor (α) and Markov chain length (M). Each factor level should be fixed in such a way that the highest response could be achieved. Table 11 and the main effects plot of Figure 2 show that the optimal solution is obtained when Tinit, α and M are set to 30, 0.75 and 20 respectively. Table 11. Response table Levels 1 2 3 δ Rank Tinit 0.9438 0.9448 0.9566 0.0127 1 α 0.9553 0.9439 0.9460 0.0114 2 M 0.9526 0.9464 0.9462 0.0063 3 5.2 Convergence Analysis Convergence analysis of the SA procedure is quite simple for the supplier selection problem. The convergence curve during iterations of the proposed metaheuristic technique is presented in Figure 3. For the first iteration the fitness score attained a value of 0.9622. Since the SA procedure is designed to maximize the fitness function with the iteration counts therefore at 4th iteration it attained the value of 0.9633, an increase of 0.1% which is the final optimal solution. Based on the experimentation reported in this article, it is observed that the fitness score is increased with the iteration counts till it reaches the best fitness score at some iteration and thereafter the fitness score continues to remain constant even if the number of iterations is increased. Therefore the convergence property is established. For the test problem in hand the proposed approach is executed for 63 iterations and took 1.9281 CPU seconds to attain the best solution which proves its computational efficiency. 12 Figure 2. Main effects plot Figure 3. Convergence curve of SA algorithm 5.3 Computational Results Data for six vendors are collected from the construction firm for a periodic demand of 600 metric ton TMT steel bar and depicted in Table 12. The strategy of the procurement department is to distribute the order among the best three vendors to avoid the biasness. Each supplier has certain capacity to supply materials as given in table 12. Each supplier supply faulty materials according to the percentage of defective items calculated and each of them supply according to their own pace and the percentages of delays in delivery are also provided in the table 12 for each supplier. The firm’s project schedule is moderately rigid and therefore delay in delivery incurs the overall cost of the project. Delay cost is calculated by the experts of the firm, which is closely 2.5 Lacs INR per day. The cost incurs due to substandard supply are generally converted into delay cost as stated by the firm’s manager and calculated using the first component of equation (4). The AHP method depicts that supplier 4, 1, 2 are the best three suppliers. The total procurement cost is obtained for these three vendors are outlined in Table 13. Thereafter the SA procedure is executed and different result is obtained which states that supplier 3, 1, 4 are the best. The total procurement cost computed for SA method is found to be less than the AHP result and shown in Table 13. Although total faulty materials supplied and total delay days are almost identical for both the methods but the SA method attains closely 0.3% better solution than that of the AHP method. In monetary term SA recovers 1.25 Lacs INR for the firm. This observation indicates that the SA technique is efficient and less complex because of its simplicity in simulation. The solution is obtained with negligible 13 computational time (< 2 seconds). Thus the proposed SA method is shown to outperform the AHP technique. Table 12. Collected vendor details data from the firm Vendors v1 v2 v3 v4 v5 v6 max order quantity (metric ton) 150 300 250 200 750 450 unit cost (INR)/kg 58.75 62 61.5 65 64.5 63.5 Percent defective/metric ton 3.2 3.8 4.5 2 5.7 6.2 percent delay delivery 2.28 2.92 3.12 1.16 3.29 4.5 Table 13. Comparison of SA result with AHP result AHP result SA result vendor v4 v1 v2 order 200 150 250 600 defective 4 5 10 19 late delivery 3 4 8 15 quality cost delay cost Total procurement Cost cost (INR) 13000000 8812500 15500000 37312500 1750000 3750000 43812500 vendor v3 v1 v4 order 250 150 200 600 defective 11 5 4 20 late delivery 8 4 3 15 quality cost delay cost Total procurement Cost cost (INR) 15375000 8812500 13000000 37187500 1750000 3750000 43687500 6. Managerial Implications The study accomplished in this research has significant managerial implications. The soft computing approach proposed in this article can be exploited as a critical managerial decision making tool. This is beneficial in optimizing the vendor network, successful resource allocation for vendor improvement curriculums. In optimizing the vendor network, managers can employ the method to choose vendors without having any biasness for any particular vendor. This further reduces the chance of failure in supplier network. By doing this it would help every supplier to grow evenly. The management of the firm can deliver these suppliers with possible standards for enhancement and target time could be anticipated in complementing them. Another managerial insight of this study affirms that, an already recognized supplier selection methodology (AHP) may not be the best methodology and other state-of-the-art techniques (metaheuristics) can substantially attain better solution and maximize the profit of the firm which is the firms’ main objective. 14 7. Conclusions This article portrays a novel SA based metaheuristic algorithm to select the supplier for a particular Indian firm which is an NP-complete problem in nature. The problem is formulated using multiobjective mathematical model which reflects the essential optimization criteria of this research. The initial feasible solution to the proposed SA based technique is obtained using AHP technique in order to quicken the computation. This work further exploits Taguchi’s robust design approach to select optimal set of parameters to SA algorithm which is crucial in influencing the performance of the technique. The uniqueness of this work lies in practicing two different decision making techniques in solving this MCDM problem model. In past literature such metaheuristic approach to evaluate and enhance the AHP ranking of vendors has never been carried out. To perform the said analysis authors have collected industrial data from a national construction firm. Computational results presented in Section 5 demonstrate that the SA method outperforms the AHP technique performing better than the AHP method for the supplier selection problem. The proposed SA procedure produces nearly 0.3% improved solution. This work is an experimental study which considered the main criteria of the problem such as cost, delivery and delay. However many other intricate sub-criteria could also be considered to make this work more realistic. Future work can be accomplished by utilizing this technique in more complex supplier selection problems incorporating more conflicting criteria and sub-criteria by considering risk factors or suppliers profiles and other related issues and that is the possible extension of this research. Acknowledgement The authors are grateful to the anonymous reviewers for their valuable comments and suggestions in improving the quality of the paper. References Aarts, E., and Korst, J. (1990) Simulated Annealing and the Boltzmann Machine, John Wiley & Sons, New York, USA. Aissaoui, N., Haouari, M., and Hassini, E. (2007) ‘Supplier selection and order lot sizing modeling: A review’, Computers & Operations Research, Vol. 34, No. 12, pp. 3516-3540. Akarte, M.M. (2001) ‘Web based casting supplier evaluation using analytic hierarchy process’, Journal of the Operational Research Society, Vol. 52, No. 5, pp. 511-522. Amid, A. Ghodsypour, S.H., and O’Brien, C. (2009) ‘A weighted additive fuzzy multiobjective model for the supplier selection problem under price breaks in a supply Chain’, International Journal of Production Economics, Vol. 121, No. 2, pp. 323-332. Amid, A. Ghodsypour, S.H., and O’Brien, C. (2010) ‘A weighted max–min model for fuzzy multiobjective supplier selection in a supply chain’, International Journal of Production Economics, doi:10.1016/j.ijpe.2010.04.044 (article in press). Arunkumar, N., Karunamoorthy, L., and Makeshwaraa, N.U. (2007) An optimization technique for vendor selection with quantity discounts using genetic algorithm’, Journal of Industrial Engineering International, Vol. 3, No. 4, pp. 1-13. Bello, M.J.S. (2003) A case study approach to the supplier selection process, M.Sc. thesis, submitted to University of the Puerto Rico Mayaguez. Zhao, G., and Bross, M.E. (2004) Supplier selection process in emerging markets - The Case Study of Volvo Bus Corporation in China, M.Sc. thesis, submitted to Göteborg University, Sweden. Chen, C.T., Lin, C.T., and Huang, S.F. (2006) ‘A fuzzy approach for supplier evaluation and selection in supply chain management’, International Journal of Production Economics, Vol. 102, No. 2, pp. 289–301. 15 Choy, K.L., Lee, W.B., and Lo, V. (2002) ‘An intelligent supplier management tool for benchmarking suppliers in outsource manufacturing’, Expert Systems with Applications, Vol. 22, No, 3, pp. 213224. Dickson, G.W. (1966) ‘An Analysis of Vendor Selection: Systems and Decisions’, Journal of Purchasing, Vol. 1, No. 2, pp. 5-17. Ding, H., Benyoucef, L., and Xie, X. (2003) ‘A simulation-optimization approach using genetic search for supplier selection’, Proceedings of the 2003 Winter Simulation Conference, pp. 12601267. Ellram, L.M. (1990) ‘The Supplier Selection Decision in Strategic Partnerships’, Journal of Purchasing and Materials Management, Vol. 26, No. 4, pp. 8-14. Gnanasekaran, S., Velappan S., and Manimaran, P. (2010) ‘An integrated model for supplier selection using fuzzy analytical hierarchy process: a steel plant case study’, International Journal of Procurement Management, Vol. 3, No. 3, pp. 292-315. Handfield, R., Walton, S.V., and Sroufe, R. (2002) ‘Applying environmental criteria to supplier assessment: A study in the application of the Analytical Hierarchy Process’, European Journal of Operational Research, Vol. 141, No. 1, pp. 70-87. Kanagaraj, G., and Jawahar, N. (2009) ‘A simulated annealing algorithm for optimal supplier selection using the reliability-based total cost of ownership model’, International Journal of Procurement Management, Vol. 2, No. 3, pp. 244-266. Kubat, C., and Yuce, B. (2006) ‘Supplier Selection with Genetic Algorithm and Fuzzy AHP’, Proceedings of 5th International Symposium on Intelligent Manufacturing Systems, pp. 13821401. Lee, A.H.I, Chen, W.C., and Chang, C. J. (2008) ‘A fuzzy AHP and BSC approach for evaluating performance of IT department in the manufacturing industry in Taiwan’, Expert Systems with Applications, Vol. 34, No. 1, pp. 96-107. Li, C.C., Fun, Y.P., and Hung, J.S. (1997) ‘A new measure for supplier performance evaluation’, IIE Transactions, Vol. 29, No. 9, pp. 753-758. Liu, F.H.F., and Hai, H.L. (2005) ‘The voting analytic hierarchy process method for selecting supplier’, International Journal of Production Economics, Vol. 97, No. 3, pp. 308-317. Ng, W.L. (2008) ‘An efficient and simple model for multiple criteria supplier selection problem’, European Journal of Operational Research, Vol. 186, No. 3, pp. 1059–1067. Petroni, A., and Braglia, M. (2000) ‘Vendor selection using principal component analysis’, Journal of Supply Chain Management, Vol. 36, No. 2, pp. 63-69. Rajkumar, C.J., Kannan, S.M., Jayabalan, V. (2009) ‘Multicriteria-based customer-focused vendor selection through the analytic hierarchy process’, International Journal of Procurement Management, Vol. 2, No. 2, pp. 143-162. Rezaei, J., and Davoodi, M. (2008) ‘A deterministic, multi-item inventory model with supplier selection and imperfect quality’, Applied Mathematical Modelling, Vol. 32, No. 10, pp. 2106– 2116. Saaty, T.L. (1980) The Analytic Hierarchy Process: Planning, Priority Setting, Resource Allocation. New York: McGraw-Hill. Saen, R.F. (2008) ‘Using super-efficiency analysis for ranking suppliers in the presence of volume discount offers’, International Journal of Physical Distribution & Logistics Management, Vol. 38, No. 8, pp. 637-651. Saen, R.F., and Zohrehbandian, M. (2008) ‘A data envelopment analysis approach to supplier selection in volume discount environments’, International Journal of Procurement Management, Vol. 1, No. 4, pp. 472-488. Saen, R.F. (2009) ‘Using data envelopment analysis for ranking suppliers in the presence of nondiscretionary factors’, International Journal of Procurement Management, Vol. 2, No. 3, pp. 229-243. Saen, R.F. (2010) ‘Restricting weights in supplier selection decisions in the presence of dual-role factors’, Applied Mathematical Modelling, Vol. 34, No. 10, pp. 2820–2830. Schooner, S. (2003) ‘Commercial Purchasing: The Chasm between the United State Government’s Evolving Policy and Practice’, In: Arrowsmith, S., and Trybus, M. (Eds.), Public Procurement: The Continuing Revolution, (pp. 137-169), Kluwer Law International, The Hague. 16 Scott, S. (1995) ‘Best Value Contracts: Lessons Learned in Paving the Road to Quality’, National Contract Management Journal, Vol. 26, No. 2, pp. 1-11. Taguchi, G. (1994) ‘Taguchi on Robust Technology Development: Bringing Quality Engineering Upstream’, Journal of Electronic Packaging, Vol. 116, No. 2, pp. 161. Tahriri, F., Osman, M.R., Ali, A., and Yusuff, R.M. (2008) ‘A Review of Supplier Selection Methods in Manufacturing Industries’, Suranaree Journal of Science and Technology, Vol. 15, No. 3, pp. 201-208. Tam, M.C.Y, and Tummala, V.M.R. (2001) ‘An application of the AHP in vendor selection of a telecommunications system’, Omega, Vol. 29, No. 2, pp. 171-182. Thompson K.N. (1990) ‘Vendor Profile Analysis’, Journal of Purchasing and Materials Management, Vol. 26, No. 1, pp. 11-18. Timmerman, M. (1986) ‘An approach to vendor performance evaluation’, Engineering Management Review, IEEE, Vol. 15, No. 3, pp. 14-20. Weber C., Current, J.R., and Benton, W.C. (1991) ‘Vendor Selection Criteria and Methods’, European Journal of Operational Research, Vol. 50, No. 1, pp. 2-18. Wu, D. (2009) ‘Supplier selection: A hybrid model using DEA, decision tree and neural network’, Expert Systems with Applications, Vol. 36, No. 5, pp. 9105–9112. Yahya, S., and Kingsman, B. (1999) ‘Vendor rating for an entrepreneur development programme: a case study using the analytic hierarchy process method’, Journal of the Operational Research Society, Vol. 50, No. 5, pp. 916-930. Yu, X., and Jing. S. (2004) ‘A Decision Model for Supplier Selection Considering Trust’, Chinese Business Review, Vol. 3, No. 6, pp. 15-20. 17	9cs.NE
TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS arXiv:1412.7611v3 [math.NT] 23 Aug 2016 JÁN MINÁČ AND NGUYỄN DUY TÂN A BSTRACT. We show that the absolute Galois group of any field has the vanishing triple Massey product property. Several corollaries for the structure of maximal pro-p-quotient of absolute Galois groups are deduced. Furthermore, the vanishing of some higher Massey products is proved. 1. I NTRODUCTION Let F be a field and Fs a separable closure of F. The Galois group GF := Gal( Fs /F) is called the absolute Galois group of F. Every such Galois group is a profinite group. One may ask what special properties absolute Galois groups have among all profinite groups. This is a difficult problem, and at this moment only few properties have been found. However those discovered properties are of great interest and considerable depth. In the classical papers [AS1, AS2] published in 1927, E. Artin and O. Schreier developed a theory of real fields, and they showed in particular that the only non-trivial finite subgroups of absolute Galois groups are groups of order 2. More recently, in some remarkable work M. Rost and V. Voevodsky proved the Bloch-Kato conjecture, thereby establishing a very special property of Galois cohomology of absolute Galois groups. Relatively recently, two new conjectures, the Vanishing n-Massey Conjecture and the Kernel n-Unipotent Conjecture were proposed ( see [MT1] and [MT2]). These conjectures are based on a number of previous considerations. One motivation is coming from topological considerations. (See [DGMS] and [HW].) Another motivation is a program to describe various n-central series of absolute Galois groups as kernels of simple Galois representations. (See [Ef, EM1, EM2, MSp, Vi].) In this paper we shall consider only the special case of the Vanishing n-Massey Conjecture when n = 3 with the exception of Section 5 when we consider n > 3 as well. In the papers [MT1, MT2], the Vanishing n-Massey Conjecture was formulated only in the case when the base field F contains a primitive p-th root of unity. In this paper we consider a stronger version of this conjecture when there is no condition on the field F. For the sake of simplicity we shall recall below the definition of n-Massey products only in the case n = 3, and we refer the reader to [MT1, Sections 2 and 3] for the more general case. See [MT1, Definition 3.3] for a definition of the vanishing n-fold Massey product JM is partially supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) grant R0370A01. NDT is partially supported by the National Foundation for Science and Technology Development (NAFOSTED) grant 101.04-2014.34. 1 2 JÁN MINÁČ AND NGUYỄN DUY TÂN property. Also see Section 5 for reviews of some basic definitions and facts related to n-fold Massey products. Conjecture 1.1 (Vanishing n-Massey Conjecture). Let p be a prime number and n ≥ 3 an integer. Let F be a field. Then the absolute Galois group GF of F has the vanishing n-fold Massey product property with respect to F p . In this paper we show that the conjecture is true when n = 3. We will use the word “triple” instead of “3-fold”. Theorem 1.2 (= Theorem 4.15). Let F be any field and p a prime number. Then the absolute Galois group GF of F has the vanishing triple Massey product property with respect to F p Below, we briefly recall the definition of triple Massey products and the corresponding vanishing property. Let G be a profinite group and p a prime number. We consider the finite field F p as a trivial discrete G-module. Let C • = (C• (G, F p ), ∂, ∪) be the differential graded algebra of inhomogeneous continuous cochains of G with coefficients in F p (see [NSW, Ch. I, §2] and [MT1, Section 3]). We write H i (G, F p ) for the corresponding cohomology groups. We denote by Z1 (G, F p ) the subgroup of C1 (G, F p ) consisting of all 1-cocycles. Because we use trivial action on the coefficients F p , we have Z1 (G, F p ) = H 1 (G, F p ) = Hom(G, F p ). Let χ1 , χ2 , χ3 be elements in H 1 (G, F p ). Assume that χ1 ∪ χ2 = χ2 ∪ χ3 = 0 ∈ H 2 (G, F p ). In this case we say that the triple Massey product hχ1 , χ2 , χ3 i is defined. Then there exist cochains D12 and D23 in C1 (G, F p ) such that ∂D12 = χ1 ∪ χ2 and ∂D23 = χ2 ∪ χ3 in C2 (G, F p ). Then we say that D := {χ1 , χ2 , χ3 , D12 , D23 } (or sometimes for simplicity { D12 , D23 }) is a defining system for the triple Massey product hχ1 , χ2 , χ3 i. Observe that ∂(χ1 ∪ D23 + D12 ∪ χ3 ) = 0, hence χ1 ∪ D23 + D12 ∪ χ3 is a 2-cocycle. We define the value hχ1 , χ2 , χ3 iD of the triple Massey product hχ1 , χ2 , χ3 i with respect to the defining system D to be the cohomology class of χ1 ∪ D23 + χ3 ∪ D12 in H 2 (G, F p ). The set of all values hχ1 , χ2 , χ3 iD when D runs over the set of all defining systems, is called the triple Massey product hχ1 , χ2 , χ3 i ⊆ H 2 (G, F p ). If 0 ∈ hχ1 , χ2 , χ3 i, then we say that our triple Massey product vanishes. Definition 1.3. We say that G has the vanishing triple Massey product property (with respect to F p ) if every triple Massey product hχ1 , χ2 , χ3 i, where χ1 , χ2 , χ3 ∈ H 1 (G, F p ), vanishes whenever it is defined. The Vanishing 3-Massey Conjecture then claims that for any field F and any prime p, the absolute Galois group GF has the vanishing triple Massey product property. It was proved by M. Hopkins and K. Wickelgren in [HW] that if F is a global field of characteristic not 2 and p = 2, then GF has the vanishing triple Massey product property. In TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 3 [MT1] it was proved that the result of [HW] is valid for any field F. In [MT3] it was proved that GF has the vanishing triple Massey product property with respect to F p for any global field F containing a primitive p-th root of unity. In [EMa1], I. Efrat and E. Matzri provided alternative proofs for the above mentioned results in [MT1] and [MT3]. In [Ma], E. Matzri proved that for any prime p and for any field F containing a primitive p-th root of unity, GF has the vanishing triple Massey product property. In this paper we shall provide a cohomological proof to the main result in [Ma] (see Theorem 4.10). We also remove the assumption that F contains a primitive p-th root of unity (see Theorem 4.15). Thus every absolute Galois group has the vanishing triple Massey product property. This is a fundamental new restriction on absolute Galois groups. See Subsection 4.3 for some significant consequences on the structure of quotients of absolute Galois groups. Indeed in Subsection 4.3 we are able to describe some strong restrictions on the shape of relations in Zassenhaus filtration S(2) modulo S(4) for any maximal pro-p-quotient of any absolute Galois group GF . This shape of relations excludes the possibility of certain triple commutators occurring in these relations. These are new significant restrictions on the shape of relations in canonical quotients of absolute Galois groups. This description of relations in S(2) modulo S(4) appears to be close to an optimal description. The structure of our paper is as follows. In Section 2 we recall basic material on the cohomology of bicyclic groups. In Section 3 we discuss Heisenberg extensions. In Section 4, by using the material developed in Sections 2 and 3, we provide an alternative cohomological proof of the main result of [Ma] on the vanishing of triple Massey products with respect to F p when F contains a primitive p-th root of unity. Before we received a very nice preprint [Ma] from E. Matzri, we planned to make such a proof but we completed this proof only after we received his preprint. We also want to notice that by referring directly to some results in [Ti] (see Remark 4.6 for a brief explanation), one can avoid using material in Section 2. However we think that Section 2 might be of independent interest. E.Matzri used in his work tools from the theory of central simple algebras. In our paper we use cohomological techniques instead. In Remark 4.11 we provide yet another short direct variant of the key part of the proof of Theorem 4.10. Considering the details of our proof makes it possible to prove the vanishing of triple Massey products in a more general setting. Consider a formation { G, { G (K )}, N }, where G is a profinite group, { G (K )} is a collection of open subgroups of G indexed by a set Σ = {K }, and N is a discrete G-module. (See [Se1, Chapter XI, §1] or [We, Chapter VI, Section 6.1] for a definition of a formation.) We shall call such a formation { G, { G (K )}, N } a p-Kummer field formation if it satisfies two axioms: (1) For each open normal subgroup G (K ) of G, H 1 (G/G (K ), N G (K ) ) = {1}. (Here N G (K ) is the set of elements of N which are fixed under the action of every element σ in G (K ).) (2) There is a short exact sequence of G-module x 7→ x p 1 −→ Z/pZ −→ N −→ N −→ 1, 4 JÁN MINÁČ AND NGUYỄN DUY TÂN where G acts trivially on Z/pZ, and N is written in a multiplicative way. Recall that axiom (1) above guarantees that each p-Kummer field formation is also a field formation. (See [We, Chapter 6, Section 6.2].) Our main interest is when G = GF , the absolute Galois group of F, and Σ is the set of all finite separable extensions of F, and N = Fs× . But our approach is valid in this more general setting which may have applications in anabelian geometry. Also this approach clarifies the key properties of G which are sufficient for our proofs to go through. Nearly simultaneously with our arXiv posting of the first version of our paper, I. Efrat and E. Matzri posted [EMa2] on arXiv. The paper [EMa2] is a replacement of [Ma]. In [EMa2], I. Efrat and E. Matzri also provide a cohomological approach to Theorem 4.10. Their approach has a similar flavor to our proofs in this paper, but it is still different. We feel that both papers taken together provide a definite complementary insight to the new fundamental property of absolute Galois groups. As we mentioned above, in Remark 4.11 we provide the second alternative proof of the vanishing of triple Massey products. In this proof we are able to show a specific element of the triple Massey product which vanishes. Some results of this paper (e.g. the results on Heisenberg extensions in Subsection 3.2, Lemma 4.2) have already been used in the construction of important Galois groups. Namely in [MT4] we succeeded in extending the crucial ideas in this paper together with further ideas in Galois theory to find explicit constructions of Galois extensions L/F with Gal( L/F) ≃ U4 (F p ), for all fields F and all primes p. For example Theorem 4.12 and its proof play an√im√ × × p p portant role in finding a crucial submodule of the E /(E ) where E := F( a, p c), which is needed in our construction of a required Galois extension L/F containing E/F and having Galois group isomorphic to U4 (F p ). In Section 5 we prove the vanishing of (k + 1)-fold Massey products of the form hχb , χa , . . . , χa i (k copies of χa ) and the vanishing of (k + 2)-fold Massey products of the form hχ a , χb , χ a , . . . , χ a i (k + 1 copies of χ a ), where k < p. (See Theorem 5.9.) The first vanishing can be deduced also from the results in [Sha2], but the second vanishing appears to be new. (See also [Wic, Corollary 3.17] for another result on the vanishing of some n-fold Massey products.) Acknowledgements: We are grateful to Jochen Gärtner and Adam Topaz with whom we initially began correspondence on the vanishing triple Massey products pursuing a different strategy. We are also grateful to Eliyahu Matzri for sending us his beautiful preprint [Ma] shortly before his arXiv posting. Although we did not yet discuss details of our paper with Ido Efrat, Eliyahu Matzri, Danny Neftin and Kirsten Wickelgren, we thank them for their interest and great encouragement. We are grateful to an anonymous referee for his/her careful reading of our paper and for providing us with insightful comments and valuable suggestions which we used to improve our exposition. TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 2. C OHOMOLOGY 5 OF BICYCLIC GROUPS In this section we study the cohomology of cyclic and bicyclic groups. A number of basic results which we will need subsequently in this paper, are recalled here. Our main references are [CKM, pages 16-19] and [Me, pages 694-697]. 2.1. Cohomology of cylic groups. If G is abelian group and g ∈ G is an element of order n, we denote D g := g − 1 and Ng := 1 + g + · · · + gn−1 ∈ Z [ G ]. Let G be a cyclic group of order n. We choose a generator s of G. Recall that ǫ : Z [ G ] → Z is the augmentation homomorphism with ǫ( g) = 1 for all g ∈ G. Then we have the following resolution of the trivial G-module Z: d d ǫ 0 1 · · · −→ Z[ G] −→ Z [ G ] −→ Z [ G ] −→ Z −→ 0, where di is multiplication by Ds (resp. Ns ) if i is even (resp. odd). For any G-module M, the above resolution determines a complex d∗ d∗ 0 1 Hom G (Z [ G ], M) −→ HomG (Z [ G ], M) −→ HomG (Z [ G ], M) −→ · · · We make the natural identification HomG (Z [ G ], M) = M. Then the above complex becomes ∗ ∗ ∗ d0 d d 1 2 M −→ M −→ M −→ M −→ · · · This implies in particular that H 2 (G, M) ≃ Ĥ 0 (G, M) := ker Ds /imNs = Ms /Ms A. As explained in [Se1, Chapter VIII, §4], the above isomorphism does depend on the choice of generator s and can be described as below. The choice of s defines a homomorphism χs : G → Q/Z such that χs (s) = 1/n. The coboundary δ : H 1 (G, Q/Z ) → H 2 (G, Z ) associated to the short exact sequence of trivial G-modules 0 → Z → Q → Q/Z → 0, sends χs to an element θs = δχs ∈ H 2 (G, Z ). Then the isomorphism Ms /Ns M ≃ H 2 (G, M), is the map which sends x ∈ M to x ∪ δχs . 2.2. Cohomology of bicyclic groups. Let G be a bicyclic group. We choose two generators s, of order m, and t, of order n. JÁN MINÁČ AND NGUYỄN DUY TÂN 6 We define a chain complex L• = ( Li ) as follows: Li = Z [ G ]i +1 for all i ≥ 0, and di : Li +1 → Li are defined by the following conditions d2i e2j d2i e2j+1 d2i +1 e2j d2i +1 e2j+1 = Ns e2j−1 + Dt e2j , = Ds e2j − Nt e2j+1, = Ns e2j−1 + Nt e2j , = Ds e2j − Dt e2j+1, here, for convenience, we put e−1 = 0 and (e0 , . . . , ei ) is the canonical basis of Li = Z [ G ]i +1 . Then we obtain a free resolution of the trivial G-module Z: (1) d d d ǫ 0 2 1 · · · −→ Z[ G]4 −→ Z [ G ]3 −→ Z [ G ]2 −→ Z [ G ] −→ Z −→ 0. We define I := ker ǫ, the augmentation ideal of Z [ G ]; and J := ker d0 . Then we obtain the following exact sequence of Z [ G ]-modules f 0 −→ J −→ Z [ G ]2 −→ I −→ 0, (2) where f ( x, y) = d0 ( x, y) = Dt x + Ds y. We also consider the following exact sequence g h 0 −→ Z [ G ]/ZNG −→ J −→ Z2 −→ 0, (3) where NG = ∑σ∈ G σ, g( x + ZNG ) = ( Dt x, − Ds x ) and h( x, y) = (ǫ( x )/n, ǫ(y)/m). Now let M be any G-module. The resolution (1) yields the following complex d0∗ d2∗ d1∗ HomG (Z [ G ], M) → Hom G (Z [ G ]2 , M) → HomG (Z [ G ]3 , M) → HomG (Z [ G ]4 , M) → · · · We make the natural identifications Hom G (Z [ G ]i , M) = Mi . Then the above complex becomes ∗ ∗ ∗ d0 d d 1 2 M −→ M2 −→ M3 −→ M4 −→ · · · The explicit descriptions of some maps di∗ in matrix form are given below:     Dt 0 0 Nt 0  Ds − Nt 0  Dt . d0∗ = , d1∗ =  Ds − Dt  , d2∗ =  0 Ns Dt  Ds 0 Ns 0 0 Ds In particular we have Z2 (G, M) := ker d2∗ = {( x, y, z) ∈ M3 \| Dt ( x ) = 0, Ds (z) = 0, Ds ( x ) = Nt (y), Dt (z) = − Ns (z)}, B2 (G, M) :=imd1∗ = {( x, y, z) ∈ A3 \| ∃(c, d) ∈ A2 : x = Nt (c), y = Ds (c) − Dt (d), z = Ns (d)}, H 2 (G, M) ≃ Z2 (G, M)/B2 (G, M). TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 7 The exact sequences (2) and (3) yield the following commutative diagram () HomG (Z2 , M) ❙❙❙ ❙❙❙ u ❙❙❙ ❙❙❙ ❙❙) / Ext1 ( I, M ) ( J, M) HomG (Z [ G ]2 , M) / Hom G ❯❯❯❯ ❯❯❯❯ v ❯❯❯❯ ❯❯❯❯ ❯ / G 0 = Ext1G (Z [ G ]2 , M) HomG (Z [ G ]/ZNG , M) Ext1G (Z2 , M) This diagram implies that we have a natural injection η : coker(u) ֒→ coker(v). Note that η is an isomorphism if and only if H 1 (G, M) = 0 since Ext1G (Z2 , M) = H 1 (G, M)2 . Under the natural identifications HomG (Z2 , M) = ( M G )2 and HomG (Z [ G ]i , M) = Mi , we shall describe explicitly all objects and maps in the diagram. First we have HomG (Z [ G ]/ZNG , M) = { x ∈ M \| NG x = 0} =: NG M. The map v becomes v : M2 −→ NG M, which sends ( x, y) to Dt x − Ds y. This follows from the observation that the map v is obtained by applying the functor HomG (·, M) to the composite Z [ G ]/ZNG → J → Z [ G ]2 , which maps x mod ZNG to ( Dt x, − Ds x ). On the other hand, the surjection d1 : Z [ G ]3 ։ J yields an injection Hom G ( J, M) ֒→ HomG (Z [ G ]3 , M) = M3 . If we identify Hom G ( J, M) with its image in M3 , then HomG ( J, M) = ker d2∗ =: Z2 (G, M), and Ext1G ( I, M) = ker d2∗ /imd1∗ = H 2 (G, M). The map u becomes u : ( M G )2 −→ H 2 (G, M), which can be described explicitly as follows. We consider the composite d Then h 1 ϕ : Z [ G ]3 −→ J −→ Z2 . ϕ( x, y, z) = h( Nt x + Ds y, − Dt y + Ns z) = (ǫ( Nt x + Ds y)/n, ǫ(− Dt y + Ns z)/m) = (ǫ( x ), ǫ(z)). The Hom-dual of ϕ is the map ϕ∗ : ( M G )2 = HomG (Z2 , M) −→ HomG ( J, M) ֒→ Hom G (Z [ G ]3 , M) = M3 , . JÁN MINÁČ AND NGUYỄN DUY TÂN 8 which is given by ϕ∗ ( x, z) = ( x, 0, z). The map u is then given by u( x, z) = [( x, 0, z)], where [( x, 0, z)] is the class of ( x, 0, z) in H 2 (G, M). Let σ ∈ G/hti denote the class of s modulo hti. Then G/hti = hσi is of order m. We have the natural identification Ĥ 0 (G/hti, Mt ) = M G /Nσ ( Mt ) = H 2 (G/hti, M) by identifying a with the cup product a ∪ θσ . (See Subsection 2.1.) Let χs : G → Q/Z be a homomorphism such that χs (s) = 1/n, χs (t) = 0, and set θs := δχs ∈ H 2 (G, Z ). Then we have [(0, 0, z)] = z ∪ θs = z ∪ δχs ∈ H 2 (G, M). This follows from an observation in [CKM, page 18] that the inflation map infG/hti : H 2 (G/hti, M) ֒→ H 2 (G, M), is given by χs (Note also that = infG/hti Similarly, we have infG/hti (z) = [(0, 0, z)]. (χσ ) and θs = infG/hti θσ .) [( x, 0, 0)] = x ∪ θt = x ∪ δχt ∈ H 2 (G, M). Here χt : G → Q/Z is a homomorphism such that χt (t) = 1/n, χt (s) = 0, and θt := δχt ∈ H 2 (G, Z ). Therefore u( x, z) = x ∪ δχt + z ∪ δχs . Observe that we have the following commutative diagram Z[G] x 7→(0,x,0) Z [ G ]3 / g Z [ G ]/Z [ G ] NG / d1 J. This induces the following commutative diagram HomG (Z [ G ]3 , M) = M3 ( x,y,z)7→y O ? HomG ( J, M) = Z2 (G, M) / / HomG (Z [ G ], M) = M O ? HomG (Z [ G ]/ZNG , M) = NG M TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 9 In summary, under the identifications HomG (Z [ G ]i , M) = Mi , HomG ( J, M) = Z2 (G, M), Ext1G ( I, M) = Z2 (G, M)/B2 (G, M) = H 2 (G, M), etc., the diagram () becomes ( M G )2 () M2 ❏ . ◆◆◆ ◆◆◆u ◆◆◆ ◆◆' 2 / Z ( G, M ) / H 2 ( G, M ) ❏❏ ❏❏ v ❏❏ ❏❏ ❏❏ % / 0 NG M H 1 (G, M)2 Here v( x, y) = Dt x − Ds y. u( x, z) = x ∪ δχt + z ∪ δχs . The natural injection η : coker(u) ֒→ coker(v) is given by η ([( x, y, z)]) = [y]. 3. H EISENBERG EXTENSIONS 3.1. Norm residue symbols. Let F be a field containing a primitive p-th root of unity ξ. For any element a in F× , we shall write χ a for the character corresponding to a via the Kummer map F× → H 1 (GF , Z/pZ )√= Hom(GF , Z/pZ). From now on we assume is a Galois extension with Galois group that a is not in ( F× ) p . The extension√F( p a)/F √ p p hσa i ≃ Z/pZ, where σa satisfies σa ( a) = ξ a. The character χ a defines a homomorphism χ a ∈ Hom(GF , 1p Z/Z ) ⊆ Hom(GF , Q/Z ) by the formula 1 χa = χa . p × Let b be any element in F . Then the norm residue symbol can be defined to be (a, b) := (χa , b) := b ∪ δχa . The cup product χ a ∪ χb ∈ H 2 (GF , Z/pZ ) can be interpreted as the norm residue symbol (a, b). More precisely, we consider the exact sequence x 7→ x p 0 −→ Z/pZ −→ Fs× −→ Fs× −→ 1, where Z/pZ has been identified with the group of p-th roots of unity µ p via the choice of ξ. As H 1 (GF , Fs× ) = 0, we obtain i ×p 0−→ H 2 (GF , Z/pZ ) −→ H 2 (GF , Fs× ) −→ H 2 (GF , Fs× ). JÁN MINÁČ AND NGUYỄN DUY TÂN 10 Then one has i (χ a ∪ χb ) = (a, b) ∈ H 2 (GF , Fs× ). (See [Se1, Chapter XIV, Proposition 5].) The following fact ([Se1, Chapter XIV, Proposition 2]) will also be used frequently in the sequel. Proposition 3.1. We have 2 × res ker χ a 2 × √ ker H (GF , Fs ) −→ H (GF ( p a) , Fs ) = {(a, b) \| b ∈ F× }. 3.2. Heisenberg extensions. In this subsection we provide a short alternative version of some material in [Ma, Section 5]. (See also [Sha1, Chapter 2, Section 2.4].) Assume that a, b are elements in F× , which are linearly independent modulo ( F× ) p . √ √ p K/F is a Galois extension whose Galois √ group is√ generated Let K = F( p a, b). Then √ √ √ √ √ √ p p p p p p p p by σa and σb . Here σa ( b) = b, σa ( a) = ξ a; σb ( a) = a, σb ( b) = ξ b. Let U3 (Z/pZ ) be the group of all upper-triangular unipotent 3 × 3-matrix  withen1 x z tries in Z/pZ. We consider a map U3 (Z/pZ ) → (Z/pZ )2 which sends 0 1 y  to 0 0 1 ( x, y). Then we have the following embedding problem GF ρ̄ 0 / Z/pZ / U3 (Z/pZ ) / (Z/pZ)2 / 1, where ρ̄ is the map (χ a , χb ) : GF → Gal(K/F) ≃ (Z/pZ )2 . (The last isomorphism Gal(K/F) ≃ (Z/pZ )2 is the one which sends σa to (1, 0) and σb to (0, 1).) Assume that √ χa ∪ χb = 0. Then the norm residue symbol (a, b) is trivial. Hence there exists α in F( p a) such that NF ( √ p a ) /F (α ) = b (see [Se1, Chapter XIV, Proposition 4 (iii)]). We set p−2 √ p−2 p−1 p−2 A0 = α σa (α ) · · · σa (α) = ∏ σai (α p−i −1 ) ∈ F( p a). i =0 Lemma 3.2. Let f a be an element in F× . Let A = f a A0 . Then we have NF ( √ p a ) /F (α ) b σa ( A) = = p. p A α α σa ( A) σa ( A0 ) Proof. Observe that = . The lemma then follows from the identity A A0 ( s − 1) p−2 ∑ ( p − i − 1) s i =0 i = p−1 ∑ si − ps0 . i =0 For any representation ρ : GF → U3 (Z/pZ ) and 1 ≤ i < j ≤ 3, let ρij : GF → Z/pZ be the composition of ρ with the projection from U3 (Z/pZ ) to its (i, j)-coordinate. TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 11 Proposition 3.3. Assume that χ a ∪ χb = 0. Let f a be an element in F× . Let A = f a A0 be defined as above. Then the homomorphism ρ̄ := (χ a , χb ) : GF → Z/pZ × Z/pZ lifts to a Heisenberg extension ρ : GF → U3 (Z/pZ ) with resker χ a (ρ13 ) = χ A . Proof. From σa ( A)/A = b/α p ∈ (K × ) p , and σb ( A) = A, we see that σ( A)/A ∈ (K × ) p √ p for every σ ∈ Gal(K/F). This implies that the extension L := K ( A)/F is Galois. Let σ̃a ∈ Gal( L/F ) (resp. σ̃√ σa (resp. σb ). Since σb ( A) = A, b ∈ Gal( L/F )) be an extension of √ √ p √ p p p p j we have σ̃b ( A ) = ξ A, for some j ∈ Z. Hence σ̃b ( A ) = A. This implies that σ̃b is of order p. On the other hand, we have √ b p σ̃a ( A ) p = σa ( A) = A p . α √ √ √ p p p Hence σ̃a ( A) = ξ i A αb , for some i ∈ Z. Then √ √ b p p p p √ = σ̃a ( A) = A A. √ NF ( p a)/F (α) This implies that σ̃a is also of order p. We have √ √ √ pb √ p j p i+ j p A σ̃a σ̃b ( A ) = σ̃a (ξ A ) = ξ , α √ √ √ ξpb √ pb √ p i+ j p i p )=ξ . A σ̃b σ̃a ( A ) = σ̃b (ξ A α α We set σ̃A := σ̃a σ̃b σ̃a−1 σ̃b−1 . Then √ √ p p σ̃A ( A) = ξ A. This implies that σ̃A is of order p and that Gal( L/F) is generated by σ̃a and σ̃b . We also have σ̃a σ̃A = σ̃A σ̃a , and σ̃b σ̃A = σ̃A σ̃b . We can define an isomorphism ϕ : Gal( L/F) → U3 (Z/pZ ) by letting       1 1 0 1 0 0 1 0 1 σ̃a 7→ x := 0 1 0 , σ̃b 7→ y := 0 1 1 , σ̃A 7→ z := 0 1 0 . 0 0 1 0 0 1 0 0 1 ϕ Then the composition ρ : GF → Gal( L/F) → U3 (Z/pZ ) is the desired lifting of ρ̄. Corollary 3.4. Let the notation be as in Proposition 3.3. Let ϕ ab be the map −ρ13 : GF → Z/pZ. Then dϕ ab = χ a ∪ χb and res ker χ a ( ϕ ab ) = −χ A . Proof. Since ρ : GF → U3 (Z/pZ ) is a homomorphism, we obtain dϕ ab (σ, τ ) = ρ13 (στ ) − ρ13 (σ) − ρ13 (τ ) = ρ12 (σ)ρ23 (τ ) = (χ a ∪ χb )(σ, τ ). Therefore dϕ ab = χ a ∪ χb , as desired. JÁN MINÁČ AND NGUYỄN DUY TÂN 12 4. T RIPLE M ASSEY PRODUCTS 4.1. Triple Massey products over fields containing primitive p-th roots of unity. In this subsection we assume that F is a field containing a primitive p-th root of unity. Let a, b and c be elements in F× . Assume further that the triple Massey product hχ a , χb , χc i is defined, i.e., we have χ a ∪ χb = 0 = χb ∪ χc . Until Theorem 4.10, we always assume further that a and c are linearly independent modulo ( F× ) p , and that a and b are linearly independent modulo ( F× ) p . We also fix two elements f a and f c in F× . Let A0 be the element defined right before Lemma 3.2. Let A = f a A0 . By Corollary 3.4, there is a map ϕ ab : GF → Z/pZ such that dϕ ab = χ a ∪ χb ∈ C2 (GF , Z/pZ ) and res ker χ a ( ϕ ab ) = −χ A . Since χb ∪ χc = 0 in H 2 (GF , Z/pZ ), there exists a map ϕbc : GF → Z/pZ such that Then dϕbc = χb ∪ χc ∈ C2 (GF , Z/pZ ). hχa , χb , χc i ϕ := χa ∪ ϕbc + ϕ ab ∪ χc is an element in the triple Massey product hχ a , χb , χc i. We consider the following commutative diagram H 2 (GF , Z/pZ ) i / H 2 (GF , Fs× ) res ker χ a H 2 ( GF ( √ p a ) , Z/pZ ) res ker χ a i / × H 2 ( GF ( √ p a ) , Fs ). Lemma 4.1. We have (1) resker χ a (hχ a , χb , χc i ϕ ) = res ker χ a (χc ) ∪ χ A ∈ H 2 (GF ( √ p a ) , Z/pZ ). 2 × √ (2) resker χ a (i (hχ a , χb , χc i ϕ )) = (c, A) ∈ H (GF ( p a) , Fs ). Proof. (1) We have resker χ a (hχ a , χb , χc i ϕ ) = res ker χ a (χ a ∪ ϕbc + ϕ ab ∪ χc ) = res ker χa (χa ) ∪ res ker χa ( ϕbc ) + res ker χa ( ϕ ab ) ∪ res ker χa (χc ) = −χ A ∪ res ker χa (χc ) = res ker χa (χc ) ∪ χ A . (2) This follows from (1) and the commutativity of the above diagram. √ √ p c). Since χ a ∪ χb = 0 = χb ∪ χc , we have (a, b) = (b, c) = 0. Thus Let E = F( p a, √ √ p p there are α in F( a ) and γ in F( c) such that √ NE/F ( √ p a ) (α ) = b = NE/F ( p c) (γ ). TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 13 Let G be the Galois group Gal(E/F).√Then G = hσa , σc i, where σa ∈ G √ (respectively p σc ∈ G) is an extension of σa ∈ Gal( F( a )/F) (respectively σc ∈ Gal( F( p c)/F)). We define p−2 √ C0 = ∏ σci (γ p−i −1 ) ∈ F( p a), i =0 C := f c C0 , and define B := γ/α. Lemma 4.2. We have σa ( A) = Nσc ( B). (1) A σc (C) = Nσa ( B)−1 . (2) C Proof. (1) From Lemma 3.2, we have b Nσc (γ) σa ( A) = p = = Nσc ( B). A α Nσc (α) (2) From Lemma 3.2, we have σc (C) b Nσa (α) = p = = Nσa ( B)−1 . C γ Nσa (γ) We consider E× as a G-module under the Galois action. The diagram () in Subsection 2.2 becomes ( F × )2 (**) ( E × )2 ◆◆◆ ◆◆◆u ◆◆◆ ◆◆' / Z2 ( G, E× ) / H 2 ( G, E× ) ▲▲▲ ▲▲▲v ▲▲▲ ▲% NG E × 0 Here σc ( x ) y , x σa (y) u( x, z) = x ∪ δχσc + z ∪ δχσa . The natural isomorphism η : coker(u) ֒→ coker(v) is given by v( x, y) = η [( x, y, z)] = [y]. / . 0 14 JÁN MINÁČ AND NGUYỄN DUY TÂN In the following results in this subsection (except in Theorem 4.10) we use the presentation of cohomology classes as in Section 2. Corollary 4.3. The triple ( A, B, C) is an element in Z2 (G, E× ). Proof. This follows immediately from the previous lemma. Lemma 4.4. The element [( A, B, C)] is in the image of u. Proof. Since η is an isomorphism, it suffices to show that η [( A, B, C)] is in the image of v. We have Nσa (α) b Nσa σc (α) = = = 1. Nσa σc ( B) = Nσa σc (γ) Nσc (γ) b × Hence by Hilbert’s Theorem 90, there exists e ∈ E such that σa σc (e) σc (σa (e)) e−1 . = e σa (e) σa (e−1 ) Therefore B is in imv, as desired. B= Corollary 4.5. There exists x, y ∈ F× such that inf([( A, B, C)]) = (a, x ) + (c, y) ∈ H 2 (GF , Fs× ). Proof. Since [( A, B, C)] is in the image of u, there exist x and y in F× such that This implies that [( A, B, C)] = x ∪ δχσa + y ∪ χσc . inf([( A, B, C)]) = x ∪ δ(inf(χσa )) + y ∪ δ(inf(χσc )) = x ∪ δχa + y ∪ δχc = ( x, a) + (y, b), as desired. Ĥ −1 (G, E× ). Remark 4.6. Observe that coker(v) = We can identify naturally the group coker(v) with group Ub (G, E× ) (the notation being as in [Ti]). (See [Ti, Remark after Corollary 1.6].) Then the composition map η H 2 (G, E× ) → coker(u) → coker(v) = Ub (G, E× ), is exactly the map which was denoted by ǫ in [Ti, page 423], ǫ : H 2 (G, E× ) → Ub (G, E× ). The proof of Lemma 4.4 shows that the element [( A, B, C)] is in the kernel of ǫ. We have the following exact sequence inf res 0 −→ H 2 (G, E× ) −→ H 2 (GF , Fs× ) −→ H 2 (GE , Fs× ). If we make the natural identification H 2 (GF , Fs× ) = Br(F) and H 2 (GE , Fs× ) = Br(E), then one can check the map ≃ inf : H 2 (G, E× ) −→ Br(E/F) := ker(Br( F) → Br(E)), TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 15 is the natural isomorphism π : H 2 (G, E× ) −→ Br(E/F) mentioned in [Ti, page 427]. Then Corollary 4.5 follows from [Ti, Proposition 1.5]. We consider the following commutative diagram H 2 (G, E× ) res G/h σai inf / H 2 (GF , Fs× ) res / H 2 (GE , Fs× ) res ker χ a H 2 (hσc i, E× ) inf / × H 2 ( GF ( √ p a ) , Fs ) res / H 2 (GE , Fs× ). Lemma 4.7. We have (1) res G/hσa i ([ A, B, C]) = A ∪ δχσc ∈ H 2 (hσc i, E× ). × (2) res ker χ a (inf([ A, B, C])) = (c, A) in H 2 (GF ( √ p a ) , Fs ). Proof. (1) By [CKM, page 17] and under the identification (E× )σc /Nσc (E× ) = H 2 (hσc i, E× ) via identifying x with the cup product x ∪ δ(χσc ), we have res G/hσa i ([ A, B, C]) = [ A] = A ∪ δχσc . (2) From the commutativity of the above diagram, we obtain res ker χ a (inf([ A, B, C])) = inf(res G/hσa i ([ A, B, C])) = inf( A ∪ δ(χσc )) = A ∪ δ(inf(χσc )) = A ∪ δχc = (c, A), as desired. Corollary 4.8. There exists x ∈ F× such that i (hχ a , χb , χc i ϕ ) = inf([ A, B, C]) + (a, x ) ∈ H 2 (GF , Fs× ). Proof. From Lemma 4.1 and Lemma 4.7 we have × res ker χ a (i (hχ a , χb , χc i ϕ )) = res ker χ a (inf([ A, B, C])) ∈ H 2 (GF ( √ p a ), Fs )). The statement then follows from Proposition 3.1. Corollary 4.9. There exist x and y in F× such that hχa , χb , χc i ϕ = χa ∪ χx + χc ∪ χy ∈ H 2 (GF , Z/pZ). In particular, the triple Massey product hχ a , χb , χc i contains 0. Proof. By Corollaries 4.5 and 4.8, there are x1 , x2 and y in F× such that i (hχ a , χb , χc i ϕ ) = inf([ A, B, C]) + (a, x2 ) = (a, x1 ) + (c, y) + (a, x2 ) ∈ H 2 (GF , Fs× ). 16 JÁN MINÁČ AND NGUYỄN DUY TÂN Let x := x1 x2 . Since i : H 2 (GF , Z/pZ ) → H 2 (GF , Fs× ) is injective, we obtain hχ a , χb , χc i ϕ = χ a ∪ χ x + χc ∪ χy , as desired. For the last statement, we replace ϕ ab by ϕ′ab = ϕ ab + χy , and ϕbc by ϕ′bc = ϕbc − χ x . (Here we consider χ x and χy as elements in Z1 (GF , Z/pZ ).) Then ϕ′ = { ϕ′ab , ϕ′bc } is also a defining system for the triple Massey product hχ a , χb , χc i. Hence hχ a , χb , χc i contains hχa , χb , χc i ϕ′ = hχa , χb , χc i ϕ − χa ∪ χx − χc ∪ χy = 0. Theorem 4.10. Let F be a field containing a primitive p-th root of unity. Let a, b and c be elements in F× . The triple Massey product hχ a , χb , χc i contains 0 whenever it is defined. Proof. Assume that hχ a , χb , χc i is defined. We can also assume that a, b and c are not in ( F× ) p . Case 1: Assume that a and c are linearly dependent modulo ( F× ) p . Let ϕ = { ϕ ab , ϕbc } be a defining system for hχ a , χb , χc i. We have resker χ a (hχ a , χb , χc i ϕ ) = res ker χ a (χ a ∪ ϕbc + ϕ ab ∪ χc ) = res ker χa (χa ) ∪ res ker χa ( ϕbc ) + res ker χa ( ϕ ab ) ∪ res ker χa (χc ) = 0 ∪ res ker χa ( ϕbc ) + res ker χa ( ϕ ab ) ∪ 0 = 0. By Proposition 3.1, hχ a , χb , χc i ϕ = χ a ∪ χ x for some x ∈ F× . We replace ϕbc by ϕ′bc = ϕbc − χ x . (Here we consider χ x as an element in Z1 (GF , Z/pZ ).) Then ϕ′ = { ϕ ab , ϕ′bc } is also a defining system for the triple Massey product hχ a , χb , χc i. Hence hχ a , χb , χc i contains hχa , χb , χc i ϕ′ = hχa , χb , χc i ϕ − χa ∪ χx = 0. Case 2: Assume that a and b are linearly dependent modulo ( F× ) p . Then χb = λχ a , for some λ ∈ Z/pZ. By the linearity of Massey products, we have λhχ a , χ a , χc i ⊆ hχ a , λχ a , χc i = hχ a , χb , χc i. (See [Fe, Lemma 6.2.4 (ii)].) So it is enough to show that hχ a , χ a , χc i contains 0. If p = 2 then hχ a , χb , χc i contains 0 by [MT1, Theorem 1.2]. So we may and shall assume that p > 2. Since then our coefficients are F p of characteristic not 2, we can divide by 2 and define χ2 ϕ aa := − a . 2 Then it is straightforward to check that dϕ aa (σ, τ ) = ϕ aa (σ) + ϕ aa (τ ) − ϕ aa (στ ) = χ a (σ)χ a (τ ) = χ a ∪ χ a (σ, τ ). TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 17 We pick any map ϕ ac : GF → Z/pZ such that dϕ ac = χ a ∪ χc . Then { ϕ aa , ϕ ac } is a defining system for hχ a , χ a , χc i. We have resker χ a (hχ a , χ a , χc i ϕ ) = res ker χ a (χ a ∪ ϕ ac + ϕ aa ∪ χc ) = res ker χa (χa ) ∪ res ker χa ( ϕ ac ) + res ker χa ( ϕ aa ) ∪ res ker χa (χc ) = 0 ∪ res ker χa ( ϕ ac ) + 0 ∪ res ker χa (χc ) = 0. By Proposition 3.1, hχ a , χ a , χc i ϕ = χ a ∪ χ x for some x ∈ F× . Use the same argument as in Case 1, this implies that hχ a , χ a , χc i contains 0. (This also follows from a more general result [Sha2, Proposition 4.2]. See also Theorem 5.9.) Case 3: Assume that a and c are linearly independent and that a and b are linearly independent. Then Corollary 4.9 says that the triple Massey product hχ a , χb , χc i contains 0. Remark 4.11. In Case 3 in the above proof, we can show that a specific element of the Massey triple product vanishes. This leads to another proof for this case and hence for Theorem 4.10, which avoids using Corollary 4.5. Theorem 4.12. There exists f a , f c in F× such that ( f a A0 , B, f c C0 ) is a coboundary. Proof. As in the proof of Lemma 4.4, there exists e ∈ E× such that B = By Lemma 4.2, we have σa ( A0 ) = Nσc ( B) = Nσc A0 σa σc (e) e = σa σc (e) . e σa ( Nσc (e)) . Nσc (e) This implies that Hence Nσc (e) = σa A0 Nσc (e) A0 . √ √ Nσc (e) ∈ F ( p c )× ∩ F ( p a )× = F × . A0 Therefore, there exists f a ∈ F× such that Nσc (e) = A0 f a . Similarly, by Lemma 4.2, we have σc (C0 ) σa σc (e−1 ) σc ( Nσa (e−1 )) −1 = Nσa ( B ) = Nσa . = C0 e −1 Nσa (e−1 ) This implies that Nσa (e−1 ) Nσa (e−1 ) = σc . C0 C0 Hence √ √ Nσa (e−1 ) ∈ F ( p a )× ∩ F ( p c )× = F × . C0 JÁN MINÁČ AND NGUYỄN DUY TÂN 18 Therefore, there exists f c ∈ F× such that Nσa (e−1 ) = C0 f c . Let C1 = σc (e) and C2 = e−1 . Then we have σa (σc (e)) e−1 σa (C1 ) C2 σa σc (e) = , = e σc (e) σc (e−1 ) C1 σc (C2 ) Nσc (C1 ) = Nσc (σc (e)) = Nσc (e) = A0 f a , B= Nσa (C2 ) = Nσa (e−1 ) = C0 f c . This implies that ( A0 f a , B, C0 f c ) is in B2 (G, E× ). As a consequence of Theorem 4.12, we obtain another proof for Case 3 in the proof of Theorem 4.10. Another proof for Case 3 in Proof of Theorem 4.10. Let A = f a A0 and C = f c C0 so that we use the same notation as before. We replace ϕbc by ϕ′bc = ϕbc − χ x . Then ϕ′ = { ϕ ab , ϕ′bc } is also a defining system for the triple Massey product hχ a , χb , χc i. Corollary 4.8 implies that i (hχ a , χb , χc i) contains i (hχ a , χb , χc i ϕ′ ) = i (hχ a , χb , χc i ϕ ) − i (χ a ∪ χ x ) = inf([ A, B, C]) + (a, x ) − (a, x ) = inf([ A, B, C]). Hence hχ a , χb , χc i contains i −1 (inf([ A, B, C])) = 0. 4.2. Embedding problems and triple Massey products over an arbitrary field. A weak embedding problem E for a profinite group Π is a diagram E := Π α U f / Ū which consists of profinite groups U and Ū and homomorphisms α : Π → Ū, f : U → Ū with f being surjective. (All homomorphisms of profinite groups considered in this paper are assumed to be continuous.) If in addition α is also surjective, we call E an embedding problem. A weak solution of E is a homomorphism β : Π → U such that f β = α. We call E a finite weak embedding problem if group U is finite. The kernel of E is defined to be N := ker( f ). Let U4 (F p ) be the group of all upper-triangular unipotent 4 × 4-matrices with entries in F p . Let Z be the subgroup of all such matrices with all off-diagonal entries being 0 except at position (1, 4). We may identify U4 (F p )/Z with the group Ū4 (F p ) of all upper-triangular unipotent 4 × 4-matrices with entries over F p with the (1, 4)-entry omitted. For any representation ρ : G → U4 (F p ) and 1 ≤ i < j ≤ 4, let ρij : G → F p be the composition of ρ with the projection from U4 (F p ) to its (i, j)-coordinate. We use similar notation for representations ρ̄ : G → Ū4 (F p ). Note that for each i = 1, 2, 3, ρi,i +1 (resp., ρ̄i,i +1) is a group homomorphism. TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 19 Recall that we have the following result ([MT2, Lemma 3.1]), which is a direct consequence of [Dwy, Theorem 2.4]. Lemma 4.13. Let G be a profinite group, and p a prime number. Then the following statements are equivalent: (1) G has the vanishing triple Massey product property with respect to F p . (2) For every homomorphism ρ̄ : G → Ū4 (F p ), the finite weak embedding problem 0 M / t yt t (ρ̄12 ,ρ̄23 ,ρ̄34 ) U 4 (F p ) / tG tt (F p )3 / 1, / has a weak solution, i.e., (ρ̄12 , ρ̄23 , ρ̄34 ) can be lifted to a homomorphism ρ : G → U4 (F p ). Proposition 4.14. Let G be a profinite group, and p a prime number. Let H be an open subgroup of G whose index is coprime to p. Assume that H has the vanishing triple Massey product property with respect to F p , then G also has the vanishing triple Massey product property with respect to F p . Proof. We shall prove the condition (2) in Lemma 4.13 for the group G. Let ρ̄ : G → Ū4 (F p ) be any homomorphism. We consider the weak embedding problem (E ) G 0 / M / U 4 (F p ) / (ρ̄12 ,ρ̄23 ,ρ̄34 )= :φ (F p )3 / 1. Then by the assumption and by Lemma 4.13, the weak embedding problem (E \| H ) (E \| H ) 0 / M / yt t t t tt H U 4 (F p ) / φ\| H (F p )3 / 1, which is induced from (E ), has a weak solution. Let ǫ be the cohomology class in H 2 ((F p )3 , M) which corresponds to the extension ( a12 ,a23 ,a34 ) 1 −→ M −→ U4 (F p ) −−−−−−→ (F p )3 −→ 1. We have the following commutative diagram H 2 ((F p )3 , M) φ∗ / H 2 (G, M) res H 2 ((F p )3 , M ) (φ\| H )∗ / H 2 ( H, M). JÁN MINÁČ AND NGUYỄN DUY TÂN 20 In particular, (φ \| H )∗ (ǫ) = res(φ∗ (ǫ)). Since the weak embedding problem (E \| H ) has a weak solution, we see that (φ \| H )∗ (ǫ) = 0 by Hoechsmann’s lemma ([NSW, Chapter 3, §5, Proposition 3.5.9]). (Note that the statement of Hoechsmann’s lemma in [NSW] deals with embedding problems, but its proof goes well with weak embedding problems.) Since [ G : H ] is coprime to p and the order of M is a p-power, we see that the restriction map res : H 2 (G, M) −→ H 2 ( H, M) is injective by [Se2, Chapter I, §2, Corollary to Proposition 9]. Hence φ∗ (ǫ) = 0. Hoechsmann’s lemma implies that the weak embedding problem (E ) has a weak solution, and we are done. Theorem 4.15. Let F be any field and p a prime number. Then the absolute Galois group GF of F has the vanishing triple Massey product property with respect to F p Proof. If charF = p, then the maximal pro-p-quotient GF ( p) of GF is a free pro-p-group. Therefore GF ( p) and GF has the vanishing triple Massey product property with respect to F p . Now we assume that charF 6= p. Let ξ be a primitive p-th root of unity, and let K = F(ξ ). Then K/F is a finite extension of degree d, and d divides p − 1. This implies that [ GF : GK ] = d is coprime to p. Since GK has the vanishing triple Massey product property by Theorem 4.10, it follows that GF also has the vanishing triple Massey product property by Proposition 4.14. 4.3. Some consequences. In this subsection we assume that p is an odd prime number. The case when p = 2 is treated in [MT1, Theorem 1.3 and Theorem 1.4]. Recall that for a profinite group G and a prime number p, the Zassenhaus (p-)filtration (G(n) ) of G is defined inductively by G(1) = G, p G(n) = G(⌈n/p⌉) ∏ [ G( i ) , G( j ) ] , i + j=n where ⌈n/p⌉ is the least integer which is greater than or equal to n/p. (Here for two closed subgroups H and K of G, [ H, K ] means the smallest closed subgroup of G containing the commutators [ x, y] = x −1 y−1 xy, x ∈ H, y ∈ K. Similarly, H p means the smallest closed subgroup of G containing the p-th powers x p , x ∈ H.) Let ( I, <) be a well-ordered set. Let S be a free pro-p-group on a set of generators xi , i ∈ I (see [NSW, Definition 3.5.14]). Let S(i ) , i = 1, 2, . . . be the p-Zassenhaus filtration of S. Then any element r in S(2) may be written uniquely as  3ai bij   ∏ [[ xi , x j ], xk ]]cijk · r′ , if p = 3,  ∏ [ xi , x j ] ∏ xi i∈ I i< j i < j,k≤ j (1) r= b ij  ∏ [[ xi , x j ], xk ]]cijk · r′ , if p 6= 3,   ∏ [ xi , x j ] i< j i < j,k≤ j where ai , bij , cijk ∈ {0, 1, . . . , p − 1} and r ′ ∈ S(4) . For convenience we call (1) the canonical decomposition modulo S(4) of r (with respect to the basis ( xi )) and we also set uij = bij if i < j, and uij = b ji if j < i. TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 21 We denote by GF ( p) the maximal pro-p-quotient of an absolute Galois group GF of a given field F. Theorem 4.16. Let R be a set of elements in S(2) . Assume that there exists an element r in R and distinct indices i, j, k with i < j, k < j such that: (i) In (1) the canonical decomposition modulo S(4) of r, uij = ukj = uki = ukl = u jl = 0 for all l 6= i, j, k, and cijk 6= 0; and (ii) for every s ∈ R which is different from r, the factors [ xk , xi ], [ xi , xk ] and [ xi , x j ] do not occur in the canonical decomposition modulo S(4) of s. Then G = S/hRi is not realizable as GF ( p) for any field F. Proof. This follows immediately from Theorem 4.15, [MT1, Theorem 7.8] and [MT1, Corollary 3.5]. Theorem 4.17. Let R be a set of elements in S(2) . Assume that there exists an element r in R and distinct indices i < j such that: (i) In (1) the canonical decomposition modulo S(4) of r, uij = uil = u jl = 0, for all l 6= i, j and ciji 6= 0 (respectively, cijj 6= 0); and (ii) for every s ∈ R which is different from r, the factor [ xi , x j ] does not occur in the canonical decomposition modulo S(4) of s. Then G = S/hRi is not realizable as GF ( p) for any field F. Proof. This follows immediately from Theorem 4.15, [MT1, Theorem 7.12] and [MT1, Corollary 3.5]. 5. VANISHING OF SOME HIGHER M ASSEY PRODUCTS 5.1. Massey products. Let G be a profinite group and p a prime number. We consider the finite field F p as a trivial discrete G-module. Let C • = (C• (G, F p ), ∂, ∪) be the differential graded algebra of inhomogeneous continuous cochains of G with coefficients in F p [NSW, Ch. I, §2]. We write H i (G, F p ) for the corresponding cohomology groups. We denote by Z1 (G, F p ) the subgroup of C1 (G, F p ) consisting of all 1-cocycles. Because we use trivial action on the coefficients F p , Z1 (G, F p ) = H 1 (G, F p ) = Hom(G, F p ). (See [MT1, MT2] and references therein for more general setups.) Let n ≥ 3 be an integer. Let a1 , . . . , an be elements in H 1 (G, F p ) = Z1 (G, F p ) ⊆ C1 (G, F p ). Definition 5.1. A collection M = { aij \| 1 ≤ i < j ≤ n + 1, (i, j) 6= (1, n + 1)} of elements aij of C 1 (G, F p ) is called a defining system for the n-fold Massey product ha1 , . . . , an i if the following conditions are fulfilled: (1) ai,i +1 = ai for all i = 1, 2 . . . , n. j −1 (2) ∂aij = ∑l =i +1 ail ∪ al j for all i + 1 < j. 22 JÁN MINÁČ AND NGUYỄN DUY TÂN Then ∑nk=2 a1k ∪ ak,n+1 is a 2-cocycle. Its cohomology class in H 2 is called the value of the product relative to the defining system M, and is denoted by ha1 , . . . , an iM . The nfold Massey product ha1 , . . . , an i itself is the subset of H 2 (G, F p ) consisting of all elements which can be written in the form ha1 , . . . , an iM for some defining system M. When n = 3 we will speak about a triple Massey product. Note that in this case the triple Massey product ha1 , a2 , a3 i is defined if and only if a1 ∪ a2 = a2 ∪ a3 = 0 in H 2 (G, F p ). For some convenience, we introduce the following definition. Definition 5.2. Let n ≥ 1 be an integer. Let a1 , . . . , an be elements in H 1 (G, F p ). A collection M = { aij \| 1 ≤ i < j ≤ n + 1} of elements aij of C 1 (G, F p ) is called a complete defining system for the n-tuple (a1 , . . . , an ) if the following conditions are fulfilled: (1) ai,i +1 = ai for all i = 1, 2 . . . , n. j −1 (2) ∂aij = ∑l =i +1 ail ∪ al j for all i + 1 < j. Note that for n = 1, M = { a12 := a1 } is a complete defining system for (a1 ). For n = 2, (a1 , a2 ) has a complete defining system if and only if a1 ∪ a2 = 0. For n ≥ 3, (a1 , . . . , an ) has a complete defining system if and only if the n-fold Massey product ha1 , . . . , an i is defined and contains 0. 5.2. Vanishing of some higher Massey products. Let A be a unital commutative ring. Let n be a positive integer. Assume that every integer 1 ≤ k ≤ n is invertible in A. We have the following binomial polynomials in the ring A[ X ] of polynomials in one variable X and with coefficients in A: X ( X − 1) · · · ( X − k + 1) X X = = 1 and , 1 ≤ k ≤ n. k 0 k! We have the following elementary lemma. Lemma 5.3. Let the notation be as above. For 1 ≤ k ≤ n, one has k Y X X+Y ∈ A[ X, Y ]. =∑ k−l l k l =0 Here A[ X, Y ] is the ring of polynomials with coefficients A in two variables X and Y. Remark 5.4. Let the notation be as above. Then for 1 ≤ k ≤ n, one also has k X l Y k− l ( X + Y )k =∑ ∈ A[ X, Y ]. k! l! ( k − l ) ! l =0 Xk And all results presented below can be stated and proved by using (with some obk! X . vious modifications) instead of using k TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 23 1 1 For any element χ in (G, F p ) = Z (G, F p ) = Hom(G, F p ) and for each k = H χ ∈ C1 (G, F p ) by 0, . . . , p − 1, we define k χ(σ ) χ , ∀σ ∈ G. (σ ) = k k Corollary 5.5. Let χ be an element in Hom(G, F p ). Let k be an integer with 1 ≤ k < p. Then k−1 χ χ χ ∈ C2 (G, F p ). ∪ =−∑ d k − l l k l =1 Proof. By Lemma 5.3, we have χ χ χ χ (στ ) (τ ) − (σ ) + (σ, τ ) = d k k k k χ(στ ) χ(τ ) χ(σ ) − + = k k k χ(σ ) + χ(τ ) χ(τ ) χ(σ ) − + = k k k k χ(τ ) χ(σ ) χ(τ ) χ(σ ) −∑ + = k−l l k k l =0 k−1 χ χ =−∑ (σ, τ ), ∀σ, τ ∈ G, ∪ k−l l l =1 as desired. Corollary 5.6. Let χ be an element in H 1 (G, F p ) = Hom(G, F p ). Let k < p be a positive integer. Then the system χ \| 1 ≤ i < j ≤ k+1 M= − j−i is a complete defining system for the n-tuple (−χ, −χ, . . . , −χ) (k copies of −χ). Proof. This follows immediately from Corollary 5.5. Proposition 5.7. Let F be a field containing a primitive p-th root of unity. Let a and b be elements which are linearly independent modulo ( F× ) p . Let k < p be a positive integer. Assume that χ a ∪ χb = 0. Then the (k + 1)-fold Massey product h−χb , −χ a , . . . , −χ a i (k copies of −χa ) is defined and has a complete defining system of the form M = {aij ∈ C1 (G, F p ) \| 1 ≤ i < j ≤ k + 2}, where χa , for all 2 ≤ i < j ≤ k + 2. aij = − j−i JÁN MINÁČ AND NGUYỄN DUY TÂN 24 Proof. By Corollary 5.6, the system { aij := −( jχ−ai) \| 2 ≤ i < j ≤ k + 2} is a complete defining system for the k-tuple (−χ a , . . . , −χ a ). We set a1,2 = −χb . We shall prove by induction on j = 3, 4, . . . , k + 2 that there exist a13 , a14 , . . . , a1j ∈ C1 (GF , F p ) such that da1r = r −1 ∑ a2l ∪ alr , ∀r = 3, 4, . . . , j. l =2 Then this will imply immediately that the system { aij \| 1 ≤ i < j ≤ k + 2} is a complete defining sytem for the (k + 1)-tuple (−χb , −χ a , . . . , −χ a ). If j = 3, then since χ a ∪ χb = 0, there exists ϕba ∈ C1 (GF , F p ) such that dϕba = χb ∪ χ a . We set a1,3 := ϕba . Then da1,3 = χb ∪ χ a = a1,2 ∪ a2,3 . Now assume that j > 3. By induction hypothesis there exist b13 , b14 , . . . , b1j ∈ C1 (GF , F p ) such that db1r = Then the system r −1 ∑ a1l ∪ alr , ∀r = 3, 4, . . . , j. l =2 N := {a12 , b13 , . . . , b1j , and aµν with 2 ≤ µ < ν ≤ j + 1} is a defining system for the j-fold Massey product h−χb , −χ a , . . . , −χ a i. Then the value of this Massey product with respect to N is hχb , χa , . . . , χa iN = a12 ∪ a2,j+1 + b13 ∪ a3,j+1 + · · · + b1j ∪ a j,j+1 χa χa χa . + · · · + b1j ∪ − + b13 ∪ − = −χb ∪ − 1 j−2 j−1 For the ease of notation, we denote res ker χ a just by res a . Then we have res a (h−χb , −χ a , . . . , −χ a iN ) χa χa χa ) ) + · · · + res a (b1j ) ∪ res a (− ) + res a (b13 ) ∪ res a (− = res a (−χb ) ∪ res a (− 1 j−2 j−1 = res a (−χb ) ∪ 0 + res a (b13 ) ∪ 0 + · · · + res a (b1j ) ∪ 0 = 0. Then by Proposition 3.1, there exists x ∈ F× such that χa χa χa χa ∈ H 2 (G, F p ). = −χx ∪ − + · · · + b1j ∪ − + b13 ∪ − −χb ∪ − 1 1 j−2 j−1 Hence there exists a1,j+1 in C1 (G, F p ) such that χa χa χa χa = da1,j+1 . + (b1j + χx ) ∪ − + · · · + b1,j−1 ∪ − + b13 ∪ − χb ∪ − 1 2 j−2 j−1 TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 25 We set a1r := b1r for r = 3, 4, . . . , j − 1, and a1j := b1j + χ x . Then we have da1r = db1r = r −1 ∑ a1l ∪ alr , ∀r = 3, 4, . . . , j, l =2 and j da1,j+1 = ∑ a1l ∪ al,j+1, l =2 as desired. Proposition 5.8. Let F be a field containing a primitive p-th root of unity. Let a and b be elements which are linearly independent modulo ( F× ) p . Let k < p be a positive integer. Assume that χ a ∪ χb = 0. Then the (k + 2)-fold Massey product h−χ a , −χb , −χ a , . . . , −χ a i (k + 1 copies of −χ a ) is defined and has a complete defining system of the form M = { aij ∈ C1 (G, F p ) \| 1 ≤ i < j ≤ k + 3}, where χa , for all 3 ≤ i < j ≤ k + 3. aij = − j−i Proof. By Proposition 5.7, there exists a system { aij ∈ C1 (G, F p ) \| 2 ≤ i < j ≤ k + 3} such that it is a complete defining system for the k + 1-tuple (−χb , −χ a , . . . , −χ a ) and that χa , for all 3 ≤ i < j ≤ k + 3. aij = − j−i We set a1,2 = −χ a . We shall prove by induction on j = 3, 4, . . . , k + 3 that there exist a13 , a14 , . . . , a1j ∈ C1 (GF , F p ) such that da1r = r −1 ∑ a2l ∪ alr , ∀r = 3, 4, . . . , j. l =2 Then this will imply immediately that the system { aij \| 1 ≤ i < j ≤ k + 3} is a complete defining system for the (k + 2)-tuple (−χ a , −χb , −χ a , . . . , −χ a ). If j = 3, then since χ a ∪ χb = 0, there exists ϕ ab ∈ C1 (GF , F p ) such that dϕ ab = χ a ∪ χb . We set a1,3 := ϕ ab . Then da1,3 = χ a ∪ χb = a1,2 ∪ a2,3 . Now assume that j > 3. By induction hypothesis there exist b13 , b14 , . . . , b1j ∈ C1 (GF , F p ) such that db1r = Then the system r −1 ∑ a1l ∪ alr , ∀r = 3, 4, . . . , j. l =2 N := {a12 , b13 , . . . , b1j , and aµν with 2 ≤ µ < ν ≤ j + 1} JÁN MINÁČ AND NGUYỄN DUY TÂN 26 is a defining system for the j-fold Massey product h−χb , −χ a , . . . , −χ a i. Then the value of this Massey product with respect to N is h−χa , −χb , −χa , . . . , −χa iN = a12 ∪ a2,j+1 + b13 ∪ a3,j+1 + · · · + b1j ∪ a j,j+1 χa χa χa . + · · · + b1j ∪ − + b13 ∪ − = −χa ∪ a2,j+1 + b13 ∪ − 1 j−2 j−1 For the ease of notation, we denote res ker χ a just by res a . Then we have res a (h−χ a , −χb , −χ a , . . . , −χ a iN ) χa χa ) ) + · · · + res a (b1j ) ∪ res a (− = res a (−χa ) ∪ res a (a2,j+1 ) + res a (b13 ) ∪ res a (− 1 j−2 = 0 ∪ res a (a2,j+1 ) + res a (b13 ) ∪ 0 + · · · + res a (b1j ) ∪ 0 = 0. By Proposition 3.1, there exists x ∈ F× such that χa χa χa ∈ H 2 (G, F p ). = −χx ∪ − + · · · + b1j ∪ − −χa ∪ a2,j+1 + b13 ∪ − 1 1 j−2 Hence there exists a1,j+1 in C1 (G, F p ) such that χa χa χa = da1,j+1 . + (b1j + χx ) ∪ − + · · · + b1,j−1 ∪ − −χa ∪ a2,j+1 + b13 ∪ − 1 2 j−2 We set a1r := b1r for r = 3, 4, . . . , j − 1, and a1j := b1j + χ x . Then we have da1r = db1r = r −1 ∑ a1l ∪ alr , ∀r = 3, 4, . . . , j, l =2 and j da1,j+1 = as desired. ∑ a1l ∪ al,j+1, l =2 Theorem 5.9. Let F be a field containing a primitive p-th root of unity. Let a and b be elements which are linearly independent modulo ( F× ) p . Let k < p be a positive integer. Assume that χ a ∪ χb = 0. Then we have (1) The (k + 1)-fold Massey products hχb , χ a , . . . , χ a i and hχ a , . . . , χ a , χb i (k copies of χ a ) are defined and contain 0. (2) The (k + 2)-fold Massey products hχ a , χb , χ a , . . . , χ a i and hχ a , . . . , χ a , χb , χ a i (k + 1 copies of χ a ) are defined and contain 0. Proof. We recall the following formal property of Massey products. If ha1 , a2 , . . . , an i is defined, then han , an−1 , . . . , a1 i is defined and ha1 , a2 , . . . , an i = ±han , an−1 , . . . , a1 i. TRIPLE MASSEY PRODUCTS VANISH OVER ALL FIELDS 27 (See [Kra, Theorem 8].) Observe also that −χ a = χ a−1 for every a ∈ F× . The statement then follows from the two previous propositions. R EFERENCES [AS1] [AS2] [CKM] [Dwy] [Ef] [EMa1] [EMa2] [EM1] [EM2] [Fe] [DGMS] [HW] [Kra] [Ma] [Me] [MSp] [MT1] [MT2] [MT3] [MT4] [NSW] [Se1] [Se2] E. Artin and O. Schreier, Algebraische Konstruktion reeller Körper, Abh. Math. Sem. Univ. Hamburg 5 (1927) pp. 85-99. Reprinted in: Artin’s Collected Papers (Eds. S. Lang and J. Tate), Springer-Verlag, New York, 1965, pp. 258-272. E. Artin and O. Schreier, Eine Kennzeichnung der reell abgeschlossenen Körper, Abh. Math. Sem. Univ. Hamburg 5 (1927) pp. 225-231. Reprinted in: Artin’s Collected Papers (Eds. S. Lang and J. Tate), Springer-Verlag, New York, 1965, pp. 289-295. J.-P. Colliot-Thélène, D. Kanevsky, J.P. Sansuc, Arithmétique des surfaces cubiques diagonales, Diophantine approximation and transcendence theory (Bonn, 1985), 1-108, Lecture Notes in Math., 1290, Springer, Berlin, 1987. W. G. Dwyer, Homology, Massey products and maps between groups, J. Pure Appl. Algebra 6 (1975), no. 2, 177-190. I. Efrat, The Zassenhaus filtration, Massey products, and representations of profinite groups, Adv. Math. 263 (2014), 389-411. I. Efrat and E. Matzri, Vanishing of Massey products and Brauer groups, Can. Math. Bull. 58 (2015), 730-740. I. Efrat and E. Matzri, Triple Massey products and absolute Galois groups, to appear in J. Eur. Math. Soc., arXiv:1412.7265. I. Efrat and J. Mináč, On the descending central sequence of absolute Galois groups, Amer. J. Math. 133 (2011), no. 6, 1503-1532. I. Efrat and J. Mináč, Galois groups and cohomological functors, to appear in Trans. Amer. Math. Soc., arXiv:1103.1508. R. Fenn, Techniques of Geometric Topology, London Math. Soc. Lect. Notes 57 Cambridge 1983. P. Deligne, P. Griffiths, J. Morgan and D. Sullivan, Real homotopy theory of Kähler manifolds, Invent. Math. 29 (1975), 245-274. M. Hopkins and K. Wickelgren, Splitting varieties for triple Massey products, J. Pure Appl. Algebra 219 (2015), 1304-1319. D. Kraines, Massey higher products, Trans. Amer. Math. Soc. 124 (1996), 431-449. E. Matzri, Triple Massey products in Galois cohomology, preprint (2014), arXiv:1411.4146. A. Merkurjev, Essential p-dimension of PGL( p2 ), J. Amer. Math. Soc. 23 (2010), no. 3, 693-712. J. Mináč and M. Spira, Witt rings and Galois groups, Ann. of Math. (2) 144 (1996), no. 1, 35-60. J. Mináč and N. D. Tân, Triple Massey products and Galois theory, to appear in J. Eur. Math. Soc., arXiv:1307.6624. J. Mináč and N. D. Tân, The Kernel Unipotent Conjecture and Massey products on an odd rigid field (with an appendix by I. Efrat, J. Mináč and N. D. Tân), Adv. Math. 273 (2015), 242-270. J. Mináč and N. D. Tân, Triple Massey products over global fields, Doc. Math 20 (2015) 1467-1480. J. Mináč and N. D. Tân, Construction of unipotent Galois extensions and Massey products, preprint (2015) arXiv:1501.01346. J. Neukirch, A. Schmidt and K. Wingberg, Cohomology of number fields, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], 323, Springer-Verlag, Berlin, 2000. J.-P. Serre, Local fields, translated from the French by M. J. Greenberg, Graduate Texts in Mathematics, 67, Springer-Verlag, New York-Berlin, 1979. J.-P. Serre, Galois cohomology, translated from the French by P. Ion and revised by the author, corrected reprint of the 1997 English edition, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2002. JÁN MINÁČ AND NGUYỄN DUY TÂN 28 [Sha1] [Sha2] [Ti] [Vi] [We] [Wic] R. Sharifi, Twisted Heisenberg representations and local conductors, Ph.D. thesis, The University of Chicago (1999). R. Sharifi, Massey products and ideal class groups, J. Reine Angew. Math. 603 (2007), 1-33. J.-P. Tignol, Produits croisés abéliens, J. Algebra 70 (1981), no. 2, 420-436. F. R. Villegas, Relations between quadratic forms and certain Galois extensions, a manuscript, Ohio State University, 1988, http://www.math.utexas.edu/users/villegas/osu.pdf. E. Weiss, Cohomology of groups, Pure and Applied Mathematics, Vol. 34 Academic Press, New York-London, 1969. K. Wickelgren, n-nilpotent obstruction to π1 sections of P1Q − {0, 1, ∞} and Massey products, Galois Teichmüller Theory and Arithmetic Geometry, proceedings for conferences in Kyoto (October 2010) (H. Nakamura, F. Pop, L. Schneps and A. Tamagawa, eds.), Advanced Studies in Pure Mathematics, vol. 63, Mathematical Society of Japan, 2012, 579-600. D EPARTMENT OF M ATHEMATICS , W ESTERN U NIVERSITY, L ONDON , O NTARIO , C ANADA N6A 5B7 E-mail address: minac@uwo.ca D EPARTMENT M ATHEMATICS , W ESTERN U NIVERSITY, L ONDON , O NTARIO , C ANADA N6A 5B7 M ATHEMATICS , V IETNAM A CADEMY OF S CIENCE AND T ECHNOLOGY, 18 H OANG Q UOC V IET, 10307, H ANOI - V IETNAM E-mail address: duytan@math.ac.vn OF AND I NSTITUTE OF	4math.GR
1 Energy Storage Arbitrage in Real-Time Markets via Reinforcement Learning Hao Wang, Baosen Zhang Department of Electrical Engineering, University of Washington, Seattle, WA 98195 arXiv:1711.03127v2 [cs.SY] 23 Feb 2018 Email: {hwang16,zhangbao}@uw.edu Abstract In this paper, we derive a temporal arbitrage policy for storage via reinforcement learning. Real-time price arbitrage is an important source of revenue for storage units, but designing good strategies have proven to be difficult because of the highly uncertain nature of the prices. Instead of current model predictive or dynamic programming approaches, we use reinforcement learning to design an optimal arbitrage policy. This policy is learned through repeated charge and discharge actions performed by the storage unit through updating a value matrix. We design a reward function that does not only reflect the instant profit of charge/discharge decisions but also incorporate the history information. Simulation results demonstrate that our designed reward function leads to significant performance improvement compared with existing algorithms. I. I NTRODUCTION Energy storage can provide various services (e.g., load shifting, energy management, frequency regulation, and grid stabilization) [1] to the power grid and its economic viability is receiving increasing attention. One of the most discussed revenue sources for energy storage is real-time temporal arbitrage (i.e., charging at low prices and discharging at higher prices), where storage units take advantage of the price spreads in real-time electricity market prices to make profits over time [2]. This application has received significant attention from the research community, especially since the growing penetration of intermittent renewable generations are resulting in more volatile real-time electricity market prices [3]. However, even with this increase in price spread, it remains nontrivial to design arbitrage policies that make significant (or even positive) profit [4]. The difficulties come from the fact that future prices are unknown, difficult to forecast and may even be non-stationary [5, 6]. In this paper, we aim to develop an arbitrage policy for energy storage in a data-driven framework by using reinforcement learning [7]. For example, arbitrage using energy storage has been studied in [2, 8–11] (and see the references within). The authors in [8] studied using sodium-sulfur batteries and flywheels for arbitrage in NYISO found the batteries can be potentially profitable using data from 2001 to 2004. The authors in [2] analyzed a generic storage system in the PJM real-time market and discovered that the arbitrage value was nearly doubled from 2002 to 2007 due to higher price variations. The authors in [9] formulated a linear optimization problem to compare the arbitrage profits of 14 energy storage technologies in several major U.S. real-time electric markets. Similar studies have also been carried out in different markets, e.g., Australian national electricity market [10] and European electricity markets [11]. February 27, 2018 DRAFT 2 Crucially, all of these studies assumed perfect knowledge of electricity prices and therefore cannot be implemented as real-time arbitrage strategies. Some recent works [12–15] have started to explicitly take the electricity price uncertainty into account when designing arbitrage strategies. The authors in [12] proposed a stochastic dynamic program to optimally operate an energy storage system using available forecast. The authors in [13] formulated a stochastic optimization problem for a storage owner to maximize the arbitrage profit under uncertainty of market prices. Both studies need to forecast electricity prices and their performances heavily rely on the quality of the forecast. However, real-time market prices are highly stochastic and notoriously difficult to forecast well [16]. To overcome the reliance on price predictions, the authors in [14] employed approximate dynamic programming to derive biding strategy for energy storage in NYISO real-time market without requiring prior knowledge of the price distribution. However, this strategy is often highly computationally expensive. The authors in [15] proposed an online modified greedy algorithm for arbitrage which is computationally straightforward to implement and does not require the full knowledge of price distributions. But it needs to estimate the bounds of prices and assume that storages are “big enough”, which is not always true in practice. The aforementioned challenges motivate us to develop an easily implementable arbitrage policy using reinforcement learning (RL). This policy is both price-distribution-free and outperforms existing ones. Without explicitly assuming a distribution, our policy is able to operate under constantly changing prices that may be non-stationary. Over time, by repeatedly performing charge and discharge actions under different real-time prices, the proposed RL-based policy learns the best strategy that maximizes the cumulative reward. The key technical challenge turns out to be the design of a good reward function that will guide the storage to make the correct decisions. Specifically, we make the following two contributions in this paper: 1) We formulate the energy storage operation as a Markov decision process (MDP) and derive a Q-learning policy [17] to optimally control the charge/discharge of the energy storage for temporal arbitrage in the real-time market. 2) We design a reward function that does not only reflect the instant profit of charge/discharge decisions but also incorporate historical information. Simulation results demonstrate that the designed reward function leads to significant performance improvements compared to the natural instant reward function. In addition, using real historical data, we show the proposed algorithm also leads to much higher profits than existing algorithms. The remainder of the paper is ordered as follows. In Section II, we present the optimization problem for energy storage arbitrage. In Section III, we provide a reinforcement learning approach to obtain the arbitrage policy. Numerical simulations using real data are discussed in Section IV. Section V concludes this paper. II. A RBITRAGE M ODEL AND O PTIMIZATION P ROBLEM We consider an energy storage (e.g., a battery) operating in a real-time electricity market over a finite operational horizon T = {1, ..., T }. The objective of the energy storage is to maximize its arbitrage profit by charging at low prices and discharging when prices are high. We assume the energy storage is a price taker, and its operation will not affect the market prices. We denote dt as the discharged power from the storage at time t and ct as the charged February 27, 2018 DRAFT 3 power into the storage at time t. Let the real-time prices be denoted as pt . We formulate the Arbitrage Maximization Problem (AMP) as follows: max T X 1 pt ηd dt − ct ηc t=1 subject to Et = Et−1 + ct − dt , ∀t ∈ T (AMP) (1) Emin ≤ Et ≤ Emax , ∀t ∈ T (2) 0 ≤ ct ≤ Cmax , t ∈ T (3) 0 ≤ dt ≤ Dmax , t ∈ T (4) variables: ct , dt , ∀t ∈ T , where ηc ∈ (0, 1) and ηd ∈ (0, 1) denote the charge/discharge efficiencies. The constraint in (1) specifies the dynamics of energy level Et over time, (2) constraints the amount of energy in the storage to be between Emin and Emax , (3) and (4) bounds the maximum charge and discharge rates (denoted by Cmax and Dmax , respectively) of the storage. The optimization problem in AMP is a linear program, and we characterize its optimal solution in the next lemma. Lemma 1. The optimal charge and discharge profiles {c?t , d?t , ∀t ∈ T } satisfy 1) At least one of c?t or d?t is 0 at any time t; 2) c?t = {0, min{Cmax , Emax − Et−1 }}, d?t = {0, min{Dmax , Et−1 − Emin }}. Lemma 1 states that the energy storage will not charge and discharge at the same time. Also, the optimal charge and discharge power will hit the boundary per the operational constraints (1)-(4). Specifically, when the storage decides to charge, it will charge either at the maximum charge rate Cmax or reaching the maximum energy level Emax . Similarly, the discharge power will be either the maximum discharge rate Dmax or the amount to reach the minimum energy level Emin . This binary charging/discharging structure will be important when we design the reinforcement learning algorithm in the next section. If the future prices are known, the optimization problem in AMP can be easily solved to provide an offline optimal strategy for the charge/discharge decisions. However, the offline solution is only practical if a good price forecast is available. In reality, future prices are not known in advance and the energy storage needs to make decisions based on only the current and historical data. In other words, the charge/discharge decisions {ĉt , dˆt } are functions of price information up to the current time slot t, denoted by {p1 , ..., pt }: {ĉt , dˆt } = π(p1 , ..., pt ), (5) where π(·) is the arbitrage policy for maximizing the profit. Therefore, AMP is a constrained sequential decision problem and can be solved by dynamic programming [18]. But the potentially high dimensionality the state space February 27, 2018 DRAFT 4 makes dynamic programming computationally expensive, and potentially unsuitable for applications like real-time price arbitrage. Moreover, price forecast in real-time markets is extremely challenging, as the mismatch between power supply and demand can be attributed to many different causes. III. R EINFORCEMENT L EARNING A LGORITHM To solve the online version of AMP, we use reinforcement learning (RL). Reinforcement learning is a general framework to solve problems in which [7]: (i) actions taken depend on the system states; (ii) a cumulative reward is optimized; (iii) only the current state and past actions are known; (iv) the system might be non-stationary. The energy storage arbitrage problem has all of the four properties: (i) different electricity prices lead to different actions (e.g., charge/discharge), and the future energy storage level depends on past actions; (ii) the energy storage aims at maximizing the total arbitrage profit; (iii) the energy storage does not have a priori knowledge of the prices, while it knows the past history; (iv) the actual price profiles are non-stationary. In the following, we describe the RL setup for AMP in more detail. A. State Space We define the state space of the energy arbitrage problem as a finite number of states. To be specific, the system’s state can be fully described by the current price pt and previous energy level Et−1 . We discretize the price into M intervals, and energy level into N intervals. S = {1, ..., M } × {1, ..., N }, where {1, ..., M } represents M even price intervals from the lowest to the highest, and {1, ..., N } denotes N energy level intervals ranging from Emin to Emax . B. Action Space Per Lemma 1, the energy storage will not charge and discharge at the same time. Moreover, the optimal charge and discharge power always reach their maximum allowable rates. We denote the maximum allowable charge/discharge rates as D̃max = min{Dmax , Et−1 − Emin } and C̃max = min{Cmax , Emax − Et−1 }. Therefore, the action space of the energy storage consists of three actions: charge at full rate, hold on, and discharge at full rate: A = {−D̃max , 0, C̃max }, where action a = −D̃max denotes discharge either at maximum rate Dmax or unitl the storage hits the minimum level Emin . Action a = C̃max denotes charge at maximum rate Cmax or until the storage reaches the maximum level Emax . C. Reward At time t, after taking an action a ∈ A at state s ∈ S, the energy storage will receive a reward, such that the energy storage knows how good its action is. According to the objective function of AMP, the energy storage aims February 27, 2018 DRAFT 5 to maximize the arbitrage profit by charging at low prices and discharge at high prices. Therefore, we can define the reward as    −pt C̃max    rt1 = 0     p D̃ t max if charge (Reward 1) if hold on if discharge which is the instant reward of charge or discharge. If the energy storage charges at the rate of C̃max at time t, it will pay at the spot price and reward is negative, i.e., −pt C̃max . In contrast, the energy storage discharges at the rate of D̃max and will earn a revenue of pt D̃max . Reward 1 is a straightforward and natural design, but is actually not very effective. The reason is that the negative reward for charge makes the energy storage perform conservatively in the learning process and thus the arbitrage opportunity is under explored. This motivates us to develop a more effective reward. To avoid conservative actions, we introduce an average price in the reward. The idea comes from the basic principle of arbitrage: to charge at low prices and discharge at high prices. The average price works as a simple indicator to determine whether the current price is low or high compared to the historical values.    (pt − pt )C̃max    rt2 = 0     (p − p )D̃ t max t Specifically, the new reward is defined as if charge if hold on (Reward 2) if discharge where the average price pt is calculated by pt = (1 − η)pt−1 + ηpt , (6) in which η is the smoothing parameter. Note that pt is not a simple average that weighs all past prices equally. Instead, we use moving average in (6), such that we not only leverage the past price information but also adapt to the current price change. We see from Reward 2 that when the energy storage charges at a price lower than the average price (i.e., pt < pt ), it will get a positive reward (pt − pt )C̃max > 0, otherwise it will receive a loss if the spot price is greater. Similarly, Reward 2 encourages the energy storage to discharge at high price by giving a positive reward, i.e., (pt − pt )D̃max > 0. Reward 2 outperforms Reward 1 in exploring more arbitrage opportunities and achieving higher profits. It also mitigates the non-stationarity of prices, since it weights the current price much heavier than prices in the more distant past. We will show the numerical comparisons in Section IV. D. Q-Learning Algorithm With the state, action and reward defined, we obtain the real-time charge and discharge policy using Q-learning (a popular subclass of RL algorithms [7]). Here the energy storage maintains a state-action value matrix Q, where each entry Q(s, a) is defined for each pair of state s and action a. When the energy storage takes a charge/discharge action under a spot price, the value matrix is updated as follows: Q(s, a)t = (1 − α)Q(s, a)t−1 + α[rt + γ max Q(s0 , a0 )], 0 a February 27, 2018 (7) DRAFT 6 where the parameter α ∈ (0, 1] is the learning rate weighting the past value and new reward. γ ∈ [0, 1] is the discount rate determining the importance of future rewards. After taking an action a, the state transits from s to s0 , and the energy storage updates the value matrix incorporating the instant reward rt (e.g., Reward 1 or 2) and the future value maxa0 Q(s0 , a0 ) in state s0 . Over time, the energy storage can learn the value each action in all states. When Q(s, a) converges to the optimal state-action values, we obtain the optimal arbitrage policy. Specifically, the Q-learning algorithm can derive an arbitrage policy for (5) as a? = π(s) = arg max Q(s, a), a (8) which is the optimal arbitrage policy guranteed for finite MDP [17]. For any state s, the energy storage always chooses the best action a? which maximizes the value matrix Q(s, a). Algorithm 1 Q-learning for energy storage arbitrage 1: Initialization: In each time slot t ∈ {1, ..., T }, set the iteration count k = 1, α = 0.5, α = 0.9, and = 0.9. Initialize the Q-matrix, i.e., Q = 0. 2: repeat 3: Step1: Observe state s based on price and energy level; 4: Step2: Decide the best action a (using -greedy method) based on Q(s, a); 5: Step3: Calculate the reward (using Reward 1 or 2); 6: Step4: Update Q(s, a) according to (7) and energy level in (1); 7: s ← s0 and k ← k + 1; 8: until end of operation, i.e., t = T . 9: end The step-by-step Q-learning algorithm for energy arbitrage is presented in Algorithm 1. To avoid the learning algorithm getting stuck at sub-optimal solutions, we employ -greedy [17]. The algorithm not only exploits the best action following (8) but also explores other actions, which could be potentially better. Specifically, using -greedy, the algorithm will randomly choose actions with probability ∈ [0, 1], and choose the best action in (8) with probability 1 − . IV. N UMERICAL R ESULTS In this section, we evaluate two reward functions and also compare our algorithm to a baseline in [15] under both synthetic prices and realistic prices. For synthetic prices, we generate i.i.d. (independent and identically distributed) prices, and for the realistic price, we use hourly prices from ISO New England real-time market [19] from January 1, 2016 to December 31, 2017. The realistic price is depicted in Figure 1. We see that the averaged price is flat but the instantaneous prices fluctuate significantly with periodic spikes. February 27, 2018 DRAFT 7 1200 Real-time price Average price Price ($/MW) 1000 800 600 400 200 0 -200 0 1000 2000 3000 4000 5000 6000 7000 8000 8784 Time (hour) Fig. 1: PJM Real-time price. A. Synthetic Price We first evaluate the two reward functions under synthetic price, which is uniformly distributed in [0, 1] over 1500 hours. We set Cmax = Dmax = 1, Emin = 0 and Emax = 1. The cumulative profits for both rewards are depicted in Figure 2. Both profits stay flat over the first 300 hours, as the algorithm is exploring the environment with different prices. Afterwards, the algorithm using Reward 2 starts to make profit and achieves 166% more than Reward 1 in the end. To further understand the how Reward 1 and Reward 2 affect the storage operation, we plot the evolution of energy level over a 48 hour horizon in Fig. 3. We see that algorithm using Reward 1 performs conservatively while Reward 2 makes the algorithm actively charge and discharge to take advantage of price spread. Therefore, Reward 2 leads to a more profitable arbitrage strategy. B. Real Historical Price We evaluate the two reward functions using realistic prices from ISO New England real-time market in 2016. We plot the cummulative profits of two rewards during training in Figure 4. We see that Reward 1 fails to make profit while using Reward 2 produces a high profit. This demonstrates the effectiveness of our designed reward: it is able to adapt to price changes and makes profit continuously. Cumulative profit ($) 150 100 50 Reward 1 Reward 2 0 0 500 1000 1500 Time (hour) Fig. 2: Cumulative profits under synthetic prices. February 27, 2018 DRAFT 8 Price ($/MW) 1 0.5 0 0 6 12 18 24 30 36 42 48 36 42 48 Time (hour) Energy level (MWh) (a) Synthetic price. 1 0.5 0 0 6 12 18 24 30 Time (hour) Energy level (MWh) (b) Energy level of algorithm using Reward 1. 1 0.5 0 0 6 12 18 24 30 36 42 48 Time (hour) (c) Energy level of algorithm using Reward 2. Fig. 3: Price and energy levels over a 48 hour period using reward 1 and reward 2 under synthetic prices. We also plot the evolution of energy levels over a 48-hours operational horizon in Figure 5. We see that algorithm using Reward 1 cannot capture the price differences but makes charge/discharge when the real-time price is flat. In contrast, our algorithm using Reward 2 is able to charge at low prices at hours 2 and 29, hold the energy when prices are low, and discharge at hours 12 and 44, respectively, when the price reaches a relatively high point. Cumulative profit ($) 8000 Reward 1 Reward 2 6000 4000 2000 0 -2000 0 1000 2000 3000 4000 5000 6000 7000 8000 8784 Time (hour) Fig. 4: Cumulative profits under real-time prices. February 27, 2018 DRAFT 9 Price ($/MW) 100 50 0 0 6 12 18 24 30 36 42 48 36 42 48 Time (hour) Energy level (MWh) (a) Real-time price. 1 0.5 0 0 6 12 18 24 30 Time (hour) Energy level (MWh) (b) Energy level of algorithm using Reward 1. 1 0.5 0 0 6 12 18 24 30 36 42 48 Time (hour) (c) Energy level of algorithm using Reward 2. Fig. 5: Price and energy levels over a 48 hour horizon for reward 1 and reward 2 under historical data. C. Comparison with baseline algorithm Above discussion demonstrates that Reward 2 performs much better than Reward 1, and thus we stick to Reward 2 and compare our algorithm with a baseline algorithm called online modified greedy algorithm in [15]. This algorithm uses a thresholding strategy to control charge and discharge in an online fashion. We configure the parameters for the baseline according to [15]. The arbitrage profits of two algorithms are simulated on an 8−MWh battery, with a charge/discharge rate of 1MW as depicted in Figure 6. The baseline algorithm can only get $5, 845, while our algorithm earns $28, 027 that is 4.8 times of the baseline profit. The profit of the baseline decreases when the charge/discharge rate increases to 2MW. But our algorithm achieves even a higher profit, i.e., $39, 690, which is 8.6 times of the baseline profit $4, 603. The reason is that the baseline algorithm relies on the off-line estimate of the price information and lacks adaptability to the real-time prices. Our algorithm updates the average price to adapt to the price changes and thus performs better. February 27, 2018 DRAFT 10 10 4 3 Baseline algorithm Our algorithm Cumulative profit ($) Cumulative profit ($) 2.5 2 1.5 1 0.5 0 -0.5 0 10 4 4 1000 2000 3000 4000 5000 6000 7000 8760 Baseline algorithm Our algorithm 3 2 1 0 -1 Time (hour) 0 1000 2000 3000 4000 5000 6000 7000 8760 Time (hour) (a) 8MWh-1MW battery. (b) 8MWh-2MW battery. Fig. 6: Cumulative profits of the baseline algorithm in [15] and our algorithm. V. C ONCLUSION In this paper, we derive an arbitrage policy for energy storage operation in real-time markets via reinforcement learning. Specifically, we model the energy storage arbitrage problem as an MDP and derive a Q-learning policy to control the charge/discharge of the energy storage. We design a reward function that does not only reflect the instant profit of charge/discharge decisions but also incorporate the history information. Simulation results demonstrate our designed reward function leads to significant performance improvement and our algorithm achieves much more profit compared with existing baseline method. We will consider self-discharge and degradation of battery in our future work. ACKNOWLEDGMENT This work was partially supported by the University of Washington Clean Energy Institute. R EFERENCES [1] J. Eyer and G. Corey, “Energy storage for the electricity grid: Benefits and market potential assessment guide,” Sandia National Laboratories, vol. 20, no. 10, p. 5, 2010. [2] R. Sioshansi, P. Denholm, T. Jenkin, and J. Weiss, “Estimating the value of electricity storage in pjm: Arbitrage and some welfare effects,” Energy economics, vol. 31, no. 2, pp. 269–277, 2009. [3] C.-K. Woo, I. Horowitz, J. Moore, and A. Pacheco, “The impact of wind generation on the electricity spot-market price level and variance: The texas experience,” Energy Policy, vol. 39, no. 7, pp. 3939–3944, 2011. [4] R. H. Byrne and C. A. Silva-Monroy, “Estimating the maximum potential revenue for grid connected electricity storage: Arbitrage and regulation,” Sandia National Laboratories, 2012. [5] T. T. Kim and H. V. Poor, “Scheduling power consumption with price uncertainty,” IEEE Transactions on Smart Grid, vol. 2, no. 3, pp. 519–527, 2011. [6] S. Borenstein, “The long-run efficiency of real-time electricity pricing,” The Energy Journal, pp. 93–116, 2005. [7] R. S. Sutton and A. G. Barto, “Reinforcement learning: An introduction,” 2011. February 27, 2018 DRAFT 11 [8] R. Walawalkar, J. Apt, and R. Mancini, “Economics of electric energy storage for energy arbitrage and regulation in new york,” Energy Policy, vol. 35, no. 4, pp. 2558–2568, 2007. [9] K. Bradbury, L. Pratson, and D. Patiño-Echeverri, “Economic viability of energy storage systems based on price arbitrage potential in real-time us electricity markets,” Applied Energy, vol. 114, pp. 512–519, 2014. [10] D. McConnell, T. Forcey, and M. Sandiford, “Estimating the value of electricity storage in an energy-only wholesale market,” Applied Energy, vol. 159, pp. 422–432, 2015. [11] D. Zafirakis, K. J. Chalvatzis, G. Baiocchi, and G. Daskalakis, “The value of arbitrage for energy storage: Evidence from european electricity markets,” Applied Energy, vol. 184, pp. 971–986, 2016. [12] K. Abdulla, J. De Hoog, V. Muenzel, F. Suits, K. Steer, A. Wirth, and S. Halgamuge, “Optimal operation of energy storage systems considering forecasts and battery degradation,” IEEE Transactions on Smart Grid, 2016. [13] D. Krishnamurthy, C. Uckun, Z. Zhou, P. Thimmapuram, and A. Botterud, “Energy storage arbitrage under day-ahead and real-time price uncertainty,” IEEE Transactions on Power Systems, 2017. [14] D. R. Jiang and W. B. Powell, “Optimal hour-ahead bidding in the real-time electricity market with battery storage using approximate dynamic programming,” INFORMS Journal on Computing, vol. 27, no. 3, pp. 525–543, 2015. [15] J. Qin, Y. Chow, J. Yang, and R. Rajagopal, “Online modified greedy algorithm for storage control under uncertainty,” IEEE Transactions on Power Systems, vol. 31, no. 3, pp. 1729–1743, 2016. [16] R. Weron, “Electricity price forecasting: A review of the state-of-the-art with a look into the future,” International Journal of Forecasting, vol. 30, no. 4, pp. 1030 – 1081, 2014. [17] C. J. Watkins and P. Dayan, “Q-learning,” Machine learning, vol. 8, no. 3-4, pp. 279–292, 1992. [18] R. A. Howard, Dynamic programming and markov processes. Oxford, England: John Wiley, 1960. [19] “Hourly real-time lmp.” [Online]. Available: https://www.iso-ne.com February 27, 2018 DRAFT	3cs.SY
Interdependent Security Games on Networks under Behavioral Probability Weighting∗ arXiv:1510.09106v3 [cs.GT] 12 Aug 2016 Ashish R. Hota and Shreyas Sundaram† Abstract We consider a class of interdependent security games on networks where each node chooses a personal level of security investment. The attack probability experienced by a node is a function of her own investment and the investment by her neighbors in the network. Most of the existing work in these settings considers players who are risk-neutral. In contrast, studies in behavioral decision theory have shown that individuals often deviate from risk-neutral behavior while making decisions under uncertainty. In particular, the true probabilities associated with uncertain outcomes are often transformed into perceived probabilities in a highly nonlinear fashion by the users, which then influence their decisions. In this paper, we investigate the effects of such behavioral probability weightings by the nodes on their optimal investment strategies and the resulting security risk profiles that arise at the Nash equilibria of interdependent network security games. We characterize graph topologies that achieve the largest and smallest worst case average attack probabilities at Nash equilibria in Total Effort games, and equilibrium investments in Weakest Link and Best Shot games. 1 Introduction Interdependent security games are a class of strategic games where the security risk faced by a player (often manifested as the probability of a successful attack) depends on her personal investment in security and the investments by other interacting players [Laszka et al., 2014, Kunreuther and Heal, 2003]. This is a broad framework to capture security interdependencies between independent stakeholders in (networked) cyber-physical systems. There is a large literature on this class of problems [Laszka et al., 2014, Kunreuther and Heal, 2003, Varian, 2004, Manshaei et al., 2013] motivated by applications in cybersecurity, airline security and epidemic risks. Much of the work in interdependent security games considers players who are risk-neutral, or are risk averse in the sense of classical expected utility theory [Laszka et al., 2014]. On the other hand, there is a rich literature in decision theory and behavioral economics showing that human behavior consistently and significantly deviates from the predictions of classical expected utility theory [Camerer et al., 2011]. While there have been some studies highlighting the significance of biases and irrationalities in human decision-making in information security domains [Christin, 2011, Schneier, 2008, Garg and Camp, 2013], theoretical analyses of deviations from classical notions of rational behavior are scarce in the literature on interdependent security games. Empirical investigations are also limited [Christin et al., 2011]. The goal of this paper is to initiate a rigorous investigation of the impact of behavioral decisiontheoretic models in interdependent security games. In the context of security games, one of the ∗ A preliminary version of this work appeared in the proceedings of GameSec 2015 [Hota and Sundaram, 2015]. The authors are with the School of Electrical and Computer Engineering, Purdue University. E-mail: {ahota,sundara2}@purdue.edu. This work is supported by a grant from the Purdue Research Foundation. † 1 most important behavioral deviations from the classical expected utility framework is the way individuals perceive the probability of an uncertain outcome (e.g., cyber attack).1 In particular, empirical studies show that individuals tend to overweight small probabilities and underweight large probabilities. Thus, the true probabilities are typically transformed in a highly nonlinear fashion into perceived probabilities, which are then used for decision-making [Tversky and Kahneman, 1992, Gonzalez and Wu, 1999]. These transformations are captured in the form of probability weighting functions. In this paper, we analyze the effects of behavioral probability weighting on players’ equilibrium strategies in interdependent security games on networks. We consider three canonical manifestations of the security risk in the forms of Total Effort, Weakest Link and Best Shot games. These game-theoretic models were first introduced in [Varian, 2004] and have been studied extensively in the literature to model several scenarios in the cybersecurity domain, as described below. In Total Effort games on networks, the probability of a successful attack on a node is an (affine) decreasing function of the average of the security investments by the node and her neighbors. The total effort externality has been studied as an abstraction of several cybersecurity problems [Grossklags et al., 2008, Grossklags and Johnson, 2009]. For instance, when an attacker tries to slow down file transfers in peer-to-peer networks, the success of the attack depends on the aggregate effort of all the participating agents [Grossklags et al., 2008]. Similar externalities arise when underinvestment in security by a user potentially causes increasing spam activity for others who communicate with her [Laszka et al., 2014]. In [Miura-Ko et al., 2008, Nguyen et al., 2009], the authors consider a similar formulation, where the security risk faced by a node is a weighted linear combination of her own investment and the investments by her neighbors. The authors discuss multiple settings where such externalities arise, such as in web authentication and spam verification. Amin et. al [Amin et al., 2013] study a related setting, where a set of independent control systems interact over a shared communication network, and the failure probability is a function of the number of controllers who have invested in security.2 In the Weakest Link game, a node is only as secure as the least secure node in her neighborhood, while in the Best Shot game, the player with the maximum investment in the neighborhood must be successfully attacked for the attack on a node to be successful. Weakest link externalities are prevalent in cybersecurity domains; successful breach of one subsystem often increases the vulnerability of connected subsystems by giving the attacker increased access to otherwise restricted parts. Best shot externalities arise in cyber-physical systems that have built-in redundancies. The attacker must breach the most secure subsystem for the attack to be successful. Best shot externalities also arise in censorship resilient communication, where information is available to a node as long as one of her neighbors possesses that information [Johnson et al., 2010]. We summarize our main findings on the effects of network structure and probability weighting on players’ strategies and equilibrium attack probabilities for all three games in Table 1. Security investments often exhibit characteristics of a public good [Varian, 2004], and the different risk externalities described above have indeed been studied in the context of public good games [Hirshleifer, 1983]. There is also a growing interest in networked public goods in recent years 1 There are also various behavioral characteristics that affect the perceived values of gains and losses [Tversky and Kahneman, 1992, Hota et al., 2016]. 2 Security externalities similar to the Total Effort formulation have also been studied in the broader interdependent security game literature in the context of inefficiency of equilibria [Jiang et al., 2011], incomplete information about the network topology [Pal and Hui, 2011] and cyber insurance [Schwartz et al., 2013]. In all these settings, the security risk faced by a node is determined by the actions of the node and her immediate neighbors. This is different from the line of work that models epidemic risks and cascading failures spreading over the network [Nowzari et al., 2016]. 2 Externality Weakest Link Impact of Network Structure The expected fraction of nodes that are successfully attacked at a PNE is highest in degree-regular graphs, among all graphs with a given average degree. In a connected graph, all nodes have identical investments at a PNE. Best Shot Nodes with nonzero investments at a PNE form a maximal independent set. Total Effort Impact of Weighting Function In degree-regular graphs, the interior PNE is more secure under probability weighting when the graph is sufficiently dense. The attack probabilities that arise at a PNE have values close to 1 and/or 0, depending on the game parameters. Under probability weighting, the PNE is never fully insecure, i.e., there always exist node(s) with nonzero investment(s). Table 1: Summary of main results of the pure Nash equilibrium (PNE) characteristics for different attack probability functions. [Galeotti et al., 2010, Bramoullé et al., 2014]. Our analysis of behavioral probability weighting complements this line of research. 2 Probability Weighting As discussed in the previous section, our focus in this paper will be on understanding the effects of nonlinear weighting of true probabilities by individuals while making decisions under risk. Such weightings have been comprehensively studied in the behavioral economics and psychology literature [Camerer et al., 2011], and more recently in wireless communications [Li and Mandayam, 2014] and the smart grid [Saad et al., 2016]. Behavioral perception of probabilities by human decision makers have certain fundamental characteristics, including (i) possibility effect: overweighting of probabilities very close to 0, (ii) certainty effect: underweighting of probabilities very close to 1, and (iii) diminishing sensitivity from the end points 0 and 1. These characteristics are usually captured by an inverse S-shaped weighting function w : [0, 1] → [0, 1]; prominent parametric forms of such weighting functions were proposed by Kahneman and Tversky [Tversky and Kahneman, 1992], Gonzalez and Wu [Gonzalez and Wu, 1999], and Prelec [Prelec, 1998], and are illustrated in Figure 1. For our general analysis of the Nash equilibrium in Total Effort games in Section 4 and Weakest Link and Best Shot games in Section 7, we will assume that the derivative of the weighting function satisfies the following properties. Assumption 1. The probability weighting function w(x) has the following properties. 1. w′ (x) has a unique minimum for x ∈ (0, 1) denoted as xmin,w := argminx∈[0,1] w′ (x), and w′′ (xmin,w ) = 0. 2. w(x) is strictly concave for x ∈ [0, xmin,w ), and is strictly convex for x ∈ (xmin,w , 1]. 3. w′ (ǫ) → ∞ as ǫ → 0, and w′ (1 − ǫ) → ∞ as ǫ → 0. We impose the following additional requirements on the weighting functions in order to obtain certain results on the effects of network structure on the equilibria in Section 5. 3 Perceived Probability (w(x)) 1 GW Pr KT 0.8 0.6 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 True Probability (x) Figure 1: Shape of the probability weighting function for different parametric forms proposed by Gonzalez and Wu (GW) [Gonzalez and Wu, 1999], Prelec (Pr) [Prelec, 1998] and Kahneman and Tversky (KT) [Tversky and Kahneman, 1992]. Assumption 2. The probability weighting function w(x) has the following properties. 1. w ′′ (x) w ′ (x) < 1 1−x for x ∈ (xmin,w , 1). 2. w′ (x) is strictly convex for x ∈ (xmin,w , 1). Remark 1. Assumption 1 and 2 hold true for the parametric forms of the weighting functions proposed by Kahneman and Tversky [Tversky and Kahneman, 1992], Gonzalez and Wu [Gonzalez and Wu, 1999], and Prelec [Prelec, 1998] for the ranges of parameter values under which these functions have an inverse-S shape (Figure 1). In particular, the assumptions hold for the ranges of parameter values estimated from empirical studies on human subjects; [Booij et al., 2010] contains a review of several such studies. For our results on the effect of the intensity of overweighting and underweighting of the weighting functions on the equilibrium attack probabilities in Section 6, we will need to consider a specific parametric form of the weighting function. For that purpose, we use the single-parameter Prelec weighting function from [Prelec, 1998] due to its analytical tractability. In particular, if the true probability of an outcome is x, the Prelec weighting function is given by w(x) = exp(−(− ln(x))α ), x ∈ [0, 1], (1) where exp(·) is the exponential function. The parameter α ∈ (0, 1) controls the curvature of the weighting function. For α = 1, we have w(x) = x, i.e., the weighting function is linear. For smaller α, the function w(x) has a sharper overweighting of low probabilities and underweighting of high probabilities. A useful property of this function is that regardless of the value of α, xmin,w = 1e , and w( 1e ) = 1e . In other words, w′′ (x) = 0 at x = 1e . The minimum value of w′ (x) is w′ ( 1e ) = α. 3 Interdependent Security Games In this paper, we consider interdependent security games on networks. Let G = {V, E} denote an undirected network (or graph) with V being the set of nodes with \|V\| = n. Each node is an independent decision-maker (player of the game) representing, for instance, an entity in a cyber-physical system. The security investment by node i is denoted as si , with si ∈ [0, 1]. The security risk or attack probability experienced by a node is a function of her investment si , and the investment by her direct neighbors. We denote the set of neighbors of node i as N (i), and the investment 4 profile of all nodes in N (i) as the vector sN (i) . The true probability of a successful attack on node i is given by fi (si , sN (i) ) ∈ [0, 1], for some function fi . Node i incurs a cost-per-unit of security investment of ci ∈ R≥0 , and if the attack is successful, she incurs a loss of Li ∈ R>0 . Her expected utility (under the true probability of successful attack) is then Eui (si , sN (i) ) = −Li fi (si , sN (i) ) − ci si . (2) For ease of notation, we define the extended neighborhood of node i, denoted N̄ (i), as the set of nodes including herself and her neighboring nodes, i.e., N̄ (i) , N (i) ∪ {i}. We denote the size of the extended neighborhood of a node i as, di , 1 + \|N (i)\|. (3) In this work, we consider three canonical models of interdependent security games initially presented in [Grossklags et al., 2008]. The models differ in the attack probability function fi (si , sN (i) ) as described below. P • Total Effort: fi (si , sN (i) ) = 1 − d1i ( j∈N̄ (i) sj ). • Weakest Link: fi (si , sN (i) ) = 1 − minj∈N̄ (i) sj . • Best Shot: fi (si , sN (i) ) = 1 − maxj∈N̄ (i) sj . Note that in prior works (e.g. [Grossklags et al., 2008]), the games were defined over complete graphs, i.e., N̄ (i) = V, while we consider more general graph topologies. Most of our analysis will focus on the Total Effort attack probability function, since the results also have potential implications for other classes of security games considered in the literature [Miura-Ko et al., 2008, Nguyen et al., 2009]. Since the focus of the present work is to understand the effects of behavioral probability weighting functions and node degrees on the Nash equilibrium security levels, we only focus on the case where the security risk of a node is influenced in an identical way by all of her neighbors’ investments. We formally define the notion of an equilibrium and best response in interdependent security games as follows. Definition 1. A strategy profile {s∗1 , s∗2 , . . . , s∗n } is a pure Nash equilibrium (PNE) if for every player i ∈ {1, 2, . . . , n}, Eui (s′i , s∗N (i) ) ≤ Eui (s∗i , s∗N (i) ), ∀s′i ∈ [0, 1]. Definition 2. The best response of player i at a given investment profile sN (i) by her neighbors is the set bi (sN (i) ) , argmaxsi ∈[0,1] Eui (si , sN (i) ). A strategy profile {s∗1 , s∗2 , . . . , s∗n } is a pure Nash equilibrium if and only if s∗i ∈ bi (s∗N (i) ) for every player i ∈ {1, 2, . . . , n}. In other words, a PNE exists if the vector of best response mappings [b1 (·), b2 (·), . . . , bn (·)] possesses a fixed point [Osborne and Rubinstein, 1994]. 3.1 Equilibria without probability weighting To establish a baseline, we present the following proposition that describes the main results from [Grossklags et al., 2008] regarding the properties of the best response of a player with w(x) = x in a Total Effort game; while that paper only considered complete graphs, the result extends directly to general graphs. We refer to a player with w(x) = x as a true expectation maximizer as she maximizes her expected utility (2) without any behavioral probability weighing. 5 Proposition 1. Consider a player i with wi (x) = x in a Total Effort game with extended neighborhood size di . Then, her best response is s∗i = 1 when dLi cii < 1, and s∗i = 0 if dLi cii > 1. In the special case where dLi cii = 1, any investment si ∈ [0, 1] is an optimal strategy. Note that, except for the pathological case where dLi cii = 1, the best response of a player is to either fully protect herself or remain completely unprotected. As the size of her neighborhood increases, her best response jumps from investing 1 to investing 0. Such behavior arises since the marginal utility of a true expectation maximizer in (2) is independent of her strategy and the strategies of the players in her neighborhood. When the cost parameters ci and Li are homogeneous across players, only a set of nodes with small enough degrees make an investment of 1, while the high degree nodes invest zero in any Nash equilibrium of the Total Effort game. Furthermore, in degree regular graphs, the only equilibrium that arises has all players investing 0 or investing 1, i.e., all players are either fully secure or fully unprotected. In our analysis, we will show that under behavioral probability weighting, both the best responses and the equilibria have much richer structural properties and vary more smoothly with the weighting parameters and the network structure. 4 Pure Nash Equilibria in Total Effort Games with Behavioral Probability Weighting In this section, we consider Total Effort games on networks with player-specific probability weighting functions wi (·) (satisfying Assumption 1) and cost parameters ci and Li . We first prove the existence of a PNE in this class of games by establishing the existence of a fixed point of the best response mapping (see Definition 2). With probability weighting, the expected utility of player i with investment si ∈ [0, 1] is given by si + s̄−i − ci si , (4) Eui (si , sN (i) ) = −Li wi 1 − di P where s̄−i = j∈N (i) sj is the total investment in security by the neighbors of i. The quantity di is the size of the extended neighborhood of player i as defined in (3). The marginal utility is given by Li ′ si + s̄−i ∂Eui = wi 1 − − ci . (5) ∂si di di The solutions of ∂Eui ∂si = 0 (if any) satisfy the first order necessary condition of optimality, and are −i is the true attack therefore candidate solutions for player i’s best response. Note that 1 − si +s̄ di probability faced by player i (without probability weighting). We illustrate the nature of solutions of the first order condition (5) in Figure 2 for a Prelec weighting function with parameter α = 0.4. One can see from the figure and (5) that if dLi cii ≤ i wi′ (xmin,wi ) = minx∈[0,1] wi′ (x), we have ∂Eu ∂si ≥ 0 for si ∈ [0, 1], and investing 1 is the only best response of a player irrespective of the strategies of her neighbors.3 Now suppose dLi cii > wi′ (xmin,wi ). In this case, the first order condition wi′ (x) = dLi cii has two distinct interior solutions corresponding to true attack probabilities Vi < xmin,w and Xi > xmin,w , as illustrated in Figure 2. Note that as the degree of the node increases, Xi − Vi increases as well. Eui = 0 at x ∂ Eui When dLi cii = wi′ (xmin,wi ), there is a unique solution to ∂∂s > 0 for all other si . The min,wi , and ∂s i i player prefers to invest 1 in this case. The proof is similar to Case 4 in the proof of Lemma 1. 3 6 2 dw(x)/dx 1.5 1 0.5 V 0 0 X 0.2 0.4 0.6 0.8 1 x Figure 2: Interior solutions of w′ (x) = dc are denoted by V and X. In this example, L is shown by the horizontal line. The Prelec weighting parameter is α = 0.4. dc L = 0.8 and When s̄−i is the total investment by the neighbors of player i, player i’s strategy can change her true attack probability in the interval s̄−i 1 + s̄−i ,1 − . (6) X (s̄−i ) , 1 − di di In other words, when the extended neighborhood size is di , each player can directly change the probability of successful attack by at most d1i . We will make the following assumptions. Assumption 3. For any player i with dLi cii > wi′ (xmin,wi ), let the size of the extended neighborhood (di ), and the cost parameters (ci and Li ) be such that, (1) Xi − Vi > d1i , (2) Vi < d1i and (3) wi ( d1i ) < Lcii . The first condition implies that at a given s̄−i , player i does not contain both Vi and Xi in X (s̄−i ). This is required for maintaining continuity of the best response as s̄−i varies. We will discuss the implications of the last two assumptions later in the analysis. Unless otherwise stated, the results of this section hold under Assumption 1 for the probability weighting functions and Assumption 3 for the neighborhood sizes and the cost parameters. We start with the following characterization of the best response of a player. In particular, we show that under Assumptions 1 and 3, the best response of a player is unique, continuous, and monotonically decreasing in the aggregate investment by other nodes in her neighborhood. The proof is presented in Appendix A. Lemma 1. Suppose dLi cii > minx∈[0,1] wi′ (x) for a player i. Then, for a given aggregate investment s̄−i by the neighboring nodes, the best response of player i is given by,   when s̄−i ≤ di (1 − Xi ) − 1 1 bi (s̄−i ) = 0 when s̄−i ≥ di (1 − Xi ),   di (1 − Xi ) − s̄−i otherwise. where Xi is the solution to wi′ (x) = di ci Li for x ∈ (xmin,wi , 1]. We use the properties of the best response proven in the above lemma to establish the existence of PNE in Total Effort games on networks under behavioral probability weighting. Theorem 1. Consider a Total Effort game on a graph where the weighting functions of the players satisfy Assumption 1 and the cost parameters and neighborhood sizes satisfy Assumption 3. Then this game admits a pure Nash equilibrium. 7 Proof. For a player i, if dLi cii ≤ wi′ (xmin,wi ), her best response is to invest 1 regardless of the investments by her neighbors. Otherwise, the best response is unique and continuous in the strategies of her neighbors (from Lemma 1). In addition, the strategy space of each player is [0, 1], which is compact and convex. Thus, according to Brouwer’s fixed point theorem [Ok, 2007], there exists a fixed point of the best response mapping, which corresponds to a PNE. As a consequence of Lemma 1, at a PNE, the investments in the extended neighborhood of any player i for whom dLi cii > w′ (xmin,wi ) can be expressed as, s∗i = min(max(di (1 − Xi ) − s̄∗−i , 0), 1). (7) In particular, we have 1 + s̄∗−i < di (1 − Xi ) =⇒ s∗i = 1 s∗i + s̄∗−i = di (1 − Xi ) =⇒ s∗i ∈ [0, 1] s̄∗−i where wi′ (Xi ) = s∗i ∈ (0, 1). di ci Li > di (1 − Xi ) =⇒ s∗i (8) = 0, with Xi ∈ (xmin,w , 1). The converse of the second identity above holds when Remark 2 (Equilibrium computation). Recently, the authors in [Gharesifard et al., 2016] showed that when the best responses of the players are given by Lemma 1, continuous best response dynamics converge to a PNE strategy profile. Furthermore, a strategy profile that satisfies (8) can be computed by solving a Linear Complementarity Program; a similar result was obtained in [Miura-Ko et al., 2008]. We present an expanded discussion in Appendix B. In general, strategy profiles that satisfy equation (7) need not be unique, and therefore, we need not have a unique PNE. However, in the special case of complete graphs (the classical setting for Total Effort games [Grossklags et al., 2008]), we show that the strategy profiles at the Nash equilibria are unique up to the true equilibrium attack probability experienced by the players. Proposition 2. Consider a Total Effort game on a complete graph where each player i has a playerspecific weighting function wi (·) satisfying Assumption 1, and cost ratio Lcii satisfying Assumption 3. Then all Nash equilibria have the same (true) probability of successful attack at the nodes. The proof is presented in Appendix A. If the strategy profile at each neighborhood satisfies the second identity of (8), then we refer to such a PNE as an interior equilibrium. In other words, at an interior equilibrium, the strategy profile s∗ satisfies, (A + In )s∗ = d ◦ (1 − X), (9) where A is the adjacency matrix of the graph, d is the vector of neighborhood sizes, 1 is the allones vector, X is the vector of Xi ’s of the players, and “◦” denotes the Hadamard (element-wise) product. Thus, at every interior equilibrium, the true attack probability faced by each node i is Xi . The existence of an interior equilibrium is not always guaranteed, except in certain special cases, such as when players are homogeneous and all nodes in the graph have identical degrees (i.e., the graph is degree-regular). Proposition 3. Consider a degree-regular graph with degree d − 1 and homogeneous players such ′ that dc L > w (xmin,w ). Then the symmetric strategy profile where each node invests 1− X constitutes an interior PNE, where w′ (X) = dc L , X ∈ (xmin,w , 1]. The proof is straightforward by substituting s∗ = 1 − X for every player in equation (8). 8 4.1 Existence of secure equilibrium Assumption 3 ensures that the best response of a player remains continuous in the strategies of other players, which helps us establish the existence of PNE in Theorem 1. It is possible to also show the existence of a PNE where all players invest 1 at equilibrium, when the second and third conditions of Assumption 3 do not hold for all the players. Proposition 4. Suppose that for every player i in a Total Effort game, either Vi ≥ ci Li . Then there exists a PNE where all players invest 1. Proof. Consider a player i with Vi ≥ ∂Eui ∂si 1 di or wi ( d1i ) > 1 di , and assume that all of her neighboring nodes are investing h i of player i is nonnegative over the interval X = 0, d1i , and her 1. Then, the marginal utility best response is to invest 1. Similarly, when Vi < d1i , and wi ( d1i ) > Lcii for player i, then her optimal investment is 1 when s̄−i = di − 1. This follows from Case 4 of in the proof of Lemma 1 as substituting s̄i = di − 1 yields Eui (1, sN (i) ) − Eui (0, sN (i) ) > 0 when wi ( d1i ) > Lcii . Both Vi ≥ d1i and wi ( d1i ) > Lcii occur when the weighting function sufficiently overweights very small attack probabilities. Overweighting of small attack probabilities discourages a player from reducing her investment, even when all other players are fully secure, as it would lead to a large perceived increase in attack probabilities from a relatively secure state. This results in a fully secure equilibrium. While such a fully secure PNE possibly coexists with other equilibria, identifying conditions under which such a PNE exists has potential implications for designing incentive mechanisms to encourage users to achieve a secure PNE. 5 Effects of Network Structure In this section, we focus on understanding the effect of network structure, vis-a-vis the degrees of the nodes, on the security investments and attack probabilities at the PNE of Total Effort games. We consider players with homogeneous weighting functions and cost parameters in order to isolate the effects of their degrees on their investments. We will use the characterization of the PNE strategy profile given in equation (8). For the analysis in this section, we assume that the weighting functions satisfy both Assumptions 1 and 2. When the weighting functions and the cost parameters are homogeneous across the players, then the quantity di (1 − Xi ) is only a function of the size of the extended neighborhood di . Recall that w′ (Xi ) = dLi c , and Xi > xmin,w . The properties of Xi as a function of di are the basis of the analytical results of this section. Remark 3. Note that Xi is an increasing function of di , as w′ (x) is strictly increasing in x for x ∈ (xmin,w , 1], as illustrated in Figure 2. 5.1 Investments by nodes with overlapping neighborhoods Without probability weighting, Proposition 1 indicated that a lower degree node always invests at least as much as a higher degree node. This monotonicity does not hold in general under behavioral probability weighting. Nonetheless, we can prove certain monotonicity properties when d(1 − X) is a monotonically decreasing function of d, which holds under Assumption 2 as shown below. 9 Lemma 2. Let c L > minx∈[0,1] w′ (x) so that the quantity X is defined for d ≥ 1. If the probabil- ity weighting function satisfies decreasing in d. w ′′ (x) w ′ (x) < 1 1−x for x ∈ (xmin,w , 1), then d(1 − X) is monotonically The proof of Lemma 2 is presented in Appendix C. We now prove the following result using Lemma 2. Recall that N̄ (i) denotes the extended neighborhood of node i. Proposition 5. Consider a Total Effort game on a network with homogeneous players whose weighting functions satisfy Assumptions 1 and 2. Let the cost parameters satisfy Assumption 3, and let Lc > minx∈[0,1] w′ (x). If N̄ (i) ⊂ N̄ (j), then at any PNE, s∗i ≥ s∗j . Proof. Consider any equilibrium strategy profile where player j invests s∗j > 0, for otherwise, the result holds trivially. Then from equation (8), and from the fact that N̄ (i) ⊂ N̄ (j), we obtain X X s∗k ≤ s∗k ≤ dj (1 − Xj ) < di (1 − Xi ), k∈N̄ (i) k∈N̄ (j) which follows from the monotonicity of d(1 − X) shown in Lemma 2. As a result, from (8), we must have s∗i = 1. From Lemma 2, we know that d(1 − X) is a decreasing function of d. From the PNE characterization in (8), node i invests such that the aggregate investment in her extended neighborhood is di (1−Xi ) subject to her investment being within [0, 1]. For a node with a few neighbors, the desired investment d(1 − X) is larger, and therefore, the node has to increase her personal investment. On the other hand, for a node with large degree, d(1 − X) is smaller and she has a large number of neighbors to rely on to meet the target investment. 5.2 Upper bound on average probability of successful attack at equilibrium For a PNE strategy profile s∗ , we denote the average (true) probability of successful attack as ! P n ∗ X s 1 j j∈ N̄ (j) 1− . (10) Φ(s∗ ) := n di i=1 Note that Φ(s∗ ) is also equal to the expected fraction of nodes that are successfully attacked under the PNE strategy profile s∗ . We obtain an upper bound on Φ(s∗ ) when 1 − Xi < d1i for every player i, with the proof presented in Appendix C. Proposition 6. Consider a Total Effort game on a graph with homogeneous players who satisfy Assumptions 1, 2 and 3. In addition, suppose dLi c > minx∈[0,1] w′ (x) and 1 − Xi < d1i for every player i. Then at any PNE with strategy profile s∗ , 1. the attack probability at node i is at most Xi , and P 2. Φ(s∗ ) ≤ n1 ni=1 Xi . Furthermore, if there exists an interior PNE with strategy profile s∗I , then Φ(s∗I ) = 1 n Pn i=1 Xi . Remark 4. If there is a leaf node i with di = 2 that satisfies Assumption 3, then we must have 1−Xi < 21 ; otherwise Xi −Vi < 12 , which violates the first condition of Assumption 3. Since d(1−X) is decreasing in d (according to Lemma 2 under Assumption 2), this implies that 1 − Xj < d1j holds for every node j in the graph. 10 We now state the main result of this section. We show that degree-regular graphs achieve the 1 Pn highest n i=1 Xi over all graph topologies with the same average degree. Theorem 2. Consider a Total Effort game on a graph under Assumptions 1, 2 and 3. Let dLi c > P w′ (xmin,w ) for every player i, and let davg , n1 ni=1 di be the average of all extended neighborhood sizes. Then, n 1X Φ(s∗ ) ≤ Xi ≤ Xavg , n i=1 where s∗ is a PNE strategy profile and Xavg ∈ (xmin,w , 1] is such that w′ (Xavg ) = davg b L . Proof. Let the function h : R+ → [xmin,w , ∞) be defined as the inverse of w′ (·), i.e., h( dLi c ) , Xi . From Assumptions 1 and 2, we know that w′ (x) is strictly increasing and is strictly convex for x ∈ (xmin,w , 1]. Therefore, the inverse function h(·) is strictly increasing concave in di and strictly davg c di c 1 Pn , which yields for di ∈ R. From the strict concavity of h(·), we have n i=1 h L ≤ h L 1 Pn i=1 Xi ≤ Xavg . Equality holds when di = davg for every node i. n The above result states that graphs with identical node degrees have a larger worst case Φ(s∗ ). Furthermore, as a graph becomes more dense, the bound on Φ(s∗ ) grows with the average degree of the nodes. 5.3 Graphs with smallest average attack probability bound In this subsection, we answer the complementary question regarding graph topologies that have the smallest upper bound on Φ(s∗ ). In order to highlight the dependence of Xi on di , we will use a slightly modified notation in the proof of the following result. In particular, we denote Xi as Xdi for player i. Proposition 7. Consider Total Effort games with n homogeneous players that satisfy Assump′ ) for d = 2. Then, among all connected graphs with n nodes, tions 1, 2 and 3. Let dc L > w (xmin,wP the star graph achieves the smallest ni=1 Xdi . Proof. Recall from Remark 3 that Xdi is an increasing function of di . If the graph is not a tree, then we can remove a setP of edges until the resulting subgraph is a tree. This reduces the neighborhood sizes and decreases ni=1 Xdi . P It remains to show that among all trees, the star graph minimizes ni=1 Xdi . Consider a tree that is not a star graph. Then, maxni=1 di < n. Consider a node u with highest extended neighborhood size du < n. Since the graph is not a star, there must exist a leaf node l which is connected to a node other P than u. Let the neighbor of l be denoted as v, with neighborhood size dv ≤ du . We argue that ni=1 Xdi decreases if we remove the edge between l and v, and add an edge between l and u. Under this operation, Xdu increases to Xdu +1 for node u and Xdv decreases Pn to Xdv −1 for node v. For all other nodes, Xdi remains unchanged. We compute the change in i=1 Xdi as, Xdu +1 − Xdu + Xdv −1 − Xdv <(Xdu − Xdu −1 ) − (Xdv − Xdv −1 ) ≤(Xdv − Xdv −1 ) − (Xdv − Xdv −1 ) = 0. (11) The first inequality follows from Theorem 2 with two players having extended neighborhood sizes du − 1 and du + 1, i.e., Xdu +1 + Xdu −1 < 2Xdu for du ≥ 3. The second inequality holds for dv ≤ du 2. Therefore, for following the strict concavity of h(·) (where h( dc L ) = Xd ) as shown in Theorem P any tree which is not a star graph, we can construct another tree which reduces ni=1 Xdi . 11 2 α=0.8 α=0.4 dw(x)/dx 1.5 1 0.5 0 X̄ 0 0.2 0.4 0.6 0.8 1 x Figure 3: wi′ (x) for Prelec weighting functions with parameters α1 = 0.4 and α2 = 0.8. At x = X̄, w1′ (x) = w2′ (x). The proof of Proposition 7 relies on the properties of the quantities Xi which arise due to the nature of the behavioral weighting functions. In addition, the result that star graphs achieve the smallest security risk upper bound is an artifact of the Total Effort risk externalities, which assume that the attack on each node only depends on the average security investment in her neighborhood. In other words, the attacks are not targeted (in the sense of trying to disconnect the network). An interesting avenue for future work is to characterize network topologies that maximize more complex network value functions (such as those considered in [Gueye et al., 2010, Schwartz et al., 2011, Cerdeiro et al., 2015]) under behavioral probability weighting and targeted attacks. 6 Comparative Statics in Weighting Functions In this section, we compare the effect of the intensity of probability weighting on Φ(s∗ ) (the average equilibrium probability of successful attack) in degree-regular graphs. Analyzing degree-regular graphs not only isolates the effects of heterogeneity in degrees from the effects of the weighting function, but as we showed in the previous section, such graphs possess an interior equilibrium that upper bounds the Φ(s∗ ) that arise in a broader class of graphs. In order to compare two different weighting functions, we use the parametric form (1) proposed by Prelec with parameter α ∈ (0, 1]. When α = 1, the weighting function is linear; as α decreases from 1, the magnitudes of overweighting and underweighting increase. We consider two Total Effort games on a (d − 1)−regular graph, i.e., the size of the extended neighborhood of the nodes is d. Let the cost parameters c and L be the same among the players across the two games. The first game has homogeneous players with weighting parameter α1 , and the second game has homogeneous players with weighting parameter α2 , with α1 < α2 < dc L . In other words, the players in the first game have a more significant overweighting and underweighting of the true probabilities compared to the players in the second game. Let X1 and X2 be the solutions to the equations wi′ (x) = dc L , i ∈ {1, 2}, such that Xi > xmin,w . Here wi (x) is the Prelec weighting function (1) with weighting parameter αi . From Proposition 3, the attack probability at each node is equal to Xi at the respective interior PNEs and Φi (s∗ ) = Xi . As we illustrate in Figure 3 for α1 = 0.4 and α2 = 0.8, w1′ (x) is initially smaller than w2′ (x) as x starts to increase from 1e , until the quantity x = X̄ (which depends on the values of α1 and α2 ) at which w1′ (x) = w2′ (x). For x > X̄, w1′ (x) > w2′ (x). We state this formally in the following lemma. The proof is presented in Appendix D. Lemma 3. Consider two Prelec weighting functions w1 (·) and w2 (·) with parameters α1 and α2 , respectively and let α1 < α2 . Then there exists a unique X̄ > 1e such that (1) w1′ (X̄) = w2′ (X̄), (2) for x ∈ ( 1e , X̄), w1′ (x) < w2′ (x), and (3) for x ∈ (X̄, 1), w1′ (x) > w2′ (x). 12 Now we present the main result of this section, which follows from the above lemma. Theorem 3. Let Xi , i ∈ {1, 2} be the true probability of successful attack at a node at the interior PNE of a (d − 1)-regular graph where the nodes have a Prelec weighting function with parameter αi , i ∈ {1, 2}, with α1 < α2 . Let X̄ be the intersection point defined in Lemma 3 for α1 and α2 . Then, 1. if dc L > w′ (X̄), we have X̄ < X1 < X2 , 2. otherwise, if α2 < 3. if dc L dc L < w′ (X̄), we have X̄ > X1 > X2 , and = w′ (X̄), we have X̄ = X1 = X2 . The above result shows that when dc L is such that the attack probability at a node at PNE is high enough (greater than X̄), the players with substantial underweighting of probabilities (i.e., with smaller α) view increased security investments to be highly beneficial in terms of reducing the perceived attack probabilities. As a result, the attack probability at a node at the PNE is smaller in the game with weighting parameter α1 . On the other hand, when the attack probability is less than X̄, the players with smaller α do not find the perceived reduction in attack probabilities to be sufficient to make a high investment. However players with weighting functions closer to linear observe greater perceived reduction in probability due to increased investment, and as a result, these players have smaller average attack probability at the interior equilibria. When we keep Lc fixed and increase the neighborhood size d, then we eventually end up in the regime where the equilibrium attack probability is greater than X̄. Thus, the expected fraction of nodes that are successfully attacked at equilibrium (Φi (s∗ )) is smaller under behavioral probability weighting when the graph is sufficiently dense (i.e., has a large number of edges for a given number of nodes), and vice versa. 7 Weakest Link and Best Shot Games The Nash equilibrium strategies in Weakest Link and Best Shot games have very special properties, as the security level at each neighborhood is determined by the investment of a single player (the one with the smallest and largest investment, respectively). Therefore, we first state the following results that characterize the security investment by an isolated player (i.e., d = 1). The proofs of the following results are in Appendix E. Lemma 4. Consider a weighting function that satisfies Assumption 1. Let z > xmin,w be such that w(x) ′ w′ (z) = w(z) z . Then i) z exists and is unique, and ii) for x > z, w (x) > x . Proposition 8. Let z be as defined in Lemma 4. If Lc < w′ (z), the optimal investment of a single player is s∗ = 1. Otherwise, the optimal investment is s∗ = 1 − X, where w′ (X) = Lc , X ∈ (xmin,w , 1]. We now analyze the PNE in Weakest Link games. Proposition 9. Consider a Weakest Link game on a connected graph with homogeneous players. Then at any PNE, all nodes make identical investments. If Lc ≥ w′ (z), where z is as defined in Lemma 4, then there is a continuum of pure Nash equilibria where the successful attack probabilities at the nodes are greater than or equal to X. When w′ (xmin,w ) < Lc < w′ (z), then there are additional equilibria (including the ones in the previous case) with attack probabilities close to 0. 13 Proof. Consider a node v1 with security investment s1 , and let ŝ < s1 be the smallest investment in her neighborhood. Then the attack probability on node v1 is fi (si , sN (i) ) = 1 − ŝ > 1 − s1 . If node v1 reduces her investment to ŝ, then her cost of investment decreases, while the attack probability remains unchanged. Therefore, at a PNE, the investment by any node must be equal to the minimum investment in her extended neighborhood. As a result, in connected graphs, all nodes must make identical security investments at a PNE. When Lc ≥ w′ (z), Proposition 8 states that a single node investing in isolation would prefer to invest s∗ = 1 − X. Now suppose all nodes have identical security investment s ≤ s∗ , i.e., the true attack probability at any node is 1 − s ≥ X. Since for each player w′ (x) > Lc for x > X (i.e., the marginal utility is positive), no player would unilaterally deviate to make a smaller investment. Therefore, any investment s ≤ 1 − X by all the nodes would result in a PNE. When w′ (xmin,w ) < Lc < w′ (z), the optimal investment of a single player is to invest 1. Therefore, a strategy profile where each node invests s = 1 − ǫ for sufficiently small ǫ > 0 such that Eu(s) > Eu(1 − X), is a PNE, with attack probability ǫ at every node. For this strategy profile, all players have positive marginal utility, and prefer to invest 1 − ǫ over 1 − X due to the continuity of the utility functions. These equilibria exist in addition to the set of equilibria with attack probabilities at least X. Note that any investment s ∈ (V, X) by the nodes is not a PNE since ∂Eu ∂s < 0 for s ∈ (V, X). Note that as long as the graph remains connected, its structure plays no role on the equilibrium investments and attack probabilities. The first part of the above result (i.e., identical investments by all players) holds for true expectation maximizers (players with w(x) = x) as well. The main differences with weighting functions are twofold. First, for large enough Lc , the only possible equilibrium with true expectation maximizers is when all players invest 0 (Proposition 1 with d = 1), while with probability weighting, there is a range of possible equilibrium investments, with resulting attack probabilities greater than X. Second, for w′ (xmin,w ) < Lc < 1, any investment by the players can give rise to a PNE for true expectation maximizers, while with probability weighting, there exist attack probabilities that can be supported at a PNE are either close to 0 or at least X. Finally, when Lc ≤ w′ (xmin,w ), X is not defined, though the optimal investment by a single player is still 1. In this case, any investment by the players can give rise to a PNE. We finally discuss the PNEs that arise in Best Shot games. Proposition 10. Consider a Best Shot network security game with homogeneous players. Then a strategy profile is a PNE if and only if there is a set of nodes who form a maximal independent set and invest according to Proposition 8, and all other nodes invest 0. Proof. In a Best Shot game, the attack probability on a node is a function of the highest investment in her extended neighborhood. Suppose at a neighborhood there is a player making an investment in accordance with Proposition 8. Since this is the optimal investment a single node can make, her neighbors do not find investing more than this level (and thereby reducing their attack probabilities) profitable. Therefore, their optimal strategy is to invest 0, as it eliminates their security investment cost csi , while the attack probability is unchanged. Therefore, any two nodes who make a nonzero investment must not be adjacent to each other. Furthermore, every node that is making a zero investment must have a neighbor who invests a nonzero amount. Therefore, the set of nodes making a nonzero investment must belong to a maximal independent set. The converse is also true; any maximal independent set with investments determined by Proposition 8, and all other nodes investing 0 constitute a PNE. The independent set characterization holds for true expectation maximizers as well; nonlinear weighting functions change the level of investment. With true expectation maximizers, the nonzero 14 1 4 2 3 9 5 7 6 8 10 Figure 4: Graph topology analyzed in Example 1 investment level is at one of the boundary points, either 0 or 1 (Proposition 1 with d = 1). With probability weighting, there are no equilibria that are entirely unprotected, and the nonzero equilibrium investment is at one of the interior solutions when Lc > w′ (z). 8 Numerical Examples We illustrate our theoretical characterizations of the impacts of network structure and probability weighting on the equilibrium investments in two numerical examples presented below. Example 1. Consider the graph shown in Figure 4 with n = 10 nodes. Each node is a decision maker with a Prelec weighting function with parameter α = 0.6 and Lc = 0.45. The extended neighborhood sizes range from 2 to 5. These parameters satisfy Assumptions 1, 2 and 3. Under the Total Effort externalities, sequential best response dynamics converged to a PNE strategy profile in this example. In this PNE, the leaf nodes, 1, 2, 3, 9, 10 have an investment 0.4095, node 7 invests 0.1442 and nodes 4, 5, 6, 8 invest 0. These investments satisfy (8). Note that node 7 has a larger neighborhood size compared to node 8, yet it invests more than node 8 at the PNE. This is in contrast with the equilibria that arise without probability weighting (Proposition 1), where the equilibrium investments were either 0 or 1 and the investment of a node was at most that of a node with a smaller degree. Furthermore, the investments of the leaf nodes are larger than their neighbors, as shown in Proposition 5. Under the Best Shot externality, a set of nodes which form a maximal independent set have a nonzero investment (Proposition 9). In the graph shown in Figure 4, nodes 1, 2, 3, 7, 9, 10 form a maximal independent set. Another such set consists of nodes 4, 2, 3, 8. For the Prelec weighting function with α = 0.6, w′ (z) = 0.8304 for the value of z defined in Lemma 4. Accordingly, if c L ≤ 0.8304, an equilibrium investment profile is when the nodes in a maximal independent set invest 1, while other nodes invest 0. On the other hand, if Lc > 0.8304, the equilibrium investment by the nodes is 1 − X, where w′ (X) = Lc , X ∈ ( 1e , 1]. In the above example, we discussed the effects of network structure on PNE investments in Total Effort and Best Shot games. The impact of the intensity of overweighting and underweighting of probabilities in degree-regular graphs is illustrated in the following example and corroborates our theoretical findings in Section 6. Example 2. We consider two degree-regular graphs on n = 6 nodes as shown in Figure 5. The nodes in the graph in Figure 5a (respectively, Figure 5b) have extended neighborhood sizes d = 3 (respectively, d = 5). We consider two types of players with Prelec weighting functions with parameters α1 = 0.4 and α2 = 0.8, respectively. The derivative of the weighting functions for such players was shown in Figure 3. Let Lc = 0.3 for all players. First we consider the 2-regular graph shown in Figure 5a. When all the players have weighting functions with parameter α1 = 0.4, then the attack probability at every node at the interior 15 1 2 6 1 3 5 2 6 4 3 5 (a) 2-regular graph on 6 nodes 4 (b) 4-regular graph on 6 nodes Figure 5: Graph topologies analyzed in Example 2 equilibrium is equal to 0.8588, while for players with parameter α2 = 0.8, the corresponding attack probability is 0.6912. In this case, the equilibrium is more secure for the players whose weighting functions are closer to linear. In contrast, in the 4-regular graph shown in Figure 5b, the attack probabilities at every node at the interior PNEs are equal to 0.9325 and 0.9643 for players with weighting function parameters α1 and α2 , respectively. As we noted in Theorem 3 and the succeeding discussion, the above example illustrates that the equilibrium is more secure for players with a substantial degree of overweighting and underweighting (captured by a smaller parameter value α) when the graph is sufficiently dense. 9 Discussion and Conclusion We studied a class of interdependent security games on networks where the players exhibit certain behavioral attributes vis-a-vis their perception of attack probabilities, while making security investment decisions. We analyzed three canonical interdependent security game models, (i) Total Effort, (ii) Weakest Link, and (iii) Best Shot games. The Nash equilibria in Total Effort games have much richer structural properties under behavioral probability weighting than the corresponding equilibria with players that are true expectation maximizers. A sharper overweighting of small probabilities may disincentivize a node from reducing her investment when her neighbors make high investments. Underweighting of large probabilities leads to equilibria where the nodes are not completely unprotected, as opposed to equilibria without probability weighting. The effect of behavioral probability weighting is most beneficial (in terms of reducing the security risk) when the probability of successful attack is sufficiently high at the PNE. On the other hand, if the attack probability is only moderately high, players with weighting functions that are closer to linear have a more secure equilibrium. We obtained an upper bound on the expected fraction of nodes that are successfully attacked at any PNE, in terms of the average degree of the nodes. Furthermore, among the class of graphs with a given average degree davg , the davg -regular graph has a PNE which achieves this upper bound. Conversely, star graphs achieve the smallest average security risk upper bound among all connected graphs. When the graph is connected, all nodes make identical investments in a Weakest Link game, and there is a continuum of attack probabilities, independent of the graph structure, that can arise at a PNE. On the other hand, a strategy profile is a PNE in a Best Shot game if and only if the nodes making nonzero security investments form a maximal independent set. In both cases, equilibria are never completely unprotected under behavioral probability weighting. References Saurabh Amin, Galina A Schwartz, and S Shankar Sastry. Security of interdependent and identical networked control systems. Automatica, 49(1):186–192, 2013. 16 Adam S Booij, Bernard MS Van Praag, and Gijs Van De Kuilen. A parametric analysis of prospect theory’s functionals for the general population. Theory and Decision, 68(1-2):115–148, 2010. Yann Bramoullé, Rachel Kranton, and Martin D’amours. Strategic interaction and networks. The American Economic Review, 104(3):898–930, 2014. Colin F Camerer, George Loewenstein, and Matthew Rabin. Advances in behavioral economics. Princeton University Press, 2011. Diego Cerdeiro, Marcin Dziubinski, and Sanjeev Goyal. Contagion risk and network design. Working paper, 2015. Nicolas Christin. Network security games: combining game theory, behavioral economics, and network measurements. In Decision and Game Theory for Security, pages 4–6. Springer, 2011. Nicolas Christin, Serge Egelman, Timothy Vidas, and Jens Grossklags. Its all about the Benjamins: An empirical study on incentivizing users to ignore security advice. In Financial Cryptography and Data Security, pages 16–30. Springer, 2011. Richard W Cottle, Jong-Shi Pang, and Richard E Stone. The linear complementarity problem, volume 60. SIAM, 1992. Andrea Galeotti, Sanjeev Goyal, Matthew O Jackson, Fernando Vega-Redondo, and Leeat Yariv. Network games. The Review of Economic Studies, 77(1):218–244, 2010. Vaibhav Garg and Joseph Camp. Heuristics and biases: Implications for security design. IEEE Technology and Society Magazine, 32(1):73–79, 2013. Bahman Gharesifard, Behrouz Touri, Tamer Başar, and Jeff Shamma. On the convergence of piecewise linear strategic interaction dynamics on networks. IEEE Transactions on Automatic Control, 61(6):1682–1687, 2016. Richard Gonzalez and George Wu. On the shape of the probability weighting function. Cognitive Psychology, 38(1): 129–166, 1999. Jens Grossklags and Benjamin Johnson. Uncertainty in the weakest-link security game. In Game Theory for Networks, pages 673–682, 2009. Jens Grossklags, Nicolas Christin, and John Chuang. Security and insurance management in networks with heterogeneous agents. In 9th ACM Conference on Electronic Commerce, pages 160–169, 2008. Assane Gueye, Jean C Walrand, and Venkat Anantharam. Design of network topology in an adversarial environment. In Decision and Game Theory for Security, pages 1–20. Springer, 2010. Jack Hirshleifer. From weakest-link to best-shot: The voluntary provision of public goods. Public Choice, 41(3): 371–386, 1983. Ashish R Hota and Shreyas Sundaram. Interdependent security games under behavioral probability weighting. In Decision and Game Theory for Security, pages 150–169. Springer, 2015. Ashish R Hota, Siddharth Garg, and Shreyas Sundaram. Fragility of the commons under prospect-theoretic risk attitudes. Games and Economic Behavior, 98:135–164, 2016. Libin Jiang, Venkat Anantharam, and Jean Walrand. How bad are selfish investments in network security? IEEE/ACM Transactions on Networking, 19(2):549–560, 2011. Benjamin Johnson, Jens Grossklags, Nicolas Christin, and John Chuang. Are security experts useful? Bayesian Nash equilibria for network security games with limited information. In ESORICS, pages 588–606. Springer, 2010. Howard Kunreuther and Geoffrey Heal. Interdependent security. Journal of Risk and Uncertainty, 26(2-3):231–249, 2003. Aron Laszka, Mark Felegyhazi, and Levente Buttyan. A survey of interdependent information security games. ACM Computing Surveys (CSUR), 47(2):23, 2014. 17 Tianming Li and Narayan B Mandayam. When users interfere with protocols: Prospect theory in wireless networks using random access and data pricing as an example. IEEE Transactions on Wireless Communications, 13(4): 1888–1907, 2014. Mohammad Hossein Manshaei, Quanyan Zhu, Tansu Alpcan, Tamer Başar, and Jean-Pierre Hubaux. Game theory meets network security and privacy. ACM Computing Surveys (CSUR), 45(3):25, 2013. R Miura-Ko, Benjamin Yolken, John Mitchell, and Nicholas Bambos. Security decision-making among interdependent organizations. In 21st Computer Security Foundations Symposium, pages 66–80. IEEE, 2008. Kien C Nguyen, Tansu Alpcan, and Tamer Başar. Stochastic games for security in networks with interdependent nodes. In International Conference on Game Theory for Networks, pages 697–703. IEEE, 2009. Cameron Nowzari, Victor M Preciado, and George J Pappas. Analysis and control of epidemics: A survey of spreading processes on complex networks. IEEE Control Systems Magazine, 36(1):26–46, 2016. Efe A Ok. Real analysis with economic applications, volume 10. Princeton University Press, 2007. Martin J Osborne and Ariel Rubinstein. A course in game theory. MIT press, 1994. Ranjan Pal and Pan Hui. Modeling internet security investments: Tackling topological information uncertainty. In Decision and Game Theory for Security, pages 239–257. Springer, 2011. Drazen Prelec. The probability weighting function. Econometrica, pages 497–527, 1998. Walid Saad, Arnold L Glass, Narayan B Mandayam, and H Vincent Poor. Toward a consumer-centric grid: A behavioral perspective. Proceedings of the IEEE, 104(4):865–882, 2016. Bruce Schneier. The psychology of security. In Progress in Cryptology–AFRICACRYPT 2008, pages 50–79. Springer, 2008. Galina Schwartz, Saurabh Amin, Assane Gueye, and Jean Walrand. Network design game with both reliability and security failures. In Allerton Conference on Communication, Control, and Computing, pages 675–681, 2011. Galina Schwartz, Nikhil Shetty, and Jean Walrand. Why cyber-insurance contracts fail to reflect cyber-risks. In Allerton Conference on Communication, Control, and Computing, pages 781–787. IEEE, 2013. Amos Tversky and Daniel Kahneman. Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5(4):297–323, 1992. Hal Varian. System reliability and free riding. In Economics of Information Security, pages 1–15. Springer, 2004. A Proofs Pertaining to Equilibrium Characterization Proof of Lemma 1: Proof. For ease of notation, we will drop the subscript i since the proof holds for every player. Let s̄ ∈ [0, d − 1] be the total security investment by the neighbors and d be the size of the extended neighborhood of the player. Under Assumption 3, the interval X (s̄) in (6) falls into one of four different cases. Case 1: X < 1 − 1+s̄ d (i.e., s̄ < d(1 − X) − 1) In this case, the interval X (s̄) lies to the right of X. Therefore, at any attack probability x ∈ X (s̄), Ld w′ (x) > c (from Figure 2). Thus, ∂Eu ∂s \|s=x > 0 and consequently, b(s̄) = 1. s̄ 1+s̄ Case 2: 1 − d ≤ X ≤ 1 − d In this case, X ∈ X (s̄), and therefore, the player has a feasible investment strategy s∗ = d(1 − X) − s̄ at which the first order condition is satisfied with equality. For any investment y < s∗ , ∂Eu L ′ we have the resulting attack probability 1 − y+s̄ d > X and ∂s = d w (X) − c > 0. As a result, no 18 value of y < s∗ would satisfy the first order necessary condition of optimality. On the other hand, 1 for any y > s∗ , we have 1 − y+s̄ d < X. However, from our assumption that X > d + V , we would y+s̄ ∂Eu ∗ ∗ have 1 − d > V . Therefore, ∂s < 0 for any y > s . As a result, x = s is the only candidate for optimal investment, and it also satisfies the second order sufficient condition since w′′ (X) > 0 ∗ as X > xmin,w . Since the optimal solution s∗ must have the property that 1 − s d+s̄ = X, it is continuous and linearly decreasing in s̄, with boundary values at 1 and 0 for s̄ = d(1 − X) − 1 and s̄ = d(1 − X), respectively. In the remaining two cases, we have s̄ > d(1 − X). s̄ Case 3: V < 1 − 1+s̄ d and X > 1 − d In this case, the interval X (s̄) lies in the region between V and X. Therefore, for any true attack probability x ∈ X (s̄), ∂Eu ∂s < 0. As a result, b(s̄) = 0. 1+s̄ Case 4: 1 − d ≤ V ≤ 1 − ds̄ In this case, there are three candidate solutions for utility maximization, s = 1, s = d(1 − V ) − s̄ and s = 0. We have 1 − s+s̄ d as the true attack probability resulting from the strategies of the players. First we show that the player would always prefer to invest 1 over investing s∗ = d(1 − V ) − s̄. We denote the resulting attack probability while investing 1 as Y1 = 1 − 1+s̄ d . Note that V − Y1 = 1 ∗ ). Now using (4) we compute (1 − s d Eu(1, sN ) − Eu(s∗ , sN ) =L(w(V ) − w(Y1 )) − c(1 − s∗ ) =L(w(V ) − w(Y1 )) − cd(V − Y1 ) w(V ) − w(Y1 ) dc =L(V − Y1 ) − (V − Y1 ) L dc >L(V − Y1 ) w′ (V ) − = 0, L where the inequality is due to the fact that w(x) is strictly concave for x ∈ [0, xmin,w ), and Y1 < V < xmin,w . Therefore, the player will always prefer the boundary solution s = 1 to the potential interior solution s∗ . This leads to the possibility that the best response might have a discontinuous jump from 0 to 1 at some value of s̄ in this region. However, we show that under the second and third conditions of Assumption 3, the player would always prefer to invest 0 over investing 1. We compute Eu(1, sN ) − Eu(0, sN ) 1 + s̄ s̄ −w 1− −c =L w 1 − d d 1 1 =L w λ + − w(λ) − c ≤ L w − w(0) − c, d d 1 where λ = 1 − 1+s̄ d . The last inequality follows because the function l1 (λ) , w(λ + d ) − w(λ) is a strictly decreasing function of λ for λ ∈ [0, V ]. Indeed, l1′ (λ) = w′ (λ + d1 ) − w′ (λ) < 0, as 1 dc 1 1 c ′ w′ (λ) ≥ w′ (V ) = dc L , and w (λ + d ) < L , as V ≤ λ + d < X. Therefore, if w( d ) < L , the player would always prefer to invest 0 over investing 1, regardless of the value of s̄ (including when s̄ = d − 1). Proof of Proposition 2: 19 Proof. Without loss of generality, let players be ordered such that X1 ≤ X2 ≤ X3 ≤ . . . ≤ Xn , nci ′ i where Xi is the largest solution to wi′ (x) = nc Li . When there is no solution to wi (x) = Li , i.e., i minx∈[0,1] w′ (x) > nc Li for a player, then we define Xi = 0 for that player. When Xn = 0, then all players investing 1 is the only PNE. The true attack probability at the PNE must be at least 0. When X1 > 0, no objective attack probability 0 ≤ X < X1 would be a PNE, since there would always exist a player with positive investment who would prefer to invest 0. Now suppose there are two PNEs with different probabilities of successful attack X ∗ and Y ∗ , with X ∗ < Y ∗ . Consider the strategy profile with the smaller attack probability X ∗ . Note that we have ruled out the possibility of X ∗ < X1 above. There are two exhaustive cases: either Xl < X ∗ < Xl+1 for some player l, or Xl = X ∗ for some player l. Let Xl < X ∗ < Xl+1 for some player l. By the definition of the quantities Xi , we have ′ ∗ i wi (X ∗ ) < nc Li for i ∈ {l + 1, . . . , n}, and therefore, si = 0 for i ∈ {l + 1, . . . , n}. Similarly, l ∗ ∗ i wi′ (X ∗ ) > nc Li and si = 1 for i ∈ {1, . . . , l}. In this case, X = 1 − n . Now at the second PNE ∗ ∗ with true attack probability Y > X , the players in {1, . . . , l} would continue to invest 1, with the possibility of more players investing nonzero amounts if Y ∗ ≥ Xl+1 . But then the true attack probability would decrease from X ∗ , contradicting the assumption that Y ∗ > X ∗ . The proof of the case where X ∗ = Xl for some player l follows identical arguments. B Linear Complementarity Problem Formulation We can exploit the structure of the best response as presented in (8) to represent the Nash equilibrium strategy profile as a solution to a Linear Complementarity Problem (LCP).4 Proposition 11. Consider a total effort game on a graph with n vertices with adjacency matrix A, and let dLi cii > wi′ (xmin,wi ) for every player i. Then, under Assumptions 1 and 3, a strategy profile is a Nash equilibrium of this game if and only if it is a solution of the LCP (q, M), where q ∈ R2n and M ∈ R2n×2n are given by −d(1 − X2 ) A + In In q= , and M = −In 0n×n 1n×1 where the jth entry of q is dj (1 − X2j ) for j ∈ {1, 2, . . . , n}. s Proof. The solution to the LCP (q, M) is a vector z ∈ which we can write as z = , where µ s is the vector of investments by the players (strategy profile) and µ is a set of variables which assume nonzero values when a corresponding investment si = 1. From the definition of a LCP, the solution z ≥ 0, i.e., s ≥ 0 and µ ≥ 0. Secondly, from q + Mz ≥ 0, for each player i, we have X si + sj + µi ≥ di (1 − X2i ), and R2n j∈N (i) − si + 1 ≥ 0. The second inequality ensures that si ≤ 1. Therefore, the investment by any player at any solution of the LCP is feasible, i.e., si ∈ [0, 1]. 4 Formally, given a vector q ∈ R n , and a matrix M ∈ R n×n , the LCP (q, M ) is the problem of finding a solution vector z ∈ R n such that i) z ≥ 0, ii) q + M z ≥ 0, and iii) z T (q + M z) = 0. 20 Finally, zT (q + Mz) = 0 gives us X si (si + sj + µi − di (1 − X2i )) = 0, and j∈N (i) µi (−si + 1) = 0, for every node i. As a result, when si ∈ (0, 1), we have µi = 0, and therefore, the investments satisfy the second part of equation (8). When si = 0, we also have µi = 0, and the resulting investments satisfy the third part of equation (8). Finally, when si = 1, we have µi > 0 such that the first part of equation (8) holds. This concludes the proof. A comprehensive discussion on LCPs and different solutions algorithms can be found in [Cottle et al., 1992]. The structure of the LCP often determines the performance of different algorithms. We show that the LCP defined in Proposition 11 satisfies certain properties, as proven in Proposition 12, that guarantee the convergence of the Lemke’s pivotal method to converge to a solution (i.e., a PNE strategy profile), if the problem is non degenerate. The convergence result is due to [Cottle et al., 1992]. Proposition 12. For the LCP (q, M) defined in Proposition 11, 1. the matrix M is copositive and 2. q ∈ SOL(0, M)∗ . Proof. For the proof of the first statement, consider any vector xT = [x1 i = 1, 2. Then we have, T T Ax1 + x1 + x2 x Mx = [x1 x2 ] −x1 x2 ]T where xi ≥ 0, for T = xT 1 Ax1 + x1 x1 ≥ 0. y1 For the second part, consider any solution y = of the LCP (0, M). Then, we must have y2 (A + In )y1 + y2 y ≥ 0, and My = ≥ 0. As a result, we must have y1 = 0. −y1 For any vector yT = y1T y2T , with y1 = 0 and y2 ≥ 0, we have qT y = 1T n×1 y ≥ 0. Thus, q ∈ SOL(0, M)∗ . C Proofs Pertaining to the Effects of Network Structure Proof of Lemma 2: Proof. We denote the inverse of w′ (·) as the function h(·), i.e., h : R+ → [xmin,w , ∞) such that 21 h( dc L ) = X. Now, consider the function dc φ(d) , d(1 − X) = d 1 − h L ′ dc dc −d h =⇒ φ′ (d) = 1 − h L L dc 1 dc = 1−h − ′′ L L w (h dc L ) dc 1 = 1−X − . L w′′ (X) The second inequality follows from differentiating the inverse of the w′ function. Now for x ∈ (xmin,w , 1), the weighting function w satisfies 1 w′′ (x) < w′ (x) 1−x dc 1 w′ (X) = =⇒ φ′ (d) < 0. =⇒ (1 − X) < ′′ w (X) L w′′ (X) Since φ(d) is strictly decreasing in d, we have d(1 − X) strictly decreasing when d is interpreted as the size of the extended neighborhood of a node. Proof of Proposition 6: Proof. We first show that when di (1 − Xi ) < 1, the attack probability at node i is at most Xi at a PNE. Recall from Lemma 1 that the best response of player i is bi (s̄−i ) = 1 when the aggregate investment by her neighbors s̄−i ≤ di (1 − Xi ) − 1. When di (1 − Xi ) < 1, the above condition is not satisfied since s̄−i ≥ 0. As a result, the investment of player i at any PNE lies in [0, 1). When s̄−i ≥ di (1 − Xi ), the investment by node i is 0, and the resulting attack probability is at most Xi . Otherwise, the sum of the investments by node i and her neighbors satisfy the first order condition (5) with equality, in which case the resulting attack probability is exactly Xi . The upper bound on average attack probability follows by averaging Xi ’s over all nodes. At an interior PNE, each player experiences an attack probability exactly equal to Xi and therefore the bound holds with equality. D Proofs Pertaining to the Impact of Weighting Function We first state and prove the following lemma whose result will be useful in the proof of Lemma 3. Lemma 5. The function g(x) , (− ln(x))α2 −α1 exp((− ln(x))α1 − (− ln(x))α2 ), with α1 < α2 < 1 is strictly decreasing for x ∈ [ 1e , 1]. 22 Proof. We compute g′ (x) = exp((− ln(x))α1 − (− ln(x))α2 )× −1 (α1 (− ln(x))α1 −1 − α2 (− ln(x))α2 −1 ) × x α2 −α1 (− ln(x)) α2 −α1 −1 1 − (α2 − α1 )(− ln(x)) x 1 = exp((− ln(x))α1 − (− ln(x))α2 )(− ln(x))α2 −α1 −1 x × [−(α1 (− ln(x))α1 + α1 + α2 (− ln(x))α2 − α2 ] . When x > 1e , (− ln(x)) < 1 and thus 1 > (− ln(x))α1 > (− ln(x))α2 . As a result, 1 − (− ln(x))α1 < 1 − (− ln(x))α2 . Since α1 < α2 , this implies α1 (− ln(x))α1 − α1 > α2 (− ln(x))α2 − α2 . Therefore, g′ (x) < 0 for x ∈ [ 1e , 1]. Proof of Lemma 3: Proof. The first derivative of the Prelec weighting function is given by w′ (x) = w(x) αx (− ln(x))α−1 . Therefore, if at a given x, w1′ (x) = w2′ (x), we have w1 (x)α1 (− ln(x))α1 = w2 (x)α2 (− ln(x))α2 w2 (x)(− ln(x))α2 α1 = =⇒ α2 w1 (x)(− ln(x))α1 α1 = (− ln(x))α2 −α1 exp((− ln(x))α1 − (− ln(x))α2 ) =⇒ α2 α1 = g(x). =⇒ α2 From the definition of g(x), g( 1e ) = 1 > αα12 . Furthermore, as x → 1, g ′ (x) → −∞. As a result, g(x) becomes smaller than αα12 for some x < 1. Thus there exists X̄ at which w1′ (x) = w2′ (x). The uniqueness of X̄ follows from the strict monotonicity of g(x) as proved in Lemma 5. In order to prove the second and third parts of the lemma, it suffices to show that w1′′ (X̄) > w2′′ (X̄). Therefore, we compute w′′ (x) = w′ (x) [1 + ln(x) + α((− ln(x))α − 1)]. −x ln(x) From the previous discussion in Lemma 5, α1 (− ln(x))α1 − α1 > α2 (− ln(x))α2 − α2 , and w′′ (x) > 0 for x > 1e . Therefore at X̄, w1′′ (X̄) > w2′′ (X̄). This concludes the proof. E Proofs Pertaining to Weakest Link and Best Shot Games Proof of Lemma 4: Proof. Consider the function l2 (x) , w′ (x) − w(xmin,w ) − xmin,w w(x) x for x ∈ [xmin,w , 1). At x = xmin,w , l2 (x) = < 0. Furthermore, l2 (1 − ǫ) → ∞ as ǫ → 0. Therefore, l2 (x) must have a root for x ∈ [xmin,w , 1). 23 Suppose there exist z1 , z2 with z1 < z2 such that w′ (zi ) = convex for x ∈ (xmin,w , 1], we have w(zi ) zi , i ∈ {1, 2}. Since w(x) is strictly w(z1 ) > w(z2 ) + w′ (z2 )(z1 − z2 ) = z1 w′ (z2 ) w(z1 ) > w′ (z2 ) > w′ (z1 ), =⇒ z1 which is a contradiction. Thus, z is unique. For the second part, suppose that there exists an x > z at which w′ (x) ≤ convexity of w, we have w(x) x . Again from the w(z) > w(x) + w′ (x)(z − x) ≥ zw′ (x) w(z) > w′ (x) > w′ (z), =⇒ z which contradicts the definition of z. Proof of Proposition 8: Proof. When there is a single player investing in isolation, we have three candidate solutions for utility maximization, s1 = 1 − V , s2 = 1 − X or s3 = 1. Here X and V are solutions to w′ (x) = Lc , as defined in Section 4. Note that since Lc is finite, we have ∂Eu ∂s > 0 at s = 0, so investing 0 in security is not a utility maximizer. From the analysis in Case 4 of Lemma 1 with s̄ = 0 and d = 1, we have Eu(1) > Eu(1 − V ). Therefore, between the potential interior solution s1 = 1 − V that satisfies the first order condition, and the boundary solution s3 = 1, the player always prefers the boundary solution. Now, to compare the utilities at the solutions s2 and 1, we compute Eu(1) − Eu(s2 ) = Lw(1 − s2 ) − c(1 − s2 ) w(1 − s2 ) ′ − w (1 − s2 ) , = L(1 − s2 ) 1 − s2 where Lc = w′ (1 − s2 ). Thus, from Lemma 4, when w′ (1 − s2 ) = Lc > w′ (z), then, 1 − s2 > z and the player prefers to invest s2 . Otherwise, the optimal investment is 1. 24	3cs.SY
1 Link Selection in Hybrid RF/VLC Systems under Statistical Queueing Constraints Marwan Hammouda, Sami Akın† , Anna Maria Vegni, Harald Haas, and Jürgen arXiv:1711.08175v1 [cs.IT] 22 Nov 2017 Peissig Abstract The co-deployment of radio frequency (RF) and visible light communications (VLC) technologies has been investigated in indoor environments to enhance network performances and to address specific quality-of-service (QoS) constraints. In this paper, we explore the benefits of employing both technologies when the QoS requirements are imposed as limits on the buffer overflow and delay violation probabilities, which are important metrics in designing low latency wireless networks. Particularly, we consider a multi-mechanism scenario that utilizes RF and VLC links for data transmission in an indoor environment, and then propose a link selection process through which the transmitter sends data over the link that sustains the desired QoS guarantees the most. Considering an ON-OFF data source, we employ the maximum average data arrival rate at the transmitter buffer and the non-asymptotic bounds on data buffering delay as the main performance measures. We formulate the performance measures under the assumption that both links are subject to average and peak power constraints. Furthermore, we investigate the performance levels when either one of the two links is used for data transmission, or when both are used simultaneously. Finally, we show the impacts of different physical layer parameters on the system performance through numerical analysis. Index Terms Visible light communications, quality-of-service constraints, buffering delay bound, link selection, multi-mechanisms. M. Hammouda, S. Akın, and J. Peissig are with the Institute of Communications Technology, Leibniz Universität Hannover, 30167 Hanover, Germany. E-mails: {marwan.hammouda, sami.akin, and peissig}@ikt.uni-hannover.de. A. M. Vegni is with COMLAB laboratory, Department of Engineering, Roma TRE University, 00146 Roma, Italy. E-mail: annamaria.vegni@uniroma3.it. H. Haas is with the Institute for Digital Communications, Li-Fi Research and Development Centre, the University of Edinburgh, Edinburgh, UK, EH9 3JL. E-mail: h.haas@ed.ac.uk. † This work was supported by the European Research Council under Starting Grant–306644. 2 I. I NTRODUCTION The ever-growing demand for mobile communications triggered a quest for technical solutions that will support stringent quality-of-service (QoS) constraints. Thanks to the significant advances in white light emitting diodes (LEDs) research, and the availability of an extensive unregulated spectrum, visible light communication (VLC) has emerged as a promising technology. We can utilize LEDs simultaneously for data transmission and illumination, since they have many unique aspects compared to the other communication technologies [1]. Moreover, we can improve data security because light does not penetrate the surrounding walls. We can also sustain an all-important green agenda and minimize the installation costs because we do not require an extensive infrastructure. Nevertheless, attention must be paid to certain limitations and challenges in VLC systems, e.g., smaller coverage, strong dependence on line-of-sight components and achievable rates that vary with spatial fluctuations [2]. In order to overcome these constraints, researchers proposed hybrid RF/VLC systems [1], [3]–[13], where end users can benefit from the wide coverage area that RF systems support and the stable rates that VLC systems provide. Such networks are practically feasible as RF and VLC systems can coexist without causing interference on each other and operate in the same environment, such as offices and rooms. Comparing hybrid RF/VLC systems with systems that employ either RF or VLC only, the authors in [1], [3], [4], [7], [14] demonstrated a remarkable increase in data transmission throughput, energy efficiency and delay performance in hybrid RF/VLC systems. Moreover, the authors in [5], [6] projected a hybrid system in which they use VLC links for down-link communication and RF links for up-link communication. In such a system, the authors in [8]– [10] and the ones in [12], [13], [15] investigated handover and load balancing mechanisms, respectively. Alternatively, considering an outdoor environment, the authors in [16] studied a point-to-point transmission scenario in which the system can switch between RF links and VLC links after comparing the signal-to-noise ratio levels in each link. Regarding the same system setting, the authors in [17] assumed that both RF and VLC links have the same transmission rates, and then proposed a diversity-based transmission scheme such that the transmitter sends data by employing both links simultaneously. The aforementioned studies analyzed the hybrid RF/VLC systems mostly from the physical layer perspective, i.e., they did not concentrate on the data link layer metrics, such as limits on the buffer overflow and buffering delay probabilities, as much as needed. Noting the dramatic 3 increase in the demand for reliable delay-sensitive services in recent years (mobile video traffic making up 55% of the total global data traffic at the end of 2014 [18]), in addition to the physical layer performance metrics, we need QoS metrics that can be a cross-layer analysis tool between the physical layer features and the performance levels in data-link layer. In this context, the authors in [19]–[25] performed cross-layer analysis between physical and data-link layers in many different RF scenarios. Regarding a Markovian data arrival process at a transmitter buffer and a statistically varying data service process in a wireless channel, the authors in [25] characterized the maximum average data arrival rate at the transmitter buffer with the presence of statistical constraints on the buffer overflow probability. On the other hand, to the best of our knowledge, there are only a few studies that investigated cross-layer performance levels in VLC systems. For example, the authors employed effective capacity as a performance measure in resource allocation schemes in VLC systems [26] and heterogeneous networks composed of VLC and RF links [27]. Here, we note that the authors in [27] concentrated on the case of constant data arrival rates at the transmitter buffer, which is not realistic in certain practical settings. For more details in effective capacity, we refer to [20]. In this paper, assuming an ON-OFF modeled data arrival process at the transmitter buffer and that the transmitter can use both RF and VLC channels for data transmission, we investigate the performance gains achieved by a hybrid RF/VLC system. This operates under statistical QoS constraints, which are inflicted as limits on the buffer overflow and delay violation probabilities. We perform a cross-layer analysis of these hybrid systems in physical and data-link layers. We employ first the maximum average data arrival rate at the transmitter buffer considering the asymptotic buffer overflow probability approximation, and then non-asymptotic buffering delay violation probability as the main performance measures. We propose a mathematical toolbox to system designers for performance analysis in hybrid RF/VLC systems that work under low latency conditions. To summarize, the main contributions of this paper are as follows: 1) Assuming that both RF and VLC links are subject to average and peak power constraints, we express the maximum average data arrival rate at the transmitter that the data service process from the transmitter to the receiver can support under QoS constraints when either the RF or VLC link is used, or both links are simultaneously used for data transmission. 2) We propose three different link usage strategies. We base two of the proposed strategies on the assumption that the receiver does not have a multihoming capability, thus data transmission is possible over only one link, either the RF or VLC link. In the third strategy, 4 we assume that link aggregation is possible and data can be transmitted over both links simultaneously following a power sharing policy. 3) We obtain the non-asymptotic data backlog and buffering delay violation probability bounds considering the proposed link usage strategies. Particularly, we provide a rudimentary model for multi-mechanisms in communication systems that operate under QoS constraints. We employ RF and VLC links as two different mechanisms. Our model can be easily invoked in settings with more than two different mechanisms as well. The reason behind multi-mechanisms in communications is to boost performance levels through the increased degree of freedom. Therefore, in order to introduce our model smoothly and make it easier for readers to understand our objective, we also benefit from the existing literature in RF and VLC studies. However, to the best of our knowledge, the analytical framework provided in this paper, in which we investigate the QoS performance, is not addressed in other studies. One aspect of this hybrid system is that VLC links provide time-invariant transmission rates, while RF links provide rates that vary over time. A communication setting that depends solely on an RF link may suffer low transmission rates, and have longer data backlogs in the transmitter buffer. However, a communication setting that can utilize both RF and VLC links, for instance, can take advantage of the constant transmission rate in the VLC link when the transmission rate in the RF link falls below a certain level. The rest of the paper is organized as follows. We introduce the hybrid RF/VLC system in Section II. We provide detailed descriptions of both RF and VLC channels. In Section III, we provide the performance analysis and the link selection process. In Section IV, we substantiate our results with numerical demonstrations. We conclude the paper in Section V and relegate the proofs to the Appendix. II. S YSTEM M ODEL We consider a network access controller that provides a connection to a user1 , through either an RF access point or a VLC access point, which are positioned at different locations in an indoor environment as seen in Figure 1. Herein, we assume a down-link scenario, i.e., the network access controller acts as a transmitter and the user acts as a receiver. In the sequel, we use transmitter and receiver instead of network access controller and user, respectively. Initially, 1 The analytical framework provided in this paper can easily be extended to a multi-user scenario. For more details, we refer to Remark 4 in Section III. 5 Fig. 1. Hybrid RF/VLC system. the transmitter receives data from a source (or sources) and stores it as packets in its buffer. Subsequently, it sends the data packets in frames of T seconds to the receiver following a given transmission strategy. The receiver is considered to be equipped with an RF front-end and a photo-diode. We also note that the VLC coverage area is generally smaller than the RF coverage area, and that they overlap2 . Finally, we consider a power-limited system and assume that the network controller is constrained by a fixed average power budget, denoted as Pavg , for data transmission. In the sequel, we initially introduce the RF and VLC channels, and then describe the data source model. A. RF Channel Model During the data transmission in the flat-fading RF channel, the input-output relation at time instant t is expressed as yr (t) = xr (t)h(t) + wr (t), (1) where xr (t) and yr (t) are the complex channel input and output at the RF access point of the transmitter and the RF front-end of the receiver, respectively. The complex channel input is subject to an average power constraint, Pavg,r , i.e., E{\|xr (t)\|2 } ≤ Pavg,r , and a peak power constraint, Ppeak,r , i.e., \|xr (t)\|2 ≤ Ppeak,r . Above, h(t) is the complex channel fading gain with an arbitrary distribution having a finite average power, i.e., E{\|h(t)\|2 } < ∞. Furthermore, we consider a block-fading channel and assume that the fading gain stays constant during one transmission frame (T seconds), i.e., h(lT ) = h(lT + B1r ) = · · · = h((l + 1)T − B1r ) = hl , where the available bandwidth is Br Hz in the channel and hl is the channel fading gain in the lth time frame. Note that there are T Br symbols transmitted in one time frame. Moreover, the fading gain changes independently from one frame to another. Meanwhile, wr (t) is the additive noise at 2 This model is also considered in [9]. 6 the RF front-end of the receiver, which is a zero-mean, circularly symmetric complex Gaussian random variable with variance σr2 , i.e., E{\|wr (t)\|2 } = σr2 < ∞. The noise samples {wr (t)} are assumed to be independent and identically distributed. We assume that a reliable data transmission exists as long as the transmission rate in the channel is lower than or equal to the instantaneous mutual information between the channel input and output3 . In particular, when the transmission rate in the lth time frame (i.e., Rl bits per frame) is lower than or equal to the instantaneous mutual information (i.e., Cr,l bits per frame) between the channel input [xr (lT ), xr (lT + B1r ), · · · , xr ((l + 1)T − B1r )] and the channel output [yr (lT ), yr (lT + 1 ), · · · Br , yr ((l + 1)T − 1 )], Br a reliable data transmission occurs and Rl bits are decoded correctly by the receiver. Here, we assume that T Br is large enough so that the decoding error probability is negligible when Rl ≤ Cr,l . In [30], a lower bound on the maximum mutual information (or channel capacity) is provided, where the input has a two-dimensional circularly truncated Gaussian distribution. Therefore, we set the instantaneous data transmission rate to the lower bound and assume that the input has a two-dimensional circularly truncated Gaussian distribution. Specifically, we have a reliable transmission in the lth time frame when bPavg,r 2\|hl \|2 exp −1 ≤ Cr,l bits per frame, (2) Rl = T Br log2 1 + aσr2 2 where a and b are the solutions of the following equations: bPpeak,r a 1 − exp − = 1, b 2 (2a) and bPpeak,r bPpeak,r Pavg,r a −1 2 (bPpeak,r ) 1 − exp − 1+ = . b 2 2 Ppeak,r (2b) Above, Cr,l is time-dependent and changes from one time frame to another because the maximum instantaneous mutual information in each frame is a function of the channel fading gain. B. VLC Channel Model We assume that the transmitter employs intensity modulation/direct detection. Principally, the VLC access point of the transmitter is equipped with an LED and the data is modulated on the intensity of the emitted light. The receiver that collects light using a photo-diode generates an 3 This assumption is based on the known result in the literature that transmitting data at rates less than or equal to the instantaneous mutual information has a high reliability, thus the decoding error is negligible [28, Eq. (10)] [29]. 7 electrical current or voltage proportional to the intensity of the received light. Besides, we know that VLC channels are typically composed of line-of-sight as well as multi-path components. However, the majority of the collected energy at photo-diodes (more than 95%) comes from the line-of-sight components in typical indoor scenarios [31]. Therefore, we can assume that the VLC channel is flat with a dominant line-of-sight component [32]–[34], and the channel gain does not vary during the data transmission as long as the receiver is stationary4 . Accordingly, the input-output relation in the VLC channel between the VLC access point of the transmitter and the photo-diode of the receiver at time instant t is given as follows: yv (t) = Ωxv (t)g + wv (t), (3) where xv (t) and yv (t) are the real-valued channel input and output, respectively. Above, Ω is the optical-to-electrical conversion efficiency (or detector responsivity) of the photo-diode in amperes per watt and wv (t) is the additive noise at the photo-diode of the receiver, which is a zero-mean, real Gaussian random variable with variance σv2 , i.e., E{wv2 } = σv2 . The noise samples {wv } are independent and identically distributed. Moreover, g is the time-invariant optical channel gain. Recall that the data transmitted over the VLC link is modulated on the light that illuminates the environment. Hence, assuming that the operation range of the radiated optical power is limited between Pmin and Pmax when the light is on, we modulate the data between the power levels Pmin and Pmax , i.e., Pmin ≤ xv (t) ≤ Pmax . As a result, the data bearing symbol, x̃v = xv (t) − Pmin is limited as follows: x̃v ≤ Pmax − Pmin = Ppeak,v . Hence, we can re-express the input-output relation in (3) as ỹv (t) = yv (t) − ΩPmin g = Ωx̃v (t)g + wv (t). (4) Now, assuming that the expected value of x̃v (t) is bounded by Pavg,v , i.e., E{x̃v (t)} ≤ Pavg,v , and that the available bandwidth in the optical channel is Bv Hz, we set the transmission rate in the channel in bits per frame to the lower bound on the channel capacity, which is defined as follows [36]: V = 4 T Bv log2 2 ( ) −µ? 2 Pavg,v ? Ωg 1−e 2 exp 2 ≤ Cv , 1 + Ppeak,v µ −1 2 2πσv Ppeak,v µ? 2 2 (5) Small-scale variations in VLC channels (i.e., fading) is mitigated since the area of a photo-diode is much larger than the light wavelength [35, Sec. 2.5]. Thus, VLC channels are known as time-invariant. This fact is almost true regardless of the frame duration or the user being stationary or mobile in indoor environments, because the users are either stationary or move very slowly. 8 Fig. 2. State transition model of the data arrival process. when 0 ≤ Pavg,v Ppeak,v < 21 , and T Bv ω 2 g 2 −1 2 V = ≤ Cv , log2 1 + Ppeak,v e 2 2πσv2 when 1 2 ≤ Pavg,v Ppeak,v (6) ≤ 1, where µ? is the unique solution to Pavg,v 1 e−µ = − . Ppeak,v µ 1 − e−µ (7) Above, Cv and V are constant values, because the mutual information in the VLC link does not change by time, i.e., due to the strong line-of-sight channel component, the channel gain, g, does not change. As for the input distribution, we refer to [36, Eq. 42]. C. Source Model Regarding the data arrival process at the transmitter buffer, we consider a two-state discretetime Markov process5 with ON and OFF states in each time frame. When the source is in the ON state in one time frame, the data from a source (or sources) arrives at the transmitter buffer. In the ON state, we consider a constant data arrival rate, i.e., λ bits per frame. The number of bits arriving at the transmitter buffer is zero in the OFF state. As shown in Fig. 2, the transition probability from the ON state to the OFF state is denoted by α and the transition from the OFF state to the ON state is denoted by β. The probability of staying in the ON state is 1 − α and the probability of staying in the OFF state is 1 − β. Hence, the state transition matrix becomes   1−α α . J = (8) β 1−β Now, let pON and pOFF be the steady-state probabilities of the data arrival process being in the ON and OFF states, respectively, where pON + pOFF = 1. Then, we have the following equality: 5 We can project certain data arrival models on ON-OFF Markov processes. For instance, voice sources are generally modeled with ON and OFF states [37]. 9 [pON pOFF ] = [pON pOFF ]J. Subsequently, we have pON = average data arrival rate at the transmitter buffer is βλ α+β β α+β and pOFF = α , α+β and hence, the bits per frame. We finally note that the following analysis can easily be extended to other source models. III. P ERFORMANCE A NALYSIS In this section, we investigate the performance levels that the aforementioned system achieves by opportunistically exploiting the RF and VLC channels for data transmission. Herein, because the data is initially stored in the transmitter buffer before transmission, we assume that certain constraints are applied on the amount of the data in the buffer and the buffering delays. Therefore, we examine the system under QoS constraints that are associated with buffer overflow (data backlog) and buffering delay, and express the decay rate of the tail distribution of the queue length as [19, Eq. (63)] loge Pr{Q ≥ q} , q→∞ q θ = − lim (9) where Q is the stationary queue length, Q(t) (see Fig. 1), and q is the buffer overflow threshold. Above, θ > 0 denotes the decay rate of the tail distribution of the data backlog, Q. Accordingly, we can approximate the buffer overflow probability for a large threshold, qmax , as Pr{Q ≥ qmax } ≈ e−θqmax . Notice that the buffer overflow probability decays exponentially with a rate controlled by θ, which is also defined as the QoS exponent. Basically, larger θ implies stricter QoS constraints, whereas smaller θ corresponds to looser constraints. Recall that the outgoing service from the transmitter queue is Rl , given in (2), when the data is sent through the RF channel and it is V , given in (5) and (6), when the data is sent through the VLC channel, while the data arrival is ON-OFF Markov process with λ bits per frame in the ON state and zero bits in the OFF state. Hence, assuming that the buffer size is infinite and considering independent data arrival and work-conserving data service processes, there exist unique θr? > 0 and θv? > 0 such that Λa (θr? ) = −Λr (−θr? ) and Λa (θv? ) = −Λv (−θv? ), respectively, where Λa (θ), Λr (θ) and Λv (θ) are the asymptotic log-moment generating functions of the total amount of bits arriving at the transmitter buffer, the total service from the transmitter in the RF channel and the total service from the transmitter in the VLC channel, respectively [38, Theorem 10 2.1]. In particular, the asymptotic log-moment generating functions for any θ are q    1 − β + (1 − α)eθλ + [1 − β + (1 − α)eθλ ]2 − 4(1 − α − β)eθλ  Λa (θ) = loge ,   2 Λr (θ) = loge Eh eθRl and Λv (θ) = θV. (10) (11) We refer to [39, Example 7.2.7] for obtaining the log-moment generating functions. Using Λa (θr? ) = −Λr (−θr? ), we can express the maximum average data arrival rate at the transmitter buffer that the service process in the RF channel can sustain for any θ > 0 as −2Λr (−θ) e − (1 − β)e−Λr (−θ) β ρr (θ) = loge (α + β)θ (1 − α)e−Λr (−θ) − (1 − α − β) ( ) 1 − (1 − β)Eh e−θRl β = loge , (α + β)θ (1 − α)Eh {e−θRl } − (1 − α − β)E2h {e−θRl } (12) where Rl is given in (2). For the derivation of ρr (θ), we refer to Appendix A. Likewise, using Λa (θv? ) = −Λv (−θv? ) and following the steps in Appendix A, we can also express the maximum average data arrival rate at the transmitter buffer that the service process in the VLC channel can sustain for any θ > 0 as e−2Λv (−θ) − (1 − β)e−Λv (−θ) (1 − α)e−Λv (−θ) − (1 − α − β) β e2θV − (1 − β)eθV = loge , (α + β)θ (1 − α)eθV − (1 − α − β) β ρv (θ) = loge (α + β)θ (13) where V is given in (5) and (6) accordingly with the relation between Pavg and Ppeak . Moreover, in the special case where α = 0 and β = 1, the expressions in (12) and (13) provide the effective capacity, which is the maximum sustainable constant data arrival rate by the channel process given the QoS constraints [20]. In another special case where α +β = 1, i.e., the state-transitions are independent of the past and current states, we have ) ( 2θVl 1 − αEh e−θRl β e − αeθVl β and ρv (θ) = loge . ρr (θ) = loge θ βEh {e−θRl } θ βeθVl (14) A. Link Selection Policy In this section, we focus on the channel selection process that the transmitter employs. We set the maximum average data arrival rate under QoS constraints as the objective in the channel selection process. In particular, the transmitter chooses the channel in which the service process maximizes the average data arrival rate at the transmitter buffer. Notice that the transmission 11 rate in the VLC channel is constant, whereas the transmission rate in the RF channel varies due to the changes in the channel fading gain. Now, due to the fact that the channel fading gains are known by the receiver as well as the transmitter, we provide the following proposition. Proposition 1: In the aforementioned RF/VLC system, the transmitter sends data to the receiver over the VLC link when the following condition for a given QoS exponent, θ, holds: q   2   [1 − β + (1 − α)ξ] − 4(1 − α − β)ξ 1 − β + (1 − α)ξ + 1 V ≥ loge ,   θ 2 where 1 − (1 − β)Eh e−θRl . ξ= (1 − α)Eh {e−θRl } − (1 − α − β)E2h {e−θRl } (15) (16) Proof: See Appendix B. Proposition 1 states that if the maximum attainable transmission rate in the VLC channel is greater than the right-hand side of (15), the transmitter should perform transmission over the VLC link because the statistical variations in the RF channel deteriorates the buffer stability. Meanwhile, in the special case when α + β = 1, we re-express (15) as 1 V ≥ − loge Eh e−θRl θ . (17) Specifically, the constant rate in the VLC channel should be greater than the effective capacity of the RF channel such that the VLC channel is chosen for data transmission. In the following, we present two transmission strategies, such that (i) the data is transmitted over the link with the highest instantaneous transmission rate in one time frame, and (ii) the data is transmitted over both links simultaneously. Hybrid-Type I Transmission Strategy: In the above analysis, we obtain the performance levels when the transmitter chooses either of these two channels for data transmission following a link selection process based on the maximum average data arrival rates that the service processes in the channel can support. On the other hand, if there exists a fast and stable handover mechanism between the transmitter and the receiver, the transmitter will forward the data to the receiver over the link that provides the maximum lower bound on the instantaneous mutual information in the corresponding channel. For instance, when the lower bound on the instantaneous mutual information in the RF channel in the lth time frame, Rl , is greater than the lower bound on the instantaneous mutual information in the VLC channel, V , i.e., Rl > V , the transmitter sends the data over the RF link in the corresponding time frame only. Otherwise, it prefers sending the 12 data over the VLC link. Respectively, we can establish the channel selection criterion as follows: The transmitter sends the data over the RF link when bPavg,r aσr2 TVB 2 \|hl \| > 2 r − 1 exp 1 − = κ. 2 2 (18) Otherwise, it sends the data over the VLC link. The aforementioned selection test can also be considered as the outage condition in the RF channel, i.e., the RF link is in outage when \|hl \|2 ≤ κ. Noting that the channel fading gain changes independently from one time frame to another in the RF channel, the log-moment generating function of the service process becomes Λrv (θ) = loge E\|h\|2 >κ eθRl + Pr{\|hl \|2 ≤ κ}eθV , (19) where E\|h\|2 >κ {eθRl } is the conditional expectation given that \|h\|2 > κ, i.e., E\|h\|2 >κ {eθRl } = R ∞ θR 1 e l f\|h\|2 (\|h\|2 )d\|h\|2 , where f\|h\|2 (\|h\|2 ) is the probability density function of \|h\|2 . Pr{\|hl \|2 >κ} κ Hence, the maximum average data arrival rate at the transmitter buffer that the hybrid service process can sustain under the QoS constraints specified by θ > 0 becomes β 1 − (1 − β)D loge ρrv (θ) = , (α + β)θ (1 − α)D − (1 − α − β)D2 where D = E\|h\|2 >κ e−θRl + Pr{\|hl \|2 ≤ κ}e−θV . (20) Hybrid-Type II Transmission Strategy: Different than the aforementioned protocols, we consider a transmitter that sends data over both links simultaneously in each frame. We assume that the receiver has a multihoming capability. We further assume a multiplexing-based transmission scheme such that the data streams transmitted over the RF and VLC links are different and independent from each other. Indeed, this scenario is feasible since the light and RF waves do not cause interference on each other. We further assume a power allocation policy between the two links, i.e., the average power constraint in the RF link is set to Pavg,r = γl Pavg , and the one in the VLC link is set to Pavg,v = (1 − γl )Pavg , where Pavg is the total average power constraint, and 0 < γl < 1. Then, the instantaneous transmission rate in the RF channel in the lth time frame becomes bγl Pavg 2\|hl \|2 exp −1 Rl = T Br log2 1 + aσr2 2 bits per frame, (21) where a and b are the solutions of (2a) and (2b). Similarly following (5)–(6), the transmission rate in the VLC channel becomes ( ) 2 2 −µ? 2 (1 − γ )P Ω g 1 − e T Bv l avg ? 2 V = log2 1 + Ppeak,v exp 2 µ −1 , 2 2πσv2 Ppeak,v µ? (22) 13 when 0 ≤ (1−γl )Pavg Ppeak,v < 12 , and T Bv Ω2 g 2 −1 2 V = log2 1 + Ppeak,v , e 2 2πσv2 when 1 2 ≤ (1−γl )Pavg Ppeak,v (23) ≤ 1, where µ? is the unique solution of (7). It follows that the total transmission rate in each frame is the sum of the transmission rates in both links, i.e., Rl + V . Then, the log-moment generating function can be readily expressed as Λsrv (θ) = loge E{eθ(Rl +V ) } = θV + loge E{eθRl }, and the maximum average data arrival rate is equal to 1 − (1 − β)D β loge , ρsrv (θ) = (α + β)θ (1 − α)D − (1 − α − β)D2 (24) (25) where D = e−θV Eh {e−θRl }. We finally remark that the optimal value of γl that maximizes the sum transmission rate, i.e., maxγl {Rl + V }, can be obtained numerically. Remark 1: Employing the aforementioned strategies requires the perfect knowledge of both RF and VLC channels at the transmitter side in each frame. We assume that the channel estimation is performed at the receiver, and forwarded to the transmitter in a delay-free and error-free feedback channel. This increases the signaling overhead. Moreover, applying Hybrid-Type II Transmission Strategy increases the implementation complexity because power sharing should be performed in each transmission frame. From the implementation perspective, exploiting Hybrid-Type II Transmission Strategy is limited because the receiver should have a multihoming capability that enables it to perform link aggregation and receive data from different transmission technologies simultaneously. On the other hand, the maximum average data arrival rate is a steady-state performance measure. Thus, the selection process explained in Proposition 1 is considered as a large-timescale operation, which can be performed over periods of multiple transmission frames, and therefore, the implementation complexity decreases. Such a large-timescale operation is also proposed in [40]. Remark 2: The selection process can also be employed in cases the access point controller has to choose between more than two transmission links. For example, let there be N transmission links, and let {ρ1 (θ), . . . , ρN (θ)} be the maximum average data arrival rates at the transmitter buffer sustained by these links. Then, Proposition 1 in the paper will be updated with the solution of the following maximization problem: Transmission Link = max {i : ρi (θ)} . (26) 14 Similarly, let {R1 (l), · · · , RN (l)} be the instantaneous transmission rates provided by each link in the lth time frame. Then, the link selection criterion in Hybrid-Type I Transmission Strategy will be updated with the solution of the following maximization problem: Transmission Link = max {i : Ri } . (27) Subsequently, the log-moment generating function of the service process will be Λ1,...,N (θ) = loge Pri EeθRi (l) , (28) where Pri is the probability that the transmitter chooses link i for data transmission. Finally, if the receiver has a multihoming capability, Hybrid-Type II Transmission Strategy can be applied with power sharing to maximize the total transmission rate in each frame. B. Impacts of Handover Delay Handover delay occurs when the transmitter moves from one link to the other, which is the case in Hybrid-Type I Transmission Strategy. In this strategy, data transmission in one time frame is performed over the link that provides the maximum instantaneous transmission rate in that frame. Particularly, the transmitter switches from one link to the other or stay in the same link at the end of any time frame after comparing the instantaneous transmission rates in both links. Now, given that TH denotes the duration of one single handover phase, let us initially assume that the frame duration is larger than the handover phase, i.e., T > TH . For the sake of simplicity, we divide the frame duration into n sub-frames that are equal to TH , i.e., T = n × TH , where n ∈ N and n > 1. Particularly, a series of n sub-frames of data transmission phase is followed by one sub-frame of handover process if the transmitter changes the transmission link, or by another series of n sub-frames of data transmission phase if the transmitter stays in the same transmission link. For an analytical representation, we model the buffer activity at the end of each sub-frame as a discrete-time Markov process. As shown in Fig. 3, we have 2n + 2 states. The first n states, i.e., {State 1,· · · ,State n}, represent the data transmission sub-frames in the VLC link, and State (n + 1) represents the handover process from the VLC link to the RF link. Similarly, the subsequent n states, i.e., {State (n + 2),· · · ,State (2n + 1)}, represent the data transmission sub-frames in the RF link, and State (2n + 2) represents the handover process from the RF link to the VLC link. Notice that the state transition probability from State i to State 15 i + 1 is 1 for i ∈ {1, · · · , n − 1} and for i ∈ {n + 2, · · · , 2n} because the data transmission in each link is completed in n sub-frames and the link change may occur at the end of the nth and (2n + 1)th sub-frames. On the other hand, in State n, either the transmitter changes the link and the system enters State n + 1 with probability 1 − δν , or the transmitter stays in the same link and the system enters State 1 with probability δν . Similarly, in State (2n + 1), either the transmitter changes the link and the system enters State (2n + 2) with probability 1 − δr , or the transmitter stays in the same link and the system enters State (n + 2) with probability δr . Finally, the system moves from State (n + 1) to State (n + 2) and from State (2n + 2) to State 1 with probability 1 because at the end of one handover phase, the transmitter starts data transmission. As seen in (18), if \|hl \|2 > κ, the transmitter sends the data over the RF link. Otherwise, it sends the data over the VLC link. Therefore, for this specific case, we have δr = Pr{\|hl \|2 > κ} = δ and δv = 1 − δ. Then, we can express    1,        1,    Γ = [pji ], where pji = δ,       1 − δ,      0, the (2n + 2) × (2n + 2) transition matrix as for j = i + 1 and 1 ≤ i ≤ n − 1 or n + 2 ≤ i ≤ 2n, for (i, j) = (n + 1, n + 2) or (i, j) = (2n + 2, 1), for (i, j) = (n, n + 1) or (i, j) = (2n + 1, n + 2), for (i, j) = (n, 1) or (i, j) = (2n + 1, 2n + 2), otherwise, (29) where pji is the state transition probability from State i to State j. Then, we can re-characterize the log-moment generating function of the hybrid system for any θ > 0, which is provided in (19), as follows [41, Chap. 7, Example 7.2.7]: Λrv (θ) = loge sp(Φ(θ)Γ), (30) where sp(Φ(θ)Γ) is the spectral radius of the matrix Φ(θ)Γ, and Φ(θ) is a diagonal matrix whose components are the moment generating functions of the processes in 2n+2 states. Notice that the transmitted bits are removed from the transmitter buffer only at the ends of the nth and (2n + 1)th frames. Therefore, the moment generating functions are eθV and E\|h\|2 >κ {eθRl } in the nth and (2n + 1)th frames, respectively. However, there are no bits removed in the other states, i.e., the service rates in the other states are effectively zero. Hence, the moment generating functions are 1 16 Fig. 3. State transition model of the hybrid scenario with handover. in other states. Moreover, the unique QoS exponent, θ? , is obtained when Λa (θ? ) = −nΛrv (−θ? ). C. Non-asymptotic Bounds The aforementioned results provide the performance analysis in the steady-state. Particularly, the analysis is obtained when the number of time frames is very large. On the other hand, nonasymptotic bounds regarding the statistical queueing and delay characterizations at the transmitter buffer are of interest for system designers as well. Therefore, we address the framework of Network Calculus [39], [42], [43], and consider [44, Theorem 2], which states that a minimal bound on the queue length can be found for a given buffer overflow probability. Particularly, given the RF-based data service process in Section II-A and the two-state Markov modeled data arrival process in Section II-C, the buffer threshold, q, is expressed as q = inf {qr + qa }, (31) c>0 for a given buffer overflow probability, Pr{Q ≥ q} = ε, where 1 Λr (−θ) loge {−εr [Λr (−θ) + θc]} for max 0, − − < θ, qr = − sup θ cεr c θ (32) and qa = − sup θ loge {εa [θc − supt>0 {Λa (θ, t)}]} θ for 0 < θ < 1 supt>0 {Λa (θ, t)} + , (33) cεa c with    (t−1)        eθλ 0 1 − α α eθλ 0 1  1           . Λa (θ, t) = loge [pON pOFF ]  t   0 1 β 1−β 0 1 1  (34) Above, the buffer violation probability is ε = εr + εa , and θ and c are free parameters. Notice also that the log-moment generating function in (32), i.e., Λr (−θ), is time-invariant because the service process is memory-less, whereas the log-moment generating function in (33), i.e., 17 Λa (θ, t), is time-variant because the arrival process depends on its current state. Moreover, we remark that when t goes to infinity, we have limt→∞ Λa (θ, t) = Λa (θ), where Λa (θ) is expressed in (10). Herein, we refer to [44] for further calculation details in (32) and (33). Likewise, when the VLC-based data service process in Section II-B is employed, we have q = inf {qv + qa }, (35) c>0 and qv = − sup θ loge {−εr [Λv (−θ) + θc]} loge {−εr [−θV + θc]} = − sup . θ θ θ (36) Notice that q in (35) is minimized when c = V . Moreover, because qv cannot be smaller than zero, we have q = qa and c = V . Therefore, when the VLC-based service channel is chosen as the service process, we have q = qa = − sup θ loge {εa [θV − supt>0 {Λa (θ, t)}]} θ and ε = εa . (37) Now, assuming a fast and stable handover mechanism between the transmitter and the receiver, and a service channel selection process as described in Hybrid-Type I Transmission Strategy, we can characterize the delay bound as follows: q = inf c>0 {qa + qrv } where loge {−εr [Λrv (−θ) + θc]} 1 Λrv (−θ) qrv = − sup for max 0, − − < θ, θ cεr c θ (38) and Λrv (θ) is given in (19). Remark 3: Let us assume a first-come first-served protocol exists at the transmitter buffer. Then, the minimal bound on the buffering delay is expressed as follows [44, Theorem 1]: qr + qa qv + qa qrv + qa , or d = inf , or d = inf , (39) d = inf c>0 c>0 c>0 c c c when the RF-based service process, or the VLC-based service process, or Hybrid-Type I Transmission Strategy is employed. Remark 4: We consider a user-related performance measure, i.e., the maximum average data arrival rate at the data buffer, and formulate the link selection by employing the transmission rates provided by the RF and VLC links. Our analytical framework can easily be extended to a more general multi-user scenario by regarding the rate allocations for each user in the transmission links and the receiver-oriented data arrival processes at the transmitter buffer. In this regard, we refer to Fig. 7 in Section IV, where we employ frequency-division multiple access (FDMA) and time-division multiple access (TDMA) protocols for numerical illustrations. Moreover, our 18 Fig. 4. The VLC channel model. paper is different than [13], [27], [40], [45]. The system sum throughput is maximized in [13], [27], and the system average power consumption is minimized in [40], [45]. In these studies, a framework in which a joint resource allocation and link assignment process is employed is not provided, and the optimization problems are in principle mixed integer and non-linear programming problems, which are mathematically intractable. Therefore, the main optimization problems are decomposed into solvable sub-problems, and iterative algorithms are provided. Remark 5: The link selection process can easily be adopted into scenarios where the receiver is mobile. All that one needs is to consider the log-moment generating function in the VLC link for changing transmission rates, i.e., Λv (θ) = loge Ev {eθVl } , and base the channel selection process on the link that increases the maximum average data arrival rate at the transmitter buffer. In Hybrid-Type I Transmission Strategy and Hybrid-Type II Transmission Strategy, because the channel fading gains in the RF and VLC links are instantaneously known at the transmitter in each time frame, the aforementioned analysis will not be different than what we currently have in the paper even if the transmission rates vary due to mobility. We also note that the user mobility is normally low in indoor scenarios. In this regard, we refer to [46]. IV. N UMERICAL R ESULTS In this section, we present numerical results that substantiate our theoretical findings. Unless otherwise specified, we set the transmission time frame to 0.1 milliseconds, i.e., T = 10−4 . We assume that the LED at the VLC access point of the transmitter has a Lambertian radiation pattern, and that the transmitter and receiver planes are parallel to each other. We further assume that the transmitter is directed downwards, while the receiver is directed upwards, as depicted in Fig. 4. However, our theoretical results can easily be adopted into different positional settings. 19 TABLE I S IMULATION PARAMETERS VLC System ◦ LED half intensity viewing angle, φ1/2 {30 , 45◦ , 60◦ } PD field of view (FOV), ψC 90◦ PD physical area, A 1 cm2 Channel bandwidth, Bv 10 MHz PD opt.-to-elect. conversion efficiency, Ω 0.53 A/W PD optical concentrator gain, u(ψ) 2.5 m Vertical distance, dv Noise power spectral density, Nv 1 −21 10 A2 / Hz RF System Channel bandwidth, Br 10 MHz Path loss exponent, q 1.8 Rician factor, K 10 dB Log-normal standard deviation, σ 3.6 dB Ambient Temperature, Tt ◦ 280 K Herein, the line-of-sight channel gain, i.e., g, is given as follows [47]: (s + 1)Au(ψ)ds+1 v g= rect(ψ/ψC ), s+3 2πd1 (40) where A is the surface area of the photo-diode, d1 is the distance between the LED at the transmitter and the photo-diode at the receiver, and dv is the normal distance between the transmitting and receiving planes. Moreover, u(ψ) is the optical concentrator gain at the photodiode, ψ is the angle of incidence, and ψC is the field of view, i.e., the maximum angle at which the light emitted by the LED is detected by the photo-diode. In addition, s = −1 log2 (cos(φ1/2 )) is the Lambertian index, where φ1/2 is the LED half intensity viewing angle, and rect(x) = 1 if \|x\| ≤ 1 and rect(x) = 0 otherwise. Finally, the thermal noise power at the photo-diode is σv2 = Nv Bv , where Nv is the noise power spectral density. Regarding the RF channel, we consider a Rician fading distribution with the Rician factor, K, where the channel realizations, {hl }, are independent and identically distributed, circularly q −L(d0 )/10 K symmetric complex Gaussian random variables with mean µ = e K+1 and variance σh2 = e−L(d0 )/10 . K+1 Setting K to a reasonable value, we can reflect the channel characteristics in millimeter wave range communications as well [2]. Here, L(d0 ) is the large-scale path loss in decibels as a function of the distance between the RF access point at the transmitter and the RF front-end at the receiver, d0 , and it is given by [48] L(d0 ) = L(dref ) + 10q loge d0 dref + Xσ , (41) where L(dref ) = 40 dB is the path loss at a reference distance, dref = 1 m, and an operating frequency of 2.4 GHz. In addition, q is the path loss exponent and Xσ represents the shadowing 20 (a) θ = 0.01 (b) θ = 0.1 Fig. 5. Maximum average arrival rates of VLC and RF links as a function of the average power limit, Pavg , for different values of the average-to-peak power ratio ν and the QoS exponent, θ. Here, d0 = 15 m, d1 = 3 m, φ1/2 = 60◦ , α = 0.3, and β = 0.7. {bpf : bits per frame}. effect, which is assumed to be a zero-mean Gaussian random variable with a standard deviation σ expressed in decibels6 . Finally, the thermal noise power at the RF front-end of the receiver is σr2 = κB Tt Br , where κB is the Boltzmann constant and Tt is the ambient temperature [14]. Table I summarizes the simulation parameters, unless otherwise stated. Finally, setting the average transmission power constraint to Pavg in all the transmission strategies, we define ν as the average-to-peak power ratio in the RF and VLC links, i.e., ν = Pavg Ppeak,r = Pavg . Ppeak,v A. Transmission Strategies We consider the scenario in which a transmitter has a VLC access point and an RF access point, as shown in Fig. 1. The receiver is located at a distance d1 = 3 m from the VLC access point, i.e., the user is located at a horizontal distance dh ≈ 1.6 m from the cell center, where φ1/2 = 60◦ and dv = 2.5 m, and at a distance d0 = 15 m from the RF access point. In Fig. 5, we plot the maximum average data arrival rates at the transmitter buffer, ρv (θ) and ρr (θ), as functions of the average power constraint, Pavg , with different average-to-peak power ratios, i.e, ν ∈ {0.1, 0.3, 0.7, 1}, when the VLC and RF links are employed, respectively. We have the results for θ = 0.01 in Fig. 5(a) and for θ = 0.1 in Fig. 5(b). We observe that the maximum average data arrival rates increase faster with the increasing average power constraint in the VLC 6 Empirical values of K, q, and σ 2 in different indoor scenarios were provided in [48]–[50]. For instance, the value of q ranges from 1.2 to 1.7, while σ varies between 3.6 dB and 4.0 dB inside buildings [48]. 21 (a) Pavg = 24 dBm (b) Pavg = 30 dBm Fig. 6. Maximum average arrival rates of VLC and RF links as a function of the QoS exponent, θ, for different values of the average power limit Pavg and the source statistics, α and β. Here, d0 = 15 m, d1 = 3 m, α = 0.3, φ1/2 = 60◦ , and ν = 0.7. {bpf : bits per frame}. link than they do in the RF link. For instance, when ν = 0.3 and θ = 0.01, the maximum average data arrival rate increases more than 145 bits per frame with Pavg increasing from 27 dBm to 28 dBm in the VLC link, whereas it increases 2 bits per frame in the RF link. We observe the same behavior when the peak power constraint increases. We can explain this result with the stochastic nature of the transmission rates in the RF channel. Particularly, when the instantaneous transmission rate in the RF channel becomes very low, more data packets are accumulated in the transmitter buffer. Therefore, in order to sustain the QoS constraints, the transmitter buffer should accept data at lower arrival rates. On the other hand, because the transmission rate in the VLC channel is constant, the maximum average data arrival rate increases almost linearly with the increasing transmission rate in the VLC channel. Moreover, the performance in the RF link is better than the VLC link when the average power constraint is lower, and the performance in the VLC link is better than the RF link when the average power constraint is higher. In Fig. 6, we plot ρr (θ) and ρv (θ) as functions of the QoS exponent, θ, for different values of α and β when ν = 0.7. We have Pavg = 24 dBm in Fig. 6(a) and Pavg = 30 dBm in Fig. 6(b). The performance in the RF link is better than the VLC link when θ is low, whereas the performance in the RF link decreases faster with increasing θ and becomes less than the performance in the VLC link. In other words, the stochastic nature of the RF channel prevents the RF link from supporting data arrival rates at the transmitter buffer when the QoS constraints are stringent. Indeed, the maximum average data arrival rate that the RF link supports approaches zero exponentially with increasing θ regardless of the average power constraint and the source statistics, i.e., α and β. However, the RF link can support higher data arrival rates if the QoS 22 (a) FDMA (i.e., power and bandwidth) (b) TDMA (i.e., time) Fig. 7. Per-user maximum average arrival rates of VLC and RF links as a function of the number of served receivers (or equivalently the receiver allocated resources) and for different values of the LED viewing angle, φ1/2 . Here, α = 0.3, β = 0.7, ν = 1, and Pavg = 24 dBm. {bpf : bits per frame}. constraints are looser. We further observe that increasing β (or decreasing α), results in better performance values in the VLC and RF links because the steady-state probability of the ON state, i.e., PON = β , α+β and the average arrival rate at the transmitter buffer increase. However, the effect of the source statistics on the performance values is much less than that of the QoS constraints, especially in the RF link. In other words, the randomness in the service process has a higher impact on the system performance than the randomness in the arrival process has. In typical indoor scenarios, VLC and RF access points serve multiple receivers. This is applied by sharing available resources (i.e., power, time, and bandwidth) among the served receivers. Herein, we assume that the transmitter employs the commonly known FDMA and TDMA schemes in both links. In Fig. 7, we plot the maximum average data arrival rate per user given that all the users are uniformly positioned within the coverage area of the VLC access point. We observe that the performance per user in the RF link is generally much higher than the performance per user in the VLC link when the number of receivers is above 4. Basically, the system can serve more users in the RF link than the VLC link when the QoS constraints are of interest. In addition, the results in Fig. 7 agree with the results in Fig. 5, where the performance of the VLC link is highly affected by the decreasing average power constraint. We finally see that with the decreasing LED viewing angle, the performance in the VLC link becomes better because the transmission power is concentrated in smaller areas. Notice also that the VLC channel gain is affected by φ1/2 through the Lambertian index. Herein, we show the performance sensitivity of the RF and VLC links to the allocated transmission resources such as power, time, and bandwidth, given that the available resources are equally shared among the 23 users. We also show that our framework can easily be invoked in a multi-user scenario. We further explore the system performance with respect to the receiver location. We set (xv , yv , zv ) = (0, 0, 0) as the Cartesian coordinates of the VLC access point, (xr , yr , zr ) = (10, 0, 0) as the coordinates of the RF access point, and (xu , yu , zu ) as the coordinates of the receiver, where zu = −dv . Particularly, we consider the following strategies: 1) RF-only Strategy: The transmitter sends data over the RF link only, and the maximum average arrival rate, ρr (θ), is expressed in (12). 2) VLC-only Strategy: The transmitter sends data over the VLC link only, and the maximum average arrival rate, ρv (θ), is expressed in (13). 3) Hybrid-type I Transmission Strategy: The transmitter sends data over the link that provides the highest transmission rate, and the maximum average arrival rate, ρrv (θ), is expressed in (20). 4) Hybrid-type II Transmission Strategy: The transmitter sends data over both links simultaneously following a power allocation policy to maximize the transmission rate, and the maximum rate, ρsrv (θ), is expressed in (25). In Fig. 8, we plot the maximum average arrival rate as a function of the xu and yu coordinates of the receiver location for different QoS constraints, where xu ∈ [−5, 5] and yu ∈ [−5, 5]. As seen in Fig. 8(a) and Fig. 8(e), the position of the receiver does not impact the performance levels in the RF link necessarily, i.e., the performance level stays almost constant when the receiver stays in the defined range. However, the performance levels in the other strategies are affected by the position of the receiver, and the maximum average data arrival rate increases as the receiver gets closer to the point (xv , yv , zv ) = (0, 0, −dv ) because the constant transmission rate from the VLC access point to the receiver increases and the stochastic nature of the RF link is mitigated more with the increasing rate in the VLC link. As seen in Fig. 8(b) and Fig. 8(f), the maximum average data arrival rate goes to zero in VLC-only Strategy as the receiver goes out of the coverage area of the VLC access point. Similarly, as seen in Fig. 8(c), Fig. 8(d), Fig. 8(g) and Fig. 8(h), the maximum average data arrival rate becomes equal to the one in RF-only Strategy as the receiver goes out of the the coverage area of the VLC access point. Furthermore, Hybrid-type I Transmission Strategy provides higher performance levels than RF-only Strategy and VLConly Strategy do because the transmitter, employing Hybrid-type I Transmission Strategy, sends the data over the RF link when the instantaneous transmission rate in the RF link is higher than the rate in the VLC link, and mitigates the lower transmission rates in the RF link by 24 (a) RF-only,, θ = 0.01 (b) VLC-only, θ = 0.01 (c) Hybrid-type I, θ = 0.01 (d) Hybrid-type II, θ = 0.01 (e) RF-only,, θ = 0.1 (f) VLC-only, θ = 0.1 (g) Hybrid-type I, θ = 0.1 (h) Hybrid-type II, θ = 0.1 Fig. 8. Maximum average arrival rates for different selection strategies as a function of the receiver position in terms of xu and yu and for different values of θ. Here, Pavg = 24 dBm, ν = 0.7, α = 0.3, β = 0.7 and φ1/2 = 60◦ . {bpf : bits per frame}. 25 (a) (xu , yu , zu ) = (0, 0, −dv ) i.e., cell center (b) (xu , yu , zu ) = (dc , 0, −dv ) i.e., cell edge Fig. 9. Maximum average arrival rate as a function of n and for different values of the QoS exponent θ and user position in terms of xu . Here, φ1/2 = 60◦ , α = 0.3, β = 0.7, Pavg = 24 dBm, and ν = 1. sending the data in the VLC link. Finally, Hybrid-type II Transmission Strategy outperforms all the other strategies. However, the performance gap between Hybrid-type I Transmission Strategy and Hybrid-type II Transmission Strategy is not necessarily large. Hence, it is more advantageous to employ Hybrid-type I Transmission Strategy in order to avoid the hardware complexity that follows the addition of multihoming capability in Hybrid-type II Transmission Strategy. In Fig. 9, assuming that the handover process causes a transmission delay, where the handover process takes TH = n1 T seconds for n ∈ N and n > 1, we plot the maximum average data arrival rate in Hybrid-type I Transmission Strategy, ρrv (θ), as a function of n considering different user locations. Noting that smaller n means a longer handover period, we observe that the transmission performance is highly affected by the handover process. With increasing n, the performance levels approach the values that are obtained when there is no handover delay. Moreover, the maximum average data arrival rates are higher in Fig. 9(a) than in Fig. 9(b), because the constant transmission rate in the VLC link is higher when the user is at the center. Subsequently in Fig. 10, we plot the maximum average data arrival rates as functions of the vertical distance between the VLC access point and the receiver. We set the position of the receiver to (xu , yu , zu ) = (0.8, 0, −dv ), i.e., the horizontal distance to the VLC cell center is dh = 0.8 m. When dv ≤ 0.6 m, the performance level in the VLC link is zero because the cell area is very small and does not cover the point where the user stands, i.e., dc = dv tan(θ1/2 ) < dh = 0.8 m, and dc is the cell radius. The performance levels in all strategies except RF-only Strategy increase up to a value, and then decrease with increasing dv . This is because the increase in the LED viewing angle is relatively less than the increase in dv at the beginning. Therefore, 26 (a) Pavg = 24 dBm (b) Pavg = 30 dBm Fig. 10. Maximum average arrival rate of different transmission strategies as a function of the vertical distance and for different average power limit Pavg . Here, φ1/2 = 60◦ , (xu , yu , zu ) = (0.8, 0, −dv ), α = 0.3, β = 0.7, θ = 0.1, and ν = 1. {bpf : bits per frame}. the user is effectively getting closer to the cell center and having more rate in the VLC link. In other words, the gain achieved by getting closer to the cell center is higher than the expected degradation due to the increasing cell radius. However, with dv increasing beyond a certain value, the user gets far away from the VLC access point, and hence, the radiated power spreads over more area, which eventually leads to decreased transmission rates in the VLC link. Therefore, the gain in the VLC link vanishes as the distance to the VLC access point becomes larger. B. Non-asymptotic Delay Bounds In the aforementioned results, we analyze the system performance in the steady-state. In the following, we provide results regarding the non-asymptotic bounds, i.e., the bounds on the buffering delay experienced by the data in the transmitter buffer. Particularly, we plot the delay bound as a function of the state transition probability from the OFF state to the ON state in the data arrival process, β, for different α values, where α is the state transition probability from the ON state to the OFF state in the data arrival process. We set the average data arrival rate, λpON , to a value very close to the average data service (transmission) rate in the transmission channel. We note that the average data service rate in the transmission channel depends on the chosen transmission strategy. The delay bound is the highest in RF-only Strategy as seen in Fig. 11(a), whereas it is the lowest in VLC-only Strategy as seen in Fig. 11(b). However, the arrival rate that RF-only Strategy supports is higher than the rate that VLC-only Strategy supports. More interestingly, the hybrid strategies can support higher arrival rates with less delay bounds, and Hybrid-type II Transmission Strategy outperforms all the others. Herein, the system takes 27 (a) RF-only Transmission, λ = 21 kbpf (b) VLC-only Transmission, λ = 7 kbpf (c) Hybrid-Type I Transmission, λ = 24 kbpf (d) Hybrid-Type II Transmission, λ = 26 kbpf Fig. 11. Delay Bounds for different transmission strategies as a function of the transition probability β and for different values of α. Here, d0 = 10 m, d1 = 3 m, Pavg = 30 dBm, ν = 0.7, and φ1/2 = 60◦ . {bpf : bits per frame} advantage of the occasional higher rates in the RF links, and mitigates the lower rates in the RF link by the constant transmission rate in the VLC link. Moreover, increasing β and decreasing α cause the delay bound to increase. Finally, we explore the effects of the data arrival rate, λ, on the delay bound performance in Fig. 12. We set α = 0.3 and β = 0.7, and consider different average power constraints, i.e., Pavg = 24, 30, and 35 dBm. The delay bounds increase asymptotically as the average arrival rate approaches the average data service rates in the channels, because when the average data arrival rate is greater than the average data service rate in one channel, the system becomes unstable, and long buffering periods are expected. Moreover, as seen in Fig. 12(b), the delay bounds are the minimum in VLC-only Strategy but the ranges of the average data arrival rates are smaller than the other strategies, and as seen in Fig. 12(a), the delay bounds are the maximum in RF-only Strategy. However, the hybrid strategies again outperform the others i.e., the hybrid strategies take advantage of the VLC link when the rate in the RF link goes down drastically, and utilize the RF link when the channel conditions are better again. While the VLC link provides stability 28 (a) RF-only Transmission (b) VLC-only Transmission (c) Hybrid-Type I transmission (d) Hybrid-Type II transmission Fig. 12. Delay Bounds for different transmission strategies as a function of the arrival rate λ and for different values of Pavg . Here, α = 0.3, β = 0.7, d0 = 10 m, d1 = 3 m, ν = 0.7, and φ1/2 = 60◦ . and decreases the delay bounds, the RF link increases the range of average data arrival rate that can be supported. Finally, in Fig. 12 we see that increasing the average power can potentially improve the system performance in terms of the buffering delay. V. C ONCLUSIONS In this paper, we have analyzed the performance of a hybrid RF/VLC system when statistical QoS constraints are inflicted as limits on the buffer overflow and delay violation probabilities. We have provided a cross-layer analysis regarding physical and data-link layers by employing the maximum average arrival rate at the transmitter buffer and the non-asymptotic delay bounds as the main performance measures. We have proposed and analyzed three strategies in which RF and VLC links are utilized for data transmission. We have further formulated the performance levels achieved by each of the proposed strategies. Through numerical results, we have shown that RF technology can be beneficial when there are lower average power constraints and/or looser QoS requirements. Moreover, we have shown that utilizing the VLC technology for data 29 transmission, either alone or in a hybrid transmission strategy, can potentially enhance the system performance in terms of delay performance. It lowers the buffering delay bounds, when compared to the RF technology. Particularly, when data arrival rates at the transmitter buffer is low, VLC links provide lower queueing delays than RF links do, but RF links support higher data arrival rates at the transmitter buffer. A PPENDIX A. Derivation of ρr (θ) in (12) Recall that λ is the data arrival rate at the buffer when the data source is in the ON state, and that pON = β α+β is the steady-state probability of the ON state. Thus, the average data arrival β λ. Moreover, we have Λa (θr? ) = −Λr (−θr? ) in steady-state, i.e., rate is λpON = α+β q    1 − β + (1 − α)eθr? λ + [1 − β + (1 − α)eθr? λ ]2 − 4(1 − α − β)eθr? λ  ? loge = − loge Eh e−θr Rl .   2 (42) Solving the aforementioned equation for λ with any given θ > 0, we obtain ( ) 1 − (1 − β)Eh e−θRl 1 . λ = loge θ (1 − α)Eh {e−θRl } − (1 − α − β)E2h {e−θRl } As a result, we formulate the maximum average data arrival rate as ρr (θ) = β λ α+β (43) for any θ > 0. As for ρv (θ), we set Λa (θv? ) = −Λv (−θv? ) and follow the same steps. B. Proof of Proposition 1 Based on the link selection process, the VLC link is selected only when ρv (θ) > ρr (θ), where ρv (θ) is the maximum average arrival rate supported by the VLC link given in (13) and ρr (θ) is the maximum average arrival rate supported by the RF link given in (12). Since logarithm is a monotonic increasing function, this condition is satisfied when we have 1 − (1 − β)Eh e−θRl e2θV − (1 − β)eθV > , (1 − α)eθV − (1 − α − β) (1 − α)Eh {e−θRl } − (1 − α − β)E2h {e−θRl } (44) Now, let χ = Eh {e−θRl } and O = eθV . Then, (44) can be expressed by the following quadratic inequality: O2 − (1 − β + (1 − α)ξ)O + ξ(1 − α − β) > 0. (45) 30 where ξ = 1−(1−β)χ . (1−α)χ−(1−α−β)χ2 O1 = and O2 = Solving the above equation results in two solutions: q 1 − β + (1 − α)ξ − [1 − β + (1 − α)ξ]2 − 4(1 − α − β)ξ (46) 2 q 1 − β + (1 − α)ξ + [1 − β + (1 − α)ξ]2 − 4(1 − α − β)ξ (47) 2 where O2 > O1 . V has two ranges 0 < V < loge {O1 } and V ≥ loge {O2 }. Setting O = 1 in (45), we have β(1 − ξ) > 0. Note that ξ > 1 because ρr (θ) = β (α+β)θ loge {ξ} > 0. Hence, we have β(1 − ξ) < 0, which implies that O1 < 1 and loge {O1 } < 0. Therefore, we have only V ≥ loge {O2 } as the solution region, which completes the proof. R EFERENCES [1] M. B. Rahaim, A. M. Vegni, and T. D. Little, “A hybrid radio frequency and broadcast visible light communication system,” in Proc. IEEE Global Telecommun. (GLOBECOM) Workshops, 2011, pp. 792–796. [2] D. A. Basnayaka and H. Haas, “Design and analysis of a hybrid radio frequency and visible light communication system,” IEEE Trans. Commun., 2017. [3] H. Chowdhury and M. Katz, “Cooperative data download on the move in indoor hybrid (radio-optical) WLAN-VLC hotspot coverage,” Trans. on Emerging Telecommun. Technol., vol. 25, no. 6, pp. 666–677, 2014. [4] D. A. Basnayaka and H. Haas, “Hybrid RF and VLC systems: Improving user data rate performance of VLC systems,” in Proc. IEEE Veh. Technol. Conf. Spring (VTC-Spring), 2015, pp. 1–5. [5] X. Bao, X. Zhu, T. Song, and Y. Ou, “Protocol design and capacity analysis in hybrid network of visible light communication and OFDMA systems,” IEEE Trans. Veh. Technol., vol. 63, no. 4, pp. 1770–1778, 2014. [6] S. Shao, A. Khreishah, M. B. Rahaim, H. Elgala, M. Ayyash, T. D. Little, and J. Wu, “An indoor hybrid WiFi-VLC internet access system,” in Proc. IEEE Int. Conf. on Mobile Ad Hoc and Sensor Sys., 2014, pp. 569–574. [7] M. Kashef, M. Ismail, M. Abdallah, K. A. Qaraqe, and E. Serpedin, “Energy efficient resource allocation for mixed RF/VLC heterogeneous wireless networks,” IEEE J. Sel. Areas Commun, vol. 34, no. 4, pp. 883–893, 2016. [8] A. M. Vegni and T. D. Little, “Handover in VLC systems with cooperating mobile devices,” in Proc. Int. Conf. Computing, Netw. Commun. (ICNC), 2012, pp. 126–130. [9] F. Wang, Z. Wang, C. Qian, L. Dai, and Z. Yang, “Efficient vertical handover scheme for heterogeneous VLC-RF systems,” J. Opt. Commun. and Netw., vol. 7, no. 12, pp. 1172–1180, 2015. [10] S. Liang, H. Tian, B. Fan, and R. Bai, “A novel vertical handover algorithm in a hybrid visible light communication and LTE system,” in Proc. IEEE Veh. Technol. Conf. Fall (VTC-FALL), 2015, pp. 1–5. [11] S. Shao and A. Khreishah, “Delay analysis of unsaturated heterogeneous omnidirectional-directional small cell wireless networks: The case of RF-VLC coexistence,” IEEE Trans. on Wireless Commun., vol. 15, no. 12, pp. 8406–8421, 2016. [12] X. Li, R. Zhang, and L. Hanzo, “Cooperative load balancing in hybrid visible light communications and WiFi,” IEEE Trans. Commun., vol. 63, no. 4, pp. 1319–1329, 2015. [13] Y. Wang and H. Haas, “Dynamic load balancing with handover in hybrid Li-Fi and Wi-Fi networks,” IEEE J. of Lightwave Tech., vol. 33, no. 22, pp. 4671–4682, 2015. 31 [14] I. Stefan, H. Burchardt, and H. Haas, “Area spectral efficiency performance comparison between VLC and RF femtocell networks,” in Proc. IEEE Int. Conf. Commun. (ICC), 2013, pp. 3825–3829. [15] Y. Wang, D. A. Basnayaka, X. Wu, and H. Haas, “Optimization of load balancing in hybrid LiFi/RF networks,” IEEE Trans. Commun., vol. 65, no. 4, pp. 1708–1720, 2017. [16] H. Kazemi, M. Uysal, and F. Touati, “Outage analysis of hybrid FSO/RF systems based on finite-state markov chain modeling,” in Proc. IEEE Int. Workshop Opt. Wireless Commun. (IWOW), 2014, pp. 11–15. [17] N. D. Chatzidiamantis, G. K. Karagiannidis, E. E. Kriezis, and M. Matthaiou, “Diversity combining in hybrid RF/FSO systems with PSK modulation,” in Proc. IEEE Int. Conf. Commun. (ICC), 2011, pp. 1–6. [18] “Cisco visual networking index: Global mobile data traffic forecast update 20142019 white paper,” http://www.cisco.com/ c/en/us/solutions/collateral/service-provider/visual-networking-index-vni/white paper c11-520862.html. [19] C.-S. Chang, “Stability, queue length, and delay of deterministic and stochastic queueing networks,” IEEE Trans. Autom. Control, vol. 39, no. 5, pp. 913–931, 1994. [20] D. Wu and R. Negi, “Effective capacity: a wireless link model for support of quality of service,” IEEE Trans. Wireless Commun., vol. 2, no. 4, pp. 630–643, 2003. [21] G. L. Choudhury, D. M. Lucantoni, and W. Whitt, “Squeezing the most out of ATM,” IEEE Trans. Commun., vol. 44, no. 2, pp. 203–217, 1996. [22] F. Kelly, “Notes on e ective bandwidths,” Stochastic networks: theory and applications, pp. 141–168, 1996. [23] S. Akin and M. C. Gursoy, “Effective capacity analysis of cognitive radio channels for quality of service provisioning,” IEEE Trans. Wireless Commun., vol. 9, no. 11, pp. 3354–3364, 2010. [24] M. Hammouda, S. Akin, and J. Peissig, “Effective capacity in cognitive radio broadcast channels,” in Proc. IEEE Global Telecommun. Conf. (GLOBECOM), 2014, pp. 1071–1077. [25] M. Ozmen and M. C. Gursoy, “Wireless throughput and energy efficiency with random arrivals and statistical queuing constraints,” IEEE Trans. Inf. Theory, vol. 62, no. 3, pp. 1375–1395, 2016. [26] F. Jin, X. Li, R. Zhang, C. Dong, and L. Hanzo, “Resource allocation under delay-guarantee constraints for visible-light communication,” IEEE Access, vol. 4, pp. 7301–7312, 2016. [27] F. Jin, R. Zhang, and L. Hanzo, “Resource allocation under delay-guarantee constraints for heterogeneous visible-light and rf femtocell,” IEEE Transactions on Wireless Communications, vol. 14, no. 2, pp. 1020–1034, 2015. [28] N. Letzepis and A. G. I. Fabregas, “Outage probability of the Gaussian MIMO free-space optical channel with PPM,” IEEE Trans. on Commun., vol. 57, no. 12, 2009. [29] L. H. Ozarow, S. Shamai, and A. D. Wyner, “Information theoretic considerations for cellular mobile radio,” IEEE trans. Veh. Technol., vol. 43, no. 2, pp. 359–378, 1994. [30] S. Shamai and I. Bar-David, “The capacity of average and peak-power-limited quadrature Gaussian channels,” IEEE Trans. Inf. Theory, vol. 41, no. 4, pp. 1060–1071, 1995. [31] T. Komine and M. Nakagawa, “Fundamental analysis for visible-light communication system using LED lights,” IEEE Trans. on Consumer Electronics, vol. 50, no. 1, pp. 100–107, 2004. [32] J. M. Kahn and J. R. Barry, “Wireless infrared communications,” Proceedings of the IEEE, vol. 85, no. 2, pp. 265–298, 1997. [33] V. Pohl, V. Jungnickel, and C. Von Helmolt, “A channel model for wireless infrared communication.” in PIMRC, 2000, pp. 297–303. [34] V. Jungnickel et al., “A European view on the next generation optical wireless communication standard,” in Proc. IEEE Conf. Standards for Commun. and Networking (CSCN), Oct 2015, pp. 106–111. 32 [35] S. Dimitrov and H. Haas, Principles of LED Light Communications: Towards Networked Li-Fi. Cambridge University Press, 2015. [36] A. Lapidoth, S. M. Moser, and M. A. Wigger, “On the capacity of free-space optical intensity channels,” IEEE Trans. Inf. Theory, vol. 55, no. 10, pp. 4449–4461, 2009. [37] H. Heffes and D. Lucantoni, “A Markov modulated characterization of packetized voice and data traffic and related statistical multiplexer performance,” IEEE J. Sel. Areas Commun., vol. 4, no. 6, pp. 856–868, 1986. [38] C.-S. Chang and T. Zajic, “Effective bandwidths of departure processes from queues with time varying capacities,” in Proc. IEEE INFOCOM, vol. 3, 1995, pp. 1001–1009. [39] C.-S. Chang, Performance Guarantees in Communication Networks. Springer Science & Business Media, 2000. [40] H. Yu, M. H. Cheung, L. Huang, and J. Huang, “Power-delay tradeoff with predictive scheduling in integrated cellular and Wi-Fi networks,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 735–742, 2016. [41] C.-S. Chang, Performance guarantees in communication networks. Springer Science & Business Media, 2012. [42] M. Fidler and A. Rizk, “A guide to the stochastic network calculus,” IEEE Commun. Surveys & Tutorials, vol. 17, no. 1, pp. 92–105, 2015. [43] F. Ciucu, A. Burchard, and J. Liebeherr, “Scaling properties of statistical end-to-end bounds in the network calculus,” IEEE Trans. Inf. Theory, vol. 52, no. 6, pp. 2300–2312, 2006. [44] M. Fidler, R. Lübben, and N. Becker, “Capacity–delay–error boundaries: A composable model of sources and systems,” IEEE Trans. Wireless Commun., vol. 14, no. 3, pp. 1280–1294, 2015. [45] W. Wu, F. Zhou, and Q. Yang, “Dynamic network resource optimization in hybrid VLC and radio frequency networks,” in Selected Topics in Mobile and Wireless Networking (MoWNeT), 2017 International Conference on, 2017, pp. 1–7. [46] M. Dehghani Soltani, X. Wu, M. Safari, and H. Haas, “On limited feedback resource allocation for visible light communication networks,” in Proceedings of the 2nd International Workshop on Visible Light Communications Systems. ACM, 2015, pp. 27–32. [47] J. R. Barry, J. M. Kahn, W. J. Krause, E. Lee, D. G. Messerschmitt et al., “Simulation of multipath impulse response for indoor wireless optical channels,” IEEE J. Sel. Areas Commun., vol. 11, no. 3, pp. 367–379, 1993. [48] R. G. Akl, D. Tummala, and X. Li, Indoor propagation modeling at 2.4 GHz for IEEE 802.11 networks. International Association of Science and Technology for Development, 2006. [49] T. S. Rappaport and C. D. McGillem, “UHF fading in factories,” IEEE J. Sel. Areas Commun., vol. 7, no. 1, pp. 40–48, 1989. [50] A. Neskovic, N. Neskovic, and G. Paunovic, “Modern approaches in modeling of mobile radio systems propagation environment.” IEEE Communications Surveys and Tutorials, vol. 3, no. 3, pp. 2–12, 2000.	7cs.IT
Developing Bug-Free Machine Learning Systems With Formal Mathematics Daniel Selsam 1 Percy Liang 1 David L. Dill 1 arXiv:1706.08605v1 [cs.SE] 26 Jun 2017 Abstract Noisy data, non-convex objectives, model misspecification, and numerical instability can all cause undesired behaviors in machine learning systems. As a result, detecting actual implementation errors can be extremely difficult. We demonstrate a methodology in which developers use an interactive proof assistant to both implement their system and to state a formal theorem defining what it means for their system to be correct. The process of proving this theorem interactively in the proof assistant exposes all implementation errors since any error in the program would cause the proof to fail. As a case study, we implement a new system, Certigrad, for optimizing over stochastic computation graphs, and we generate a formal (i.e. machine-checkable) proof that the gradients sampled by the system are unbiased estimates of the true mathematical gradients. We train a variational autoencoder using Certigrad and find the performance comparable to training the same model in TensorFlow. 1. Introduction Machine learning systems are difficult to engineer for many fundamental reasons. First and foremost, implementation errors can be extremely difficult to detect—let alone to localize and address—since there are many other potential causes of undesired behavior in a machine learning system. For example, an implementation error may lead to incorrect gradients and so cause a learning algorithm to stall, but such a symptom may also be caused by noise in the training data, a poor choice of model, an unfavorable optimization landscape, an inadequate search strategy, or numerical instability. These other issues are so common that it is often assumed that any undesired behavior is caused by one of them. As a result, actual implementation errors can persist 1 Stanford University, Stanford, CA. Correspondence to: Daniel Selsam <dselsam@stanford.edu>. Proceedings of the 34 th International Conference on Machine Learning, Sydney, Australia, PMLR 70, 2017. Copyright 2017 by the author(s). Standard methodology: test it empirically Debug Program Test Our methodology: verify it mathematically Debug Specify Program Prove Figure 1. A high-level comparison of our methodology with the standard methodology for developing machine learning systems. Instead of relying on empirical testing to expose implementation errors, we first formally specify what our system is required to do in terms of the underlying mathematics, and then try to formally prove that our system satisfies its specification. The process of proving exposes implementation errors systematically and the (program → prove → debug) loop eventually terminates with a bug-free system and a machine-checkable proof of correctness. indefinitely without detection.1 Errors are even more difficult to detect in stochastic programs, since some errors may only distort the distributions of random variables and may require writing custom statistical tests to detect. Machine learning systems are also difficult to engineer because it can require substantial expertise in mathematics (e.g. linear algebra, statistics, multivariate analysis, measure theory, differential geometry, topology) to even understand what a machine learning algorithm is supposed to do and why it is thought to do it correctly. Even simple algorithms such as gradient descent can have intricate justifications, and there can be a large gap between the mechanics of an implementation—especially a highly-optimized one—and its intended mathematical semantics. 1 Theano (Bergstra et al., 2010) has been under development for almost a decade and yet there is a recent GitHub issue (https://github.com/Theano/Theano/issues/4770) reporting a model for which the loss continually diverges in the middle of training. Only after various experiments and comparing the behavior to other systems did the team agree that it is most likely an implementation error. As of this writing, neither the cause of this error nor the set of models it affects have been determined. Developing Bug-Free Machine Learning Systems With Formal Mathematics In this paper, we demonstrate a practical methodology for building machine learning systems that addresses these challenges by enabling developers to find and eliminate implementation errors systematically without recourse to empirical testing. Our approach makes use of a tool called an interactive proof assistant (Gordon, 1979; Gordon & Melham, 1993; Harrison, 1996; Nipkow et al., 2002; Owre et al., 1992; Coq Development Team, 2015-2016; de Moura et al., 2015), which consists of (a) a programming language, (b) a language to state mathematical theorems, and (c) a set of tools for constructing formal proofs of such theorems. Note: we use the term formal proof to mean a proof that is in a formal system and so can be checked by a machine. In our approach, developers use the theorem language (b) to state a formal mathematical theorem that defines what it means for their implementation to be error-free in terms of the underlying mathematics (e.g. multivariate analysis). Upon implementing the system using the programming language (a), developers use the proof tools (c) to construct a formal proof of the theorem stating that their implementation is correct. The first draft of any implementation will often have errors, and the process of interactive proving will expose these errors systematically by yielding impossible proof obligations. Once all implementation errors have been fixed, the developers will be able to complete the formal proof and be certain that the implementation has no errors with respect to its specification. Moreover, the proof assistant can check the formal proof automatically so no human needs to understand why the proof is correct in order to trust that it is. Figure 1 illustrates this process. Proving correctness of machine learning systems requires building on the tools and insights from two distinct fields: program verification (Leroy, 2009; Klein et al., 2009; Chlipala, 2013; Chen et al., 2015), which has aimed to prove properties of computer programs, and formal mathematics (Rudnicki, 1992; Gonthier, 2008; Gonthier et al., 2013; Hales et al., 2015), which has aimed to formally represent and generate machine-checkable proofs of mathematical theorems. Both of these fields make use of interactive proof assistants, but the tools, libraries and design patterns developed by the two fields focus on different problems and have remained largely incompatible. While the methodology we have outlined will be familiar to the program verification community, and while reasoning formally about the mathematics that underlies machine learning will be familiar to the formal mathematics community, proving such sophisticated mathematical properties of large (stochastic) software systems is a new goal and poses many new challenges. To explore these challenges and to demonstrate the prac- W1 ∗ x W2 softplus N (·, I) ∗ sigmoid cost L(W1 , W2 ) = Ez∼N (softplus(W1 x),I) [cost(x, σ(W2 z))] Figure 2. An example stochastic computation graph representing a simple variational autoencoder. Stochastic nodes are indicated by rounded rectangles. The loss function for the graph is the expected value of the cost node over the stochastic choices, which in this case is a single sample from a Gaussian distribution. ticality of our approach, we implemented a new machine learning system, Certigrad, for optimizing over stochastic computation graphs (Schulman et al., 2015). Stochastic computation graphs extend the computation graphs that underly systems like TensorFlow (Abadi et al., 2015) and Theano (Bergstra et al., 2010) by allowing nodes to represent random variables and by defining the loss function for a graph to be the expected value of the sum of the leaf nodes over the stochastic choices. See Figure 2 for an example of a stochastic computation graph. We implement our system in the Lean Theorem Prover (de Moura et al., 2015), a new interactive proof assistant still under active development for which the integration of programming and mathematical reasoning is an ongoing design goal. We formally state and prove functional correctness for the stochastic backpropagation algorithm: that the sampled gradients are indeed unbiased estimates of the gradients of the loss function with respect to the parameters. We note that provable correctness need not come at the expense of computational efficiency: proofs need only be checked once during development and they introduce no runtime overhead. Although the algorithms we verify in this work lack many optimizations, most of the running time when training machine learning systems is spent multiplying matrices, and we are able to achieve competitive performance simply by linking with an optimized library for matrix operations (we used Eigen (Guennebaud et al., 2010)).2 To demonstrate practical feasibility empirically, we trained an Auto-Encoding Variational Bayes (AEVB) model (Kingma & Welling, 2014) on MNIST using ADAM (Kingma & Ba, 2014) and found the performance comparable to training the same model in TensorFlow. 2 Note that the validity of our theorem becomes contingent on Eigen’s matrix operations being functionally equivalent to the versions we formally proved correct. Developing Bug-Free Machine Learning Systems With Formal Mathematics Informal R R f (x, y) dx f (x, y) ∇θ f (g(θ))\|θ0 ∇f (g(θ)) We summarize our contributions: 1. We present the first application of formal (i.e. machine-checkable) proof techniques to developing machine learning systems. 2. We describe a methodology that can detect implementation errors systematically in machine learning systems. 3. We demonstrate that our approach is practical by developing a performant implementation of a sophisticated machine learning system along with a machinecheckable proof of correctness. 2. Motivation When developing machine learning systems, many program optimizations involve extensive algebraic derivations to put mathematical expressions in closed-form . For example, suppose you want to compute the following quantity efficiently: Z 2 N (x; µ, Diag(σ )) log N (x; 0, In×n ). (1) x You expand the density functions, grind through the algebra by hand and eventually derive the following closed form expression: " n # 1 X 2 2 σ − µi + n log 2π − 2 i=1 i Formal R (λ x, f x y) ? ∇ (λ θ, f (g θ)) θ0 ? Figure 3. Translating informal usages of the integral and gradient to our formal representation. Note that whereas some of the informal examples are too ambiguous to interpret without additional information, the Lean representation is always unambiguous. language. All three capabilities are provided by the new interactive proof assistant Lean (de Moura et al., 2015). Lean is an implementation of a logical system known as the Calculus of Inductive Constructions (Coquand & Huet, 1988), and its design is inspired by the better-known Coq Proof Assistant (Coq Development Team, 2015-2016). Our development makes use of certain features that are unique to Lean, but most of what we present is equally applicable to Coq, and to a lesser extent, other interactive theorem provers such as Isabelle/HOL (Nipkow et al., 2002). To explain and motivate the relevant features of Lean, we will walk through applying our methodology to a toy problem: writing a program to compute the gradient of the softplus function. We can write standard functional programs in Lean, such as softplus: def splus (x : R) : R := log (1 + exp x) We can also represent more abstract operations such as integrals and gradients: (2) You implement a procedure to compute this quantity and include it as part of a larger program, but when you run your first experiment, your plots are not as encouraging as you hoped. After ruling out many other possible explanations, you eventually decide to scrutinize this procedure more closely. You implement a naı̈ve Monte Carlo estimator for the quantity above, compare it against your procedure on a few random inputs and find that its estimates are systematically biased. What do you do now? If you re-check your algebra carefully, you might notice that the sign of µ2i is wrong, but wouldn’t it be easier if the compiler checked your algebra for you and found the erroneous step? Or better yet, if it did the algebra for you in the first place and could guarantee the result was error-free? 3. Background: The Lean Theorem Prover To develop software systems with no implementation errors, we need a way to write computer programs, mathematical theorems, and mathematical proofs all in the same R (f : R → R) : R ∇ (f : R → R) (θ : R) : R R Here the intended meaning of f is the integral of the function f over all of R, while the intended meaning of ∇ f θ is the gradient (i.e. the derivative) of the function f at the point θ. Figure 3 shows how to represent common idioms of informal mathematics in our formal representation; note that whereas some of the informal examples are too ambiguous to interpret without additional information, the Lean representation is always unambiguous. We can represent mathematical theorems in Lean as well. For example, we can use the following predicate to state that a particular function f is differentiable at a point θ: is_diff (f : R → R) (θ : R) : Prop The fact that the return type of is_diff is Prop indicates that it is not a computer program to be executed but rather that it represents a mathematical theorem. We can also state and assume basic properties about the gradient, such as linearity: ∀ (f g : R → R) (θ : R), is_diff f θ ∧ is_diff g θ → ∇ (f + g) θ = ∇ f θ + ∇ g θ Developing Bug-Free Machine Learning Systems With Formal Mathematics Returning to our running example, we can state the theorem that a particular function f computes the gradient of the softplus function: def gsplus_spec (f : R → R) : Prop := ∀ x, f x = ∇ splus x Suppose we try to write a program to compute the gradient of the softplus function as follows: why a proof is correct in order to trust that it is.3 Although we cannot execute functions such as gsplus directly in the core logic of Lean (since a real number is an infinite object that cannot be stored in a computer), we can execute the floating-point approximation inside Lean’s virtual machine: vm_eval gsplus π −− answer: 0.958576 def gsplus (x : R) : R := 1 / (1 + exp x) The application gsplus_spec gsplus represents the proposition that our implementation gsplus is correct, i.e. that it indeed computes the gradient of the softplus function for all inputs. We can try to formally prove theorems in Lean interactively: theorem gsplus_correct : gsplus_spec gsplus := lean: ` gsplus_spec gsplus user: expand_def gsplus_spec, lean: ` ∀ x, gsplus x = ∇ splus x user: introduce x, lean: x : R ` gsplus x = ∇ splus x user: expand_defs [gsplus, splus], lean: x : R ` 1 / (1 + exp x) = ∇ (λ x, log (1 + exp x)) x user: simplify_grad, lean: x : R ` 1 / (1 + exp x) = exp x / (1 + exp x) The lines beginning with lean show the current state of the proof as displayed by Lean, which at any time consists of a collection of goals of the form assumptions ` conclusion. Every line beginning with user invokes a tactic, which is a command that modifies the proof state in some way such that Lean can automatically construct proofs of the original goals given proofs of the new ones. Here the simplify_grad tactic rewrites exhaustively with known gradient rules—in this case it uses the rules for log, exp, addition, constants, and the identity function. The final goal is clearly not provable, which means we have found an implementation error in gsplus. Luckily the goal tells us exactly what gsplus x needs to return: gsplus x = exp x / (1 + exp x). Once we fix the implementation of gsplus, the proof script that failed before now succeeds and generates a machine-checkable proof that the revised gsplus is bug-free. Note that we need not have even attempted to implement gsplus before starting the proof, since the process itself revealed what the program needs to compute. We will revisit this phenomenon in §4.5. In the process of proving the theorem, Lean constructs a formal proof certificate that can be automatically verified by a small stand-alone executable, whose soundness is based on a well-established meta-theoretic argument embedding the core logic of Lean into set theory, and whose implementation has been heavily scrutinized by many developers. Thus no human needs to be able to understand 4. Case Study: Certified Stochastic Computation Graphs Stochastic computation graphs are directed acyclic graphs in which each node represents a specific computational operation that may be deterministic or stochastic (Schulman et al., 2015). The loss function for a graph is defined to be the expected value of the sum of the leaf nodes over the stochastic choices. Figure 2 shows the stochastic computation graph for a simple variational autoencoder. Using our methodology, we developed a system, Certigrad, which allows users to construct arbitrary stochastic computation graphs out of the primitives that we provide. The main purpose of the system is to take a program describing a stochastic computation graph and to run a randomized algorithm (stochastic backpropagation) that, in expectation, provably generates unbiased samples of the gradients of the loss function with respect to the parameters. 4.1. Overview of Certigrad We now briefly describe the components of Certigrad, some of which have no analogues in traditional software systems.4 Mathematics libraries. There is a type that represents tensors of a particular shape, along with basic functions (e.g. exp, log) and operations (e.g. the gradient, the integral). There are assumptions about tensors (e.g. gradient rules for exp and log), and facts that are proved in terms of those assumptions (e.g. the gradient rule for softplus). There is also a type that represents probability distributions over vectors of tensors, that can be reasoned about mathematically and that can also be executed procedurally using a pseudo-random number generator. Implementation. There is a data structure that represents stochastic computation graphs, as well as an implementation of stochastic backpropagation. There are also functions that optimize stochastic computation graphs in various ways (e.g. by integrating out parts of the objective 3 This appealing property can be lost when an axiom is assumed that is not true. We discuss this issue further in §4.3. 4 The complete development can be found at www.github.com/ dselsam/certigrad. Developing Bug-Free Machine Learning Systems With Formal Mathematics function), as well as basic utilities for training models (e.g. stochastic gradient descent). Specification. There is a collection of theorem statements that collectively define what it means for the implementation to be correct. For Certigrad, there is one main theorem that states that the stochastic backpropagation procedure yields unbiased estimates of the true mathematical gradients. There are also other theorems that state that individual graph optimizations are sound. Proof. There are many helper lemmas to decompose the proofs into more manageable chunks, and there are tactic scripts to generate machine-checkable proofs for each of the lemmas and theorems appearing in the system. There are also tactic programs to automate certain types of reasoning, such as computing gradients or proving that functions are continuous. Optimized libraries. While the stochastic backpropagation function is written in Lean and proved correct, we execute the primitive tensor operations with the Eigen library for linear algebra. There is a small amount of C++ code to wrap Eigen operations for use inside Lean’s virtual machine. The rest of this section describes the steps we took to develop Certigrad, which include sketching the high-level architecture, designing the mathematics libraries, stating the main correctness theorem and constructing the formal proof. Though many details are specific to Certigrad, this case study is designed to illustrate our methodology and we expect other projects will follow a similar process. Note: Certigrad supports arbitrarily-shaped tensors, but doing so introduces more notational complexity than conceptual difficulty and so we simplify the presentation that follows by assuming that all values are scalars. 4.2. Informal specification The first step of applying our methodology is to write down informally what the system is required to do. Suppose g is a stochastic computation graph with n nodes and (to simplify the notation) that it only takes a single parameter θ. Then g, θ together define a distribution over the values at the n nodes (X1 , . . . , Xn ). Let cost(g, X1:n ) be the function that sums the values of the leaf nodes. Our primary goal is to write a (stochastic) backpropagation algorithm bprop such that for any graph g, Eg,θ [bprop(g, θ, X1:n )] = ∇θ (Eg,θ [cost(g, X1:n )]) (3) While this equation may seem sufficient to communicate the specification to a human with a mathematical background, more precision is needed to communicate it to a computer. The next step is to formalize the background mathematics, such as real numbers (tensors) and probability distributions, so that we can state a formal analogue of Equation 3 that the computer can understand. Although we believe it will be possible to develop standard libraries of mathematics that future developers can use off-the-shelf, we needed to develop the mathematics libraries for Certigrad from scratch. 4.3. Designing the mathematics libraries Whereas in traditional formal mathematics the goal is to construct mathematics from first principles (Gonthier et al., 2013; Hales et al., 2015), we need not concern ourselves with foundational issues and can simply assume that standard mathematical properties hold. For example, we can assume that there is a type R of real numbers without needing to construct them (e.g. from Cauchy sequences), and likewise can assume there is an integration operator on the R reals (f : R → R) : R that satisfies the well-known properties without needing to construct it either (e.g. from Riemann sums). Note that axioms must be chosen with great care since even a single false axiom (perhaps caused by a single missing precondition) can in principle allow proving any false theorem and so would invalidate the property that all formal proofs can be trusted without inspection.5 However, there are many preconditions that appear in mathematical theorems, such as integrability, that are almost always satisfied in machine learning contexts and which most developers ignore. Using axioms that omit such preconditions will necessarily lead to proving theorems that are themselves missing the corresponding preconditions, but in practice a non-adversarial developer is extremely unlikely to accidentally construct vacuous proofs by exploiting these axioms. For the first draft of our system, we purposely omitted integrability preconditions in our axioms to simplify the development. Only later did we make our axioms sound and propagate the additional preconditions throughout the system so that we could fully trust our formal proofs. Despite the convenience of axiomatizing the mathematics, designing the libraries was still challenging for two reasons. First, there were many different ways to formally represent the mathematical objects in question, and we needed to experiment to understand the tradeoffs between the different representations. Second, we needed to extend several traditional mathematical concepts to support reasoning about executable computer programs. The rest of this sub5 For example, the seemingly harmless axiom ∀x, x/x = 1 without the precondition x 6= 0 can be used to prove the absurdity (0 = 0 ∗ 1 = 0 ∗ (0/0) = (0 ∗ 0)/0 = 0/0 = 1). If a system assumes this axiom, then a formal proof of correctness could not be trusted without inspection since the proof may exploit this contradiction. Developing Bug-Free Machine Learning Systems With Formal Mathematics section illustrates these challenges by considering the problem we faced of designing a representation of probability distributions for Certigrad. Representing probability distributions. Our challenge is to devise a sufficiently abstract representation of probability distributions that satisfies the following desiderata: we can reason about the probability density functions of continuous random variables, we have a way to reason about arbitrary deterministic functions applied to random variables, we can execute a distribution procedurally using a pseudorandom number generator (RNG), the mathematical and procedural representations of a distribution are guaranteed to correspond, and the mathematics will be recognizable to somebody familiar with the informal math behind stochastic computation graphs. We first define types to represent the mathematical and procedural notions of probability distribution. For mathematics, we define a Func n to be a functional that takes a realvalued function on Rn to a scalar: def Func (n : N) : Type := ∀ (f : Rn → R), R The intended semantics is that if p : Func n represents a distribution on Rn , then p f is the expected value of f over p, i.e. Ex∼p [f (x)]. For sampling, we define an Prog n to be a procedure that takes an RNG and returns a vector in Rn along with an updated RNG: def Prog (n : N) : Type := RNG → Rn × RNG We also assume that there are primitive (continuous) distributions (PrimDist := Func 1 × Prog 1) that consist of a probability density function and a corresponding sampling procedure. In principle, we could construct all distributions from uniform variates, but for expediency, we treat other well-understood distributions as primitive, such as the Gaussian (gauss µ σ : PrimDist). Finally, we define a type of distributions (Dist n) that abstractly represents programs that may mix sampling from primitive distributions with arbitrary deterministic computations. A Dist n can be denoted to a Func n (with the function E) to reason about mathematically, and to an Prog n (with the function run) to execute with an RNG. For readers familiar with functional programming, our construction is similar to a monad. We allow three ways of constructing a Dist n, corresponding to sampling from a primitive distribution (sample), returning a value deterministically (det), and composing two distributions (compose): sample ((pdf, prog) : PrimDist) : Dist 1 det (xs : Rn ) : Dist n compose (d1 : Dist m) (d2 : Rm → Dist n) : Dist n Dist n : Type E {n : N} (d : Dist n) (f : Rn → R) : R SCG n : Type SCG.to_dist {n : N} (g : SCG n) (θ : R) : Dist n cost {n : N} (g : SCG n) (xs : Rn ) : R Figure 4. The basic types and functions we will need to formally state the specification. Dist n represents a distribution over Rn , E is the expected value function, SCG n represents a computation graph on n nodes, SCG.to dist is the function that samples from an SCG n and yields a distribution over the values at the nodes, and cost sums the values at the leaf nodes of a graph. Curly braces around an argument indicates that it can be inferred from context and need not be passed explicitly. The mathematical semantics of all three constructors are straightforward: R E (sample (pdf, prog)) f = (λ x, pdf x * f x) E (det xs) f = f xs E (compose d1 d2 ) f = E d1 (λ x, (E (d2 x) f)) as are the procedural semantics: run (sample (pdf, prog)) rng = prog rng run (det xs) rng = (xs, rng) run (compose d1 d2 ) rng = let (x, rng’) := run d1 rng in run (d2 x) rng’ We have defined E and run to correspond; we consider a stochastic program correct if we can prove the relevant theorems about its Func denotation, and we sample from it by passing an RNG to its Prog denotation. 4.4. Formal specification With the background mathematics in place, the next step is to write down the formal specification itself. First, we design types for every other object and function appearing in the informal description. To start, we need a type SCG n to represent stochastic computation graphs on n nodes, and a function SCG.to_dist that takes an SCG n and a scalar parameter θ to a distribution over n real numbers (Dist n). We also need a function cost that takes a graph and the values at each of its nodes and sums the values at the leaf nodes. Figure 4 provides the full types of all objects that will appear in the specification. Now we can write down a type-correct analogue of the informal specification presented in Equation 3: def bprop_spec (bprop : ∀ {n}, SCG n → R → Rn → R) : Prop := ∀ (n : N) (g : SCG n) (θ : R), E (SCG.to_dist g θ) (λ xs, bprop g θ xs) = ∇ (λ θ, E (SCG.to_dist g θ) (λ xs, cost g xs)) θ Given the mathematics libraries, implementing the other Developing Bug-Free Machine Learning Systems With Formal Mathematics objects and functions appearing in the specification such as SCG n and SCG.to_dist is straightforward functional programming. 4.5. Interactive proof While conventional wisdom is that one would write their program before trying to prove it correct, the interactive proof process provides so much helpful information about what the system needs to do that we began working on the proof immediately after drafting the specification. We split the proof into two steps. First, we implemented the simplest possible function that satisfied the specification (that only computed the gradient for a single parameter at a time and did not memoize at all) and proved that correct. Second, we implemented a more performant version (that computed the gradient for multiple parameters simultaneously using memoization) and proved it equivalent to the first one. For the first step, we started with a placeholder implementation that immediately returned zero and let the interactive proof process guide the implementation. Whenever the proof seemed to require induction on a particular data structure, we extended the program to recurse on that data structure; whenever the proof showed that a branch of the program needed to return a value with a given expectation, we worked backwards from that to determine what value to return. Proving the first step also exposed errors in our specification in the form of missing preconditions. For the specification to hold, we needed to make additional assumptions about the graph, e.g. that the identifier for each node in the graph is unique, and that each leaf node is a scalar (WellFormed g). We also needed to assume a generalization of the differentiability requirement mentioned in Schulman et al. (2015), that a subset of the nodes determined by the structure of the graph must be differentiable no matter the result of any stochastic choices (GradsExist g θ). For the second step, we wrote the memoizing implementation before starting the proof and used the process of proving to test and debug it. Although the code for memoizing was simple and short, we still managed to make two implementation errors, one conceptual and one syntactic. Luckily the process of proving necessarily exposes all implementation errors, and in this case made it clear how to fix both of them. We completed the main proof of correctness before proving most of the lemmas that the proof depends on, but the lemmas turned out to be true (except for a few missing preconditions) and so proving them did not expose any additional implementation errors. We also completed the main proof while our axioms were still unsound (see §4.3). When we made our axioms sound and propagated the changes we def bprop_spec (bprop : ∀ {n}, SCG n → R → Rn → R) : Prop := ∀ (n : N) (g : SCG n) (θ : R), WellFormed g ∧ GradsExist g θ ∧ IntegralsExist g θ ∧ CanDiffUnderInts g θ → E (SCG.to_dist g θ) (λ xs, bprop g θ xs) = ∇ (λ θ, E (SCG.to_dist g θ) (λ xs, cost g xs)) θ Figure 5. The final specification for the simplified problem with only scalars (as opposed to tensors) and only a single parameter θ. Our actual system supports arbitrarily-shaped tensors and differentiating with respect to multiple parameters at once. found that our specification required two additional preconditions: that all functions that are integrated over in the theorem statement are indeed integrable (IntegralsExist g θ), and that the many preconditions needed for pushing the gradient over each integral in the expected loss are satisfied (CanDiffUnderInts g θ). However, tracking these additional preconditions did not lead to any changes in our actual implementation. Figure 5 shows the final specification. 4.6. Optimizations We can also use our methodology to verify optimizations that involve mathematical reasoning. When developing machine learning models, one often starts with an easyto-understand model that induces a gradient estimator with unacceptably high variance, and does informal mathematics by hand to derive a new model that has the same objective function but that induces a better gradient estimator. In our approach, the user can write both models and use the process of interactive proving to confirm that they induce the same objective function. Common transformations can be written once and proved correct so that users need only write the first model and the second can be derived and proved equivalent automatically. As part of Certigrad, we wrote a program optimization that integrates out the KL-divergence of the multivariate isotropic Gaussian distribution and we proved once and for all that the optimization is sound. We also verified an optimization that reparameterizes a model so that random variables do not depend on parameters (and so need not be backpropagated through). Specifically, the optimization replaces a node that samples from N (µ, Diag(σ 2 )) with a graph of three nodes that first samples from N (0, In×n ) and then scales and shifts the result according to σ and µ respectively. We applied these two transformations in sequence to a naı̈ve variational-autoencoder to yield the Auto-Encoding Variational Bayes (AEVB) estimator (Kingma & Welling, 2014). Developing Bug-Free Machine Learning Systems With Formal Mathematics at each epoch against the same model and optimization procedure in TensorFlow, both running on 2 CPU cores. We found that the expected losses decrease at the same rate, and that Certigrad takes only 7% longer per epoch (Figure 6). (a) Expected loss vs epoch (b) Running time vs epoch Figure 6. Results of running our certified procedure on an AEVB model, compared to TensorFlow. Our system trains just as well and takes only 7% longer per epoch. 4.7. Verifying backpropagation for specific models Even though we proved that bprop satisfies its formal specification (bprop_spec), we cannot be sure that it will compute the correct gradients for a particular model unless we prove that the model satisfies the preconditions of the specification. Although some of the preconditions are technically undecidable, in practice most machine learning models will satisfy them all for simple reasons. We wrote a (heuristic) tactic program to prove that specific models satisfy all the preconditions and used it to verify that bprop computes the correct gradients for the AEVB model derived in §4.6. 4.8. Running the system We have proved that our system is correct in an idealized mathematical context with infinite-precision real numbers. To actually execute the system we need to replace all real numbers in the program with floating-point numbers. Although doing so technically invalidates the specification and can introduce numerical instability in some cases, this class of errors is well understood (Higham, 2002), could be ruled out as well in principle (Harrison, 2006; Boldo et al., 2015; Ramananandro et al., 2016) and is conceptually distinct from the algorithmic and mathematical errors that our methodology is designed to eliminate. To improve performance, we also replace all tensors with an optimized tensor library (Eigen). This approximation could introduce errors into our system if for whatever reason the Eigen methods we use are not functionally equivalent to ones we formally reason about; of course developers could achieve even higher assurance by verifying their optimized tensor code as well. 4.9. Experiments Certigrad is efficient. As an experiment, we trained an AEVB model with a 2-layer encoding network and a 2layer decoding network on MNIST using the optimization procedure ADAM (Kingma & Ba, 2014), and compared both the expected loss and the running time of our system 5. Discussion Our primary motivation is to develop bug-free machine learning systems, but our approach may provide significant benefits even when building systems that need not be perfect. Perhaps the greatest burden software developers must bear is needing to fully understand how and why their system works, and we found that by formally specifying the system requirements we were able to relegate much of this burden to the computer. Not only were we able to synthesize some fragments of the system (§4.5), we were able to achieve extremely high confidence that our system was bug-free without needing to think about how all the pieces of the system fit together. In our approach, the computer— not the human—is responsible for ensuring that all the local properties that the developer establishes imply that the overall system is correct. Although using our methodology to develop Certigrad imposed many new requirements and increased the overall workload substantially, we found that on the whole it made the development process less cognitively demanding. There are many ways that our methodology can be adopted incrementally. For example, specifications need not cover functional correctness, not all theorems need to be proved, unsound axioms can be used that omit certain preconditions, and more traditional code can be wrapped and axiomatized (as we did with Eigen). When developing Certigrad we pursued the ideal of a complete, machinecheckable proof of functional correctness, and achieved an extremely high level of confidence that the system was correct. However, we realized many of the benefits of our methodology—including partial synthesis and reduced cognitive demand—early in the process before proving most of the lemmas. Although we could not be certain that we had found all of the bugs before we made our axioms sound and filled in the gaps in the formal proofs, in hindsight we had eliminated all bugs early in the process as well. While a pure version of our methodology may already be cost-effective for high-assurance applications, we expect that pragmatic use of our methodology could yield many of the benefits for relatively little cost and could be useful for developing a wide range of machine learning systems to varying standards of correctness. Developing Bug-Free Machine Learning Systems With Formal Mathematics Acknowledgments We thank Jacob Steinhardt, Alexander Ratner, Cristina White, William Hamilton, Nathaniel Thomas, and Vatsal Sharan for providing valuable feedback on early drafts. We also thank Leonardo de Moura, Tatsu Hashimoto, and Joseph Helfer for helpful discussions. This work was supported by Future of Life Institute grant 2016-158712. References Abadi, Martı́n, Agarwal, Ashish, Barham, Paul, Brevdo, Eugene, Chen, Zhifeng, Citro, Craig, Corrado, Greg S., Davis, Andy, Dean, Jeffrey, Devin, Matthieu, Ghemawat, Sanjay, Goodfellow, Ian, Harp, Andrew, Irving, Geoffrey, Isard, Michael, Jia, Yangqing, Jozefowicz, Rafal, Kaiser, Lukasz, Kudlur, Manjunath, Levenberg, Josh, Mané, Dan, Monga, Rajat, Moore, Sherry, Murray, Derek, Olah, Chris, Schuster, Mike, Shlens, Jonathon, Steiner, Benoit, Sutskever, Ilya, Talwar, Kunal, Tucker, Paul, Vanhoucke, Vincent, Vasudevan, Vijay, Viégas, Fernanda, Vinyals, Oriol, Warden, Pete, Wattenberg, Martin, Wicke, Martin, Yu, Yuan, and Zheng, Xiaoqiang. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL http://tensorflow. org/. Software available from tensorflow.org. Bergstra, J., Breuleux, O., Bastien, F., Lamblin, P., Pascanu, R., Desjardins, G., Turian, J., Warde-Farley, D., and Bengio, Y. Theano: a CPU and GPU math expression compiler. In Python for Scientific Computing Conference, 2010. de Moura, Leonardo, Kong, Soonho, Avigad, Jeremy, Van Doorn, Floris, and von Raumer, Jakob. The Lean theorem prover (system description). In Automated Deduction-CADE-25, pp. 378–388. Springer, 2015. Gonthier, Georges. Formal proof–the four-color theorem. Notices of the AMS, 55(11):1382–1393, 2008. Gonthier, Georges, Asperti, Andrea, Avigad, Jeremy, Bertot, Yves, Cohen, Cyril, Garillot, François, Le Roux, Stéphane, Mahboubi, Assia, O’Connor, Russell, Biha, Sidi Ould, et al. A machine-checked proof of the odd order theorem. In Interactive Theorem Proving, pp. 163– 179. Springer, 2013. Gordon, Michael JC. Edinburgh lcf: a mechanised logic of computation. 1979. Gordon, Michael JC and Melham, Tom F. Introduction to hol a theorem proving environment for higher order logic. 1993. Guennebaud, Gaël, Jacob, Benoı̂t, et al. http://eigen.tuxfamily.org, 2010. Eigen v3. Hales, Thomas, Adams, Mark, Bauer, Gertrud, Dang, Dat Tat, Harrison, John, Hoang, Truong Le, Kaliszyk, Cezary, Magron, Victor, McLaughlin, Sean, Nguyen, Thang Tat, et al. A formal proof of the kepler conjecture. arXiv preprint arXiv:1501.02155, 2015. Harrison, John. Hol light: A tutorial introduction. In International Conference on Formal Methods in ComputerAided Design, pp. 265–269. Springer, 1996. Boldo, Sylvie, Jourdan, Jacques-Henri, Leroy, Xavier, and Melquiond, Guillaume. Verified compilation of floatingpoint computations. Journal of Automated Reasoning, 54(2):135–163, 2015. Harrison, John. Floating-point verification using theorem proving. In International School on Formal Methods for the Design of Computer, Communication and Software Systems, pp. 211–242. Springer, 2006. Chen, Haogang, Ziegler, Daniel, Chajed, Tej, Chlipala, Adam, Kaashoek, M Frans, and Zeldovich, Nickolai. Using crash hoare logic for certifying the fscq file system. In Proceedings of the 25th Symposium on Operating Systems Principles, pp. 18–37. ACM, 2015. Higham, Nicholas J. Accuracy and stability of numerical algorithms. SIAM, 2002. Chlipala, Adam. The bedrock structured programming system: Combining generative metaprogramming and hoare logic in an extensible program verifier. In ACM SIGPLAN Notices, volume 48, pp. 391–402. ACM, 2013. Coq Development Team. The Coq proof assistant reference manual: Version 8.5. INRIA, 2015-2016. Coquand, Thierry and Huet, Gérard. The calculus of constructions. Information and computation, 76(2-3):95– 120, 1988. Kingma, D. P. and Welling, M. Auto-encoding variational Bayes. arXiv, 2014. Kingma, Diederik and Ba, Jimmy. method for stochastic optimization. arXiv:1412.6980, 2014. Adam: A arXiv preprint Klein, Gerwin, Elphinstone, Kevin, Heiser, Gernot, Andronick, June, Cock, David, Derrin, Philip, Elkaduwe, Dhammika, Engelhardt, Kai, Kolanski, Rafal, Norrish, Michael, et al. sel4: Formal verification of an os kernel. In Proceedings of the ACM SIGOPS 22nd symposium on Operating systems principles, pp. 207–220. ACM, 2009. Leroy, Xavier. Formal verification of a realistic compiler. Communications of the ACM, 52(7):107–115, 2009. Developing Bug-Free Machine Learning Systems With Formal Mathematics Nipkow, Tobias, Paulson, Lawrence C, and Wenzel, Markus. Isabelle/HOL: a proof assistant for higherorder logic, volume 2283. Springer, 2002. Owre, Sam, Rushby, John M, and Shankar, Natarajan. Pvs: A prototype verification system. In Automated Deduction—CADE-11, pp. 748–752. Springer, 1992. Ramananandro, Tahina, Mountcastle, Paul, Meister, Benoı̂t, and Lethin, Richard. A unified coq framework for verifying c programs with floating-point computations. In Proceedings of the 5th ACM SIGPLAN Conference on Certified Programs and Proofs, pp. 15–26. ACM, 2016. Rudnicki, Piotr. An overview of the mizar project. In Proceedings of the 1992 Workshop on Types for Proofs and Programs, pp. 311–330, 1992. Schulman, John, Heess, Nicolas, Weber, Theophane, and Abbeel, Pieter. Gradient estimation using stochastic computation graphs. In Advances in Neural Information Processing Systems, pp. 3528–3536, 2015.	2cs.AI
arXiv:1512.05999v1 [math.AC] 18 Dec 2015 ON DING INJECTIVE, DING PROJECTIVE, AND DING FLAT MODULES AND COMPLEXES JAMES GILLESPIE Abstract. We characterize Ding modules and complexes over Ding-Chen rings. We show that over a Ding-Chen ring R, the Ding projective (resp. Ding injective, resp. Ding flat) R-modules coincide with the Gorenstein projective (resp. Gorenstein injective, resp. Gorenstein flat) modules, which in turn are nothing more than modules appearing as a cycle of an exact complex of projective (resp. injective, resp. flat) modules. We prove a similar characterization for chain complexes of R-modules: A complex X is Ding projective (resp. Ding injective, resp. Ding flat) if and only if each component Xn is Ding projective (resp. Ding injective, resp. Ding flat). Along the way, we generalize some results of Stovicek and Bravo-Gillespie-Hovey to obtain other interesting corollaries. For example, we show that over any Noetherian ring, any exact chain complex with Gorenstein injective components must have all cotorsion cycle modules. That is, Ext1R (F, Zn I) = 0 for any such complex I and flat module F . On the other hand, over any coherent ring, the cycles of any exact complex P with projective components must satisfy Ext1R (Zn P, A) = 0 for any absolutely pure module A. 1. introduction Let R be a Gorenstein ring in the sense of [Iwa79, Iwa80]. This is a left and right Noetherian ring R having finite injective dimension as both a left and right module over itself. Over such rings, exact chain complexes of projective, injective, and flat R-modules each have very nice homological properties. In particular, for an exact complex P of projectives, the complex HomR (P, Q) is also exact for any projective module Q. In general, such complexes are called totally acyclic complexes of projectives and the modules appearing as a cycle Zn P of such a complex are called Gorenstein projective. So over Gorenstein rings, any exact complex of projectives is totally acyclic and their cycles are all Gorenstein projective. Similar statements hold for exact complexes of injectives and flats, and the corresponding Gorenstein injective and Gorenstein flat modules. Gorenstein homological algebra is the study of these modules and complexes and the theory is particularly satisfying over Gorenstein rings. In this case, most results from traditional homological algebra have an analog in Gorenstein homological algebra. The book [EJ00] is a standard reference, but many authors have studied the subject. Stovicek recently proved the coherent analog of the above result concerning totally acyclic complexes of injectives [Sto14, Prop. 7.9]. It raises the same question for the projective and flat analogs, which is answered in this paper. Before explaining further, first recall that a (left) coherent ring is one in which all finitely Date: December 21, 2015. 1 2 JAMES GILLESPIE generated (left) ideals are finitely presented. A ring that is both left and right coherent is called coherent and these include all Noetherian and Von Neumann regular rings. A lesson learned from [Ste70] is that many results in homological algebra extend from Noetherian to coherent rings by replacing finitely generated modules with finitely presented modules. In the process, injective modules are replaced with FP-injective (i.e. absolutely pure) modules. In this way, a Ding-Chen ring is a coherent ring R which has finite absolutely pure (i.e. FP-injective) dimension as both a left and right module over itself. They were introduced in [DC93, DC96]. Since a coherent ring is Noetherian if and only if the FP-injective modules coincide with the injective modules, the Ding-Chen rings are nothing more than Gorenstein rings whenever R is Noetherian. In [Gil10], the author noted that the work of Ding, Mao and Li, see especially [DM05, DM07, DM08, DLM09], provides a natural way to extend notions from Gorenstein homological algebra from Noetherian to coherent rings. In the process, the Gorenstein modules are replaced by Ding’s modules. For example, we say a module is Ding projective if it is a cycle module of some exact complex P of projectives such that HomR (P, F ) remains exact for all flat modules F . See Definitions 4.1, 6.1, and 7.1. It admittedly seems strange to require that HomR (P, F ) remain exact for all flat modules F rather than just all projectives. After all, it feels like we are requiring too much. However, it was shown in [BGH13] that the Ding projectives are the cofibrant objects of an especially nice Quillen model structure on R-Mod, the category of (left) R-modules, whenever R is coherent. It implies that every module M can be approximated by a Ding projective and that the full subcategory of all Ding projective modules naturally form a Frobenius category. We think of the associated stable category as the (projective) stable module category of R. The analog of this for the usual Gorenstein projectives over Noetherian rings simply doesn’t seem to be true, in general. The first result of this paper is an extension, to Ding-Chen rings, of the well known result concerning Gorenstein rings described in the first paragraph. Part (2) is the result of Stovicek which motivated the analogous question for Ding projectives and Ding flats. Theorem 1.1. Let R be a Ding-Chen ring. That is, a left and right coherent ring R such that R R and RR each have finite absolutely pure dimension. Let X be an exact chain complex. Then (1) If each component Xn is projective, then each cycle Zn X is Ding projective. Indeed HomR (X, F ) remains exact for all flat modules F . (2) If each component Xn is injective, then each cycle Zn X is Ding injective. Indeed HomR (A, X) remains exact for all absolutely pure A. (3) If each component Xn is flat, then each cycle Zn X is Ding flat. Indeed A ⊗R X remains exact for all absolutely pure A. Consequently, the Gorenstein modules coincide with the Ding modules whenever R is a Ding-Chen ring. Before discussing the proof methods we note that Gorenstein chain complexes have also been studied quite a bit. For example, it has been known for some time that over Gorenstein rings, a chain complex X is Gorenstein injective (resp. Gorenstein projective) if and only if each component is Gorenstein injective (resp. Gorenstein projective). See [GR99, EE05, YL11]. In [YLL11, YLL13], Yang, Liu, DING MODULES AND COMPLEXES 3 and Liang characterize the Ding complexes over a Ding-Chen ring. The next result is a surprising refinement of their characterization. Theorem 1.2. Let R be a Ding-Chen ring and X a chain complex of R-modules. (1) X is Ding projective in Ch(R) if and only if each component Xn is a Ding projective R-module. (2) X is Ding injective in Ch(R) if and only if each component Xn is a Ding injective R-module. (3) X is Ding flat in Ch(R) if and only if each component Xn is a Ding flat R-module. Consequently, the Gorenstein complexes coincide with the Ding complexes whenever R is a Ding-Chen ring. We now briefly describe our methods used and highlight other results in the paper. First, the injective cases are proved in a completely different fashion than their projective analogs, (while the flat analogs are relatively easy). For the injective case, we generalize the approach of Stovicek to record a useful result concerning cotorsion pairs and direct limits. If (W, F ) is a cotorsion pair, then W is closed under direct limits whenever W satisfies the two out of three property on short exact sequences; see Proposition 3.1. The injective cases of Theorems 1.1 and 1.2 follow as corollaries, but the approach leads to a string of other interesting corollaries appearing in Sections 3 and 4. Perhaps the most interesting is Corollary 3.5. It implies that over any Noetherian ring, any exact chain complex with Gorenstein injective components must have all cotorsion cycle modules. That is, Ext1R (F, Zn I) = 0 for any such complex I and flat module F . The projective case requires a completely different approach than the injective one. Our proof of the projective part of Theorem 1.1 relies on a result proved in [BGH13, Theorem A.6]. To obtain the projective statement in Theorem 1.2 we in fact need to generalize this result to the category of chain complexes. Section 5 is devoted to this, with Theorem 5.9 being the main result. Finally, the following interesting result is proved in Theorem 3.6. Theorem 1.3. Let R be a coherent ring and P any exact complex of projectives. Then Ext1R (Zn P, A) = 0 for any absolutely pure module A. Regarding the structure of the paper, after some preliminaries in Section 2, we first address the injective cases in Sections 3 and 4. Sections 5 and 6 are devoted to the projective case, and Section 7 to the flat case. 2. preliminaries Throughout the paper R denotes a general ring with identity. An R-module will mean a left R-module, unless stated otherwise. The category of R-modules will be denoted R-Mod. It is an abelian category. 2.1. Chain complexes on abelian categories. Let A be an abelian category. We denote the corresponding category of chain complexes by Ch(A). In the case A = R-Mod, we denote it by Ch(R). Our convention is that the differentials of dn+1 d n Xn−1 − → · · · is a our chain complexes lower degree, so · · · − → Xn+1 −−−→ Xn −→ chain complex. We also have the chain homotopy category of A, denoted K(A). Its objects are also chain complexes but its morphisms are chain homotopy classes 4 JAMES GILLESPIE of chain maps. Given a chain complex X, the nth suspension of X, denoted Σn X, is the complex given by (Σn X)k = Xk−n and (dΣn X )k = (−1)n dk−n . For a given object A ∈ A, we denote the n-disk on A by Dn (A). This is the complex consisting 1A A concentrated in degrees n and n − 1, and 0 elsewhere. We denote only of A −−→ the n-sphere on A by S n (A), and this is the complex consisting only of A in degree n and 0 elsewhere. Given two chain complexes X, Y ∈ Ch(A) we define Hom(X, Y ) to be the comQ δn Q → plex of abelian groups · · · − → k∈Z Hom(Xk , Yk+n ) −→ k∈Z Hom(Xk , Yk+n−1 ) − · · · , where (δn f )k = dk+n fk −(−1)n fk−1 dk . We get a functor Hom(X, −) : Ch(A) − → Ch(Z). Note that this functor takes exact sequences to left exact sequences, and it is exact if each Xn is projective. Similarly the contravariant functor Hom(−, Y ) sends exact sequences to left exact sequences and is exact if each Yn is injective. It is an exercise to check that the homology satisfies Hn [Hom(X, Y )] = K(A)(X, Σ−n Y ). Being an abelian category, Ch(A) comes with Yoneda Ext groups. In particular, Ext1Ch(A) (X, Y ) will denote the group of (equivalences classes) of short exact sequences 0 − → Y − → Z − → X − → 0 under the Baer sum operation. There is a subgroup Ext1dw (X, Y ) ⊆ Ext1Ch(A) (X, Y ) consisting of the “degreewise split” short exact sequences. That is, those for which each 0 − → Yn − → Zn − → Xn − → 0 is split exact. The following lemma gives a well-known connection between Ext1dw and the above hom-complex Hom. Lemma 2.1. For chain complexes X and Y , we have isomorphisms: Ext1 (X, Σ(−n−1) Y ) ∼ = Hn Hom(X, Y ) = K(A)(X, Σ−n Y ) dw In particular, for chain complexes X and Y , Hom(X, Y ) is exact iff for any n ∈ Z, any chain map f : Σn X − → Y is homotopic to 0 (or iff any chain map f : X − → Σn Y is homotopic to 0). In the case of A = R-Mod, we recall the usual tensor product of chain complexes. Given that X (resp. Y ) is a complex of right (resp. left) R-modules, the tensor product X ⊗Y is defined by (X ⊗Y )n = ⊕i+j=n (Xi ⊗Yj ) in degree n. The boundary map δn is defined on the generators by δn (x ⊗ y) = dx ⊗ y + (−1)\|x\| x ⊗ dy, where \|x\| is the degree of the element x. 2.2. The modified Hom and Tensor complexes. Here again, Ch(R) denotes the category of chain complexes of R-modules. The above Hom is often referred to as the internal hom, for in the case that R is commutative, Hom(X, Y ) is again an object of Ch(R). Note that the cycles in degree 0 of the internal hom coincide with the external hom functor: Z0 [Hom(X, Y )] ∼ = HomCh(R) (X, Y ). This idea in fact is useful to define an alternate internal hom as follows. Given X, Y ∈ Ch(R), we define Hom(X, Y ) to be the complex Hom(X, Y )n = Zn Hom(X, Y ) with differential λn : Hom(X, Y )n − → Hom(X, Y )n−1 defined by (λf )k = (−1)n dk+n fk . Notice that the degree n component of Hom(X, Y ) is exactly HomCh(R) (X, Σ−n Y ). In this way we get an internal hom Hom which is useful for categorical considerations in Ch(R). For example, Hom(X, −) is a left exact functor, and is exact if and only if X is projective in the category Ch(R). On DING MODULES AND COMPLEXES 5 the other hand, Hom(−, Y ) is exact if and only if Y is injective in Ch(R). There i are corresponding derived functors which we denote by Ext . They satisfy that i Ext (X, Y ) is a complex whose degree n is ExtiCh(R) (X, Σ−n Y ). Similarly, the usual tensor product of chain complexes does not characterize categorical flatness. For this we need the modified tensor product and its left derived torsion functor from [EGR97] and [GR99]. We will denote it by ⊗, and it is defined in terms of the usual tensor product ⊗ as follows. Given a complex X of right R-modules and a complex Y of left R-modules, we define X⊗Y to be the complex whose nth entry is (X ⊗ Y )n /Bn (X ⊗ Y ) with boundary map (X ⊗ Y )n /Bn (X ⊗ Y ) → (X ⊗ Y )n−1 /Bn−1 (X ⊗ Y ) given by x ⊗ y 7→ dx ⊗ y. This defines a complex and we get a bifunctor −⊗− which is right exact in each variable. We denote the corresponding left derived functors by Tori . We refer the reader to [GR99] for more details. 2.3. Finitely generated projective complexes. A standard characterization of projective objects in Ch(R) is the following: A complex P is projective if and only if it is an exact complex with each cycle Zn P a projective R-module. We also recall that, by definition, a chain complex X is finitely generated if whenever X = Σi∈I Si , for some collection {Si }i∈I of subcomplexes of X, then there exists a finite subset J ⊆ I for which X = Σi∈J Si . It is a standard fact that X is finitely generated if and only if it is bounded (above and below) and each Xn is finitely generated. We say that a chain complex X is of type F P∞ if it has a projective resolution by finitely generated projective complexes. Certainly any such X is finitely presented and hence finitely generated. 2.4. Absolutely clean and level complexes; character duality. The so-called level and absolutely clean modules were introduced in [BGH13] as generalizations of flat modules over coherent rings and injective modules over Noetherian rings. The same notions in the category Ch(R) were also studied in [BG15]. Here we recall some definitions and results from [BG15]. Definition 2.2. We call a chain complex A absolutely clean if Ext1Ch(R) (X, A) = 1 0 for all chain complexes X of type F P∞ . Equivalently, if Ext (X, A) = 0 for all complexes X of right R-modules of type F P∞ . On the other hand, we call a chain complex L level if Tor1 (X, L) = 0 for all chain complexes X of right R-modules of type F P∞ . We refer to [BG15, Prop. 2.6/4.6] for proof of the following. Proposition 2.3. A chain complex A is absolutely clean if and only if A is exact and each Zn A is an absolutely clean R-module. A chain complex L is level if and only if L is exact and each Zn L is a level R-module. Recall that the character module of M is defined as M + = HomZ (M, Q/Z), and that M + is a right (resp. left) R-module whenever M is a left (resp. right) Rmodule. The construction extends to chain complexes: Given a chain complex X, we have X + = HomZ (X, Q/Z). Since Q/Z is an injective cogenerator for the category of abelian groups, the functor HomZ (−, Q/Z) preserves and reflects exactness. 6 JAMES GILLESPIE So Proposition 2.3 immediately gives us the following corollary due to the perfect character module duality between absolutely clean and level modules [BGH13, Theorem 2.10]. Corollary 2.4. A chain complex L of left (resp. right) modules is level if and only if L+ = HomZ (L, Q/Z) is an absolutely clean complex of right (resp. left) modules. And, a chain complex A of left (resp. right) modules is absolutely clean if and only if A+ = HomZ (A, Q/Z) is a level complex of right (resp. left) modules. 2.5. Cotorsion pairs. Let A be an abelian category. By definition, a pair of classes (X , Y) in A is called a cotorsion pair if Y = X ⊥ and X = ⊥ Y. Here, given a class of objects C in A, the right orthogonal C ⊥ is defined to be the class of all objects X such that Ext1A (C, X) = 0 for all C ∈ C. Similarly, we define the left orthogonal ⊥ C. We call the cotorsion pair hereditary if ExtiA (X, Y ) = 0 for all X ∈ X , Y ∈ Y, and i ≥ 1. The cotorsion pair is complete if it has enough injectives and enough projectives. This means that for each A ∈ A there exist short exact sequences 0 − →A− →Y − →X − → 0 and 0 − → Y′ − → X′ − →A− → 0 with X, X ′ ∈ X ′ and Y, Y ∈ Y. Standard references include [EJ00] and [GT06] and connections to abelian model categories can be found in [Hov02] and [Gil15]. Recall that a Grothendieck category is a cocomplete abelian category with a set of generators and such that direct limits are exact. By an injective cotorsion pair (W, F ) in a Grothendieck category G we mean a complete cotorsion pair for which W is thick and W ∩F is the class of injective objects. Since Grothendieck categories have enough injectives it turns out that such a cotorsion pair is equivalent to an injective model structure on G. By this we mean the model structure is abelian in the sense of [Hov02] and all objects are cofibrant. The fibrant objects in this case are exactly F and the trivial objects are exactly W. See [Gil12] for more on injective cotorsion pairs. We also have the dual notion of projective cotorsion pairs (C, W) which give us projective model structures on abelian categories with enough projectives. 3. Injective cotorsion pairs and direct limits Let (W, F ) be a cotorsion pair in R-Mod, or Ch(R), or some other Grothendieck category. Assume that W is thick; that is, assume W satisfies the two out of three property on short exact sequences. In this section we show that W must be closed under direct limits. This leads to several interesting corollaries. The proof of the proposition below is adapted from Stovicek’s [Sto14, Prop. 5.3]. After studying the proof, the author simply realized that it can be reinterpreted to yield the following convenient result. Proposition 3.1. Let (W, F ) be a cotorsion pair in R-Mod with W thick. Then W is closed under direct limits. Proof. As a first step we show that W is closed under direct unions. It is in fact enough to prove this for well-ordered continuous direct unions. So assume that we have a λ-sequence of module monomorphisms X0 ֒→ X1 ֒→ X2 ֒→ · · · ֒→ Xi ֒→ Xi+1 ֒→ · · · with Xi ∈ W for each i < λ. Then each Xi+1 /Xi ∈ W by the thickness hypothesis. Since the diagram of modules is assumed continuous we see that its colimit is the same thing as the transfinite extension of the X0 , Xi+1 /Xi . Thus its colimit is DING MODULES AND COMPLEXES 7 in W by Eklof’s lemma (assuming λ is a limit ordinal; if it is a successor ordinal λ = α + 1, then the colimit coincides with Xα ∈ W). Step two: We show that W is closed under direct limits. Again, it is enough to prove this for well-ordered continuous direct limits. (See [AR94, Section 1.6], especially Corollary 1.7 and the Remark that follows it where well-ordered direct limits are referred to as chains and continuous well-ordered direct limits as smooth chains.) So we consider a λ-diagram f0,1 f1,2 f2,3 fi,i+1 → ··· X0 −−→ X1 −−→ X2 −−→ · · · Xi −−−−→ Xi+1 − with each Xi ∈ W and Xγ = limi<γ {Xi , fi,i+1 } for each limit ordinal γ < λ. Our −→ job is to show that limi<λ Xi ∈ W. We will assume λ is a limit ordinal, for otherwise −→ the direct limit just equals Xλ−1 ∈ W. Following a standard way for defining direct limits (for example, see [Sten75, Prop IV.8.4, p.100]), limi<λ Xi is the cokernel of −→ L L → i<λ Xi where the first direct sum is the following homomorphism: i<j Xij − taken over all pairs i < j < λ, and Xij = Xi is just a copy of the domain of the map fij Xi −−→ Xj , and the map is defined on the ith coordinate by xi 7→ ei xi − ej fij xi where the ei denotes L the canonical injection into the coproduct. In other words, the direct limit is ( i<λ Xi )/K where K is the image of this map: K = < ei xi − ej fij xi \| xi ∈ Xi and i < j < λ > Note that since the maps ei and fij are linear we have that K is the set of all finite sums of the form ei xi − ej fij xi where the xi range through Xi and i < j ranges through all i < j < λ. There is a short exact sequence ⊆ M 0− →K− → → 0, Xi − → lim Xi − −→ i<λ i<λ and since W is closed under direct sums and is thick, it is enough to show that K ∈ W. We now show that K is a direct union (not well ordered, but still a direct union) of modules in W. So the proof will follow from the above Step 1. Thinking of λ as the set of all its smaller ordinals we define, for each finite subset J ⊆ λ with \|J\| > 1, the mapping M M Xi − → Xi φJ : i∈J\{j} i<λ where j denotes the maximum element of the finite subset J, and the map is defined on the ith coordinate via xi 7→ ei xi −ej fij xi . One now verifies each of the following: (1) S = {J ⊆ λ \| 1 < \|J\| < ω} is a directedLposet and there is a functor D:S− → R-Mod defined on objects by J 7→ i∈J\{j} Xi , and on arrows by taking an inclusion J ⊆ J ′ to the map DJJ ′ defined on the ith coordinate as follows: xi 7→ ei xi if j = j ′ (that is, just a natural inclusion if J and J ′ have the same maximal element), but via xi 7→ ei xi − ej fij xi if j < j ′ . (2) Each DJJ ′ is a monomorphism. In fact it is a split monomorphism with retraction map the canonical projection. (3) Each φJ is also L a split monomorphism with similar retraction. The image of φJ identifies i∈J\{j} Xi with the submodule KJ = < ei xi − ej fij xi \| xi ∈ Xi and i ∈ J\{j} > . 8 JAMES GILLESPIE That is KJ is the set of all finite sums of elements of the form ei xi − ej fij xi as i ranges through J\{j} and xi ranges through Xi . (4) The direct system of monomorphisms D : S − → R-Mod is isomorphic via the natural transformation {φJ } to the direct system of submodules KJ . The direct limit limi<λ D identifies with the direct union of submodules L −→ S KJ ⊆ i<λ Xi . J∈SS (5) K = J∈S KJ . L Now since each i∈J\{j} Xi ∈ W, we conclude from (3), (4), (5) and Step 1 that K ∈ W. This completes the proof. Of course generalizations of Proposition 3.1 to categories beyond R-Mod are possible and will be important for applications. The reader will note that the above proof holds in any Grothendieck category, giving the following generalization. Proposition 3.2. Let (W, F ) be a cotorsion pair in a Grothendieck category G with W thick. Then W is closed under direct limits. In particular, the class W of trivial objects is closed under direct limits whenever (W, F ) be an injective cotorsion pair. Proposition 3.2 has some interesting consequences that apply to each of the injective cotorsion pairs recently appearing in [BGH13, Gil12]. For example, it is shown in [BGH13] that for any ring R we have an injective cotorsion pair (W, GI) in R-Mod where GI is the class of Gorenstein AC-injective modules. The next corollary says that all modules of finite flat dimension are sent to zero in the corresponding stable homotopy category. Consequently, any Gorenstein AC-injective module is cotorsion and must be injective if it has finite flat dimension. Corollary 3.3. Let (W, F ) be an injective cotorsion pair in R-Mod or Ch(R). Then W contains all objects of finite flat dimension. Consequently, any fibrant object is cotorsion and must be injective whenever it is of finite flat dimension. Proof. Certainly W contains all projectives and hence all flat objects since these are precisely direct limits of projectives. Since W is thick and contains the flat objects it also contains all objects of finite flat dimension. The final statement follows since W ∩ F coincides with the class of injective objects. Recall that a chain complex is flat if it is exact and each cycle module is a flat R-module. These complexes are categorically flat in Ch(R); they are direct limits of finitely generated projective complexes [GR99, Theorem 4.1.3]. A complex X is called DG-flat if each Xn is flat and E ⊗ X is exact whenever E is an exact chain complex of (right) R-modules. E ⊗ X is the usual tensor product of chain complexes recalled in Section 2.1. Again, the book [GR99] is a standard reference for DG-flat complexes. Corollary 3.4. Let (W, F ) be an injective cotorsion pair in Ch(R). Then each of the following hold: (1) Any exact complex Y ∈ F has all cotorsion cycle modules. That is, each Zn Y , and hence each Yn , is a cotorsion module. (2) If every complex in F is exact, then each DG-flat complex is in W. Here we will use the notation from [Gil04] where dg Ce denotes the class of all complexes C with each Cn cotorsion and with the property that all chain maps F − → C are null homotopic whenever F is a flat chain complex. It turns out that DING MODULES AND COMPLEXES 9 dg Ce is precisely the right Ext orthogonal to the class of flat complexes. That is, dg Ce is really the class of cotorsion objects in the category of complexes. e Proof. For (1), we have that Y is cotorsion by Corollary 3.3. That is, Y ∈ dg C. The result now follows from [Gil04, Theorem 3.12] which tells us that the exact complexes in dg Ce are precisely the ones with cotorsion cycle modules. For (2), in the case that every complex Y ∈ F is exact we have Ext1Ch(R) (S n (P ), Y ) ∼ = Ext1R (P, Zn Y ) = 0 for each projective module P . So S n (P ) ∈ W for each projective P , and hence S n (F ) ∈ W for each flat module F . But it follows from [Gil07, Prop. 3.8] that each DG-flat complex is a retract of a transfinite extension of complexes of the form S n (F ) with F flat. Applying the well known Eklof’s lemma, we conclude that all DG-flat complexes are in W, proving (2). Going back to the Gorenstein AC-injective modules appearing before Corollary 3.3, we get the following corollary. Corollary 3.5. Any chain complex X of Gorenstein AC-injective modules is in e Consequently, any exact complex of Gorenstein AC-injectives must have codg C. torsion cycle modules. Proof. We refer to [Gil12, Proposition 7.2]. It shows that there is an injective cotorsion pair in Ch(R) whose right side consists of the class of all complexes of Gorenstein AC-injective modules. The rest follows from what we already observed above. Note that Corollary 3.5 implies Stovicek’s original result that any exact complex of injectives must have cotorsion cycle modules. However, the reader familiar with [Sto14] will realize that our methods are just a generalization of those employed by Stovicek. These methods however don’t dualize. So the following result is quite interesting. It is the dual of Stovicek’s result, assuming R is a coherent ring. Theorem 3.6. Let R be a coherent ring and P an exact complex of projectives. Then Ext1R (Zn P, A) = 0 for any absolutely pure module A. Consequently, any cycle Zn P must be a retract of a transfinite extension of finitely presented modules. Proof. Since P is exact, we have Ext1R (Zn P, A) = Ext1Ch(R) (P, S 0 (A)), and so we can instead show that Ext1Ch(R) (P, S 0 (A)) = 0 for all absolutely pure A. But it is enough to show Hom(P, S 0 (A)) = HomR (P, A) is exact whenever P is an exact complex of projectives and A is absolutely pure. Let D denote the class of all absolutely pure R-modules, and C denote the class of all flat (right) R-modules. Since R is coherent we have that (C, D) forms a perfect duality pair with respect to the character modules of Section 2.4. Therefore, by [BGH13, Theorem A.6], HomR (P, A) is exact for all absolutely pure A if and only if F ⊗R P is exact for all flat F . But certainly this is true. The last statement holds since the “set” S, of all finitely presented modules, cogenerates the complete cotorsion pair (⊥ (S ⊥ ), S ⊥ ) and S ⊥ is precisely the class of all absolutely pure modules. 10 JAMES GILLESPIE Remark. Suppose R is any ring in which all level modules have finite flat dimension. Then the above argument will extend to show Ext1R (Zn P, A) = 0 for all absolutely clean modules A and exact complexes of projectives P . 4. Applications to Ding injective modules and complexes This section is devoted to studying Ding injective modules and chain complexes over Ding-Chen rings. As in [Gil10], a Ding-Chen ring is a (left and right) coherent ring with R R and RR each of finite FP-injective (absolutely pure) dimension. We refer to [Gil10] for more on Ding-Chen rings and Ding injective modules although the definitions below should suffice for our purposes. Definition 4.1. We call an R-module M Ding injective if there exists an exact complex of injectives · · · → I1 → I0 → I 0 → I 1 → · · · with M = ker (I 0 → I 1 ) and which remains exact after applying HomR (A, −) for any absolutely pure module A. In the same way, we call a chain complex X Ding injective if there exists an exact complex of injective complexes · · · → I1 → I0 → I 0 → I 1 → · · · with X = ker (I 0 → I 1 ) and which remains exact after applying HomCh(R) (A, −) for any absolutely pure chain complex A. Recall that a chain complex I is injective (resp. absolutely pure) in Ch(R) if and only if it is exact and each cycle Zn I is an injective (resp. absolutely pure) module. The content of Corollaries 4.3 and 4.4 is that the Hom condition in the above definitions come automatically when R is a Ding-Chen ring. In fact, for the module case, we have the following stronger statement. Theorem 4.2. Let R be a Ding-Chen ring. Then M is Ding injective if and only if M = Z0 I for some exact complex I of Ding injective R-modules. That is, any exact complex of Ding injectives automatically has Ding injective cycles. Proof. Since R is Ding-Chen, the modules of finite flat dimension coincide with the modules of finite absolutely pure dimension [DC93]. Denote this class by W. We have that (W, DI) is an injective cotorsion pair, where DI is the class of all Ding injective modules; see [DM07, Theorem 3.4] and [Gil10, Corollary 4.5]. f exDI) f By [Gil12, Prop. 7.2] it lifts to another injective cotorsion pair (⊥ exDI, f is the class of all exact complexes of Ding injectives. By on Ch(R), where exDI f for each flat module F and hence Corollary 3.4 (2) we have that S n (F ) ∈ ⊥ exDI f we have for each F of finite flat dimension. Thus for all W ∈ W and I ∈ exDI 1 1 n 0 = Ext (S (W ), I) = ExtR (W, Zn I). Hence each Zn I must be Ding injective. Since injective modules are Ding injective we deduce the following corollary. Corollary 4.3. [Sto14, Prop. 7.9] Let R be a Ding-Chen ring. Then M is Ding injective if and only if M = Z0 I for some exact complex I of injective R-modules. Therefore, for any absolutely pure A, any exact complex I of injectives will remain exact after applying HomR (A, −). DING MODULES AND COMPLEXES 11 It was shown by Yang, Liu, and Liang in [YLL13] that the Ding injective complexes are precisely the complexes X for which each Xn is Ding injective and all chain maps A − → X are null homotopic whenever A is an absolutely pure (FPinjective) chain complex. We now show that the null homotopy condition is automatic when R is a Ding-Chen ring. Corollary 4.4. Let R be a Ding-Chen ring. Then a complex X is Ding injective if and only if each Xn is a Ding injective R-module. Proof. The “only if” part is easy from the definition of Ding injective. For the converse, we use that for any coherent ring R, the Ding injective cotorsion pair f dwDI) f on Ch(R), (W, DI) on R-Mod lifts to an injective model structure (⊥ dwDI, f by [Gil12, Proposition 7.2]. Here, dwDI is the class of all complexes of Ding injective modules. From Corollary 3.3 we have that each complex of finite flat f But in the case that R is Ding-Chen, it is easy to see that dimension is in ⊥ dwDI. any absolutely pure complex A has finite flat dimension. Indeed such an A is an exact complex with each cycle module Zn A absolutely pure. Thus each Zn A has finite flat dimension. Now taking a flat resolution of A, the upper bound on the f flat dimensions implies that A itself has finite flat dimension. Hence A ∈ ⊥ dwDI for each absolutely pure A. It follows that Hom(A, X) is exact for any absolutely f whenever R is Ding-Chen. pure A and any complex X ∈ dwDI, Remark. It was shown by Yang, Liu, and Liang in [YLL13] that the class of Ding f whenever R is Ding-Chen. So we now have injective complexes is precisely dg DI f f f = DI, f whenever R is a Ding-Chen ring. shown that both dg DI = dwDI and exDI (Again, we are using the notation from [Gil04].) 5. Complexes of projective complexes Our main purpose here is to prove Theorem 5.9, which will be our tool in Section 6 for characterizing Ding projective complexes. We get this by generalizing results in [BGH13, Appendix A], from modules to chain complexes. Since Ch(R) is itself an abelian category we can of course consider Ch(Ch(R)), the category of chain complexes of chain complexes. Using [Wei94, Sign Trick 1.2.5], the category Ch(Ch(R)) can be identified with the category of bicomplexes. However, for our purpose here it is somewhat easier to stick with the category Ch(Ch(R)). 5.1. Free chain complexes. A free R-module is one that is isomorphic to a direct sum of copies of R. Analogously, we say a chain complex F ∈ Ch(R) is free if it is isomorphic to a direct sum ⊕i∈I Dni (R) where each ni is some integer. Clearly this is equivalent to saying F is isomorphic to ⊕n∈Z Dn (Fn ) where each Fn is some free R-module. It is also equivalent to define F to be an exact complex with each Zn F a free R-module. However, it will be sufficient and most convenient for us to use the representations ⊕i∈I Dni (R). Evidently free complexes are closed under arbitrary direct sums. It is also easy to check that free complexes are projective objects in Ch(R), and that each projective complex P is a retract of a free complex. Lemma 5.1 (Eilenberg’s swindle). Given any projective chain complex P , there exists a free chain complex F such that P ⊕ F ∼ = F. 12 JAMES GILLESPIE Proof. We follow [Lam99, Corollary 2.7]. Since P is projective, as noted above we can find another (projective) complex Q such that P ⊕ Q is a free complex. Setting F = (P ⊕ Q) ⊕ (P ⊕ Q) ⊕ (P ⊕ Q) ⊕ · · · will produce a free complex as desired. Indeed ∼ P ⊕ (Q ⊕ P ) ⊕ (Q ⊕ P ) ⊕ (Q ⊕ P ) · · · ∼ F = = P ⊕ F. We now need to study chain complexes of projective chain complexes. That is, objects of Ch(Ch(R)) with each component a projective complex. Lemma 5.2. Let R be any ring, and P ∈ Ch(Ch(R)) a complex of projective chain complexes. Then P is a direct summand of a complex F of free chain complexes. Furthermore, if P is exact then F can be taken to be exact. Proof. By Lemma 5.1 we may, for each Pn , find a free chain complex F n such that Pn ⊕ F n ∼ = F n . We form F = P ⊕ (⊕n∈Z Dn (F n )). This is a complex of complexes which in degree n is the chain complex Pn ⊕ F n ⊕ F n+1 ∼ = F n ⊕ F n+1 , and so a free chain complex. We are done since P is a direct summand of P ⊕ (⊕n∈Z Dn (F n )). Moreover, P ⊕ (⊕n∈Z Dn (F n )) is exact whenever P is exact. Theorem 5.3. Let R be any ring, and let S be the set of all bounded above complexes of finitely generated free complexes. The cotorsion pair (⊥ (S ⊥ ), S ⊥ ) in Ch(Ch(R)), cogenerated by S, is functorially complete. Moreover, C = ⊥ (S ⊥ ) is the class of all complexes of projective complexes. Proof. Since Ch(Ch(R)) is a Grothendieck category with enough projectives it follows from [Hov02, Corollary 6.8] that S cogenerates a functorially complete cotorsion pair (⊥ (S ⊥ ), S ⊥ ) in Ch(Ch(R)). Setting C = ⊥ (S ⊥ ), our aim is to show that this is precisely the class of all complexes of projective complexes. We note that S contains a set of projective generators for Ch(Ch(R)); for example, the set {Dn (Dm (R))}n,m∈Z . So it follows from [Hov02, Corollary 6.9/Proof of 6.5] that ⊥ (S ⊥ ) is precisely the class of all direct summands of transfinite extensions of objects in S. By Lemma 5.2 we only need to show that a complex of free complexes is a transfinite extension of bounded above complexes of finitely generated free complexes. So let F be a complex of free complexes and write each Fn = ⊕i∈In Dni (Ri ) for some indexing set In with each Ri = R. We assume that F is not (isomorphic to) a complex in S, so we can certainly find a nonzero Fn , and we then take just one summand Dnj (Rj ) for some j ∈ In . We start to build a bounded above subcomplex X ⊆ F by setting Xn = Dnj (Rj ) and setting Xi = 0 for all i > n. Now note d(Dnj (Rj )) ⊆ ⊕i∈In−1 Dni (Ri ) and that we can find a finite subset Ln−1 ⊆ In−1 such that d(Dnj (Rj )) ⊆ ⊕i∈Ln−1 Dni (Ri ). We set Xn−1 = ⊕i∈Ln−1 Dni (Ri ). We can continue down in the same way finding Ln−2 ⊆ In−2 with \|Ln−2 \| finite and with d(⊕i∈Ln−1 Dni (Ri )) ⊆ ⊕i∈Ln−2 Dni (Ri ). In this way we can construct the subcomplex → ··· → ⊕i∈Ln−2 Dni (Ri ) − → ⊕i∈Ln−1 Dni (Ri ) − X = ··· − →0− → Dnj (Rj ) − Note that X is a nonzero bounded above complex of finitely generated free complexes. That is, X ∈ S. We set X0 = X. To finish the proof we will argue that we can write F as the union of a continuous chain 0 6= X0 ( X1 ( · · · ( Xα ( Xα+1 ( · · · with each Xα+1 /Xα ∈ DING MODULES AND COMPLEXES 13 S. So we consider F/X0 , and in the same way as above we find a bounded above subcomplex X1 /X0 ⊆ F/X0 consisting of finitely generated free complexes in each degree. However, we are careful to construct X1 as follows: Let L0n denote the indexing sets of the previously constructed complex X0n = ⊕i∈L0n Dni (Ri ). Note that we can identify the quotient F/X0 with a complex whose degree n entry is ⊕i∈In −L0n Dni (Ri ). And so we may take X1 so that X1n = ⊕i∈L1n Dni (Ri ) with each L1n finite and L0n ⊆ L1n ⊆ In for each n. Continuing this process, we continue to construct an increasing union 0 6= X0 ( X1 ( X2 ( · · · corresponding to a nested union of subsets L0n ⊆ L1n ⊆ L2n ⊆ · · · for each n. Assuming this process doesn’t terminate we set Xω = ∪α<ω Xα and note that Xω Dni (Ri ) where n = ⊕i∈Lω n ω α ω ω Ln = ∪α<ω Ln . Of course X and F/X are complexes of free complexes and continuing this process we can obtain an ordinal λ and a continuous chain 0 6= X0 ( X1 ( · · · ( Xα ( Xα+1 ( · · · with F = ∪α<λ Xα and X0 , Xα+1 /Xα ∈ S. 5.2. Pure exact complexes of complexes. Let E : 0 − →X − →Y − →Z− → 0 be a short exact sequence in Ch(R). We say that E is pure if for each finitely presented complex F , the sequence of abelian groups HomCh(R) (F, E), that is 0− → HomCh(R) (F, X) − → HomCh(R) (F, Y ) − → HomCh(R) (F, Z) − → 0, is also exact. Pure exact sequences of complexes can be characterized in terms of the functors Hom and ⊗ from Section 2.2. In particular, E is pure exact if and only if Hom(F, E) is a short exact sequence of complexes for each finitely presented complex F if and only if F ⊗E is a short exact sequence of complexes for each finitely presented complex F (or just any complex F ). See [GR99, Theorem 5.1.3]. This leads us to the following notion of a pure exact complex of complexes. Definition 5.4. Let X ∈ Ch(Ch(R)). We say that X is a pure exact complex of complexes if the following equivalent conditions are satisfied: (1) HomCh(R) (F, X) = Hom(S 0 (F ), X) is an exact complex of abelian groups for each finitely presented chain complex F . (2) Hom(F, X) is an exact complex (of complexes of abelian groups) for each finitely presented chain complex F . (3) F ⊗X is an exact complex (of complexes of abelian groups) for each finitely presented chain complex F . (4) F ⊗X is an exact complex (of complexes of abelian groups) for each chain complex F . Lemma 5.5. Suppose X is a bounded complex of finitely presented complexes and Y is a pure exact complex of complexes. Then every chain map f : X − → Y is chain homotopic to 0. Proof. Note that it is equivalent to show Hom(X, Y) is exact, where Hom is the usual hom-complex of Section 2.1 (but applied to the case of A = Ch(R) instead of the more typical application of A = R-Mod). Let m be the largest degree i for which Xi is nonzero, and let A be the subcomplex of X with Ai = Xi for i < m and Am = 0. We have a degreewise split short exact sequence 0− →A− →X− → S m (Xm ) − →0 which for the given Y induces the short exact sequence 0− → Hom(S m (Xm ), Y) − → Hom(X, Y) − → Hom(A, Y) − → 0. 14 JAMES GILLESPIE Now Hom(S m (Xm ), Y) is exact by condition (1) of Definition 5.4. The long exact sequence in homology now gives us the result by induction on m. Let (C, W) denote the cotorsion pair of Theorem 5.3. The result below tells us that pure exact complexes of complexes are in W. Theorem 5.6. Let R be any ring and let C be a complex of projective complexes. If Y is a pure exact complex of complexes, then Hom(C, Y) is exact, or, equivalently Y ∈ W = C ⊥. Proof. In view of Theorem 5.3, it suffices to assume that C is a bounded above complex of finitely generated free complexes and to show that any chain map f : C − →Y is chain homotopic to 0. We construct a chain homotopy Dn : Cn − → Yn+1 with dDn + Dn−1 d = fn by downwards induction on n. Since C is bounded above, we can take Dn = 0 for large n to begin the induction. So we suppose that Di has been defined for i ≥ n and that dDn+1 + Dn d = fn+1 . e n still having dDn+1 + D e n d = fn+1 , The idea is to replace Dn with a new map D e and to also find a map Dn−1 such that dDn + Dn−1 d = fn . First note that (fn − dDn )d = d(fn+1 − Dn d) = d2 Dn+1 = 0, π → so there is an induced map gn : Cn /Bn C − → Yn , such that the composite Cn − gn Cn /Bn C −→ Yn equals fn − dDn . Now consider the bounded complex X of finitely presented complexes with Xn = Cn /Bn C, Xn−1 = Cn−1 , Xn−2 = Cn−1 /Bn−1 C, and Xi = 0 for all other i. There is a chain map g : X − → Y that is gn in degree n, fn−1 in degree n − 1, and fn−2 d¯ in degree n − 2. By Lemma 5.5, this chain → Yn+1 map must be chain homotopic to 0. This gives us maps Dn′ : Cn /Bn C − and Dn−1 : Cn−1 − → Yn such that dDn′ + Dn−1 d¯ = gn . Upon composing with π en = Cn − → Cn /Bn C this becomes dDn′ π + Dn−1 d = fn − dDn . Now setting D ′ e Dn +Dn π, we still have the required relation dDn+1 + Dn d = fn+1 because dDn+1 + (Dn + Dn′ π)d = dDn+1 + Dn d + 0 = fn+1 . Moreover, we have the other promised e n + Dn−1 d = fn . Indeed dD e n + Dn−1 d = d(Dn + D′ π) + Dn−1 d = relation: dD n ′ dDn + (dDn π + Dn−1 d) = dDn + (fn − dDn ) = fn . Recall that for a complex X we have its character dual X + = HomZ (X, Q/Z). See Section 2.4. Lemma 5.7. Let C be a complex of complexes, and X a complex of right R-modules. Then X⊗C is exact if and only if Hom(C, X + ) is exact. Proof. X⊗C is exact if and only if (X⊗C)+ is exact. But also using parts (1) and (6) b) of [GR99, Proposition 4.2.1] we get (X⊗C)+ = HomZ (X⊗C, Q/Z) = Hom(X⊗C, D1 (Q/Z)) ∼ = Hom(C, X + ). = Hom(C, Hom(X, D1 (Q/Z))) ∼ We need just one more lemma before proving the main theorem. Lemma 5.8. Let Y ∈ Ch(R) be a complex, and C ∈ Ch(Ch(R)) be a complex of complexes . Then Hom(C, Y ) is exact if and only if Hom(C, S n (Y )) is exact for each n. DING MODULES AND COMPLEXES 15 Proof. Pondering definitions, we see Hom(C, Y ) is the cochain complex, of complexes of abelian groups, whose degree zero entry is the complex of abelian groups ··· ← − HomCh(R) (Σn−1 C0 , Y ) ← − HomCh(R) (Σn C0 , Y ) ← − ··· (where HomCh(R) (Σn C0 , Y ) is the degree n component of this complex). So to say that the overall complex is exact means that the cochain complex of abelian groups, HomCh(R) (Σn C, Y ), is exact for each n. On the other hand, by again following definitions, Hom(C, S n (Y )) is isomorphic to the cochain complex, HomCh(R) (Σ−n C, Y ), of abelian groups. So exactness of Hom(C, S n (Y )) for all n is equivalent to exactness of Hom(C, Y ). Now if C is a collection of chain complexes of right R-modules, and D is a collection of chain complexes of left R-modules, we say that (C, D) is a duality pair if X ∈ C if and only if X + ∈ D, and Y ∈ D if and only if Y + ∈ C. It is immediate from Corollary 2.4 that the absolutely clean and level complexes give rise to two duality pairs. One where C is the class of all absolutely clean complexes of right R-modules, and another where C is the class of all level complexes of right R-modules. Theorem 5.9. Suppose (C, D) is a duality pair in Ch(R) such that D is closed under pure quotients. Let C be a complex of projective complexes. Then X⊗C is exact for all X ∈ C if and only if Hom(C, Y ) is exact for all Y ∈ D. In particular, (1) A⊗C is exact for all absolutely clean complexes A if and only if Hom(C, L) is exact for all level complexes L. (2) L⊗C is exact for all level complexes L if and only if Hom(C, A) is exact for all absolutely clean complexes A. Proof. In view of Lemma 5.7, if Hom(C, Y ) is exact for all Y ∈ D, then X⊗C is exact for all X ∈ C. Conversely, suppose X⊗C is exact for all X ∈ C. Then if Y ∈ D, Y + ⊗C is exact, and so Lemma 5.7 tells us that Hom(C, Y ++ ) is exact. We conclude that Hom(C, D) is exact for all D ∈ D++ , and we note that D++ ⊆ D since (C, D) is a duality pair. Now, for any complex Y , the natural map Y − → Y ++ is a pure monomorphism of complexes [GR99, Proposition 5.1.4.(4)]. So if Y ∈ D, the quotient Y ++ /Y is also in D since D is closed under pure quotients. We can therefore create a resolution of Y ∈ D by elements of D++ . That is, we can find a pure exact resolution of Y by complexes of complexes in D++ . This gives us a short exact sequence 0− → S n (Y ) − →D− →P− →0 in which P is pure exact and D is a bounded above complex with entries in D++ . Theorem 5.6 tells us that Hom(C, P) is exact. Hence Hom(C, D) ∼ = Hom(C, S n (Y )). So by Lemma 5.8, it only remains to show that Hom(C, D) is exact whenever D is bounded above with components in D++ . But showing Hom(C, X) is exact for all complexes of projective complexes C, is equivalent to showing Ext1 (C, X) = 0 for all such C. By Lemma 5.8 and the first paragraph we conclude Ext1 (C, S n (D)) = 0 whenever D ∈ D++ . Now any bounded above complex D with entries in D++ can be seen to be an inverse transfinite extension of spheres S n (D) for some D ∈ D++ . By the dual of the Eklof Lemma, 16 JAMES GILLESPIE we know that Ext1 (C, −) is closed under inverse transfinite extensions [EIJ07, Theorem 1.6]. So Ext1 (C, D) = 0, whence Hom(C, D) is exact for all complexes of projective complexes C. This completes the proof. As noted above, the absolutely clean complexes and level complexes give rise to two duality pairs. Moreover, each class is closed under pure quotients by [BG15, Propositions 2.7/4.7]. 6. Applications to Ding projective modules and complexes We now wish to prove the duals to Corollaries 4.3 and 4.4. They are Theorem 6.2 and Corollary 6.3 below. Definition 6.1. We call an R-module M Ding projective if there exists an exact complex of projectives · · · → P1 → P0 → P 0 → P 1 → · · · with M = ker (P 0 → P 1 ) and which remains exact after applying HomR (−, F ) for any flat module F . In the same way, we call a chain complex X Ding projective if there exists an exact complex of projective complexes · · · → P1 → P0 → P 0 → P 1 → · · · with X = ker (P 0 → P 1 ) and which remains exact after applying HomCh(R) (−, F ) for any flat chain complex F . Recall that a chain complex P is projective (resp. flat) in Ch(R) if and only if it is exact and each cycle Zn P is a projective (resp. flat) module. Theorem 6.2. Let R be a Ding-Chen ring. Then a module M is Ding projective if and only if M = Z0 P for some exact complex P of projective R-modules. In the same way, a chain complex X is Ding projective if and only if X = Z0 P for some exact complex P of projective complexes. Proof. First, we look at the R-module case. Let P be any exact complex of projectives. We need to show that HomR (P, F ) remains exact for any flat (left) module F . But by [BGH13, Theorem 6.7], and since R is coherent, it is enough to show A ⊗R P is exact for all absolutely pure (right) modules A. Note that clearly F ⊗R P is exact for any flat (right) module F . It follows that M ⊗R P is exact for all M of finite flat dimension. But since R is a Ding-Chen ring, any absolutely pure module A does have finite flat dimension. So we are done. We note that a similar but alternate proof could be given, using Theorem 3.6 instead. Next, we see that the proof holds for chain complexes due to the work of Section 5. So now let P be any exact complex of projective complexes. We need to show that HomCh(R) (P, F ) remains exact for any flat (left) chain complex F . However, looking at the definition of Hom it is clear that this is equivalent to the statement that Hom(P, F ) remains exact for any flat complex F . Now by Theorem 5.9, and since R is coherent, this is equivalent to the statement that A⊗P is exact for all absolutely pure (right) complexes A. We note that F ⊗P is exact for any flat (right) complex F by [GR99, Proposition 5.1.2]. It follows that X⊗P is exact for all complexes X of finite flat dimension. But since R is a Ding-Chen ring, any absolutely pure complex A is an exact complex with an upper bound on the flat dimensions of the Zn A. It follows that A itself is a complex of finite flat dimension. So we are done. DING MODULES AND COMPLEXES 17 Corollary 6.3. Let R be a Ding-Chen ring. Then a complex X is Ding projective if and only if each Xn is a Ding projective R-module. Proof. The “only if” part is easy to show. For the converse, Theorem 6.2 assures us that we only need to show X equals the zero cycles of some exact complex d2 d3 P2 −→ of projective complexes. Certainly we can find an exact complex · · · −→ d1 ǫ →X − → 0 with each Pi a projective complex. So it is left to extend this P1 −→ P0 − complex to the right. First note that there is an obvious (degreewise split) short exact sequence (1,d) M n+1 −d+1 0 → X −−−→ D (Xn ) −−−−→ ΣX → 0. n∈Z Now each Xn is Ding projective. So we certainly can find a short exact sequence βn αn Qn −−→ Yn → 0 where Qn is projective and Yn is also Ding projective. 0 → Xn −−→ This gives us another short exact sequence L L n+1 n+1 M (αn ) M n+1 (βn ) M n+1 n∈Z D n∈Z D Dn+1 (Xn ) −−−−−−−−−−−→ D (Qn ) −−−−−−−−−−→ D (Yn ). n∈Z L n∈Z n∈Z (Qn ) is a projective complex and we will denote it by P 0 . Notice that n∈Z D Furthermore, let η : X → P 0 be the composite L n+1 (αn ) M n+1 (1,d) M n+1 n∈Z D X −−−→ D (Xn ) −−−−−−−−−−−→ D (Qn ). n+1 n∈Z n∈Z Then η is an monomorphism since it is the composite of two monomorphisms. Moreover, setting C 0 = cok η, it follows from snake lemma that C 0 sits in the L the n+1 0 (Yn ) → 0. In particular, C 0 short exact sequence 0 → ΣX − →C − → n∈Z D L n+1 0 is an extension of n∈Z D L(Yn ) and ΣX, and so C must be Ding projective n+1 in each degree since both of n∈Z D (Yn ) and ΣX are such. Since C 0 has the same properties as X, we may continue inductively to obtain the desired resolution η d0 d1 d2 0− →X− → P 0 −→ P 1 −→ P 2 −→ · · · d d ǫ 1 2 → X → 0 by P0 − P1 −→ Finally, we paste the resolution together with · · · − → P2 −→ setting d0 = ηǫ and we are done. 7. Ding flat modules and complexes As pointed out in [Gil10], there is also a natural notion of a Ding flat module, but it turns out to be equivalent to the notion of a Gorenstein flat module. We recall the definition. Definition 7.1. We call an R-module M Ding flat if there exists an exact complex of flat modules · · · → F1 → F0 → F 0 → F 1 → · · · with M = ker (F 0 → F 1 ) and which remains exact after applying A ⊗R − for any absolutely pure (right) R-module A. In the same way, we call a chain complex X Ding flat if there exists an exact complex of flat complexes · · · → F1 → F0 → F 0 → F 1 → · · · 18 JAMES GILLESPIE with X = ker (F 0 → F 1 ) and which remains exact after applying A⊗− for any absolutely pure complex A. Proposition 7.2. Let R be any ring. A Ding flat module is nothing more than a Gorenstein flat module. Similarly, a chain complex X is Ding flat if and only if it is Gorenstein flat. A proof of the above proposition, for modules, goes back to [DM08, Lemma 2.8]. We provide a new proof which is very quick and easy. Proof. We give the proof for complexes but the same proof works for modules. Note that Ding flat complexes are clearly Gorenstein flat since injective complexes are absolutely pure. Conversely, suppose that X is a Gorenstein flat complex. By definition, this means there exists an exact complex F = · · · → F1 → F0 → F 0 → F 1 → · · · of flat complexes with X = ker (F 0 → F 1 ) which remains exact after applying I ⊗− for any injective complex I. We will show that this complex does in fact remain exact after applying A⊗− for any absolutely pure (right) chain complex A. So let such an A be given and note that it is equivalent to show that the map of complexes A⊗Zn F − → A⊗Fn is a monomorphism for each n. Now let A ֒→ I be an embedding into an injective complex I and note that this must be a pure monomorphism since A is an absolutely pure complex. For each n we have the commutative diagram: A⊗Zn F −−−−→ I⊗Zn F     y y A⊗Fn −−−−→ I⊗Fn The two horizontal arrows are monomorphisms since A ֒→ I is pure. The right vertical arrow is also a monomorphism since I⊗F is exact. It follows that the left vertical arrow must also be a monomorphism. Proposition 7.3. Let R be a Ding-Chen ring. Then a module M is Ding flat if and only if M = Z0 F for some exact complex F of flat R-modules. In the same way, a chain complex X is Ding flat if and only if X = Zn F for some exact complex F of flat complexes. Proof. The proof is easy and very similar to the last few sentences of the proof of Theorem 6.2. Briefly, say F is an exact complex of flat complexes; we wish to show that A⊗ F is exact for any absolutely pure complex A. Certainly F ⊗ F is exact for any flat complex F , and, since R is a Ding-Chen ring, any absolutely pure complex A has finite flat dimension. We argue that A⊗F is exact for any such A. Corollary 7.4. Let R be a Ding-Chen ring. Then a complex X is Ding flat if and only if each Xn is a Ding flat R-module. Proof. Using the above proposition, the proof of Corollary 6.3 carries over. Just replace the word projective with the word flat throughout the proof. DING MODULES AND COMPLEXES 19 References [AR94] J. Adámek and J. Rosický, Locally presentable and accessible categories, Number 189 in London Mathematical Society Lecture Note Series, Cambridge University Press, 1994. [BG15] Daniel Bravo and James Gillespie, Absolutely clean, level, and Gorenstein AC-injective complexes, preprint; arXiv:1408.7089 [BGH13] Daniel Bravo, James Gillespie and Mark Hovey, The stable module category of a general ring, submitted. [DC93] N. Ding and J. Chen, The flat dimensions of injective modules, Manuscripta Math. vol. 78, no. 2, 1993, pp. 165-177. [DC96] N. Ding and J. Chen, Coherent rings with finite self-FP-injective dimension, Comm. Algebra, vol. 24, no. 9, 1996, pp. 2963-2980. [DM05] N. Ding and L. Mao, Reletive FP-projective modules, Comm. Algebra, vol. 33, no. 5, 2005, pp. 1587-1602. [DM07] N. Ding and L. Mao, Envelopes and covers by modules of finite FP-injective and flat dimensions, Comm. Algebra, vol. 35, no. 3, 2007, pp. 833–849. [DM08] N. Ding and L. Mao, Gorenstein FP-injective and Gorenstein flat modules, J. Algebra Appl. 7 (2008), no. 4, 491–506. [DLM09] N. Ding and Y. Li and L. Mao, Strongly Gorenstein flat modules, J. Aust. Math. Soc. vol. 86, no. 3, 2009, pp. 323-338. [EGR97] E. Enochs and J.R. Garcı́a-Rozas, Tensor products of chain complexes, Math J. Okayama Univ. 39, 1997, pp. 19–42. [EE05] E. Enochs, S. Estrada, and A. Iacob, Gorenstein projective and flat complexes over noetherian rings, Mathematische Nachrichten vol. 285, no. 7, 2012, pp. 834–851. [EJ00] Edgar E. Enochs and Overtoun M. G. Jenda, Relative homological algebra, de Gruyter Expositions in Mathematics, vol. 30, Walter de Gruyter & Co., Berlin, 2000. MR 2001h:16013 [EIJ07] E. Enochs, A. Iacob, and O.M.G. Jenda, Closure under Transfinite Extensions, Illinois Journal of Mathematics vol. 51, no. 2, 2007, pp.561–569. [GR99] J. R. Garcı́a-Rozas, Covers and envelopes in the category of complexes of modules, Research Notes in Mathematics no. 407, Chapman & Hall/CRC, Boca Raton, FL, 1999. [Gil04] James Gillespie, The flat model structure on Ch(R), Trans. Amer. Math. Soc. vol. 356, no. 8, 2004, pp. 3369–3390. [Gil07] James Gillespie, Kaplansky classes and derived categories, Math. Zeit. vol. 257, no. 4, 2007, pp.811–843. [Gil10] James Gillespie, Model structures on modules over Ding-Chen rings, Homology, Homotopy Appl. 12 (2010), no. 1, 61–73. MR 2607410 (2011e:18027) [Gil12] James Gillespie, Gorenstein complexes and recollements from cotorsion pairs, preprint; arXiv 1210.0196, 2012. [Gil15] James Gillespie, How to construct a Hovey triple from two cotorsion pairs, Fundamenta Mathematicae, Vol. 230, no.3, 2015, pp. 281–289. [GT06] Rüdiger Göbel and Jan Trlifaj, Approximations and Endomorphism Algebras of Modules, de Gruyter Expositions in Mathematics, vol. 41, Walter de Gruyter & Co., Berlin, 2006. [Hov02] Mark Hovey, Cotorsion pairs, model category structures, and representation theory, Mathematische Zeitschrift, vol. 241, 2002, pp.553–592. [Iwa79] Y. Iwanaga, On rings with finite self-injective dimension, Comm. Algebra, vol. 7, no. 4, 1979, pp. 393-414. [Iwa80] Y. Iwanaga, On rings with finite self-injective dimension II, Tsukuba J. Math. vol. 4, no. 1, 1980, pp. 107-113. [Lam99] T. Y. Lam, Lectures on modules and rings, Graduate Texts in Mathematics, vol. 189, Springer-Verlag, New York, 1999. MR MR1653294 (99i:16001) [Sto14] Jan Šťovı́ček,, On purity and applications to coderived and singularity categories, arXiv:1412.1615. [Ste70] Bo Stenström, Coherent rings and F P -injective modules, J. London Math. Soc. (2) 2 (1970), 323–329. MR MR0258888 (41 #3533) [Sten75] Bo Stenström, Rings of Quotients, Die Grundlehren der mathematischen Wissenschaften in Einzeldarstellungen Band 217, Springer-Verlag, New York 1975. 20 JAMES GILLESPIE [Wei94] Charles A. Weibel, An Introduction to Homological Algebra, Cambridge Studies in Advanced Mathematics vol. 38, Cambridge University Press, 1994. [YLL11] Gang Yang, Zhongkui Liu, Li Liang, Ding Projective and Ding Injective Modules, Algebra Colloquium 20 (2013), no. 04, 601–612. [YLL13] Gang Yang, Zhongkui Liu, and Li Liang, Model structures on categories of complexes over Ding-Chen rings, Communications in Algebra, vol. 41, 2013, pp. 50–69. [YL11] Gang Yang and Zhongkui Liu, Cotorsion pairs and model structures on Ch(R), Proc. Edinb. Math. Soc. (2) vol. 54, no. 3, 2011, pp. 783–797. Ramapo College of New Jersey, School of Theoretical and Applied Science, 505 Ramapo Valley Road, Mahwah, NJ 07430 E-mail address, Jim Gillespie: jgillesp@ramapo.edu URL: http://pages.ramapo.edu/~jgillesp/	0math.AC
Beyond Highway Dimension: Small Distance Labels Using Tree Skeletons∗ Adrian Kosowski and Laurent Viennot arXiv:1609.00512v2 [cs.DS] 12 Dec 2016 Inria Paris and IRIF, Université Paris Diderot, France Abstract The goal of a hub-based distance labeling scheme for a network G = (V, E) is to assign a small subset S(u) ⊆ V to each node u ∈ V , in such a way that for any pair of nodes u, v, the intersection of hub sets S(u) ∩ S(v) contains a node on the shortest uv-path. The existence of small hub sets, and consequently efficient shortest path processing algorithms, for road networks is an empirical observation. A theoretical explanation for this phenomenon was proposed by Abraham et al. (SODA 2010) through a network parameter they called highway dimension, which captures the size of a hitting set for a collection of shortest paths of length at least r intersecting a given ball of radius 2r. In this work, we revisit this explanation, introducing a more tractable (and directly comparable) parameter based solely on the structure of shortest-path spanning trees, which we call skeleton dimension. We show that skeleton dimension admits an intuitive definition for both directed and undirected graphs, provides a way of computing labels more efficiently than by using highway dimension, and leads to comparable or stronger theoretical bounds on hub set size. Key Words: Distance Labeling, Highway Dimension, Shortest Path Tree, Skeleton Dimension 1 Introduction The task of efficiently processing shortest path queries to a graph has been studied in a plethora of settings. One interesting observation is that for many real-world graphs of small degree in a geometric or geographical setting, such as road networks, it is possible to design compact data structures and schemes for efficiently answering shortest path queries. The general principle of operation of this approach consists in detecting and storing subsets of so-called transit nodes, which appear on shortest paths between many node pairs. In an attempt to explain the efficiency of variants of the transit node routing (TNR) algorithm [13,14], Abraham et al. [7] introduced the concept of highway dimension h. This parameter captures the intuition that when a map is partitioned into regions, all significantly long shortest paths out of each region can be hit by a small number of transit node vertices. The value of h is presumed to be a small constant e.g. for road networks. However, the definition of highway dimension relies on the notion of a hitting set of shortest path sets within network neighborhoods, and hence, e.g., exact computation of the parameter is known to be NP-hard even for unweighted networks [20]. This motivates us to look at other measures, which are both more locally defined and computationally tractable, while capturing essentially the same (or more) characteristics of the network’s amenability to shortest path queries. ∗ Supported by Inria project GANG, ANR project DESCARTES, and NCN grant 2015/17/B/ST6/01897. 1 Looking more precisely at the TNR algorithm, one observes that it is built around the idea that for every source node, the set of transit nodes which are the first to be encountered when going a long way from the source, is small. This is a weaker assumption than the existence of a small hitting set for the set of shortest paths in a given network neighborhood, since different source nodes could use different transit nodes resulting in an overall large number of transit nodes around a given region. This source-centered approach leads us to the definition of skeleton dimension k , to which we devote the remainder of this paper. Informally, the skeleton dimension is the maximum, taken over all nodes u of the graph and all radii r > 0, of the number of distinct nodes at distance r from u in the set of all shortest paths originating at u and having length at least 3r/2.1 In transit node parlance, it states that the paths from u that extend by r/2 at least outside the disk of radius r pass through at most k transit nodes at the disk border. This property ensures that each shortest-path spanning tree is built around a core skeleton with at most k branches at a given distance range while the rest of the branches are relatively short. Bounding tree skeletons turns out to encompass a larger class of constant-degree graphs than the shortest path cover approach used in the definition of highway dimension, while still ensuring the existence of efficient labeling schemes. Motivated by applications in distributed algorithms and distributed data representation, we will display the link between small skeleton dimension of a graph and efficient processing of shortest path queries using the framework of distance labeling. Distance labeling schemes, popularized by Gavoille et al. [21], are among the most fundamental distributed data structures for graph data. Within distance labeling, we work with the most basic framework of transit-node based schemes, namely so-called hub labelings, cf. [5] (this framework was first described in [18] under the name of 2-hop covers, and is also referred to as landmark labelings [8]). In this setting, each node u ∈ U stores the set of its distances to some subset S(u) ⊆ V of other nodes of the graph. Then, the computed distance value d0 (u, v) for a queried pair of nodes u, v ∈ V is returned as: d0 (u, v) := min d(u, w) + d(w, v), w∈S(u)∩S(v) (1) where d denotes the shortest path distance function between a pair of nodes. The computed distance between all pairs of nodes u and v is exact if set S(u) ∩ S(v) contains at least one node on some shortest u − v path. This property of the family of sets (S(u) : u ∈ V ) is known as shortest path cover. The hub-based method of distance computation is in practice effective for two reasons. First of all, for transportation-type networks it is possible to show bounds on the sizes of sets S, which follow from the network structure. Notably, considering networks of bounded highway dimension h, Abraham et al. [7] show that an appropriate cover of all shortest paths in the graph can be achieved using sets S of size e e O(h), where the O-notation conceals logarithmic factors in the studied graph parameters. Moreover, the order in which elements of sets S(u) and S(v) is browsed when performing the minimum operation is relevant, and in some schemes, the operation can be interrupted once it is certain that the minimum has been found, before probing all elements of the set. This is the principle of numerous heuristics for the exact shortest-path problem, such as contraction hierarchies and algorithms with arc flags [15, 25]. 1.1 Results and Organization of the Paper In Section 2, we formally define skeleton dimension k , and show that in the so-called continuous representation of the graph, the skeleton dimension is at most highway dimension, i.e. it satisfies the bound k ≤ h. In all cases, k = O(h) for graphs of bounded maximum degree. On the other hand, we show that 1 One may also define skeleton dimension with a different choice of constants, considering the set of shortest paths having length at least αr, where α ∈ (1, 2) is an absolute constant. The choice of α = 3/2 is subsequently necessary only for establishing relations with highway dimension. 2 skeleton dimension provides a better explanation for small hub set size in Manhattan-type networks than highway dimension. In particular, we provide a natural example of a weighted grid with k = O(log n) √ and h = Ω( n). In Section 3, we show how to construct efficient hub labelings for networks of small skeleton dimension. The hub set sizes we obtain for a graph of weighted diameter D are bounded by O(k log D) on average and O(k log log k log D) in the worst case (cf. Corollaries 1 and 3, respectively), as compared to previous best bounds of O(h log h log D) for labels computable in polynomial time based on highway dimension. Our labeling technique, based on picking hubs through a random selection process on a subtree of the shortest-path tree, allows each node to compute its hub set independently in almost-linear time, and appears to be of independent interest. In particular, as an extension of our technique, we provide in Section 4 improved bounds on label size (in general unweighted graphs) for the so-called δ-preserving distance labeling problem, in which the considered distance queries are restricted to nodes at distance at least δ from each other. The hub sets constructed using the hub-based method have average size O(n/δ). e √n), and bounded by Their worst-case size is also bounded by O(n/δ) up to some threshold δ = O( O(log δ + (n/δ) log log δ) in general (Theorem 2). This improves upon previous δ-preserving schemes, including the previously best result from [10], where hub sets of worst-case size O((n/δ) log δ) are constructed by a more direct application of the probabilistic method to sets of randomly sampled vertices. Finally, in Sections 5, 6, and 7 we provide some concluding remarks on the computability of the proposed parameter of skeleton dimension, as well as its possible generalizations and applications. 1.2 Other Related Work Distance Labelings. The distance labeling problem in undirected graphs was first investigated by Graham and Pollak [23], who provided the first labeling scheme with labels of size O(n). The decoding time for labels of size O(n) was subsequently improved to O(log log n) by Gavoille et al. [21] and to O(log∗ n) by Weimann and Peleg [29]. Finally, Alstrup et al. [11] present a scheme for general graphs with decoding in O(1) time using labels of size log2 3 n + o(n) bits. This matches up to low order terms the space of the currently best know distance oracle with O(1) time and log2 3 n2 + o(n2 ) total space in a centralized memory model, due to Nitto and Venturini [27]. For specific classes of graphs, Gavoille et √ al. [21] described a O( n log n) distance labeling for planar graphs, together with Ω(n1/3 ) lower bound √ for the same class of graphs. Additionally, O(log2 n) upper bound for trees and Ω( n) lower bound for sparse graphs were given. Distance Labeling with Hub Sets. For a given graph G, the computational task of minimizing the sizes of hub sets (S(u) : u ∈ V ) for exact distance decoding is relatively well understood. A O(log n)approximation algorithm for minimizing the average size of a hub set having the sought shortest path cover property was presented in Cohen et al. [18], whereas a O(log n)-approximation for minimizing the largest hub set at a node was given more recently in Babenko et al. [12]. Rather surprisingly, the structural question of obtaining bounds on the size of such hub sets for specific graph classes, such as graphs of bounded degree or unweighted planar graphs, is wide open. δ-preserving Labeling. The notion of δ-preserving distance labeling, first introduced by Bollobás et al. [16], describes a labeling scheme correctly encoding every distance that is at least δ. [16] presents such a δ-preserving scheme of size O( nδ log2 n). This was recently improved by Alstrup et al. [10] to a δ-preserving scheme of size O( nδ log2 δ). Together with an observation that all distances smaller than n δ can be stored directly, this results in a labeling scheme of size O( nx log2 x), where x = loglogm+n . For n sparse graphs, this is o(n). 3 Road networks. Highway dimension h guarantees the existence of distance labels of size O(h log D) where D is the weighted diameter of the graph [7]. However, when restricting to polynomial time algorithms, such labels can only be approximated within a log n factor using shortest path cover algorithms [7] or a log h factor with a more involved procedure based on VC-dimension [3]. In any case, this requires an all-pair shortest path computation. For large networks, labels can be practically computed when classical heuristics such as contraction hierarchies (CH) can be performed [4, 6, 7]. Low highway dimension guarantees that there exists an elimination ordering for CH such that the graph produced has bounded size [7]. However, it does not ensure running time faster than all pair shortest path computation. Besides highway dimension, skeleton dimension is also related to the notion of reach introduced in [24] and also used in the RE algorithm [22]. The reach of a node v on a path P is the minimum distance to an extremity of P , and the reach of v is its maximum reach over all shortest paths P containing v. Efficient algorithms are obtained by pruning nodes with small reach during Dijkstra search. Similarly, we obtain the skeleton of a tree by pruning nodes whose reach (in the tree) is less than half of their distance to the root. 1.3 Notation and Parameters We consider a connected undirected graph G and a non-negative length function ` : E(G) → R+ . Let n denote the number of nodes in V (G). We let `(P ) denote the length of a path P under the given length function. Given two nodes u and v, we assume that there is a unique shortest path Puv between them. This common assumption can be made without loss of generality, as one can perturb the input to ensure uniqueness. Given two nodes u and v, their distance is dG (u, v) = `(Puv ). Let D = maxu,v dG (u, v) denote the diameter of G. For u ∈ V (G) and r > 0, the ball BG (u, r) of radius r centered at u is the set of nodes v with dG (u, v) ≤ r. In this paper, we assume that ` is non-negative and integral. The notions presented here easily extend to non-negative real lengths, but we use integer lengths for a cleaner exposition of algorithms and theorems. We also recall two structural parameters, which have application to networks in a geometric setting or low-dimensional topological embedding: highway dimension and doubling dimension. For r > 0, let PG (r) denote the collection of all shortest paths P with 2r < `(P ) ≤ r in G. For u ∈ V (G), we consider the collection PG (u, r) = {P ∈ Pr (G) \| P ∩ BG (u, r) 6= ∅} of shortest paths around u. A hitting set for PG (u, r) is a set H of nodes such that any path in PG (u, r) contains a node in H. In [3], the highway dimension of G is defined as the smallest h such that PG (u, r) has a hitting set of size at most h for all u, r. (This definition is slightly less restrictive than that of [7] while allowing to prove similar results with improved bounds.) The notion of highway dimension is related to that of doubling dimension. Recall that a graph is h-doubling if any ball can be covered by at most h balls of half the radius. That is, for all u, r, there exists H with \|H\| ≤ h such that BG (u, r) ⊆ ∪v∈H BG (v, 2r ). It is shown in [7] that if the geometric realization of a graph G has highway dimension h, then G is h-doubling. Informally, the geometric e can be seen as the “continuous” graph where each edge is seen as infinitely many vertices realization G of degree two with infinitely small edges, such that for any uv ∈ E(G) and t ∈ [0, 1], there is a node in e at distance t`(u, v) from u on edge uv. (The proof in [7] consists in proving that any node in BG (u, r) G e and holds also for the highway dimension is at distance at most 2r from any hitting set of PGe (u, r) in G definition of [3].) 2 A Presentation of Skeleton Dimension We start by providing a standalone definition of skeleton dimension based on size of cuts in shortest path trees, and then show its relation to the previously considered parameters of highway and doubling dimension. 4 2.1 Definition of the Parameter Tree skeleton. Given a tree T rooted at node u with length function ` : E(T ) → R+ , we treat it as directed from root to leaves and consider the geometric realization Te of this directed graph. We define the reach of v ∈ V (Te) as Reach Te (v) := maxx∈V (Te) dTe (v, x). We then define the skeleton T ∗ of T as the subtree of Te induced by nodes with reach at least half their distance from the root. More precisely, T ∗ is the subtree of Te induced by {v ∈ V (Te) \| Reach Te (v) ≥ 12 dTe (u, v)}. Width of a tree. The width of a tree T with root u is defined as the maximum number of nodes (points) in Te at a given distance from its root. More precisely, the width of T is Width(T ) = maxr>0 \| Cut r (Te)\| where Cut r (Te) is the set of nodes v ∈ V (Te) with dTe (u, v) = r. Skeleton dimension. The skeleton dimension k of a graph G is defined as the maximum width of the skeleton of a shortest path tree, that is k = maxu∈V (G) Width(Tu∗ ), where Tu denotes the shortest path tree of u obtained as the union of shortest paths from u to all v ∈ V (G). We remark that, under the assumption of scale-invariance of the graph, different cuts of the tree skeleton have similar width, and the definition of the skeleton dimension is a meaningful measure of the structure of the tree. A smoothed (integrated) variant of skeleton dimension is also discussed further on, cf. Eq. (2). 2.2 Skeleton Dimension is at most (Geometric) Highway Dimension e of a graph G has highway dimension h̃, then G has skeleton Claim 1. If the geometric realization G dimension k ≤ h̃. Proof. Consider a node u and the skeleton Tu∗ of its shortest path tree Tu . For r > 0, consider the cut Cut r (Tu∗ ). For ε > 0 sufficiently small, Cut r (Tu∗ ) and Cut r−ε (Tu∗ ) have same size. Now, for v ∈ Cut r−ε (Tu∗ ), consider a node x in Tu such that dTu (v, x) = Reach Tu (v). The shortest path Pvx intersects BG (u, r) and has length `(Pvx ) = Reach Tu (v) ≥ r/2+ε > r/2. Pvw is thus in PGe (u, r). For each node in Cut r−ε (Tu∗ ), we get a similar path in PGe (u, r). All these paths are pairwise node-disjoint as they belong to disjoint sub-branches of Tu . Their number is thus upper-bounded by the size of any hitting set of PGe (u, r). We then get \|Cut r (Tu∗ )\| ≤ h̃ for all u, r and the skeleton dimension of G is at most h̃. Note that a (discrete) graph G has highway dimension h ≤ h̃, where h̃ is the highway dimension of e In road networks it is expected that the continuous and the discrete versions its geometric realization G. of highway dimension coincide almost exactly, in particular due to the constant maximum degree and bounded length of edges in these graphs. In a more general setting, one can easily show h̃ ≤ (∆ + 1)h where ∆ is the maximum degree of G, with a star being a worst-case example. (Indeed, a hitting set H of PG (u, r) may miss some shortest path P ∈ PGe (u, r). Making P longer to have extremities in V (G) transforms it into a path of PG (u, r) that is hit by H. It is thus possible to hit all PGe (u, r) by adding at most one node per edge adjacent to a node in H.) We remark that the extended tech-report version [2] of [7] introduces a modified notion of highway dimension, in a way more closely related to its geometric variant, which we can denote here as h∗ . For this modified parameter, we have: k ≤ h∗ ≤ h̃ ≤ 2h∗ , where the first inequality follows from an analysis similar to the proof of Claim 1, while the latter two are shown in [2][Section 11]. 5 2.3 Low Skeleton Dimension Implies Low Doubling Dimension It is known [7] that a graph having a geometric realization with highway dimension h̃ is at most h̃doubling. However, the relation k ≤ h̃ need not be tight, and it turns out that the link between skeleton dimension and doubling dimension holds in a slightly weaker form. Proposition 1. If a graph G has skeleton dimension k , then G is (2 k +1)-doubling. Proof. We show the stronger requirement that each ball of radius 19r/9 can be covered by 2 k +1 balls of radius r. For u ∈ V (G), consider the shortest path tree Tu of u. For r0 > 0, consider the set Cr0 of the edges containing a node in Cut r0 (Tu∗ ) and let Ir0 = {w \| vw ∈ Cr0 and dG (u, w) ≥ r0 } be the (at most k ) far extremities of edges cutting distance r0 in the skeleton of Tu . Each node v at distance greater than 23 r0 from u is descendant in Tu of a node x ∈ Ir0 by skeleton definition and is thus in BG (x, r) if dG (u, v) ≤ r0 + r. Considering r0 = 2r/3, we obtain that any node v with r < dG (u, v) ≤ 35 r is at distance at most r from a node in I2r/3 . Similarly any node v with 53 r < dG (u, v) ≤ 19 9 r is at distance at most r from a node in I10r/9 . The ball BG (u, 19r/9) is thus covered by balls of radius r centered at the at most 2 k +1 nodes in {u} ∪ I2r/3 ∪ I10r/9 . 2.4 Separating Skeleton Dimension and Highway Dimension We now provide a family of graphs which exhibit an exponential gap between skeleton and highway dimensions, in a setting directly inspired by Manhattan-type road networks. The idea is to consider the usual square grid and define edge lengths, which give priority to certain transit “arteries”. In our example, paths using edges whose coordinates are multiples of high powers of 2 have slightly lower transit times. For L > 0, let GL denote the 2L × 2L grid with length function ` defined as follows. We identify a node with its coordinates (x, y) with 1 ≤ x ≤ 2L and 1 ≤ y ≤ 2L . We consider small length perturbations pxy for every horizontal edge {(x, y), (x + 1, y)} and qxy for every vertical edge {(x, y), (x, y + 1)}, and define Q = 1 + maxx,y max{pxy , qxy }. These non-negative integers will be chosen to ensure uniqueness of shortest paths. For x = 2i x0 with 0 ≤ i ≤ L and x0 odd, we define `((x, y), (x + 1, y)) = Q((D + 2)L − i) − qxy for all y where D = 2L+3 . For y = 2j y 0 with 0 ≤ j ≤ L and y 0 odd, we define `((x, y), (x, y + 1)) = Q((D + 2)L − j) − pxy for all x. A possible choice for perturbations ensuring uniqueness of shortest paths is pxy = 0 and qxy = x for all x, y as will be clear later on. √ Proposition 2. For any L > 0, grid GL has highway dimension Ω( n) and skeleton dimension O(log n), where n = 22L is the number of nodes in GL . Proof. We first prove that the shortest paths of GL are also shortest path of the 2L × 2L grid UL with unit edge lengths, that is those paths that use a minimum number of edges. Any path P with p edges has length at most pQ(D + 2)L and at least pQ((D + 2)L − L − 1) ≥ pQDL. Given two nodes u and v, let p denote the minimum number of edges of a path from u to v and let q denote the number of edges of the shortest path Puv from u to v in GL . We then have p ≤ q ≤ p(D+2)L ≤ p + 12 since p ≤ 2L+1 = D/4. DL This implies q = p. Puv is thus a shortest path of the grid UL . Note that balls must then also be almost identical: for all u, p, we have BGL (u, pQ(D + 2)L) = BUL (u, p). √ √ This implies that the highway dimension of GL is Ω( n) since the ball of radius r = Q(D +2)L n √ √ centered at (1, 1) intersects at least n horizontal shortest paths of length QDL n > r/2 at least. We define the 2i × 2j rectangle R at (x, y) with odd x and y as the set of nodes with coordinates (x0 , y 0 ) such that 2i x ≤ x0 ≤ 2i (x+1) and 2j y ≤ y 0 ≤ 2j (y +1). Its border is the set of nodes for which one inequality at least is indeed an equality. Other nodes are said to be interior. The main argument for bounding skeleton dimension is that a shortest path from u = (x0 , y 0 ) with x0 < 2i x and y 0 < 2j y cannot 6 traverse the interior of R: a shortest path passing through an inner node of R necessarily ends inside R. The reason is that such a path necessarily passes through the lower left corner at (2i x, 2j y). It is then shorter to reach a border node by following the border rather than using edges inside the rectangle. Note that two possible choices of shortest paths could be possible when going from a corner of a 2i × 2i rectangle to the corner diagonally opposed. However the choice of perturbing lengths by decreasing the length of any vertical edge in position (x, y) by qxy = x ensures that the path through the rightmost side is preferred. j k r Now consider a node u = (x, y) and a radius r with 2i QDL ≤ r < 2i+1 QDL. Set p = QDL . According to the first part of the proof, the ball BGL (u, r) is in sandwich between the balls of radius p and p + 1 in UL : BUL (u, p) ⊆ BGL (u, r) ⊆ BUL (u, p + 1). We first consider the upper right quadrant of BUL (u, p) and the border set SU R of nodes v = (x + a, y + b) with a, b ≥ 0 and a + b = p. We now bound the number of nodes v ∈ SU R such that Reach Tu (v) > 2r where Tu denote the shortest path tree of u in GL . As 2r ≥ 2i−1 QDL, such a node v cannot be interior to a 2i−2 × 2i−2 rectangle as shortest paths interior to the rectangle have length at most (2i−1 − 4)Q(D + 2)L. The number of nodes in SU R whose x coordinate is a multiple of 2i−2 is bounded by 8 as p < 2i+1 . Similarly, the number of nodes whose y coordinate is a multiple of 2i−2 is also bounded by 8. Apart from these 16 nodes, we have to consider nodes v = (x0 , y 0 ) where x0 (resp. y 0 ) is less than the smallest multiple of 2i−2 greater than x (resp. y). Such a node cannot be interior to a 2j × 2i−2 (resp. 2i−2 × 2j ) rectangle for j ≤ i − 3. For j = i − 3, we obtain at most two nodes whose x coordinate is a multiple of 2j . By repeating the argument for j = i − 4..1, we can finally bound the number of nodes v ∈ SU R with reach greater than fu r/2 by 4(i − 3) + 16 = 4(i + 1). Nodes in the upper right quadrant that are at distance r from u in T ∗ must be on edges outgoing from nodes in SU R and Cut r (Tu ) has thus size at most 8(i + 1) = O(log n). By symmetry, the bound holds for other quadrants and the skeleton dimension of GL is O(log n). We remark that there exist different lengths functions on the grid for which the skeleton dimension √ is also as large as Θ( n). This is the case, for example, for a grid with unit lengths of all edges except for edges intersecting its major diagonal, which is configured to be a fast transit artery (it suffices to set `((x, y), (x + 1, y)) = 0.5 for all x = y). We complement this result with experimental observation in real grid like networks such as encountered in Brooklyn. We computed the skeleton dimension of the New York travel-time graph proposed in the 9th DIMACS challenge [1] which turns out to be k = 73. (Average skeleton tree width is 30, but a maximum width of 73 is encountered for a skeleton tree rooted in Manhattan.) In order to estimate the highway dimension of this graph, we have implemented a heuristic for finding a large packing of paths near a given ball, that is a set of disjoint paths intersecting the ball and having length greater than half radius. We could find a packing of 172 paths in Brooklyn. This proves that the highway dimension of this graph is 172 at least (h ≥ 172). In comparison the skeleton tree of the center of the corresponding ball has width 48, and 42 branches are cut at radius distance (see Figure 1). 3 Hub Labeling using Tree Skeletons In this section, we assign shortest-path-intersecting hub sets to a set of (terminal) nodes V of the considered network. We will assume that the length function ` on edges is integer weighted. To emulate the geometric realization of the graph, we subdivide edges into sufficiently short fragments by inserting a set of additional nodes V + into the network. For convenience of subsequent analysis, we assume that an edge vw, for v, w ∈ V , of integer length `(vw) is subdivided into 12`(vw) edges of length 1/12 each. After this, all edges have the same length, and we subsequently treat the graph as unweighted. All the parameter definitions carry over directly from the geometric setting; for the sake of precision, we formally state the assumptions on the studied setting below. 7 Figure 1: An OpenStreetMap view of Brooklyn, with a packing of 172 paths (in black) intersecting a ball of radius 720 seconds (with white border) on the left, and the skeleton tree (in black) of the center of the ball on the right. We consider an unweighted graph G = (V ∪ V + , E), with a distinguished set of terminal nodes V and where all nodes from V + have degree 2. We denote n := \|V \|. We assume that every node u ∈ V is associated with a fixed (unweighted) tree Tu ⊆ G. Throughout the section, we will denote by Pu (v, w) the unique path between the pair of nodes v and w in tree Tu , and more concisely Pu (v) := Pu (u, v). Where this does not lead to confusion, we will identify a path with its edge set, and we will also use the symbol \|P \| to denote the length of path P , i.e., the number of edges belonging to P . We write du (v) := \|Pu (v)\|. We require that the collection of trees {Tu }u∈V satisfies the following property: For any pair of nodes u, v ∈ V , we have Pu (v) = Pv (u). We will also assume that for all u, v, w ∈ V , we have that \|Pu (v, w)\| is an integer multiple of 12. We remark that if the graph G was obtained by a distance-preserving subdivision of nodes of an edge-weighted graph on node set V under some distance metric `, then each tree Tu ∈ G corresponds to the shortest path tree of node u under the original distance metric, and the assumption Pu (v) = Pv (u) corresponds to the assumption of uniqueness of shortest paths under the original metric `. An edge hub labeling is an assignment of a set of edges S(u) ⊆ E to each node u ∈ V , such that the following property is fulfilled: for every pair of nodes u, v ∈ V , there exists an edge η ∈ S(u) ∩ S(v) such that η ∈ Pu (v). The set S(u) is known as the edge hub set of u. We remark that this edge-based notion of hub sets is slightly stronger than an analogous vertex-based notion: indeed, knowing that an edge η ∈ Pu (v), we also conclude that both of the endpoints of edge η belong to Pu (v). We choose to work with edge hub sets rather than node hub sets in this Section for compactness of arguments. We restate in the setting of the family of trees {Tu }u∈V the notion of the skeleton Tu∗ = Tu [Vu∗ ] ⊆ Tu as the subtree of Tu induced by node set Vu∗ = {v ∈ V (Tu ) : Reach Tu (v) ≥ 21 du (v)}. For u ∈ V , the ∗(r) width Width(Tu∗ ) of the skeleton Tu∗ may be written as Width(Tu∗ ) = maxr∈N \| Cut u \|, where: Cut ∗(r) := {v ∈ Vu∗ : du (v) = r}. u Finally, we note that the skeleton dimension of graph G may be written as k = maxu∈V Width(Tu∗ ). 8 3.1 Construction of the Hub Sets The edge hub sets S(u), u ∈ V , are obtained by the following randomized construction. Assign to each edge e ∈ E a real value ρ(e) ∈ [0, 1], uniformly and independently at random. We condition all subsequent considerations on the event that all values ρ are distinct, \|ρ(E)\| = \|E\|, which holds with probability 1. For all u, v ∈ V , we define the central subpath Pu0 (v) ⊆ Pu (v) as the subpath of Pu (v) consisting of its middle du6(v) edges; formally, Pu0 (v) := Pu (u0 , v 0 ), where u0 , v 0 ∈ Pu (v) are nodes given by: 5 7 du (u0 ) = 12 du (v) and du (v 0 ) = 12 du (v). Next, for all u, v ∈ V , v 6= u, we define the hub edge ηu (v) ∈ Pu (v) as the edge with minimum value of ρ on the central subpath between u and v: ηu (v) = arg min ρ(e). e∈Pu0 (v) Finally, for each node u ∈ V , we adopt the natural definition of edge hub set S(u) as the set of all edge hubs of node u on paths to all other nodes: S(u) := {ηu (v) : v ∈ V, v 6= u} . Proof of Correctness. Taking into account that for all u, v ∈ V , Pu (v) = Pv (u), we observe that by symmetry of the central subpath with respect to its two endpoints, we also have Pu0 (v) = Pv0 (u). It follows directly that ηu (v) = ηv (u). Hence, we have ηu (v) ∈ S(u) ∩ S(v), and also ηu (v) ∈ Pu (v), which completes the proof of correctness of the edge hub labeling. We devote the rest of this Section to bounding the size of hub sets S(u). 3.2 Bounding Average Hub Set Size In subsequent considerations we will fix a node u ∈ V , and restrict considerations to the tree Tu . We will assume that tree Tu is oriented from its root u towards its leaves, and we will call a path P ⊆ Tu a descending path in Tu if one of its endpoints is a descendant of the other in Tu . In particular, every path Pu (v) is a descending path. For an edge e ∈ Tu , we will denote by e+ and e− the two endpoints of e, with e− being the one further away from the root (du (e− ) = du (e+ ) + 1). Likewise, for a descending path P , we denote by P + and P − its two extremal vertices, closest and furthest from the root u, respectively. We also denote by du (e) := du (e− ) the distance of edge e from the root. In order to bound the expected size of the hub set S(u), we will observe that elements of S(u) necessarily belong to the skeleton Tu∗ and satisfy certain minimality constraints with respect to descending paths of sufficiently large length, contained entirely within the skeleton Tu∗ . Lemma 1. Let η ∈ S(u), for some u ∈ V . Then, the following claims hold: (1) η ∈ E(Tu∗ ). (2) Reach Tu∗ (η − ) ≥ 71 du (η). (3) There exists a descending path P ⊆ Tu∗ , such that η = arg mine∈P ρ(e) and \|P \| ≥ 72 du (η). Furthermore, the following claims hold for any edge η ∈ E satisfying Claims (1) and (3): (4) There exists a descending path P ⊆ Tu∗ , such that η = argmine∈Pρ(e), η is one of the two extremal edges of P (i.e., P + = η + or P − = η − ), and \|P \| = 71 du (η) . ∗ ∗ (5) There exists 7 a descending path Pu (x, y), for some y ∈ Vu and x ∈ Vu satisfying du (x) = 8 du (y) , such that η = arg mine∈Pu (x,y) ρ(e) and η is one of the two extremal edges of Pu (x, y). 9 Proof. Select η ∈ S(u) arbitrarily. Let v ∈ V be any node such that η = ηu (v). We recall that for the 5 descending path Pu (u0 , v 0 ) ⊆ Pu (v), we have η = arg mine∈Pu (u0 ,v0 ) ρ(e), where du (u0 ) = 12 du (v) 7 2 0 00 00 and du (v ) = 12 du (v). Let v ∈ Pu (v) be a node such that du (v ) = 3 du (v) (we recall that 12\|du (v) by assumption). By the definition of skeleton Tu∗ , we have v 00 ∈ Vu∗ , and clearly Pu (v 00 ) ⊆ Tu∗ . We note that η ∈ Pu (v 00 ) and moreover: 2 7 1 1 1 Reach Tu∗ (η − ) ≥ \|Pu (η − , v 00 )\| ≥ \|Pu (v 0 , v 00 )\| = du (v) − du (v) = du (v) = du (v 0 ) ≥ du (η), 3 12 12 7 7 hence Claims (1) and (2) follow. To show Claim (3), we put P = Pu (u0 , v 0 ) and observe that η = arg mine∈P ρ(e) by definition, and: \|P \| = \|Pu (u0 , v 0 )\| = 5 1 2 2 7 du (v) − du (v) = du (v) = du (v 0 ) ≥ du (η). 12 12 6 7 7 Next, to show Claim (4), we observe that by Claim (3), η = arg mine∈Pu (u0 ,η− ) ρ(e) and η = arg mine∈Pu (η+ ,v0 ) ρ(e). + , v 0 )\| ≥ 1+ 2 d (η) ≥ 2 1 d (η) , Moreover, since \|Pu (u0, v 0 )\| ≥ 27 du (η), we have \|Pu(u0 , η − )\|+\|P (η u u u 7 7 and so \|Pu (u0 , η − )\| ≥ 71 du (η) or \|Pu (η + , v 0 )\| ≥ 17 du (η) . Claim (4) thus follows for anappropriate choice of descending path P ⊆ \|Pu (u0 , η − )\| or P ⊆ \|Pu (η + , v 0 )\|, respectively, with \|P \| = 17 du (η) . Finally, to show Claim (5), we consider separately the two cases from the proof of Claim (4). If P ⊆ Pu (u0 , η − ), then we set y = η − , and choose x ∈ Pu (η − ) so that du (x) = b 87 du (η − )c. We have du (x, y) = d 81 du (η − )e ≤ d 71 du (η − )e = \|P \|, hence Pu (x, y) ⊆ P , η = arg mine∈Pu (x,y) ρ(e), and the claim follows. If P ⊆ Pu (η + , v 0 ), then we set x = η + , and choose y ⊆ Pu (x, v 0 ) so that du (x) = b 78 du (y)c. Such a choice of y is always possible since, when moving with y along the path P , the value of b 78 du (y)c increases by at most 1 in every step; moreover, for the lower end node v ∗ of path P we have du (v ∗ ) > du (x) + 17 du (x), and so du (x) ≤ b 78 du (v ∗ )c. We again obtain η = arg mine∈Pu (x,y) ρ(e), and the claim follows. We remark that the remainder of our analysis is valid for any construction of hub sets S which satisfies Claims (1), (2), and (3) of Lemma 1.2 To bound the average hub set size precisely, we introduce for each node u a parameter called integrated skeleton dimension k̂ (u), defined through a sum of inverse distances to u over nodes of its tree skeleton: X 1 X \| Cut u∗(r) \| k̂ (u) := = , (2) du (v) r ∗ v∈Vu r∈N ∗(r) where the equivalence of the two definitions follows directly from the definition of cuts, Cut u = {v ∈ V ∗ (u) : du (v) = r}. ∗(r) Taking into account that \| Cut u \| ≤ Width(Tu∗ ), we have k̂ (u) = O(Width(Tu∗ ) log D(Tu∗ )), and even more roughly, we have for all u ∈ V : k̂ (u) = O(k log D), (3) where we recall that D = maxu∈V D(Tu ). One may, in particular consider an alternative construction of a hub set S + (u), defined as the set of all edges η ∈ E(Tu ) which satisfy Claims (1), (2), and (3) of the Lemma. All of our bounds on hub set size also hold in the case of {S + (u)}u∈V . The definition results in larger labels in practice: we always have S(u) ⊆ S + (u) (correctness results from this observation). On the other hand, hub sets S + (u) may sometimes be constructed more efficiently than S(u): the definition of S(u) requires a scan of the entire tree T (u), whereas hub set S + (u) may be constructed based only on the smaller skeleton T ∗ (u). 2 10 Lemma 2. The expected hub set size of a node u ∈ V satisfies the bound: E\|S(u)\| ≤ 16 k̂ (u). Proof. Fixu ∈ V arbitrarily. For y ∈ Vu∗ , we define xy as the unique node on the path Pu (y) such that 7 du (xy ) = 8 du (y) . We define random variable Qu (y) ∈ {0, 1} as the number of extreme edges e of the path Pu (xy , y) (i.e., e+ = xy or e− = y), which satisfy e = arg min ρ(Pu (xy , y)). We have: Pr[Qu (y) = 1] = 2 2 16 . = ≤ 7 \|Pu (xy , y)\| du (y) du (y) − 8 du (y) By Claim (5) of Lemma 1, it follows that by summing random variables Qu (y) exhaustively over all vertices y we count each element η ∈ S(u) at least once, hence: X \|S(u)\| ≤ Qu (y). y∈Vu∗ By linearity of expectation, it follows that: E\|S(u)\| ≤ X y∈Vu∗ EQu (y) = X Pr[Qu (y) = 1] ≤ y∈Vu∗ X y∈Vu∗ 16 . du (y) A direct application of Markov’s inequality to the bound from Lemma 2, combined with Eq. (3), gives the following Corollary. P P Corollary 1. The average hub set size satisfies n1 u∈V \|S(u)\| = O( n1 u∈V k̂ (u)) ≤ O(k log D), with probability at least 1/2 w.r.t. choice of random values ρ. Obtaining concentration bounds on the maximal size of a hub set, maxu∈V \|S(u)\|, requires some more care, and we proceed with the analysis in the following subsection. 3.3 Concentration Bounds on Hub Set Size For fixed u ∈ V , we consider the size of the hub set of u given by the random variable: X \|S(u)\| = X(η), (4) η∈E(Tu∗ ) where X(η) ∈ {0, 1} is the indicator variable for the event “η ∈ S(u)”. The random variables {X(η)}η∈E(Tu∗ ) need not, in general, be independent or negatively correlated. In the subsequent analysis, for fixed η, we make use of Claim (4) of Lemma 1 to bound random variable X(η). By the Claim of the Lemma, we can decompose X(η) into the contributions from descending paths located towards the root and away from the root with respect to η: X(η) ≤ X + (η) + X − (η), where we define: • X + (η) ∈ {0, 1} is the indicator variable for the event: “for the unique descending subpath P ⊆ Pu (η − ) of length \|P \| = 71 du (η) ending in edge η (i.e., P − = η − ), it holds that η = arg min ρ(P )”, • X − (η) ∈ {0, 1} is theindicator variable for the event: “there exists a descending path P ⊆ Tu∗ of length \|P \| = 71 du (η) starting in edge η (i.e., P + = η + ), such that η = arg min ρ(P )”. 11 Moreover, by Claim (2) of Lemma 1, we may have X(η) 6= 0 only for those edges η for which Reach Tu∗ (η − ) ≥ 71 du (η). We denote Vu∗∗ = {v ∈ Vu∗ : Reach Tu∗ (v) ≥ 17 du (v)} and Tu∗∗ = Tu [Vu∗∗ ]. We can now rewrite Eq. (4) as: X X X − (η), (5) X + (η) + \|S(u)\| ≤ η∈E(Tu∗∗ ) η∈E(Tu∗∗ ) and proceed to bound both of these sums separately. In order to be able to manipulate the sums more conveniently, we first introduce a partition of the tree according to geometrically increasing scales of distance. Partition of Tu∗ into lLayers. mWe consider a sequence of increasing integer radii (r[i] )i∈N , given as 16 i r[0] = 0 and r[i] = 15 15 , for i ≥ 1. The last non-empty layer corresponds to index imax < log16/15 D < 16 ln D. Cutting the edge set of tree Tu∗ at vertices located at distances {r[i] }i∈N from the root u yields the following partition into layers: [ E(Tu∗ ) = L∗[i] , where: L∗[i] := {e ∈ E(Tu∗ ) : r[i] < du (e) ≤ r[i+1] }. i∈N We further denote the subset of each layer restricted to edges from Tu∗∗ as L∗∗[i] := L∗[i] ∩ E(Tu∗∗ ), i ∈ N. Si P [i,j] , such that Lemma 3. For all i ≥ 1, edge set L∗∗[i] admits a partition into paths, L∗∗[i] = lj=1 ∗(r) li < 2 minr∈[r[i+1] ,r[i+2] ] \| Cut u \| ≤ 2 k , each P [i,j] is a descending path, and all internal nodes of all paths P [i,j] have degree exactly 2 in Tu∗∗ . Si P [i,j] of the edge set of the considered layer so that each path Proof. Define partition L∗∗[i] = lj=1 P [i,j] , 1 ≤ j ≤ li , is a maximal descending path whose internal nodes all have degree exactly 2 in Tu∗∗ . Let F = Tu (L∗∗[i] ) be the oriented sub-forest of Tu induced by the edges of L∗∗[i] . Let l be the number of leaves of F and c be the number of its connected components. An elementary relation between the number of leaves and the number of nodes of degree more than 2 in a forest gives li ≤ 2l − c < 2l. Moreover, by the definition of Tu∗∗ , we have that for each η ∈ L∗∗[i] , we have Reach Tu∗ (η − ) ≥ 1 − [i,j] , such that P [i,j]− is a leaf of F , can be extended along 7 du (η ). It follows that each of the l paths P a descending path in Tu∗ by a distance of 17 du (P [i,j]− ) ≥ 17 r[i] ≥ r[i+2] − r[i] − 1. It follows that each of the l leaves of F can be extended along a (independent) descending path until radius r[i+2] inclusive. ∗(r) Thus, l < minr∈[r[i+1] ,r[i+2] ] \| Cut u \| ≤ k , which completes the proof. Bounding the Sum of X + (η). Denote by P [i] the set of descending paths P in Tu which stretch precisely between the endpoints of the i-th layer: du (P + ) = r[i] , du (P − ) = r[i+1] . For a fixed path P [i,j] ⊆ L∗∗[i] , i ≥ 1, we denote by Q[i,j] the unique path in P [i−1] which extends to P [i,j] , i.e., Q[i,j] ∈ P [i−1] and Q[i,j] ⊆ Pu (P [i,j]+ ). Consider now an arbitrary edge η which does not belong to layers 0 or 1 of the tree partition, η ∈ E(Tu∗∗ ) \ (L∗∗[0] ∪ L∗∗[1] ). Taking into account the above decomposition of set Tu∗∗ into layers, and of layers into paths, there exists a unique path P [i,j] , such that η ∈ P [i,j] . Then, we observe that for the event X + (η) = 1 to hold, it is necessary that two conditions are jointly fulfilled: η must satisfy the prefix minimum condition on the path P [i,j] : η = arg min e∈P [i,j] ∩Pu (η − ) 12 ρ(e), (6) and moreover, we must have ρ(η) < min ρ(Q[i,j] ). Indeed, considering the definition of X + (η), the unique descending subpath P ⊆ Pu (η − ) of length \|P \| = 17 du (η) which ends with edge η has its other endpoint in L∗[i−2] . We have η = arg min ρ(P ), and path P includes as subpaths both the entire prefix P [i,j] ∩ Pu (η + ), and the path Q[i,j] . We denote by M +[i,j] ⊆ P [i,j] the set of all edges η ∈ P [i,j] satisfying ρ(η) < min ρ(Q[i,j] ). We further denote by η +[i,j,k] the k-th edge in M +[i,j] , when ordering edges of M +[i,j] by increasing distance from the root u, 1 ≤ k ≤ \|M +[i,j] \|. Finally, we denote by Z +[i,j,k] ∈ {0, 1} the indicator random variable for the event that “edge η +[i,j,k] satisfies the prefix minimum condition (6) on path P [i,j] ”. It follows that: \|M +[i,j] \| X X X + (η) ≤ η∈E(Tu∗∗ ) X X i=2,3,4,... 1≤j≤li k=1 X + (η) + η∈L∗∗[0] ∪L∗∗[1] Z +[i,j,k] , where we note that the ranges of sum indices i, j do not depend on the random choice of ρ in our setting. We further rewrite the above expression, roughly bounding the first sum by cardinality, and splitting the second double sum according to even and odd values of i: X + X (η) ≤ η∈E(Tu∗∗ ) 19 X \|M +[i,j] \| \| Cut ∗(r) \| u r=0 X X i=2,4,6,... 1≤j≤li k=1 + \|M +[i,j] \| Z +[i,j,k] {z \| } :=A+ even X X i=3,5,7,... 1≤j≤li k=1 + \| Z +[i,j,k] . {z :=A+ odd (7) } We subsequently consider only bounds on the summed expression A+ even for 2\|i (bounds on the expression A+ follow by identical arguments). odd We observe that for fixed i, i = 2a for some a ∈ N + , the random variables \|M +[i,j] \| depend only on the choice of random values ρ(e) for e ∈ L∗[i−1] . Now, conditioning on a choice ρodd of random values ρ(e) for e ∈ L∗[2a−1] , for all a ∈ N+ , we observe that {Z +[i,j,k] }2\|i,j,1≤k≤\|M +[i,j] \| is a set of independent random variables, with: Pr[Z +[i,j,k] = 1\|ρodd ] = 1/k. The above probability and independence follows directly from a well-known characterization of the probability that the k-th element of a uniformly random permutation (ordering) is its prefix minimum. We have: \|M +[i,j] \| E[A+ even \|ρodd ] = X X i=2,4,6,... 1≤j≤li k=1 1 < k X (1 + ln◦ \|M +[i,j] \|) + =: Neven , (8) i=2,4,6,... 1≤j≤li where ln◦ x = ln x for x > 0 and ln◦ 0 = 0. By an application of a simple Chernoff bound for the sum of variables {Z +[i,j,k] }i,j,k , we have: + −c Pr[A+ even < Neven + 3c ln n] > 1 − n , for c > 1. (9) + It now remains to provide bounds on the concentration of random variable Neven in its upper tail. If our only goal is to bound the hub set size as O(k log D log log n + log n), then obtaining such bounds becomes a relatively straightforward exercise in Chernoff bounds over individual paths P [i,j] . In this work, we go for a more pedestrian approach through a type of balls-into-bins process, optimizing bounds over larger path sets, which will eventually give us slightly tighter bounds, including a bound of O(k log D log log k ).P i Denote Mi+ := lj=1 \|M +[i,j] \|, for i ≥ 2. Then, we have the following bound. 13 L[i+1] L[i] L[i-1] P[i, 2] P[i, 1] Q[i, j={1,2,3,4,5}] P[i, 4] P[i, 3] P[i, 5] Q[i, 6] u P[i, 6] Q[i, 7] r [i-1] P[i, 7] r [i] r [i+1] r [i+2] Figure 2: Illustration of paths P [i,j] and Q[i,j] . Edges from Tu∗∗ are marked with solid lines, remaining edges from Tu∗ are marked with dashed lines. Lemma 4. Fix i ≥ 2. Then, for any c ≥ 1: Pr Mi+ > 12cli < 2−c . Si Si Q[i,j] (forming a subforest of Tu∗ , P [i,j] ∪ lj=1 Proof. For fixed i, we consider the edge set F := lj=1 contained entirely within layers L∗[i−1] and L∗[i] ). See Figure 3.3 for an illustration. Choose arbitrarily the set I of (necessarily distinct) values of ρ which appear within F , I = ρ(F ). Let I = {i1 , i2 , . . . , i\|F \| }, with i1 ≤ i2 ≤ . . . ≤ i\|F \| . We couple the sampling of values ρ on F with the following two-stage process. First, we fix set I. Then, given a choice of I, we select a uniformly random permutation to perform the assignment of values from I to edges in F (\|I\| = \|F \|). The latter permutation is defined iteratively, by assigning to successive values it , 1 ≤ t ≤ \|F \|, an as yet unoccupied edge (site) from F . The value of \|M +[i,j] \| is given as the number of elements of I which are placed in sites from P [i,j] before the smallest index tj ∈ N, such that itj ∈ ρ(Q[i,j] ). We will refer to the index t as representing moments of time, and we will then say that path P [i,j] was cut off at time tj . For successive moments of time t ∈ N , we denote by Jt the set of surviving path indices at time t, 0 i.e., Jt = {j it0 ∈ /S ρ(Q[i,j] )}. We then obtain subforest Ft ⊆ F by restricting F to its surviving S : ∀t ≤t[i,j] part, Ft = j∈Jt P ∪ j∈Jt Q[i,j] . To prove the claim, we will consider how the random variable Mi+ increases over time, until Ft = ∅. We again couple our sampling process by first deciding in each time step whether to place it in forest F \ Ft−1 or in forest Ft−1 , and only afterwards fixing for it a specific free site with uniform probability within the chosen subforest. Observe that if it is placed F \Ft−1 , then in the given step of the considered process, the value of Mi+ remains unchanged at time t. We will thus eliminate from the process all time steps t such that it ∈ F \ Ft−1 , and by a slight abuse of notation, we will relabel time indices as if these steps never occurred. Thus, in each time step t, we assume that a free site is picked for it from Ft−1 uniformly at random. Consider now the random variable φt = \|Jt−1 \| − \|Jt \|, representing the number of paths cut off in time step t. The expectation of φt can be lower-bounded, regardless of the history of the process. Claim (). E[φt \|Ft−1 ] > 1/3, for any Ft−1 6= ∅. S Proof (of Claim). Fix forest Ft−1 . We assign to each edge e ∈ j∈Jt−1 Q[i,j] a weight w(e) ∈ N+ , given as the j ∈ Jt−1 such that P [i,j] can be reached from e by a descending path. For all S number of [i,j] edges e ∈ j∈Jt−1 P , we put w(e) = 0. Now, if e is chosen as the t-th edge in the process, we have 14 φt = \|Jt−1 \| − \|Jt \| = w(e). It follows that: P P 1 \|Jt−1 \|(r[i] − r[i−1] ) e∈Ft−1 w(e) e∈Ft−1 w(e) > E[φt \|Ft−1 ] = ≥ ≥ , \|Ft−1 \ {e ∈ F : ρ(e) < it }\| \|Ft−1 \| 3 \|Jt−1 \|(r[i+1] − r[i−1] ) which completes the proof of the claim. P Moreover, taking into that φt has bounded range φt ∈ [0, li ], and that t∈N φt = li , we obtain a concentration result on the number of steps until the stopping of the process (Ft = ∅); for completeness, we provide a standalone proof. Claim (*). For T = 12cli , we have: Pr[FT 6= ∅] < 2−c . PT 0 Proof (of Claim). We define the non-negative submartingale Φ0T = t=1 φt as follows. When 0 0 0 Ft−1 6= ∅, we choose φt ∈ [0, li ] to be dominated by φt , so that φt ≤ φt and E[φt \|Ft−1 ] = 1/3. (The latter condition can always be satisfied by Claim ()). When Ft−1 = ∅, we put φ0t = 1/3. Observe that when Φ0T > li for some T , we necessarily have φ0t > φt for some t ≤ T , hence FT = ∅. To lower-bound the probability of the event Φ0T > li , we remark that the bounds φ0t ∈ [0, li ] and E[φ0t \|Ft−1 ] = 1/3 imply the following bound on variance of the process: σ 2 [φ0t \|Ft−1 ] ≤ li /3. Now, using a standard martingale bound (cf. e.g. [17][Thm. 18] applied to the process Xt = Φ0t − t/3), we obtain for any λ > 0: −λ2 /2 0 . Pr[ΦT ≤ T /3 − λ] ≤ exp T li /3 + λli /3 Substituting T = 12cli and λ = 3cli , we obtain: Pr[Φ0T ≤ cli ] ≤ e−0.9c < 2−c . Taking into account that c ≥ 1 by assumption, the claim follows directly. Recalling that in each time step t with Ft 6= ∅, the value of random variable Mi+ increases by at most 1, we obtain directly from Claim (*) that Pr[Mi+ > 12cli ] < 2−c , which completes the proof. Next, for 2\|i, let Ci ∈ N+ be a random variable defined as the smallest integer such that Mi+ ≤ 12Ci li . Since Ci depends only on random values ρ chosen with L∗[i−1] ∪ L∗[i] , the random variables {Ci }2\|i are independent. Moreover, by Lemma 4, each Ci may be stochastically dominated by a (independent) geometrically distributed random variable Γi with parameter p = 1/2. It follows that: X X Ci ≤ Γi ∼ NB(bimax /2c, 1/2), 2\|i, i≤imax 2\|i, i≤imax where the parameters of the negative binomial distribution NB(r, p) represent the number of trials with success probability p until r successes are reached. An application of a rough tail bound for NB(bimax /2c, 1/2) gives:   X Pr  Ci < 2imax + 4c ln n > 1 − n−c , for any c > 1. (10) 2\|i Recalling that Mi+ = Pli j=1 \|M +[i,j] \| ≤ 12Ci li , we may write by concavity of the logarithm function:   li li X X 1 (1 + ln◦ \|M +[i,j] \|) ≤ li + li ln◦  \|M +[i,j] \| ≤ li + li ln(12Ci ) < li ln(33Ci ). li j=1 j=1 15 + , taking into account its definition (8): We can therefore bound variable Neven X X + li ln(33Ci ). Neven = (1 + ln \|M +[i,j] \|) ≤ 2\|i 1≤j≤li (11) 2\|i We now apply a union bound over the two events given by (9) and (10), which hold w.h.p., From Eq. (9), (10), and (11), we have that for any c > 1, the following event holds w.p. at least 1 − 2n−c : X + Neven ≤ 3c ln n + max li ln(33ci ), (12) (ci ) 2\|i i≤imax P where (ci )i≤imax ,2\|i are positive integers satisfying the condition: 2\|i ci < 2imax + 4c ln n. + Returning to Eq. (7), with respect to Nodd an analogous technique gives us that w.p. at least 1−2n−c : X + Nodd ≤ 3c ln n + max (ci ) li ln(33ci ), (13) 26=i i≤imax P where (ci )i≤imax ,26\| i are likewise positive integers satisfying the condition: 26\| i ci < 2imax + 4c ln n. Combining Eq. (7) with Eq. (12) and (13) through a union bound (and substituting γi := 33ci ), we eventually obtain that w.p. at least 1 − 4n−c : X X + (η) ≤ η∈E(Tu∗∗ ) 19 X \| Cut ∗(r) \| + 6c ln n + max u (γi ) r=0 X li ln γi , (14) i≤imax where (γi )i≤imax satisfy the condition: ∀i γi ∈ N+ and X γi < 132imax + 264c ln n. (15) Bounding the Sum of X − (η). For the random variables X − (η), the main arguments required to establish the bound are similar to those in the case of X + (η); we confine ourselves to an exposition of the differences. The main difference is that for a path P [i,j] , instead of a unique predecessor path Q[i,j] in layer L∗[i−1] , we now have to deal with multiple possible descendant paths in layer L∗[i+1] ; on the other hand, the outward-branching structure of the tree means that we can show tighter concentration bounds in this case. We recall that P [i] is the set of descending paths P in Tu which stretch precisely between the endpoints of the i-th layer, and for 1 ≤ j ≤ li , we denote by R[i,j] the set of all paths in P [i+1] which are extensions of P [i,j] , i.e., R[i,j] ⊆ P [i+1] and for all R ∈ R[i,j] , we have P [i,j] ⊆ Pu (R+ ). For the event X − (η) = 1 to hold, it is now necessary that two conditions are jointly fulfilled: η must satisfy the suffix minimum condition on the path P [i,j] : η = arg min e∈P [i,j] \Pu (η + ) ρ(e), (16) and moreover, ρ(η) < maxR∈R[i,j] min ρ(R). We next denote by M −[i,j] ⊆ P [i,j] the set of all edges η ∈ P [i,j] satisfying ρ(η) < maxR∈R[i,j] min ρ(R). We further denote by η −[i,j,k] the k-th edge in M −[i,j] , when ordering edges of M −[i,j] by decreasing distance to the root u, 1 ≤ k ≤ \|M −[i,j] \|. Finally, we denote by Z −[i,j,k] ∈ {0, 1} the indicator random variable for the event that “edge η −[i,j,k] satisfies the suffix minimum condition (16) on path P [i,j] ”. The subsequent analysis proceeds as before, and we obtain direct analogues of Eq. (7), (8), and (9), replacing superscripts “+” ofPall random variables by “−”. i We next denote Mi− := lj=1 \|M −[i,j] \|, for i ≥ 2, and obtain the following analogue of Lemma 4. 16 Lemma 5. Fix i ≥ 2. Then, for any c ≥ 1: Pr Mi− > 12cli < e−4cli < 2−c . Proof. The proof follows along lines as that of Lemma 4. Sli similar Si [i] [i,j] For fixed i, let R := j=1 R . We consider the edge set F := lj=1 P [i,j] ∪ R[i] (forming a subforest of Tu∗ , contained entirely within layers L∗[i] and L∗[i+1] ). Choose arbitrarily the set I of (necessarily distinct) values of ρ which appear within F , I = ρ(F ). Let I = {i1 , i2 , . . . , i\|F \| }, with i1 ≤ i2 ≤ . . . ≤ i\|F \| . As in the proof of Lemma 4, we couple the sampling of values ρ on F with the following two-stage process. First, we fix set I. Then, given a choice of I, we select a uniformly random permutation to perform the assignment of values from I to edges in F (\|I\| = \|F \|). The latter permutation is defined iteratively, by assigning to successive values it , 1 ≤ t ≤ \|F \|, an as yet unoccupied edge (site) from F . The value of \|M −[i,j] \| is given as the number of elements of I which are placed in sites from P [i,j] before the smallest index t ∈ N, such that for all paths R ∈ R[i,j] , there exists t0 < t such that it0 ∈ ρ(R). We will refer to the index t as representing moments of time. Let R[i] = {R1 , . . . , R\|R[i] \| . We say that a path Rk ∈ R[i] was cut off at time t if t is the smallest time such that it ∈ ρ(Rk ). For successive moments of time t ∈ N , we denote by Kt the set of surviving indices k of paths Rk which have not been cut off at time t. We obtain subforest Ft ⊆ F by restricting F to its surviving part, where we treat a path P [i,j] if it extends to at least one surviving path Rk : [ [ Ft = Rk ∪ P [i,j] . k∈Kt j: ∃k∈Kt P [i,j] ⊆Pu (Rk+ ) Exactly as in the proof of Lemma 4, we will consider how the random variable Mi− increases over time, until Ft = ∅. We again couple our sampling process by first deciding in each time step whether to place it in forest F \ Ft−1 or in forest Ft−1 , and only afterwards fixing for it a specific free site with uniform probability within the chosen subforest. Observe that if it is placed F \ Ft−1 , then in the given step of the considered process, the value of Mi− remains unchanged at time t. We will thus eliminate from the process all time steps t such that it ∈ F \ Ft−1 , and by a slight abuse of notation, we will relabel time indices as if these steps never occurred. Thus, in each time step t, we assume that a free site is picked for it from Ft−1 uniformly at random. Consider now the random variable φt = \|Kt−1 \| − \|Kt \|, representing the number of paths R cut off in time step t. The expectation of φt can be lower-bounded, regardless of the history of the process. Claim (). E[φt \|Ft−1 ] ≥ 1/2, for any Ft−1 6= ∅. Proof (of Claim). Fix forest Ft−1 . When inserting it , the number of free sites in layer L∗[i+1] is at least: X \|Rk \| ≥ \|Kt−1 \|(r[i+1] − r[i] ). k∈Kt−1 On the other hand, since each surviving path P [i,j] extends to some surviving path Rk , the total number of free sites for insertion of it is upper-bounded by \|Kt−1 \|(r[i+1] − r[i−1] . Since insertion of it into layer L∗[i] means that \|Kt \| = \|Kt−1 \|, and insertion of it into layer L∗[i+1] means that \|Kt \| = \|Kt−1 \| − 1, we obtain: \|Kt−1 \|(r[i+1] − r[i] ) 1 E[φt \|Ft−1 ] ≥ ≥ , [i+1] [i−1] 2 \|Kt−1 \|(r −r ) which completes the proof of the claim. P Moreover, taking into that φt has bounded range φt ∈ {0, 1}, and that t∈N φt ≤ li , we obtain a concentration result on the number of steps until the stopping of the process (Ft = ∅) directly from the 17 Azuma-McDiarmid martingale inequality (cf. e.g. [17][Thm. 16], [26]). For parameter T = 12cli , we obtain after some transformations: Pr[F12cli 6= ∅] < e−4cli . Recalling that in each time step t with Ft 6= ∅, the value of random variable Mi− increases by at most 1, we obtain directly that Pr[Mi− > 12cli ] < e−4cli , which completes the proof. The rest of the argument proceeds as for the case of X + , applying Lemma (5) in place of Lemma (4). We eventually obtain the following analogue of Eq. (7): for any c > 1, w.p. at least 1 − 4n−c : X X − (η) ≤ η∈E(Tu∗∗ ) 19 X X \| Cut ∗(r) \| + 6c ln n + max u (γi ) r=0 li ln γi , (17) i≤imax where (γi )i≤imax satisfy condition (15). Combining Bounds. Introducing the bounds of Eq. (14) and (17) into Eq. (5) through a union bound we obtain the following statement: For any c > 1, w.p. at least 1 − 8n−c : \|S(u)\| ≤ 2 19 X \| Cut u∗(r) \| + 12c ln n + 2 max (γi ) r=0 X li ln γi , (18) i≤imax where (γi )i≤imax satisfy condition (15). Now, recalling the bounds on li from Lemma 3, the bound imax < 16 ln D, setting c = 2, and applying a union bound over all vertices u, we obtain the main technical result of the Section. We present it first in its strongest form, and then provide two more useful corollaries. Theorem 1. With probability at least 1 − 8/n, all nodes u ∈ V satisfy the following bound on hub set size: 19 X X S(u) ≤ 2 \| Cut ∗(r) \| + 24 ln n + 2 max li ln γi , (19) u (γi ) r=0 i=1,2,...,b16 ln Dc l i m ∗(r) where li ≤ 2 minr∈[r[i+1] ,r[i+2] ] \| Cut u \| with r[i] = 15 16 , and the maximum is taken over posi15 P tive integers (γi ) satisfying the condition: γi < 2112 ln D +528 ln n. We provide two more convenient corollaries of Theorem 1. For the case when the considered trees ∗(r) Tu are close to scale-free, we simply bound the size of all cuts Cut u through skeleton dimension: ∗(r) \| Cut u \| ≤ k . Bound (19) then takes the asymptotic form, for imax = b16 ln Dc: X S(u) ≤ O(k ) + O(log n) + O(k ) max ln γi , (20) (γi ) i≤imax where the latter sum can be bounded using the concavity of the logarithm function as:   X X 1 log n   γi ≤ imax O max 1, log ≤ max ln γi ≤ imax max ln imax log D (γi ) (γi ) i≤imax i≤imax log n ≤ O log D max 1, log . log D 18 We also observe the following link between the parameters k , D, and n. Since by Proposition 1 graph G has doubling dimension bounded by 2 k +1, it follows that a radius-D ball in G may only contain at most (2 k +1)dlog2 De nodes from V . Hence, (2 k +1)dlog2 De ≥ n, and we obtain: log n = O(log k · log D). (21) Thus, the O(log n) additive factor in the bound (20) on S(u) is dominated in notation by the last factor of the sum, which is stated as at least O(k log D). Combining the above, we obtain the following corollary. Corollary 2. With probability at least 1 − O(1/n), the hub set size of every node is bounded by: log n . O k log D max 1, log log D In particular, when the graph has sufficiently large diameter, D = nΩ(1) , we have that the hub set size of all nodes is bounded by O(k log D). For the general case, by introducing Eq. (21) into Corollary 2, we obtain the following statement. Corollary 3. With probability at least 1 − O(1/n), the hub set size of every node is bounded by: O (k log log k log D) . When considering the case of trees in which the width of tree Tu∗ is far from uniform over different scales of distance, tighter bounds are obtained by relating the size of S(u) toP the integrated skeleton dimension k̂ (u). To do this, we apply in Eq. (19) the rough bound: ln γi < ln i γi = O(log log n + log log D). This leaves us with an expression of the form: X S(u) ≤ O(k̂ (u))+O(log n)+O(log log n+log log D) li ≤ O(log n+k̂ (u)(log log n+log log D)), i≤imax where we used the bound rameter k̂ . P i≤imax li ≤ O(k̂ (u)), which follows easily from the definition of the pa- Corollary 4. With probability at least 1 − O(1/n), the hub set size of every node u ∈ V is bounded by O(log n + k̂ (u)(log log n + log log D)). 4 An Application to δ-preserving Distance Labeling As a slight extension of our results, we note that our technique based on analyzing tree skeletons for shortest path trees has direct application the δ-preserving distance labeling problem in unweighted graphs, for some parameter δ > 0. We recall that a scheme is called δ-preserving if for any queried pair of nodes (u, v) with δ(u, v) > δ, the value returned by the decoder is equal to δ(u, v). By analogy to the integrated skeleton dimension given by (2), we introduce a variant for this parameter which only considers cuts at distance more than δ/6. k̂ δ (u) := X v∈Vu∗ :d(u,v)>δ/6 1 = du (v) ∗(r) X r∈N,r>δ/6 \| Cut u r \| . (22) The claims of Lemma 2 and Corollary 4, which give bounds on average hub size of O(k̂ (u)) and O(log n + k̂ (u)(log log n)) in the unweighted setting, naturally translate to δ-preserving labeling. For our techniques to be directly applicable, it suffices to subdivide each edge of the graph into a path of 12 19 vertices so that all distances between u − v pairs are divisible by 12, and to choose shortest path trees so that for any pair of nodes u, v, the intersection Tu ∩ Tv contains a shortest u − v path (this may be achieved, for example, by enforcing a unique choice of shortest paths between any node pair, e.g., by choosing i.a.r. the length of each edge in the range [1 − 1/n, 1]). Then, the entire analysis holds, and we can eventually replace k̂ (u) by k̂ δ (u) in the statement of the claims. ∗(r) We remark that it is an elementary property of the tree skeleton that \| Cut u \| = O(n/r), since any node at distance r from u continues in Tu along an independent path of length at least r/2. By performing the latter sum in (22), we obtain k̂ δ (u) = O(n/δ). Thus, we obtain the following Proposition. Proposition 3. There exists a hub labeling scheme for the δ-preserving distance labeling problem with hubs of average size O(n/δ) and worst case size O(log n+(n/δ) log log n). The size of the bit representations of the corresponding labels is O((n/δ) log n) and O(log2 n+(n/δ) log n log log n), respectively. The size of the obtained δ-preserving hub-based labeling scheme is (almost) optimal, since there holds a lower bound of Ω(n/δ) on both the average and worst-case size of hub sets [10]. In fact, our scheme can be modified slightly to obtain hub sets of worst-case size O(n/δ) up to a certain threshold e √n). We present the details of this modified scheme in the following Subsection. value δ = O( 4.1 A Modified δ-preserving Labeling Scheme In this section we present an independent family of distance labeling schemes, which have the δpreserving property. Whereas the scheme and the presented analysis are valid for any value of parameter δ > 0, we obtain an improvement on the previously discussed scheme only up to some threshold value e √n). δ = O( Construction of the Labeling. Fix the value of parameter δ > 0, with 12\|δ. The basic building block of our labeling is a construction of hub sets Sδ (u) for each node u ∈ V , which allow us to handle distance queries for pairs of nodes whose distance is in the range [δ, 5δ/4]. Before providing the details of the constructions of sets Sδ (u), we first introduce some auxiliary notation. As before, for a pair of nodes u, v ∈ V , we denote by P (u, v) a fixed shortest path between u and v. In the definition of P (u, v), ties between different shortest paths should be broken in a consistent manner S over the whole graph, so that P (u, v) = P (v, u), and the set of shortest paths rooted at a node u, v∈V P (u, v), is a spanning tree of G. For a node u ∈ V , we denote by Tu the shortest path subtree of G, rooted at u, leading from u to nodes at distance in the range [δ, 5δ/4]: [ Tu = P (u, v). v∈V : d(u,v)∈[δ,5δ/4] We denote by Tu∗ the subtree (skeleton) of tree Tu , also rooted at u and truncated to its first 3δ/4 levels from the root: Tu∗ = Tu [{v ∈ V (Tu ) : d(u, v) ≤ 3δ/4}]. We remark that all descending paths in Tu∗ have reach of at least δ/4 in Tu . The set Sδ (u) will be constructed similarly as before, to include vertices from the central part of any u − v path in the tree, for vertices v with d(u, v) ∈ [δ, 5δ/4]. However, we wish to control the number of possible bad events in which a descending path in the tree Tu∗ branches out at some level into too many subpaths, from each of which some representative node will need be chosen into Sδ (u). To do this, we will partition the vertex set of tree Tu∗ into two subsets, Hu ∪ Lu , known as heavy and light vertices, respectively. A vertex w of Tu∗ belongs to Hu if the subtree of Tu∗ rooted at w has at least δ leaves (all in the last level 3δ/4), and belongs to Lu otherwise. We remark that Tu∗ [Hu ] is a (possibly empty) subtree of Tu∗ , whereas Tu∗ [Lu ] is a sub-forest of Tu∗ , in which each connected component is a tree with less than 20 Hu u 3 4δ δ 5 4δ Figure 3: Hub set selection for distance range [δ, 5δ/4]. The corresponding shortest path tree Tu for some vertex u is shown in the figure. The set of heavy vertices Hu is shaded around vertex u; all remaining vertices up to distance 34 δ belong to Lu . Vertices of L0u ⊆ Lu are also marked, with corresponding descending paths Pd shaded. δ leaves. In all the considered trees, we maintain the same ancestry relation. In particular, we speak of a descending subpath in a tree if one of its endpoints is an ancestor of the other with respect to the tree Tu∗ rooted at u. We are now ready to define the hub sets Sδ (u), u ∈ V by the following randomized construction. Assign to each node v ∈ V a real value ρ(v) ∈ [0, 1], uniformly and independently at random. We now put for all u ∈ V : Sδ (u) := Hu ∪ L0u , (23) where L0u ⊆ Lu is defined as the set of all vertices v ∈ Lu , such that there exists a descending subpath Pd in Tu∗ [Lu ], such that v ∈ Pd , \|Pd \| = δ/12, and v has the minimal value of ρ along path Pd , v = arg minw∈Pd ρ(w) ≡ arg min ρ(Pd ). See Fig. 3 for an illustration. Correctness. We start by showing that sets Sδ have the desired hub property, regardless of the choice of random values ρ (which may only affect the size of these sets). Lemma 6. For any pair of nodes u, v ∈ V such that d(u, v) ∈ [δ, 5δ/4], we have: d(u, v) = (d(u, w) + d(w, v)) . min w∈Sδ (u)∩Sδ (v) Proof. Consider the path P = P (u, v) = P (v, u). We have P ⊆ Tu and P ⊆ Tv . Moreover, \|P \| ≤ 5δ/4 and \|P ∩ Tu∗ \| = \|P ∩ Tv∗ \| = 3δ/4. Denoting by P ∗ the subpath of P which belongs to both trees Tu∗ and Tv∗ , P ∗ = P ∩ Tu∗ ∩ Tv∗ , it follows that \|P ∗ \| ≥ δ/4. We now prove the claim of the lemma by showing that at least one vertex from P ∗ has to belong to Sδ (u) ∩ Sδ (v). We achieve this by a case analysis, depending on the portions of path P ∗ which belong to the sets Lu , Hu , Lv , and Hv . • If \|P ∗ ∩Hu ∩Hv \| > 0, then there exists at least one vertex w ∈ P ∗ ∩Hu ∩Hv ⊆ P ∗ ∩Sδ (u)∩Sδ (v), which completes the proof. • If \|P ∗ ∩ Lu ∩ Hv \| ≥ δ/12, then there exists at least one descending subpath Pd of length δ/12 in Tu∗ [Lu ] which is completely contained in P ∗ ∩ Hv . Setting w = arg min ρ(Pd ), we have w ∈ L0u , and so it follows that w ∈ P ∗ ∩ L0u ∩ Hv ⊆ P ∗ ∩ Sδ (u) ∩ Sδ (v). • If \|P ∗ ∩ Hu ∩ Lv \| ≥ δ/12, we obtain the result by applying analogous considerations as in the previous case. 21 • Finally, in all other cases we must have \|P ∗ ∩ Lu ∩ Lv \| ≥ δ/12. It follows that there exists at least one subpath Pd ⊆ P ∗ of length δ/12, which is a descending subpath in both Tu∗ [Lu ] and Tv∗ [Lv ]. Setting w = arg min ρ(Pd ), we obtain w ∈ L0u and w ∈ L0v , hence w ∈ P ∗ ∩ L0u ∩ L0v ⊆ P ∗ ∩ Sδ (u) ∩ Sδ (v). Analysis. We now consider the size of sets Sδ (u). The size of set Hu is independent of the choice of random variables ρ; it can easily be bounded, taking into account that tree Tu∗ has O(n/δ) leaves. Lemma 7. For all u ∈ V , \|Hu \| ≤ 3n/δ. Proof. Let l ∈ Hu be a leaf node of Tu∗ [Hu ]. By definition of Hu , we have that the subtree of Tu∗ rooted at l has at least δ leaves. As every leaf of Tu∗ is at depth 3δ/4, and all leaves in tree Tu are at depth at least δ, it follows that the subtree of l in Tu contains at least δ disjoint descending paths of length δ/4 each, and so it has at least δ 2 /4 nodes. Since the size of tree Tu is at most n, we obtain that tree Tu∗ [Hu ] has at most 4n/δ 2 leaves. Moreover, the distance of each node of Tu∗ [Hu ] from its root u is at most 3δ/4. Hence, \|Tu∗ [Hu ]\| ≤ (3δ/4)(4n/δ 2 ) = 3n/δ. The size of set L0u depends on the choice of random variables ρ. We start by bounding the expected number of elements of L0u , belonging to specific connected components of Tu∗ [Lu ]. Suppose Tu∗ [Lu ] be (1) (k ) a forest consisting of ku trees, and let Lu = Lu ∪ . . . ∪ Lu u be the partition of its vertex set such that (i) (i) (i) Tu∗ [Lu ] represents its i-th connected component. Let lu denote the number of leaves of tree Tu∗ [Lu ]. 0(i) (i) 0(1) 0(k ) Finally, let Lu = L0u ∩ Lu . Clearly, Lu ∪ . . . ∪ Lu u is a partition of L0u . In the following, we P 0(i) u \|Lu \|, showing that it has an expectation of O(n/δ), and consider the random variable \|L0u \| = ki=1 obtaining concentration results around this expectation. First, we remark that each descending path of tree Tu∗ contributes O(1) elements in expectation to set 0(i) Tu∗ . Consequently, the expected size of set \|Lu \| can be related to the number of leaves in the considered connected component. 0(i) (i) Lemma 8. For all u ∈ V and all 1 ≤ i ≤ ku , E\|Lu \| ≤ 36lu . (i) (i) Proof. Let Pu be the set of (inclusion-wise) maximal descending paths in the tree Tu∗ [Lu ]. We remark (i) (i) that \|Pu \| ≤ lu . (i) For a path P ∈ Pu , let MP (v) be the event that there exists a subpath Pd ⊆ P , with \|Pd \| = δ/12, such that v = arg min ρ(Pd ). We have Pr[MP (v)] = 0 for v ∈ / P . If v ∈ P , then we use the following folklore probability estimation: for MP (v) to hold, one of the two (at most) descending subpaths P 0 of P of length δ/24, having v as one of their endpoints, must satisfy v = arg min ρ(P 0 ). It follows that for 0(i) v ∈ P , we have Pr[MP (v)] ≤ 48/δ. By linearity of expectation we now obtain a bound on E\|Lu \|: E\|L0(i) u \|= X v∈V Pr[v ∈ L0(i) u ]≤ X X Pr[MP (v)] ≤ \|Pu(i) \| · max \|P \| · (i) P ∈Pu (i) P ∈Pu v∈P 48 3δ 48 ≤ lu(i) · · = 36lu(i) . δ 4 δ By linearity of expectation, we can apply the claim of Lemma 8 over all connected components, obtaining the following result. Lemma 9. For all u ∈ V , E\|L0u \| ≤ 144n/δ. 22 Proof. By Lemma 8 we have: E\|L0u \| =E ku X \|L0(i) u \| ≤ 36 i=1 ku X lu(i) . i=1 P u (i) lu ≤ 4n/δ. Indeed, this sum represents the total number The claim follows when we observe that ki=1 ∗ of leaves in Tu [Lu ]. Each such leaf (located at distance 3δ/4 from u) is the upper endpoint of a distinct descending path of length at least δ/4 in the tree Tu , and \|Tu \| ≤ n, hence we obtain the bound. In order to apply Chernoff bounds to the sum of random variables \|L0u \|, we start by bounding the range of these variables. 0(i) (i) Lemma 10. For all u ∈ V and all 1 ≤ i ≤ ku , \|Lu \| ≤ \|Lu \| < δ 2 . 0(i) (i) (i) Proof. We have Lu ⊆ Lu . By the definition of set Lu , tree Tu∗ [Lu ] has less than δ leaves and all its (i) nodes are at distance at most 3δ/4 from its root. It follows that \|Lu \| < δ · 3δ/4 < δ 2 . 0(i) The above upper bound provides an estimate on the maximum value of each random variable \|Lu \|. However, in order to be able to perform a concentration analysis in a range of fairly large δ (roughly, 0(i) for n1/3 < δ < n0.5 ), we also need to bound more tightly the concentration of each \|Lu \| around its expected value. Let V0 = {v ∈ V : ρ(v) < 50 lnδn }. We start by showing that with high probability, all elements of the sets \|L0u \| belong to V0 . Lemma 11. Denote by X the “bad” event that there exists a node u ∈ V , such that L0u 6⊆ V0 . We have: Pr[X] < 1/n. Proof. Consider first the probability pv that a fixed node v ∈ V \ V0 satisfies v = arg min ρ(P ), where P is any fixed path of δ/12 nodes in G which contains v. We have (with the last inequality holding when δ ≥ 300): Y pv = (1 − ρ(v)) = (1 − ρ(v))δ/12−1 < (1 − 50 ln n/δ)δ/12−1 < n−4 . w∈P \{v} The probability of event X occurring can be upper-bounded by performing a union bound over all nodes u, all descending paths P in Tu∗ [Lu ], and all nodes v ∈ P of the event [v = arg min ρ(P )∩v ∈ / V0 ] occurring. For each node u, there are at most n such paths to consider, and less than δ possible nodes v in each path. Overall, we obtain: Pr[X] ≤ n · n · δ · pv < n3 · n−4 = 1/n. (i) Next, we show that with high probability, each connected component Lu contains at most O(δ log n) nodes from V0 . Lemma 12. Denote by Y the “bad” event that there exists a node u ∈ V and 1 ≤ i ≤ ku , such that (i) \|Lu ∩ V0 \| > 100δ ln n. We have: Pr[Y ] < 1/n. 23 Proof. Let Z(v) denote the indicator variable for node v and set V0 , i.e., Z(v) = 1 if v ∈ V0 and Z(v) = 0 otherwise. Clearly, Pr[Z(v) = 1] = 50 ln n/δ, and all random variables Z(v), v ∈ V are (i) independent. For any fixed \|Lu \|, we have: X E Z(v) ≤ \|L(i) u \| · 50 ln n/δ ≤ 50δ ln n, (i) v∈Lu (i) where we used 10 to bound \|Lu \|. Next, we proceed by apply a simple multiplicative Chernoff bound for the considered random variable: X Pr[\|L(i) Z(v) > 100δ ln n] ≤ e−50δ ln n/3 < n−16 . u ∩ V0 \| > 100δ ln n] = Pr[ (i) v∈Lu (i) Applying a union bound over Lu , for all u ∈ V and 1 ≤ i ≤ ku , gives the claim. We are now ready to apply a Chernoff-bound type analysis to the random variable \|L0u \|. √ Lemma 13. Let δ ≤ n/ ln n. Then: Pr[∀u∈V \|L0u \| < 800n/δ] = 1 − O(1/n). 0(i) 0(i) 0(i) (i) (i) Proof. Define the random variable λu as \|Lu \| when Lu ⊆ Lu ∩ V0 and \|Lu ∩ V0 \| ≤ 100δ ln n, 0(i) and fix λu = 0 otherwise. Pku 0(i) 0(i) Define λ0u = i=1 λu . All λu are independent random variables, since they are functions of (i) 0(i) disjoint sets of random variables (ρ(v) : v ∈ Lu ). Moreover, we have 0 ≤ λu ≤ 100δ ln n. An application of a simple multiplicative Chernoff bound gives for any A > 0: Pr[λ0u 144en/δ A/(100δ ln n) 400n/δ A/(100δ ln n) Eλ0u A/(100δ ln n) ≤ < , ≥ A] < e A A A where we used the bound Eλ0u ≤ E\|L0u \| ≤ 144n/δ following from Lemma 9. Putting A = 800n/δ and √ taking into account that δ ≤ n/ ln n, we get for sufficiently large n: Pr[λ0u ≥ 800n/δ] < 2−8n/(δ 2 ln n) < 1 . n2 By applying a union bound over all nodes, we obtain that Pr[∀u∈V λ0u < 800n/δ] > 1 − 1/n. Taking into account Lemmas 11 and 12, we also have: Pr[∀u∈V λ0u = \|L0u \|] > 1 − 2/n. Overall, we obtain: Pr[∀u∈V \|L0u \| < 800n/δ] > 1 − 3/n. In view of the definition of the proposed hub set labeling (Eq. (23)), Lemmas 7 and 13 complete √ the analysis of the case of δ ≤ n/ ln n, showing that our randomized construction yields with high probability hub sets of size O(n/δ) for all nodes of the graph. √ Proposition 4. For any 0 < δ ≤ n/ ln n, 12\|δ, there exists a hub labeling scheme which correctly decodes the distance between any pair of nodes lying at a distance in the range [δ, 5δ/4], using hubs of size at most O(n/δ). 24 4.2 Improved δ-preserving Distance Labeling for Arbitrary Distance For an arbitrary instance of the δ-preserving distance labeling problem, we can now construct a hub set S + (u) by combining the results of Propositions 4 and 3, for large and small scales of distance, respectively. Formally, we put: [ Sδi (u), S + (u) := Sδmax (u) ∪ (24) i=0,1,2,... δi <δmax where in the first part of the expression, the value √ δmax = max{δ, b n/ ln nc} is a suitably chosen threshold parameter, and the hub set Sδmax (u) is constructed following Proposition 4, thus providing a δmax -preserving distance labeling. In the second part of the expression, we take care of smaller distances from the range [δ, δmax ), by applying Proposition 3 over a specifically chosen distance sequence {δi }i , to obtain hub sets Sδi (u), such that each set Sδi (u) intersects with a shortest u − v path, for all nodes v such that d(u, v) ∈ [δi , 5δi /4]. To obtain a coverage of the entire distance range [δ, δmax ), we put δ0 = 12bδ/12c < δ, and choose δi+1 as the largest integer such that 12\|δi+1 and δi+1 < 5δi /4. Since the sequence {δi }i is geometrically increasing, in view of Proposition 3, we obtain the following bound: \|S + (u)\| = O(n/δ + \|Sδmax (u)\|). (25) Now, taking into account the definition of δmax and bounding \|Sδmax (u)\| by Proposition 4, we directly obtain the main result of this Section. Theorem 2. There exists a δ-preserving distance labeling scheme based on hub sets, such that: √ (i) When δ ≤ n/(ln n ln ln n), the hub sets of all nodes are of size O(n/δ), which corresponds to distance labels of size O(n log δ/δ) per node. (ii) For any δ > 0, the hub sets of all nodes are of size O(log δ + n log log δ/δ), which corresponds 2 to distance labels of size O log δ + n log δ log log δ/δ per node. (iii) For any δ > 0, the average size of a hub set, taken over all nodes, is O(n/δ), which corresponds to distance labels of average size O(n log δ/δ) per node. Furthermore, the corresponding labels can be constructed in expected polynomial time. 5 Computing Skeleton Dimension and Distance Labels Discrete skeleton representation. Given a tree T rooted at node u with length function `, a discrete representation of its skeleton T ∗ can be obtained as the sub-tree with edges vw ∈ E(T ) such that Reach T (v) ≥ 21 dT (u, v) equipped with length function `0 defined by `0 (vw) = `(vw) if Reach T (w) ≥ 1 2 0 2 dT (u, w) and ` (vw) = r − dT (u, v) with r = 3 (dT (u, w) + Reach T (w)) otherwise. The idea is that ∗ node w is a leaf of T that corresponds to the point w0 of edge vw in Te that satisfies Reach Te (w0 ) = 1 0 2 dTe (u, w ). To see this, let x be a descendant of w such dT (w, x) = Reach T (w). We then have 2 r = 3 dT (u, x). We thus get Reach Te (w0 ) = dT (v, x) − `0 (vw) = dT (u, x) − r = 31 dT (u, x) = 2r whereas dTe (u, w0 ) = dT (u, v) + `0 (vw) = r. 25 Skeleton dimension computation. Given a tree T , the reach of each node can be computed by a scan of vertices in reverse topological order. Obtaining the discrete skeleton representation is then straightforward. Its width k can be computed by scanning vertices by non-decreasing distance from the root using a priority queue for storing edges containing nodes in Cut r (T ). This can be done in O(n log log k ) time using a dedicated integer priority queue [28]. The skeleton of a graph can thus be simply obtained from an all pair shortest path computation. With integer lengths and dedicated priority queues [28], this can be done in O(nm + n2 log log n) time. We remark that faster computation of tree skeletons could be obtained in practice by using classical heuristics for bounding reach of nodes [22, 24]. The algorithm proposed in [24] alternates partial shortest-path tree computation with introduction of shortcuts to obtain efficiently reach bounds on the graph plus the added shortcuts. The computation of partial trees up to a given radius 2r allows to prune nodes with reach less than r. Shortcuts allow to reduce reach of nodes with degree 2: if a node v has two neighbors u, w, a shortcut uw with length `(uv) + `(vw) is added to by-pass v. The algorithm results in reach bounds on the graph with shortcuts. Reach bounds on the original graph can then be obtained by removing shortcuts in reverse order and updating the reach bound R(v) of a node v by-passed by shortcut uw as R(v) := max(R(v), min(R(u) + `(uv), `(vw) + R(w))) where R(u) and R(w) denote the reach bounds obtained for u and w respectively. A subtree containing the tree skeleton of a node u can then be obtained through a partial Dijkstra from u where we prune nodes whose reach is known to be less than half of the current distance from u. In practice, we believe that this would allow to compute each skeleton tree in time comparable to an RE query. Our labeling algorithm can then be adapted to take the resulting family of trees as input. Distance label computation. Computing the hub set of a tree T is more intricate as we have to emulate the subdivision of each edge of length a into 12a unweighted edges to conform to the analysis of Section 3. For the sake of notation, we number these unweighted edges of the subdivision from 1 to 12a, and let ρi , 1 ≤ i ≤ 12a, denote the associated random number which is generated for each of the edges of the subdivision. Given a sample {ρi \| 1 ≤ i ≤ 12a}, let M denote the set of indices of edges which are prefix minima or suffix minima in the sequence (ρ1 , . . . , ρ12a ). For the purpose of our selection algorithm, we only need to generate set M and the ρi -values associated with i ∈ M . We start by generating those elements of M which are prefix minima. By a slight abuse of notation, to initialise the process, we set i0 = 1 and generate ρ1 uniformly at random in [0, 1], ρ1 := rand(0, 1). For successive j = 0, 1, 2, . . ., we then generate the index ij+1 of the first edge of M after the edge with index ij (which is also the first index of M having value less than ρij ). As ij+1 − ij follows a geometric distribution j k rand(0,1) with parameter p = ρij , this can be done in constant time by setting ij+1 := ij + loglog(1−p) (see e.g. [19]). We then generate ρij+1 uniformly in [0, ρij ]. We generate in this way indices i1 , . . . , ix , until for some x ∈ N we reach the 12a bound (i.e. ix+1 > 12a). We then proceed similarly in reverse order for edges with index greater than ix , to generate those edges with suffix minimal value (note that this time we have to sample values greater than ρix , rather than greater than 0, and adapt all ranges accordingly, for consistency with the choice of prefix minima). In this way, we perform O(log a) constant-time sampling operations per edge of length a, obtaining O(log a) values (together with their positions) per edge, in expectation (and also w.h.p. with respect to a). Since random choices made for all edges of the original graph are independent, by a quick Chernoff bound, the total amortized sampling time over the whole graph is O(m log C), w.h.p., where C denotes the maximum length of an edge. We remark that when constructing a hub set for a subset of nodes, each node only relies on the random choices made in its shortest-path tree, which can be evaluated in time O(n log C), w.h.p. Our selection algorithm of the edge with minimal value in the middle window for a pair u, v will necessarily select an edge that we have generated when the window contains a (real) edge extremity. Each time a virtual unweighted edge is selected as a hub, we indeed select the real edge it belongs to. 26 We also have to manage the special case where the middle window entirely falls inside a long edge. In that case, the long edge is selected as hub. The computation of the hub set is then a matter of scanning the tree by non-decreasing distance r of generated vertices while maintaining a sliding middle window for each branch reaching distance r. Using a balanced binary search tree per window for storing the virtual edges it contains, we obtain the hub set in O(n log C log(n log C)) time. Distance labels can thus be computed in expected O(nm + n2 log C(log n + log log C)) time. Note that each of the n labels can be computed independently (e.g. in parallel) in O(m + n log C(log n + log log C)) time per label, as long as randomness is shared (e.g. using random generators with same seeds). 6 Generalizing the Definition of a Skeleton The definition of a skeleton, and the corresponding notion of skeleton dimension can be generalized in two ways, by using a different distinct distance metric to compute reach of a point in a tree, as well as by modifying the threshold value of reach required to retain a point in the skeleton. Using two metrics. Suppose the graph G is associated with two non-negative length functions, ` and `0 , of its edges. For example, in road networks, one can typically consider travel time `t and geographic distance `d , resulting in time and distance metrics, respectively. Another metric which may be of interest is hop count, corresponding to the constant function `h = 1 (i.e. `h (uv) = 1 for each edge uv). Once a shortest path tree Tu for node u and its geometric realization Teu has been computed according to length function `, distance and reach within Teu can be computed according to `0 . Formally, extending 0 the definition from Section 2.1, the skeleton Tu∗` of Tu is then defined as the subtree of Teu induced 0 0 0 0 by {v ∈ V (Teu ) \| Reach `Te (v) ≥ 21 d`e (u, v)}, where d`e and Reach `Te denote distance and reach with Tu u respect to `0 . 0 The skeleton dimension of G is then k ` Tu u 0 = maxu∈V (G) Width(Tu∗` ). The advantage of this approach is that it sometimes results in smaller skeleton dimension (depending on the choice of metric `0 ), without affecting the correctness of any of the hub labeling schemes designed in this paper. As an example, the respective skeleton dimensions of the 9-th DIMACS New York graph [1] for different choices of metrics turn out to be: k `t = 73, k `d = 66, and k `h = 56, where `t , `d , and `h denote traveltime, geographic-distance, and hop-count length functions, respectively (considering shortest path trees for the metric `t in all three cases). We remark that a similar phenomenon, also taking advantage of two metrics, was observed and used in the reach-pruning approach [24]. Modifying reach threshold. The choice of a reach threshold of 21 in the definition of skeleton is arbitrary. Indeed, for any fixed α > 0, we can define the skeleton Tuα∗ as the subtree of Teu induced by {v ∈ V (Teu ) \| Reach Teu (v) ≥ α dTeu (u, v)}, and the skeleton dimension kα of G is given as kα = maxu∈V (G) Width(Tuα∗ ). The values of skeleton dimension for different values α and β of the reach threshold are related to each other by the following Proposition. Proposition 5. For two constants α < β, the following bounds hold: kβ ≤ kα ≤ kβ k β+1 β/α−1 . Proof. The first relation is immediate since Tuβ∗ is a subtree of Tuα∗ for β ≥ α. The second relation S 0 0 is obtained by observing that Cut r (Tuα∗ ) ⊆ v∈Cut 0 (T β∗ ) Cut r−r0 (Tvβ ∗ ), with r0 = 1+α 1+β and β = u r β+1 . Indeed, for w ∈ Cut r (T α∗ ), we can consider in Teu the point v at distance r0 from u on the β/α−1 u branch leading to w. The reach of w in Teu is then at least r − r0 + Reach Teu (v) ≥ r − r0 + αr ≥ βr0 for 27 0 (1 + α)r ≥ (1 + β)r0 and v thus belongs to Tuβ∗ for r0 = 1+α 1+β . Moreover, w is at distance r − r from v 0 in Tev and has reach at least αr, implying w ∈ V (Tvβ ∗ ) for β 0 ≤ αr 0 , which is the case for β 0 = β+1 r−r when r0 = β/α−1 1+α 1+β . For β = 1 and α < 1, the second relation of the above Proposition gives kα ≤ k1 k 2α 1−α , which also implies that k1 ≤ kα ≤ k12 for α ≥ 1/3. More generally, we can derive the following bounds by repeatedly applying Proposition 5 for β = 1: dlog(1+1/α)e k1 ≤ kα ≤ k1 for α < 1. This shows that for a given graph, skeleton dimension kα grows at most polynomially with α1 . Naturally, one can also apply both of the above-described generalizations together, obtaining a new 0 skeleton dimension parameter kα` with reach metric `0 and reach threshold α. All the results of the paper about hub labelings and their computation in graphs with low skeleton dimension can be easily 0 0 generalized to use kα` instead of k, as long as α < 1 (ensuring that any two skeleton trees Tu` ∗α and 0 ` Tv` ∗α share a constant fraction of the u − v shortest path). The particular choice of k = k1/2 was made ` with the objective of clarity, and also on account of the simple relationship between k1/2 and highway dimension. 7 Conclusion In this paper, we have proposed skeleton dimension as a measure of the network’s amenability to shortest path schemes based on hub/transit nodes. We intend it as a parameter which is easy to describe and can be computed efficiently. Computations of hub sets based on skeleton dimension allow each node to individually and efficiently define its own hub set, subject only to a universal choice of random id-s. Such a construction is always correct, and gives small hub sets w.h.p. We remark that in a weighted network each node can compute its own appropriate labeling in O(m + n log C(log n + log log C)) time, where C is the length of the longest integer weight in the network. The definition of hub sets, and the obtained bounds on their size, hold both for undirected and directed graphs. For directed graphs, skeleton dimension appears to be a parameter which is more directly usable than highway dimension. Possible extensions of skeleton dimension, discussed in Section 6, include variants of skeleton dimension with other values of reach threshold, as well as skeleton dimension defined using two separate distance metrics in the graph: one corresponding to the needs of the shortest path queries (used to construct shortest path trees), and another, potentially independent metric used internally in the computation of hub labelings, chosen so as to empirically minimize the width of the skeleton. When studying average-case parameters of a network, the integrated skeleton dimension given by (2) (as well as its natural generalizations to weighted graphs) appear to be a natural parameter, which may be related to that of average highway dimension [3]. We could also use the integrated skeleton dimension averaged over all nodes to get an even more accurate bound on average label size. Finally, we remark on the interplay between skeleton and highway dimension. Skeleton dimension is always not greater than geometric highway dimension. We have also shown a clear case of separation in a weighted Manhattan-type network, where skeleton dimension is asymptotically much smaller than (geometric) highway dimension. We remark that skeleton dimension appears particularly worthy of further theoretical study in the context of scale-free models of random graphs (cf. e.g. [9] for a discussion in the context of highway dimension and reach). For geometric percolation graphs, skeleton dimension displays a close link with the coalescence exponent for geodesics. Consequently, it may be easier to show rigorous theoretical bounds for skeleton dimension than for highway dimension. 28 Acknowledgment The authors thank Przemek Uznański and Olivier Marty for inspiring discussions on closely related problems. We also thank PU and Zuzanna Kosowska-Stamirowska for their help with the figures. References [1] 9th DIMACS Implementation Challenge: The Shortest Path Problem, 2006. [2] Ittai Abraham, Daniel Delling, Amos Fiat, Andrew Goldberg, and Renato Werneck. Highway dimension and provably efficient shortest path algorithms. Technical report, September 2013. [3] Ittai Abraham, Daniel Delling, Amos Fiat, Andrew V. Goldberg, and Renato F. Werneck. VCdimension and shortest path algorithms. In ICALP, volume 6755 of Lecture Notes in Computer Science, pages 690–699. Springer, 2011. [4] Ittai Abraham, Daniel Delling, Andrew V. Goldberg, and Renato F. Werneck. A hub-based labeling algorithm for shortest paths in road networks. In SEA, volume 6630 of Lecture Notes in Computer Science, pages 230–241. Springer, 2011. [5] Ittai Abraham, Daniel Delling, Andrew V. Goldberg, and Renato F. Werneck. Hierarchical hub labelings for shortest paths. In Proceedings of the 20th Annual European Conference on Algorithms, ESA’12, pages 24–35, Berlin, Heidelberg, 2012. Springer-Verlag. [6] Ittai Abraham, Daniel Delling, Andrew V. Goldberg, and Renato F. Werneck. Hierarchical hub labelings for shortest paths. In ESA, volume 7501 of Lecture Notes in Computer Science, pages 24–35. Springer, 2012. [7] Ittai Abraham, Amos Fiat, Andrew V. Goldberg, and Renato F. Werneck. Highway dimension, shortest paths, and provably efficient algorithms. In Moses Charikar, editor, Proceedings of the Twenty-First Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2010, Austin, Texas, USA, January 17-19, 2010, pages 782–793. SIAM, 2010. [8] Ittai Abraham and Cyril Gavoille. On approximate distance labels and routing schemes with affine stretch. In In International Symposium on Distributed Computing (DISC), pages 404–415, 2011. [9] David Aldous and Karthik Ganesan. True scale-invariant random spatial networks. Proceedings of the National Academy of Sciences, 110(22):8782–8785, 2013. [10] Stephen Alstrup, Søren Dahlgaard, Mathias Bæk Tejs Knudsen, and Ely Porat. Sublinear distance labeling. In Piotr Sankowski and Christos D. Zaroliagis, editors, 24th Annual European Symposium on Algorithms, ESA 2016, August 22-24, 2016, Aarhus, Denmark, volume 57 of LIPIcs, pages 5:1– 5:15. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2016. [11] Stephen Alstrup, Cyril Gavoille, Esben Bistrup Halvorsen, and Holger Petersen. Simpler, faster and shorter labels for distances in graphs. In Robert Krauthgamer, editor, Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016, pages 338–350. SIAM, 2016. [12] Maxim A. Babenko, Andrew V. Goldberg, Anupam Gupta, and Viswanath Nagarajan. Algorithms for hub label optimization. In Fedor V. Fomin, Rusins Freivalds, Marta Z. Kwiatkowska, and David Peleg, editors, Automata, Languages, and Programming - 40th International Colloquium, ICALP 2013, Riga, Latvia, July 8-12, 2013, Proceedings, Part I, volume 7965 of Lecture Notes in Computer Science, pages 69–80. Springer, 2013. 29 [13] H. Bast, Stefan Funke, Domagoj Matijevic, Peter Sanders, and Dominik Schultes. In transit to constant time shortest-path queries in road networks. In ALENEX. SIAM, 2007. [14] Holger Bast, Stefan Funke, Peter Sanders, and Dominik Schultes. Fast routing in road networks with transit nodes. Science, 316(5824):566–566, 2007. [15] Reinhard Bauer and Daniel Delling. SHARC: Fast and robust unidirectional routing. J. Exp. Algorithmics, 14:4:2.4–4:2.29, January 2010. [16] Béla Bollobás, Don Coppersmith, and Michael Elkin. Sparse distance preservers and additive spanners. SIAM Journal on Discrete Mathematics, 19(4):1029–1055, 2005. [17] Fan R. K. Chung and Lincoln Lu. Survey: Concentration inequalities and martingale inequalities: A survey. Internet Mathematics, 3(1):79–127, 2006. [18] Edith Cohen, Eran Halperin, Haim Kaplan, and Uri Zwick. Reachability and distance queries via 2-hop labels. SIAM J. Comput., 32(5):1338–1355, May 2003. [19] Luc Devroye. Non-Uniform Random Variate Generation. Springer-Verlag, 1986. [20] Andreas Emil Feldmann, Wai Shing Fung, Jochen Könemann, and Ian Post. A 1 + -embedding of low highway dimension graphs into bounded treewidth graphs. In ICALP 2015, volume 9134 of Lecture Notes in Computer Science, pages 469–480. Springer, 2015. [21] Cyril Gavoille, David Peleg, Stéphane Pérennes, and Ran Raz. Distance labeling in graphs. J. Algorithms, 53(1):85–112, October 2004. [22] Andrew V. Goldberg, Haim Kaplan, and Renato F. Werneck. Reach for A*: Efficient point-to-point shortest path algorithms. In ALENEX, pages 129–143. SIAM, 2006. [23] R.L. Graham and H.O. Pollak. On embedding graphs in squashed cubes. In Y. Alavi, D.R. Lick, and A.T. White, editors, Graph Theory and Applications, volume 303 of Lecture Notes in Mathematics, pages 99–110. Springer Berlin Heidelberg, 1972. [24] Ronald J. Gutman. Reach-based routing: A new approach to shortest path algorithms optimized for road networks. In ALENEX/ANALCO, pages 100–111. SIAM, 2004. [25] Ekkehard Köhler, Rolf H. Möhring, and Heiko Schilling. Fast point-to-point shortest path computations with arc-flags. In 9th DIMACS Implementation Challenge, 2006. [26] Colin McDiarmid. Concentration. In Michel Habib, Colin McDiarmid, Jorge Ramirez-Alfonsin, and Bruce Reed, editors, Probabilistic Methods for Algorithmic Discrete Mathematics, pages 195– 248. Springer Berlin Heidelberg, 1998. [27] Igor Nitto and Rossano Venturini. On compact representations of all-pairs-shortest-path-distance matrices. In Paolo Ferragina and Gad M. Landau, editors, Combinatorial Pattern Matching, 19th Annual Symposium, CPM 2008, Pisa, Italy, June 18-20, 2008, Proceedings, volume 5029 of Lecture Notes in Computer Science, pages 166–177. Springer, 2008. [28] Mikkel Thorup. Integer priority queues with decrease key in constant time and the single source shortest paths problem. Journal of Computer and System Sciences, 69(3):330 – 353, 2004. [29] Oren Weimann and David Peleg. 111(14):671–673, 2011. A note on exact distance labeling. 30 Inf. Process. Lett.,	8cs.DS
Proceedings of Machine Learning Research vol 75:1–22, 2018 Minimax Distribution Estimation in Wasserstein Distance Shashank Singh SSS 1@ CS . CMU . EDU Barnabás Póczos BAPOCZOS @ CS . CMU . EDU arXiv:1802.08855v1 [math.ST] 24 Feb 2018 Machine Learning Department, Carnegie Mellon University, Pittsburgh, PA 15213, USA Abstract The Wasserstein metric is an important measure of distance between probability distributions, with several applications in machine learning, statistics, probability theory, and data analysis. In this paper, we upper and lower bound minimax rates for the problem of estimating a probability distribution under Wasserstein loss, in terms of metric properties, such as covering and packing numbers, of the underlying sample space. Keywords: Wasserstein Distance; Density Estimation; Minimax Theory; Covering Number; Packing Number 1. Introduction The Wasserstein metric is an important measure of distance between probability distributions, based on the cost of transforming either distribution into the other through mass transport, under a metric on the sample space. Originating in the optimal transport literature,1 the Wasserstein metric has, owing to its intuitive and general nature, been utilized in such diverse areas as probability theory and statistics, economics, image processing, text mining, robust optimization, and physics (Villani, 2008; Fournier and Guillin, 2015; Esfahani and Kuhn, 2015; Gao and Kleywegt, 2016). In the analysis of image data, the Wasserstein metric has been used for various tasks such as texture classification and face recognition (Sandler and Lindenbaum, 2011), reflectance interpolation, color transfer, and geometry processing (Solomon et al., 2015), image retrieval (Rubner et al., 2000), and image segmentation (Ni et al., 2009), and, in the analysis of text data, for tasks such as document classification (Kusner et al., 2015) and machine translation (Zhang et al., 2016). In contrast to a number of other popular notions of distance2 between probability distributions, such as Lp distances or Kullback-Leibler and other f -divergences (Morimoto, 1963; Csiszár, 1964; Ali and Silvey, 1966), which require distributions to be absolutely continuous with respect to each other or to a base measure, Wasserstein distance is well-defined between any pair of probability distributions over a sample space equipped with a metric.3 As a particularly important consequence, Wasserstein distances between discrete (e.g., empirical) distributions and continuous distributions are well-defined and informative (i.e., often finite). Partly for this reason, Wasserstein distances have been widely used in probability theory to prove central limit and related approximation theorems (Rüschendorf, 1985; Johnson et al., 2005; 1. The Wasserstein metric has been variously attributed to Monge, Kantorovich, Rubinstein, Gini, Mallows, and others; see Chapter 3 of Villani (2008) for detailed history. 2. We use the term distance to refer both to proper metrics as well as to similar notions, such as information divergences, that fail to satisfy symmetry or the triangle inequality. 3. For this reason, we use “distribution estimation” in this paper, rather than the more popular “density estimation”. c 2018 S. Singh & B. Póczos. D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Chatterjee et al., 2008; Rio et al., 2009, 2011; Chen et al., 2010; Reitzner et al., 2013). Within machine learning and statistics, this same property has inspired a class of so-called minimum Wasserstein distance estimates (del Barrio et al., 1999, 2003; Bassetti et al., 2006; Bernton et al., 2017) of distributions, ranging from exponential distributions (Baı́llo et al., 2016) to more exotic models such as restricted Boltzmann machines (RBMs) (Montavon et al., 2016) and generative adversarial networks (GANs) (Arjovsky et al., 2017). k-means clustering (and, similarly, k-medians clustering) also fall in this class of estimators, where the hypothesis class is taken to be discrete distributions supported on at most k points; more flexible multi-level clustering algorithms can also be expressed in this way, using a more elaborate hypothesis class (Ho et al., 2017). PCA can also be expressed and generalized to manifolds using Wasserstein distance minimization (Boissard et al., 2015). In principle, these estimators are equivalent to empirical risk minimization, taking advantage of the fact that Wasserstein distances between the empirical distribution and distributions in the relevant hypothesis class are well-defined and finite. Moreover, these estimates often perform well in practice because they are free of both tuning parameters and strong distributional assumptions. In this paper, we study, in a minimax sense, the problem of estimating a distribution using Wasserstein distance as a loss function. We make minimal assumptions, and some examples of the varied metric spaces to which our results apply are given in Section 7. Specifically, we assume that the sample space is a totally bounded metric space, but make no assumptions at all on the distribution. We then prove bounds on the minimax convergence rates of distribution estimation, utilizing covering numbers of the sample space for upper bounds and packing numbers for lower bounds. It may at first be surprising that positive results can be obtained under such mild assumptions; this highlights that the Wasserstein metric is quite a weak metric (see our Lemma 8 and the subsequent remark for some detailed discussion of this). Moreover, our results imply that, without further assumptions on the population distribution, the empirical distribution is often minimax rate-optimal. O RGANIZATION The remainder of this paper is organized as follows. Section 2 provides notation required to formally state both the problem of interest and our results, while Section 3 reviews previous work studying convergence of distributions in Wasserstein distance. Section 4 provides several lemmas on which the proofs of our main results rely; all lemmas are proven in Appendix A. Sections 5 and 6 respectively contain our main upper and lower bound results, along with their proofs. Finally, in Section 7, we apply our upper and lower bounds to identify minimax convergence rates in a number of concrete examples. Section 8 concludes with a summary of our contributions and suggested avenues for future work. 2. Notation and Problem Setting For any integer n ∈ N, [n] = {1, 2, ..., n} denotes the set of the first n positive integers. For sequences {an }n∈N and {bn }n∈N of non-negative real numbers, an . bn and, equivalently bn & an , are used to indicate the existence of a constant C > 0 such that lim supn→∞ abnn ≤ C. an ≍ bn indicates an . bn and an & bn simultaneously. 2 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE 2.1. Problem Setting For the remainder of this paper, fix a metric space (Ω, ρ), over which Σ denotes the Borel σ-algebra, and let P denote the family of all Borel probability distributions on Ω. The main object of study in this paper is the Wasserstein distance on P, defined as follows: Definition 1 (r-Wasserstein Distance) Given two Borel probability distributions P and Q over Ω and r ∈ [1, ∞), the r-Wasserstein distance Wr (P, Q) ∈ [0, ∞] between P and Q is defined by !1/r Wr (P, Q) := inf E µ∈Π(P,Q) (X,Y )∼µ [ρr (X, Y )] , where Π(P, Q) denotes all couplings between X ∼ P and Y ∼ Q; that is, Π(P, Q) := µ : Σ2 → [0, 1] for all A ∈ Σ, µ(A × Ω) = P (A) and µ(Ω × A) = Q(A) , is the set of joint probability measures over Ω × Ω with marginals P and Q. Wr is a proper metric over P; namely, it is symmetric in its arguments and satisfies the triangle inequality. Intuitively, Wr (P, Q) quantifies the r-weighted total cost of transforming mass distributed according to P to be distributed according to Q, where the cost of moving a unit mass from x ∈ Ω to y ∈ Ω is ρ(x, y). Note that Wr (P, Q) is sometimes defined in terms of equivalent (e.g., dual) formulations (especially for r = 1); however, they will not be necessary in this paper. We now give a formal statement of the problem studied in this paper: Formal Problem Statement: Suppose (Ω, ρ) is a metric space. Suppose P is an unknown IID Borel probability distribution on Ω, from which we observe n IID samples X1 , ..., Xn ∼ P . We are interested in studying the minimax rates at which P can be estimated from X1 , ..., Xn , in terms of the (r th power of the) r-Wasserstein loss. Specifically, we are interested in upper and lower bounding the quantity i h Wrr P, Pb , inf sup E Pb P ∈P X ,...,XnIID ∼ P 1 where the infimum is taken over all estimators Pb of P (i.e., all functions Pb : Ωn → P of the data). 2.2. Definitions for Stating our Results Here, we give notation and definitions needed to state our theoretical results in Sections 5 and 6 and intermediate lemmas in Section 4. Ω Let 2Ω denote the power set of Ω. Let S ⊆ 22 denote the family of all Borel partitions of Ω: ( ) [ S := S ⊆ Σ : Ω ⊆ S and ∀S, T ∈ S, S ∩ T = ∅ . S∈S We first define some metric notions that will later be useful for bounding Wasserstein distances: Definition 2 (Diameter and Separation of a Set, Resolution of a Partition) For any set S ⊆ Ω, the diameter Diam(S) of S is defined by Diam(S) := supx,y∈S ρ(x, y), and the separation Sep(S) of S is defined by Sep(S) := inf x6=y∈S ρ(x, y). If S ∈ S is a partition of Ω, then the resolution Res(S) of S is defined by Res(S) := supS∈S Diam(S). 3 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Definition 3 (Refinement of a Partition, Nested Partitions) Suppose S, T ∈ S are partitions of Ω. T is said to be a refinement of S if, for every T ∈ T , there exists S ∈ S with T ⊆ S. A sequence {Sk }k∈[K] of partitions is called nested if, for each k ∈ [K]\{K}, Sk+1 is a refinement of Sk , We now define the covering and packing number of a metric space, which are classic and widely used measures of the size or complexity of a metric space (Dudley, 1967; Haussler, 1995; Zhou, 2002; Zhang, 2002). Our main convergence results will be stated in terms of these quantities, as well as the packing radius, which acts, approximately, as the inverse of the packing number. Definition 4 (Covering Number, Packing Number, and Packing Radius of a Metric Space) The covering number N : (0, ∞) → N of (Ω, ρ) is defined for all ε by N (ε) := min {\|S\| : S ∈ S and Res(S) ≤ ε} . The packing number M : (0, ∞) → N of (Ω, ρ) is defined for all ε by M (ε) := max {\|S\| : S ⊆ Ω and Sep(S) ≥ ε} . Finally, the packing radius R : N → [0, ∞] is defined for all n ∈ N by R(n) := sup{Sep(S) : S ⊆ Ω and \|S\| ≥ n}. Sometimes, we use the covering or packing number of a metric space, say (Θ, τ ), other than (Ω, ρ); in such cases, we write N (Θ; τ ; ε) or M (Θ; τ ; ε) rather than N (ε) or M (ε), respectively. For specific ε > 0, we will also refer to N (Θ; τ ; ε) as the ε-covering number of (Θ, τ ). Remark 5 The covering and packing numbers of a metric space are closely related. In particular, for any ε > 0, we always have M (ε) ≤ N (ε) ≤ M (ε/2). (1) The packing number and packing radius also have a close approximate inverse relationship. In particular, for any ε > 0, we always have R(M (ε)) ≥ ε. (2) However, we may have R(M (ε)) > ε, and it need not, in general, be the case that, for all n ∈ N, M (R(n)) ≥ n or, for all n ∈ N, M (R(n)) ≤ n. Remark 6 We defined the covering number slightly differently from usual (using partitions rather than covers). However, the given definition is equivalentSto the usual definition, since (a) any partition is itself a cover (i.e., a set C ⊆ 2Ω such that Ω ⊆ C∈C C), and (b), for any countable cover C := {C1 , C2 , ...} ⊆ 2Ω ,Sthere exists a partition S ∈ S with \|S\| ≤ \|C\| and each Si ⊆ Ci , defined recursively by Si := Ci \ i−1 j=1 Si . S is often called the disjointification of C. 4 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE 3. Related Work A long line of work (Dudley, 1969; Ajtai et al., 1984; Dereich et al., 2013; Boissard et al., 2014; Fournier and Guillin, 2015; Weed and Bach, 2017) has studied distribution estimation in Wasserstein distance, with all the above papers focusing on analyzing convergence of the empirical distribution to the population distribution in Wasserstein distance. In terms of upper bounds, the most general and tight upper bounds are those of Weed and Bach (2017), who provide results in general metric spaces, in terms of the covering numbers of those spaces. Their main results are expressed in terms of a particular notion of dimension, which they call the Wasserstein dimension s because it allows them to derive convergence rates of order n−1/s (matching the n−1/D rate achieved on the unit cube [0, 1]D ). Our general upper bounds as a function of covering number are somewhat easier to apply than those of Weed and Bach (2017), and we use these in Section 7 to derive convergence rates even in high-dimensional cases (where D → ∞ as n → ∞) or infinite dimensional cases. To the best of our knowledge, there has been no work on deriving minimax lower bounds for distribution estimation under Wasserstein loss. As noted above, most previous works have focused on studying convergence rate of the empirical distribution to the true distribution in Wasserstein distance. For this rate, several lower bounds have been established, matching known upper bounds in many quite general cases. However, many distribution estimators besides the empirical distribution can be considered. For example, it is tempting (especially given the infinite dimensionality of the distribution to be estimated) to try to reduce variance by techniques such as smoothing or importance sampling (Bucklew, 2013). However, our lower bound results, given in Section 6, imply that the empirical distribution is already minimax optimal, up to constant factors, in many cases. 4. Preliminary Lemmas We begin by providing a few basic lemmas; these lemmas are not fundamentally novel, but they will be used in the subsequent proofs of our main upper and lower bounds, and also help provide intuition for the behavior of the Wasserstein metric and its connections to other metrics between probability distributions. Our first lemma relates Wasserstein distance to the notion of resolution of a partition. Lemma 7 Suppose S ∈ S is a countable Borel partition of Ω. Let P and Q be Borel probability measures such that, for every S ∈ S, P (S) = Q(S). Then, for any r ≥ 1, Wr (P, Q) ≤ Res(S). Our next lemma gives simple lower and upper bounds on the Wasserstein distance between distributions supported on a countable subset X ⊆ Ω, in terms of Diam(X ) and Sep(X ). Since our main results will utilize coverings and packings to approximate Ω by finite sets, this lemma will provide a first step towards approximating (in Wasserstein distance) distributions on Ω by distributions on these finite sets. Indeed, the lower bound in Inequality (3) will suffice to prove our lower bounds, although a tighter upper bound, based on the upper bound in (3), will be necessary to obtain tight upper bounds. Lemma 8 Suppose (Ω, ρ) is a metric space, and suppose P and Q are Borel probability distributions on Ω with countable support; i.e., there exists a countable set X ⊆ Ω with P (X ) = Q(X ) = 1. Then, for any r ≥ 1, X X \|P ({x}) − Q({x})\| . (3) \|P ({x}) − Q({x})\| ≤ Wrr (P, Q) ≤ (Diam(X ))r (Sep(X ))r x∈X x∈X 5 D ISTRIBUTION E STIMATION Remark 9 Recall that the term P IN WASSERSTEIN D ISTANCE \|P ({x}) − Q({x})\| in Inequality (3) is the L1 distance X kp − qk1 := \|p(x) − q(x)\| x∈X x∈X between the densities p and q of P and Q with respect to the counting measure on X , and that this same quantity is twice the total variation distance T V (P, Q) := sup \|P (A) − Q(A)\| . A⊆Ω Hence, Lemma 8 can be equivalently written as Sep(Ω) (kp − qk1 )1/r ≤ Wr (P, Q) ≤ Diam(Ω) (kp − qk1 )1/r and as Sep(Ω) (2T V (P, Q))1/r ≤ Wr (P, Q) ≤ Diam(Ω) (2T V (P, Q))1/r , bounding the r-Wasserstein distance in terms of the L1 and total variation distance. As noted in Example 1, equality holds in (3) precisely when ρ is the unit discrete metric given by ρ(x, y) = 1{x6=y} for all x, y ∈ Ω. On metric spaces that are discrete (i.e., when Sep(Ω) > 0), the Wasserstein metric is (topologically) at least as strong as the total variation metric (and the L1 metric, when it is well-defined), in that convergence in Wasserstein metric implies convergence in total variation (and L1 , respectively). On the other hand, on bounded metric spaces, the converse is true. In either of these cases, rates of convergence may differ between metrics, although, in metric spaces that are both discrete and bounded (e.g., any finite space), we have Wr ≍ T V 1/r . While Lemma 8 gave a simple upper bound on the Wasserstein distance, the factor of Diam(Ω) turns out to be too large to obtain tight rates for a number of cases of interest (such as the Ddimensional unit cube Ω = [0, 1]D , discussed in Example 2). The following lemma gives a tighter upper bound, based on a hierarchy of nested partitions of Ω; this allows us to obtain tighter bounds (than Diam(Ω)) on the distance that mass must be transported between P and Q. Note that, when K = 1, Lemma 10 reduces to a trivial combination of Lemmas 7 and 8; indeed, these lemmas are the starting point for proving Lemma 10 by induction on K. Note that the idea of such a “multi-resolution” upper bound has been utilized extensively before, and numerous versions have been proven before (see, e.g., Fact 6 of Do Ba et al. (2011), Lemma 6 of Fournier and Guillin (2015), or Proposition 1 of Weed and Bach (2017)). However, most of these versions have been specific to Euclidean space; to the best of our knowledge, only Proposition 1 of Weed and Bach (2017) applies to general metric spaces. Lemma 10 Let K be a positive integer. Suppose {Sk }k∈[K] is a nested sequence of countable Borel partitions of (Ω, ρ), with S0 = Ω. Then, for any r ≥ 1 and Borel probability measures P and Q on Ω,   K X X Wrr (P, Q) ≤ (Res(SK ))r + \|P (S) − Q(S)\| . (Res(Sk−1 ))r  (4) S∈Sk k=1 6 D ISTRIBUTION E STIMATION WASSERSTEIN D ISTANCE IN Lemma 10 requires a sequence of partitions of Ω that is not only multi-resolution but also nested. While the ε-covering number implies the existence of small partitions with small resolution, these partitions need not be nested as ε becomes small. For this reason, we now give a technical lemma that, given any sequence of partitions, constructs a nested sequence of partitions of the same cardinality, with only a small increase in resolution. Lemma 11 Suppose S and T are partitions of (Ω, ρ), and suppose S is countable. Then, there exists a partition S ′ of (Ω, ρ) such that: a) \|S ′ \| ≤ \|S\|. b) Res(S ′ ) ≤ Res(S) + 2 Res(T ). c) T is a refinement of S ′ 5. Upper Bounds In this section, we utilize the covering number bounds from the previous section to develop error bounds for density estimation using the empirical distribution under Wasserstein loss. 5.1. Expectation Bounds Here, we use Lemmas 10 and 11 to prove bounds on the expected Wasserstein distance between the empirical and true distributions. We begin with a simple technical lemma bounding the expected L1 deviation of a multinomial random variable from its mean. Lemma 12 Suppose (X1 , ..., XK ) ∼ Multinomial(n, p1 , ..., pK ). Let Z := kX − pk1 = Then, E [Z/n] ≤ p K X k=1 \|Xk − pk \| . 21K/n. Theorem 13 Let (Ω, ρ) be a metric space on which P is a Borel probability measure. Let Pb denote IID the empirical distribution of n IID samples X1 , ..., Xn ∼ P , give by n 1X Pb(S) := 1{Xi ∈S} , n i=1 ∀S ∈ Σ. Then, for any sequence {εk }k∈[K] ∈ (0, ∞)K with ε0 = Diam(Ω),  r r K K i X X p 21 r r  2j−k εj  N (εk ). E Wr (P, Pb) ≤ εK + n h k=1 7 j=k−1 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Proof By recursively applying Lemma 11, there exists a sequence {Sk }k∈[K] of partitions of (Ω, ρ) satisfying the following conditions: 1. for each k ∈ [K], \|Sk \| = N (εk ). 2. for each k ∈ [K], Res(Sk ) ≤ K X 2j−k εj . j=k 3. {Sk }k∈[K] is nested. Note that, for any k ∈ [K], the vector nPb(S) (indexed by S ∈ Sk ) follows an n-multinomial distribution over \|Sk \| categories, with means given by P (S); i.e., (nPb(S1 ), ..., nPb(Sk )) ∼ Multinomial(n, P (S1 ), ..., P (Sk )). Thus, by Lemma 12, for each k ∈ [K],   r r X 21\|S \| 21N (εk ) k P (S) − Pb(S)  ≤ = . E n n S∈Sk Thus, by Lemma 10,  r r E [Wr (P, Q)] ≤ E εK + ≤ εrK + K X k=1 K X k=1   K X j=k r  2j−k εj   X S∈Sk  \|P (S) − Q(S)\|  r   K X X  2j−k εj  E  \|P (S) − Q(S)\| S∈Sk j=k  r r K K X X p 21  N (εk ) 2j−k εj  ≤ εrK + n k=1 j=k 6. Lower Bounds In this section, we prove minimax lower bounds (over the family P of all densities on the metric space (Ω, ρ)) for density estimation in Wasserstein distance (that is, the quantity i h (5) Wrr P, Pb , inf sup E Pb P ∈P X1 ,...,Xn IID ∼ P where the infimum is over all estimators Pb of P (i.e., all functions Pb : Ωn → P)). Our bounds are in terms of the packing radius R of (Ω, ρ). 8 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Theorem 14 Suppose P ∈ P is an unknown Borel probability measure on (Ω, ρ), and suppose we IID observe n IID samples X1 , ..., Xn ∼ P . Then, there exists a universal constant CHJW > 0 such that r h i k−1 r r Wr (P, Pb) ≥ CHJW sup R (k) inf sup . E n Pb P ∈P X ,...,X IID k∈[n] ∼ P 1 n Proof Our proof is based on a reduction to a minimax lower bound for estimating the mean parameter vector of a multinomial distribution in L1 norm. Specifically, we will use Corollary 4 of Han et al. (2015), which implies the existence of a universal constant CHJW > 0 (independent of n) such that, for any k ≤ n, r k−1 , p(X1 , ..., Xn ) − pk1 ] ≥ CHJW inf sup E [kb n pb p∈∆k X1 ,...,Xn where ∆k :=    p ∈ [0, 1]k : k X j=1   pj = 1  denotes the standard (k − 1)-dimensional probability simplex, X1 , ..., Xn ∈ [k] are distributed IID according to a multinomial distribution with parameter vector p ∈ ∆k , and the infimum is taken over all estimators pb of p (i.e., all functions pb : [k]n → ∆k ). For some k ≤ n, let PD denote the class of discrete distributions over an R(k)-packing D of (Ω, ρ) with \|D\| = k. Then, Corollary 4 of Han et al. (2015) implies that r i h \|D\| − 1 kPb (X1 , ..., Xn ) − P k1 ≥ CHJW , inf sup E n Pb p∈PD X1 ,...,Xn where the infimum is taken over all estimators Pb of P (i.e., all functions Pb : Ωn → P). Recalling Lemma 8 (which implies (Sep(D))r kPb − P k1 ≤ Wrr (P, Pb)), we have i h h i inf sup E Wrr (Pb, P ) ≥ (Sep(D))r inf sup E kPb − P k1 Pb P ∈P Pb P ∈P i h ≥ (Sep(D))r inf sup E kPb − P k1 Pb P ∈PD r \|D\| − 1 r ≥ CHJW (Sep(D)) n r k−1 ≥ CHJW Rr (k) . n The theorem follows by taking the supremum over k ≤ n on both sides. 9 (6) D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE 7. Example Applications We conclude by exploring applications of our results for several metric spaces (Ω, ρ). In each of the following examples, P is an unknown Borel probability measure over the specified Ω, from which we observe n IID samples. For upper bounds, Pb denotes the empirical distribution of these samples. Example 1 (Finite Space) Consider the case where Ω is a finite set, over which ρ is the discrete metric given, for some δ > 0, by ρ(x, y) = δ1{x=y} , x, y ∈ Ω. for all Then, for any ε ∈ (0, δ), the covering number is N (ε) = \|Ω\|. Thus, setting K = 1 and sending ε1 → 0 in Theorem 13 gives r h i 21\|Ω\| r r b . E Wr (P, P ) ≤ δ n On the other hand, for any k ≤ \|Ω\|, R(k) = δ, while, for k > \|Ω\|, R(k) = 0. Thus, setting k = \|Ω\| in Theorem 14 gives r h i \|Ω\| − 1 r r b . Wr (P, P ) ≥ CHJW δ inf sup E n Pb P ∈P X ,...,X IID ∼ P 1 n Example 2 (Unit Cube, Euclidean Metric) Consider the case where Ω = [0, 1]D is the unit cube and ρ is the Euclidean metric. Suppose P is a Borel probability measure on (Ω, ρ), and let Pb denote the empirical distribution of n IID samples from P . Setting εk := 2−2k and recalling that N (ε) ≤ (2D)D ε−D , one can show that Theorem 13 gives r K h i 21(2D)D X k(D−2r) r −2Kr r+1 b 2 W (P, P ) ≤ 2 + 2 E r n k=1 If D = 2r, then this gives E and setting K ≍ 1 4r h i Wrr (P, Pb ) −2Kr ≤2 r+1 +2 r 21(2D)D K, n h i log n gives the rate E Wrr (P, Pb) . n−1/2 log n. Otherwise, if D 6= 2r, then h i r 21(2D)D 2K(D−2r) − 1 n 2D−2r − 1 i h For 2r > D, sending K → ∞ gives a rate of E Wrr (P, Pb) . n−1/2 . For 2r < D, setting i h r 1 K ≍ 2D log n gives a rate of E Wrr (P, Pb) . n− D . To summarize E Wrr (P, Pb) −2Kr ≤2 D+1 +2  −1/2 i  n r n−1/2 log n E Wr (P, Pb) .  −r/D n h 10 if 2r > D if 2r = D if 2r < D D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE (reproducing results of Fournier and Guillin (2015)). On the √ other hand, it is easy to check that the packing radius R satisfies R(n) ≥ n−1/D and R(2) ≥ D. Thus, applying Theorem 14 with k = n and k = 2, we have o h i n inf sup E Wrr (Pb, P ) ≥ CHJW max (n + 1)−r/D , D r/2 n−1/2 . Pb P ∈P Combining these upper and lower bounds gives that the minimax rate for density estimation over P in the loss Wrr is h i n−1/2 if 2r > D r b inf sup E Wr (P , P ) ≍ n−r/D if 2r < D Pb P ∈P (except for the case 2r = D, when our upper and lower bounds are separated by a factor of log n). Moreover, this rate is achieved by using the empirical distribution to estimate P . We now consider a high dimensional problem, where the size of Ω can grow with n. Example 3 (Binary Hypercube, Hamming Metric) Suppose Ω = {0, 1}D is the P D-dimensional binary hypercube, and suppose ρ is the Hamming metric given by ρδ (x, y) = δ D j=1 1{xi 6=yi } . Then, the covering number can easily be upper bounded for ε ∈ [D] by N (ε) ≤ 2D−ε . Thus, setting K = 1 and sending ε1 → 0 in Theorem 13 gives r h i 21 r r D . E Wr (P, Pb ) ≤ D 2 n The famous Varshamov-Gilbert bound (see, e.g., Lemma 25.12 of Rigollet (2015)) can be interpreted as the statement that, for any γ ∈ (0, 1/2), R ⌊eγ D ⌋ ≥ 21 − γ Dδ. Thus, setting p 2 k = ⌊eγ D ⌋ (so that k ≤ n as long as γ ≤ (1/D) log n) in Theorem 14 and applying the Varshamov-Gilbert bound gives s r h i 1 ⌊eγ 2 D ⌋ − 1 − γ Dδ . Wrr (P, Pb) ≥ CHJW sup inf sup E q 2 n Pb P ∈P X ,...,X IID log n 1 ∼ P 1 n γ∈ 0,min 2 , D To summarize, in the high-dimensional case where D, n → ∞, consistent estimation is possible if D log D ∈ o(log n) and only if D ∈ o(log n). Finally, we consider distributions over an infinite dimensional space of smooth functions. Example 4 (Hölder Ball, L∞ Metric) Suppose that, for some α ∈ (0, 1], Ω := f [0, 1]D → [−1, 1] ∀x, y ∈ [0, 1]D , \|f (x) − f (y)\| ≤ kx − ykα2 is the class of unit α-Hölder functions on the unit cube and ρ is the L∞ -metric given by ρ(f, g) = sup \|f (x) − g(x)\|, x∈[0,1]D 11 for all f, g ∈ Ω. D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE The covering and packing numbers of (Ω, ρ) are well-known to be of order exp ε−D/α (DeVore and Lorentz, 1993); specifically, there exist positive constants 0 < c1 < c2 such that, for all ε ∈ (0, 1), c1 exp ε−D/α ≤ N (ε) ≤ M (ε) ≤ c2 exp ε−D/α . It also follows from Inequality (2) that, for all k ∈ N, R(k) ≥ 1 log(k/c1 ) α D . Since Diam(Ω) = 2, setting K = 1 and ε1 = in Theorem 13 gives E h D 2D log n − αr log log n i Wrr (P, Pb ) . 1 log n αr αr D D . On the other hand, setting k = n in Theorem 14 gives inf sup Pb P ∈P X 1 αr h i 1 D 1 r Wr (P, Pb ) ≥ , E 2 log(n/c1 ) IID ,...,X ∼ P n −αr showing that distribution estimation over (P, Wrr ) has the extremely slow minimax rate (log n) D . Although we considered only α ∈ (0, 1] (due to the notational complexity of defining higher-order Hölder spaces), analogous rates hold for all α > 0. Also, since our rates depend only on covering and packing numbers of Ω, identical rates can be derived for related Sobolev and Besov classes. One might wonder why we are interested studying Wasserstein convergence of distributions over spaces of smooth functions, as in Example 4. Our main motivation comes from the fact that, historically, smooth function spaces have been widely used for modeling images and other complex naturalistic signals (Mallat, 1999; Peyré, 2011). Far more recently, empirical breakthroughs have been made in generative modeling, particularly of images, based on the principle of minimizing Wasserstein distance between the empirical distribution and a large class of models encoded by a deep neural network (Montavon et al., 2016; Arjovsky et al., 2017; Gulrajani et al., 2017). However, little is known about the theoretical properties of these methods; while there has been some work studying the optimization landscape of such models (Nagarajan and Kolter, 2017), we know of no work exploring their statistical properties. Given the extremely slow minimax convergence rate we derived above, it must be the case that the class of distributions encoded by such models is far smaller or sparser than P. An important avenue for further work is thus to explicitly identify stronger assumptions that can be made on distributions over interesting classes of signals, such as images, to bridge the gap between empirical performance and our theoretical understanding. 12 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE 8. Conclusion In this paper, we derived upper and lower bounds for distribution estimation under Wasserstein loss. Our upper bounds generalize prior results and are tighter in certain cases, while our lower bounds are, to the best of our knowledge, the first minimax lower bounds for this problem. We also provided several simple examples for which our upper and lower bounds agree. 8.1. Future Work We studied only minimax rates over the entire class P of all distributions on the metric space (Ω, ρ). It would be useful to understand how minimax rates improve when additional assumptions, such as smoothness or finite-moment assumptions are made (see Fournier and Guillin (2015) for somewhat improved upper bounds under finite-moment assumptions when (Ω, ρ) is Euclidean space). Given the rather slow convergence rates we found over P in many cases, studying minimax rates under assumptions may help to explain the relatively favorable empirical performance of popular distribution estimators based on empirical risk minimization in Wasserstein loss. Moreover, while rates over all of P are of interest only for very weak metrics such as the Wasserstein distance (as stronger metrics may be infinite or undefined), studying minimax rates under additional assumptions will allow for a better understanding of the Wasserstein metric in relation to other commonly used metrics. Acknowledgments This work was partly supported by a NSF Graduate Research Fellowship DGE-1252522 to S.S. and by NSF IIS-1563887, Darpa D3M program, and AFRL grants to B.P.. References Miklós Ajtai, János Komlós, and Gábor Tusnády. On optimal matchings. Combinatorica, 4(4): 259–264, 1984. Syed Mumtaz Ali and Samuel D Silvey. A general class of coefficients of divergence of one distribution from another. Journal of the Royal Statistical Society. Series B (Methodological), pages 131–142, 1966. Martin Arjovsky, Soumith Chintala, and Léon Bottou. arXiv:1701.07875, 2017. Wasserstein GAN. arXiv preprint Amparo Baı́llo, Javier Cárcamo, and Konstantin V Getman. The estimation of Wasserstein and Zolotarev distances to the class of exponential variables. arXiv preprint arXiv:1603.06806, 2016. Federico Bassetti, Antonella Bodini, and Eugenio Regazzini. On minimum Kantorovich distance estimators. Statistics & probability letters, 76(12):1298–1302, 2006. Espen Bernton, Pierre E Jacob, Mathieu Gerber, and Christian P Robert. Inference in generative models using the Wasserstein distance. arXiv preprint arXiv:1701.05146, 2017. Emmanuel Boissard, Thibaut Le Gouic, et al. On the mean speed of convergence of empirical and occupation measures in Wasserstein distance. In Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, volume 50, pages 539–563. Institut Henri Poincaré, 2014. 13 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Emmanuel Boissard, Thibaut Le Gouic, Jean-Michel Loubes, et al. Distributions template estimate with wasserstein metrics. Bernoulli, 21(2):740–759, 2015. James Bucklew. Introduction to rare event simulation. Springer Science & Business Media, 2013. Sourav Chatterjee et al. A new method of normal approximation. The Annals of Probability, 36(4): 1584–1610, 2008. Louis HY Chen, Larry Goldstein, and Qi-Man Shao. Normal approximation by Steins method. Springer Science & Business Media, 2010. Imre Csiszár. Eine informationstheoretische ungleichung und ihre anwendung auf beweis der ergodizitaet von markoffschen ketten. Magyer Tud. Akad. Mat. Kutato Int. Koezl., 8:85–108, 1964. Eustasio del Barrio, Evarist Giné, and Carlos Matrán. Central limit theorems for the Wasserstein distance between the empirical and the true distributions. Annals of Probability, pages 1009– 1071, 1999. Eustasio del Barrio, Evarist Giné, Carlos Matrán, et al. Correction: Central limit theorems for the Wasserstein distance between the empirical and the true distributions. The Annals of Probability, 31(2):1142–1143, 2003. Steffen Dereich, Michael Scheutzow, Reik Schottstedt, et al. Constructive quantization: Approximation by empirical measures. In Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, volume 49, pages 1183–1203. Institut Henri Poincaré, 2013. Ronald A DeVore and George G Lorentz. Constructive approximation, volume 303. Springer Science & Business Media, 1993. Luc Devroye. The equivalence of weak, strong and complete convergence in L1 for kernel density estimates. The Annals of Statistics, pages 896–904, 1983. Khanh Do Ba, Huy L Nguyen, Huy N Nguyen, and Ronitt Rubinfeld. Sublinear time algorithms for earth movers distance. Theory of Computing Systems, 48(2):428–442, 2011. Joseph L Doob. Measure theory, volume 143. Springer Science & Business Media, 2012. Richard M Dudley. The sizes of compact subsets of hilbert space and continuity of gaussian processes. Journal of Functional Analysis, 1(3):290–330, 1967. RM Dudley. The speed of mean Glivenko-Cantelli convergence. The Annals of Mathematical Statistics, 40(1):40–50, 1969. Peyman Mohajerin Esfahani and Daniel Kuhn. Data-driven distributionally robust optimization using the wasserstein metric: Performance guarantees and tractable reformulations. arXiv preprint arXiv:1505.05116, 2015. Nicolas Fournier and Arnaud Guillin. On the rate of convergence in Wasserstein distance of the empirical measure. Probability Theory and Related Fields, 162(3-4):707–738, 2015. 14 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Rui Gao and Anton J Kleywegt. Distributionally robust stochastic optimization with wasserstein distance. arXiv preprint arXiv:1604.02199, 2016. Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. In Advances in Neural Information Processing Systems, pages 5769–5779, 2017. Yanjun Han, Jiantao Jiao, and Tsachy Weissman. Minimax estimation of discrete distributions under ℓ1 loss. IEEE Transactions on Information Theory, 61(11):6343–6354, 2015. David Haussler. Sphere packing numbers for subsets of the boolean n-cube with bounded vapnikchervonenkis dimension. Journal of Combinatorial Theory, Series A, 69(2):217–232, 1995. Nhat Ho, XuanLong Nguyen, Mikhail Yurochkin, Hung Hai Bui, Viet Huynh, and Dinh Phung. Multilevel clustering via wasserstein means. arXiv preprint arXiv:1706.03883, 2017. Oliver Johnson, Richard Samworth, et al. Central limit theorem and convergence to stable laws in mallows distance. Bernoulli, 11(5):829–845, 2005. Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document distances. In International Conference on Machine Learning, pages 957–966, 2015. Stéphane Mallat. A wavelet tour of signal processing. Academic press, 1999. Grégoire Montavon, Klaus-Robert Müller, and Marco Cuturi. Wasserstein training of restricted Boltzmann machines. In Advances in Neural Information Processing Systems, pages 3718–3726, 2016. Tetsuzo Morimoto. Markov processes and the H-theorem. Journal of the Physical Society of Japan, 18(3):328–331, 1963. Vaishnavh Nagarajan and J Zico Kolter. Gradient descent GAN optimization is locally stable. arXiv preprint arXiv:1706.04156, 2017. Kangyu Ni, Xavier Bresson, Tony Chan, and Selim Esedoglu. Local histogram based segmentation using the wasserstein distance. International journal of computer vision, 84(1):97–111, 2009. Gabriel Peyré. The numerical tours of signal processing. Computing in Science & Engineering, 13 (4):94–97, 2011. Matthias Reitzner, Matthias Schulte, et al. Central limit theorems for u-statistics of poisson point processes. The Annals of Probability, 41(6):3879–3909, 2013. Phillippe Rigollet. High-dimensional statistics. Lecture notes for course 18S997, 2015. Emmanuel Rio et al. Upper bounds for minimal distances in the central limit theorem. In Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, volume 45, pages 802–817. Institut Henri Poincaré, 2009. Emmanuel Rio et al. Asymptotic constants for minimal distance in the central limit theorem. Electronic Communications in Probability, 16:96–103, 2011. 15 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover’s distance as a metric for image retrieval. International journal of computer vision, 40(2):99–121, 2000. Ludger Rüschendorf. The wasserstein distance and approximation theorems. Probability Theory and Related Fields, 70(1):117–129, 1985. Roman Sandler and Michael Lindenbaum. Nonnegative matrix factorization with earth mover’s distance metric for image analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(8):1590–1602, 2011. Justin Solomon, Fernando De Goes, Gabriel Peyré, Marco Cuturi, Adrian Butscher, Andy Nguyen, Tao Du, and Leonidas Guibas. Convolutional wasserstein distances: Efficient optimal transportation on geometric domains. ACM Transactions on Graphics (TOG), 34(4):66, 2015. Cédric Villani. Optimal transport: old and new, volume 338. Springer Science & Business Media, 2008. Jonathan Weed and Francis Bach. Sharp asymptotic and finite-sample rates of convergence of empirical measures in Wasserstein distance. arXiv preprint arXiv:1707.00087, 2017. Meng Zhang, Yang Liu, Huan-Bo Luan, Maosong Sun, Tatsuya Izuha, and Jie Hao. Building earth mover’s distance on bilingual word embeddings for machine translation. In AAAI, pages 2870– 2876, 2016. Tong Zhang. Covering number bounds of certain regularized linear function classes. Journal of Machine Learning Research, 2(Mar):527–550, 2002. Ding-Xuan Zhou. The covering number in learning theory. Journal of Complexity, 18(3):739–767, 2002. 16 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Appendix A. Proofs of Lemmas Lemma 7 Suppose S ∈ S is a countable Borel partition of Ω. Let P and Q be Borel probability measures such that, for every S ∈ S, P (S) = Q(S). Then, for any r ≥ 1, Wr (P, Q) ≤ Res(S). Proof This fact is intuitively obvious; clearly, there exists a transportation map µ from P to Q that moves mass only within each S ∈ S and therefore without moving any mass further than δ. For completeness, we give a formal construction. Let µ : Σ2 → [0, 1] denote the coupling that is conditionally independent given any set S ∈ S with P (S) = Q(S) > 0 (that is, for any A, B ∈ Σ, µ(A×B ∩S ×S)P (S) = P (A∩S)Q(B ∩S)).4 It S is easy to verify that µ ∈ C(P, Q). Since S is a countable partition and µ is only supported on S∈S S × S, Wr (P, Q) ≤ Z Ω×Ω XZ = S∈S XZ ≤ S∈S =δ 1/r ρ (x, y) dµ(x, y) r X S∈S !1/r r ρ (x, y) dµ(x, y) S×S !1/r r δ dµ(x, y) S×S !1/r µ(S × S) =δ X P (S)Q(S) P (S) S∈S !1/r =δ X !1/r Q(S) S∈S = δ. Lemma 8 Suppose (Ω, ρ) is a metric space, and suppose P and Q are Borel probability distributions on Ω with countable support; i.e., there exists a countable set X ⊆ Ω with P (X ) = Q(X ) = 1. Then, for any r ≥ 1, X X \|P ({x}) − Q({x})\| . \|P ({x}) − Q({x})\| ≤ Wrr (P, Q) ≤ (Diam(X ))r (Sep(X ))r x∈X x∈X P Proof The term x∈X \|P ({x}) − Q({x})\| = T V (P, Q) is precisely the (unweighted) amount of mass that must be transported to transform between P and Q. Hence, the result is intuitively fairly obvious; all mass moved has a cost of at least Sep(Ω) and at most Diam(Ω). However, for completeness, we give a more formal proof below. To prove the lower bound, suppose µ ∈ Π(P, Q) is any coupling between P and Q. For x ∈ X , µ({x} × {x}) + µ({x} × (Ω\{x})) = µ({x} × Ω) = P ({x}) and, similarly, µ({x} × {x}) + µ((Ω\{x}) × {x}) = µ(Ω × {x}) = Q({x}). 4. The existence of such a measure can be verified by the Hahn-Kolmogorov theorem, similarly to that of the usual product measure (see, e.g., Section IV.4 of Doob (2012)). 17 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Since P ({x}), Q({x}) ∈ [0, 1], it follows that µ({x} × (Ω\{x})) + µ(µ((Ω\{x}) × {x})) ≥ \|P ({x} − Q({x})\| . Therefore, since ρ(x, y) = 0 whenever x = y and ρ(x, y) ≥ Sep(Ω) whenever x 6= y, Z Z ρr (x, y) dµ(x, y) ρr (x, y) dµ(x, y) = X ×X Ω×Ω Z XZ r ρr (x, y) dµ(x, y) ρ (x, y) dµ(x, y) + = x∈X {x}×(Ω\{x}) ≥ (Sep(Ω))r r ≥ (Sep(Ω)) X x∈X X x∈X (Ω\{x})×{x} µ({x} × (Ω\{x})) + µ((Ω\{x}) × {x}) \|P ({x}) − Q({x})\| . Taking the infimum over µ on both sides gives X \|P ({x}) − Q({x})\| ≤ Wrr (P, Q). (Sep(Ω))r x∈X To prove the upper bound, since ρ is upper bounded by Diam(Ω), it suffices to construct a coupling µ that only moves mass into or out of each given point, but not both; that is, for each x ∈ X , min{µ({x} × (Ω\{x})), µ((Ω\{x}) × {x})} = 0. One way of doing this is as follows. Fix an ordering x1 , x2 , ... of the elements of X . For each i ∈ N, define i i X X Xi := (P (xℓ ) − Q(xℓ ))+ and Yi := (Q(xℓ ) − P (xℓ ))+ , ℓ=1 ℓ=1 and further define ji := min{j ∈ N : Xi ≤ Yj } and ki := min{k ∈ N : Xj ≥ Yi }. Then, for each i ∈ N, move Xi mass from {x1 , ..., xi } to {y1 , ..., yji } and move Yi mass from {y1 , ..., yi } to {x1 , ..., xki }. As i → ∞, by construction of Xi and Yi , the total mass moved in this way is X \|P (x) − Q(x)\| . µ((X × X )\{(x, x) : x ∈ X }) = lim Xi + Yi = i→∞ x∈X Lemma 10 Let K be a positive integer. Suppose {Sk }k∈[K] is a sequence of nested countable Borel partitions of (Ω, ρ), with S0 = Ω. Then, for any r ≥ 1 and any Borel probability distributions P and Q on Ω,   K X X (Res(Sk−1 ))r  Wrr (P, Q) ≤ (Res(SK ))r + \|P (S) − Q(S)\| . S∈Sk k=1 18 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Proof Our proof follows the same ideas as and slightly generalizes of the proof of Proposition 1 in Weed and Bach (2017). Intuitively, to prove Lemma 10 it suffices to find a transportation map such that For each k ∈ [K], recursively define Pk := P − k−1 X µk and j=0 Qk := Q − k−1 X νk , j=0 where, for each k ∈ [K], µk and νk are Borel measures on Ω defined for any E ∈ Σ by µk (E) := X (Pk (S) − Qk (S))+ Pk (E ∩ S) Pk (S) X (Qk (S) − Pk (S))+ Qk (E ∩ S) . Qk (S) S∈Sk :Pk (S)>0 and νk (E) := S∈Sk :Qk (S)>0 P By construction of µk and νk , each µk and νk is a non-negative measure and K k=1 µk ≤ P and PK k=1 νk ≤ Q. Furthermore, for each k ∈ [K − 1], for each S ∈ Sk , µk+1 (S) = νk+1 (S), and X \|P (S) − Q(S)\| . µk (Ω) = νk (Ω) ≤ S∈Sk Consequently, although µ and ν are not probability measures, we can slightly generalize the definition of Wasserstein distance by writing Wrr (µk , νk ) := µ(Ω) τ ∈Π inf ν µk , k µk (Ω) νk (Ω) [ρr (X, Y )] E (X,Y )∼µ (or Wrr (µk , νk ) = 0 if µk = νk = 0). In particular, this is convenient because we one can easily show that, by construction of the sequences {Pk }k∈[K] and {Qk }k∈[K] , Wrr (P, Q) ≤ Wrr (PK , QK ) + For each k ∈ [K], Lemma 8 implies that X Wrr (µk , νk ) ≤ (Diam(S))r S∈Sk−1 ≤ (Res(Sk−1 )) r = (Res(Sk−1 )) r X K X Wrr (µk , νk ) . k=1 X T ∈Sk :T ⊆S \|P (T ) − Q(T )\| X S∈Sk−1 T ∈Sk :T ⊆S X T ∈Sk \|P (T ) − Q(T )\| . Furthermore, for each S ∈ SK , PK = QK , Lemma 7 gives that Wrr (PK , QK ) ≤ (Res(SK ))r 19 \|P (T ) − Q(T )\| (7) D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Plugging these last two inequalities into Inequality (7) gives the desired result: Wrr (P, Q) r ≤ (Res(SK )) + K X (Res(Sk−1 ))r X S∈Sk k=1 \|P (S) − Q(S)\| . Lemma 11 Suppose S and T are partitions of (Ω, ρ), and suppose S is countable. Then, there exists a partition S ′ of (Ω, ρ) such that: a) \|S ′ \| ≤ \|S\|. b) Res(S ′ ) ≤ Res(S) + 2 Res(T ). c) T is a refinement of S ′ . Proof Enumerate the elements of S as S1 , S2 , .... Define S0′ := ∅, and then, for each i ∈ {1, 2, ...}, recursively define -   i−1 [ [ T  Sj′  , Si′ :=  j=1 T ∈T :T ∩Si 6=∅ and set S ′ = {S1′ , S2′ , ...}. Clearly, \|S ′ \| ≤ \|S\| (equality need not hold, as we may have some Si′ = ∅). By the triangle inequality, each   [ Diam(Si′ ) ≤ Diam  T  ≤ δS + 2δT . T ∈T :T ∩Si 6=∅ Finally, since T is a partition and we can write -  i−1 [ [ T  Si′ =  j=1 T ∈T T ∈T :T ∩Si 6=∅ T is a refinement of S ′ . [ :T ∩Sj′ 6=∅  T , Before proving Lemma 12, we recall a simple concentration inequality bounding the L1 deviation of a multinomial random variable from its mean: Lemma 15 (Lemma 3 P of Devroye (1983)) Suppose (X1 p , ..., XK ) ∼ Multinomial(n, p1 , ..., pK ). 20K/n, for all ε ∈ [ε0 , 1), Let Z := kX − pk1 = K \|X − p \|. Then, for ε := 0 k k k=1 P [Z > nε] ≤ 3 exp −nε2 /25 . Lemma 12 now follows as an easy corollary. 20 D ISTRIBUTION E STIMATION IN WASSERSTEIN D ISTANCE Lemma 12 Consider the setting of Lemma 15. Then, p Z ≤ 21K/n. E n Proof Since, almost surely, Z < n, by Lemma 15, Z ∞ P [Z > nε] dε [Z/n] = E 0 Z 1 ≤ ε0 + 3 exp −nε2 /25 dε Z ε0 1−ε0 exp −n (ε + ε0 )2 /25 dε 0 Z 1 p 3 2 ≤ ε0 + 3 exp −nε0 exp −nε2 /25 dε ≤ ε0 + exp −nε20 ≤ 21K/n. 2 0 = ε0 + 3 21 D ISTRIBUTION E STIMATION IN 22 WASSERSTEIN D ISTANCE	10math.ST
Intelligent Device Discovery in the Internet of Things – Enabling the Robot Society James Sunthonlap, Phuoc Nguyen, Zilong Ye arXiv:1712.08296v2 [cs.AI] 8 Jan 2018 California State University Los Angeles 5151 State University Drive, Los Angeles, CA 90032, USA Email: zye5@calstatela.edu Abstract—The Internet of Things (IoT) is continuously growing to connect billions of smart devices anywhere and anytime in an Internet-like structure, which enables a variety of applications, services and interactions between human and objects. In the future, the smart devices are supposed to be able to autonomously discover a target device with desired features and thus yield the computing service, network service and data fusion that leads to the generation of a set of entirely new services and applications that are not supervised or even imagined by human beings. The pervasiveness of smart devices, as well as the heterogeneity of their design and functionalities, raise a major concern: How can a smart device efficiently discover a desired target device? In this paper, we propose a Social-Aware and Distributed (SAND) scheme that achieves a fast, scalable and efficient device discovery in the IoT. The proposed SAND scheme adopts a novel device ranking criteria that measures the device’s degree, social relationship diversity, clustering coefficient and betweenness. Based on the device ranking criteria, the discovery request can be guided to travel through critical devices that stand at the major intersections of the network, and thus quickly reach the desired target device by contacting only a limited number of intermediate devices. We conduct comprehensive simulations on both random networks and scale-free networks to evaluate the performance of SAND in terms of the discovery success rate, the number of devices contacted and the number of communication hops. The simulation results demonstrate the effectiveness of SAND. With the help of such an intelligent device discovery as SAND, the IoT devices, as well as other computing facilities, software and data on the Internet, can autonomously establish new social connections with each other as human being do. They can formulate self-organized computing groups to perform required computing tasks, facilitate a fusion of a variety of computing service, network service and data to generate novel applications and services, evolve from the individual aritificial intelligence to the collaborative intelligence, and eventually enable the birth of a robot society. Index Terms—Internet of Things, social-aware, distributed, device discovery, computing, network and data fusion, robot society. I. I NTRODUCTION The Internet of Things (IoT) is continuously growing to connect billions of smart devices anywhere and anytime through an Internet-like structure. A recent forecast by International Data Corporation (IDC) shows that the IoT will involve more than 212 billion of objects by 2020, and the IoT and the associated ecosystem are predicted to have a $1.7 trillion market [1]. The IoT devices are capable of sensing, analyzing and evaluating the surrounding objects and people. They can collaborate and work together to provide a set of new applications and services, such as smart home, e-healthcare and intelligent transportation system. The IoT is envisioned to dramatically change and enhance the interactions between human and objects, bringing transformative benefits to the lives of human beings. As the IoT evolves, it is promising that the future smart devices may have the capability to discover a target device with desired features, and autonomously collaborate with each other to accomplish certain missions or tasks. An intelligent device discovery strategy may allow the IoT devices to efficiently establish new connections with other devices based on their need. The connections can be set up to gather a certain number of computing powers, network functions, program source codes, raw datasets, and etc to enable new services and applications. A variety of computing service, network service and data can be fused [2]-[4] by following the guidance of an intelligent device discovery strategy. For example, a glucose level monitor can send a request to find a glucose analyzer, and collaborate on evaluating a given patient’s glucose level. The glucose analyzer may also need to search for a national wide database to compare the given patient’s results with that of other patients, and then given the advisement or alert to the given patient. Here, an intelligent device discovery strategy could guide the request to the appropriate destinations for obtaining the required computing facilities, functions and data to provide the evaluation service, natually leading to a computing service, network service and data fusion. In addition, the IoT devices can also form social connections as human beings do, and they can collaborate to generate entirely new kinds of applications and services with the help of collaborative intelligence of smart devices and data, rather than the supervision of human beings. In the above future scenarios, one of the key challenges is how to achieve a fast, scalable and efficient device discovery when there are millions or billions of devices in the IoT. An intelligent device discovery should have a forwarding strategy that guide the discovery request to quickly arrive at the desired target device with minimum detours. In addition, since most of the IoT devices are power constrained, the device discovery should avoid involving too many intermediate devices in the process of data exchange. There are a few existing works on the topic of IoT device discovery. The work in [5] identified the differences between user search and machine-to-machine search, and presented the challenges and requirements for IoT device search. The authors in [6] studied device discovery on different network topologies, including a star-like centralized network, a regular mesh-like decentralized network and a hierarchical network. They focused on analyzing the effect of the network topology on the discovery success rate, rather than designing new device discovery strategies. The study in [7] proposed a centralized IoT search engine that can accurately interpret the context of the discovery request and thus make a proper management of search and usage in the IoT middleware environment. However, this centralized approach may not be scalable when the network size is large, and can be vulnerable to a single point of failure. More recent work [8] investigated a distributed device discovery strategy based on Information-Centric Networking (ICN). However, this approach may not be applicable to the IoT system since it relies on the built-in features of ICN such as cache and broadcast capabilities, which could introduce an extra cost to the resource-constrained IoT devices. In this work, we focus on investigating autonomous and intelligent device discovery in the IoT. We do not treat smart devices as independent objects or limit our study on the artificial intelligence of a single device, but we explore humanlike social behaviors and collaborative intelligence of smart devices. It is envisioned for the next episode of research in the IoT and robotics. In particular, we propose a novel IoT device discovery scheme, called Social-Aware and Distributed (SAND), which can discover a desired target device in a fast, efficient and scalable manner. In SAND, IoT devices are fair and equal as peers, and they can interact with each other in an autonomous and distributed manner as human do, as each of them only maintains the information of their local neighboring peers. Each IoT device is ranked by measuring the device’s degree, social relationship diversity, clustering coefficient and betweenness. Based on this, SAND opts to forward the discovery request to the neighboring device with the highest rank, which is likely to stand at the major intersection of the network that can fast and efficiently lead the request to the desired target device. We conduct comprehensive simulations to evaluate the proposed SAND scheme on both a random network and a scale-free network. The simulation results show that SAND can achieve a high discovery success rate and a short discovery path, while contacting a small number of intermediate devices. The rest of the paper is organized as follows. We first introduce the traditional centralized and distributed device discovery schemes in Section 2. Then, we describe the design details of the proposed SAND scheme in Section 3. In Section 4, we present the performance evaluation. In Section 5, we discuss the research challenges and opportunities related to the IoT device discovery. Finally, we conclude the paper and propose our future work in Section 6. II. T RADITIONAL I OT D EVICE D ISCOVERY S CHEMES There are two traditional solutions for addressing the device discovery in the IoT, which are the centralized scheme and the distributed scheme. The former leverages a centralized controller to resolve the discovery request, and the latter uses a simple broadcast mechanism to discover the required target device in a breadth-first search manner. The detailed description and the comparison of these two approaches are presented in the following subsections. A. Centralized IoT Device Discovery Scheme To achieve a fast and efficient IoT device discovery, a simple yet effective approach is to introduce a centralized controller which keeps track of the general information (e.g., the device’s functionality) of all the devices and maintains the shortest paths between all the device pairs. If there exists a source device which requests to discover a target device that provides a desired function, the discovery request can be sent to the centralized controller to resolve. The centralized controller can find the location of the desired target device and reply to the source device with the shortest path information to the target device. Such a centralized scheme is fast and accurate due to that the centralized controller has a global view of all the devices in the network. It is also efficient since the centralized controller can provide the shortest path between the source and target devices, which involves the minimum number of intermediate devices in transmitting the discovery request. The centralized scheme can be designed with robustness against dynamic changes in the network. When a new device joins the network, it needs to register the general information at the controller. A device that leaves the network also needs to unsubscribe at the centralized controller. In addition, the centralized controller can periodically send out heartbeat messages to all the registered devices and maintain an up-to-date list of the alive devices in the network. If a device fails to reply to the heartbeat after three attempts, it is considered as in the mode of lost and will be removed from the list of alive at the centralized controller. We can see that the centralized scheme highly relies on the centralized controller which manages the whole system and the communication between the IoT devices. B. Distributed IoT Device Discovery Scheme The centralized scheme is fast and efficient; however, it may not be scalable when the number of devices is large. In contrast, the distributed scheme offers a scalable device discovery in the IoT. In the distributed scheme, each device maintains a local neighbor list that consists of the neighbors’ general information. When a new device joins a network, it only exchanges the general information with nearby devices that are within the transmission range. When a device leaves the network, it needs to inform its neighbors for them to update their local neighbor list. Heartbeats can be exchanged between the devices to maintain the fresh live or lost status of the devices’ neighbors. The distributed discovery scheme works in a breadth-first search manner. When a source device makes a request to discover a target device with desired features, the distributed scheme simply broadcasts the discovery request to all the neighboring devices. This process iterates until the desired target device is found. Compared to the centralized scheme, the distributed scheme is more resilient against possible failures since there is no centralized control that is at the risk of single point of failure. It is also more robust when the network is dynamically changing since each device only needs to maintain their local neighbor’s information. We can also infer that the distributed scheme can find a cost-efficient communication path between the source and target devices since the breadth-first search is used. The only limitation is that a large number of intermediate devices may be involved in transmitting the discovery request when broadcast is used, which is energy consuming. This is harmful to the IoT system where most of the devices are power constrained. III. S OCIAL -AWARE AND D ISTRIBUTED I OT D EVICE D ISCOVERY – SAND In the famous small-world experiment [9], Travers and Milgram found that people are tied by a short chain of connections and any two persons can be connected in six hops by simply exploring their social networks. A more recent Facebook research [10] confirmed this observation in which they concluded that, among their 1.59 billion active users, in average a person is separated by 3.57 hops away from another person. As the IoT evolves, devices could also exhibit human-like social relationships. For example, devices can be considered as family if they are from the same manufacturer, or can be considered as colleagues if they work together to provide a service. Such social aspects can therefore be used to intelligently navigate device discovery requests through the device social network and connect two devices in a few hops as the small-world exhibited by the human social network. Based on this motivation, we propose the Social-Aware and Distributed (SAND) scheme, which can achieve a scalable, fast and efficient device discovery in the IoT. In SAND, IoT devices can autonomously establish meaningful relationships with other devices, and form an overlay device social network in addition to their communication network. It may be faster and more efficient for the devices to find their desired target device by searching among their “friends” through the overlay device social network, as oppose to simply broadcasting through the communication network as the distributed scheme works. In the following subsections, we will present the details of the SAND architecture, the device ranking criteria and the forwarding strategy, respectively. A. The SAND Architecture In SAND, the IoT devices form two layers of networks, which consists of the communication network and the overlay device social network. The communication network is at the bottom layer, which abstracts the data communication links between the IoT devices. As long as two IoT devices are within their transmission range, a communication link exists between them. Actually, the distributed scheme introduced in the previous section only considers this communication network and performs simple broadcasting for device discovery. In SAND, in addition to the communication network, an overlay device social network is constructed to help achieve an effective IoT device discovery. When two IoT devices can communicate with each other, they can exchange their device general information, such as the manufacturer, the functionality, the ownership and the location information. In the overlay device social network, there exists a link between two IoT devices if they have a valid social relation. Similar as human relationships, device social relationships typically include family, friends, neighbors, colleagues, and etc. Devices from the same manufacturer can be considered as family, e.g., an iPhone and an Apple TV. Device friendship can be described as objects that interact frequently with one another and tend to share a common theme. For example, Bob’s smartphone and his body sensors are considered as friends since they interact frequently and they both serve as key components in providing e-healthcare services. IoT devices that locate in the same room or floor can be considered as neighbors. IoT devices are considered as colleagues if they work together to provide a specific service. For example, the temperature sensors, the humidity sensors and the air conditioner are colleagues that work together to offer a comfortable living environment in a smart home. It is worth noting that, in our simulation (to be introduced in Section IV), we abstract the device general information into a number of device features, and there exists a social link between two IoT devices if they have a common feature. An illustrative example of SAND in a smart home is shown in Fig. 1. Here, we only show the overlay device social network. We assume that all the devices are connected through WiFi, so the communication network is a complete mesh network, which is not shown. In the overlay social network, the refrigerator (E), the TV (A) and the washing machine (C) are connected since they are from the same manufacturer, e.g., Samsung. The neighbor relationship exists between the vacuum (B) and the washing machine (C), both of which locate in the storage room. Friendship exists between (1) the telephone (D) and the TV (A), and (2) the refrigerator (E) and the PC (F ), which are involved in frequent interactions to serve the home owner. The refrigerator (E), the microwave (G) and the boiler (H) are considered as colleagues that work together to prepare meals. In SAND, when an IoT device joins the network, it will exchange the general information with nearby devices that are within the transmission range. Consequently, social links can be established if valid social relationships exist. Here, we assume each IoT device is social-aware and intelligent in the sense that they can establish social relationships autonomously like human. The social aspects of IoT devices give them the ability to form an overlay device social network automatically. It is worth noting that the overlay device social network is not static but dynamically changing because of the device movements or the device relationship changes (e.g., the colleague relationship may change frequently). Hence, each IoT device needs to periodically update their social connections in SAND. In SAND, IoT devices are supposed to act as human beings. The social aspects of the devices allow them to dynamically create new social connections and form new colleagueship to work together to generate new services. With the social aspects B A C F E D H G Fig. 1: SAND in a Smart Home in SAND, devices become more visible and aware to each other, thus leading to a fast and effective device discovery. In the process of device discovery, rather than simply broadcasting the discovery message to all the neighbors as the distribute schemes does, SAND will send the discovery message to each of the neighbors in a preferred order based on the the rank of the neighboring device. The higher rank the device is, the more likely and faster it is able to forward the discovery message to the target device requested by the source device. With the social aspects in SAND, devices become more visible to other devices, thus making SAND more scalable and efficient than the baseline distributed scheme. B. The Device Ranking Criteria in SAND In SAND, the device discovery adopts the depth-first search strategy, where the discovery request is forwarded to the neighboring devices in a preferred order based on their ranks. Here, the rank of a device is determined by four factors which are the device degree, the diversity, the (local) clustering coefficient and the (local) betweenness of the IoT device. The first three factors help SAND to select a device that may reach a broad community, while the betweenness leads the discovery request to a device that stands at the major intersection of multiple shortest paths in the network. Thus, the device ranking criteria makes SAND an intelligent, accurate and fast device discovery strategy. We will present the calculation details of these four factors as follows. Device Degree: In the overlay social network, the degree of device i is denoted by ki , which is defined as the number of social links it has. For example, in Fig. 1, device E has a degree of 5 since there are five social connections associated with it. The higher the device degree is, the more likely it is associated with a routing path to the desired target device since there are more outlets from this device. Diversity: In SAND, the diversity of device i is denoted by di , which is defined as the number of types of social links a device is associated with. For example, in Fig. 1, device E has a diversity of 3 since it has three different types of social links which are family (i.e., links with A and C), colleagueship (i.e., links with G and H) and friendship (i.e., link with F ). An IoT device could have a high device degree ki , but if all the connections are within the same type of social relationship, the device is still assigned with a relatively low rank since it is not diverse enough to reach distinct communities of devices. In contrast, an IoT device that has a high diversity d is involved in many different types of social relations, and may have a broad reachability, thus having a higher chance to connect with the desired target device. Clustering coefficient: The clustering coefficient shows how likely a device and its neighbors form a clique [11]. The local clustering coefficient of a device is defined as the number of the social links in its neighborhood (including the device itself and its one-hop neighbors) divided by the total number of possible links in the neighborhood. For a given device i, the local clustering coefficient ci can be calculated as: 2\| est : s, t ∈ Ni , est ∈ E \| (1) ci = ki (ki − 1) where est is a social link between device s and device t, Ni is the set of devices in the neighborhood of device i, E is the set of social links in the neighborhood of device i, and ki is the degree of device i. The higher the local clustering coefficient is, the more likely the device and its neighbors are forming a clique (e.g., fully mesh), thus the diameter of the device neighborhood is smaller, which can lead to a faster and wider dissemination for the discovery request. Betweenness: The betweenness reflects the probability that a given device stands at the critical intersection of multiple shortest paths in the network [12]. It is defined as the number of shortest paths that traverse the given device divided by the total number of possible shortest paths. In SAND, the local betweenness measures such a probability in the neighborhood that includes the given device and its one-hop neighbors. The local betweenness of device i is calculated as: P s,t∈Ni δst (i) bi = (2) ki (ki − 1) where Ni is the set of devices in the neighborhood of device i, s and t are a pair of devices in Ni , and ki is the degree of device i. Here, δst (i) is 1 if the shortest path between s and t traverses through device i; otherwise, it is 0. The higher the local betweenness is, the higher chance the given device is located at the intersection that connects major shortest paths in the neighborhood. Discovery requests that arrive at such intersection can be disseminated to anyplace in the network easily and quickly over those shortest paths. In SAND, the rank of a device Ri is defined as the multiplication of the above mentioned four factors (e.g., Eq. (3)). The higher the rank is, the more likely the device can forward the discovery request to the desired target device in a fast and efficient manner. Ri = ki ∗ di ∗ ci ∗ bi (3) C. The SAND Forwarding Strategy Based on the architecture and the device ranking criteria as described above, we propose the SAND device discovery scheme in this subsection. Rather than the simple broadcast used in the distributed scheme, SAND forwards the discovery request to a neighbor device that ranks the highest. For example, in Fig. 1, the discovery request from D to C will be forwarded to E, which has the highest rank (with a degree of 5, diversity of 3, local clustering coefficient of 0.53 and local betweenness of 0.53). The device with the highest rank is supposed to be the most effective one that leads to the desired target device. In the discovery process, SAND performs a depth-first search with limited search depth, where the search will not exhaust to the deepest level but a depth level of n (which is a tunable parameter). If the desired target device cannot be found after searching for n levels, SAND will step back to examine the other unchecked neighbors in depth n−1. The pseudocode of SAND discovery algorithm is shown in Algorithm 1. Algorithm 1 The SAND Device Discovery Algorithm Input: A discovery request rst from source s to target t Output: The communication path pst between s and t 1: pst .push(s); 2: while pst is not empty do 3: i ← pst .pop(); 4: if i is the desired target device then 5: return the reverse path of pst ; else if i has been checked before then 6: 7: continue; 8: else if pst .length() = depth n then 9: continue; 10: else 11: rank and sort the neighbor devices Ni ; 12: pst .push(Ni highest rank ); 13: end if 14: end while Similar as the distributed scheme, SAND is scalable since each IoT device only maintains information of its local connections. SAND is also resilient and robust since there is no centralized controller and each IoT device is fair and equal as a peer. Compared to the distributed scheme, SAND is more efficient because the former applies simple broadcast and thus a significant large number of devices are involved in the discovery process, while in the latter, SAND performs depthfirst search with limited search depth and the guidance of an intelligent device ranking criteria, so the number of devices involved is relatively small. Such an advantage of SAND may lead to a low energy consumption in the process of transmitting the discovery request, which is critically beneficiary to the IoT system where devices are power constrained. Hence, we can see that SAND inherits the advantages of the distributed scheme, and further improves the communication efficiency. IV. P ERFORMANCE E VALUATION In this section, we focus on evaluating the performance of the proposed SAND scheme, compared to the distributed scheme that uses simple broadcast. We conduct simulations on both a random network and a scale-free network. The random network is an irregular mesh network with social links randomly generated between IoT devices, while the scale-free network is generated by considering the power law [13] in which the fraction of devices with k connections follows a distribution P (K) ∼ k −γ (γ is usually between 2 and 3). It is meaningful and reasonable to test SAND on a scale-free network, since most of the social networks are scale-free and SAND works by considering an overlay device social network. In our simulation, we generate a network that consists of 20000 IoT devices. Each IoT device is associated with a number of connections to its neighboring devices (the number of connections follows uniform distribution and power law in the random network and the scale-free network, respectively), which forms the communication network. Each physical link in the communication network has a transmission latency uniformly distributed within [10ms, 40ms]. The distributed scheme only runs on this communication network. In SAND, for each IoT device, we randomly generate three features. Any two IoT devices that are physically connected and share the same features will be equipped with a social link, which forms the overlay device social network. The total number of features in the network is set to vary from 2000 to 10000. For all the simulation results shown below, we generate 15000 discovery requests from randomly selected source devices with randomly chosen desired features, and obtain the average results. Each discovery request has a time to live of 60s. With regards to the evaluation, we focus on three performance metrics, which are the success rate, the average number of devices contacted on success and the average number of hops of the discovery path on success. We will present the simulation results and findings in the following parts. A. Success Rate In the simulation, if the target device with the desired feature can be found before the time to live expires, it is considered as a successful discovery; otherwise, it is considered as a failure and the discovery request will be dropped. The success rate is defined as the number of successful discovery divided by the total number of discovery requests. We show the success rate as a function of the number of features in Fig. 2(a) and Fig. 2(b). Given a fixed number of IoT devices (i.e., 15000 by default) in the system, as the number of features increases, the system become more diverse with a variety of types of devices. From Fig. 2(a), we can see that in the random network the success rate of both the distributed scheme and SAND decreases as the network becomes more diverse. The distributed scheme performs better than SAND since the distributed scheme uses simple broadcast while SAND uses the depth-first search which may yield to a long discovery period that exceeds the required time to live. 100.00% 100.00% Distributed Broadcast SAND 90.00% Success Rate Success Rate 90.00% 80.00% 70.00% 80.00% 70.00% 60.00% 60.00% 50.00% 50.00% Broadcast Distributed SAND 2000 4000 6000 8000 10000 No. of Features (a) Success Rate in Random Networks 2000 4000 6000 8000 10000 No. of Features (b) Success Rate in Scale-free Networks Fig. 2: Simulation Results in the Random Network and the Scale-free Network However, in the scale-free network as shown in Fig. 2(b), we can see that SAND achieves a similar success rate as that of the distributed scheme, which is as high as more than 98%. Compared to the results in the random network, the success rate experiences a significant increase. This is primarily because in the scale-free network there are some superhub devices that have a large number of connections, which could help the discovery request reach the target device in a short time. While in the random network, each device has a similar node degree and the discovery request may take a long time to reach the target, and in some bad cases the time to live requirement is violated and thus failures occur. Another reason is that the device ranking criteria in SAND is effective and can successfully navigate the discovery request to the desired target device. Given that most of the real-world social networks are scale-free, we can infer that SAND is a practically solid solution since its success rate is above 98%. B. Average Number of Devices Contacted One of our main interests in this research is to measure, during a device discovery process, how many devices are being contacted before the desired target device is found. It is more energy efficient if less number of devices are involved in transmitting the discovery message, which is critically important for IoT system where devices are power constrained. Fig. 3(b) and Fig. 3(b) show the number of devices contacted with different number of features in the random network and the scale-free network, respectively. Here, we only count the measures on successful discovery. From the simulation results, we can observe that in both random and scale-free networks, as the number of features increases, SAND outperforms the distributed scheme, with an average performance improvement of 14.6% and 13.9%, respectively. It can also be seen that as the number of features increases, the performance improvement of SAND over the distributed scheme becomes more significant in both networks. This indicates that SAND is more efficient when the IoT system is more diverse and heterogeneous. Thanks to the adoption of depth-first search and the intelligent device ranking criteria, SAND can efficiently discover the desired target device without involving too many unnecessary intermediate devices. Thus, SAND can potentially achieve a low energy consumption in the process of transmitting the discovery request, which is critically beneficiary to the IoT where the devices are power constrained. C. Communication Hops Another interest in this research is to find out how many communication hops are there to separate a source device from a desired target device. After the discovery process is done, the source and target devices will communicate with each other over the discovery path and cooperate to perform computing tasks. The post-discovery communication can be cost-efficient if the number of hops of the discovery path is small. In Table I and II, we show the average number of hops of the discovery path in the random network and the scale-free network, respectively. It can be seen that SAND can achieve as small communication hops as that of the distributed scheme, the latter of which is supposed to be the optimal one because of using the breadth-first search. This finding further validates the effectiveness of the device ranking criteria of SAND, which can intelligently navigate the discovery to the desired target device with minimum detours. We can also see that as the number of features increases, the number of hops increases in both network scenarios. This is reasonable because it becomes more difficult and needs more hops of query to find the desired target device when the network becomes more diverse. Comparing the results in the two tables, we can also observe that from the random network to the scale-free network, the average number of hops exhibits an average increase of two hops. The reason behind is that the random network is a regular mesh network where all the devices are relatively fair and have similar number of connections, while the scale-free network has some superhub devices that have huge number of connections. In the scale-free network, a given discovery request is usually forwarded to those superhub devices first and then finds a path to the desired target device, thus resulting in a larger number of hops than that of the random network. Furthermore, we evaluate the distribution of devices over the number of hops of the discovery path in the scale-free network. As shown in Fig. 4, we plot the device distributions for SAND 4,000.00 3,000.00 2,000.00 1,000.00 Distributed Broadcast SAND Avg. No. of Devices Contacted Avg. No. of Devices Contacted 4,000.00 3,000.00 2,000.00 1,000.00 Broadcast Distributed SAND 0.00 0.00 2000 4000 6000 8000 10000 No. of Features (a) Avg. No. of Devices Contacted in Random Networks 2000 4000 6000 8000 10000 No. of Features (b) Avg. No. of Devices Contacted in Scale-free Networks Fig. 3: Simulation Results in the Random Network and the Scale-free Network TABLE I: No. of Hops in Random Networks 5000 2000 2000 4000 6000 8000 10000 Broadcast 3.67 4.16 4.43 4.53 4.82 SAND 3.66 4.14 4.28 4.40 4.69 TABLE II: No. of Hops in Scale-free Networks No. of Features 2000 4000 6000 8000 10000 Broadcast 5.41 6.15 6.37 6.47 6.70 SAND 5.41 6.14 6.36 6.41 6.69 6000 4000 No. of Devices No. of Features 10000 3000 2000 1000 0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 No. of Hops using 2000, 6000 and 10000 features. The result confirms the past research on social networks [9][10], and shows that in most of the cases any two IoT devices are separated by 5∼7 hops. We can also observe that the average number of hops is small when the network is less diverse (see the center of the peak of distribution with 2000 features), and the average number of hops increases when the network becomes more diverse (e.g., 6000 and 10000 features). We can also infer that if a device cannot be discovered in a few hops, the probability that the discovery is unsuccessful increases dramatically. The reason behind is that the unsuccessful discovery involves devices that are relatively isolated from all other devices. It is quite common that the discovery to those isolated devices will violate the time to live requirement and get dropped. In contrast, the successful discovery usually involves devices that are in a cluster or connected to the superhubs. V. R ESEARCH C HALLENGES AND O PPORTUNITIES An efficient IoT device discovery may lead to the generation of a new set of applications and services, as well as research opportunities, such as self-organized computing, service and data fusion, malware source traceback and the formation of a robot society, etc. Meanwhile, there are many open challenges with regards to these applications. In the following parts, we propose a few research opportunities and challenges related to the topic of intelligent device discovery in the IoT, and beyond. Fig. 4: Device Distribution v.s. No. of Hops in SAND with 20K nodes A. Self-Organized Computing The intelligent device discovery can enable a new format of computing services that can be self-organized by a number of smart devices. A client can initiate a computing task from any device (e.g., a smart watch), with a set of specified requirements such as computing power demand (e.g., CPU and memory), function or platform demand, and data demand (e.g., source code or raw dataset). The source IoT device is responsible for deliverying these requirements to the appropriate computing facilities in the Internet and gathering the results back. In this case, an intelligent device discovery strategy can be used to establish connections between IoT devices and the computing facilities in the Internet. For example, a user wants to issue a request to check his/her glucose level from the smart watch. The smart watch can use the intelligent device discovery to search for the nearest server that has enough computing and memory power to perform the computing tasks. Then, the server may use the intelligent device discovery to find the source code of glucose level analyzer, and then download and install the analyzer to perform the analysis. The intelligent device discovery grants IoT devices and computing facilities the capability to fetch the data and functions as they need for performing a given computing tasks. Thus, the IoT devices and computing facilities can perform the computing tasks in a self-organized manner. In this process, one of the open challenges is how to standardize the experssion of the computing requirements so that the devices and machines can initiate standard requests that can be understood by other machines. Another challenge is how to optimize the request and reply flow routes so that the network resource consumption can be minimized and the request-response latency can be minimized. B. Computing, Network and Data Fusion In the above self-organized computing scenario, a variety of computing facilities, network functions/services and data (e.g., programming source code, patient’s raw data) can be interconnected and fused by running an intelligent device discovery strategy. For a given service request, the intelligent device discovery can help with gathering the computing, switching and storage hardware from a specific site, while fetching a pool of patients’ data from a different site and obtaining the source codes of particular computing functions or network functions from another site. Here, the computing, switching and storage hardware are served as placeholders, while the source codes of the computing and network functions and the raw dataset can be discovered and fetched by the intelligent device discovery strategy and served as the input to those hardware placeholders. Data, software and hardware can be distributed in different sites, while they are coupled by the intelligent device discovery approach. Hence, an intelligent device discovery strategy is promising to enable the computing, network and data fusion, and achieve an efficient communication between the fused devices. After the computing, network and data fusion becomes true, devices and machines can easily launch new services by specifying their needs to the Internet. The Internet can resolve the needs in a standard format (e.g., service function chains) and efficiently connect those required devices and funcitons (e.g., through intelligent device discovery strategy) to provide the required new services specified by the devices and machines. With the computing, network and data fusion, it is promising to generate novel machine-initiative services that are not supervised by human beings. Fusion can be considered as an intermediate phase that prepares and enables the formation of a robot society. To achieve this, an open challenge is how to develop a platform that can address the IoT interoperability issue and facilitate a computing, network service and data fusion platform. Another challenge is how to efficiently discover and deliver the required data and source codes to the selected hardware sites in a timely manner such that a seamless service can be provided for the users. C. Traceback Cyber Attack Source Another application of intelligent device discovery is in the area of network security. The sources of malicious attacks are usually spoofed to avoid being detected [14]. The intelligent device discovery can be used to traceback the actual source of malware attacks based on the characteristics and features of the malicious attacks. The security protection system can issue a traceback request for the malware source generator, with certain key features of the malicious traffic. The overlay social network on top of the communication network can be constructed based on the model in Section 3.1. Then, an enhanced model that can accurately predict the hidden links/relationships between machines can be constructed using the method in [15]. Machine learning or deep learning [16] techniques can also be applied here to improve the accruacy of the prediction. After that, the proposed intelligent device discovery strategy can run on this enhanced overlay social network (with hidden links based on prediction) to traceback the malicious source attacker. Even if the malicious source are spoofed and hidden behind some links or relationships, the intelligent device discovery can still guide the traceback route based on the pattern and features of malicious attacks. D. Robot Society The intelligent device discovery can be used to initiate the connection between a given device and its desired device for the first time. After the first contact, the two devices can be considered to know each other and maintain a certain type of social relationship as human does. As the communications between different devices continue to grow, a variety of computing service, network service and data can be fused, and a robot society could be formed which consists of IoT devices, servers, robots and all the machines in the Internet. The device relationship in the robot society can be the same as the human relationship in the human society, such as friend, colleague and family relationships (which are considered in this paper). There is also a high chance that some unique machine type of relationships may exist in the robot society. One of the challenging and meaningful work is to analyze a large dataset of network traffic data to categorize the relationships between machines/devices in the Internet. This can be useful to give us an insight of the relationships and communication purpose between machines. It is also helpful for us to formulate the foundamental rules of how to establish the social relationship between machines/devices. Another open challenge is how to ensure the security of the connection before the two machines/devices establish their social relationship. It is necessary to avoid establishing relationships with potential malicious devices. To achieve this, we may need to develop machine learning based solutions (e.g., decision tree classification, convolutional neural networks or recurrent neural networks) to train the machines/devices to identify their own roles and their contact’s roles in the robot society, to reduce the chance of establishing connections with malicious devices. In general, a promising research area is to investigate the social behavior and the collaborative intelligence in the robot society, in addition to the artificial intelligence in a single device/machine. VI. C ONCLUSION AND F UTURE W ORK The Internet of Things (IoT) grows continuously and connects billions of smart devices with heterogeneous functionalities, purposes and platforms. As the IoT evolves, it is promising that human-like social relationships could be autonomously established by the smart devices. With the help of an intelligent device discovery strategy, IoT devices can easily discover the resources (e.g., source codes) and devices (e.g., computing facilities) that meet their needs. Each device/machine is associated with one or more resources or capabilities of either computing power, network function, source code or data. The intelligent device discovery can be used to glue and couple the resources from distributed sites to achieve the computing, network service and data fusion. Each device/machine acts like a robot and can establish social connections with each other autonomously to collaborate and create new applications and services, which could eventually lead to the formation of a robot society. The goal of this research is to address the impending scalability issue in the process of device discovery in IoT. In this paper, we have proposed a novel approach that leverages such social aspects of IoT devices to achieve a scalable and efficient IoT device discovery. The proposed Social-Aware and Distributed (SAND) scheme applies the depth-first search and forwards the discovery request to neighboring devices in a preferred order defined by a novel social network-aware device ranking criteria. More specifically, the ranking criteria takes into consideration the device’s degree, diversity and clustering coefficient, which could potentially navigate the discovery request to reach a broad community. Furthermore, the local betweenness is considered in the ranking criteria, which prioritizes the devices that stands at the critical intersections of multiple shortest paths of the network. With the guidance of such an intelligent device ranking criteria, the discovery request can find the desired target device in a fast and efficient manner. We have conducted comprehensive simulations to validate the effectiveness of SAND on both a random network and a scale-free network. Simulation results have shown that SAND can achieve the near-optimal success rate and communication hops as that of the distributed scheme which uses broadcast. In addition, we have found that the SAND scheme contacts a much smaller number of intermediate devices than that of the distributed scheme during the discovery process, thus potentially leading to a much less energy consumption. As for future work, we plan to extend our study in the following aspects. First, we plan to extend SAND with multicast feature, which allows one discovery request to fetch multiple replies. This could be useful in wireless sensor network applications, such as in the temperature monitor system. In addition, rather than specifying only one feature of the desired target device, we can enhance the SAND scheme by allowing it to claim for multiple features to achieve a more accurate discovery. We also plan to implement SAND on a physical testbed and conduct experimental studies. Secondly, we plan to analyze various network traffic datasets to understand the meanings and purposes behind those communications, and we can use machine learning techniques to categorize the social relationships in the IoT. This can be used to further improve the discovery speed and the discovery accuracy of SAND, and can also provide the basic insights for exploring the formation of the robot society. Thirdly, we plan to investigate the interoperability issues in the IoT and explore a standard platform that enables a variety of computing, network service and data fusion. R EFERENCES [1] A. Zaslavsky and P. P. Jayaraman, “Discovery in the internet of things,” in ACM Ubiquity Magazine, October 2015, Article No. 2. [2] G. Sun, V. Chang, G. Yang and D. Liao, “The cost-efficient deployment of replica servers in virtual content distribution networks for data fusion,” in Information Sciences (2017), http://dx.doi.org/10.1016/j.ins.2017.08.021 [3] G. Sun, D. Liao, D. Zhao, Z. Sun, V. Chang, “Towards provisioning hybrid virtual networks in federated cloud data centers,” in Future Generation Computer Systems, 2017, https://doi.org/10.1016/j.future.2017.09.065 [4] C. Li, F. Darema and V. Chang, “Distributed behavior model orchestration in cognitive internet of things solution,” in Journal of Enterprise Information Systems, 2017, pp. 1-21. [5] P. Barnaghi and A. Sheth, “On searching the internet of things: requirements and challenges,” in IEEE Intelligent Systems, vol. 31, no. 6, 2016, pp. 71-75. [6] P. C. Ccori, L. C. De Biase, M. K. Zuffo and F. S. Da Silva, “Device discovery strategies for the IoT,” in Proc. of IEEE International Symposium on Consumer Electronics, 2016, pp. 97-98. [7] W. T. Lunardi, E. De Matos, R. Tiburski, et. al., “Context-based search engine for industrial IoT: discovery, search, selection and usage of devices,” in Proc. of IEEE Conference on Emerging Technologies and Factory Automation, 2015, pp. 8-11. [8] L. Dong, R. Ravindran and G. Wang, “ICN based distributed IoT resource discovery and routing,” in Proc. of IEEE International Conference on Telecommunications, 2016, pp. 1-7. [9] J. Travers and S. Milgram, “An experimental study of the small world problem,” in Sociometry, vol. 32, no. 4, 1969, pp. 425-443. [10] S. Edunov, C. Diuk, I. O. Filiz and S. B. Burke, “Three and a half degrees of separation,” 2016. [Online]. Available: https://research.fb.com/threeand-a-half-degrees-of-separation. [11] D. J. Watts and Steven Strogatz, “Collective dynamics of small-world networks,” in Nature, vol. 393, June 1998, pp. 440-442. [12] M. E. J. Newman, “Networks: an introduction,” Oxford University Press, 2010, ISBN 978-0199206650. [13] A. Barabsi and R. Albert, “Emergence of scaling in random networks,” in Science, vol. 286, 1999, pp. 509-512. [14] Zhen Ling ; Junzhou Luo ; Kui Wu ; Wei Yu ; Xinwen Fu, “TorWard: Discovery, Blocking, and Traceback of Malicious Traffic Over Tor,” in IEEE Transactions on Information Forensics and Security, vol. 10, no. 12, pp. 2515-2530, 2015. [15] Berlusconi G, Calderoni F, Parolini N, Verani M, Piccardi C, “Link Prediction in Criminal Networks: A Tool for Criminal Intelligence Analysis,” in PLoS ONE 11(4): e0154244. https://doi.org/10.1371/journal.pone.0154244 [16] R. Chauhan, H. Kaur and V. Chang, “Advancement and applicability of classifiers for variant exponential model to optimize the accuracy for deep learning,” in Journal of Ambient Intelligence and Humanized Computing, 2017, pp. 1-10.	2cs.AI
SOFIC ENTROPY OF GAUSSIAN ACTIONS arXiv:1509.07835v2 [math.DS] 31 Dec 2015 BEN HAYES Abstract. Associated to any orthogonal representation of a countable discrete group is a probability measure-preserving action called the Gaussian action. Using the Polish model formalism we developed before, we compute the entropy (in the sense of Bowen, Kerr-Li) of Gaussian actions when the group is sofic. Computations of entropy for Gaussian actions has only been done when the acting group is abelian and thus our results are new even in the amenable case. Fundamental to our approach are methods of noncommutative harmonic analysis and C ∗ -algebras which replace the Fourier analysis used in the abelian case. Contents 1. Introduction 2. Preliminaries on Sofic Entropy 3. Generating Sets and Tightness 4. Entropy of Gaussian Actions 4.1. Definition of Gaussian Actions 4.2. Preliminaries on the Group von Neumann Algebra and Embedding Sequences 4.3. Preliminaries on Real Subspaces of the Left Regular Representation 4.4. Sofic Entropy of Gaussian Actions References 1 3 4 7 7 10 12 16 27 1. Introduction This paper is concerned with giving new computations for sofic entropy, specifically in computing entropy of Gaussian actions. Entropy for actions of Z is classical and goes back to work of Kolmogorov and Sinaı̌. Entropy roughly measures how chaotic the action of Z is. Kieffer in [20] showed that one can generalize entropy to actions of amenable groups. An amenable group is a group which has a sequence of nonempty finite sets which are almost invariant under translation by elements of the group. Abelian groups, nilpotent groups and solvable groups are amenable, but the free group on r letters is not if r ≥ 2. Entropy for amenable groups has been studied by many people and is a useful invariant in ergodic theory: it can be computed in many cases and positivity of entropy implies interesting structural properties. Fundamental examples in [24] led many to believe that there cannot be a good entropy theory beyond the realm of amenable groups. In groundbreaking and seminal work Bowen in [1] defined a notion of entropy for actions of sofic groups. The class of sofic groups is considerably larger than that of amenable groups: it contains all amenable groups, all residually finite groups, all linear groups and is closed under free products with amalgamation over amenable subgroups (see [11],[10],[12],[25],[28]). Since the subject is fairly young, not as much is known about sofic entropy as entropy for actions of amenable groups but some structure is beginning to emerge. It can be calculated for some interesting examples such as Bernoulli shifts (see [1],[19]) as well as algebraic actions (see [2],[5],[15],[19]). Additionally, recent work in [22],[16] shows that one can deduce structural properties of an action from assumptions of positive sofic entropy. Our goal in this paper is to add to the list of computations of sofic entropy by computing the entropy of Gaussian actions. The Gaussian action construction is a way to associate, in a functorial way, to any orthogonal representation ρ : Γ → O(H) a probability measure-preserving action called the Gaussian action. Date: April 11, 2018. 2010 Mathematics Subject Classification. 37A35,28D15, 37A55, 22D25,43A65. Key words and phrases. Sofic entropy, Gaussian actions, noncommutative harmonic analysis. 1 2 BEN HAYES We refer the reader to Section 4.1 for the precise definition. For intuition we mention that if H is finitedimensional, then this action is just the induced action of Γ on H with the Gaussian measure. The Gaussian action is a natural generalization of this construction to the case of infinite-dimensional representations. Recall that the left regular representation λR : Γ → ℓ2 (Γ, R) is given by (λR (g)ξ)(h) = ξ(g −1 h), for ξ ∈ ℓ2 (Γ, R), g, h ∈ Γ. It is known that the Gaussian action associated to the left-regular representation Γ y ℓ2 (Γ, R) is the Bernoulli action with base (R, ν) where ν is the Gaussian measure. Thus Gaussian actions are a class of actions which are similar to the class of Bernoulli shifts. To the best of our knowledge, entropy for Gaussian actions has only been computed when the acting group is abelian (see [21] and [9]). To state our result, we need to introduce the following decomposition of representations. We define singularity of orthogonal representations exactly as in the unitary case. Given orthogonal representations ρ1 , ρ2 of Γ, we say that ρ1 is singular with respect to ρ2 and write ρ1 ⊥ ρ2 if no nonzero subrepresentation of ρ1 embeds into ρ2 . We say that ρ1 is absolutely continuous with respect to ρ2 , and write ρ1 ≪ ρ2 , if ρ1 embeds into ρ⊕∞ . For general pairs of orthogonal representations ρ1 , ρ2 we can 2 always write ρ1 = ρ1,a ⊕ ρ1,s where ρ1,s , ρ1 are singular and ρ1,a ≪ ρ2 . If Γ is abelian, this reduces to the Lebesgue decomposition in measure theory. We thus regard this as a noncommutative analogue of the Lebesgue decomposition. Theorem 1.1. Let Γ be a countable discrete sofic group with sofic approximation Σ. Let ρ : Γ → O(H) be an orthogonal representation with H separable. Let Γ y (Xρ , µρ ) be the induced Gaussian action; write ρ = ρa ⊕ ρs where ρs and λΓ,R are singular and ρa ≪ λΓ,R . Then   −∞, if hΣ,µρs (Xρs , Γ) = −∞ hΣ,µρ (Xρ , Γ) = 0, if ρa = 0 and hΣ,µρs (Xρs , Γ) 6= −∞   ∞, if ρa 6= 0 and hΣ,µρs (Xρs , Γ) 6= −∞. Since it appears to be new, we specifically mention the amenable case. Corollary 1.2. Let Γ be a countable discrete amenable group. Let ρ : Γ → O(H) be an orthogonal representation and Γ y (Xρ , µρ ) the induced Gaussian action. Write ρ = ρ1 ⊕ ρ2 where ρ2 is embeddable into Γ y ℓ2 (Γ, R)⊕∞ , and HomΓ (ρ1 , λΓ,R ) = {0}. Then hµρ (Xρ , Γ) = ∞ if and only if ρ2 6= 0, and if ρ2 = 0, then hµρ (Xρ , Γ) = 0. In [15] we gave a formula for entropy of a probability measure-preserving action Γ y (X, µ) when X is a Polish space, µ is a Borel probability measure, and Γ y X by homeomorphisms. Given an arbitrary probability measure-preserving action Γ y (Y, ν), a probability measure-preserving action Γ y (X, µ) where X is Polish, µ is a Borel measure on X and Γ y X by homeomorphisms is called a Polish model for Γ y (Y, ν) if Γ y (X, µ) ∼ = Γ y (Y, ν) as probability measure-preserving actions. Our definition of entropy in terms of a Polish model took into account the topology of X in a nontrivial way like the definition for entropy in the presence of a compact model developed by Kerr-Li in [19]. Our computation for the entropy of Gaussian actions goes through Polish models. This model is associated to a family of generators for the representation, and the measure is canonically defined in terms of the representation. Although one can write down a compact model, it is unnatural and the measure is not expressed nicely in terms of the representation. We mention that entropy in the sofic case is roughly a measure of “how many” finitary simulations a probability measure-preserving action has. The typical way to prove existence of these simulations is through a probabilistic argument. For the Gaussian action, our probabilistic argument uses Gaussian measures on finite-dimensional spaces. A consequence of our methods is that, associated to any sofic approximation of a group, we have a natural way of describing the Gaussian measure for a subrepresentation of the left regular representation as a limit of finite-dimensional Gaussian measures. Let us sketch how we are able to handle the case ρ is singular with respect to λR when Γ is nonabelian. In the case Γ = Z, the problem of showing that Γ y (Xρ , µρ ) has zero entropy reduces to the fact that if µ, ν are two singular measures on T = R/Z, then there is a f ∈ C(T) with 0 ≤ f ≤ 1 so that f is “close” to 1 in L2 (µ) and close to zero in L2 (ν). As the representation theory of a group is captured by its universal C ∗ -algebra it is natural to replace C(T) with C ∗ (Γ) in the nonabelian case. This statement about singularity SOFIC ENTROPY OF GAUSSIAN ACTIONS 3 of representations then becomes a statement about singularity of measures and we can prove an analogous characterization of singularity of representations in terms of elements of C ∗ (Γ). In short, we abstract the harmonic analysis present in the abelian case to noncommutative harmonic analysis using the framework of C ∗ -algebras. We mention that if Γ is amenable and is an infinite conjugacy class group (i.e. every nontrivial conjugacy class is infinite) then there is a easy proof that if ρ : Γ → O(H) embeds into λ⊕∞ Γ,R , then hµρ (Xρ , Γ) = ∞. Namely, one can show in this case that there is an n so that ρ⊕n contains λΓ,R (this is a consequence of the theory of II1 factors). In this case (Xρ , µρ ) will factor onto (Xλ , µλ ) which is isomorphic to Γ y (R, η)Γ where η is the Gaussian measure. Since entropy for amenable groups decreases under factor maps, we are done. This proof fails disastrously in the non-amenable case. It is very far from true that sofic entropy decreases under factor maps for non-amenable groups. In fact, it can be shown that for every nonamenable group there is a α > 0 so that if (X, µ) is a measure space with H(X, µ) ≥ α then Γ y (X, µ)Γ factors onto every nontrivial Bernoulli shift (see [4] Corollary 1.6), even a Bernoulli shift with infinite entropy. Moreover, if Γ contains a free group then α can be taken to be any positive number [3]. In fact, even more is true: a recent result of Seward (see [29], Theorem 1.1) implies that for every nonamenable sofic group Γ, there is an α > 0 so that any probability measure-preserving action of Γ is a factor of an action having sofic entropy less than α. Thus, there is no simple proof in the nonamenable case based on factors, and we must use a direct proof. Even in the amenable case this argument relies on the group being an infinite conjugacy class group and one needs general methods to handle the general case. 2. Preliminaries on Sofic Entropy We use Sn for the symmetric group on n letters. If A is a set, we will use Sym(A) for the set of bijections of A. Definition 2.1. Let Γ be a countable discrete group. A sofic approximation of Γ is a sequence Σ = (σi : Γ → Sdi ) of functions (not assumed to be homomorphisms) so that lim udi ({1 ≤ k ≤ di : σi (g)σi (h)(k) = σi (gh)(k)}) = 1, for all g, h ∈ Γ i→∞ lim udi ({1 ≤ k ≤ di : σi (g)(k) 6= σi (h)(k)}) = 1, for all g 6= h in Γ. i→∞ We will call Γ sofic if it has a sofic approximation. All amenable groups and residually finite groups are sofic. Also, it is known that soficity is closed under free products with amalgamation over amenable subgroups (see [12],[25],[11],[10], [28]). Additionally, residually sofic groups and locally sofic groups are sofic. Thus by Malcev’s Theorem we know all linear groups are sofic. It is known that graph products of sofic groups are sofic by [7]. If Λ is a subgroup of Γ, and Λ is sofic and Γ y Γ/Λ is amenable (in the sense of having a Γ-invariant mean) then Γ is sofic. One can argue this by the same methods of Theorem 1 of [12] (for example consider the observations after Definition 12.2.12 of [6]). We recall the definition of entropy in the presence of a Polish model given in [16]. Let X be a Polish space and let Γ be a countable, discrete group with Γ y X by homeomorphisms. We say that a continuous pseudometric ∆ on X is dynamically generating if for every x ∈ X and every neighborhood U of x in X there exists a finite F ⊆ Γ and a δ > 0 so that if y ∈ X and max ∆(gx, gy) < δ g∈F then y ∈ U. Notice that our definition includes the hypothesis that ∆ is continuous. We use Cb (X) for the Banach space of bounded, continuous functions on X with norm kf k = sup \|f (x)\|. x∈X Given a pseudometric space (X, ∆), A, B ⊆ X and ε > 0, we say that A is ε-contained in B and write A ⊆ε B if for any a ∈ A, there is a b ∈ B so that ∆(a, b) ≤ ε. We say that A ⊆ X is ε-dense if X ⊆ε A. We use Sε (X, ∆) for the minimal cardinality of an ε-dense subset of X. We say that A ⊆ X is ε-separated if for any a1 6= a2 in A we have ∆(a1 , a2 ) > ε. We use Nε (A, ∆) for the maximal cardinality of an ε-separated subset of X. We leave it as an exercise to show that (1) N2ε (A, ∆) ≤ Sε (A, ∆) ≤ Nε (A, ∆), 4 BEN HAYES and that if δ > 0 and A ⊆δ B then S2(ε+δ) (A, ∆) ≤ Sε (B, ∆). We use ∆2 for the metric on X n defined by n 1X ∆(φ(j), ψ(j))2 . ∆2 (φ, ψ)2 = n j=1 If X is a Polish space we use Prob(X) for the space of all Borel probability measures on X. If Γ is a countable, discrete group with Γ y X by homeomorphisms we use ProbΓ (X) for the space of all Γ-invariant elements of Prob(X). We are now ready to state our definition of sofic entropy from [16]. It is defined by counting the exponential growth of maps from {1, . . . , di } → X which approximately preserve the measure-theoretic structure of X and are approximately equivariant. We call these maps microstates (this is a heuristic term and will not be defined rigorously). Definition 2.2. Let Γ be a countable, discrete, sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let X be Polish space with Γ y X by homeomorphisms. Fix a bounded, continuous pseudometric on X. For a finite F ⊆ Γ and δ > 0 we let Map(∆, F, δ, σi ) be all φ ∈ X di so that max ∆2 (gφ, φ ◦ σi (g)) < δ. g∈F Notice that Map(∆, F, δ, σi ) only accounts for the group action and not the measure-theoretic structure of X. Recall that Cb (X) denotes the space of bounded continuous functions on X. For µ ∈ Prob(X), a finite L ⊆ Cb (X), and a δ > 0 we let Z Z \ UL,δ (µ) = ν ∈ Prob(X) : f dµ − f dν < δ . f ∈L The sets UL,δ (µ) form a basis of open sets for a topology called the weak topology. We use this topology to account for the measure-theoretic structure in our microstates. Definition 2.3. Let Γ be a countable, discrete, sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let X be Polish space with Γ y X by homeomorphisms and fix µ ∈ ProbΓ (X). For finite F ⊆ Γ, L ⊆ Cb (X) and δ > 0 we set Mapµ (∆, F, δ, L, σi ) to be the set of all φ ∈ Map(∆, F, δ, L, σi ) so that φ∗ (udi ) ∈ UL,δ (µ). Definition 2.4. Let Γ be a countable, discrete, sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let X be Polish space with Γ y X by homeomorphisms and let µ ∈ ProbΓ (X). Fix a bounded, dynamically generating pseudometric on X. For finite F ⊆ Γ, L ⊆ Cb (X) and δ, ε > 0 we set 1 hΣ,µ (∆, F, δ, L, ε) = lim sup log Nε (Mapµ (∆, F, δ, L, σi ), ∆2 ), d i→∞ i hΣ,µ (∆, ε) = inf F ⊆Γ finite, L⊆Cb (X) finite,δ>0 hΣ,µ (∆, F, δ, L, ε), hΣ,µ (X, Γ) = sup hΣ,µ (∆, ε). ε>0 We call hΣ,µ (X, Γ) the entropy of Γ y (X, µ) with respect to Σ. It is shown in [16] that this agrees with entropy as defined by [1],[19]. 3. Generating Sets and Tightness Since Cb (X) is not separable, we would like to reduce checking the approximate measure-preserving property of our microstates from all functions in Cb (X) to a smaller class of functions. For example Cb (X) is separable in the topology of uniform convergence on compact sets, so if we require that we have a family of functions dense in this topology then this will give us a sufficiently small family of functions to deal with. However, for this to work we need to modify our microstates so that they have some uniform tightness. We proceed with the definitions. Definition 3.1. Let X be a Polish space. A family L ⊆ Cb (X) is said to be generating if there is a A > 0 so that for every g ∈ Cb (X), for every compact K ⊆ X and for every ε > 0 there is a f ∈ Span(L) so that kf K − g K k < ε and kf k ≤ Akgk. SOFIC ENTROPY OF GAUSSIAN ACTIONS 5 We now proceed with our modified version of sofic entropy in the case of a generating set of functions. Definition 3.2. Let Γ be a countable discrete group and σ : Γ → Sd a function for some d ∈ N. Let X be a Polish space with Γ y X by homeomorphisms preserving a Borel probability measure µ. Let ∆ be a dynamically generating pseudometric on X. For an open subset U of X, for η, δ > 0 and finite L ⊆ Cb (X), F ⊆ Γ we let MapU,η µ (∆, F, δ, L, σ) be the set of φ ∈ Mapµ (∆, F, δ, L, σ) so that φ∗ (ud )(U ) > 1 − η. Suppose now that Γ is sofic. For a dynamically generating pseudometric ∆, L ⊆ Cb (X) and a sofic approximation Σ = (σi : Γ → Sdi ) set: 1 hU,η log Nε (MapU,η µ (∆, F, δ, L, σi )), Σ,µ (∆, ε, F, δ, L) = lim sup i→∞ di hK,η Σ,µ (∆, ε, F, δ, L) = hηΣ,µ (∆, ε, F, δ, L) = hΣ,µ (∆, ε, L) = inf open U ⊇ K hU,η Σ,µ (∆, ε, F, δ, L), sup hK,η Σ (∆, ε, F, δ, L) K⊆X compact inf hηΣ,µ (∆, ε, F, δ, L) η,δ>0, F ⊆Γ finite, L⊆L finite hΣ,µ (∆, L) = sup hΣ,µ (∆, ε, L). ε>0 In the above definition, the necessary trick is to add more quantifiers to the definition of sofic entropy. The reader may be concerned already by the number of quantifiers involved in the original definition of sofic entropy. When we compute sofic entropy of Gaussian actions in Section 4.1 it will be clear that this is the correct tradeoff. The difficulty involved in the computation will not be in dealing with quantifiers but instead that we can only show the approximate measure-preserving property on a class of functions which is not norm dense. However, one can easily see that this class of functions is generating and so we will use the above definition. Theorem 3.3. Let Γ be a countable discrete sofic group with sofic approximation Σ. Let X be a Polish space with Γ y X by homeomorphisms preserving a Borel probability measure µ. For any dynamically generating pseudometric ∆ on X, L ⊆ Cb (X) generating, and ε > 0 In particular, hΣ,µ (∆, ε) = hΣ,µ (∆, ε, L). hΣ,µ (X, Γ) = hΣ,µ (∆, L). Proof. The “in particular” part follows from Theorem 3.12 of [16]. We first show that hΣ,µ (∆, ε) ≤ hΣ,µ (∆, ε, L). Let η > 0. Since X is Polish, we may find a compact K ⊆ X with µ(X \ K) < η. Fix an open subset U of X containing K. By Urysohn’s Lemma we may find a f ∈ Cb (X) with Note that if ν ∈ Prob(X) and δ > 0 with χK ≤ f ≤ χU . Z X we have f dµ − Z f dν < δ, X ν(U ) > 1 − η − δ. We now see that for all finite L ⊆ L, F ⊆ Γ, 0 < δ < η we have MapU,η+δ (∆, F, δ, L, σi ) ⊇ Map(∆, F, δ, L ∪ {f }, σi ). Thus for all δ < η, and all finite F ′ ⊇ F, L′ ⊇ L ∪ {f } < and δ ′ < δ, hΣ,µ (∆, ε, F ′ , δ ′ , L′ ) ≤ hU,2η Σ,µ (∆, ε, F, δ, L). 6 BEN HAYES Taking the infimum over F ′ , δ ′ , L′ we find hΣ,µ (∆, ε) ≤ hU,2η Σ,µ (∆, ε, F, δ, L). Taking the infimum over all U ⊇ K we find that K,2η hΣ,µ (∆, ε) ≤ hΣ,µ (∆, ε, F, δ, L). Now take the supremum over all K and then the infimum over all η, δ, F, L ⊆ L to find that hΣ,µ (∆, ε) ≤ hΣ,µ (∆, ε, L). We now show that hΣ,µ (∆, ε, L) ≤ hΣ,µ (∆, ε). Let A > 0 be as in the definition of generating. Let F ⊆ Γ and L ⊆ Cb (X) be given finite sets and δ > 0. Let κ > 0 be sufficiently small depending upon F, L, δ in a manner to be determined later. Since X is Polish, we may find a compact K0 ⊆ X so that µ(K0 ) ≥ 1 − κ. Let K ⊆ X compact be given with K ⊇ K0 . Let L = {f1 , . . . , fl }. For 1 ≤ j ≤ l we can find a gj ∈ Span(L) so that kgj K − fj K k < δ and kgj k ≤ Akfj k. We may find an open neighborhood U0 of K so that kgj U0 − fj k U0 Cb (U0 ) < δ, for j = 1, . . . , l. Let δ ′ > 0 and F ′ ⊆ Γ and L′ ⊆ Cb (X) be finite sets which will depend upon F, δ, L in a manner to be ′ ′ ′ determined shortly. Let U be an open neighborhood of K contained in U0 . For φ ∈ MapU,η µ (∆, F , δ , L , σi ) we have Z Z Z Z Z Z fj dµ − fj dφ∗ (udi ) ≤ gj dφ∗ (udi ) + fj dφ∗ (udi ) − gj dµ − gj dφ∗ (udi ) X X X X X X Z Z + fj dµ gj dµ − X X Z Z ≤ η(1 + A)kfj k + κ(1 + A)kfj k + gj dφ∗ (udi ) fj dφ∗ (udi ) − U U Z Z Z Z + gj dφ∗ (udi ) + gj dµ − fj dµ gj dµ − X X U U Z Z ≤ η(1 + A)kfj k + κ(1 + A)kfj k + 2δ + gj dµ − gj dφ∗ (udi ) X If we choose δ ′ > 0 sufficiently small, and F ′ , L′ sufficiently large we may force Z Z gj dφ∗ (udi ) < δ. gj dµ − X X If we choose max 2η(1 + A)kfj k < δ j max κ(1 + A)kfj k < δ j then we have Thus MapU,η (∆, F ′ , δ ′ , L′ , σi ) ⊆ Map(∆, F, 7δ, L, σi ). U,η ′ ′ ′ ′ ′ ′ hK,η Σ,µ (∆, ε, F , δ , L ) ≤ hΣ,µ (∆, ε, F , δ , L ) ≤ hΣ,µ (∆, ε, F, 7δ, L). Since K ⊇ K0 was arbitrary, we may take the supremum over all K to see that hΣ,µ (∆, ε, L) ≤ hηΣ,µ (∆, ε, F ′ , δ ′ , L′ ) ≤ hΣ,µ (∆, ε, F, 7δ, L). X SOFIC ENTROPY OF GAUSSIAN ACTIONS 7 Now taking the infimum over all F ′ , δ ′ , L′ proves that hΣ,µ (∆, ε, L) ≤ hΣ,µ (∆, ε). 4. Entropy of Gaussian Actions Gaussian actions are a natural class of actions induced by orthogonal representations of a group. When the representation is the left regular representation, the Gaussian action is simply the Bernoulli action on (R, ν)Γ where ν is the Gaussian measure. In [16], we prove some results which give structural results of the Koopman representation Γ y L2 (X, µ) of a probability measure-preserving action Γ y (X, µ) of a sofic group Γ, under the assumption that this action has positive entropy. For example, we could show that if Γ y (X, µ) has positive entropy then Γ y L2 (X, µ) must contain a “piece” of the left regular representation (for more precise statements see [16] Theorem 1.1 and Corollary 1.2). We will exploit the connections to representation theory to apply our spectral consequences of positive entropy in [16] to Gaussian actions. We will also exploit the similarity to Bernoulli shifts to compute the entropy of Gaussian actions. Let Γ be a countable discrete group. Recall that an orthogonal representation of Γ on a real Hilbert space H is a homomorphism ρ : Γ → O(H) where O(H) is the group of orthogonal transformations of H (i.e. the set of O ∈ B(H) so that hOξ, Oηi = hξ, ηi for ξ, η ∈ H.) We let HC = H ⊗R C = H + iH be the complexification of H equipped with the unique sesquilinear inner product extending the one on H. We let ρC : Γ → U(H + iH) be the complexification of ρ, i.e. ρC (g) for g ∈ Γ is the unique unitary transformation so that ρC (g)(ξ) = ρ(g)ξ for ξ ∈ H. 4.1. Definition of Gaussian Actions. The most natural way to define Gaussian actions is by von Neumann algebras. Definition 4.1. Let H be a complex Hilbert space. A von Neumann algebra on H is a ∗-subalgebra of B(H) containing the identity and closed in the weak operator topology. We say that a vector ξ ∈ H is cyclic for M if M ξ = H. by Suppose that (X, µ) is a standard probability space. For f ∈ L∞ (X, µ), let Mf ∈ B(L2 (X, µ)) be given (Mf ξ)(x) = f (x)ξ(x). The map f 7→ Mf allows us to view L∞ (X, µ) as a von Neumann algebra. It turns out (see [8] Theorem 14.5) that if M ⊆ B(H) is a commutative von Neumann algebra with cyclic vector ξ with kξk2 = 1 and H is separable, then there is a standard probability space (X, µ) and a unitary U : H → L2 (X, µ) so that U (ξ) = 1 U M U ∗ = {Mf : f ∈ L∞ (X, µ)}. Additionally, if φ : M → C is a linear functional so that φ {T ∈M:kT k≤1} is weak operator topology continuous, then there is a complex measure ν ≪ µ so that if U T U ∗ = Mf then Z φ(T ) = f dν. X ∗ We leave it to reader the verify that if U T U = Mf then Z f dµ. hT ξ, ξi = X 8 BEN HAYES Definition 4.2. Let H be a real Hilbert space. The Gaussian algebra associated to H, denoted A(H) is a commutative von Neumann algebra with cyclic vector Ω with kΩk = 1, which is generated by unitaries {u(ξ) : ξ ∈ H} satisfying u(ξ1 + ξ2 ) = u(ξ1 )u(ξ2 ) for ξ ∈ H hu(ξ)Ω, Ωi = exp(−πkξk2 ). For a ∈ A(H), we let φ(a) = haΩ, Ωi. Suppose that Γ is a countable discrete group and ρ : Γ → O(H) is a representation. Then there is a φpreserving action α on A(H) determined uniquely by α(g)(u(ξ)) = u(ρ(g)ξ), for ξ ∈ H. This action is called the Gaussian action. By [27] the Gaussian algebra exists and is unique up to φ-preserving isomorphism. See also [27] for the existence and uniqueness of the Gaussian action. Let us sketch an alternate construction of the Gaussian algebra. First consider the case that H is finite−πkxk2ℓ2 (n) dimensional. Say H = Rn . In this case we can simply take A(H) = L∞ (Rn , ν) where dν = e dx, and Z φ(a) = a(x) dν(x), Rn u(ξ)(x) = exp(2πihξ, xi). e In general, consider the universal ∗-algebra A(H) generated by unitaries u e(ξ), ξ ∈ H satisfying the relation u e(ξ1 + ξ2 ) = u e(ξ1 )e u(ξ2 ) for ξ1 , ξ2 ∈ H. Here one can make sense of a unitary in a ∗-algebra by saying that it is an element u so that u∗ u = uu∗ = 1. One then has to check that there is a well-defined linear function e φe : A(H) → C defined by e u(ξ)) = exp(−πkξk2 ). φ(e Checking that φe is well-defined reduces to showing that if we take ξ1 , . . . , ξn ,ζ1 , . . . , ζm ∈ H and λ1 , . . . , λn , c1 , . . . , cm ∈ C with m n X X cj u e(ζj ), λj u e(ξj ) = j=1 then n X j=1 λj exp(−πkξj k2 ) = i=1 m X j=1 cj exp(−πkζj k2 ). Replacing H with the span of {ξ1 , · · · , ξn , ζ1 , · · · , ζm } reduces the verification that φe is well-defined to the case that H is finite-dimensional, where we have already shown that φe is well-defined (by explicitly exhibiting it as integration with respect to the Gaussian measure as explained above). By similar reasoning, one checks e ∗ a) ≥ 0 for a ∈ A(H). e that φ(a One now runs the GNS construction. That is, we let K be the completion of e A(H) under the inner product e ∗ a). ha, bi = φ(b e We then have an induced homomorphism ρ : A(H) → B(K) given by SOT e (ρ(a))(b) = ab for b ∈ A(H). e We then let A(H) = ρ(A(H)) , and Ω = u e(0), u(ξ) = ρ(e u(ξ)). It is then straightforward to verify that A(H), Ω have the desired properties. By our remarks before this definition if ρ, Γ are as in the definition then there is a standard probability space (Xρ , µρ ) and a measure-preserving action Γ y (Xρ , µρ ) so that ∼ Γ y (A(H), φ). Γ y L∞ (Xρ , µρ ) = Furthermore, this action is uniquely determined up to isomorphism. Note that for two orthogonal representations ρj : Γ → O(Hj ), j = 1, 2 SOFIC ENTROPY OF GAUSSIAN ACTIONS 9 we have Γ y (Xρ1 ⊕ρ2 , µρ1 ⊕ρ2 ) ∼ = (Xρ1 × Xρ2 , µρ1 ⊗ µρ2 ). The definition via von Neumann algebras may be abstract, so let us mention a simple version of the definition in the case of a cyclic representation. Recall that if t ∈ cc (Γ, R) and x ∈ RΓ we use X t·x = t(g)x(g). g∈Γ Proposition 4.3. Let Γ be a countable discrete group and ρ : Γ → O(H) an orthogonal representation. Suppose that there is a vector ξ ∈ H so that H = Span{ρ(g)ξ : g ∈ Γ}. Then the Gaussian action is isomorphic to the shift action on RΓ with measure determined by   2 Z X   exp(2πit · x) dµξ (x) = exp −π t(g)ρ(g)ξ  for all t ∈ cc (Γ, R). RΓ g∈Γ Proof. Choose a realization Γ y (A(H), φ) ∼ = Γ y L∞ (Xρ , µρ ). By [30] Proposition 5.3, a sequence an ∈ A(H) converges in the strong operator topology to a ∈ A(H) if and only if φ((a − an )∗ (a − an )) → 0. Thus, t → 7 u(tξ) is a strongly continuous one-parameter group in A(H). Fix ζ ∈ H. By Stone’s Theorem there is a closed, densely-defined self-adjoint operator ω(ζ) on L2 (Xρ , µρ ) so that for all s ∈ R exp(2πisω(ζ)) = u(sζ), in the sense of functional calculus. Recall that L∞ (Xρ , µρ ) can be viewed as a von Neumann algebra (by multiplication operators). It is not hard to see that the elements u ∈ L∞ (Xρ , µρ ) whose multiplication operators are unitaries are almost everywhere equal to measurable functions u : X → {z ∈ C : \|z\| = 1}. Thus we can identify u(sζ) as a measurable function Xρ → {z ∈ C : \|z\| = 1}. For similar reasons, we can identify ω(ζ) with a measurable function Xρ → R. Thus for every s ∈ R, it is true that for almost every x ∈ Xρ we have exp(2πisω(ζ)(x)) = u(sζ)(x). Define Φ : X ρ → RΓ by Φ(x)(g) = ω(ξ)(g −1 x). Then it is not hard to see that for every s ∈ R and for all g ∈ Γ we have exp(2πisΦ(x)(g)) = u(sρ(g)ξ)(x) for almost every x ∈ Xρ . Define µξ = Φ∗ µρ . Since the unitaries of the form u(sρ(g)ξ), s ∈ R, g ∈ Γ generate all of A(H) we see that Φ gives an Γ-equivariant isomorphism (RΓ , µξ ) ∼ = (Xρ , µρ ). 10 BEN HAYES Additionally for all t ∈ cc (Γ, R) Z RΓ exp(2πit · x) dµξ (x) = Z X  exp 2πi   = φ ω   X g∈Γ  = exp −π X g∈Γ  t(g)Φ(x)(g) dµρ (g)  t(g)ρ(g)ξ  X 2 t(g)ρ(g) g∈Γ   . 4.2. Preliminaries on the Group von Neumann Algebra and Embedding Sequences. For our purposes we will need to “linearize” a sofic approximation to approximations of algebras associated to Γ. Let C(Γ) be the ring of finite formal linear combinations of elements of Γ with addition defined naturally and multiplication defined by   ! ! X X X X  ag g  bh h = ah bh−1 g g. g∈Γ h∈Γ g∈Γ h∈Γ We will also define a conjugate-linear involution on C(Γ) by  ∗ X X  ag−1 g. ag g  = g∈Γ g∈Γ P Given a sofic approximation Σ = (σi : Γ → Sdi ) and α = g∈Γ αg g ∈ C(Γ) we define σi (α) ∈ Mdi (C) by X σi (α) = αg σi (g). g∈Γ In order to talk about the asymptotic properties of this extended sofic approximation, we will need a more analytic object associated to Γ. Let λ : Γ → U(ℓ2 (Γ)) be the left regular representation defined by (λ(g)ξ)(h) = ξ(g −1 h). We will continue to use λ for the linear extension to C(Γ) → B(ℓ2 (Γ)). The group von Neumann algebra of Γ is defined by λ(C(Γ)) W OT where WOT denotes the weak operator topology. We will use L(Γ) to denote the group von Neumann algebra. Define τ : L(Γ) → C by τ (x) = hxδe , δe i. We leave it as an exercise to the reader to verify that τ has the following properties. 1: τ (1) = 1, 2: τ (x∗ x) ≥ 0, with equality if and only if x = 0, 3: τ (xy) = τ (yx), for all x, y ∈ M, 4: τ is weak operator topology continuous. We call the third property the tracial property. We will typically view C(Γ) as a subset of L(Γ). In particular, we will use τ as well for the functional on C(Γ) which is just the restriction of τ on L(Γ). In order to state our linearization of a sofic approximation properly, we give a general definition. By definition, ∗-algebra is a complex algebra equipped with an involution ∗ which is conjugate linear and antimultiplicative. Definition 4.4. A tracial ∗-algebra is a pair (A, τ ) where A is a ∗-algebra equipped with a linear functional τ : A → C so that SOFIC ENTROPY OF GAUSSIAN ACTIONS 11 1: τ (1) = 1, 2: τ (x∗ x) ≥ 0, with equality if and only if x = 0, 3: τ (xy) = τ (yx), for all x, y ∈ M, 4: For all a ∈ A, there is a Ca > 0 so that for all x ∈ A, \|τ (x∗ a∗ ax)\| ≤ Ca τ (x∗ x). For a, b ∈ A we let ha, bi = τ (b∗ a) and we let kak2 = τ (a∗ a)1/2 . Define L2 (A, τ ) to be the Hilbert space completion of A in this inner product. By condition 4 of the definition, we have a representation λ : A → B(L2 (A, τ )) defined densely by λ(a)x = ax for x ∈ A. We let kak∞ = kλ(a)k. We make Mn (C) into a tracial ∗-algebra using tr = n1 Tr where Tr is the usual trace. In particular, we use kAk2 = tr(A∗ A)1/2 and kAk∞ will denote the operator norm. We let C[X1 , . . . , Xn ] be the free ∗-algebra on n-generators X1 , . . . , Xn . We will call elements of C[X1 , . . . , Xn ] ∗-polynomials in n indeterminates. For a ∗-algebra A, for elements a1 , . . . , an ∈ A, there is a unique ∗homomorphism C[X1 , . . . , Xn ] → A sending Xj to aj . If P ∈ C[X1 , . . . , Xn ] we use P (a1 , . . . , an ) for the image of P under this ∗-homomorphism. Definition 4.5. Let (A, τ ) be a tracial ∗-algebra. An embedding sequence is a sequence Σ = (σi : A → Mdi (C)) such that sup kσi (a)k∞ < ∞, for all a ∈ A i kP (σi (a1 ), . . . , σi (an )) − σi (P (a1 , . . . , an ))k2 → 0, for all a1 , . . . , an ∈ A and all P ∈ C[X1 , . . . , Xn ] tr(σi (a)) → τ (a) for all a ∈ A. We will frequently use the following fact: if x1 , . . . , xn ∈ A and P ∈ C[X1 , . . . , Xn ] then (2) kP (σi (x1 ), . . . , σi (xn ))k2 → kP (x1 , . . . , xn )k2 . To see this, we notice that since kP (σi (x1 ), . . . , σi (xn )) − σi (P (x1 , . . . , xn ))k2 → 0, it suffices to handle the case n = 1 and P (X) = X. Then, kσi (x)k22 = tr(σi (x)∗ σi (x)) and since kσi (x)∗ σi (x) − σi (x∗ x)k2 → 0 we have \| tr(σi (x)∗ σi (x)) − tr(σi (x∗ x))\| → 0. As tr(σi (x∗ x)) → kxk22 , we have proved (2). The proof of the next two propositions will be left to the reader. Proposition 4.6. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Extend Σ to maps σi : C(Γ) → Mdi (C) linearly. Then Σ is an embedding sequence of (C(Γ), τ ). Proposition 4.7. Let (A, τ ) be a tracial ∗-algebra and Σ = (σi : A → Mdi (C)) be an embedding sequence. If Σ′ = (σi′ : A → Mdi (C)) is another sequence of functions so that sup kσi′ (a)k∞ < ∞, for all a ∈ A, i then Σ′ is an embedding sequence. kσi (a) − σi′ (a)k2 → 0, for all a ∈ A, We will in fact need to extend our sofic approximation to the group von Neumann algebra. For this, we use the following. Lemma 4.8 (Lemma 5.5 in [17]). Let Γ be a countable discrete group. Then any embedding sequence for C(Γ) extends to one for L(Γ). We will use the preceding lemma when Γ is sofic, in combination with Proposition 4.6. 12 BEN HAYES 4.3. Preliminaries on Real Subspaces of the Left Regular Representation. We define the Fourier algebra of Γ as all functions φ : Γ → C so that there is a linear functional Φ : L(Γ) → C with Φ {x∈L(Γ):kxk∞ ≤1} being weak operator topology continuous so that Φ(λ(g)) = φ(g). We will call Φ the continuous extension of φ (note that by continuity and Kaplansky’s density theorem, Φ as above must be unique). We let A(Γ)+ consist of all such φ where the continuous extension of φ is a positive linear functional. We let kφkA(Γ) = kΦk. By [30] Theorem II.2.6 k · kA(Γ) is a norm on A(Γ) which makes A(Γ) a Banach space. By [30] Theorem V.3.15 we have that A(Γ) consists of all functions of the form g 7→ hλ(g)ξ, ηi where ξ, η ∈ ℓ2 (Γ). Moreover kφk = inf kξkkηk where infimum is over all ξ, η so that φ(g) = hλ(g)ξ, ηi. For intuition, we leave it to the reader to verify that when Γ is abelian, then A(Γ) consists of all fb b and that A(Γ)+ consists of all fb where f ∈ L1 (Γ) b and f ≥ 0. We state a few basic (and where f ∈ L1 (Γ), well-known) properties of A(Γ) in the following proposition. Lastly, if we are given x ∈ L(Γ) we set x b = xδe ∈ ℓ2 (Γ). Proposition 4.9. Let Γ be a countable discrete group. (1) If ρ : Γ → U(H) is a unitary representation with ρ ≪ λΓ , and ξ ∈ H, then φ(g) = hρ(g)ξ, ξi is in A(Γ)+ . In particular there is a ζ ∈ ℓ2 (Γ) with hρ(g)ξ, ξi = hλ(g)ζ, ζi. (2) If φ ∈ A(Γ), then φ ∈ A(Γ) and kφk = kφk. If φ ∈ A(Γ)+ , then so is φ. (3) For x ∈ L(Γ) we define φx : Γ → C by φx (g) = τ (xλ(g)), then φx ∈ A(Γ), and φx ∈ A(Γ)+ if and only if x ∈ L(Γ)+ . Additionally kφx k = τ (\|x\|), and {φx : x ∈ L(Γ)+ } is dense in A(Γ)+ . (4) If x, y ∈ L(Γ) and φ(g) = hλ(g)b x, x bi, ψ(g) = hλ(g)b y , ybi then kb x − ybk22 ≤ kφ − ψkA(Γ) . Proof. (1): The assumption ρ ≪ λΓ implies that we may extend ρ to a ∗-homomorphism ρ : L(Γ) → B(H) so that ρ {x∈L(Γ):kxk∞ ≤1} is weak operator topology continuous. Thus the continuous extension Φ of φ is given by Φ(x) = hρ(x)ξ, ξi. The “in particular” part follows from the discussion preceding the proposition. (2): Write φ(x) = hλ(x)ξ, ηi, 2 with ξ, η ∈ ℓ (Γ). Then φ(x) = hλ(x)ξ, ηi the conclusion follows easily from this equality. (3): If x ∈ L(Γ)+ , then the continuous extension Φ of φx is given by Φ(y) = τ (xy). For y ≥ 0, Φ(y) = τ (xy) = τ (y 1/2 xy 1/2 ) ≥ 0 as x ≥ 0. Conversely, if φx ∈ A(Γ)+ let Φ be the continuous extension of φx to L(Γ). Then for all y ∈ L(Γ) hxyδe , yδe i = τ (y ∗ xy) = τ (xyy ∗ ) = Φ(yy ∗ ) ≥ 0. Since L(Γ)δe = ℓ2 (Γ), we see that x ∈ L(Γ)+ . The norm equality and density statement are contained in [31] Lemma IX.2.12. SOFIC ENTROPY OF GAUSSIAN ACTIONS 13 (4): This is the Powers-Størmer inequality (see [6] Proposition 6.2.4 for a proof which generalizes to our situation). In order to compute the entropy of Gaussian actions, we will need to discuss real subspaces of ℓ2 (Γ), and for this we need a real version of L(Γ). We let LR (Γ) be all x ∈ L(Γ) so that x b ∈ ℓ2 (Γ, R). Recall that the Γ convolution between f ∈ cc (Γ), g ∈ C is defined by X f ∗ g(x) = f (h)g(h−1 x). h∈Γ Γ We similarly define f ∗ g if f ∈ C and g ∈ cc (Γ). If f ∈ cc (Γ), we use X fˇ = f (g)g ∈ C(Γ). g∈Γ Note that for all f ∈ cc (Γ), x ∈ L(Γ) we have xf = x b ∗ f. Thus kxk∞ = sup f ∈cc (Γ),kf k2 =1 kb x ∗ f k2 . c∗ . Thus for all f ∈ cc (Γ), b=x Define J : ℓ2 (Γ) → ℓ2 (Γ) by (Jf )(x) = f (x−1 ), and observe that J x c∗ ∗ f k2 ≤ kx∗ k∞ kf k2 = kxk∞ kf k2 . kf ∗ x bk2 = kJ x b ∗ Jf k2 = kx Replacing x with x∗ we see that kxk∞ = sup f ∈cc (Γ),kf k2 =1 kf ∗ x bk2 . Thus there is a unique bounded operator on ℓ2 (Γ) extending f 7→ f ∗ x b of norm equal to that of x. We write the image of ξ ∈ ℓ2 (Γ) under this operator as ξx. We use ℓ2 (Γ)x = {ξx : ξ ∈ ℓ2 (Γ)}. By Theorem 43.11 and Proposition 43.10 of [8], we have ( {b x : x ∈ L(Γ)} = 2 ξ ∈ ℓ (Γ) : sup f ∈cc (Γ),kf k≤1 ) kξ ∗ f k2 < ∞ . For ξ ∈ ℓ2 (Γ), f ∈ cc (Γ) we have kξ ∗ f k2 = kξ ∗ f k2 where bar denotes complex conjugation. It follows that there is a weak operator topology continuous, norm-one operator C : L(Γ) → L(Γ) so that c =x Cx b. Lemma 4.10. Let Γ be a countable discrete group and ρ : Γ → O(H) an orthogonal representation on a real separable Hilbert space H with ρC ≪ λΓ . Suppose that ρ has a cyclic vector. Then there is an orthogonal projection p ∈ LR (Γ) so that ΓyH∼ = Γ y ℓ2 (Γ, R)p. Proof. Let ξ be a cyclic vector for H, and define φ : Γ → R by φ(g) = hρ(g)ξ, ξi, then φ(g) ∈ A(Γ)+ . By (3) of the preceding Proposition, we may find xn ∈ L(Γ)+ so that Letting yn = xn +Cxn 2 kφ − φxn k → 0. and using that φ = φ we find that kφ − φyn k → 0. d 1/2 Let ζn = yn . Approximating the square root function by polynomials, we see that ζn ∈ ℓ2 (Γ, R). Note that φyn (g) = τ (λ(g)yn1/2 yn1/2 ) = τ (yn1/2 λ(g)yn1/2 ) = hλ(g)ζn , ζn i. 14 BEN HAYES By the Powers-Størmer inequality (i.e. (4) of the preceding Proposition), kζn − ζm k ≤ kφyn − φym k1/2 so that ζn is a Cauchy sequence. Hence ζn converges to a ζ ∈ ℓ2 (Γ, R). Also, φ(g) = lim φyn (g) = lim hλ(g)ζn , ζn i = lim hλ(g)ζ, ζi. n→∞ Thus 2 n→∞ n→∞ ΓyH∼ = Γ y Span{λ(g)ζ : g ∈ Γ}. Let P be the projection from ℓ (Γ, R) onto Span{λ(g)ζ : g ∈ Γ} and let PC denote its complexification as an operator on ℓ2 (Γ). As PC commutes with λ(g), Theorem 43.11 of [8] shows that there is a unique orthogonal projection p ∈ L(Γ) so that for all ξ ∈ ℓ2 (Γ) we have PC ξ = ξp. Moreover, so p ∈ LR (Γ). Thus  X  g∈Γ pb(g) = hb p, δg i = hP (δe ), δg i ∈ R   f (g)λ(g)ζ : f ∈ cc (Γ, R) = P (ℓ2 (Γ, R)) = ℓ2 (Γ, R)p.  We will need to extend a sofic approximation to an embedding sequence of L(Γ) as in Lemma 5.5 in [17], however we will also want σi (LR (Γ)) ⊆ Mdi (R). Proposition 4.11. Let Γ be a countable discrete group and Σ = (σi : Γ → Sdi ) a sofic approximation. (i): There exists an embedding sequence Σ = (σi : L(Γ) → Mdi (C)) such that X σi (α) = α b(g)σi (g) g∈Γ and σi (LR (Γ)) ⊆ Mdi (R). (ii) If Σ = (σi : L(Γ) → Mdi (C)) is as in (i), and p ∈ LR (Γ) is an orthogonal projection, then there are orthogonal projections pi ∈ Mdi (R) so that kσi (p) − pi k2 → 0. Proof. (i): By Lemma 4.8 we may extend to some embedding sequence Σ = (σi : L(Γ) → Mdi (C)) so that X σi (α) = α b(g)σi (g). g∈Γ By Proposition 4.7, it suffices to show that for all x ∈ LR (Γ) \ R(Γ), there are xi ∈ Mdi (R) so that sup kxi k∞ < ∞ i and kxi − σi (x)k2 → 0. For A ∈ Mn (C) define A ∈ Mn (C) by Aij = Aij , then kAk∞ = kAk∞ It suffices to show that kAk2 = kAk2 . kσi (Cx) − σi (x)k2 → 0 for all x ∈ L(Γ). Indeed, assuming we have the above convergence we may then redefine σi (x) for x ∈ LR (Γ) by 21 (σi (x) + σi (x)). So let x ∈ L(Γ), let ε > 0, and let α ∈ C(Γ) be such that kxδe − λ(α)δe k2 < ε. Since σi (Cλ(α)) = σi (α), kσi (Cx) − σi (x)k2 ≤ kσi (Cx) − σi (Cλ(α))k2 + kσi (x) − σi (λ(α))k2 = kσi (Cx) − σi (Cλ(α))k2 + kσi (x) − σi (λ(α))k2 . SOFIC ENTROPY OF GAUSSIAN ACTIONS 15 Letting i → ∞ and using that σi is a sofic approximation we find that lim sup kσi (Cx) − σi (x)k2 ≤ kCxδe − Cλ(α)δe k + kxδe − λ(α)δe k2 < 2ε. i→∞ Letting ε → 0 proves (i). (ii): Since σi is an embedding sequence we have k\|σi (p)\|2 − σi (p)k2 → 0, k\|σi (p)\|2 − \|σi (p)\|4 k2 → 0. By functional calculus, kχ[1/2,3/2] (\|σi (p)\|2 ) − \|σi (p)\|2 )k22 = kχ{t:\|t−1\|>1/2} (\|σi (p)\|2 )k22 = tr(χ{t:\|t−1\|>1/2} (\|σi (p)\|2 )) ≤ 16 tr(\|\|σi (p)\|2 − \|σi (p)\|4 \|2 ) = 16k\|σi (p)\|2 − \|σi (p)\|4 k22 → 0. Thus setting pi = χ[1/2,3/2] (\|σi (p)\|2 ) completes the proof. Lastly, we will need an analogous definition of singularity, as in the unitary case. If ρj : Γ → O(Hj ), j = 1, 2 are two orthogonal representations, we use HomΓ (ρ1 , ρ2 ) for the space of real, linear, bounded, Γ-equivariant maps from H1 → H2 . We say that ρ1 , ρ2 are mutually singular, written ρ1 ⊥ ρ2 , if whenever ρ′j is a nonzero subrepresentation of ρj for j = 1, 2 we have that ρ′1 , ρ′2 are not isomorphic. Similarly, we say that ρ1 ≪ ρ2 if ρ1 is embeddable into ρ⊕∞ . 2 Lemma 4.12. Let Γ be a countable discrete group, and ρj : Γ → O(Hj ) two orthogonal representations. The following are equivalent. (i): ρ1 ⊥ ρ2 (ii): HomΓ (ρ1 , ρ2 ) = {0}, (iii): HomΓ (ρ2 , ρ1 ) = {0}. Proof. The proof of (ii) equivalent to (iii) and (ii) implies (i) is the same as in Proposition 4.2 of [16]. We can copy the proof of (i) implies (ii) in Proposition 4.2 of [16] provided we prove an analogue of the Polar decomposition. So let T : H1 → H2 be a bounded operator. Let TC : H1,C → H2,C be the complexification of T. Using T t for the real Banach space transpose of T, TC∗ Thus H2 = T t. TC∗ TC (H1 ) ⊆ H1 , approximating the square root function by polynomials, we see that \|TC \|(H1 ) ⊆ H1 . Let be the polar decomposition. As TC = U \|TC \| U = SOT − lim TC (\|TC \| + ε)−1 , ε→0 we find that U (H1 ) ⊆ H2 . The rest is as in Proposition 4.2 of [16]. We now show that in the case of the left regular representation, that the concepts of singularity and absolute continuity in the real case are related to the complex case. 16 BEN HAYES Lemma 4.13. Let Γ be a countable discrete group. Let ρ : Γ → O(H) be an orthogonal representation. Then ρC ≪ λΓ (in the unitary sense) if and only if ρ ≪ λΓ,R . Similarly ρC ⊥ λΓ if and only if ρ ⊥ λΓ,R . Proof. Suppose ρC ≪ λΓ . Applying Zorn’s Lemma to write ρ as a direct sum of cyclic representations and applying Lemma 4.10 we see that ρ ≪ λΓ,R . The converse is even easier. Suppose ρC ⊥ λΓ . If K ⊆ H is a closed, linear, Γ-invariant subspace so that ρ K embeds into λΓ,R , then by complexification we have that ρC K embeds into λΓ . Conversely, suppose T ∈ HomΓ (ρC , λΓ ). Then T H C is a Γ-equivariant, bounded, real-linear map H → ℓ2 (Γ). As ℓ2 (Γ) ∼ = ℓ2 (Γ, R)⊕2 as a real representation we see that T HC (as a complex vector space) we see that T = 0. H = 0, by the previous Lemma. Since H spans The following is proved in the same way as Proposition 4.3 of [16]. Proposition 4.14. Let Γ be a countable discrete group and ρj : Γ → O(Hj ), j = 1, 2 be two orthogonal representations. Then ρ1 = ρ1,s ⊕ ρ1,c where ρ1,s ⊥ ρ2 , and ρ1,c ≪ ρ2 . 4.4. Sofic Entropy of Gaussian Actions. In this section we compute the entropy of Gaussian actions. Let us first start with a very simple corollary of Theorem 4.4 of [16]. Corollary 4.15. Let Γ be a countable discrete sofic group with sofic approximation Σ. Let ρ : Γ → O(H) be an orthogonal representation on a separable Hilbert space H. Suppose that ρ ⊥ λΓ,R . Then if Γ y (Xρ , µρ ) is the corresponding Gaussian action we have hΣ,µρ (Xρ , Γ) ≤ 0. Proof. As in proposition 4.3, for any ξ ∈ H we can find a unique ω(ξ) : Xρ → R so that for all t ∈ R we have exp(2πitω(ξ)(x)) = u(tξ)(x) for almost every x ∈ Xρ . By uniqueness, we have that ω(ξ + η) = ω(ξ) + ω(η) almost everywhere. By [27] we have that ω(ξ) ∈ L2 (Xρ , µρ ) for all ξ ∈ H and that in fact Γ y Span{ω(ξ) : ξ ∈ H} ∼ = Γ y H ⊥ λΓ . We have that L∞ (Xρ , µρ ) is generated as a von Neumann algebra by u(tρ(g)ξ) = exp(2πitω(ρ(g)ξ)) for t ∈ R, ξ ∈ H, g ∈ Γ. From this, it is not hard to argue that the sigma-algebra generated by {ω(ξ)−1 (E) : E ⊆ C is Borel, ξ ∈ H} is all measurable sets (up to measure zero). Thus by Theorem 4.4 of [16] and Lemma 4.13 we know that hΣ,µρ (Xρ , Γ) ≤ 0. We turn to the computation of sofic entropy of Gaussian actions in the case that ρC ≪ λΓ . We will need the following general Lemmas. SOFIC ENTROPY OF GAUSSIAN ACTIONS 17 Lemma 4.16. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let X be a Polish space with a bounded, compatible (not necessarily complete) metric ∆′ , and fix x0 ∈ X. Let Γ y X Γ be the Bernoulli action, and give X Γ the dynamically generating pseudometric ∆(x, y) = ∆′ (x(e), y(e)). (E) For a finite E ⊆ Γ and x ∈ X di define φx : {1, . . . , di } → X Γ by ( x(σi (g)−1 (j)), if g ∈ E (E) φx (j)(g) = x0 , if g ∈ / E. For any finite F ⊆ Γ, there is a finite E0 ⊆ Γ so that if E ⊆ Γ is a finite set containing E0 , then for all δ > 0 and for all large i we have φx(E) ∈ Map(∆, F, δ, σi ). Proof. For h ∈ F, we have ∆2 (hφx(E) , φx(E) ◦ σi (h))2 = If E ⊇ F ∪ F −1 ∪ {e}, then we have ∆2 (hφx(E) , φx(E) ◦ σi (h))2 = 1 di di 1 X ∆′ (φx(E) (j)(h−1 ), φx(E) (σi (h)(j))(e))2 . di j=1 X ∆′ (x(σi (h−1 )−1 (j)), x(σi (h)(j)))2 . j:σi (h−1 )−1 (j)6=σi (h)(j) Let M be the diameter of (X, ∆′ ) then we have ∆2 (hφx(E) ◦ σi (h), hφx(E) )2 ≤ M udi ({1 ≤ j ≤ di : σi (h−1 )−1 (j) 6= σi (h)(j)}) → 0 by soficity. Setting E0 = F ∪ F −1 ∪ {e} completes the proof. Lemma 4.17. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). By Lemma 4.8 extend Σ to an approximation sequence Σ = (σi : L(Γ) → Mdi (C)) so that for α ∈ C(Γ) X σi (α) = α b(g)σi (g). g∈Γ Fix a finite F ⊆ Γ and a x ∈ L(Γ), and let xi ∈ Mdi (C) be such that sup kxi k∞ < ∞, i kσi (x) − xi k2 → 0. Then the following statements hold. (i): There is a sequence Ci ⊆ {1, . . . , di } so that udi (Ci ) → 1 with lim i→∞ sup t∈RF ,s∈RF , j∈Ci kσi (ť)xi σi (š)ej k2ℓ2 (di ) − kλ(ť)xλ(š)δe k22 ktk2ℓ2 (F ) ksk2ℓ2 (F ) = 0. (ii): There is a sequence Ai ⊆ {1, . . . , di }2 so that udi ⊗ udi (Ai ) → 1, and lim i→∞ sup t∈RF ,s∈RF , (j,k)∈Ai kσi (ť)∗ xi ej − σi (š)∗ xi ek k2ℓ2 (di ) − kλ(ť)∗ xδe k22 − kλ(š)xδe k22 (ktkℓ2 (F ) + kskℓ2 (F ) )2 = 0. 18 BEN HAYES Proof. (i): We first handle the case that x ∈ C(Γ) and xi = σi (x). We may choose a Ci so that udi (Ci ) → 1, σi (ťxš)ej = σi (ť)σi (x)σi (š)ej , for j ∈ Ci . By soficity, we may also force that σi (g)ej 6= σi (h)ej for g 6= h in supp(ťxš), and j ∈ Ci , while still having udi (Ci ) → 1. In this case it is easy to see that kσi (ť)σi (x)σi (š)ej k2ℓ2 (di ) = kσi (ťxš)ej k2ℓ2 (di ) = kλ(ť)xλ(š)δe k2ℓ2 (di ) for all j ∈ Ci . Now we handle the general case. Let ε > 0, and choose α ∈ C(Γ) so that k(λ(α) − x)δe k2 < ε. Since σi is an approximation sequence by (2) we have for all g, h ∈ Γ that lim kσi (g)σi (α)σi (h) − σi (g)xi σi (h)k2 = kλ(g)(λ(α) − x)λ(h)δe k2 < ε. i→∞ Thus for all large i we have di X 1 X X kσi (g)σi (α)σi (h)ej − σi (g)xi σi (h)ej k2ℓ2 (di ) = kσi (g)σi (α)σi (h) − σi (g)σi (α)σi (h)k2L2 (Md (C),tr) i di j=1 g,h∈F g,h∈F < ε\|F \|2 . For such i, we may find a Ci ⊆ {1, . . . , di } with so that udi (Ci ) ≥ (1 − √ ε\|F \|2 ) kσi (g)σi (α)σi (h)ej − σi (g)xi σi (h)ej k2ℓ2 (di ) < √ ε for all g, h ∈ F and all large i, and j ∈ Ci . Thus for all j ∈ Ci , and all s, t ∈ RF X kσi (ť)xi σi (š)ej − σi (ť)σi (α)σi (š)ej kℓ2 (di ) ≤ \|t(g)s(h)\|kσi (g)xi σi (h)ej − σi (g)σi (α)σi (h)ej kℓ2 (di ) g,h∈Γ ≤ ε1/4 \|F \|2 ktkℓ2 (F ) kskℓ2 (F ) . Therefore for j ∈ Ci and all large i, and all t, s ∈ RF \|kσi (ť)xi σi (š)ej kℓ2 (di ) − kσi (ť)σi (α)σi (š)ej k2 \| ≤ ε1/4 \|F \|2 ktkℓ2 (F ) kskℓ2 (F ) . By the first part, we can find a Ci′ ⊆ {1, . . . , di } with and udi (Ci′ ) → 1, kσi (ť)σi (α)σi (š)ej k2ℓ2 (di ) = kλ(ť)λ(α)λ(š)δe k22 , for all j ∈ Ci′ , and t, s ∈ RF . Thus for all large i and all t, s ∈ RF , j ∈ Ci ∩ Ci′ , \|kσi (ť)xi σi (š)ej kℓ2 (di ) − kλ(ť)xλ(š)δe kℓ2 (Γ) \| ≤ ε1/4 \|F \|2 ktkℓ2 (F ) kskℓ2 (F ) + εkλ(ť)k∞ kλ(š)k∞ ≤ ε1/4 \|F \|2 ktkℓ2 (F ) kskℓ2 (F ) + εktkℓ1 (F ) kskℓ1 (F ) Since ≤ ε1/4 \|F \|ktkℓ2 (F ) kskℓ2 (F ) + \|F \|εktkℓ2 (F ) kskℓ2 (F ) √ udi (Ci ∩ Ci′ ) ≥ (1 − 2 ε)\|F \| for all large i, we can use a diagonal argument to complete the proof of (i). (ii): Let Ci be as (i). Note that kσi (ť)∗ xi ej − σi (š)∗ xi ek k2ℓ2 (di ) = kσi (ť)∗ xi ej k2ℓ2 (di ) + kσi (š)∗ xi ek k2ℓ2 (di ) − 2 Re(hx∗i σi (š)σi (ť)∗ xi ej , ek iℓ2 (di ) ). SOFIC ENTROPY OF GAUSSIAN ACTIONS 19 We look for Ai with Ai ⊆ Ci × Ci . By (i) it is enough to find an Ai ⊆ Ci × Ci with lim sup i→∞ s,t∈RF ,(j,k)∈A i \|hx∗i σi (š)σi (ť)∗ xi ej , ek iℓ2 (di ) \| = 0, (ktkℓ2 (F ) + kskℓ2 (F ) )2 udi ⊗ udi (Ai ) → 1. We prove that 1 2 i→∞ d2 2 (ktk ℓ (F ) + kskℓ2 (F ) ) i (3) lim X 1≤j,k≤di \|hx∗i σi (š)σi (ť)∗ xi ej , ek iℓ2 (di ) \| = 0, which will clearly prove the existence of such an Ai . Again, as in (i) we first do this for xi = σi (α) for α ∈ C(Γ). Let Di be the set of all j so that σi (α)∗ σi (š)σi (ť)∗ σi (α)ej = σi (α∗ st∗ α)ej . We use o(1) for any expression which tends to 0 as i → ∞. Then lim udi (Di ) = 1, i→∞ and 1 X d2i 1≤j,k≤di \|hσi (α)∗ σi (š)σi (ť)∗ σi (α)ej , ek iℓ2 (di ) \|2 ≤ o(1) + 1 d2i X X (α∗ šť∗ α)(g)(α∗ šť∗ α)(h)hσi (g)ej , ek iℓ2 (di ) hek , σi (h)ej iℓ2 (di ) 1≤j,k≤di ,j∈Di g,h∈Γ = o(1) + 1 XX ∗ ∗ \|(α šť α)(g)\|2 d2i j∈Di g∈Γ 1 = o(1) + udi (Di )kα∗ šť∗ αk22 di 1 ≤ o(1) + udi (Di )kαk∞ ksk2ℓ1 (F ) ktk2ℓ1 (F ) kαk∞ di 1 ≤ o(1) + \|F \|2 ktk2ℓ2 (F ) ksk2ℓ2 (F ) kαk2∞ (1 + o(1)) di → 0. This proves (3) by the Cauchy-Schwartz inequality. The general case follows by approximation as in (i). For notation, if E ⊆ Γ is finite, and f : RE → C is measurable, we let f ⊗ 1RΓ\E : RΓ → C be defined by f ⊗ 1RΓ\E (x) = f (x E ). We say that f ∈ S(RΓ ) if there is a finite E ⊆ Γ and a Schwartz function f0 : RE → C so that f = f0 ⊗ 1RΓ\E . By standard Fourier analysis, there is a θ ∈ S(RE ) so that Z exp(2πit · x) θ(t) dt. f (x) = RE We prove that if we choose x ∈ Rdi with respect to the Gaussian measure on pi Rdi then with high prob(E) ability, the microstate φx will approximately preserve the measure µpδe when integrated against Schwartz functions. We need the following notation: if φ ∈ X di , ψ ∈ Y di we define φ ⊗ ψ ∈ (X × Y )di by (φ ⊗ ψ)(j) = (φ(j), ψ(j)). 20 BEN HAYES Lemma 4.18. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let p ∈ LR (Γ) be an orthogonal projection. Fix a sequence of orthogonal projections pi ∈ Mdi (R) such that kpi − σi (p)k2 → 0. Define the Gaussian measure νi on pi ℓ2R (di ) by d νi (x) = e −πkxk2ℓ2 (d i) dx. (E) Here dx is the Lebesgue measure on pi ℓ2R (di ). Let φx be defined as in Lemma 4.16 for X = R, x0 = 0. Let µpδe be defined as in Proposition 4.3. Let F ⊆ E be finite subsets of Γ and δ > 0. Then for any compact Hausdorff space Y, sequence ψi ∈ Y di , and f ∈ S(RF ) , g ∈ C(Y ), Z Z Z x ∈ p i Rd i : νi f ⊗ 1RΓ\F ⊗ g d(φx(E) ⊗ ψi )∗ (udi ) − f ⊗ 1RΓ\F dµpδe g d(ψi )∗ (udi ) > δ → 0. RΓ Proof. Define G : pi Rdi → R by G(x) = Z f ⊗ 1RΓ\E ⊗ g d(φx(E) ⊗ ψi )∗ (udi ). We will show that (4) Z Z G dνi − Z \|G\|2 dνi − pi Rdi X f ⊗ 1RΓ\E dµpδe Z Y g d(ψi )∗ (udi ) → 0 and (5) pi Rdi As G− Z Z X 2 f ⊗ 1RΓ\E dµpδe 2 = G dνi Z pi Rdi L2 (dνi ) Z Y 2 2 g d(ψi )∗ (udi ) \|G\| dνi − Z → 0. 2 pi Rdi G dνi , the Lemma will then follow from Chebyshev’s inequality. Write Z exp(2πit · x)θ(t) dt f (x) = RF F with θ ∈ S(R ). Note that by Proposition 4.3 and by the fact that θ ∈ L1 (RF ), Z Z Z Z f ⊗ 1RΓ\E dµpδe = θ(t) exp(−πkλ(ť)pk22 ) dt. (6) exp(2πit · x)θ(t) dµpδe dt = RΓ RF RΓ RF We have Z pi Rdi G(x) dνi (x) = = Z di Z 1 X θ(t) exp(2πit · φx(E) (j))g(ψi (j)) dt dνi (x) di j=1 pi Rdi RF Z di Z 1 X θ(t) exp(2πit · φx(E) (j))g(ψi (j)) dνi (x) dt di j=1 RF pi Rdi the interchanges of integrals being valid as g is bounded and θ ∈ L1 (RF ). If E ⊇ F, then t · φx(E) (j) = hσi (ť)x, ej iℓ2 (di ) . SOFIC ENTROPY OF GAUSSIAN ACTIONS 21 Thus Z pi Rdi G(x) dνi (x) = Z di Z 1 X θ(t) exp(2πihσi (ť)x, ej iℓ2 (di ,R) )g(ψi (j)) dνi (x) dt di j=1 RF pi Rdi Z di Z 1 X = θ(t) exp(2πihx, pi σi (ť)∗ ej iℓ2 (di ,R) )g(ψi (j)) dνi (x) dt di j=1 RF pi Rdi di Z 1 X θ(t) exp(−πkpi σi (ť)∗ ej k2ℓ2 (di ) )g(ψ(j)) dt. = di j=1 RF Here we are using that Z pi Rdi exp(2πihx, pi σi (t)∗ ej iℓ2 (di ,R) ) dνi (x) = exp(−πkpi σi (ť)∗ ej k2ℓ2 (di ) ), (this is an obvious generalization of Proposition 8.24 in [13]). By the preceding Lemma, there is a Ci ⊆ {1, . . . , di } with \|Ci \| →1 di (7) and kpi σi (ť)∗ ej k2ℓ2 (di ) − kλ(ť)pk22 lim sup sup i→∞ j∈Ci ,t∈RF ktk22 As we have (8) Z sup j∈Ci RF = 0. kλ(ť)pδe k22 = τ (pλ(ť)∗ λ(ť)p) = τ (λ(ť)pλ(ť)∗ ) = kpλ(ť)∗ δe k22 , θ(t) exp(−πkpi σi (ť)∗ ej k2ℓ2 (di ) ) dt − Z RF θ(t) exp(−πkλ(ť)pδe k2 ) dt ≤ o(1) Z RF \|θ(t)\|ktk2ℓ2 (F ) dt, where we use o(1) for any expression that goes to zero as i → ∞. Additionally, Z (9) sup θ(t) exp(−πkpi σi (ť)∗ ej k2ℓ2 (di ) ) ≤ kθk1 . j∈{1,...,di }\Ci Since RF Z RF θ(t)ktk2ℓ2 (F ) dt < ∞ equations (8), (9), (7) and the fact that g is bounded imply that Z Z Z θ(t) exp(−πkλ(ť)pk22 ) dt g dψ∗ (udi ) → 0. G(x) dνi (x) − pi Rdi RF By (6) we have proved (4). We now turn to the proof of (5). By the same computations as above, Z Z Z Z 1 X \|G\|2 dνi = 2 θ(t)θ(s)e2πi(hσi (ť)x,ej iℓ2 (di ) −hσi (š)x,ek iℓ2 (di ) ) g(ψi (j))g(ψi (k)) dt ds dνi (x) d d d F F i i R pi R R i 1≤j,k≤di pi R Z Z X 1 −π(kpi σi (ť)∗ ej −pi σi (š)∗ ek k2ℓ2 (d ) ) i g(ψ (j))g(ψ (k)) dt ds. θ(t)θ(s)e = 2 i i di RF RF 1≤j,k≤di Again the interchanges of integrals are valid as θ ∈ L1 (RF ). By the preceding Lemma, there are Ai ⊆ {1, . . . , di }2 so that (10) udi ⊗ udi (Ai ) → 1 22 BEN HAYES and lim kpi σi (ť)∗ ej − pi σi (š)∗ ej k22 − kλ(ť)pδe k22 − kλ(š)pδe k22 dt = 0. (ktkℓ2 (F ) + kskℓ2 (F ) )2 sup i→∞ (j,k)∈A , i t,s∈RF Thus (11) sup (j,k)∈Ai as Z RF Z θ(t)θ(s)e −π(kpi σi (ť)∗ ej −pi σi (š)∗ ek k2ℓ2 (d ) ) i RF Z Z Z Z RF Additionally, (12) sup (j,k)∈{1,...,di }2 \(Ai ∩Ci ×Ci ) RF RF Z dt ds − θ(t)e (−πkpi σi (ť)∗ ej k2ℓ2 (d ) ) i 2 dt RF → 0, \|θ(t)\|\|θ(s)\|(ktkℓ2 (F ) + kskℓ2 (F ) )2 dt ds < ∞. θ(t)θ(s)e −π kpi σi (t)∗ ej k2ℓ2 (d ) −2hσi (s)pi σi (t)∗ ej ,ek iℓ2 (d ) +kpi σi (t)ek k2ℓ2 (d ) i i i RF ≤ kθk21 . Equations (8), (12), (10), (7) and the fact that g is bounded imply that Z Z 2 1 X g(ψi (j))g(ψi (k)) → 0. \|G\|2 dνi − θ(t) exp(−πkλ(ť)pδe k22 ) dt · 2 di pi Rdi RF 1≤j,k≤di Since 1 d2i we have proved (5). X g(ψi (j))g(ψi (k) = 1≤j,k≤di Z 2 g dψ∗ (udi ) A A Lemma 4.19. Let A be an infinite set. Then the space S(R ) generates Cb (R ) in the sense of Definition 3.1. Proof. Let ε > 0, f ∈ Cb (RA ) and K ⊆ RA compact. We say that a function ψ : K → C depends upon finitely many coordinates if there is a finite E ⊆ A so that if x, y ∈ K and x(a) = y(a) for all a ∈ E then f (x) = f (y). By the Stone-Weierstrass Theorem there is a function ψ : K → C depending upon finitely many coordinates so that kψkC(K) ≤ kf K kC(K) ≤ kf kCb(X) and kψ − f K kC(K) < ε. Let E ⊆ A be such that if x, y ∈ K and then ψ(x) = ψ(y). Let x(a) = y(a) for all a ∈ E KE = {x ∈ RE : there is a y ∈ K with x(a) = y(a) for all a ∈ E}. Then KE is a compact, being the continuous image of K under the projection map RA → RE . There is a well-defined function ψe : KE → C such that e ψ(x) = ψ(x′ ) whenever x′ ∈ RΓ has x′ (g) = x(g) for all g ∈ E. It is well known that there is a φe ∈ Cc∞ (RE ) so that kφe K − ψe E C(KE ) <ε SOFIC ENTROPY OF GAUSSIAN ACTIONS 23 e C (RE ) ≤ kf kC (RA ) . kφk b b Let φ = φe ⊗ 1RA\E . Since φe is a Schwartz function, we have that φ ∈ S(RA ). Finally, kφ This completes the proof. K −f k K C(K) ≤ kφ K − ψkC(K) + kψ − f k K C(K) < 2ε. Theorem 4.20. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let ρ : Γ → O(H) be an orthogonal representation on a real, separable Hilbert space H with ρ ≪ λΓ,R . Let Γ y (Xρ , µρ ) be the corresponding Gaussian action. Let (Y, ν) be a standard probability space and Γ y (Y, ν) a measure-preserving action with hΣ,ν (Y, Γ) ≥ 0. Then hΣ,µρ ⊗ν (Xρ × Y, µρ ⊗ ν) = ∞. Proof. We shall first reduce to the case that ρ is cyclic, i.e there is a vector ξ ∈ H so that H = Span{ρ(g)ξ : g ∈ Γ}, so suppose we can prove the Theorem in the cyclic case. Let ρ = ρ1 ⊕ ρ2 , where ρ1 is cyclic and ρ2 ≪ λΓ,R . Since we are assuming the theorem in the cyclic case and Γ y (Xρ × Y, µρ ⊗ ν) ∼ = Γ y (Xρ1 × Xρ2 × Y, µρ1 ⊗ µρ2 ⊗ ν) it suffices to show that hΣ,µρ2 ⊗ν (Xρ2 × Y, Γ) ≥ 0, (13) as ρ1 is cyclic. Since H is separable, we may write ρ2 = ∞ M ρ2,n n=1 where ρ2,n is cyclic. For n ∈ N, let ρ2,≤N = N M ρ2,n . n=1 By the cyclic case and induction, we have hΣ,µρ2,≤N ⊗ν (Xρ2,≤N × Y, Γ) = ∞ for any N ∈ N. Using that Γ y (Xρ2,≤N × Y, µρ2,≤N ⊗ ν), Γ y (Xρ2 × Y, µρ2 ⊗ ν) ∼ = lim ←− it is not hard to show (13). We use the preceding Lemma and Theorem 3.3. By Proposition 4.3 we may regard Γ y (Xρ , µρ ) as RΓ with the measure µpδe defined by Z exp(2πit · x) dµpδe (x) = exp(−πkλ(t)pδe k22 ) RΓ for t ∈ cc (Γ, R) and some orthogonal projection p ∈ LR (Γ). Extend Σ to an approximation sequence of L(Γ) still denoted σi : L(Γ) → Mdi (C). By Proposition 4.11, there exists a sequence pi of orthogonal projections in Mdi (R) so that kpi − σi (p)k2 → 0. di Let νi be the Gaussian measure on pi R defined by dνi (x) = e −πkxk2ℓ2 (d i) dx. Choosing a compact model, we may assume that Y is a compact metrizable space and that Γ y Y by homeomorphisms. Let ∆Y be a compatible metric on Y. Let ∆RΓ be the dynamically pseudometric on RΓ defined by ∆RΓ (x, y) = min(\|x(e) − y(e)\|, 1). 24 BEN HAYES We shall use the generating set L = {f ⊗ g : f ∈ S(RΓ ), g ∈ C(Y )}. We use the dynamically generating pseudometric on RΓ × Y defined by ∆((x1 , y1 ), (x2 , y2 )) = ∆RΓ (x1 , x2 ) + ∆Y (y1 , y2 ). Let ε, δ, η > 0 and finite F ⊆ Γ,L ⊆ L be given. Let e = g1 ,g2 , . . . be an enumeration of the elements of Γ. Inductively find positive real numbers M1 , M2 , . . . so that l X Γ 2−j η. µpδe {x ∈ R : \|x(gj )\| ≤ Mj , j = 1, . . . , l} ≥ 1 − j=1 Define M : Γ → (0, ∞) by M (gj ) = Mj . Set KM = {x ∈ RΓ : \|x(g)\| ≤ M (g) for all g ∈ Γ} and note that KM is compact. Let U be an open neighborhood of KM in RΓ . Let L1 ⊆ S(RΓ ), L2 ⊆ C(Y ) be finite sets so that L ⊆ {f ⊗ g : f ∈ L1 , g ∈ L2 }. Since hΣ,ν (Y, Γ) ≥ 0, for infinitely many i, there is a ψi ∈ Mapν (∆Y , F, δ, L2 , σi ). For these i, by Lemma 4.18, there is a finite set E of Γ containing the identity and an Bi ⊆ pi Rdi so that νi (Bi ) → 1 φx(E) ⊗ ψi ∈ Mapµ⊗ν (∆, F, δ, L, σi ) for all x ∈ Bi . We will also assume that U ⊇ πE (KM ) × RΓ\E . This may be done by enlarging E, as KM is compact. Choose f ∈ S(RE ) so that Z f d(πE )∗ µρ ≥ −η + (πE )∗ µ(πE (KM )) RE and 0 ≤ f ≤ χπE (KM ) . We may assume that φx(E) ⊗ ψi ∈ Mapµ⊗ν (∆, F, δ, L ∪ {f ⊗ 1RΓ\E ⊗1 }, σi ), for all x ∈ Bi . Note that in this case (φx(E) ⊗ ψi )∗ (udi )(U × Y ) ≥ (φx(E) ⊗ ψi )∗ (udi )(πE (KM ) × RΓ\E × Y ) Z ≥ −δ + f ⊗ 1RΓ\E dµp RΓ ≥ −δ − η + (πE )∗ µ(πE (KM )) ≥ −δ − 2η + 1. So if δ < η we have Thus U×Y,3η φx(E) ⊗ ψi ∈ Mapµ⊗ν (∆, F, δ, L ∪ {f ⊗ 1RΓ\E ⊗ 1}, σi ). 1 log Sε (Bi , ∆2 ). i→∞ di Suppose that S ⊆ Bi is ε-dense. Then for every y ∈ Bi , there exists x ∈ S and a C ⊆ {1, . . . , di } so that √ \|x(j) − y(j)\| < ε, for j ∈ C, √ \|C\| ≥ (1 − ε)di . U×Y,η hΣ,µ (∆, ε, F, δ, L ∪ {f ⊗ 1RΓ\E ⊗ 1}) ≥ lim sup p ⊗ν SOFIC ENTROPY OF GAUSSIAN ACTIONS Using m for Lebesgue measure, νi (B) ≤ ≤ X x∈S, A⊆{1,...,d √ i }, \|A\|≥(1− ε)di X νi ((xA + 25 √ c ε Ball(ℓ2R (A, uA ))) × RA ) √ m( ε Ball(ℓ2R (A, uA ))) x∈S, A⊆{1,...,d √ i }, \|A\|≥(1− ε)di X = ε\|A\|/2 m(Ball(ℓ2R (A, uA ))) x∈S, A⊆{1,...,d √ i }, \|A\|≥(1− ε)di By [13] Corollary 2.55 and Stirling’s Formula, there is a R > 0 so that \|A\|/2 di m(Ball(ℓ2R (A, uA ))) ≤ R\|A\|/2 . \|A\| Set H(t) = ( Then for some D > 0 we have νi (B) ≤ 0, −t log(t) − (1 − t) log(1 − t), X di ε 2 R\|A\|/2 x∈S, A⊆{1,...,d √ i }, \|A\|≥(1− ε)di di \|A\| \|A\|/2 √ √ ≤ \|S\|D εdi Rdi /2 exp(H( ε)di ) the last line following by Stirling’s Formula. Thus 1 1 lim sup log \|S\| ≥ log 2 i→∞ di So U×Y,3η hΣ,µ (∆, ε, F, δ, L) p ⊗ν ≥ U×Y,3η hΣ,µ (∆, ε, F, δ, L p ⊗ν if t = 0, 1 for 0 < t < 1. ε √ (1 − ε) d2i , √ √ 1 1− ε − log R − H( ε). ε 2 ∪ {f ⊗ R Γ\E 1 ⊗ 1}) ≥ log 2 √ √ 1 1− ε − log R − H( ε), ε 2 taking the infimum over all U ⊇ KM , then the supremum over all KM , then the infimum over all F, δ, η and L ⊆ L we find that √ √ 1 1− ε 1 − log R − H( ε). hΣ,µpδe ⊗ν (∆, ε, L) ≥ log 2 ε 2 If we now let ε → 0 we see that hΣ,µpδe ⊗ν (X × Y, Γ) = ∞. We now prove a general formula for the entropy of Gaussian actions. Corollary 4.21. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). Let ρ : Γ → O(H) be an orthogonal representation. By Proposition 4.14 write ρ = ρ1 ⊕ ρ2 where ρ1 ≪ λΓ,R and ρ2 ⊥ λΓ,R . Let Γ y (Xρ , µρ ), Γ y (Xρj , µρj ), j = 1, 2 be the corresponding Gaussian actions. Then: (i): hΣ,µρ2 (Xρ2 , µρ2 ) ∈ {0, −∞}, (ii):   −∞, if hΣ,µρ2 (Xρ2 , µρ1 ) = −∞, hΣ,µρ (Xρ1 , µρ1 ) = 0, if ρ1 = 0 and hΣ,µρ2 (Xρ2 , µρ2 ) = 0,   ∞, if ρ1 6= 0 and hΣ,µρ2 (Xρ2 , µρ2 ) = 0. 26 BEN HAYES Proof. Statement (i) is just a direct corollary of Corollary 4.15 and the fact that sofic entropy is always nonnegative or −∞. The first case of statement (i) follows from the general fact that if Γ y (X, µ), Γ y (Y, ν) are two measure-preserving actions on standard probability spaces, then hΣ,µ (X, Γ) = −∞ implies that hΣ,µ⊗ν (X × Y, Γ) = −∞. The second case of statement (ii) is also just a special case of statement (i). The last case of statement (ii) follows from Theorem 4.20. We give some examples to show that hΣ,Xρ (Xρ , µρ ) can be −∞. Let Γ be a countable discrete sofic group with sofic approximation Σ = (σi : Γ → Sdi ). We say that Σ is ergodic if whenever Ai ⊆ {1, . . . , di } are such that lim udi (Ai ∆σi (g)Ai ) = 0 for all g ∈ Γ i→∞ then lim udi (Ai )(1 − udi (Ai )) = 0. i→∞ The following is a folklore result and we include a proof for completeness. Proposition 4.22. Let Γ be a countable discrete sofic group with an ergodic sofic approximation Σ. Then, if Γ y (X, µ) is a nonergodic probability measure-preserving action on a standard probability space, we have hΣ,µ (X, Γ) = −∞. Proof. Let Σ = (σi : Γ → Sdi ). Let E ⊆ X be a Γ-invariant set with 0 < µ(E) < 1. Let α : X → {0, 1} be the finite observable given by α(x) = χE (x). Let ε > 0, by a diagonalization argument it is easy to see that there is a δ > 0 and a finite F ⊆ Γ containing e so that if Bi ⊆ {1, . . . , di } has lim sup max udi (Bi ∆σi (g)Bi ) < δ, (14) i→∞ g∈F then lim sup udi (Bi )(1 − udi (Bi )) < ε. (15) i→∞ e = {a ∈ {0, 1} : a(e) = 1}. If δ is sufficiently small, and φ ∈ AP(α, F, δ ′ , σi ) then since µ(gE∆E) = 0, Set E we have e e <δ E)) max udi (σi (g)φ−1 (E)∆φ( F g∈F and we always have If δ ′ , ε are sufficiently small, then e ≥ µ(E) − δ ′ . µ(E) + δ ′ ≥ udi (φ−1 (E)) e > ε. e − udi (φ−1 (E))) udi (φ−1 (E))(1 Thus we see from (14), (15) that AP(α, F, δ ′ , σi ) = ∅. Thus hΣ,µ (X, Γ) = −∞, using Kerr’s definition of entropy via partitions. Combining with Theorem 2.8 of [27] we have the following. Corollary 4.23. Let Γ be a countable discrete sofic group with an ergodic sofic approximation Σ. Let ρ : Γ → O(H) be an orthogonal representation which is not weakly mixing (e.g. ρ could be compact). Then if Γ y (X, µ) is the associated Gaussian action, hΣ,µ (X, Γ) = −∞. SOFIC ENTROPY OF GAUSSIAN ACTIONS 27 We mention an example of an ergodic sofic approximation when Γ = F2 is the free group on two generators a, b. Here we can choose a sofic approximation randomly. Namely let φ = (φ1 , φ2 ) ∈ Sn2 be chosen uniformly at random, and let σφ : F2 → Sn be the unique homomorphism so that σφ (a) = φ1 , σφ (b) = φ2 . It is known that with high probability σφ is a sofic approximation (see [23], as well as [18] Lemma 3.1). It is also known from the theory of expanders that with high probability σφ is a ergodic sofic approximation (see [14], and the remarks in Section 5 of [26]). If we take an orthogonal representation ρ : F2 → O(H) which is not weakly mixing (e.g. take H to be finite-dimensional) then we have hΣ,µρ (Xρ , Γ) = −∞. References [1] L. Bowen. Measure conjugacy invariants for actions of countable sofic groups. J. Amer. Math. Soc, 23:217–245, 2010. [2] L. Bowen. Entropy for expansive algebraic actions of residually finite groups. Ergodic Theory Dynam. Systems, 31(3):703– 718, 2011. [3] L. Bowen. Weak isomorphisms between Bernoulli shifts. Israel J. Of Math, 183, 2011. [4] L. Bowen. Every countably infinite group is almost Ornstein. In Dynamical systems and group actions, volume 567 of Contemp. Math, pages 67–78. Amer. Math. Soc, Providence, RI, 2012. [5] L.Bowen and H.Li. Harmonic models and spanning forests of residually finite groups. J. Funct. Anal., 263(7):1769–1808, 2012. [6] N. Brown and N. Ozawa. C ∗ -Algebras and Finite-Dimensional Approximations. Cambridge University Press, 1994. [7] L. Ciobanu, D. Holt, and S. Rees. Sofic groups: graph products and graphs of groups. Pacific J. of Math, 271(1):53–64, November 2014. [8] J. Conway. A Course in Operator Theory. Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 1991. [9] A. Dooley and V. Golodets. The spectrum of completely positive entropy actions of countable amenable groups. J. Funct. Anal., 196(1):1–18, 2002. [10] K. Dykema, D. Kerr, and M. Pichot. Orbit equivalence and sofic approximation. arXiv:1102.2556. [11] K. Dykema, D. Kerr, and M. Pichot. Sofic dimension for discrete measurable groupoids. Trans. Amer. Math. Soc, 366(2):707–748, 2013. [12] G. Elek and E. Szabo. On sofic groups. J. Group Theory, 9(2):161–171, 2006. [13] G. Folland. Real Analysis: Modern Techinques and Their Applications. John Wiley & Sons, Hoboken, NJ, second edition, 1999. [14] J. Friedman. A proof of Alon’s second eigenvalue conjecture and related problems. Memoirs of AMS, 195:1–100, 2008. [15] B. Hayes. Fuglede-Kadison determinants and sofic entropy. arXiv:1402.1135. [16] B. Hayes. Polish models and sofic entropy. J. Inst. Math. Jussieu. To appear. [17] B. Hayes. An lp -version of von Neumann dimension for Banach space representations of sofic groups. J. Funct. Anal., 266(2):989–1040, 2014. [18] D. Kerr. Bernoulli actions of sofic groups have completely positive entropy. Israel J. Math. To appear. [19] D. Kerr and H. Li. Topological entropy and the variational principle for actions of sofic groups. Invent. Math, 186:501–558, 2011. [20] J. Kieffer. A generalized Shannon-McMillan theorem for the action of an amenable group on a probability space. Ann. Prob, 3(6):1031–1037, 1975. [21] M. Lemańczyk. Entropy of Gaussian actions for countable abelian groups. Fund. Math., 157:277–286, 1998. [22] T. Meyerovitch. Positive sofic entropy implies finite stabilizer. arXiv:1504.08137. [23] A. Nica. Asymptotically free families of random unitaries in symmetric groups. Pacific J. Math, 157:295–310, 1993. [24] D. Ornstein and B. Weiss. Entropy and isomorphism theorems for actions of amenable groups. J. Anal. Math., 48:1–141, 1987. [25] L. Paunescu. On sofic actions and equivalence relations. J. Funct. Anal., 261(9):2461–2485, November 2011. [26] L. Paunescu. Convex structures revisited. Ergodic Theory Dynam. Systems. To appear. [27] J. Peterson and T. Sinclair. On cocycle superrigidity for Gaussian actions. Ergodic Theory Dynam. Systems, 32(1):249–272, 2012. [28] S. Popa. Independence properties in sublagebras of ultraproduct II1 factors. J. Funct. Anal., 266(9):5818–5846, 2014. [29] B. Seward. Every action of a non-amenable group is the factor of a small action. J. Mod. Dyn., 8(2):251–270, 2014. [30] M. Takesaki. The Theory of Operator Algebras I. Springer-Verlag, Berline-Heidelberg-New York, 2002. [31] M. Takesaki. Theory of Operator Algebras II, volume 125 of Encyclopaedia of Mathematical Sciences. Springer, New York, NY, 2003. Stevenson Center, Nashville, TN 37240 E-mail address: benjamin.r.hayes@vanderbilt.edu	4math.GR
arXiv:1708.05761v1 [math.AC] 18 Aug 2017 EXTENDED PLUS CLOSURE IN COMPLETE LOCAL RINGS RAYMOND C HEITMANN AND LINQUAN MA Abstract. The (full) extended plus closure was developed as a replacement for tight closure in mixed characteristic rings. Here it is shown by adapting André’s perfectoid algebra techniques that, for complete local rings, this closure has the colon-capturing property. In fact, more generally, if R is a (possibly ramified) complete regular local ring of mixed characteristic, I and J are ideals of R, and the local domain S is a finite R-module, then (IS : J) ⊆ (I : J)S epf . A consequence is that all ideals in regular local rings are closed, a fact which implies the validity of the direct summand conjecture and the Briançon-Skoda theorem in mixed characteristic. 1. Introduction In [7], the first author introduced several closure operations for mixed characteristic rings, most notably the full extended plus closure. These closures were proposed as possibilities to play the same role in mixed characteristic as the tight closure does in characteristic p. It was hoped that they might allow us to prove some of the homological conjectures that remained unresolved in mixed characteristic and indeed, the full extended plus closure did figure in the proof of the direct summand conjecture in dimension 3 [8]. Of course, the direct summand conjecture has been resolved [2] without the use of this closure operation. However, it still remains of interest to note to what extent the full extended plus closure fills the void left open by the absence of tight closure in mixed characteristic. In this paper, we note a number of properties that were shown in [7] and add several others. In particular, we deal with two fundamental questions that were left open in [7]. We shall see that the full extended plus closure has the colon-capturing property in complete local rings and that ideals in all regular local rings are closed. The latter fact is sufficient to imply the direct summand conjecture. This is not truly an alternate proof as our work makes use of André’s notation and techniques. For the most part, this article is about the full extended plus closure in complete local rings. Since we do not know whether or not completing, closing, and contracting occasionally gives a larger closure, only Theorem 3.16 gives new information about rings which are not complete. The primary result in this article is that, in a sense, the full extended plus closure captures obstructions to flatness. Our principal theorems are: Theorem 1.1 (Corollary 3.11). Let S be a complete local domain of mixed characteristic and let y1 , . . . , yn be parameters in S. Then ((y1 , . . . , yn−1 )S : yn ) ⊆ (y1 , . . . , yn−1 )S epf . Date: August 22, 2017. 1 2 RAYMOND C HEITMANN AND LINQUAN MA Theorem 1.2 (Theorem 3.12). Let R be a complete regular local ring of mixed characteristic and let the integral domain S be a finite extension of R. If I is an ideal of R and J ⊆ R, then (IS : J) ⊆ (I : J)S epf . An equivalent formulation of Theorem 1.2, divorced from the language of extended plus closure, is Theorem 1.3 (Theorem 3.14). Let R be a complete regular local ring of mixed characteristic p and let the integral domain S be finite over R. Let M be an RN module. Then there exist c ∈ R such that the natural map cǫ TorR −→ i (S, M )/p R + N Tori (S , M )/p is zero for all i ≥ 1 and for every N, ǫ > 0. Consequently, for + N ǫ every α ∈ TorR i (R , M )/p , there exist c ∈ R such that c α = 0 for every ǫ. This paper is organized as follows. In Section 2, we will give a brief introduction to extended plus closure. The results discussed here are generally known although Proposition 2.10 is actually new. We will also give a short introduction to our terminology and conventions. Section 3 will be devoted to our new results. 2. basics Notation 2.1. If I is an ideal of R and J ⊆ R, (I : J) = {r ∈ R \| rJ ⊆ I}. The ring involved will always be clear from the ideal and there will be no need for the more cumbersome notation (I :R J). In [7], the first author introduced four closure operations for mixed characteristic rings. Subsequently, only one of these, the full extended plus closure, has received any attention. That will also be the case in this article. In fact, while we shall preserve the notation I epf , the f for full, we shall mostly refer to the closure as the extended plus closure. It should be noted that the original definition and early results allow R to be either mixed characteristic or characteristic p. However, the characteristic p case is less interesting as the full extended plus closure trivially contains the tight closure. Definition 2.2. Let R be an integral domain with p ∈ J(R) and I an ideal of R. Then an element x ∈ R is said to be in the (full) extended plus closure of I (designated x ∈ I epf ) provided there exists c ∈ R such that for every ǫ > 0 and every positive integer N , cǫ x ∈ (I, pN )R+ . Remark 2.3. The extended plus closure is actually defined for all Noetherian rings with p ∈ J(R). Just as with tight closure, one may compute the extended plus closure by computing it modulo each minimal prime ideal and then taking the intersection of the liftings back to the original ring. The more general setting does not hold great interest and in this article, we will focus on the case of integral domains. We first shall highlight some of the basic results from [7]. First we see that it behaves as a reasonable closure operation. Proposition 2.4. [7, Proposition 2.1] Let I, J be ideals in a Noetherian ring R. (1) I epf is an ideal. (2) I ⊆ I epf . (3) If I ⊆ J, I epf ⊆ J epf . (4) (I epf )epf = I epf . EXTENDED PLUS CLOSURE IN COMPLETE LOCAL RINGS 3 (5) (I epf J epf )epf = (IJ)epf . (6) (I : J)epf ⊆ (I epf : J). In particular, if I is closed, then (I : J) is closed. Remark 2.5. It is known that if R is local and S = Rh , the henselization of R, then IS epf ∩ R = I epf [7, Theorem 3.4]. However, this result is unfortunately not known b of R, even when R is an excellent domain. For if we replace S by the completion R this reason our main theorems on extended plus closure do not obviously generalize to non-complete rings. Finally, we have the two most interesting theorems from the earlier work - two theorems which match consequential theorems from tight closure theory. First full extended plus closure gives a Briançon-Skoda type theorem. Theorem 2.6. [7, Theorem 4.2] Let R be a ring, let I be an ideal of R generated by n elements, and suppose y is in the integral closure of I n+d for some integer d ≥ 0. Then y ∈ (I d+1 )epf . We note that this Briançon-Skoda type theorem is a simple consequence of [6, Theorem 2.13]. We include that result next as it serves as a lemma we will need in our proof of Lemma 3.3 and Theorem 3.17. Proposition 2.7. p Let R be an integral domain and I = (x1 , . . . , xn )R an ideal of R. Suppose p ∈ (x1 , x2 )R and z ∈ I n+k with k ≥ 0. Then there exists an integral extension S of R with z ∈ I k+1 S. Recall that the classical Briançon-Skoda theorem says that in a regular ring, if an ideal I is generated by n elements, then the integral closure of I n+d is contained in I d+1 for every integer d [14]. We can recover this result by proving that in a regular local ring, every ideal is closed under the extended plus closure; see Theorem 3.16. We give two proofs of this theorem: it can be obtained from our earlier work on the vanishing conjecture for maps of Tor [9], and it also follows from our main result in this article and the following result in [7]: Theorem 2.8. [7, Theorem 4.4] Let R be a regular local ring with p in the maximal ideal. If the full extended plus closure has the colon-capturing property for finite extensions of complete regular local rings, then I epf = I for every ideal I of R. Remark 2.9. The actual statement of this result in [7] presumes that the coloncapturing property holds for finite extensions of all regular local rings. However, the statement of [7, Lemma 4.3] makes it clear that one may restrict to complete regular local rings. Then, by faithful flatness, the fact that I epf = I holds in complete regular local rings implies it holds in all regular local rings. One may observe that these results are all well known for tight closure. In fact, an alternate definition of tight closure closely resembles that of extended plus closure. Instead of requiring cǫ x ∈ IR+ , the tight closure requires cǫ x ∈ IRF where RF is the integral closure of R in the largest purely inseparable extension of its quotient field. Superficially, this means the tight closure could be smaller. However, Hochster and Huneke [12] introduced the dagger closure (which must contain the extended plus closure) and showed [12, Theorem 3.1] that, for characteristic p complete local rings, the tight closure and dagger closure were equal. Thus Proposition 2.10. If R is a complete local ring of equal characteristic p, the extended plus closure coincides with tight closure for ideals of R. 4 RAYMOND C HEITMANN AND LINQUAN MA 2.1. Perfectoid algebras. We will freely use the language of perfectoid spaces [16] and almost mathematics [5]. In this paper we will always work in the following situation: for a perfect field k of characteristic p > 0, we let W (k) be the ring of ∞ Witt vectors with coefficients in k. Let K ◦ be the p-adic completion of W (k)[p1/p ] and K = K ◦ [1/p]. Then K is a perfectoid field in the sense of [16] with K ◦ ⊆ K its ring of integers. A perfectoid K-algebra is a Banach K-algebra R such that the set of powerbounded elements R◦ ⊆ R is bounded and the Frobenius is surjective on R◦ /p. A K ◦ -algebra S is called integral perfectoid if it is p-adically complete, p-torsion free, satisfies S = S∗ 1 and the Frobenius induces an isomorphism S/p1/p −→ S/p. These two categories are equivalent to each other [16, Theorem 5.2] via the functors R −→ R◦ and S −→ S[1/p]. Almost mathematics in this article will be measured with respect to a flat ideal ∞ (c1/p ). There are two cases that we will use: c = p or c = pg for some nonzero element g. This will be stated explicitly throughout. 3. Main Results For R a Noetherian integral domain of mixed characteristic p, we let R+ denote the absolute integral closure of R and let R++ denote the p-adic completion of R+ . For a nonzero element c ∈ R, we will use Rc to denote the integral closure of R++ in R++ [c−1 ]. Lemma 3.1. Let R be an integral domain of mixed characteristic p and I an ideal of R. For x ∈ R, if there exists c ∈ R such that for every ǫ > 0 and every positive integer N , cǫ x ∈ (I, pN )R++ , then x ∈ I epf . N ++ Proof. We need only show that if cǫ x ∈ (I, cǫ x ∈ (I, pN )R+ . HowPps )R , then N ǫ N ++ ǫ ever, if c x ∈ (I, p )R , we have c x = i=1 yi ui + p u0 where each Psyi ∈ I and each ui is the limit of a Cauchy sequence {uij }. It follows that {cǫ x − i=1 yi uij − pN u0j } is a Cauchy sequence which converges to zero. Thus, P P for sufficiently large j, we have cǫ x − si=1 yi uij − pN u0j ∈ pN R+ . Then cǫ x = si=1 yi uij + pN (u0j + v) ∈ (I, pN )R+ . Lemma 3.2. Let R be a universally catenarian Noetherian integral domain of mixed characteristic p and suppose (p, g)R is a height two ideal. Then we have ∩n (g, pn )R++ = gR++ . Proof. Pick z ∈ ∩n (g, pn )R++ ; we can write z = a1 g + b1 p with a1 ∈ R+ and b1 ∈ R++ . Now b1 p ∈ ∩n (g, pn )R++ and so we can write b1 p = a2 g + b2 p2 with a2 ∈ R+ and b2 ∈ R++ . But then b2 p2 ∈ ∩n (g, pn )R++ , so we can repeat the above process to write b2 p2 = a3 g + b3 p3 and keep going: z = a1 g + b 1 p = a1 g + a2 g + b 2 p 2 = a1 g + a2 g + a3 g + b 3 p 3 = · · · . Note that we have ai g ∈ pi−1 R++ for all i, thus ai g = 0 as an element in R+ /pi−1 = R++ /pi−1 . Since R is universally catenary, (p, g) is a height two ideal in any integral k 1S = {x ∈ S[1/p]\|S 1/p · x ∈ S for all k}. Hence S is almost isomorphic to S with respect ∗ ∗ ∞ to (p1/p ), thus in practice we will often ignore this distinction since one can always pass to S∗ without affecting the issue. EXTENDED PLUS CLOSURE IN COMPLETE LOCAL RINGS 5 extension of R. Since R+ is a direct limit of normal integral extensions of R, g is a nonzerodivisor on R+ /pi−1 and so ai ∈ pi−1 R+ . Therefore, in R++ we have z = a1 g + a2 g + a3 g + · · · = g(a1 + a2 + a3 + · · · ) ∈ gR++ . This finishes the proof. Lemma 3.3. Let R be a universally catenarian Noetherian integral domain of mixed characteristic p of dimension at least two and I an ideal of R. For x ∈ R, if there exists a nonzero c ∈ R such that for every ǫ > 0 and every positive integer N , cǫ x ∈ (I, pN )Rc , then x ∈ I epf . √ Proof. We can find an element g ∈ R such that pg ∈ cR and (p, g)R is a height two ideal and likewise so is (p, g)R′ for every integral extension R′ of R (since R is universally catenary). The hypothesis of the lemma clearly holds with pg in place of c (note that Rpg is larger than Rc ) and so we may assume c = pg. Now this lemma will follow immediately from Lemma 3.1 if we can show that the injection ∞ R++ −→ Rc is an almost isomorphism with respect to (c1/p ). Suppose u ∈ Rc . Then u = (pg)−d v for v ∈ R++ and some integer d. As R++ is just the p-adic completion of R+ , we may write v = w1 + pd w2 with w1 ∈ R+ and 1 1 is integral over R++ . This forces w to be integral w2 ∈ R++ . It follows that w pd pd + over the integrally closed R . So without loss of generality, we may assume w1 = 0 and so reduce to the case u = g −d v. Fix t > 0 and let ǫ = 1/pt . To complete the proof, it suffices to show v ∈ d−ǫ n (g , p )R++ for all positive integers n by Lemma 3.2. Now we fix an n > 0. We may write v = v1 + pnN v2 with v1 ∈ R+ , v2 ∈ R++ , and N = dpt = d/ǫ. As v is integral over g d R++ , v1 is integral over (g d , pnN )R++ . This means there exists some integer k such that (v1 , g d , pnN )k+1 R++ = (v1 , g d , pnN )k (g d , pnN )R++ . Since both ideals contain a power of p, it is easy to see that (v1 , g d , pnN )k+1 R+ = (v1 , g d , pnN )k (g d , pnN )R+ and thus v1 is integral over (g d , pnN )R+ . It follows that v1 ∈ (g ǫ , pn )N R+ . As R+ is absolutely integrally closed, Proposition 2.7 gives v1 ∈ (g ǫ , pn )N −1 R+ ⊆ (g d−ǫ , pn )R+ . Therefore v = v1 + pnN v2 ∈ (g d−ǫ , pn )R++ as desired. Now suppose R is a complete and unramified regular local ring of mixed characteristic, and R −→ S is a module-finite local extension (so S is a complete local domain). By Cohen’s structure theorem R ∼ = W (k)[[x1 , . . . , xd−1 ]] where W (k) is a coefficient ring of R, which is a complete and unramified DVR with residue field k. Let k̄ denote an algebraic closure of k, let A0 = R ⊗W (k) W (k̄), and let A be the p-adic completion of A0 . Then A is p-adically complete and formally smooth over W (k̄) mod p. We next point out that A is actually Noetherian (though we don’t need this). Lemma 3.4. With the notation as above, A is a Noetherian ring. Proof. First of all, in general, if T is p-adically complete, then T is Noetherian if and only if T /pT is Noetherian: suppose Q is an infinitely generated prime ideal of T , then since T /pT is Noetherian, Q does not contain p. Now Q + pT is finitely generated, say by f1 + ptP 1 , . . . , fn + pt P Pn . Suppose f ∈PQ ⊆ Q + pT ; we can write fP= a1j (fj + ptj ) = a1j fj P + p a1j tP a1j tj P ∈ Q but p P ∈ / Q, so j . Now p a1j tj ∈ Q ⊆ Q + pT and thus a t = a (f + pt ) = a f + p a2j tj . 1j j 2j j j 2j j P P This gives f = (a1j + pa2j )fj + p2 a2j tj . We can repeat this process to find 6 RAYMOND C HEITMANN AND LINQUAN MA PP aij and since T is p-adically complete, we have f = ( i pi−1 aij )fj . Therefore Q = (f1 , . . . , fn ) is finitely generated. Now to see A is Noetherian, it is enough to show A/pA = k̄ ⊗ k[[x1 , . . . , xd−1 ]] is Noetherian by the previous paragraph. We prove this by induction on d. The case d = 1 is clear. We now prove that every prime ideal Q in k̄ ⊗ k[[x1 , . . . , xd−1 ]] is finitely generated. Since k[[x1 , . . . , xd−1 ]] −→ k̄ ⊗ k[[x1 , . . . , xd−1 ]] is integral, we have P = Q ∩ k[[x1 , . . . , xd−1 ]] with ht P = ht Q. If ht Q = 0 then Q = 0 since k̄ ⊗ k[[x1 , . . . , xd−1 ]] is a domain. If ht Q = ht P > 0, then Q̄ is a minimal prime of k̄ ⊗ (k[[x1 , . . . , xd−1 ]]/P ). But we have k[[y1 , . . . , yd′ −1 ]] −→ k[[x1 , . . . , xd−1 ]]/P for some d′ < d by Cohen’s structure theorem; hence k̄ ⊗ (k[[x1 , . . . , xd−1 ]]/P ), being a finite extension of k̄ ⊗ k[[y1 , . . . , yd′ −1 ]], is Noetherian by our induction hypothesis. Therefore Q̄ is finitely generated and thus Q is finitely generated. Returning to the development of our set-up, we next set B = S ⊗R A. Since R −→ S is module-finite and generically étale, there exists a nonzero g ∈ R with ht(p, g)R = 2 such that R[(pg)−1 ] −→ S[(pg)−1 ] is finite étale. Thus A[(pg)−1 ] −→ ∞ B[(pg)−1 ] is finite étale. Let K ◦ be the p-adic completion of W (k̄)[p1/p ] and ∞ ∞ 1/p 1/p∞ K = K ◦ [1/p]. Let A∞,0 be the p-adic completion of A[p1/p , x1 , . . . , xd−1 ], which is an integral perfectoid K ◦ -algebra. Furthermore, for g ∈ R ⊆ A as above, we let A∞,0 −→ A∞ be André’s construction of integral perfectoid K ◦ -algebras (for example see [3, Theorem 1.4]): it is the integral perfectoid K ◦ -algebra of functions ∞ ∞ on the Zariski closed subset of Spa(A∞,0 hT 1/p i[1/p], A∞,0 hT 1/p i) defined by the ideal T − g. More explicitly, A∞ can be described as the p-adic completion of the ∞ ∞ integral closure of A∞,0 hT 1/p i/(T − g) inside (A∞,0 hT 1/p i/(T − g))[1/p]. André m proved that A∞ is almost faithfully flat over A∞,0 mod p for every m (for example, see [3, Theorem 2.3]). More importantly, André proved the following remarkable result: Theorem 3.5. [1, Theorem 0.3.1] Let A be a perfectoid algebra over a perfectoid field K of residue characteristic p. Suppose g ∈ A◦ is a nonzerodivisor and A contains a compatible system of p-power roots of g. Let B ′ be a finite étale A[ g1 ]algebra. Then (1) There exists a larger perfectoid algebra B between A and B ′ such that the inclusion A −→ B is continuous. We have B[ g1 ] = B ′ and B ◦ is contained ∞ ∞ in the integral closure of g −1/p A◦ in B ′ and it is a (pg)1/p -almost isomorphism. ∞ (2) For every m, B ◦ /pm is (pg)1/p -almost finite étale over A◦ /pm . ∞ Remark 3.6. It is easy to see that B ◦ is (pg)1/p -almost isomorphic to the integral ∞ closure of A◦ in B ′ : clearly this integral closure is (pg)1/p -almost contained in B ◦ . Now suppose we have an equation y n + a1 y n−1 + · · · + an = 0 ∞ where n = pe and a1 , . . . , an belongs to g −1/p A◦ . Then after multiplication by t t+e ∞ g 1/p , the above equation tells us that g 1/p y is integral over A◦ and thus g 1/p B ◦ ∞ is (pg)1/p -almost contained in the integral closure of A◦ in B ′ . Notation 3.7. We will frequently use the following notation throughout the rest of the article. R will be a complete and unramified regular local ring and S a modulefinite domain extension of R. Fix a nonzero element g ∈ R with ht(p, g)R = 2 such EXTENDED PLUS CLOSURE IN COMPLETE LOCAL RINGS 7 that R[(pg)−1 ] −→ S[(pg)−1 ] is finite étale. We construct A as in the paragraph after Lemma 3.3, and construct B, A∞,0 , A∞ as in the paragraph before Theorem 3.5. A will be the perfectoid algebra A∞ [p−1 ] and B ′ will be B ⊗A A∞ [(pg)−1 ]. By Theorem 3.5 we have B and B ◦ . The next lemma is well-known to some experts. We record it here for completeness. ∞ Lemma 3.8. Using the terminology of Notation 3.7, B ◦ is (pg)1/p -almost flat over R. ∞ Proof. By Theorem 3.5 we know B ◦ /pm is (pg)1/p -almost flat over A∞ /pm , and ∞ hence (pg)1/p -almost flat over A/pm by construction. Since A is flat over R,2 ∞ B ◦ /pm is (pg)1/p -almost flat over R/pm for all m. But we also know that B ◦ is p-torsion-free and p-adically complete since it is integral perfectoid. Now the same argument as in [15, Lemma 2.2] (simply replace A by R and A∞ by B ◦ ) shows that ∞ B ◦ is (pg)1/p -almost flat over R. ∞ Lemma 3.9. Using the terminology of Notation 3.7, there exists a (pg)1/p -almost map from B ◦ to S pg = Rpg . Proof. As W (k̄) is, up to p-adic completion, integral over W (k), we get an embedding W (k̄) −→ R++ and hence a map A −→ R++ (recall that A is the p-adic com∞ 1/p∞ 1/p∞ pletion of R⊗W (k) W (k̄)). We next fix an embedding A[p1/p , x1 , . . . , xd−1 ] −→ R++ . This in turn induces an embedding of the p-adic completions of the rings, i.e., an embedding α : A∞,0 −→ R++ = S ++ . We next extend this to a homomor∞ k k phism β : A∞,0 hT 1/p i −→ S ++ which sends T 1/p to g 1/p . In particular, since ∞ β(T − g) = 0, we get an induced homomorphism A∞,0 hT 1/p i/(T − g) −→ S ++ . ∞ Finally, since S ++ is integral perfectoid, it is (at least (p1/p )-almost) isomorphic to the integral closure of S ++ in S ++ [1/p]. Hence by the description of A∞ , we get ∞ a (p1/p )-almost map γ : A∞ −→ S ++ . This induces 1 ⊗ γ : S ⊗R A∞ −→ S ++ . But S ⊗R A∞ = S ⊗R (A ⊗A A∞ ) = B ⊗A A∞ . So this map extends to an almost map δ from the integral closure of B ⊗A A∞ in B ⊗A A∞ [(pg)−1 ] to S pg . As B ◦ ∞ is (pg)1/p -almost isomorphic to the integral closure of A∞ in B ⊗A A∞ [(pg)−1 ] ∞ by Remark 3.6, we get a (pg)1/p -almost map from B ◦ to S pg . The proof is complete. We are ready to prove our main theorem in the unramified case. Theorem 3.10. Using the terminology of Notation 3.7, if I is an ideal of R and ∞ J ⊆ R, then (pg)1/p (IS : J) ⊆ (I : J)B ◦ and consequently (IS : J) ⊆ (I : J)S epf . Proof. Because there exist a chain of maps: S −→ B −→ B ⊗ A∞ −→ B ◦ , ∞ we have (IS : J) ⊆ (IB ◦ : J) and the later one is (pg)1/p -almost isomorphic ∞ to (I : J)B ◦ because B ◦ is (pg)1/p -almost flat over R by Lemma 3.8. Thus ∞ ∞ (pg)1/p (IS : J) ⊆ (I : J)B ◦ . Next, by Lemma 3.9 there is a (pg)1/p -almost map ∞ from B ◦ to S pg = Rpg . This implies that (pg)1/p (IS : J) ⊆ (I : J)S pg . 2In general, if A −→ B is a map of p-adically complete and p-torsion-free rings with A Noetherian and A/p −→ B/p flat, then A −→ B is flat, see Remark 4.31 of [4]. 8 RAYMOND C HEITMANN AND LINQUAN MA So for every z ∈ (IS : J) and every ǫ = z ∈ (I : J)S epf by Lemma 3.3. 1 pt , (pg)ǫ z ∈ (I : J)S pg , and thus Now we can prove the usual form of colon-capturing: Corollary 3.11. Using the terminology of Notation 3.7, suppose y1 , . . . , yn are ∞ parameters in S. Then (pg)1/p ((y1 , . . . , yn−1 )B ◦ : yn ) ⊆ (y1 , . . . , yn−1 )B ◦ . As a consequence, we have ((y1 , . . . , yn−1 )S : yn ) ⊆ (y1 , . . . , yn−1 )S epf . ∞ Proof. We first prove (pg)1/p ((y1 , . . . , yn−1 )B ◦ : yn ) ⊆ (y1 , . . . , yn−1 )B ◦ . Assume k of the elements yi actually belong to R. If k = n, the claim is true because ∞ B ◦ is (pg)1/p -almost flat over R by Lemma 3.8. We now prove by induction on n − k; the case n − k = 0 is done. Suppose we have a counterexample with n − k minimal. First suppose yn ∈ / R. Clearly (y1 , . . . , yn )S ∩ R is not contained in any minimal prime of (y1 , . . . , yn−1 )S; so there exists wn ∈ R such that ((y1 , . . . , yn−1 )B ◦ : wn ) ⊇ ((y1 , . . . , yn−1 )B ◦ : yn ) and y1 , . . . , yn−1 , wn are param∞ eters in S. By the minimality assumption, we have (pg)1/p ((y1 , . . . , yn−1 )B ◦ : ∞ yn ) ⊆ (pg)1/p ((y1 , . . . , yn−1 )B ◦ : wn ) ⊆ (y1 , . . . , yn−1 )B ◦ and so we did not have a counterexample. Thus we know yn ∈ R. Without loss of generality, we may now assume y1 ∈ / R. If an ∈ ((y1 , . . . , yn−1 )B ◦ : yn ), we have an equation a1 y1 + · · · + an yn = 0 with each ai ∈ B ◦ . Then a1 ∈ ((y2 , . . . , yn )B ◦ : y1 ). The ∞ ◦ argument we used above nowP shows (pg)1/p aP 1 ∈ (y2 , . . . , yn )B . Consider any n n t ǫ ǫ ǫ = 1/p . We have (pg) a1 = i=2 bi yi . Then i=2 (bi y1 + (pg) ai )yi = 0 and so bn y1 + (pg)ǫ an ∈ ((y2 , . . . , yn−1 )B ◦ : yn ). Since we have decreased n but not k, n − k has been decreased and so (pg)ǫ (bn y1 + (pg)ǫ an ) ∈ (y2 , . . . , yn−1 )B ◦ . Thus ∞ (pg)2ǫ an ∈ (y1 , . . . , yn−1 )B ◦ . Hence (pg)1/p an ∈ (y1 , . . . , yn−1 )B ◦ . This finishes the proof of the first conclusion. Finally, for every z ∈ ((y1 , . . . , yn−1 )S : yn ) and every ǫ = p1t , z ∈ ((y1 , . . . , yn−1 )B ◦ : yn ) and so we have (pg)ǫ z ∈ (y1 , . . . , yn−1 )B ◦ . Now by Lemma 3.9, (pg)ǫ z ∈ (y1 , . . . , yn−1 )S pg and thus z ∈ (y1 , . . . , yn−1 )S epf by Lemma 3.3. Our next objective is to extend Theorem 3.10 to the ramified case. Theorem 3.12. Let R′ be a (possibly ramified) complete regular local ring of mixed characteristic p and let the integral domain S be a finite extension of R′ . If I is an ideal of R′ and J ⊆ R′ , then (IS : J) ⊆ (I : J)S epf . Proof. R′ is a finite extension of an unramified complete local ring R. We choose g so that R[(pg)−1 ] −→ S[(pg)−1 ] is finite étale, and (p, g)R is a height two ideal. We now use the standard framework of Notation 3.7. We first note that, just as in the proof of Theorem 3.10, our Lemma 3.9 and Lemma 3.3 will give us the ∞ desired result provided we can prove (pg)1/p (IB ◦ : J) ⊆ (I : J)B ◦ . Thus we will ∞ be done if we can show B ◦ is (pg)1/p -almost flat over R′ . This follows formally from the fact that R′ is regular and every system of parameters of R′ (in fact S) ∞ is a (pg)1/p -almost regular sequence on B ◦ by Corollary 3.11. We give a detailed argument below. ∞ Claim. B ◦ is (pg)1/p -almost flat over R′ . ′ ◦ 1/p∞ Proof of Claim. It is enough to show TorR -almost zero for all i (N, B ) is (pg) finitely generated R′ -modules N and every i > 0. We use descending induction. ′ ◦ This is clearly true when i > dim R′ since R′ is regular. Now suppose TorR k+1 (N, B ) ′ ∞ ◦ is (pg)1/p -almost zero for all finitely generated N . We want to show TorR k (N, B ) EXTENDED PLUS CLOSURE IN COMPLETE LOCAL RINGS 9 ∞ is (pg)1/p -almost zero for all finitely generated N . By considering a prime cyclic filtration of N , it is enough to prove this for N = R′ /P . Let h = ht P . We can pick a regular sequence x1 , . . . , xh ∈ P . Now P is an associated prime of (x1 , . . . , xh ) and so we have 0 −→ R′ /P −→ R′ /(x1 , . . . , xh ) −→ C −→ 0. The long exact sequence for Tor gives: ′ ′ ′ R R ◦ ′ ◦ ′ ◦ TorR k+1 (C, B ) −→ Tork (R /P, B ) −→ Tork (R /(x1 , . . . , xh ), B ). ′ ′ ◦ ◦ 1/p Now TorR -almost zero bek (R /(x1 , . . . , xh ),∞B ) = Hk (x1 , . . . , xh , B ) is (pg) 1/p ◦ cause x1 , . . . , xh is a (pg) -almost regular sequence on B by Corollary 3.11, and ′ ′ ◦ 1/p∞ ′ ◦ TorR -almost zero by induction. It follows that TorR k+1 (C, B ) is (pg) k (R /P, B ) ∞ 1/p is (pg) -almost zero. ∞ Remark 3.13. The analog of Theorem 3.10 and Theorem 3.12 in equal characteristic p > 0 is true if we replace extended plus closure by tight closure (in fact, even by the plus closure): suppose R −→ S is a module-finite extension of excellent local domains of characteristic p > 0 with R regular, let I, J be ideals of R, then we have (IS : JS) ⊆ (I : J)S ∗ . To see this, let z ∈ (IS : JS), then since we have R −→ S −→ R+ , we know that z ∈ (IR+ : JR+ ). Since R is regular and R+ is a balanced big Cohen-Macaulay algebra in characteristic p > 0, R+ is faithfully flat over R and thus (IR+ : JR+ ) = (I : J)R+ . Hence z ∈ (I : J)R+ ∩ S = (I : J)S + ⊆ (I : J)S ∗ . Theorem 3.14. Let R′ be a (possibly ramified) complete regular local ring of mixed characteristic p and let the integral domain S be finite over R′ . Let M be an R′ ′ N module. Then there exists c ∈ R′ such that the natural map cǫ TorR i (S, M )/p −→ ′ R + N Tori (S , M )/p is zero for all i ≥ 1 and every N, ǫ > 0. Consequently, for every ′ N ′ ǫ ′+ α ∈ TorR i (R , M )/p , there exists c ∈ R such that c α = 0 for every ǫ. Proof. As before, R′ is a finite extension of an unramified complete regular local ring R. We choose g so that R[(pg)−1 ] −→ S[(pg)−1 ] is finite étale, and (p, g)R is a height two ideal. We now use the standard framework of Notation 3.7. By ∞ the Claim in the proof of Theorem 3.12, B ◦ is (pg)1/p -almost flat over R′ . Thus ′ ′ ∞ ∞ R (pg)1/p Tori (B ◦ , M ) = 0 for all i > 0. It follows that (pg)1/p TorR i (S, M ) −→ ′ ′ pg ◦ TorR through TorR i (S , M ) is zero since it factors i (B , M ) by Lemma 3.9. At this pg 1/p∞ point, we note that S is (pg) -almost isomorphic to S ++ by Lemma 3.3; thus ′ ′ ∞ R ++ the map (pg)1/p Tori (S, M ) −→ TorR , M ) is also zero. Finally, tensoring i (S the commutative diagram / S+ 0 ·pN / S+ / S + /pN / S ++ / S ++ /pN /0 = / S ++ 0 N ·p with M induces a commutative diagram: ′ + N TorR i (S , M )/p /0 ′ + N / TorR i (S /p , M ) = ′ ++ TorR , M )/pN i (S / TorR′ (S ++ /pN , M ) i 10 RAYMOND C HEITMANN AND LINQUAN MA By the injectivity of the map on the first line of the above diagram, in order to show ′ ∞ R′ + N the image of (pg)1/p TorR i (S, M ) is zero in Tori (S , M )/p , it suffices to prove ′ + N its image is zero in TorR i (S /p , M ), which is clear because it factors through ′ R′ ++ N ++ Tori (S , M )/p and we already know the image is zero in TorR , M ). i (S ′+ N Remark 3.15. We cannot assert that TorR is almost zero or that i (R , M )/p + ′ N R /p is almost flat because each S requires its own c. ′ The next theorem is a simple consequence of our results, following directly from Corollary 3.11 and Theorem 2.8. However, as our proof relies heavily on André’s very deep perfectoid Abhyankar lemma, we would like to offer a more direct proof based on our earlier work on the vanishing conjecture for maps of Tor. Theorem 3.16. Let R be a regular local ring of mixed characteristic and let I be an ideal of R. Then I epf = I. Proof. We first claim that for every c 6= 0 in R, there exists e such that the natural e map R −→ R+ sending 1 to c1/p is pure. We prove this by contradiction. Suppose e there exists c 6= 0 such that for every e, the map R −→ R+ sending 1 −→ c1/p is ·c1/p e not pure. Let x1 , . . . , xd be a regular system of parameters of R. Since R −−−−→ R+ ·c1/p e d d d (R) −−−−→ Hm is not pure, the map Hm (R) to (R+ ) sends the socle element of Hm 0, i.e., there exists t (depending on e) such that (1) e t+1 t+1 + c1/p (x1 x2 · · · xd )t ∈ (xt+1 1 , x2 , . . . , xd )R . Next we claim that for every t > 0, (2) t+1 t+1 + t + ((xt+1 1 , x2 , . . . , xd )R : (x1 x2 · · · xd ) ) ⊆ (x1 , . . . , xd )R . t+1 t+1 + To see this, suppose u(x1 x2 · · · xd )t ∈ (xt+1 but u is not in 1 , x2 , . . . , xd )R (x1 , . . . , xd )R+ . Then there exists a domain S module-finite over R such that t+1 t+1 u(x1 x2 · · · xd )t ∈ (xt+1 1 , x2 , . . . , xd )S. Say u(x1 x2 · · · xd )t = r1 xt+1 + · · · + rd xt+1 1 d . Then (r1 , . . . , rd , −u) represents a class in TorR 1( R t+1 (xt+1 1 , . . . , xd , (x1 x2 · · · xd )t ) , S). Since u ∈ / (x1 , . . . , xd )R+ , u ∈ / (x1 , . . . , xd )S. By the valuation criterion for integral closure, we know that there exists a mixed characteristic DVR V such that u ∈ / (x1 , . . . , xd )V . This means (r1 , . . . , rd , −u) represents a nonzero class in t+1 t+1 t TorR 1 (R/(x1 , . . . , xd , (x1 x2 · · · xd ) ), V ). But this contradicts the vanishing conjecture for maps of Tor in mixed characteristic e [9, Theorem 1.3] applied to R −→ S −→ V . Now (1) and (2) shows that c1/p ∈ (x1 , . . . , xd )R+ for all e > 0. But this is clearly impossible by computing valuations. Finally, suppose x ∈ I epf , then for some c 6= 0, cǫ x ∈ (I, pN )R+ for every N, ǫ. e Since the map R −→ R+ sending 1 −→ c1/p is pure for some e, we know that x ∈ (I, pN )R for every N . This implies x ∈ I and hence I epf = I. We also have completed a closure-based proof of the Briançon-Skoda Theorem. EXTENDED PLUS CLOSURE IN COMPLETE LOCAL RINGS 11 Theorem 3.17. [14] Let R be a regular ring and let I be an ideal of R generated by n elements. Then the integral closure of I d+n is contained in I d+1 . Proof. It suffices to prove the result locally and so we can assume that we have either an equicharacteristic regular local ring or a mixed characteristic regular local ring. Hochster and Huneke have proved the result in the equal characteristic case. See [10, Theorem 5.4] for equal characteristic p case and [11] for equal characteristic 0 case (or [13, section 9.6] for generalizations in the equal characteristic 0 case using multiplier ideals). We can now handle the mixed characteristic case. By Theorem 2.7, we know that the integral closure of I d+n is contained in (I d+1 )epf . Theorem 3.16 completes the proof. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] Y. André, Le lemme d’Abhyankar perfectoide, arXiv:1609.00320. Y. André, La conjecture du facteur direct, arXiv:1609.00345. B. Bhatt, On the direct summand conjecture and its derived variant, arXiv:1608.08882. B. Bhatt, M. Morrow and P. Scholze, Integral p-adic Hodge theory, arXiv:1602.03148. O. Gabber and L. Romero, Almost ring theory, Lecture Notes in Mathematics, vol. 1800, Springer-Verlag, Berlin, 2003. R. Heitmann, The plus closure in mixed characteristic, J. of Algebra 193 (1997), 688–708. R. Heitmann, Extensions of plus closure, J. of Algebra 238 (2001), 801–826. R. Heitmann, The direct summand conjecture in dimension three, Ann. of Math. 156 (2002), 695–712. R. Heitmann and L. Ma, Big Cohen-Macaulay algebras and the vanishing conjecture for maps of Tor in mixed characteristic, arXiv:1703.08281. M. Hochster and C. Huneke, Tight closure, invariant theory, and the Briançon-Skoda theorem, J. Amer. Math. Soc. 3 (1990), 31–116. M. Hochster and C. Huneke, Tight closure in equal characteristic zero, preprint. M. Hochster and C. Huneke, Tight closure and elements of small order in integral extensions, J. of Pure & Appl. Algebra 71 (1991), 233–247. R. Lazarsfeld, Positivity in algebraic geometry. II, Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics], vol. 49, Springer-Verlag, Berlin, 2004. J. Lipman and A. Sathaye, Jacobian ideals and a theorem of Briançon-Skoda, Michigan Math. J. 28 (1981), 199–222. L. Ma and K. Schwede, Perfectoid multiplier/test ideals in regular rings and bounds on symbolic powers, arXiv:1705.02300. P. Scholze, Perfectoid spaces, Publ. Math. Inst. Hautes Études Sci. 116 (2012), 245–313.	0math.AC
arXiv:1711.08410v1 [math.GR] 22 Nov 2017 LATTICE ENVELOPES URI BADER, ALEX FURMAN, AND ROMAN SAUER Abstract. We introduce a class of countable groups by some abstract grouptheoretic conditions. It includes linear groups with finite amenable radical and finitely generated residually finite groups with some non-vanishing ℓ2 Betti numbers that are not virtually a product of two infinite groups. Further, it includes acylindrically hyperbolic groups. For any group Γ in this class we determine the general structure of the possible lattice embeddings of Γ, i.e. of all compactly generated, locally compact groups that contain Γ as a lattice. This leads to a precise description of possible non-uniform lattice embeddings of groups in this class. Further applications include the determination of possible lattice embeddings of fundamental groups of closed manifolds with pinched negative curvature. 1. Introduction 1.1. Motivation and background. Let G be a locally compact second countable group, hereafter denoted lcsc1. Such a group carries a left-invariant Radon measure unique up to scalar multiple, known as the Haar measure. A subgroup Γ < G is a lattice if it is discrete, and G/Γ carries a finite G-invariant measure; equivalently, if the Γ-action on G admits a Borel fundamental domain of finite Haar measure. If G/Γ is compact, one says that Γ is a uniform lattice, otherwise Γ is a nonuniform lattice. The inclusion Γ ֒→ G is called a lattice embedding. We shall also say that G is a lattice envelope of Γ. The classical examples of lattices come from geometry and arithmetic. Starting from a locally symmetric manifold M of finite volume, we obtain a lattice embedding π1 (M ) ֒→ Isom(M̃ ) of its fundamental group into the isometry group of its universal covering via the action by deck transformations. The real Lie group Isom(M̃ ) is semi-simple if M̃ has no Euclidean direct factors, and the lattice Γ < Isom(M̃ ) is uniform if and only if M is compact. The fact that semi-simple real Lie groups are algebraic and can be expressed as zero sets of polynomials allows to construct arithmetic lattices, such as SLd (Z) < SLd (R). 2000 Mathematics Subject Classification. Primary 20565; Secondary 22D05. Key words and phrases. Locally compact groups, lattices, geometric group theory. In part, U.B. and A.F. were supported by BSF grant 2008267 and A.F. by NSF grant DMS 1611765. All three authors thank MSRI for its support during the Geometric Group Theory Program. 1If G is, in addition, totally disconnected we call G a tdlc group. 1 2 URI BADER, ALEX FURMAN, AND ROMAN SAUER Analogous constructions exist for products of real and p-adic algebraic semi-simple groups, such as SLd (Z[ 1 1 , . . . , ]) < SLd (R) × SLd (Qp1 ) × · · · × SLd (Qpk ). p1 pk Our notations and conventions for arithmetic lattices are defined in §4.1 below. A central theme in the study of lattices are the connections between the lattices and the ambient group. Mostow’s Strong Rigidity ([52] for the uniform case, [45,57] for the non-uniform case) asserts that an irreducible lattice in a semi-simple real Lie group G 6≃ SL2 (R) determines the ambient Lie group among all Lie groups and determines its embedding in the ambient Lie group uniquely up to an automorphism of G. It is natural to ask to what extent lattices determine their lattice envelopes among all lcsc groups, and which countable groups have only trivial lattice envelopes. To make the question precise we introduce the notion of virtual isomorphism between lattice embeddings in Definition 3.1, which generalizes the notion of weak commensurability or virtual isomorphism of abstract groups (see Definition 2.2). A lattice embedding is called trivial if it is virtually isomorphic to the identity homomorphism of a countable discrete group, which can be regarded as a lattice embedding. Virtual isomorphism is an equivalence relation. We refer to §3.1 for more information. Problem 1.1. Given a countable group Γ, describe all its possible lattice envelopes up to virtual isomorphism. In the study of lattices, the non-uniform lattices are often harder to work with when compared to uniform ones. For example, the condition of integrability of lattices (e.g. required for different purposes in [4, 37, 46, 50, 64]) and the spectral gap condition (also known as weak cocompactness, which is e.g. required in [41]) are automatically satisfied by uniform lattices and can be proven for some examples of non-uniform lattices, often using elaborate arguments. This motivates the following. Problem 1.2. How prevalent are non-uniform lattice embeddings? 1.2. Main result. To state our main result, Theorem A, towards Problem 1.1 we need to put some conditions on the group Γ. Definition 1.3. We say that a countable group Γ has property • (BT) if there is an upper bound on the order of finite subgroups of Γ. • (Irr) if Γ is not virtually isomorphic to a product of two infinite groups. • (CAF) if every amenable commensurated subgroup of Γ is finite. • (NbC) if for every normal subgroup N ⊳ Γ and every commensurated subgroup M < Γ such that N ∩ M = {1} there is a finite index subgroup of M that commutes with N . The abbreviations stand for Bound on Torsion, Irreduciblity, Commensurated Amenable is Finite, and Normal by Commensurated. For the definition of commensurated subgroup see §2.1. In §2 we prove the following result that shows that large classes of groups satisfy the above conditions, and we also describe these classes in more detail. LATTICE ENVELOPES 3 Theorem 1.4. The following countable groups have property (CAF): (1) Groups with a non-vanishing ℓ2 -Betti number in some degree. (2) Linear groups with finite amenable radical. (3) Groups in the class Dreg introduced in [68]. This class contains all acylindrically hyperbolic groups and thus non-elementary convergence groups. Further, the following countable groups have property (NbC): (1’) Groups with non-vanishing first ℓ2 -Betti number. (2’) Linear groups. (3’) Finitely generated residually finite groups. (4’) Groups in the class Dreg . The groups in (1’) and (4’) also have the property (Irr). The statement about (NbC) for the groups of case (3’) is due to CapraceKropholler-Reid-Wesolek [20, Corollary 18]. Of course, linear groups fall into this case provided they are finitely generated. The next theorem was announced in [5] with a slightly stronger formulation of the (NbC) and (CAF) condition. It gives a partial answer to Problem 1.1 that concerns possible lattice envelopes of a given group Γ. We impose no restriction on the lattice envelope, except for it being compactly generated. This condition can also be removed if Γ is assumed to be finitely generated (see Lemma 3.6). Theorem A (Structure of possible lattice envelopes). Let Γ be a countable group with properties (Irr), (CAF), and (NbC), and let Γ < G be a non-trivial lattice embedding into a compactly generated lcsc group G. Then the lattice embedding Γ ֒→ G is virtually isomorphic to one of the following: (1) an irreducible lattice in a connected, center-free, semi-simple real Lie group without compact factors, or (2) the inclusion of a subgroup commensurable with H(O[S]) into a (possibly trivial) tree extension of HK,S , where S, being a finite set of places compatible with H, and H, K are as in §4.1. (3) a lattice in a non-discrete totally disconnected locally compact group without normal amenable subgroups. If in addition Γ has (BT), then the lattice in (3) is uniform. For the notion of tree extension see Definition 4.3. The three cases are distinct (see Remark 4.2). We show in Example 2.14 the necessity of the condition (NbC) for the theorem above. The classical case of lattices in semi-simple real Lie groups includes both arithmetic and non-arithmetic examples (non-arithmetic lattices can appear only in SO(n, 1) and SU(n, 1)). The case of S-arithmetic lattices in the above statement refers to irreducible lattices in a product of finitely many real semi-simple Lie groups and p-adic ones, where factors of both types are present. Case (3) contains a large class of examples that includes lattices in p-adic groups, and fundamental groups Γ of some finite cubical complexes whose universal cover has a non-discrete group of automorphisms. This last case remains quite mysterious although a structure theory of non-discrete, simple, totally disconnected groups emerged in the last decade (in this regard, Caprace’s survey [19] is recommended). However assuming property (BT) the last case does not allow non-uniform lattices. Hence we obtain the following partial solution to Problem 1.2. 4 URI BADER, ALEX FURMAN, AND ROMAN SAUER Corollary 1.5 (Classification of non-uniform lattices). Let Γ be a countable group with properties (BT), (Irr), (CAF), and (NbC). Then every non-uniform lattice embedding Γ ֒→ G of Γ is virtually isomorphic to either (1) an irreducible lattice in a connected, center free, semi-simple real Lie group without compact factors, or (2) a (possibly trivial) tree extension of an S-arithmetic lattice with both real and p-adic factors present. In particular, every lattice embedding of Γ is weakly compact, and, with the exception of lattice embeddings into Lie groups isogenous to SL2 (R) or SL2 (C), every lattice embedding of Γ is square integrable. This is a direct corollary except for the last statement which is a consequence of results about weak compactness and square integrability of classical lattices: Weak compactness of lattices in connected semi-simple Lie groups is stated in [47, III.1.12 on p. 113] and proved in [8, Lemma 3]. Weak compactness of lattices in S-arithmetic groups as in Setup 4.1 follows from the stronger property (T) [47, III.5]. The statement about square integrability is taken from Shalom’s work (see [64, §2] for the S-arithmetic and higher rank case and [65, Theorem 3.7] for the rank 1 case). 1.3. Applications. As applications of our main result, we now present more precise classification results of lattice envelopes for specific groups. Their proofs can be found in §6. Theorem B (Mostow rigidity with locally compact targets). Let Γ < H be either an irreducible lattice in a center-free, real semi-simple Lie group H without compact factors and H 6≃ SL2 (R), or an S-arithmetic lattice H(O[S]) < HK,S as in Setup 4.1. Then any other lattice embedding of Γ < G into an lcsc group G is virtually isomorphic to Γ < H. This theorem may be viewed as a generalization of Mostow’s Strong Rigidity, corresponding to the special case of Theorem B, where H is a real semi-simple Lie group and G is also known to be a semi-simple real Lie group (in most accounts one even assumes G = H and the focus is on aligning two given lattice embeddings of Γ in G by an automorphism of G). In fact, Mostow considered the case of uniform lattice embeddings [52]; the non-uniform were later obtained by Prasad [57] and Margulis [45]. In [33] the second author proved Theorem B for the case of a simple real Lie group H with rkR (H) ≥ 2 and a general lattice envelope Γ < G, and for the case of rank-one real Lie group H and a uniform lattice envelope Γ < G. The case of H = PSL2 (R) is excluded from Theorem B for two reasons. First, strong rigidity does not hold in this setting – the moduli space of embeddings of a surface group in the same group G = H = PSL2 (R) is multidimensional. Secondly, non-uniform lattices in PSL2 (R), that are virtually free groups of finite rank 2 ≤ n < ∞, can be embedded as a lattice in completely different lcsc groups, such as PGL2 (Qp ), or the automorphism group Aut(T ) of a locally finite tree T . Our next results show that all examples are related to the ones mentioned above. Theorem C (Lattice embeddings of free groups). Let Γ be a finite extension of a finitely generated non-abelian free group Fn and let Γ < G be a non-trivial lattice embedding. LATTICE ENVELOPES 5 (1) If Γ < G is non-uniform, then it is virtually isomorphic to a non-uniform lattice in PSL2 (R). (2) If Γ < G is uniform, then it is virtually isomorphic to a lattice in a closed cocompact subgroup of the automorphism group of a tree. Note that the second possibility in the Theorem includes such examples as PGL2 (Qp ) or other rank-one non-Archimedean groups. For uniform lattices in PSL2 (R), such as surface groups, the possibilities for lattice embeddings are even more restricted. Theorem D (Lattice embeddings of surface groups). Let Γ be a uniform lattice in PSL2 (R). Any other non-trivial lattice embedding of Γ is virtually isomorphic to a uniform lattice embedding into PSL2 (R). Let M be a manifold that admits a Riemannian metric with strictly negative sectional curvature, but is not homeomorphic to a locally symmetric one. We conjecture that fundamental groups of such manifolds Γ = π1 (M ) have no nontrivial lattice envelopes. Under an additional pinching assumption, we are able to prove this. Theorem E (Lack of lattice embeddings in pinched negative curvature). The fundamental group of a closed Riemannian manifold M of dimension n ≥ 5 whose sectional curvatures range in 1 2 ) , −1 −(1 + n−1 does not admit a non-trivial lattice embedding unless M is homeomorphic to a closed hyperbolic manifold. Some remarks about this result are in order. Gromov and Thurston [38] construct infinitely many examples of negatively curved manifolds with the above pinching in dimensions n ≥ 4 that are not homeomorphic to hyperbolic manifolds. By taking connected sums of a hyperbolic manifold with an exotic sphere Farrell and Jones construct closed smooth manifolds that are homeomorphic to a hyperbolic manifold but whose smooth structure does not support a hyperbolic Riemannian metric [30]. We recommend the survey [31] by Farrell, Jones and Ontaneda on these issues. Finally, we obtain the following surprising characterization of free groups. (2) Theorem F (Non-uniform lattice embeddings of groups with β1 > 0). Let Γ be a group with non-vanishing first ℓ2 -Betti number that has an upper bound on the order of its finite subgroups. If Γ possesses a non-uniform compactly generated lattice envelope, then Γ has a non-abelian free subgroup of finite index. 1.4. Structure of the paper. We devote most of §2 to the proof of Theorem 1.4. We discuss an example by Burger and Mozes in §2.4 that shows the necessity of the condition (NbC); if one drops (NbC) while still keeping (CAF) then there are exotic lattice embeddings that are not covered by the three cases of Theorem A. The definition of virtually isomorphism of lattice embeddings and the tools of the trade for the proof of Theorem A are provided in §3. The most difficult proof in §3 is the result about outer automorphisms of S-arithmetic lattices (Theorem 3.13). While this result is known or expected to hold by the experts there was no proof so far in the required generality. We hope its proof provides a useful reference. 6 URI BADER, ALEX FURMAN, AND ROMAN SAUER The bulk of the paper is devoted to the proof of Theorem A in §5. The paper [5] in which we announced the results of this paper and provided proofs in special cases might be helpful for the reader. In the first step of the proof we rely on property (CAF) and the positive solution of Hilbert’s 5th problem to show that the given lattice embedding Γ < G is virtually isomorphic to one into a product of a semi-simple Lie group L and a totally disconnected group D. So we may assume that G = L × D. Let U < D be a compact-open subgroup. In the second step of the proof we split L further as L = L0 × L1 in a way so that the projection of Γ to L1 is dense. Depending on the finiteness of Γ ∩ (L × U ) and Γ ∩ ({1} × L1 × U ), we distinguish three cases in the third step which correspond to the cases (1), (2) or (3) of the statement of Theorem A. The second case is the most sophisticated. Here we have to identify the lattice Γ ∩ ({1} × L1 × D′ ) < L1 × D′ , where D′ < D is a certain closed cocompact subgroup, as an S-arithmetic lattice, and D′ with the corresponding non-Archimedean factor. This identification is an arithmeticity theorem which is proved in our companion paper [6] and heavily relies on Margulis’ commensurator rigidity. We actually prove in that paper a result of greater generality which does not assume compact generation of the ambient group. The version needed here is only slightly more general than the arithmeticity theorem by Caprace and Monod in [23, Theorem 5.20]. Our basic setup for S-arithmetic groups and a detailed explanation of the arithmeticity theorem is provided in §4. Further, we use Margulis’ normal subgroup theorem and the aforementioned result on outer automorphism groups to conclude that L = L1 and L0 = {1} in the second case. The condition on (NbC) appears just once in the whole proof, namely in the third step of the proof, in Lemma 5.9. The final step of the proof identifies the difference between L × D′ and L × D as a tree extension in the sense of Setup 4.1. Important ingredients here are the fact that D and D′ are quasi-isometric, the quasi-isometry rigidity theorem of KleinerLeeb [42] and the work of Mosher-Sageev-White [51]. In the proof of the applications, that is Theorems B–F, a major step is always to identify the most mysterious case, namely (3), in Theorem A. To this end, we appeal to ideas of geometric group theory as in the final step of the proof of Theorem A. This is based on the fact that a lattice with property (BT) is quasi-isometric to any of its totally disconnected lattice envelopes. 2. The properties (BT), (Irr), (CAF), and (NbC) We describe the classes in Theorem 1.4 in more detail. In §2.3 we prove Theorem 1.4. In §2.4 we describe an exotic lattice embedding of a group that does not have property (NbC). 2.1. Commensurated and commensurable groups. We recall the following well known definitions. Definition 2.1. A subgroup A of Γ is commensurated by a subset S ⊂ Γ if s−1 As∩ A has finite index in s−1 As and A for every s ∈ S. If A < Γ is commensurated by all of Γ we just say that A < Γ is commensurated. Definition 2.2. Two groups Γ and Λ are commensurable if there are subgroups Γ′ < Γ and Λ′ < Λ of finite index such that Γ′ ∼ = Λ′ . In this case we write Γ ≃ Λ. Further, Γ and Λ are weakly commensurable or virtually isomorphic if there are LATTICE ENVELOPES 7 subgroups Γ′ < Γ and Λ′ < Λ of finite index and finite normal subgroups N ⊳ Γ′ and M ⊳ Λ′ such that Γ′ /N ∼ = Λ′ /M . The notion of virtual isomorphism is generalized to lattice embeddings in Definition 3.1. The proof of the following easy lemma is left to the reader. Lemma 2.3. Being commensurated is preserved in the following situations: (1) Preimages of commensurated subgroups under homomorphisms are commensurated. (2) Intersections of commensurated subgroups are commensurated. (3) A finite index subgroup of a commensurated subgroup is itself commensurated. Both (CAF) and (NbC) are not necessarily preserved if one passes to a weakly commensurable group. They are, however, preserved by passing to quotients by finite kernels. Lemma 2.4. Let K ⊳ Γ be a finite normal subgroup. If Γ has property (CAF), then Γ/K has (CAF). Proof. Let p : Γ → Γ/K be the projection. Let A < Γ/K be an amenable, commensurated subgroup. Then p−1 (A) < Γ is commensurated by Lemma 2.3. Since p−1 (A) is a finite extension of A, it is amenable too. By the property (CAF) the group p−1 (A) is finite, thus A is finite. Lemma 2.5. Let K ⊳ Γ be a finite normal subgroup. If Γ has property (NbC), then Γ/K has (NbC). Proof. Let M, N be subgroups of Γ/K such that M ∩ N = {1}, M ⊳ Γ/K, and M < Γ/K is commensurated. Then p−1 (M ) < Γ and p−1 (N ) < Γ satisfy the same properties (see Lemma 2.3). Hence there is a finite index subgroup of p−1 (M ) that commutes with p−1 (N ). Its projection is a finite index subgroup of M that commutes with N . 2.2. Groups in the class Dreg and acylindrically hyperbolic groups. The class Dreg was introduced by Thom [68]. It is closely related to the class Creg of Monod-Shalom. Both Dreg and Creg are an attempt to define negative curvature for groups in a cohomological way. The definition of Dreg is analytical. Let π : Γ → U (H) be a unitary representation. A map c : Γ → H is a quasi-1cocycle for π if Γ × Γ ∋ (g, h) 7→ π(g)c(h) − c(gh) + c(g) ∈ H is uniformly bounded on Γ × Γ. The vector space of quasi-1-cocycles modulo the bounded ones forms a group QH 1 (Γ, H). The class Dreg is the class of groups for which QH 1 (Γ, ℓ2 (Γ)) has non-zero L(Γ)-dimension in the sense of Lück [44, Section 6.1]. More concrete is the description of non-amenable groups in Dreg : Proposition 2.6 ([68, Corollary 2.5]). A non-amenable group Γ is in Dreg if and only if there is an unbounded quasi-1-cocycle Γ → ℓ2 (Γ). The following theorem was proved by Bestvina-Fujiwara [9] and Hamenstädt [39] for classes of groups that were later identified as acylindrically hyperbolic groups by Osin [54]. 8 URI BADER, ALEX FURMAN, AND ROMAN SAUER Theorem 2.7 ([54, Theorem 8.3]). The class Dreg strictly contains the class of acylindrically hyperbolic groups. We also mention the following result of Sun [67]. Theorem 2.8. Non-elementary convergence groups are acylindrically hyperbolic. Remark 2.9. In the sequel, we show properties (NbC) and (CAF) for groups in Dreg . Using the difficult Theorem 2.7 this implies (NbC) and (CAF) for acylindrically hyperbolic groups. We emphasize, however, that one can deduce (CAF) and (NbC) directly and quite easily for acylindrically hyperbolic groups. For example, (CAF) follows directly from [54, Corollary 1.5] and the non-amenability of acylindrically hyperbolic groups. We show that groups in Dreg satisfy the (NbC) property for the trivial reason that the situation in which to apply (NbC) cannot happen. That means, we show the following stronger property for groups in Dreg . Definition 2.10. A group Γ has property (NbC)1 if it satisfies the following: If N ⊳ Γ is a normal subgroup and M < Γ is a commensurated subgroup and N ∩ M = {1}, then N is finite or M is finite. It is obvious that (NbC)1 implies (NbC). Lemma 2.11. Let Λ be a group with a normal subgroup N ⊳Λ and a commensurated subgroup M < Λ such that Λ = N · M and N ∩ M = {1}. To every finite subset F ⊂ Λ we can assign subgroups NF < N and MF < M in such a way that the following holds: (1) NF is normalized by M ; (2) MF ⊳ M is normal and has finite index; (3) F ⊂ ΛF := NF M and NF MF < ΛF has finite index; (4) MF commutes with NF , thus NF MF ∼ = NF × M F ; (5) F ⊂ F ′ implies NF ⊂ NF ′ and MF ⊃ MF ′ . In particular, if N is finitely generated, then M has a finite index subgroup that commutes with N . Proof. For every F ⊂ Λ let MF be the normal core of the subgroup \ λM λ−1 ⊂ M. MF′ := λ∈F ∪{1} MF′ Since M is commensurated, < M , hence MF < M , are of finite index. In other words, MF is the largest normal subgroup of M with the property λMF λ−1 ⊂ M for all λ ∈ F . Let q : Λ → N be the map q(nm) = n; it is well defined because of N ∩ M = {1}. ′ Clearly, MF′ = Mq(F ) , thus MF = Mq(F ) . Since N ⊳ Λ is normal, the commutator −1 −1 [n, m] = n(mn m ) lies in N . For m ∈ MF = Mq(F ) and n ∈ q(F ) we also have [n, m] = (nmn−1 )m−1 ∈ M , hence [n, m] ∈ M ∩ N = {1}. So MF commutes with the subset q(F ). Let [ S := mq(F )m−1 ⊂ N. m∈M Because of normality MF ⊳ M the group MF commutes with the subset S as well. The subgroups NF := hSi < N and MF satisfy all the required properties. LATTICE ENVELOPES 9 Theorem 2.12. Every group in Dreg is (Irr), (CAF) and (NbC)1 . Proof. The properties (Irr) and (CAF) are shown in [68, Theorem 3.4]. Let Γ ∈ Dreg , and let N ⊳ Γ and M < Γ normal and commensurated subgroups, respectively, with trivial intersection. Assume by contradiction that both subgroups are infinite. According to loc. cit. N is non-amenable. Let N = {n1 , n2 , . . .} be an enumeration of N . We refer to the notation of Lemma 2.11 applied to our situation. The group N is the increasing union of subgroups NFi < N for Fi = {n1 , . . . , ni }. Hence there is some k ∈ N such that N ′ := NFk is non-amenable, thus infinite. The group N ′ commutes with the finite index subgroup M ′ := MFk of M . Since M < Γ is commensurated, also M ′ < Γ is commensurated. By [68, Lemma 3.3] the restriction maps QH 1 (Γ, ℓ2 (Γ)) → QH 1 (M ′ , ℓ2 (Γ)) and QH 1 (N ′ M ′ , ℓ2 (Γ)) → QH 1 (M ′ , ℓ2 (Γ)) to commensurated subgroups are injective. Hence the restriction map QH 1 (Γ, ℓ2 (Γ)) → QH 1 (N ′ M ′ , ℓ2 (Γ)) is injective. We reach a contradiction as the latter module vanishes by [68, Theorem 3.4] and the fact that N ′ M ′ is a product of two infinite groups. 2.3. Conclusion of proof of Theorem 1.4. Proof of property (CAF). We refer to the cases in the statement of Theorem 1.4. (1) Since all ℓ2 -Betti numbers of an infinite amenable group vanish by a result of Cheeger-Gromov [24], every group with some non-vanishing ℓ2 -Betti number is (CAF) according to [7, Corollary 1.4]. (2) This is exactly [14, Theorem 6.8]. (3) See Theorem 2.12. For the classes included in Dreg see Theorems 2.7 and 2.8. Proof of property (NbC). We refer to the cases in the statement of Theorem 1.4. (1’) Let Γ be a group with positive first ℓ2 -Betti number. Assume by contradiction that Γ does not have (NbC)1 . Let M, N < Γ be subgroups with M ∩N = {1} such that N ⊳ Γ is normal and M is commensurated by Γ. Asssume that neither M nor N are finite. (2) We have to show that β1 (Γ) = 0. Since Λ := N M is commensurated by (2) Γ, the vanishing of β1 (Λ) would imply this according to [7, Corollary 1.4]. By Lemma 2.11 the group Λ is an ascending union of groups Λn such that Λn is virtually a product of infinite groups. A product of infinite groups has vanishing first ℓ2 -Betti number because of the Künneth formula for ℓ2 -Betti numbers and the fact that the zeroth ℓ2 -Betti number of an infinite group (2) vanishes. The same is true for a virtual product, so β1 (Λn ) = 0 for large n. This follows from the fact that the vanishing of ℓ2 -Betti numbers is an invariant of virtual isomorphism which can be easily deduced from their basic properties [44]. Or one may cite the much more general result that the vanishing of ℓ2 -Betti numbers of lcsc groups is a coarse invariant [61]. (2) (2) Finally, β1 (Λn ) = 0 for all n implies β1 (Λ) = 0 by [44, Theorem 7.2 (2)]. (2’) Let Γ < GLn (k) be a linear group, and let N, M < Γ subgroups with trivial intersection such that N ⊳ Γ is normal and M < Γ is commensurated by Γ. Let F 7→ NF , MF be an assignment of finite subsets to subgroups as in Lemma 2.11. By the Noetherian property of the algebraic group GLn (k) 10 URI BADER, ALEX FURMAN, AND ROMAN SAUER there exists some finite set F1 ⊂ N such that the Zariski closure M̄1 of MF1 is minimal among all finite subsets of N . Thus, for every n ∈ N , MF0 ∪{n} is Zariski dense in M̄1 . Since [n, MF0 ∪{n} ] = 1 we get that [n, M̄1 ] = 1. In particular, the finite index subgroup MF1 of M commutes with N . This implies (NbC) for linear groups. (3’) As mentioned before, this is exactly [20, Corollary 18]. (4’) See Theorem 2.12. Proof of property (Irr). By the Künneth formula for ℓ2 -Betti numbers and the fact from the previous proof that their vanishing is unaffected by passing to virtually isomorphic groups it follows that groups with positive first ℓ2 -Betti numbers have property (Irr). That groups in Dreg have (Irr) is contained in Theorem 2.12. 2.4. Examples of groups without property (NbC). Among the conditions required for our main result, the (NbC) property is the most opaque. We present two examples of groups that do not have property (NbC). The first example is a wreath product. The second one comes with a lattice envelope that shows the necessity of the (NbC) condition in Theorem A. Example 2.13. For an infinite residuallyQfinite group M and a countable group H we consider the product group H M = M H endowed with the natural (shift) action of M . Let N be the subgroup in H M consisting of periodic elements, that is, elements having an M -stabilizer of finite index in M . N is clearly a countable subgroup on which M acts. Let Γ = M ⋉ N . Observe that Γ = N M , N ⊳ Γ, M is commensurated by Γ and N ∩ M = {e}. The subgroup N does not commute with any finite index subgroup of M , yet any finitely generated subgroup of N does so (cf. Lemma 2.11). Example 2.14. Let Λ be an irreducible lattice in SLn (Qp )× SLn (R). Let X be the symmetric space of SLn (R), and let B be the 1-skeleton of the Bruhat-Tits building associated with SLn (Qp ). Then the action of Λ on B × X can be extended to an action of an extension, denoted by Γ, of Λ by the fundamental group π1 (B), which is an infinitely generated free group, on the universal covering B̃ × X. Note that T := B̃ is a tree. Moreover, Γ is an irreducible lattice in Aut(T ) × SLn (R). This type of example originates from the work of Burger-Mozes [18, 1.8]. More general constructions in this direction are discussed in Caprace-Monod [22, 6.C]. The group Γ has the properties (BT), (Irr), and (CAF), but it lacks property (NbC). Let us show property (CAF) first. Let A < Γ be a commensurated subgroup. The image p(A) under the projection p : Γ → Λ is an amenable commensurated subgroup of Γ. It follows from Theorem 1.4 that Λ has (CAF). Thus p(A) is finite. Since π1 (B) satisfies (CAF) by loc. cit., the intersection A ∩ π1 (B) is finite as well, which implies that A is finite. Let U ⊂ Aut(T ) be the stabilizer of a vertex; it is a compact open subgroup. Then it is easy to verify that the normal subgroup N = π1 (B) ⊳ Γ and the subgroup M = Γ ∩ (U × SLn (R)), which is commensurated by Lemma 3.5, violate property (NbC): No finite index subgroup of M commutes with N . LATTICE ENVELOPES 11 3. General facts about locally compact groups and their lattices In §3.1 we introduce the analog of weak commensurability or virtual isomorphism of groups for lattice embeddings. In §3.2 and §3.3 we have a look at broader classes of subgroups in lcsc groups than lattices, discrete subgroups and closed subgroups of finite covolume. In §3.4 we recall the definition of the amenable radical and study its behaviour under passage to closed subgroups of finite covolume. Finally, in §3.5 and §3.6 we study automorphisms of lcsc groups and their lattices in different categories: outer automorphism groups and quasi-isometry groups. 3.1. Virtual isomorphism of lattice embeddings. Definition 3.1. Two lattice embeddings Γ ֒→ G and Γ′ ֒→ G′ are virtually isomorphic if there are open finite index subgroups H < G and H ′ < G′ and compact normal subgroups K ⊳H and K ′ ⊳H ′ and a topological isomorphism H/K → H ′ /K ′ that restricts to an isomorphism of lattices (Γ ∩ H)/(Γ ∩ K) → (Γ′ ∩ H ′ )/(Γ′ ∩ K ′ ), that is, there is a commutative square: (Γ ∩ H)/(Γ ∩ K) H/K ∼ = (Γ′ ∩ H ′ )/(Γ′ ∩ K ′ ) ∼ = H ′ /K ′ A lattice embedding is trivial if it is virtually isomorphic to the identity homomorphism of a countable discrete group. In particular, if Γ ֒→ G and Γ′ ֒→ G′ are virtually isomorphic then Γ and Γ′ are virtually isomorphic in the sense of Definition 2.2. To justify the above definition one needs that the horizontal maps in the above diagram are lattice embeddings. This follows from Lemma 3.3 and Lemma 3.4. Proposition 3.2. Virtual isomorphism is an equivalence relation on the class of lattice embeddings. Proof. The only non-obvious claim is transitivity. Its proof is similar to the proof that weak commensurability among groups is transitive. Let jl : Γl ֒→ Gl for l ∈ {1, 2, 3} be lattice embeddings such that j1 , j2 and j2 , j3 are virtually isomorphic. So for l ∈ {1, 2, 3} there are finite index subgroups Hl < Gl and H2′ < G2 and compact normal subgroups Kl ⊳ Hl and K2′ ⊳ H2′ and isomorphisms ∼ = ∼ = H1 /K1 − → H2 /K2 and H2′ /K2′ − → H3 /K3 that restrict to isomorphisms (Γ1 ∩ H1 )/(Γ1 ∩ K1 ) ∼ = (Γ2 ∩ H2 )/(Γ2 ∩ K2 ) and ′ ∼ ′ (Γ2 ∩ H )/(Γ2 ∩ K ) = (Γ3 ∩ H3 )/(Γ3 ∩ K3 ), 2 2 respectively. The subgroup H2′′ := H2 ∩ H2′ < G2 has finite index in G2 . Let pr : H2′′ → Q2 := H2′′ /(K2 ∩ H2′ ) be the projection. Let K2′′ be the kernel of the composition of quotient maps H2′′ → Q2 → Q2 / pr(K2′ ∩ H2 ). 12 URI BADER, ALEX FURMAN, AND ROMAN SAUER Note that K2′′ contains K2 ∩ H2′ = K2 ∩ H2′′ and K2′ ∩ H2 = K2′ ∩ H2′′ . Since K2 and pr(K2′ ∩ H2 ) are compact, K2′′ ⊳ H2′′ is a compact normal subgroup of H2′′ . Consider the compositions of quotient maps and isomorphisms as above ∼ = ∼ = f1 : H1 → H1 /K1 − → H2 /K2 and g3 : H2′ → H2′ /K2′ − → H3 /K3 . Similarly, using the inverses of the above isomorphisms we obtain homomorphisms ∼ = ∼ = g1 : H2 → H2 /K2 − → H1 /K1 and f3 : H3 → H3 /K3 − → H2′ /K2′ . We set Ci := gi (H2′′ ) and Ei := gi (K2′′ ) ⊳ Ci for i ∈ {1, 3}; further, let C̃i < Hi and Ẽi < Hi be the preimages of Ci and Ei under the quotient maps. Then C̃i has finite index in Hi , thus also in Gi , and Ẽi is a compact normal subgroup of C̃i for i ∈ {1, 3}. The homomorphism C̃i /Ẽi → H2′′ /K2′′ , [x] 7→ [fi (x)] is an isomorphism for i ∈ {1, 3} with inverse map [y] 7→ [gi (y)]. Since fi and gi preserve the lattices, this isomorphism restricts to an isomorphism (Γi ∩ C̃i )/(Γi ∩ Ẽi ) ∼ = (Γ2 ∩ H2′′ )/(Γ2 ∩ K2′′ ) for i ∈ {1, 3}. This finishes the proof of transitivity. 3.2. Discrete subgroups of locally compact groups. We collect some important, albeit easy facts used in the proof of the main result. Lemma 3.3 (cf. [21, 2.C]). Let Γ < G be a lattice in an lcsc group. Let U < G be an open subgroup. Then Γ ∩ U is a lattice in U . Lemma 3.4 ([59, Theorem 1.13 on p. 23]). Let G be an lcsc group and N ⊳ G a normal closed subgroup. The projection of a lattice Γ < G to G/N is discrete if and only if Γ ∩ N is a lattice in N . If so, then the projection of Γ to G/N is also a lattice. The following statement is our main source of commensurated subgroups. Lemma 3.5. Let G be an lcsc group and U < G be a compact open subgroup. Then U is commensurated by G. For any subgroup Λ < G the intersection Λ ∩ U is commensurated in Λ. Proof. For any g ∈ G, U ∩ gU g −1 is an open subgroup of the compact group U (resp. in gU g −1 ), hence it has finite index in U (resp. in gU g −1 ). The second statement follows by taking g ∈ Λ and taking intersections with Λ. Lemma 3.6. Any lattice envelope of a finitely generated group is compactly generated. Proof. Let Γ < G be a lattice in an lcsc group, and let Γ be generated by a finite set F ⊂ Γ. Let V ⊂ G be an open neighborhood of the identity with compact closure. We claim that the compact set K = V ∪ F generates a subgroup G′ of finite index in G. Indeed, G′ contains the group generated by V , hence G′ is open. Thus G/G′ is a countable discrete space. But Γ ⊂ G′ , hence G/G′ is finite. Hence by adding a finite set to K one obtains a compact generating set for G. LATTICE ENVELOPES 13 3.3. Closed subgroups of finite covolume. The notion of a lattice, that is a discrete subgroup of finite covolume, can be generalized to closed, but not necessarily discrete, subgroups as follows. A closed subgroup H < G of an lcsc group G is said to have finite covolume in G if G/H carries a finite G-invariant Borel measure. If H has finite covolume in G, then G is unimodular if and only if H is unimodular [59, Lemma 1.4 on p. 18]. Lemma 3.7 ([59, Lemma 1.6 on p. 20]). Let G be an lcsc group, and H < L < G closed subgroups. Then H < G has finite covolume if and only if both L < G and H < L have finite covolumes. It is well known that a torsion-free lattice in a tdlc group is uniform. A mild generalization is given by the following. Lemma 3.8 ([2, Corollary 4.11]). Let G be a unimodular tdlc group. Let H < G be a closed subgroup of finite covolume. Let mH be the Haar measure of H. If there is an upper bound on the Haar measures of compact open subgroups in H, i.e. sup mH (U ) \| U is a compact open subgroup in H < ∞, then G/H is compact. Lemma 3.9. Let G be a connected, semi-simple, center-free, real Lie group without compact factors. Every closed subgroup H < G of finite covolume has the form H = ∆ × G2 , where G2 is a direct factor of G, G = G1 × G2 , and ∆ < G1 is a lattice. Proof. The Lie algebra h of the connected component H 0 is Ad(H)-invariant. Since H is Zariski dense in G by Borel’s density theorem [10], it follows that h is an ideal in the semisimple Lie algebra g of G. Let g = g1 ⊕ · · · ⊕ gn be the decomposition into simple Lie algebras. Let I1 = {i ∈ {1, . . . , n} \| gi ∩ h = {0}}, and let I2 = {1, . . . , n}\I1 . Then G = G1 × G2 with Gk being the Lie subgroup with Lie algebra M gi . i∈Ik Since gi is simple we either have gi ∩ h = {0} or gi ∩ h = gi . This implies H = (G1 × {1} ∩ H) × G2 where ∆ := G1 × {1} ∩ H is discrete. Since H has finite covolume, ∆ is a lattice. 3.4. The amenable radical. Definition 3.10. The amenable radical of an lcsc group G is the maximal closed normal amenable subgroup. We denote it by Radam (G) ⊳ G. Lemma 3.11. Let G be an lcsc group, and let H < G be a closed subgroup of finite covolume. Then Radam (H) = Radam (G) ∩ H. This lemma can be deduced from the following result. 14 URI BADER, ALEX FURMAN, AND ROMAN SAUER Lemma 3.12 (Furstenberg [35]). Let G be an lcsc group and G y M a continuous action that is minimal and strongly proximal. Let H < G be a closed subgroup of finite covolume. Then the restriction of the G-action to H is also minimal and strongly proximal. In [35] Furstenberg was interested in the restriction to lattices, but the proof applies to general closed subgroups of finite covolume. Proof of Lemma 3.11 from Lemma 3.12. The very definition of Radam (G) yields the inclusion Radam (G) ∩ H < Radam (H). For the converse inclusion, we use an equivalent characterization of the amenable radical as the common kernel of all minimal, strongly proximal actions on compact metrizable spaces [34, Proposition 7]. Thus for the converse inclusion it suffices to show that given an arbitrary minimal and strongly proximal G-action G → Homeo(M ), its restriction to H is also minimal and strongly proximal, because it would show that Radam (H) acts trivially in every minimal strongly proximal G-action, which is the content of Lemma 3.12. 3.5. (Outer) automorphism groups. We describe various results about groups of automorphisms and groups of outer automorphisms of certain groups. We are confident that the main result of this section, Theorem 3.13, is known to experts, but it does not seem to be in the literature in the appropriate generality. Theorem 3.13. Let K1 , . . . , Kn be number fields, and let Hi be connected, noncommutative, adjoint, absolutely simple Ki -groups. Let Si ⊂ Vi be finite compatible sets of places of Ki as in §4.1. Then any group that is (abstractly) commensurable with n Y Hi (Oi [Si ]) i=1 has a finite outer automorphism group. The proof of Theorem 3.13 will be given by the end of this subsection. This theorem will be used in §5 together with Lemma 3.14 to split certain extensions of groups. For a subgroup B < A of a group A, let Aut(A), AutB (A), and Out(A) denote the automorphism group of A, the subgroup of Aut(A) preserving B set-wise, and the outer automorphism group of A. For a σ-compact lcsc group G we denote the group of continuous automorphisms by Autc (G) and the group of continuous outer automorphisms by Outc (G) = Autc (G)/G. Note that a continuous automorphism of G is a homeomorphism by the open mapping theorem [12, IX. §5.3]. We begin by quoting two auxiliary results which are elementary. Lemma 3.14 ([16, Corollary (6.8) and Exercise 1 on p. 106]). Let B ⊳ A be a normal subgroup in a group A. Assume that B has trivial center. Then the right hand square in the commutative diagram B A A/B ρ B Aut(B) Out(B) LATTICE ENVELOPES 15 whose rows are exact and whose vertical homomorphisms are induced by conjugation is a pullback diagram of groups. Thus, if ρ is trivial, then the group extension B ֒→ A ։ A/B splits as a direct product. Lemma 3.15 ([72]). Let B ⊳ A be a normal subgroup in a group A. Let C be the center of B. Then the sequence of groups (3.1) j π 1 → Z 1 (A/B; C) − → AutB (A) − → Aut(B) × Aut(A/B) is exact. Here the left group is the abelian group of 1-cocycles of A/B in C, j maps a cocycle c to a 7→ c([a])a, and π(f ) = (f \|B , [f ]). The following three results are immediate consequences of Lemma 3.15. Lemma 3.16. Every topological automorphism of a product G = G0 × Gtd of a center-free connected group G0 and a tdlc group Gtd is a product of a topological automorphism of G0 and a topological automorphism of Gtd . Proof. We apply Lemma 3.15 to A = G and B = G0 . In this situation the map π in (3.1) has an obvious right inverse s. Both π and s preserve continuity of automorphisms. Let α : G → G be a continuous automorphism. Then α(G0 ) = G0 . Since G0 has trivial center, π is injective. So α = s(π(α)). Lemma 3.17. Let A be a group, and let B ⊳ A be a characteristic subgroup of finite index. If B has a finite center, then the restriction map Aut(A) → Aut(B) has a finite kernel. Proof. Denote by K the kernel of Aut(A) = AutB (A) → Aut(B) and by C the center of B. Since Aut(A/B) is finite, it is enough to show that the kernel K ′ of K → Aut(A/B) is finite. By Lemma 3.15, K ′ is isomorphic to Z 1 (A/B; C) which is finite since A/B and C are finite. Corollary 3.18. If a group A has a characteristic subgroup of finite index that has a finite center and a finite outer automorphism group, then also A has a finite outer automorphism group. Proof. Let B ⊳ A be characteristic with finite center and finite outer automorphism group. Since the group Aut(A)/B surjects onto Aut(A)/A = Out(A), it is enough to show that the former is finite. This follows by the exactness of 1 → ker Aut(A) → Aut(B) → Aut(A)/B → Out(B), as Out(B) is finite by assumption, and the kernel is finite by Lemma 3.17. For the purposes of this subsection, we introduce the following definition. Definition 3.19. A group A is said to be strongly irreducible if for every homomorphism π : B1 × B2 → A with finite cokernel, either B1 < ker π or B2 < ker π. A group is said to be strongly outer finite if the outer automorphism group of any of its finite index subgroups is finite. Lemma 3.20. A Zariski dense subgroup of the K-points of a connected, noncommutative, adjoint, K-simple K-algebraic group is strongly irreducible. Proof. Let H be a connected, non-commutative, adjoint, K-simple K-group, and let Γ < H(K) be a Zariski dense subgroup. 16 URI BADER, ALEX FURMAN, AND ROMAN SAUER Let π : B1 × B2 → Γ be a homomorphism with finite cokernel. Since Γ < H(K) is Zariski dense, the Zariski closure of im(π) has to be H as a finite index subgroup of the connected group H. The Zariski closure of π(Bi ) define K-subgroups Hi , i ∈ {1, 2} that commute in H and together generate H. By the simplicity of H we conclude that for some i ∈ {1, 2}, Hi is trivial. It follows that Bi < ker π. Theorem 3.21. Let B be a group that is (abstractly) commensurable to a product of a finite family of finitely generated, strongly irreducible, strongly outer finite groups. Then B has a finite outer automorphism group. The proof of Theorem 3.21 will be preceded by some preparation. First we make the following ad-hoc definition, which is not standard. Q Definition Q 3.22. A standard subgroup of a product i∈I Ai of groups is a subgroup of the form i∈I Bi with each Bi < Ai being a finite index subgroup. Lemma 3.23. Let (Ai )i∈I be a finite family of strongly irreducible groups, and let (Bj )j∈J be a finite family of non-trivial groups with \|J\| ≥ \|I\|. If Y φ Y → Bj − Ai j∈J i∈I is a monomorphism with finite cokernel, then there are a bijection f : J → I and, for each j ∈ J, a monomorphism φj : Bj → Af (j) with finite cokernel such that φ is the product of the maps φj . In particular, the image im(φ) is a standard subgroup. Proof. Let A and B be the product of (Ai )i∈I and (Bj )j∈J , respectively. The proof is by induction on \|I\|. If \|I\| = 1, there is nothing to prove. Let \|I\| > 1. Then \|J\| > 1, and Qwe can decompose it non-trivially as J = J1 ∪ J2 . For k ∈ {1, 2} we set B (k) = j∈Jk Bj . We let Y Ik = {i ∈ I \| B2−k < ker πi } and A(k) = Ai , i∈Ik where πi : A → Ai is the i-th projection. By the strong irreducibility of the Ai ’s, I = I1 ∪ I2 is a partition. Hence the composition of φ : B → A with the projection A → A(k) factors through B (k) ∼ = B/B (2−k) , thus giving a homomorphism φk : B (k) → (k) A . We observe that these φk are injective homomorphisms with finite cokernels. Without loss of generality, \|I1 \| ≤ \|J1 \|. By induction hypothesis, \|I1 \| = \|J1 \| and therefore \|I2 \| ≤ \|J2 \|. Applying the induction hypothesis to both φk , the statement follows. Applying Lemma 3.23 to automorphisms φ, we obtain the following. Corollary 3.24. Let (Ci )i∈I be a finite family of strongly irreducible groups. Then the obvious embedding Y Y Aut(Ci ) → Aut Ci i∈I i∈I has a finite cokernel. Lemma 3.25. Let (Ai )i∈I be a finite family of strongly irreducible Q groups, and let A be its product. Let B < A be a subgroup of finite index. Then i∈I (B ∩ Ai ) is a characteristic subgroup of B of finite index. LATTICE ENVELOPES 17 Proof. Let us call a subgroup of B that is simultaneously a standard subgroup of A a substandard subgroup. By Lemma 3.23, every φ ∈ Aut(B) preserves substandard Q subgroups. Hence i∈I (B ∩ Ai ) is characteristic in B as it is the unique maximal element in the collection of substandard subgroups. Proof of Theorem 3.21. By assumption, B has a finite index subgroup C ′ < B that is isomorphic to a finite index subgroup of the product A of a finite family (Ai )i∈I of finitely generated, strongly irreducible, strongly outer finite groups. From now on we will identify C ′ and its various subgroups with their images in A under this isomorphism. As each Ai is finitely generated, A, C ′ , and thus B are finitely generated too. Thus C ′ has a finite index subgroup C ′′ that is characteristic in B (e.g. the intersection of all subgroups of B with index [B : C ′ ]). By Lemma 3.25 C ′′ has Q a characteristic subgroup C of finite index which is standard in A. That is, C = i∈I Ci , where Ci < Ai are of finite index. Strongly irreducibility passes to finite index subgroups. So each Ci is strongly irreducible. Being characteristic is transitive, so C is characteristic in B. Q By Corollary 3.24, the obvious embedding i∈I Aut(Ci ) → Aut(C) has a finite cokernel. By the strong outer finite assumption, Out(Ci ) is finite for each i ∈ I. Hence Out(C) is finite. By the strong irreducibility assumption C has a finite center. Thus, Corollary 3.18 implies that Out(B) is finite. The following proposition is well known. Proposition 3.26. Let K be a number field, and let H be non-commutative, connected, adjoint, absolutely simple Q K-group. Let S ⊂ V be a finite compatible set of places of K as in §4.1. Then ν∈S H(Kν ) has a finite group of continuous outer automorphisms. Proof. By Lemma 3.20 each group H(Kν ) is strongly irreducible. By Corollary 3.24, the natural embedding Y Y j Aut(H(Kν )) − → Aut( H(Kν )) ν∈S ν∈S has a finite cokernel. Further, if f ∈ im(j) is a continuous automorphism then f is a product of continuous automorphisms. This implies that the restriction of j Y Y Autc (H(Kν )) → Autc ( H(Kν )) ν∈S ν∈S has a finite cokernel, too. It is thus enough to show that Outc (H(Kν )) is finite for every ν ∈ S. Since every continuous field automorphism of Kν is trivial on the closure of Q over which Kν is finite, the field Kν has a finite group of continuous field automorphisms. Moreover, the group H has a finite group of algebraic outer automorphisms, given by Dynkin diagram automorphisms. Thus, Outc (H(Kν )) is indeed finite by [47, Chapter I, 1.8.2(IV)]. Proposition 3.27. Let K be a number field, and let H be a non-commutative, connected, adjoint, absolutely simple K-group. Let S ⊂ V be a finite compatible set of places of K as in §4.1. Then H(O[S]) is finitely generated, strongly irreducible and strongly outer finite. 18 URI BADER, ALEX FURMAN, AND ROMAN SAUER Proof. By [47, Chapter IX, Theorem 3.2] H(O[S]) is finitely generated, and by Lemma 3.20 H(O[S]) is strongly irreducible. We are left to prove that H(O[S]) is strongly outer finite. To this end, consider a subgroup Γ < H(O[S]) of finite index. Q The group Γ is an irreducible lattice in H = ν∈S H(Kν ). Let A = Autc (H). The conjugation homomorphism H → A is injective as H is center-free and has a finite cokernel by Proposition 3.26. Then Γ ֒→ A embeds as a lattice. We claim: (1) The normalizer N = NA (Γ) of Γ in A is discrete in A. (2) The conjugation action N → Aut(Γ) is an isomorphism. Let us first conclude the finiteness of Out(Γ) from (1) and (2). Since N contains the lattice Γ and N is discrete in A, N < A is also a lattice. Hence [N : Γ] < ∞. By (2), Out(Γ) is isomorphic to N/Γ, hence finite. Regarding (1), discreteness will follow once we show that N is countable because N < A is closed. Since Γ is finitely generated, Aut(Γ) is countable. Hence N/ZA (Γ) is countable, where ZA (Γ) denotes the centralizer of Γ in A. Each factor of H is center-free, so ZH (Γ) is trivial by the density of the projection of Γ in each factor. Since H < A is of finite index, ZA (Γ) is finite, and the countability of N follows. To verify (2), it is enough to show that every automorphism of Γ induces a continuous automorphism of H. This is a consequence of the Mostow-Prasad strong rigidity theorem for H [47, Theorem VII 7.1]. Proof of Theorem 3.13. By Proposition 3.27 the groups Hi (Oi [Si ]) are finitely generated, strongly irreducible and strongly outer finite. Thus the proof follows at once from Theorem 3.21. Finally, we need a classical result from the work of Tits and Borel-Tits. Theorem 3.28. Let K be a number field, and let ν a finite place of K. Let H be a non-commutative, connected, adjoint, absolutely simple K-group of Kν -rank ≥ 2. Then the automorphism group of the Bruhat-Tits building associated to H and ν contains H(Kν )+ as a subgroup of finite index. Proof. In higher rank the automorphism group of the associated Bruhat-Tits building X is isomorphic to Aut(H)(Kν ) which is isomorphic to an extension of the group Aut(H(Kν )) of algebraic automorphisms over Kν by a subgroup of field automorphisms of Kν . This follows from the results in [69]. See [3, Proposition C.1] for an elaboration. The group Aut(H(Kν )) contains H(Kν )+ as a subgroup of finite index, and Kν has only finitely many field automorphisms since every field automorphism is continuous [66] and trivial on the closure of the rationals over which Kν is finite. This implies the statement. 3.6. Quasi-isometries of locally compact groups. Recall that a map between metric spaces f : (X, dX ) → (Y, dY ) is an (L, C)-quasi-isometry if 1 · dX (x, x′ ) − C < dY (f (x), f (x′ )) < L · dX (x, x′ ) + C L for all x, x′ ∈ X and for every y ∈ Y there is x ∈ X so that dY (y, f (x)) < C. Two maps f, g : (X, dX ) → (Y, dY ) have bounded distance if sup dY (f (x), g(x)) < ∞. x∈X The quasi-isometry group QI(X, dX ) of (X, dX ) is the group (under composition) of equivalence classes of quasi-isometries (X, dX ) → (X, dX ) under bounded distance. LATTICE ENVELOPES 19 Lemma 3.29. Let G be an lcsc group containing a closed subgroup H < G such that G/H is compact, the projection G → H\G has locally a continuous cross-section, and H is compactly generated. Let H → Isom(X, d) be a properly discontinuous cocompact action on some proper geodesic metric space (X, d). Then the natural homomorphism φ : H → Isom(X, d) → QI(X, d) extends to a homomorphism φ̄ : G → QI(X, d) with the following property. There exist constants L, C > 0 so that each φ̄(g), g ∈ G, is represented by a (L, C)quasi-isometry fg : X → X such that for every bounded set B ⊂ X there is a neighbourhood of the identity V ⊂ G with ∀g∈V ∀x∈B d(fg (x), x) ≤ C. (3.2) Remark 3.30. The projection G → H\G of an lcsc group G to the coset space of a closed subgroup H < G has locally a continuous cross-section if G is finitedimensional with respect to covering dimension [40, Theorem 2]. Tdlc groups, which are the ones we will apply Lemma 3.29 to, are zero-dimensional. Proof of Lemma 3.29. Since H is compactly generated, and H\G is compact, G is also compactly generated. Thus G and H possess left-invariant word metrics that are unique up to quasi-isometry. The inclusion j : H → G is a quasi-isometry. We construct a quasi-inverse j ′ : G → H of j, which is unique up to bounded distance, in the following way. Let pr : G → H\G be the projection. Let s : U0 → G be a local continuous crosssection of pr defined on some compact neighbourhood U0 ⊂ H\G of H1 such that s(H1) = 1. We can extend s to a global cross-section s : H\G → G with relatively compact image s(H\G), which is not necessarily continuous. Then j ′ : G → H, g 7→ g · s(pr(g))−1 is a quasi-inverse of j. The assignment f 7→ j ◦ f ◦ j ′ yields an isomorphism j∗ : QI(H) → QI(G). Let x0 ∈ X. By the Švarc-Milnor theorem the map h → hx0 is a quasi-isometry f : H → X. We pick a quasi-inverse f ′ : X → H of f . Similarly, we obtain an isomorphism f∗ : QI(H) → QI(X). Let H → QI(H) be the homomorphism that sends a group element to its left translation. Similarly, G → QI(G). It is easy to see that the map φ coincides with the composition f∗ H → QI(H) −→ QI(X). We define φ̄ to be the composition j −1 f∗ ∗ G → QI(G) −− → QI(H) −→ QI(X). Clearly, φ̄ extends φ. Let lgG : G → G denote left translation by g ∈ G in G. Similarly, let lhH : H → H be left translation by h ∈ H. Then φ̄(g) is represented by the quasi-isometry fg := f ◦ j ′ ◦ lgG ◦ j ◦ f ′ . Clearly, there are constants C, L > 0 only depending on j, j ′ , f , and f ′ such that fg is a (L, C)-quasi-isometry. Let B ⊂ X be a bounded subset. Let D := sup d x, f (f ′ (x)) = sup d x, f ′ (x)x0 < ∞. x∈X x∈X 20 URI BADER, ALEX FURMAN, AND ROMAN SAUER The set f ′ (x) \| x ∈ B ⊂ H is bounded, thus relatively compact. Let K ⊂ H be its closure. The continuous map H × G → H\G, (h, g) 7→ pr(gh) sends H × {1} to H1. By continuity and compactness of K there is an identity neighbourhood V0 ⊂ G such that K × V0 is mapped to U0 ⊂ H\G. Then the function ψ : K × V0 → R, (h, g) 7→ d ghs(pr(gh))−1 x0 , hx0 is continuous and maps K × {1} to 0. By continuity and compactness of K there is an identity neighborhood V ⊂ V0 such that ψ(K × V ) ⊂ [0, 1]. Now let x ∈ B and g ∈ V . Let h = f ′ (x) ∈ K. Then d fg (x), x = d f (j ′ (gj(f ′ (x)))), x) = d j ′ (gf ′ (x))x0 , x) ≤ d j ′ (gh)x0 , hx0 ) + D = d ghs(pr(gh))−1 x0 , hx0 ) + D ≤ D + 1. Upon replacing the constant C by max{C, D + 1}, the statement follows. 4. Arithmetic lattices, tree extensions, and the arithmetic core In §4.1 we introduce the notion of tree extensions of S-arithmetic lattices, which appears in our main result, Theorem A. In §5 we discuss an arithmeticity result for lattices in a product of a semisimple Lie group and a tdlc group, which identifies the second case in Theorem A. 4.1. Arithmetic lattices and tree extensions. Setup 4.1. Let K be a number field, O its ring of integers, and let H be a connected, noncommutative, absolutely simple adjoint K-group. Let V be the set of inequivalent valuations (places) of K, let V ∞ denote the archimedean ones, and V fin = V − V ∞ the non-archimedean ones (finite places). For ν ∈ V we denote by Kν the completion of K with respect to ν; it is a local field. Let S ⊂ V be a finite subset of places that is compatible with H in the sense of [6]. Explicitly, this means the following: (1) For every ν ∈ S, H is Kν -isotropic. (2) S contains all ν ∈ V ∞ for which H is Kν -isotropic. (3) S contains at least one finite and one infinite place. Let S ∞ = S ∩ V ∞ and S fin = S ∩ V fin . Let O[S] ⊂ K be the ring of S-integers. For ν ∈ V let H(Kν )+ < H(Kν ) be the normal subgroup as defined in [11, Section 6]. If Kν is perfect, which is the case in our situation, then H(Kν )+ is the subgroup generated by all unipotent elements and it has finite index in H(Kν ). If ν ∈ V ∞ , then H(Kν )+ is just the connected component of the identity in the real Lie group H(Kν ). Define Y Y HK,S := H(Kν ) H+ H(Kν )+ . K,S := ν∈S The quotient HK,S /H+ K,S is finite. ν∈S LATTICE ENVELOPES 21 By reduction theory after Borel and Harish-Chandra the diagonal embedding realizes H(O[S]) is a lattice in HK,S . Note that HK,S = HK,S ∞ × HK,S fin is the splitting into a semi-simple real Lie group and a totally disconnected locally compact group. We thus make the following remark. Remark 4.2. If S is finite set of places compatible with H, then H(O[S]) is a lattice in a product of at least two lcsc group, one of them being a Lie group and one of them being a tdlc group. We now exhibit a generalization of the previous type of lattices. Definition 4.3. Let S be finite. Let S1fin denote the finite places ν ∈ S fin such that the Kν -rank of H is 1; denote by Tν the associated Bruhat-Tits tree and by Aut(Tν ) the tdlc group of automorphisms of this simplicial tree. Any closed intermediate group Y Y Y H+ Aut(Tν ) H(Kµ ) × H(Kν )+ < H ∗ < K,S = ν∈S µ∈S−S1fin ν∈S1fin is called a tree extension of HK,S . Q + ∗ Since H+ K,S = ν∈S H(Kν ) < H is cocompact, the inclusion of a subgroup commensurable to H(O[S]) into a tree extension of HK,S is a lattice embedding by Lemma 3.7. Remark 4.4. Denote by Xν the Bruhat-Tits buildings or the symmetric space of + H(Kν ). For ν ∈ S − S1fin the group Isom(X Q ν ) contains H(Kν ) as a subgroup of finite index Q by Theorem 3.28. Further, ν∈S Isom(Xν ) is a finite index subgroup of Isom( ν∈S Xν ). This follows from [15, Proposition I.5.3(4) on p. 56]. Thus any closed intermediate subgroup Y ∗ H+ Xν ). K,S < H < Isom( ν∈S is, up to passing to a finite index subgroup, a tree extension of HK,S . 4.2. Arithmetic core theorem. Next we state an arithmeticity result for lattices in products that provides a key step in the proof of Theorem A. Theorem 4.5 (Arithmetic Core Theorem). Let H be a connected, center-free, semi-simple, real Lie group without compact factors, D be a tdlc group and Γ < H × D be a lattice. Assume that (1) the projection Γ → D has a dense image, (2) the projection Γ → H has a dense image, (3) the projection Γ → H is injective, (4) D is compactly generated. Then there exist number fields K1 , . . . , Kn , connected, non-commutative, adjoint, absolutely simple Ki -groups Hi , finite sets Si ⊂ Vi of places of Ki compatible with Hi as in §4.1 with the following properties: There is a topological isomorphism H∼ = n Y (Hi )0Ki ,Si∞ i=1 22 URI BADER, ALEX FURMAN, AND ROMAN SAUER and a continuous epimorphism D ։ Q with compact kernel such that Q is a closed intermediate subgroup n n Y Y (Hi )Ki ,Sifin (Hi )+ < Q < fin K ,S i i=1 i i=1 and the image of Γ in H × Q is commensurable with the image of n n n Y Y Y (Hi )Ki ,Sifin . (Hi )Ki ,Si∞ × Hi (Oi [Si ]) −→ i=1 i=1 i=1 This result is quite close to Theorem 5.20 in the paper [23] of Caprace and Monod. In fact, if H is assumed to be simple instead of semi-simple, Theorem 4.5 is essentially their theorem. In our companion paper [6] we prove Theorem 4.5 deducing it from a more general statement, in which D is not assumed to be compactly generated. In this more general case the sets Si of places of Ki might be infinite, and the gap between H+ K,S and HK,S becomes large. In the current paper we use the above arithmetic core theorem 4.5 in §5, Step 3, case (iii). At this point in the proof we do not know yet that Γ (called Γ3 there) is finitely generated, but do know compact generation of D (called Gtd 3 there) inherited from the compact generation assumption on G. For reader’s convenience we sketch the idea of the proof of Theorem 4.5 as it appears in [6]. Apart from step (1) below our approach differs from that taken by Caprace and Monod in [23]. In the following sketch we ignore some important details for the sake of transparency. (1) Choose a compact open subgroup U < D, and observe that the projection ∆ of ΓU := Γ ∩ (H × U ) to H is a lattice in H, commensurated by the projection of all of Γ to H. If ∆ is irreducible, Margulis’ commensurator super-rigidity and arithmeticity theorems provide a number field K and a simple K-algebraic group H, so that the semi-simple real Lie group H is isogenous to HK,S ∞ and ∆ is commensurable to the arithmetic lattice H(O); here S ∞ is the set of infinite places ν ∈ V ∞ for which H(Kν ) is non-compact. If ∆ is reducible, it is commensurable to a product ∆1 × · · · × ∆n < H1 × · · · × Hn where ∆i < Hi are irreducible lattice, leading to fields Ki and Ki -simple groups Hi , i = 1, . . . , n. For clarity we continue with n = 1 case. (2) We now view Γ as a subgroup of the commensurator CommenH (H(O)) of the arithmetic lattice H(O) ≃ ∆. If one glosses over the difference between the simply-connected and adjoint forms of H (such as SLd and PGLd ) then the above commensurator is the subgroup of rational points H(K), and therefore Γ < H(K). Define S fin to be the set of those non-archimedean places ν ∈ V fin for which the image of Γ in H(Kν ) is unbounded. (For example, for SLd (Z) < Γ < PSL(Q) the set S fin would consist of primes p that appear with arbitrarily high powers in denominators of entries of γ ∈ Γ). Then Γ is commensurable to a subgroup of H(O[S]). (3) For ν ∈ S fin the image of Γ in the tdlc group H(Kν ) is not precompact. It can be showed using Howe-Moore’s theorem (alternatively, use [58]) that this implies that the image of Γ is dense in (an open subgroup of finite LATTICE ENVELOPES 23 index in) H(Kν ). One can even show that the image of Γ is dense in (an open subgroup of finite index in) HK,S fin . (4) One now considers the closure L of the diagonal imbedding of Γ in D × HK,S fin . Using the fact that the projections to both factors are dense and that the closure of ∆ ≃ H(O) in both projections is open compact, one shows that L is a graph of a continuous epimorphism D → HK,S fin with compact kernel. (5) Finally, the fact that Γ < H×D is a lattice, implies that its image, contained in the S-arithmetic lattice H(O[S]), is a lattice in H × HK,S fin ≃ HK,S . Thus Γ ≃ H(O[S]). If D is compactly generated, then so is HK,S fin , which implies that S fin (hence also all of S) is finite. It also follows that Γ is finitely generated. 5. Proof of Theorem A The starting point of the proof of Theorem A is a consequence of Hilbert’s 5th problem which was observed by Burger-Monod [17, Theorem 3.3.3]. Theorem 5.1. Every locally compact group H contains an open, normal, finite index subgroup H ′ containing Radam (H) such that the quotient H ′ / Radam (H) is isomorphic to a direct product of a connected, center-free, semi-simple, real Lie group without compact factors and a tdlc group with trivial amenable radical. Step 1. Reduction to a lattice in a product. It is only the very first step of the proof where we shall take advantage of property (CAF). Theorem 5.2. The amenable radical of a lattice envelope of a group with property (CAF) is compact. Proof. Let G′ < G be an open, normal, finite index subgroup so that G1 := G′ / Radam (G) ∼ = Gss × Gtd 1 1 Gss 1 as in Theorem 5.1, namely is a connected, center-free, semi-simple, real Lie td group without compact factors, and Gtd 1 is a tdlc group with Radam (G1 ) = {1}. td Pick a compact open subgroup U < G1 . We consider the following commutative diagram: =:A }\| { z Γ ∩ p−1 (j({1} × U )) G p−1 G (j({1} × U )) Γ G Γ ∩ G′ G′ pG pG′ j({1} × U ) G/ Radam (G) G′ / Radam (G) j ∼ = td Gss 1 × G1 The arrows are the obvious inclusions and projections. Moreover, j is defined by requiring commutativity. First we show that A is commensurated by Γ which is equivalent to j({1}×U ) being commensurated by G/ Radam (G) by Lemma 2.3. Let g ∈ G/ Radam (G). Since G′ is normal in G and thus G′ / Radam (G) is normal G/ Radam (G), conjugation by 24 URI BADER, ALEX FURMAN, AND ROMAN SAUER td ∼ ss td g is a continuous automorphism of j(Gss 1 × G1 ) = G1 × G1 . By Lemma 3.16 ss ss this automorphism is the product of continuous automorphisms css g : G1 → G1 and td td td td cg : G1 → G1 . Hence conjugation by g maps j({1} × U ) to j({1} × cg (U )). Since td td the subgroup ctd g (U ) < G1 is open and compact, the intersection cg (U ) ∩ U has finite index in U . This implies that j({1} × U ) is commensurated by G/ Radam (G) from which we conclude that A is commensurated by Γ. Let G′U < G′ be the preimage of Gss 1 × U under pG′ , and let pG′ ss K := ker G′U −− → Gss 1 × U → G1 . The group K is an extension of the amenable group Radam (G) by the compact group U . Hence K is amenable. The group Γ ∩ G′U is a lattice in G′U by Lemma 3.3. We now invoke a deep result of Breuillard and Gelander [13, §9]. It follows from [13, Theorem 9.5] and K being a closed amenable subgroup that the projection of the lattice Γ ∩ G′U in G′U /K is discrete. The result of Breuillard and Gelander is a generalization of Auslander’s theorem where K is a solvable Lie group. By Lemma 3.4, the group ′ A = Γ ∩ p−1 G′ (j({1} × U )) = K ∩ (Γ ∩ GU ) −1 is a lattice in K, thus is amenable. Note that we used p−1 G (j({1}×U )) = pG′ (j({1}× U )) here which follows from the diagram. On the other hand A is commensurated by Γ and hence finite. This implies that K and its closed subgroup Radam (G) are compact. Let us summarize the situation after the first step: Proposition 5.3. Let Γ have property (CAF), and let Γ < G be a lattice embedding. Then the amenable radical K := Radam (G) of G is compact. By passing to an open normal finite index subgroup G′ < G containing K and to the finite index subgroup Γ′ := Γ ∩ G′ < Γ and taking quotients by compact or finite normal subgroups G1 = G′ /K and Γ1 := Γ′ /(Γ′ ∩ K) one obtains a lattice embedding td Γ1 < G1 = Gss 1 × G1 , virtually isomorphic to Γ < G, into a product of a connected, center free, semisimple real Lie group without compact factors Gss 1 and a tdlc group with trivial td amenable radical Gtd 1 . Furthermore, G is compactly generated if and only if G1 is compactly generated. Step 2: Projection to the semi-simple factor is irreducible. td Let Γ1 < G1 = Gss 1 × G1 be a lattice in a product as in Proposition 5.3. Since compact generation of G is assumed in Theorem A, the group G1 is compactly generated. The center-free, semi-simple, real Lie group Gss 1 splits as a direct product of the simple factors: Gss 1 = S1 × · · · × Sn . For a subset J = {j1 , . . . , jk } ⊂ {1, . . . , n} we denote by SJ = Sj1 × · · · × Sjk the sub-product, which can be viewed both as a subgroup and as a factor of the semi-simple group Gss 1 . Set S∅ = {1}. Given a subset J ⊂ {1, . . . , n} consider the image of Γ1 under the projection prJ : G1 −→ Gss 1 −→ SJ . LATTICE ENVELOPES 25 Note that it is possible that pr{i} (Γ1 ) is dense in Si for each i ∈ {1, . . . , n}, but prJ (Γ1 ) is discrete for some non-empty proper subset J ⊂ {1, . . . , n}. Lemma 5.4. There is a unique maximal subset J ⊂ {1, . . . , n} for which the projection prJ (Γ1 ) is discrete in SJ . Proof. It suffices to show that the collection of all subsets J ⊂ {1, . . . , n} with discrete projection to SJ , is closed under union. Let J, K ⊂ {1, . . . , n} be subsets such that prJ (Γ1 ) is discrete in SJ and prK (Γ1 ) is discrete in SK . Let V ⊂ SJ , W ⊂ SK be open neighborhoods of the identity with prJ (Γ1 ) ∩ V = {1}, prK (Γ1 ) ∩ W = {1}. View SJ∪K as a subgroup in SJ × SK . Then U = SJ∪K ∩ (V × W ) is an open neighborhood of the identity with prJ∪K (Γ1 ) ∩ U = {1}. Therefore prJ∪K (Γ1 ) is discrete in SJ∪K . This proves the claim. Let J ⊂ {1, . . . , n} be the maximal subset as in the claim. Denote Y Y Si . L = SJ = Sj , H = SJ c = j∈J We have Gss 1 i∈J / = L × H, and consider the projection prL : G1 = L × H × Gtd 1 → L, Γ0 = prL (Γ1 ). Define Γ2 to be the kernel of this projection Γ2 = Ker(prL : Γ1 −→ Γ0 ). Then Γ0 < L and Γ2 < H × Gtd 1 are lattices (Lemma 3.4). Consider the projections p : H × Gtd 1 −→ H, td q : H × Gtd 1 −→ G1 . Lemma 5.5. The projection p(Γ2 ) of Γ2 to H is dense. Proof. Let H ′ = p(Γ2 ) be the closure of the projection of Γ2 to H. Then H ′ × Gtd 1 ′ td is a closed subgroup of H × Gtd 1 containing the lattice Γ2 . By Lemma 3.7, H × G1 td ′ is a closed subgroup of finite covolume in H × G1 . It follows that H has finite covolume in H. By Lemma 3.9 H ′ has the form H ′ = ∆ × H2 , where H2 is a direct factor of H, H = H1 × H2 , and ∆ < H1 is a lattice. In our setting such a splitting has to be trivial: H ′ = H2 = H, H1 = ∆ = {1}. ss Indeed, otherwise the semi-simple group G1 splits as Gss 1 = (L × H1 ) × H2 , where the projection of Γ1 to the L × H1 -factor lies in Γ0 × ∆ which is discrete. This contradicts the maximality of L as such a factor. This completes the proof of the lemma. td We now define an lcsc group G2 to be G2 = Gss 2 × G2 , where Gss 2 = H, td Gtd 2 = q(Γ2 ) < G1 . td Being a closed subgroup of a tdlc group Gtd 1 , the group G2 itself is tdlc. The group ss G2 = H is a connected, center-free, semi-simple, real Lie group. td The group G2 is a closed subgroup of H × Gtd 1 containing a lattice Γ2 < H × G1 . td By Lemma 3.7, Γ2 forms a lattice in G2 , and G2 = H × G2 is a finite covolume td td subgroup of H × Gtd 1 . Thus G2 is a subgroup of finite covolume in G1 . Since 26 URI BADER, ALEX FURMAN, AND ROMAN SAUER td Radam (Gtd 1 ) = {1} we deduce that Radam (G2 ) = {1} using Lemma 3.11. We summarize: td Proposition 5.6. Let Γ1 < G1 = Gss 1 × G1 be as in Proposition 5.3. Then there ss td is a splitting G1 = L × H and a closed subgroup Gtd 2 < G1 such that by setting ss ss td G2 = H and G2 = G2 × G2 we have: (1) (2) (3) (4) (5) (6) (7) The projection Γ0 = prL (Γ1 ) of Γ1 is a lattice in L. prL td The kernel Γ2 = Ker(Γ1 −→ Γ0 ) is a lattice in G2 = Gss 2 × G2 . ss ss The projection pr2 (Γ2 ) is dense in G2 = H. td The projection prtd 2 (Γ2 ) is dense in G2 . td td G2 < G1 is a closed subgroup of finite covolume. Gtd 2 has trivial amenable radical. Gtd 2 is compactly generated. Most of the conditions needed for an application of Theorem 4.5 are satisfied with the exception of injectivity of the projection ss prss 2 : Γ2 −→ G2 = H. This is our next topic of concern. Step 3. Identifying the lattice embedding. We still denote the projections of G1 and G2 to their Lie and tdlc factors by ss prss i : Gi → Gi td and prtd for i ∈ {1, 2}. i : Gi → Gi td The images of a subgroup in Gi under prss i and pri will be indicated by superscripts ss and td, respectively. For a subgroup H < G let ZG (H) := {g ∈ G \| ∀h∈H gh = hg} denote the centralizer of H in G. td Let U < Gtd 1 be a compact open subgroup in the tdlc group G1 such that Γ1 ∩ ({1} × U ) = {1}. We define the following groups: Ni := Γi ∩ ({1} × Gtd i ) for i ∈ {1, 2}, M1 := Γ1 ∩ (Gss 1 × U ), (5.1) M2 := Γ2 ∩ M1 , Γ3 := Γ2 /N2 , M3 := p(M2 ) where p : Γ2 ։ Γ3 is the projection, ss Gss 3 := G2 , td td Gtd 3 := G2 /N2 , td G3 := Gss 3 × G3 . Remark 5.7. Since N2td is normalized by the dense image of Γ2 in Gtd 2 and the normalizer of a closed subgroup is closed, the subgroup N2td is normal in Gtd 2 . This justifies the last definition. Moreover, since Gss 1 commutes with N2 it follows that N2 is normal in G2 . Note that since compact generation passes to quotients, Gtd 3 is compactly generated. Remark 5.8. At this point, we do not claim triviality of the amenable radical of Gtd 3 , even though it will follow from later analysis. Lemma 5.9. The group N1 commutes with some finite index subgroup of M1 . LATTICE ENVELOPES 27 Proof. At this point we want to apply property (NbC) of the original group Γ. Note that property (NbC) does not pass to finite index subgroups in general, so we have to argue more specifically. We refer to the notation of Proposition 5.3. Note that Γ1 = (Γ ∩ G′ )/(Γ ∩ K) is a subgroup of Γ/(Γ ∩ K). We show that N1 is normal in Γ/(Γ ∩ K) and M1 is commensurated by Γ/(Γ ∩ K). From this the claim follows since Γ/(Γ ∩ K) has (NbC) by Lemma 2.5. ss The group {1} × Gtd 1 is a topologically characteristic subgroup of G1 = G1 × td ′ ′ td G1 = G /K by Lemma 3.16. Since G /K < G/K is normal, {1} × G1 is normal in G/K. Since Γ1 = (Γ ∩ G′ )/(Γ ∩ K) is a normal subgroup of Γ/(Γ ∩ K), the group ′ N1 is normal in Γ/(Γ ∩ K). The subgroup Gss 1 × U < G1 = G /K is commensurated ′ by G/K since G /K ⊳ G/K is normal and every (topological) automorphism of G1 ss is a product of an automorphism of Gtd 1 and one of G1 by Lemma 3.16. Since Γ1 ⊳ Γ/(Γ ∩ K) is normal, M1 is commensurated by Γ/(Γ ∩ K). By the previous lemma and upon making U smaller we may and will assume that M1 itself centralizes N1 and M1 ∩ N1 = {1}. It follows that M2 centralizes N2 . We record that (5.2) Mi ∩ Ni = {1} and [Mi , Ni ] = {1} for i ∈ {1, 2}. In particular, the subgroup Mi Ni < Γi is isomorphic to Mi × Ni . Lemma 5.10. The inclusions Miss ֒→ Gss i for i ∈ {1, 2, 3} and Γ3 ֒→ G3 are lattice embeddings. td Proof. By Lemma 3.3, M1 is a lattice in Gss 1 ×U since U < G1 is an open subgroup. ss ss Since U is also compact, the image M1 of M1 in G1 is a lattice of Gss 1 as well. ss ss ss ss Similarly, M2 < Gss 2 × U and M2 < G2 are lattices. Since G2 = G3 by definition ss ss one easily sees that the images of M2 and M3 in Gss 2 coincide, so M3 = M2 . In ss ss particular, M3 < G3 is a lattice. Since N2 ⊳ Γ2 is also normal in G2 (see the remark below (5.1)), the quotient Γ3 td td ∼ is a lattice in G3 = Gss 2 × G2 /N2 = G2 /N2 by Lemma 3.4. We are finally in a position to identify the lattice embedding Γ1 < G1 , which is virtually isomorphic to the original lattice embedding Γ < G. We distinguish three cases depending on the finiteness of the groups M1 and M2 . Case (i): M1 is finite (Γ1 is a lattice in a tdlc group). In this case, the connected real Lie group Gss 1 without compact factors has a finite td group M1ss as a lattice, thus Gss 1 must be trivial. Thus G1 = G1 is a tdlc group with trivial amenable radical that contains Γ1 as a lattice. If Γ is assumed to have property (BT), the same applies to Γ1 , and by Lemma 3.8, Γ1 < G1 is a uniform lattice. Either G1 is discrete which means that the original lattice embedding Γ ֒→ G was trivial or we are in case (3) of the main theorem. Case (ii): M1 is infinite, but M2 is finite (Γ1 is a lattice in a semi-simple Lie group). ss In this case M2ss is finite and a lattice in Gss 2 . As above we conclude that G2 = H is trivial. Therefore Γ2 = N2 , and td td td Gtd 2 = Γ2 = N 2 = N 2 . The last equality follows from the discreteness of N2td . 28 URI BADER, ALEX FURMAN, AND ROMAN SAUER ss Since M1ss = prss 1 (M1 ) is a lattice in G1 = L × H = L and a subgroup of the ss lattice Γ0 (see Proposition 5.6), M1 < Γ0 must have finite index. We have a short exact sequence prL Γ0 −→ 1. 1 −→ Γ2 −→ Γ1 −→ ss The prL -preimage of M1 which equals Γ2 · M1 = N2 M1 ∼ = N2 × M1 is a finite index subgroup of Γ1 . Since M1 is infinite, the (Irr) condition forces Γ2 = N2 to be td td finite. So Gtd 2 is finite. As G2 has finite covolume in G1 , the latter is compact. In td fact, G1 has to be trivial, because it has a trivial amenable radical. We conclude that Γ1 is a classical lattice in a connected, center-free, semi-simple, real Lie group L = Gss 1 = G1 . This lattice is irreducible due to assumption (Irr). This corresponds to case (1) in the main theorem. Case (iii): M2 is infinite (Γ1 is an S-arithmetic lattice). ss Recall that the projection M2ss of M2 to Gss 2 = G3 = H is a lattice there. The ss assumption that M2 is infinite, means that G3 is non-trivial. The inclusion td Γ3 < G3 = Gss 3 × G3 is a lattice embedding by Lemma 5.10. At this point we may and will apply the Arithmetic Core Theorem 4.5 to Γ3 < G3 to deduce that, up to dividing Gtd 3 by a compact normal subgroup C, one has a product of S-arithmetic lattices with nontrivial semi-simple and totally disconnected factors. (It will become clear below that C is trivial and that there is only one irreducible S-arithmetic lattice). More precisely, there are number fields K1 , . . . , Kn , absolutely simple Ki -algebraic groups Hi , and finite sets Si ⊂ Vi of places of Ki compatible with Hi such that, denoting ◦ the connected real Lie groups by Hi = (Hi )+ Ki ,S ∞ = (Hi )Ki ,S ∞ , we have i i td Gss 3 × G3 /C = n Y Hi × i=1 n Y Qi i=1 for certain closed intermediate groups (Hi )+ K fin i ,Si < Qi < (Hi )Ki ,Sifin . Moreover, Γ3 is commensurable to the product of the Si -arithmetic lattices: There are finite index subgroups Γ3,i < Hi (Oi [Si ]) such that (5.3) Γ3,1 × · · · × Γ3,n < Γ3 , [Γ3 : Γ3,1 × · · · × Γ3,n ] < ∞ has finite index. Upon passing to smaller finite index subgroup, we may assume that (5.3) is the inclusion of a normal subgroup. By Lemma 5.10, M3ss < Gss 3 = H1 × · · · × Hn is a lattice. It contains a product of irreducible lattices ∆i < Hi as a finite index subgroup: ∆1 × · · · × ∆n < M3ss , [M3ss , ∆1 × · · · × ∆n ] < ∞. Relying on the (NbC) condition we showed (see (5.2)) that M2 < Γ2 commutes with N2 , so (5.4) M2 < ZΓ2 (N2 ). The abelian subgroup ZΓ2 (N2 ) ∩ N2 = Z(N2 ) is, being a characteristic subgroup td td of N2td ⊳ Gtd 2 , normal in G2 . Since G2 has a trivial amenable radical, (5.5) ZΓ2 (N2 ) ∩ N2 = {1}. LATTICE ENVELOPES 29 Hence the restriction p\|ZΓ2 (N2 ) of the quotient map p : Γ2 → Γ3 = Γ2 /N2 is injective. The image p(ZΓ2 (N2 )) of the normal subgroup ZΓ2 (N2 ) is normal in Γ3 and contains M3 . The subgroup n Y Γ3,i T3 := p(ZΓ2 (N2 )) ∩ i=1 has finite index in p(ZΓ2 (N2 )) and is normal in Γ3,1 × · · · × Γ3,n . The commutative diagram below summarizes the various relations between the groups. Since ∆i and M3 ∩ T3 ∩ Γ3,i are commensurable, the normal subgroup T3 ∩ Γ3,i ⊳ Γ3,i is infinite. Since each Γ3,i is an irreducible Si -arithmetic lattice, Margulis’ normal subgroup theorem implies that each T3 ∩ Γ3,i < Γ3,i is of finite index, thus p(ZΓ2 (N2 )) < Γ3 is of finite index. By the injectivity of p\|ZΓ2 (N2 ) , the group ZΓ2 (N2 ) is thus (abstractly) commensurable with a product of S-arithmetic lattices. By Theorem 3.13, ZΓ2 (N2 ) has a finite outer automorphism group. Further, the center of ZΓ2 (N2 ) ⊳ Γ2 is trivial since it lies in the amenable radical of Γ2 which is trivial by Proposition 5.6 and Lemma 3.11. Qn M2 ∼ = ZΓ2 (N2 ) p ∼ = i=1 Hi f.i. p(ZΓ2 (N2 )) M3 Gss 3 × {1} Γ3 td Gss 3 × G3 ⊳ ⊳ Γ2 Qn ∆i i=1 p Γ2 /N2 ⊳ f.i. T3 Qn ⊳ i=1 Qn Γ3,i i=1 Hi × Q i f.i. Qn i=1 Hi (Oi [Si ]) Qn i=1 (Hi )Ki ,Si ∞ × (Hi )Ki ,Sifin Since N2 is the intersection of Γ2 ⊳ Γ1 and N1 ⊳ Γ1 the subgroup N2 is normal in Γ1 which implies that ZΓ2 (N2 ) is normal in Γ1 . By Lemma 3.14 the finite index subgroup Γ′1 = ker Γ1 → Out(ZΓ2 (N2 )) < Γ1 splits as a direct product: Γ′1 ∼ = ZΓ2 (N2 ) × (Γ′1 /ZΓ2 (N2 )). Assumption (Irr) implies that one of the factors is finite; and by M2 being infinite and (5.4), we obtain that ZΓ2 (N2 ) is of finite index in Γ1 , hence Γ2 < Γ1 is of finite index. So Γ0 is finite. Because of [Γ1 : ZΓ2 (N2 )] < ∞ and (5.5) we obtain that N1 and N2 are finite. The group L, being a lattice envelope of Γ0 , is compact, thus L and Γ0 are trivial. ss ss So Gss 1 = H = G2 = G3 . Using condition (Irr) we can now also deduce that there was only one irreducible factor, hence n = 1 in (5.3). Furthermore, the group N2 ∼ = N2td , that is now known to be finite, has to be td td trivial, because N2 is normal in Gtd 2 , while Radam (G2 ) = {1}. It follows that td G3 = G2 and Radam (G3 ) = {1}. In particular, the compact normal subgroup C ⊳ Gtd 3 is actually trivial. 30 URI BADER, ALEX FURMAN, AND ROMAN SAUER We deduce that there is a number field K, a connected, non-commutative, adjoint, absolutely simple K-groups H, and a finite set S = S ∞ ∪ S fin of places as in §4.1 so that Γ1 = Γ2 = Γ3 ≃ H(O[S])+ is an irreducible S-arithmetic lattice in the lcsc group + + td G2 = Gss 2 × G2 = HK,S ∞ × HK,S fin . It remains to identify the lcsc group G1 . We have established the triviality of Γ0 and of L, hence + ss ss Gss 1 = H × L = H = G2 = G3 = HK,S ∞ . It remains to identify the totally disconnected component Gtd 1 that contains Y + + td H(Kν ) G2 = HK,S fin = ν∈S fin as a closed subgroup of finite covolume. + td Step 4: Lattice envelope Gtd 1 of G2 = HK,S fin . Let us enumerate the elements in S fin = {ν1 , . . . , νn } in such a way that H has Kνi -rank at least two for 1 ≤ i ≤ k and has Kνi -rank 1 for k < i ≤ n. The extreme cases k = n or k = 0 are, of course, possible. Let Xi be the Bruhat-Tits building associated to H(Kνi ), and let Hi := H(Kνi )+ . We write H = H1 × · · · × Hn and X = X1 × · · · × Xn . With that notation, H = Gtd 2 . The spaces Xi are irreducible Euclidean buildings with cocompact affine Weyl group and Moufang Tits boundary. Each group Hi acts by automorphisms on the simplicial complex Xi , such that the action on vertex set has finitely many orbits and the vertex stabilizers are maximal compact subgroups of Hi . For i ∈ {k + 1, . . . , n} the building Xi is a tree. Theorem 5.11. Let G be a tdlc group that has trivial amenable radical and contains H = Gtd 2 as a closed subgroup of finite covolume. Then G/H is compact, and there is an open subgroup G′ < G of finite index that contains H such that the natural embedding H → Isom(X) extends to an injective continuous homomorphism G′ → Isom(X) with closed image. We apply this result to G = Gtd 1 whose amenable radical is trivial. According to Remark 4.4 und upon passing to a closed subgroup of G′ of finite index the group G′ is a tree extension of H in the sense of Definition 4.3. This finishes the proof of Theorem A based on Theorem 5.11. The proof of Theorem 5.11 proceeds in several steps. We heavily use techniques from geometric group theory, inspired by the approach taken in [33, Section 3]. Compactness of G/H. By the Bruhat-Tits fixed-point theorem (see [15, Corollary 2.8 on p. 179] for a more general result on CAT(0)-spaces) every compact subgroup of H fixes a vertex of X. But since there are only finitely many H-orbits of vertices there are only finitely many vertex stabilizer groups up to conjugation. In particular, H endowed with Haar measure m has an upper bound on the Haar measure of its open compact subgroups: sup m(U ) \| U is an open compact subgroup of H < +∞. LATTICE ENVELOPES 31 Thus Lemma 3.8 applies to H < G and implies compactness of G/H. Definition of the subgroup G′ < G. Applying Lemma 3.29 in combination with Remark 3.30 to H acting on X, one obtains a homomorphism of G to the quasiisometry group of X φ : G → QI(X). Further, there are constants C, L > 0 such that for every g ∈ G there is a (L, C)quasi-isometry fg : X → X that represents the class φ(g) ∈ QI(X) and has the following property: For every bounded set B ⊂ X there is a neighborhood of the identity V ⊂ G so that (5.6) ∀g∈V ∀x∈B dX (fg (x), x) < C. We now apply the following splitting theorem by Kleiner and Leeb. Theorem 5.12 ([42, Theorem 1.1.2]). For every C, L > 0 there are L′ , C ′ , D > 0 such that every (L, C)-quasi-isometry X → X is within distance D from a product of (L′ , C ′ )-quasi-isometries Xi → Xπ(i) between the factors for some permutation π ∈ Symn . At the expense of increasing the constants C and L, we may hence assume (i) that each fg is a product of (L, C)-quasi-isometries fg : Xi → Xπg (i) for some permutation πg ∈ Symn . Another consequence of Theorem 5.12 is that the product QI(X1 ) × · · · × QI(Xn ) embeds into QI(X) as a subgroup of index at most n! = \|Symn \|. We define the finite index subgroup G′ < G as the φ-preimage of QI(X1 ) × · · · × QI(Xn ). Note that G′ contains H. So there are homomorphisms φi : G′ −→ QI(Xi ) for i ∈ {1, . . . , n} such that φ\|G′ = φ1 × · · · × φn . Further, there are (L, C)(i) (i) quasi-isometries fg : Xi → Xi for every i ∈ {1, . . . , n} and g ∈ G′ such that fg represents φi (g). The restriction of φi to H agrees with the homomorphism pr i H −−→ Hi → Isom(Xi ) → QI(Xi ), where pri is the projection to the i-th factor. In view of (5.6) and the quantitative statement in Theorem 5.12 and at the expense of increasing C, L > 0 once more, the following holds true: For every i ∈ {1, . . . , n} and for every bounded set B ⊂ Xi there is a neighborhood of the identity V ⊂ G so that (5.7) ∀g∈V ∀x∈B dX (fg(i) (x), x) < C. Openness of G′ < G. The group G′ is defined as the kernel of a homomorphism of G to a finite group. But we do not know at this point that the homomorphism is continuous, which would imply that G′ < G is open, thus an lcsc group itself. Next we provide a direct argument that shows openness of G′ < G. Let dX and dXk denote the metric on X and Xk , respectively. Let Bi ⊂ Xi be a bounded subset whose diameter exceeds 3C, and let B = B1 × · · · × Bn . Let V ⊂ G be a neighborhood of the identity that satisfies (5.6). Next we show that V is contained in G′ and so G′ < G is an open subgroup. Suppose it is not. Then there is g ∈ V such that πg is non-trivial. Let i ∈ {1, . . . , n} be such that j := πg (i) 6= i. Let xj , x′j ∈ Bj be points whose distance is at least 3D. Pick (i) points xl ∈ Bl for l ∈ {1, . . . , j − 1, j + 1, . . . , n}. Then either dXj (fg (xi ), xj ) > C 32 URI BADER, ALEX FURMAN, AND ROMAN SAUER (i) or dXj (fg (xi ), x′j ) > C. Without loss of generality assume the first case. Let x = (x1 , . . . , xn ) ∈ B. Then dX (fg (x), x) ≥ dXj (fg(i) (xi ), xj ) > C, contradicting (5.6). Therefore V ⊂ G′ . Mapping G′ to Isom(Xi ) for the higher rank factors Xi . The rigidity theorem for higher rank irreducible buildings (such as X1 , . . . , Xk ) by Kleiner and Leeb [42] is the next important ingredient. Theorem 5.13 ([42, Theorem 1.1.3]). Let i ∈ {1, . . . , k}. For every C > 0 and L > 0 there is a constant D > 0 such that every (L, C)-quasi-isometry Xi → Xi is within distance D from a unique isometry Xi → Xi . Moreover, no two distinct isometries Xi → Xi for i ∈ {1, . . . , k} are within bounded distance from each other. Hence the natural homomorphism Isom(Xi ) → QI(Xi ) is an isomorphism for i ∈ {1, . . . , k} and composing its inverse with φi we obtain homomorphisms ψi : G′ −→ Isom(Xi ) pr i for which the restriction to H < G is the homomorphism H −−→ Hi → Isom(Xi ). (i) The statement (5.7) and Theorem 5.13 applied to fg show that there is a constant E > 0 such that for every bounded subset B ⊂ Xi there is a neighbourhood of the identity V ⊂ G such that (5.8) ∀g∈V ∀x∈B dXi (ψi (g)(x), x) < E. Lemma 5.14. The map ψi is continuous for i ∈ {1, . . . , k}. Proof. Let U ⊂ Isom(Xi ) be an open neighborhood of the identity. We have to show that there is open neighborhood of the identity of G that is contained in ψi−1 (U ). We rely on the following geometric fact about buildings that follows, for example, from [60]: For every constant D > 0 and any open neighborhood W of the identity in Isom(Xi ), there is a bounded set B ⊂ Xi , depending on D and W so that (5.9) θ ∈ Isom(Xi ) \| sup dXi (θ(x), x) < D ⊂ W. x∈B We apply this general fact to the constant E from (5.8) and the identity neighborhood U . Let us fix a bounded subset B = B(E, U ) ⊂ Xi such that (5.10) θ ∈ Isom(Xi ) \| sup dXi (θ(x), x) < E ⊂ U. x∈B Applying the statement (5.8) to this specific subset B provides us with a neighborhood V = V (E, B) ⊂ G of the identity such that (5.11) ∀g∈V ∀x∈B dXi (ψi (g)(x), x) < E. Since (5.10) and (5.11) mean that ψi−1 (U ) ⊂ V , continuity of ψi follows. LATTICE ENVELOPES 33 Mapping G to Homeo(∂Xj ) for the rank 1 factors Xj . Next we turn to the tree factors Xj with j ∈ {k + 1, . . . , n}. For every tree of bounded degree T , such as Xj with k < j ≤ n, the space of ends ∂T is compact, and one has an embedding of groups QI(T ) ֒→ Homeo(∂T ). For j ∈ {k + 1, . . . , n} let ψj be the composition φj ψj : G′ −→ QI(Xj ) ֒→ Homeo(∂Xj ). Lemma 5.15. The map ψj is continuous and has a closed image for every j ∈ {k + 1, . . . , n}. Proof. The continuity statement is proved in [33, Theorem 3.5]. The argument is analogous to the one for Lemma 5.15. However, instead of the geometric fact (5.9) one uses the following fact which is a consequence of the Mostow-Morse lemma: For a tree T of bounded degree and constants L, C, D > 0 and for any identity neighbourhood U ⊂ Homeo(∂T ) there is a compact subset B ⊂ T so that the image in Homeo(∂T ) of any (L, C)-quasi-isometry f : T → T with sup{dT (f (x), x) \| x ∈ B} < D belongs to U . Since H is cocompact in G′ and the image of H in Homeo(∂Xj ), which equals the image of Hj in Homeo(∂Xj ), is closed, also the image of G′ is closed. Identifying the image of ψj . Since ψi is continuous for i ∈ {1, . . . , n}, Ker(ψi ) is a closed subgroup of G. We set Gi := G′ / Ker(ψi ). Regarding Hi as a subgroup of H and thus of G′ , the inclusion Hi ֒→ G induces a monomorphism Hi ֒→ Gi since G′ restricts to an injective map Hi → Isom(Xi ) or Hi → Homeo(∂Xi ), respectively. By Theorem 3.28 Isom(Xi ) contains Hi as a closed subgroup of finite index for i ∈ {1, . . . , k}. Thus the ψi -image of G′ for i ∈ {1, . . . , k} contains Hi as a subgroup of finite index. Let us consider the ψj -image for the tree factors. Fix an index j ∈ {k +1, . . . , n}. The map φj describes a quasi-action of G′ , thus a faithful quasi-action of Gj , (j) on the tree Xj since every g ∈ G acts by a (L, C)-quasi-isometry fg such that (j) (j) (j) fg1 ◦ fg2 is within uniformly bounded distance from fg1 g2 . Next we resort to the work of Mosher, Sageev and Whyte on quasi-actions on trees. Their main result says: Theorem 5.16 ([51, Theorem 1]). A cobounded quasi-action of a group on a bushy tree of bounded degree is quasi-conjugated to an isometric action of the same group on a – possibly different – tree. Since Xj has degree at least 3 and Hj acts cocompactly on Xj the assumptions are fulfilled for the quasi-action of Gj on Xj . Hence there is a tree Tj on which Gj acts by isometries and a quasi-conjugacy from Xj to Tj , that is, a quasi-isometry η : Xj → Tj such that dTj η φj (g)(x) , gη(x) < ∞. sup g∈Gj ,x∈Xj 34 URI BADER, ALEX FURMAN, AND ROMAN SAUER Some preliminary remarks for the proof of Lemma 5.18 below are in order. We rely on the group-theoretic description of Bruhat-Tits buildings (see e.g. [1, Corollary 11.35 on p. 563 and Theorem 11.38 on p. 564]) which we only need in dimension 1: The Bruhat-Tits tree Xj as a 1-dimensional Hj -equivariant CW-complex consists of two Hj -orbits of vertices Hj /K1 and Hj /K2 , where K1 and K2 are nonconjugated maximal compact-open subgroups of Hj , and one Hj -orbit Hj /(K1 ∩K2 ) of edges. Further, K1 and K2 generate Hj . In this description we are using that Hj = H(Kνj )+ is the image of the Kνj -points of the simply connected form of H; the whole group H(Kνj ) might have only one orbit of vertices. An action of a group on a tree is called minimal if there are no proper subtrees invariant under the group action. If an action on a locally finite tree is minimal then the induced action on the boundary of the tree is minimal in the sense of dynamical systems: there are no proper closed invariant subsets. This is a consequence of the limit set being the minimal closed invariant subset of the boundary (see [26, Théorème 5.1] for a more general result, which is attributed to Gromov, in the context of Gromov-hyperbolic spaces). Conversely, if a group G is acting by automorphisms on a locally finite tree T , and L ⊂ ∂T is a minimal G-invariant set consisting of more than two points, then the union of all geodesics in T with end points in L, called the convex hull TL of L, is a minimal G-invariant sub-tree of T . Any G-invariant sub-tree T ′ has ∂T ′ = L, and therefore contains TL . Moreover there exists a retraction r : T ′ → TL given by nearest point projection to TL . This retraction is equivariant with respect to Aut(T ). Returning to our situation recall that the action of Hj on the tree Xj is minimal, hence the Hj -action on ∂Xj is minimal. Since ∂η is an equivariant homeomorphism, the action of Hj on ∂Tj is also minimal. In particular, the Gj -action on the convex hull of ∂Tj gives a minimal Gj -invariant tree. Without loss of generality we may assume Tj itself is the convex hull of its boundary ∂Tj , i.e. that Tj does not have vertices of degree one. The action of Gj on ∂Tj defines a proper action on the space ∂ (3) Tj of triples of distinct points of ∂Tj . Thus the Gj -action on ∂Tj is minimal and strongly proximal. The subgroup Hj < Gj has finite covolume. Thus the action of Hj on ∂Tj is also minimal and strongly proximal by Lemma 3.12. Minimality of the Hj -action on ∂Tj implies that Tj , which is the convex hull of its boundary, is minimal also as an Hj -invariant tree. Lemma 5.17. The action homomorphism Gj → Isom(Tj ) is injective, continuous and has a closed image. Proof. The map η induces a Gj -equivariant homeomorphism ∂η : ∂Xj → ∂Tj . Since both the isometry group and the quasi-isometry group of either Xj or Tj inject into the homeomorphism group of ∂Xj or ∂Tj , respectively, and φj is injective, the map Gj → Isom(Tj ) is injective. Since Isom(Tj ) embeds as a closed subgroup into Homeo(∂Tj ) continuity and closedness follow from the continuity and closedness of the composition Gj → Isom(Tj ) → Homeo(∂Tj ). But the latter is conjugated via ∂η to the map Gj → Homeo(∂Xj ) whose continuity and closedness is implied by Lemma 5.15. LATTICE ENVELOPES 35 Lemma 5.18. There is c ∈ N and an Hj -equivariant cellular homeomorphism f : Xj → Tj such that f is a homothety with stretch factor c. Proof. Upon subdividing we may assume that Hj acts on Tj without inversion. By the continuity part of Lemma 5.17 the orbits of the compact subgroups K1 and K2 in Tj are bounded. By the Bruhat-Tits fixed point theorem [15, Corollary 2.8 on p. 179], there is a vertex in Tj that is fixed under K1 or K2 , respectively. Choose a pair (v1 , v2 ) of vertices with Ki vi = vi that has minimal distance. The Hj -orbit of the geodesic segment c between v1 and v2 is connected, thus a Hj -invariant subtree of Tj , because Hj is generated by K1 and K2 . By Hj minimality of Tj the Hj -orbit of c is Tj . Moreover, c is fixed by K1 ∩ K2 . The map that sends the vertex 1 · Kl to vl for l ∈ {1, 2} and the edge e = 1 · (K1 ∩ K2 ) onto c by an affine homeomorphism can be extended to an Hj -equivariant cellular map f : Xj → Hj . Clearly, f is surjective. Next we show that f is locally injective. It is enough to show local injectivity at the vertices 1K1 and 1K2 of Xj . By symmetry let us consider 1K1 only. Consider its neighbours 1K2 and hK2 for some h ∈ Hj . Since K1 and K2 are not conjugated the f -images v1 , v2 , v2′ of 1K1 , 1K2 and hK2 are pairwise distinct. Let w be the center of the tripod given by v1 , v2 , v2′ . We want to show that w = v1 which implies local injectivity of f at 1K1 . Both K1 ∩ K2 and K1 ∩ hK2 h−1 are contained in the stabilizer S of w. Since K1 ∩ K2 is maximal in K1 and K1 ∩ hK2 h−1 6= K1 ∩ K2 by the group theoretic description of the Bruhat-Tits tree [1, Theorem 11.38 on p. 564] it follows that S = K1 . So w = v1 since we chose (v1 , v2 ) to have minimal distance among the vertices fixed by K1 or K2 , respectively. As a surjective, locally injective map between trees f is a homeomorphism. Let c be the distance between v1 and v2 . Obviously, f is locally a homothety with stretch factor c. Thus it is globally so. Conjugating with a homothety as in Lemma 5.18 we obtain a topological isomorphism Isom(Tj ) ∼ = Isom(Xj ) compatible with the two embeddings of Hj . Hence, Gj can be embedded as a closed intermediate subgroup Hj < Gj < Isom(Xj ), and the homomorphism ψj can be regarded as G′ → Gj ֒→ Isom(Xj ) composed with the natural embedding ′ Isom(Xj ) → Homeo(∂X Qn j ). ′Thus we also denote the map G → Gj → Isom(Xj ) ′ by ψj . Let ∆ : G → i=1 G be the diagonal embedding. We obtain a continuous homomorphism with closed image n Q ψi ◦∆ Y Isom(Xi ) < Isom(X) ψ : G′ −−−−−→ i=1 whose restriction to H is the natural embedding H ֒→ Isom(X). Since G/H is compact, ker(ψ) is compact. Since the amenable radical of G, thus of G′ by Lemma 3.11, is trivial, the map ψ is injective. This completes the proof of Theorem 5.11, and therefore also the proof of the main classification result – Theorem A. 6. Proofs of Theorems B–F Proof of Theorem B. Let Γ < H be a lattice embedding, where we assume that 36 URI BADER, ALEX FURMAN, AND ROMAN SAUER (i) either H is a connected, center-free, real Lie group without compact factors, H 6≃ SL2 (R) and Γ < H is an irreducible lattice, (ii) or H = HK,S as in Setup 4.1 and Γ is commensurable to the S-arithmetic lattice H(O[S]). Such Γ satisfies all the assumptions of Theorem A. Indeed, being a lattice Γ is Zariski dense in the semi-simple algebraic group H (Borel’s Density Theorem, cf. [47, Theorem II.(4.4)]). Thus Theorem 1.4.(2) ensures that Γ satisfies conditions (CAF) and (NbC). Being an irreducible lattice, Γ satisfies (Irr). Lattices in semisimple groups are known to be finitely generated (cf. [47, §IX.(3.2)]). By Selberg’s lemma Γ contains a torsion-free subgroup of finite index, in particular condition (BT) applies. Let G be some lcsc group and Γ < G be a lattice embedding. Theorem A states that there is an open subgroup of finite index G′ < G and a compact normal subgroup K ⊳ G′ so that, upon replacing G by G′ /K and Γ by Γ ∩ G′ (we assume Γ to be torsion-free), one has the following three possibilities: (1) G is a connected, center-free, semi-simple, real Lie group without compact factors and Γ < G is an irreducible lattice. (2) There is a number field K ′ , a finite set S ′ of places, including infinite and finite ones, a K ′ -group G so that Γ is a commensurable to an S ′ -arithmetic lattice G(O[S ′ ]) in G = GK ′ ,S ′ as in Setup 4.1. (3) G is a tdlc group with trivial amenable radical and Γ < G is a uniform lattice. Cases (1) and (2) are covered by the following Theorem 6.1 (Strong Rigidity Theorem, Mostow, Prasad, Margulis). Let i : Γ → G and j : Γ → H be embeddings of Γ as irreducible lattices into centerfree semi-simple Lie groups, or arithmetic lattices as in (1), (2) and (i), (ii) above. Then there exists a continuous group isomorphism π : G → H such that j = π ◦ i. Strong Rigidity Theorem was proved by Mostow [52] for all the cases where G and H are semi-simple real Lie Lie groups and i(Γ) < G and j(Γ) < H are uniform irreducible lattices. Prasad [57] extended this result to some non-uniform lattices, including all non-uniform lattices in rank one real Lie groups G and H. Higher rank cases, including the S-arithmetic situations, for both uniform and non-uniform lattice embeddings, follow from Margulis super-rigidity theorem, see [47, Theorem VII.(7.1)]. Therefore it remains to prove that if Γ < H is as in (i) or (ii), and Γ < G is as in case (3), then G is a discrete group containing Γ as a as subgroup of finite index. To this end we will use use the homomorphism ρ : G −→ QI(Γ) provided by Lemma 3.29. We shall prove below that Claim 6.2. The kernel U := ker ρ is an open subgroup in G. Let us first show that the claim suffices to complete the proof of the theorem. Let mG denote a Haar measure of G. Since U is open, we have mG (U ) > 0. Therefore the restriction of mG to U is a Haar measure mU on U . For a lattice Γ < H in a center-free semi-simple group, the homomorphism Γ → QI(Γ) is injective. Thus U ∩ Γ = {1} and one may include U in a Borel fundamental domain of Γ in G. It LATTICE ENVELOPES 37 follows that mG (U ) = mU (U ) is finite, and therefore U is a compact group. As G is assumed to have trivial amenable radical, we deduce that U = {1}. But since U is open in G, it follows that G is discrete and that its lattice Γ is just a subgroup of finite index. We are left with proving Claim 6.2, that will be deduced from quasi-isometric rigidity results about lattices in semi-simple Lie groups; we refer to Farb’s survey [28] for background and references. First assume that Γ is a non-uniform irreducible lattice in H. In this case QI(Γ) is known to coincide with the commensurator of Γ in H, which is a countable group: ρ : G −→ QI(Γ) = CommenH (Γ). This was first proved by Schwartz [62] for rank-one real Lie groups, and then by Schwartz [63], Farb-Schwarz [29], and Eskin [27] for higher rank real Lie groups, and by Wortman [73] for S-arithmetic cases. Hence the homomorphism ρ has countable image. The kernel U = ker ρ is clearly a Borel subgroup, and since countably many of its cosets cover G, one has mG (U ) > 0. A Borel subgroup U of positive Haar measure is open because of U = U −1 · U and U −1 · U contains an identity neighborhood. This proves Claim 6.2 in case Γ < H is non-uniform. Next assume Γ is a uniform lattice in H. Let H = H1 × · · · × Hn be the decomposition of H into simple factors and Xi = Hi /Ki be the associated symmetric spaces and/or Bruhat-Tits buildings. At least one of the factors Hi is a simple real Lie group, and we shall assume H1 is such. We have ρ : G −→ QI(Γ) = QI(X1 × · · · × Xn ). Arguing as in Step 4 in Section 5 (in particular Theorem 5.12), we observe that, upon replacing G by an open subgroup of finite index (and replacing Γ by a finite index subgroup), we may assume that ρ(G) takes value in QI(X1 ) × · · · × QI(Xn ) < QI(X1 × · · · × Xn ). Consider the homomorphism pr1 ◦ρ : G −→ QI(X1 ). We shall show that it factors through a continuous homomorphism G −→ Isom(X1 ) by considering the following cases. • X1 is a symmetric space of higher rank, corresponding to rkR (H1 ) ≥ 2. In this case the natural homomorphism Isom(X1 ) → QI(X1 ) is a bijection (Kleiner-Leeb [42]), and the group homomorphism ρ G −→ QI(X1 ) = Isom(X1 ) is continuous by Lemma 5.14. m 2 • X1 is a quaternionic hyperbolic space HH or the Cayley plane HO , corresponding to H1 ≃ Sp(m, 1) or F4 . Also in this case the natural homomorphism Isom(X1 ) → QI(X1 ) is a bijection (Pansu [56]), and the homomorphism ρ G −→ QI(X1 ) = Isom(X1 ) is continuous by Lemma 5.15. 38 URI BADER, ALEX FURMAN, AND ROMAN SAUER • X1 is real hyperbolic space Hm with m ≥ 3, corresponding to H1 = PO(m, 1)0 . The boundary at infinity ∂X1 is a sphere S m−1 with a natural conformal structure that is preserved by Isom(X1 ). In fact, the well known isomorphism Isom(Hm ) = Conf(S m−1 ) extends to an isomorphism ∼ QConf(S m−1 ) QI(Hm ) = between the group of quasi-isoemtries and the group of quasi-conformal homeomorphisms of the sphere. By a result of Tukia [70], a subgroup of QConf(S m−1 ) that is uniformly quasi-conformal and acts cocompactly on triples of points of S m−1 is conjugate into Conf(S m−1 ). This can be applied to ρ(G). Indeed, by Lemma 3.29 there exist constants C, L so that all all ρ(g) are represented by (L, C)-quasi-isometries, and therefore, by Mostow-Morse lemma, induce a uniformly quasi-conformal group of mappings of S m−1 . This group contains Γ, that acts cocompactly on X1 , and therefore also cocompactly on triples of points in S m−1 . Therefore we get a homomorphism G −→ Isom(Hm ) = Conf(S m−1 ) < QConf(S m−1 ) = QI(Hm ). m The topology on Isom(HR ) coincides with that of Conf(S m−1 ) inherited from the Polish topology Homeo(S m−1 ). Hence the continuity of the homomorphism G −→ Isom(Hm ) follows from Lemma 5.15. m • X1 is a complex hyperbolic space HC , corresponding to PU(m, 1)0 . In m this case the boundary sphere ∂HC has a natural conformal class of a subRiemannian Carnot-Caratheodory structure. Carnot-Caratheodory analogues of the theory of quasi-conformal mappings and a result analogous to Tukia’s theorem allow to prove (see [25]) that ρ(G) is quasi-conformally conjugate into m m Conf(∂HC ) = Isom(HC ). m The continuity of the resulting homomorphism G −→ Isom(HC ) follows from Lemma 5.15. This shows that the homomorphism ρ : Γ → QI(Γ) factors through a continuous homomorphism into a simple real Lie group ρ1 : G −→ Isom(X1 ) ≃ H1 . Let V be an open compact subgroup of a tdlc group G. Its image ρ1 (V ) is a profinite group subgroup in Isom(X1 ), which is a real Lie group. Therefore ρ1 (V ) must be finite, and V0 := V ∩ ker(ρ1 ) be an open subgroup. Since U = ker ρ contains an open subgroup V0 , U is also an open subgroup of G. This completes the proof of Claim 6.2, and thereby the proof of Theorem B. Proof of Theorem C. Let Γ be a finite extension of a finitely generated non-abelian free group, and let Γ < G be a lattice envelope. Since Γ satisfies all the assumptions of Theorem A by Theorem 1.4, we only need to discuss three possibilities. (1) Γ < G is virtually isomorphic to an irreducible lattice in a center-free semisimple real Lie group H. The first ℓ2 -Betti number of Γ is positive. By Olbrich’s work [53] (see also Chapter 5 about computations of ℓ2 -invariants LATTICE ENVELOPES 39 of locally symmetric spaces in Lück’s book [44]) the only such lattices with positive first ℓ2 -Betti numbers are the ones in PSL2 (R). In this case, Γ is a non-uniform lattice, since, for example, Γ has virtual cohomological dimension one. (2) Γ < G is virtually isomorphic to an S-arithmetic lattices as in Setup 4.1. In particular, it is virtually isomorphic to a lattice embedding into a product of two non-compact lcsc groups. This can be ruled by the non-vanishing of the first ℓ2 -Betti number of Γ (see the argument in the proof of Theorem F). Alternatively, this situation can ruled out by the fact that S-arithmetic lattices as in §4 are not Gromov hyperbolic. (3) Γ < G is virtually isomorphic to a uniform lattice Λ < H in a tdlc group H with trivial amenable radical. In particular, Λ is virtually isomorphic to a free non-abelian group, and by passing to a finite index subgroup we may assume that F := Λ is a non-abelian free group. Let TF be the Cayleygraph of F which is a regular tree of degree at least 3. The natural action homomorphism F → Isom(TF ) → QI(TF ) of F extends by Lemma 3.29 to a homomorphism H → QI(TF ) which describes a cobounded quasi-action of H on TF (see the remark preceding Theorem 5.16). Next we appeal to the work of Mosher, Sageev and Whyte [51] as in Step 4 in the proof of Theorem A. By Theorem 5.16, which is taken from their work, the quasi-action of H on TF is quasi-conjugated to an isometric (automatically also cobounded, i.e. cocompact) action φ : H → Isom(T ) on a possibly different tree T . Using the arguments for Lemma 5.15 and Lemma 5.17 verbatim, one concludes that φ is injective, continuous and has closed image. By the open mapping theorem [12, IX. §5.3], H is topologically isomorphic to im(φ), and the latter is a cocompact subgroup of Isom(T ) by the cocompactness of the action. Proof of Theorem D. Let Γ be a uniform lattice in PSL2 (R), and let Γ < G be another lattice embedding. As a Gromov hyperbolic group, Γ satisfies all the assumptions of Theorem A according to Theorem 1.4. So we only need to discuss three possibilities. (1) Γ < G is virtually isomorphic to an irreducible lattice is a semisimple real Lie group H. As in the proof above, by positivity of the first ℓ2 -Betti number of Γ one is reduced to the case of H ≃ PSL2 (R). In this case, Γ has to be cocompact, but the two embeddings of Γ → PSL2 (R) need not be conjugate. (2) Γ < G is virtually isomorphic to an S-arithmetic lattices as in Setup 4.1. This case is ruled out for the same reason as case (2) in the proof of Theorem C. (3) Γ < G is virtually isomorphic to a uniform lattice in a tdlc group. This is covered by [33, Theorem C]: the only possibility is a trivial lattice embedding. Proof of Theorem E. Let M be a closed Riemannian manifold of dimension at least 5 with sectional 40 URI BADER, ALEX FURMAN, AND ROMAN SAUER curvatures ranging in 1 2 , −1 . n−1 Let Γ = π1 (M ) be the fundamental group, and let Γ < G be a lattice envelope of Γ. Since Γ is finitely generated, G is compactly generated by Lemma 3.6. The group Γ is Gromov hyperbolic, so it has the properties (CAF), (NbC), and (Irr) according to Theorem 1.4. As a fundamental group of a closed aspherical manifold, Γ is torsionfree. In particular, Γ also has property (BT). By Theorem A there are the following three possibilities for the lattice embedding Γ < G up to virtual isomorphism. Since Γ is torsionfree, we may assume upon replacing Γ by a finite index subgroup, G by a finite index subgroup and a quotient by a compact normal subgroup, and M by a finite cover that we still have Γ = π1 (M ) and Γ < G is isomorphic (not just virtually) to one of the following cases. (1) Γ < G is an (irreducible) lattice embedding into a semisimple real Lie group G. (2) Γ < G is an S-arithmetic lattice in the sense of Setup 4.1. (3) Γ < G is a uniform lattice and G is a tdlc group with trivial amenable radical. We need to show that the only possibility is case (3) with G being discrete unless M is homeomorphic to a hyperbolic manifold. Since Γ is Gromov hyperbolic we can rule out case (2). By the same reason, case (1) is only possible if G has real rank 1 and Γ < G is uniform; note here that non-uniform lattices contain a free abelian subgroup of rank 2. Let us analyze the situation where Γ < G is a uniform lattice in a simple Lie group of real rank 1. Let X = G/K be the associated symmetric space which is thus real, complex, quaternionic or Cayley hyperbolic. As aspherical spaces with the same fundamental group, the locally symmetric space Γ\X and M are homotopy equivalent. If X is real hyperbolic, then M is homeomorphic to the closed hyperbolic manifold Γ\X by the following striking result of Farrell and Jones [32] in their work on the Borel conjecture. So this possibility is ruled out by assumption. − 1+ Theorem 6.3 (Topological rigidity [32]). Let Y be a closed non-positively curved manifold of dimension 6= 3, 4. If a closed manifold Z is homotopy equivalent to Y , then Y and Z are homeomorphic. The possibilities that X is complex, quaternionic or Cayley hyperbolic are ruled out by applying the following result by Mok, Siu, and Yeung [49] to Y = Γ\X and Z = M and the pinching assumption for M . Theorem 6.4 (Geometric rigidity [49, Theorem 1]). Let Y and Z be homotopy equivalent closed Riemannian manifolds. Assume that Z is negatively curved and Y is complex, quaternionic or Cayley hyperbolic. Then Y and Z are isometric up to scaling. In particular, the sectional curvatures of Z cannot range in [−a, −1] for a < 4. Summing up, case (2) never occurs, and case (1) can only occur if M is homeomorphic to hyperbolic manifold. Finally, let us consider case (3) in which we have to show that the tdlc group G is discrete upon dividing out a compact normal subgroup. Consider the natural homomorphism φ : Γ → Isom(M̃ ) → QI(M̃ ) → Homeo(∂ M̃ ) = Homeo(S n−1 ). LATTICE ENVELOPES 41 Of course, φ coincides with the natural action Γ → Homeo(∂Γ) of Γ on its Gromov boundary. We can refer to Lemma 3.29 or, in this situation, to [33, Theorem 3.5] to conclude that the map φ extends to a homomorphism φ̄ : G → Homeo(S n−1 ). In the latter reference it is also shown that φ̄ is continuous, the kernel K := ker(φ̄) is compact, and the image of φ̄ is locally compact in the subspace topology. By the open mapping theorem [12, IX. §5.3] G is topologically isomorphic to im(φ̄). Since G has trivial amenable radical, K is trivial. Since the tdlc group G acts continuously and faithfully on the sphere S n−1 , a positive solution of the Hilbert-Smith conjecture would imply that G is discrete and thus finish the proof. The general Hilbert-Smith conjecture remains open but in our situation we can appeal to the work of Mj [48] that contains the following result. Theorem 6.5 (Hilbert-Smith conjecture for boundary actions [48, Corollary 2.2.5]). Let Λ be a Gromov hyperbolic Poincaré duality group. Let Q < Homeo(∂Λ) be a subgroup that is finite dimensional and locally compact in the subspace topology and lies in the image of the natural homomorphism QI(Λ) → Homeo(∂Λ). Let h be the Hausdorff dimension of ∂Λ with respect to the visual metric, and let t be the topological dimension of ∂Λ. We assume that h < t + 2. Then Q is a Lie group. We apply this theorem to Q = im(φ̄) ∼ = G and Λ = Γ. The topological dimension of ∂Γ = S n−1 is t = n − 1. It remains to verify that we have h < t + 2 = n + 1. The Hausdorff dimension h equals the volume entropy hvol of M̃ (see [74, Theorem C]), that is, 1 h = hvol := lim sup log vol BM̃ (x, R) . R→∞ R As a direct consequence of the Bishop-Gunther comparison theorem [36, Theorem 3.101 on p. 169] we obtain that if the n-dimensional manifold M̃ has sectional curvature in [−b2 , −a2 ], then (n − 1)a ≤ hvol ≤ (n − 1)b. In the present situation, b = 1 + 1/(n − 1), so we obtain that 1 h = hvol ≤ (n − 1)(1 + ) = n < n + 1 = t + 2. n−1 So the assumptions in Theorem 6.5 are satisfied and we conclude that G is at the same time a Lie group and a tdlc group, thus discrete. Proof of Theorem F. (2) Let Γ be group with positive first ℓ2 -Betti number β1 (Γ) > 0 and property (BT). Assume that Γ admits a non-uniform compactly generated lattice envelope G. Since, by Theorem 1.4, Γ satisfies all the conditions of Theorem A and (BT), Γ < G is virtually isomorphic to some lattice embedding Γ′ < G′ that is one of the first two types in Theorem A; the totally disconnected case is ruled out by (BT). Non-vanishing of the first ℓ2 -Betti number is preserved under virtual isomorphism (this easily deduced from their basic properties), even under quasi(2) isometry [61, Theorem 1]. So β1 (Γ′ ) > 0. Hence Γ′ cannot be a lattice in a 42 URI BADER, ALEX FURMAN, AND ROMAN SAUER (2) product of two non-compact lcsc groups. That is because β1 (Γ′ ) > 0 implies that the first ℓ2 -Betti number of any lattice envelope of Γ′ is positive [43, Theorem B], and the first ℓ2 -Betti number of a product of two non-compact unimodular lcsc groups is zero [55, Theorem 6.5 on p. 61]. This excludes that Γ′ < G′ is an Sarithmetic lattice as in Setup 4.1, so type (2) is ruled out. Hence Γ′ < G′ is a lattice in a connected center-free semisimple Lie group without compact factors. By Olbrich’s work [53] (see also Chapter 5 about computations of ℓ2 -invariants of locally symmetric spaces in Lück’s book [44]) the only such lattices with positive first ℓ2 -Betti numbers are the ones in PSL2 (R). Since Γ′ < G′ is also non-uniform, the group Γ′ , hence Γ, is virtually isomorphic to a free group. References [1] P. Abramenko and K. S. Brown, Buildings, Graduate Texts in Mathematics, vol. 248, Springer, New York, 2008. Theory and applications. [2] U. Bader, P.-E. Caprace, T. Gelander, and S. Mozes, Lattices in amenable groups. ArXiv preprint: 1612.06220. [3] U. Bader, P.-E. Caprace, and J. Lecureux, On the linearity of lattices in affine buildings and ergodicity of the singular Cartan flow. ArXiv preprint: 1608.06265. [4] U. Bader, A. Furman, T. Gelander, and N. Monod, Property (T) and rigidity for actions on Banach spaces, Acta Math. 198 (2007), no. 1, 57–105. [5] U. Bader, A. Furman, and R. Sauer, On the structure and arithmeticity of lattice envelopes, C. R. Math. Acad. Sci. Paris 353 (2015), no. 5, 409–413. , An adelic arithmeticity theorem for lattices in products. ArXiv preprint: 1705.08158. [6] , Weak Notions of Normality and Vanishing up to Rank in ℓ2 -Cohomology, Int. Math. [7] Res. Not. IMRN 12 (2014), 31773189. [8] M. B. Bekka, On uniqueness of invariant means, Proc. Amer. Math. Soc. 126 (1998), no. 2, 507–514. [9] M. Bestvina and K. Fujiwara, Bounded cohomology of subgroups of mapping class groups, Geom. Topol. 6 (2002), 69–89. [10] A. Borel, Density and maximality of arithmetic subgroups, J. Reine Angew. Math. 224 (1966), 78–89. [11] A. Borel and J. Tits, Homomorphismes “abstraits” de groupes algébriques simples, Ann. of Math. (2) 97 (1973), 499–571 (French). [12] N. Bourbaki, General topology. Chapters 5–10, Elements of Mathematics (Berlin), SpringerVerlag, Berlin, 1998. Translated from the French; Reprint of the 1989 English translation. [13] E. Breuillard and T. Gelander, A topological Tits alternative, Ann. of Math. (2) 166 (2007), no. 2, 427–474. [14] E. Breuillard, M. Kalantar, M. Kennedy, and N. Ozawa, C ∗ -simplicity and the unique trace property for discrete groups (2016). ArXiv preprint: 1410.2518. [15] M. R. Bridson and A. Haefliger, Metric spaces of non-positive curvature, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 319, Springer-Verlag, Berlin, 1999. [16] K. S. Brown, Cohomology of groups, Graduate Texts in Mathematics, vol. 87, Springer-Verlag, New York, 1994. Corrected reprint of the 1982 original. [17] M. Burger and N. Monod, Continuous bounded cohomology and applications to rigidity theory, Geom. Funct. Anal. 12 (2002), no. 2, 219–280. [18] M. Burger and S. Mozes, Groups acting on trees: from local to global structure, Inst. Hautes Études Sci. Publ. Math. 92 (2000), 113–150 (2001). [19] P.-E. Caprace, Non-discrete, simple locally compact groups (2016). to appear in the Proceedings of the 7th European Congress of Mathematics. [20] P.-E. Caprace, P. Kropholler, C. D. Reid, and P. Wesolek, On the residual and profinite closures of commensurated subgroups (2017). ArXiv preprint: 1706.06853. [21] P.-E. Caprace and N. Monod, A lattice in more than two Kac-Moody groups is arithmetic, Israel J. Math. 190 (2012), 413–444. LATTICE ENVELOPES [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] 43 , Isometry groups of non-positively curved spaces: structure theory, J. Topol. 2 (2009), no. 4, 661–700. , Isometry groups of non-positively curved spaces: discrete subgroups, J. Topol. 2 (2009), no. 4, 701–746. J. Cheeger and M. Gromov, L2 -cohomology and group cohomology, Topology 25 (1986), no. 2, 189–215. R. Chow, Groups quasi-isometric to complex hyperbolic space, Trans. Amer. Math. Soc. 348 (1996), no. 5, 1757–1769. M. Coornaert, Mesures de Patterson-Sullivan sur le bord d’un espace hyperbolique au sens de Gromov, Pacific J. Math. 159 (1993), no. 2, 241–270 (French, with French summary). A. Eskin, Quasi-isometric rigidity of nonuniform lattices in higher rank symmetric spaces, J. Amer. Math. Soc. 11 (1998), no. 2, 321–361. B. Farb, The quasi-isometry classification of lattices in semisimple Lie groups, Math. Res. Lett. 4 (1997), no. 5, 705–717. B. Farb and R. Schwartz, The large-scale geometry of Hilbert modular groups, J. Differential Geom. 44 (1996), no. 3, 435–478. F. T. Farrell and L. E. Jones, Negatively curved manifolds with exotic smooth structures, J. Amer. Math. Soc. 2 (1989), no. 4, 899–908. F. T. Farrell, L. E. Jones, and P. Ontaneda, Negative curvature and exotic topology, Surveys in differential geometry. Vol. XI, Surv. Differ. Geom., vol. 11, Int. Press, Somerville, MA, 2007, pp. 329–347. F. T. Farrell and L. E. Jones, Topological rigidity for compact non-positively curved manifolds, Differential geometry: Riemannian geometry (Los Angeles, CA, 1990), Proc. Sympos. Pure Math., vol. 54, Amer. Math. Soc., Providence, RI, 1993, pp. 229–274. A. Furman, Mostow-Margulis rigidity with locally compact targets, Geom. Funct. Anal. 11 (2001), no. 1, 30–59. , On minimal strongly proximal actions of locally compact groups, Israel J. Math. 136 (2003), 173 – 187. H. Furstenberg, Rigidity and cocycles for ergodic actions of semisimple Lie groups (after G. A. Margulis and R. Zimmer), Bourbaki Seminar, Vol. 1979/80, Lecture Notes in Math., vol. 842, Springer, Berlin, 1981, pp. 273–292. S. Gallot, D. Hulin, and J. Lafontaine, Riemannian geometry, 3rd ed., Universitext, SpringerVerlag, Berlin, 2004. T. Gelander, A. Karlsson, and G. A. Margulis, Superrigidity, generalized harmonic maps and uniformly convex spaces, Geom. Funct. Anal. 17 (2008), no. 5, 1524–1550. M. Gromov and W. Thurston, Pinching constants for hyperbolic manifolds, Invent. Math. 89 (1987), no. 1, 1–12. U. Hamenstädt, Bounded cohomology and isometry groups of hyperbolic spaces, J. Eur. Math. Soc. (JEMS) 10 (2008), no. 2, 315–349. T. Karube, On the local cross-sections in locally compact groups, J. Math. Soc. Japan 10 (1958), 343–347. D. Y. Kleinbock and G. A. Margulis, Bounded orbits of nonquasiunipotent flows on homogeneous spaces, Sinaı̆’s Moscow Seminar on Dynamical Systems, Amer. Math. Soc. Transl. Ser. 2, vol. 171, Amer. Math. Soc., Providence, RI, 1996, pp. 141–172. B. Kleiner and B. Leeb, Rigidity of quasi-isometries for symmetric spaces and Euclidean buildings, Inst. Hautes Études Sci. Publ. Math. 86 (1997), 115–197 (1998). D. Kyed, H. D. Petersen, and S. Vaes, L2 -Betti numbers of locally compact groups and their cross section equivalence relations, Trans. Amer. Math. Soc. 367 (2015), no. 7, 4917–4956. W. Lück, L2 -invariants: theory and applications to geometry and K-theory, Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics], vol. 44, Springer-Verlag, Berlin, 2002. G. A. Margulis, Non-uniform lattices in semisimple algebraic groups, Lie groups and their representations (Proc. Summer School on Group Representations of the Bolyai János Math. Soc., Budapest, 1971), Halsted, New York, 1975, pp. 371–553. , Finiteness of quotient groups of discrete subgroups, Funktsional. Anal. i Prilozhen. 13 (1979), no. 3, 28–39 (Russian). 44 [47] [48] [49] [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] [62] [63] [64] [65] [66] [67] [68] [69] [70] [71] [72] [73] [74] URI BADER, ALEX FURMAN, AND ROMAN SAUER , Discrete subgroups of semisimple Lie groups, Ergebnisse der Mathematik und ihrer Grenzgebiete (3) [Results in Mathematics and Related Areas (3)], vol. 17, Springer-Verlag, Berlin, 1991. M. Mj, Pattern rigidity and the Hilbert-Smith conjecture, Geom. Topol. 16 (2012), no. 2, 1205–1246. N. Mok, Y. T. Siu, and S.-K. Yeung, Geometric superrigidity, Invent. Math. 113 (1993), no. 1, 57–83. N. Monod, Superrigidity for irreducible lattices and geometric splitting, J. Amer. Math. Soc. 19 (2006), no. 4, 781–814. L. Mosher, M. Sageev, and K. Whyte, Quasi-actions on trees. I. Bounded valence, Ann. of Math. (2) 158 (2003), no. 1, 115–164. G. D. Mostow, Strong rigidity of locally symmetric spaces, Princeton University Press, Princeton, N.J., 1973. Annals of Mathematics Studies, No. 78. M. Olbrich, L2 -invariants of locally symmetric spaces, Doc. Math. 7 (2002), 219–237. D. Osin, Acylindrically hyperbolic groups, Trans. Amer. Math. Soc. 368 (2016), no. 2, 851– 888. H. D. Petersen, L2 -Betti numbers of locally compact groups, Ph.D. thesis, Department of Mathematical Sciences, Faculty of Science, University of Copenhagen, 2012. 139 p. P. Pansu, Métriques de Carnot-Carathéodory et quasiisométries des espaces symétriques de rang un, Ann. of Math. (2) 129 (1989), no. 1, 1–60 (French, with English summary). G. Prasad, Strong rigidity of Q-rank 1 lattices, Invent. Math. 21 (1973), 255–286. , Elementary proof of a theorem of Bruhat-Tits-Rousseau and of a theorem of Tits, Bull. Soc. Math. France 110 (1982), no. 2, 197–202 (English, with French summary). M. S. Raghunathan, Discrete subgroups of Lie groups, Springer-Verlag, New York-Heidelberg, 1972. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 68. C. Ramos-Cuevas, On the displacement function of isometries of Euclidean buildings, Indag. Math. (N.S.) 26 (2015), no. 2, 355–362. R. Sauer and M. Schrödl, Vanishing of ℓ2 -Betti numbers of locally compact groups as an invariant of coarse equivalence (2017). ArXiv preprint:1702.01685. R. E. Schwartz, The quasi-isometry classification of rank one lattices, Inst. Hautes Études Sci. Publ. Math. 82 (1995), 133–168 (1996). , Quasi-isometric rigidity and Diophantine approximation, Acta Math. 177 (1996), no. 1, 75–112. Y. Shalom, Rigidity of commensurators and irreducible lattices, Invent. Math. 141 (2000), no. 1, 1–54. , Rigidity, unitary representations of semisimple groups, and fundamental groups of manifolds with rank one transformation group, Ann. of Math. (2) 152 (2000), no. 1, 113–182. F. K. Schmidt, Mehrfach perfekte Körper, Math. Ann. 108 (1933), no. 1, 1–25 (German). B. Sun, Non-elementary convergence groups are acylindrically hyperbolic (2016). ArXiv preprint: 1612.00134. A. Thom, Low degree bounded cohomology and L2 -invariants for negatively curved groups, Groups Geom. Dyn. 3 (2009), no. 2, 343–358. J. Tits, Buildings of spherical type and finite BN-pairs, Lecture Notes in Mathematics, Vol. 386, Springer-Verlag, Berlin-New York, 1974. P. Tukia, On quasiconformal groups, J. Analyse Math. 46 (1986), 318–346. J. Tits, Sur le groupe des automorphismes d’un arbre, Essays on topology and related topics (Mémoires dédiés à Georges de Rham), Springer, 1970. C. Wells, Automorphisms of group extensions, Trans. Amer. Math. Soc. 155 (1971), 189–194. K. Wortman, Quasi-isometric rigidity of higher rank S-arithmetic lattices, Geom. Topol. 11 (2007), 995–1048. C. Yue, The ergodic theory of discrete isometry groups on manifolds of variable negative curvature, Trans. Amer. Math. Soc. 348 (1996), no. 12, 4965–5005. LATTICE ENVELOPES Weizmann Institute, Rehovot E-mail address: uri.bader@gmail.com University of Illinois at Chicago, Chicago E-mail address: furman@uic.edu Karlsruhe Institute of Technology E-mail address: roman.sauer@kit.edu 45	4math.GR
NEURAL NETWORK APPROACH TO RAILWAY STAND LATERAL SKEW CONTROL Peter Mark Beneš1, Matouš Cejnek1, Jan Kalivoda2 and Ivo Bukovsky1 1 Department of Instrumentation and Control Engineering, Czech Technical University, Prague, Czech Republic {PeterMark.Benes;Matous.Cejnek;Ivo.Bukovsky}@fs.cvut.cz 2 Department of Automobiles, Internal Combustion Engines and Railway Vehicles, Czech Technical University, Prague, Czech Republic Jan.Kalivoda@fs.cvut.cz ABSTRACT The paper presents a study of an adaptive approach to lateral skew control for an experimental railway stand. The preliminary experiments with the real experimental railway stand and simulations with its 3-D mechanical model, indicates difficulties of model-based control of the device. Thus, use of neural networks for identification and control of lateral skew shall be investigated. This paper focuses on realdata based modelling of the railway stand by various neural network models, i.e; linear neural unit and quadratic neural unit architectures. Furthermore, training methods of these neural architectures as such, real-time-recurrent-learning and a variation of back-propagation-through-time are examined, accompanied by a discussion of the produced experimental results. KEYWORDS Roller Rig, Linear Neural Unit, Quadratic Neural Unit, Real-Time-Recurrent-Learning (RTRL), Back-Propagation-Through-Time (BPTT) 1. INTRODUCTION An ongoing problem, which is currently under research in the railway industry, is that of lateral skew control of railway carriage wheel sets with Independently Rotating Wheels (IRW). In particular, the control of position of the leading wheel set of a railway carriage bogie, such to achieve central positioning of the wheel set, with respect to the span of the rail track. Or furthermore, to control the lateral deviation of the wheel set to follow a given desired set point. The primary necessity for such control arises from the need to improve behaviour of IRW wheel sets, as currently under study at CTU, on an experimental railway stand (roller rig). Particularly, for minimisation of wheel flange and rail head wearing as well as lateral forces and, furthermore, optimal stability at higher speeds of the wheel set. To date, various methods for control of the wheel set lateral position of IRW wheel sets have been under investigation, with both mechanical and electrical means of control [2], each featuring their own, individual drawbacks. At CTU, the latest study features active control of the roller rig through manipulation of the yaw torque of the rig wheel set [1], via a state feedback and cascade PID control for a linearization of the model of the CTU roller rig. However, the results from this initial experimentation shows, that such method is not suitable for real time control. Thus, this paper aims to investigate the possible use of a neural network approach for lateral control of such railway wheel sets. The suitability of application to this problem is motivated by promising theoretical studies of higherorder neural units (HONUs), especially the quadratic neural unit for engineering problems for [3]-[7]. These studies are focused on the use of supervised learning based approaches for polynomial structured neural units, also known as a class of HONUs, for adaptive identification and control of real engineering systems. Further motivation arises from the successful implementation of a quadratic neural unit controller (Neuro-controller), for control of a bathyscaphe system located in the automatic control laboratories of CTU [12]. Where, here, such controller adhered more closely to the desired behaviour of the system than the conventionally used PID controller. An extension on this result can be recalled in the work [11]. Where, further study was made via introduction of a new software for adaptive identification and control, along with further testing on both a theoretical system and the previously mentioned bathyscaphe system. Given this, in this paper we aim to investigate use of neural network (NN) approaches in the following manner. We will begin by explaining in more depth the problem behind the recently employed state feedback and cascade PID control of the linearization of the CTU roller rig. The proceeding section will then describe the principles and control schemes behind the various experimented methods, for adaptive identification and control. Following this, an experimental analysis of the various approaches of NN, focusing firstly on adaptive identification of the CTU roller rig system and then to test the various NN methods for control. The final component of this paper will then be, to analyse and compare the produced experimental results, with a conclusion to be drawn at the end. 2. PROBLEM DESCRIPTION This section aims to describe in more depth, the functionality behind the previously introduced experimental railway stand (CTU roller rig). More importantly, for the scope of this paper, we will elaborate on the necessity for control of such roller rig and the issues of conventional linear PID control or state feedback control, for controlling the lateral skew. 2.1 Experimental Setup of the Roller Rig at CTU The below figure (Figure 1) depicts the experimental setup of the CTU roller rig, along with the 3-D model for design and simulation purposes. This rig features 5 motor drives, two pairs of 0.5m diameter rollers are independently driven, via the largest of the set of drives (FM1 & 2). To simulate both straight and curved tracks, similar to those present on real railway networks. A central servo motor (FM4) was introduced, to yaw the lower roller pair for replicating curved track motion. This setup however, assumes simulation of a rail pair without effects of rail buckling. For manipulation of the wheel set yaw, a separate servo motor, central to the wheel set (FM3) is installed. For the scope of this paper, this is the most crucial component, as it is the action of this servo motor that is used for control actuation for the lateral skew of the wheel sets. Where, the discussed control setups in this paper will analyse control via manipulation of this servo motor torque. Finally, a fifth drive FM5 is located between the front roller set, as a control differential. Figure 1: CTU Roller Rig (real left, PTC Pro/Engineer 3-D model on the right) 2.2 Problem of the State Feedback and Cascade PID Control In the work [1], the method of lateral skew control via state feedback with cascade PID setup is introduced. In this presented setup, a linearization of the roller rig is used, via translation to a state space representation, achieved via a 12x12 matrix of dynamics. However, according to theory, this linearization of the roller rig system appeared to be uncontrollable and unobservable via analysis by common linear algebra and control approaches. Figure 2: Control loop on linearization of the roller rig with State feedback + Cascade PID control The above figure (Figure 2), depicts the control scheme for this setup. After numerous testing, it was found that a simulation step of approximately < 1E-6 was necessary, in order to achieve stability of the continuous control loop, which cannot be used for practical application. Hence, it is for this reason that an alternative approach is necessary for the lateral skew control of such railway wheel set. Thus, this paper aims on investigating the possibility of adaptive identification and control of the stand, based on measured data. For the purposes of this paper the data will be generated via a 3-D simulation model via software SIMPACK, linked via SIMAT toolbox with MATLAB Simulink, to provide real time simulation of the roller rig as a plant system for the investigated control approaches. Then, this data will be used as the training data for neural network models, to simplify otherwise complex simulation models and to allow a constant sampling for real time usability, furthermore, investigation into the potentials for real-time neural-network based control. 3. APPLIED METHODS In this section, the various NN approaches, aimed for adaptive identification and control of the previously introduced lateral skew problem, will be described. These approaches are based on the well know Gradient Descent (GD) method used as a tool for defining the learning rule of the applied neural units. These applied neural units (adaptive models) are trained via two methods, which will be focused on in this paper. These methods are namely, the method of incremental training or Real Time Recurrent Learning (RTRL) [8] applicable for dynamic adaptive models or a batch form of training, which is a variation of the back propagation through time or (BPTT) training [10], as an extension of RTRL training in combination with the famous LevenbergMarquardt equation. 3.1 Preliminaries This subsection aims to review the important theories from works [11]-[13], behind the GD method and structures of the neural units used within this paper. Firstly, we will begin by introducing the famous GD algorithm for the linear and quadratic neural units. To understand this, we must begin with the polynomial models of the linear and quadratic (LNU and QNU) neural units respectively, as follows n y = ∑ xi wi = w0 .x0 +w1.x1 + ... + wn .xn = w.x (1) y = ∑∑ xi x j wi , j = w0,0 x0 x0 +w0,1 x0 x1 + ... + wn,n xn 2 = rowx.colW (2) i =0 n n i =0 j =0 Where, y represents the output of the LNU (1) and QNU (2) respectively. In regards to the LNU, w stands for an updatable vector of neural weights and x, represents the input values of the engineering system in the case of a purely static model and in the sense of a dynamic model, a combination of inputs of the real system and neural model outputs. Looking at equation (2), rowx is a long-vector representation of the utilised input vector. Where, colW represents a weight matrix of the quadratic neural unit in general. From this, we may then understand the GD algorithm as applied to such neural units. wi +1 = wi + µ .e(k ). ∂y (k ) ∂wi colW(k +1) = colW(k ) + µ.e(k ). ∂y (k + 1) ∂colW (3) (4) Here, equation (3) & (4) depicts the GD algorithm for both the LNU and QNU respectively. Here the output of the GD algorithm is to incrementally over sample-by-sample, update the neural weights such to adaptively teach the LNU or QNU model, to learn the engineering system. Here, µ represents the learning rate of the GD algorithm. This is analogical to humans where, setting µ relatively high, corresponds to faster learning of the human and consequently, means the less detail the human can remember and digest from their learning. Furthermore, setting this parameter to a smaller value, corresponds to a slowing rate of learning i.e. the human may remember the information learned, quite well, but less information overall about the subject. The next parameter is e(k) (k representing the number of the sample), this represents the current error between the real and calculated output of the model. The final term ∂y (k ) , ∂wi corresponds to the partial derivatives of the neural unit output, with respect to the individual neural weights. Regarding the QNU in equation (4), we see that the GD algorithm is analogical, with exception of updating the matrix of neural weights as opposed to a vector in the sense of the LNU. Till now, the structures of GD for LNU and QNU were reviewed in the format of RTRL or sample-by-sample method of learning. Where, the neural weights are updated over each new sample of the engineering system data. However for certain engineering processes, it is more advantageous to use the BPTT batch method of training these neural weights, over runs of the algorithm rather than over each sample. This is because the RTRL method focuses on the contemporary governing law of the system as opposed the BPTT which focuses more so on the main governing law of the input and outputs of the engineering system. The BPTT method is achieved via an extension of the RTRL method with combination of the famous LevenbergMarquardt equation. It is also important to note that this method is more preferable in cases where the data may be affected by noise. The following equation denotes the weight update rule for the BPTT method; Δw = (JT .J + 1 µ .I)−1.JT .e (5) Here, the neural weights are updated over each run of the algorithm (or batch trained) in the following way w=w+Δw. Here equation (5) describes the change, necessary for the update of the batch trained weights. Where, J represents the Jacobian matrix of derivatives for the neural unit. This may be the complete partial derivatives of the neural model with respect to each neural weights, or in practical applications it seems useful to simply introduce this Jacobian matrix as the input vector or matrix itself, being x and colx for LNU and QNU respectively. Furthermore, it is important to note that colx = rowxT . Often in such adaptive neural units, it is apparent that a modification of the normalised learning rate may be used to solve issues associated with instability of learning. In practise, it is possible to employ the simplified normalised learning rate as presented in the work [11], as follows; η= µ x(k)x(k)T + 1 (6) Where, equation (6) represents the normalised learning rate in the sense of LNU. This is analogically represented in the QNU, by replacement of x(k) (i.e. each k th sample of the input vector x) with colx(k). It should be noted that the above representation holds for RTRL training for dynamic adaptive models, where the algorithms used in this paper for BPTT method of training, take the learning rate µ itself. 3.2 Adaptive Identification and control The previous section focused on usage of neural units in the sense of adaptive identification of a real engineering system. In this subsection, we extend on these neural units as a method of control, with brief review from works [11]-[13], and extension as applied to the lateral skew problem, which is indeed our focus within this paper. Figure 3: Adaptive Identification with supervised learning of neural networks (where, w.x=LNU, rowX.colW=QNU) The above figure (Figure 3), depicts the identification scheme for the previously reviewed neural network architectures. Where, for the scope of this paper a simulation of the real roller rig will be used for data generation of the investigated neural network approaches for control. u, represents the input data of the roller rig, in our case this is the yaw torque of the servo motor system for manipulating the lateral skew of the wheel set. The output yreal, is the simulated output from the real time 3-D SIMPACK model and y, being the output of the neural unit, and the difference being the error, e. Figure 4: Adaptive control loop for experimental study of a neural network controller (where, w.x=LNU, rowX.colW=QNU) The above (Figure 4), shows an extension of the discussed neural architectures for application to lateral skew control of the roller rig system. Here, once the neural unit as a model is identified, it may be utilised as the foundation for a neural network based control setup. For the scope of this paper, we will propose an offline tuned control scheme, as the goal here, is to investigate the potentials of applying a neural network based control, for this application. However, the extension to online control is indeed the ultimate aim of our research to this problem, beyond this paper. Figure 4, depicts use of a second neural unit as a controller. Similarly to the previously mentioned architectures, as a controller here too the neural unit may take shape to that of the LNU or QNU adaptive models. However, in this case the adaptive neural weights are tuned differently to that of those used for the neural unit as a model and hence, depicted as v. Analogically, should a QNU structure be applied, these neural weights would be represented in a matrix form. Further to this, the above figure (Figure 4) introduced a new vector ξ. This vector comprises of a combination of outputs from the neural unit as a model and the difference between the desired behaviour (in our case desired yaw torque or lateral skew of the roller rig) and the output of the neural model. v. ξ or collectively, the variable q, thus serves as a manipulator for the newly feed samples of the neural unit as an identified model. Here the GD algorithms are employed in the following manner to achieve sample-by-sample adaptation of the neural weights for the controller, as follows; vi +1 = vi + µ.ereg (k ). ∂y (k ) ∂vi (7) Where vi , are adaptable neural weights of the neural unit as a controller and ereg(k) is the error between the desired value of the real system (in our case the roller rig, where the desired value will be denoted as d) and the real system output value at sample k. ∂y (k ) , is the partial ∂vi derivative of the output of the neural unit as a model, with respect to the individual adaptive neural weights of the neural unit as a controller. An extension of this weight update scheme for BPTT training would result in the following form v=v+Δv, where the change of neural weights for each batch would be analogical to equation (5). 4. EXPERIMENTAL ANALYSIS This section aims to analyse the previously reviewed methods of adaptive identification and control, via the introduced neural network architectures. Particularly, we aim to investigate the applicability of the discussed neural architectures for this problem of the lateral skew control of railway wheel sets. In this section, we will simulate our results using simulation data of the previously described 3-D SIMPACK model (at mid-range speed) linked with MATLAB Simulink in real time simulation. The data was produced with 0.001 sampling over a 10 second interval. First, we will analyse the results for identification of the roller rig as a system, via the different methods of the neural network models, followed by an extension of control, where the various combinations of neural network architectures will be tested. DLNU with RTRL Almost exact identification of Roller Rig data (lines are superimposed) Deviation of the model with the Roller Rig data, over each sample Figure 5: Testing of the adaptive identification, where the plant (roller rig) is represented by blue colour and green being the neural model. Trained by DLNU with RTRL Training; µ=1, epochs =10, For ny=3 (previous samples of model output) and nu=5 (previous samples of real system input) The above figure (Figure 5) and following figure (Figure 6), illustrates the adaptive identification process of the roller rig system data. We can note, that both methods of the dynamic LNU or (DLNU) and dynamic QNU or (DQNU), for RTRL learning methods, achieved almost exact identification. Where, the DQNU performed slightly faster in terms of convergence to minima of sum of square errors, as opposed to the DLNU. DQNU with RTRL. Excellent identification with slightly faster learning than the DLNU of same parameters Error of model is still within an excellent range Sum of square errors decreases more quickly with the DQNU for same parameters. Figure 6 : Testing of the adaptive identification, where the plant (roller rig) is represented by blue colour and green being the neural model. Trained by DQNU with RTRL Training; µ=1, epochs =10, For ny=3 (previous samples of model output) and nu=5 (previous samples of real system input) Desired (black dashed) Roller Rig system data (blue) Neural controller (magenta) applied to identified Roller Rig data Error plot illustrates deviation of the Neural Controller and desired behaviour. Sum of Square Errors reaches close to 10-4 after 50 epochs Figure 7: Testing of the adaptively tuned control loop, where the roller rig is represented by trained DLNU with constant and previously trained parameters (Figure 5) and the adaptive feedback controller is trained via LNU, after 10 epochs Adaptive Identification (of DLNU with RTRL) + Neural Unit Controller for lateral skew control of roller rig system - Using LNU with Incremental training µ=0.1, Data re-sampling = each 5th sample, epochs =200, for nqy=0 (previous samples of neural model output) and nqe=2 (previous samples of difference of the neural model output and desired behaviour of the system. Closer to desired behaviour of the system with QNU + BPTT The difference of the Neural Controller and desired behaviour is within +0.2 to -0.2 mm. Sum of Square Errors reaches close to 10-4 after only few epochs, with QNU and BPTT training. Figure 8: Testing of the adaptively tuned control loop, where the roller rig is represented by trained DQNU with constant and previously trained parameters (Figure 6) and the adaptive feedback controller is trained via QNU, after 10 epochs Adaptive Identification (of DQNU with RTRL) + Neural Unit Controller for lateral skew control of roller rig system - Using QNU with BPTT Training µ=0.00818, epochs=5, for nqy=0 and nqe=3 Here, Figure 7 & Figure 8 show the application of the various neural units for control of the lateral skew. The black dotted line, illustrates the chosen desired behaviour of the roller rig lateral skew. It was found that the BPTT training method used in combination with the QNU had the best performance. Providing closer control to the desired behaviour, as compared with the incremental training methods of both LNU and QNU architectures. Desired lateral skew (black dashed line) set to ideal position of zero Model provides almost exact response to the desired after the first few samples. Figure 9: Testing of the adaptively tuned control loop, where the roller rig is represented by trained DQNU with constant and previously trained parameters (Figure 6) and the adaptive feedback controller is trained via QNU, after 10 epochs Adaptive Identification (of DQNU with RTRL) + Neural Unit Controller for lateral skew control of roller rig system - Using QNU with BPTT Training µ=0.001, epochs=5, for nqy=3, nqe=3 Figure 9 depicts a theoretical test of the roller rig system under an idealised case of zero lateral skew. This is included in our results to investigate the capability of the used neural controller. From this figure, we see that after several small samples at the beginning of the controller’s application the controller provides almost exactly zero lateral skew. This result however, must be reasoned with what is really possible in terms of actuation of the real system and real conditions of the rig, such to achieve this ideal of a result. 5. DISCUSSION OF EXPERIMENTAL RESULTS In the previous section, the experimental results of the various neural network architectures for adaptive identification and control were presented. From the first set of figures, regarding the adaptive identification for the roller rig system, it was found that the most suitable learning method for both the DLNU and DQNU was the RTRL training method. Here, relatively few epochs were adequate for both architectures to identify the system data, with almost exact identification. The DQNU model was slightly faster in learning the behaviour of the system, thus, was used as the identified neural unit as a model, in the control section. The extension of these various architectures for control showed more distinguishable results. The application of both LNU and QNU with incremental learning for the neural controller was only possible when a higher number of epochs were run, or higher learning of the model with a small learning rate. Few runs of the algorithm were unable to match the desired behaviour even moderately. However, after the first few samples of application as in Figure 7, reasonable behaviour of the controller was achieved. The combination of identification via the DQNU followed by extension of a QNU trained by BPTT on the neural controller, showed to follow more closely to the desired behaviour of the roller rig system than the other tested combinations. Where, the setup of the QNU with BPTT neural controller featured input vector ξ, being comprised only of several errors or differences of the desired behaviour, with the output of the neural unit as a model. Further to this, the training was substantially quicker than the incremental training method, showing that only several epochs were sufficient for the model to achieve its optimal behaviour with respect to the desired of the system. In the above results, two forms of the desired behaviour were used. The first was under proposal that the lateral derivation of the wheel set would be within a small tolerance range, which is most realistic in real application of such control. The presented desired data is relative to the small range of lateral skew in the simulation data. Where, within this exampled range, we could be sure the wheel flanges would not contact the rails of the real roller rig. The desired behaviour is however relative to the setup of the wheel sets and rails, thus the desired deviation and lateral skew outputs may indeed vary for real railway applications. The second was an idealised situation where no lateral skew at all would occur. This is principally unrealistic for the neural controller to achieve zero lateral skew, due to a combination of natural factors on real railway wheel sets, however for demonstrational purposes, we may note the capability of the neural controller, providing almost exact behaviour of the idealised lateral skew, as applied on the identified roller rig system. 6. CONCLUSION Referring back to the original control loop depicted in Figure 1 (as presented in work [1]), of the Simulink model of the linearization of the roller rig system and state feedback with cascade PID controller. We can recall that a minimum sampling of <1E-6 seconds for numerical stability was necessary, which is not possible for practical application. After investigation of the presented neural network approaches, we may conclude that it has promising potential for real application on controlling the lateral skew of IRW railway wheel sets. Now, we note that the achieved sampling can be within order of 0.005 (5E-3) seconds, for adequate functionality of the neural network based, adaptive identification and control system as applied to the roller rig. The tested models in this paper were DLNU and DQNU for adaptive identification models and LNU or QNU architectures as a controller. The experimental results of these architectures showed that both DLNU and DQNU, precisely approximated the complex SIMPACK model of the roller rig system. This result however could vary when using more or, real training data measured from the rig itself. In terms of control of the lateral skew, the QNU architecture showed better results to keep the lateral skew within close limits to the desired behaviour of the system, which is also in general consensus with our previous findings on using QNU for control. We also can note the significantly faster tuning of the control loop via QNU as compared to the LNU, more particularly, the BPTT training method as applied to the controller compared to the incremental method. Where, more desirable control was shown, in much fewer epochs or runs of the controller algorithm. Thus, this case study shows that with proper tuning of the investigated neural units, application to the real system is theoretically possible, to achieve adequate control for this investigated problem. Thus, real application of such adaptive control loop of neural network based architecture, is indeed the very next step of our research. ACKNOWLEDGEMENTS The authors would like to acknowledge the following grant for support during this work: SGS12/177/OHK2/3T/12SGS12/177/OHK2/3T/12, Non-conventional and Cognitive Signal Processing Methods of Dynamic Systems. And also the Technology Agency of Czech Republic. Project No: TE01020038 “Competence Center of Railway Vehicles”. REFERENCES [1] J. Kalivoda, P. Bauer (2013) “Curving Behaviour of 2-Axle Bogie with Independently Rotating Wheels – Simulations and Scaled Roller Rig Tests”- 23rd International Symposium on Dynamics of Vehicles on Roads and Tracks, Qingdao, China, August 19th-23rd. [2] S. Bruni, R. M. Godall, T.X. Mei, H. Tsunashima (2007) “Control and monitoring for railway vehicle dynamics”, Vehicle System Dynamics, vol. 45 issue 7-8, pp. 743-779 [3] Gupta, M., M., Bukovsky, I., Homma, N., Solo M. G. A., Hou Z.-G. (2012): “Fundamentals of Higher Order Neural Networks for Modeling and Simulation“, in Artificial Higher Order Neural Networks for Modeling and Simulation, ed. M. Zhang, IGI Global,. [4] Bukovsky, I., Bila, J., Gupta, M., M, Hou, Z-G., Homma, N. (2010): “Foundation and Classification of Nonconventional Neural Units and Paradigm of Nonsynaptic Neural Interaction” in Discoveries and Breakthroughs in Cognitive Informatics and Natural Intelligence within the series of the Advances in Cognitive Informatics and Natural Intelligence (ACINI), ed. Y. Wang, IGI Publishing, Hershey PA, USA, ISBN: 978-1-60566-902-1, pp.508-523. [5] Bukovsky, I., Homma, N., Smetana, L., Rodriguez, R., Mironovova M., Vrana S. (2010): “Quadratic Neural Unit is a Good Compromise between Linear Models and Neural Networks for Industrial Applications”, ICCI 2010 The 9th IEEE International Conference on Cognitive Informatics, Tsinghua University, Beijing, China, July 7-9. [6] Bukovsky I., S. Redlapalli and M. M. Gupta (2003): Quadratic and Cubic Neural Units for Identification and Fast State Feedback Control of Unknown Non-Linear Dynamic Systems, Fourth International Symposium on Uncertainty Modeling and Analysis ISUMA 2003, IEEE Computer Society, , Maryland USA, ISBN 0-7695-1997-0, p.p.330-334 [7] Rodriguez , R., Bukovsky, I., Homma, N.: “Potentials of Quadratic Neural Unit for Applications”, in International Journal of Software Science and Computational Intelligence (IJSSCI) ,vol 3, issue 3, IGI Global, Publishing, Hershey PA, USA ISSN. [8] Bukovsky, I., Lepold, M., Bila J. (2010): “Quadratic Neural Unit and its Network in Validation of Process Data of Steam Turbine Loop and Energetic Boiler”,WCCI 2010, IEEE Int. Joint. Conf. on Neural Networks IJCNN, Barcelona,Spain. [9] R. J. Williams and D. Zipser (1989) “A learning algorithm for continually running fully recurrent neural networks,” Neural Computation, vol. 1, pp. 270–280. [10] P. J.Werbos (Oct. 1990) “Backpropagation through time: What it is and how to do it” Proc. IEEE, vol. 78, no. 10, pp. 1550–1560. [11] Peter M. Benes (2013): Software Application for Adaptive Identification and Controller Tuning Student's Conference STC, Faculty of Mechanical Engineering, CTU in Prague. [12] Ladislav Smetana (2008): “Nonlinear Neuro-Controller for Automatic Control Laboratory System”, Master’s Thesis, Czech Tech. Univ. in Prague, [13] Laboratory System Batyscaphe, Automatic Control Laboratory, Dpt. of Instr. and Cont. Eng., FME, Czech Technical University in Prague, online http://vlab.fsid.cvut.cz/en/ulohy/batyskaf.php, 2010. Short Biography of the Authors Peter Mark Beneš received his Bachelor’s degree with honours at Czech Technical University (CTU) in Prague in 2012. Currently he is a Master’s student, with expected PhD studies to follow in 2014. His research focuses on non-conventional neural networks for adaptive identification and control of industrial systems including hoist mechanisms and skew control of rail-based mechanisms as such cranes and railway vehicles. Peter’s work has been awarded in local and international student competitions also with an industrial BOSCH award in 2013. Matouš Cejnek received his Bachelor’s degree at Czech Technical University (CTU) in Prague in 2012. Currently he is a Master’s student at Czech Technical University in Prague. His research focuses on non-conventional neural networks for adaptive systems and novelty detection in time series and biomedical applications. Matous’s work has been awarded in local and international student competitions 2013. Jan Kalivoda graduated from Czech Technical University (CTU) in Prague where he received his Master’s degree with honours in 1996 and Ph.D. in the field of Machines and Equipment for Transportation in 2006. Currently he is a teacher and active researcher in the Department of Automotive, Combustion Engine and Railway Engineering at CTU. His research interests include MBS models of railway vehicles, Mechatronics for railway vehicles, Active control of railway vehicle suspensions and wheel sets. Ivo Bukovsky graduated from Czech Technical University in Prague (CTU) where he received his Ph.D. in the field of Control and System Engineering in 2007 and became associate professor since 2013. His research interests include higher-order neural networks, adaptive evaluation of time series and systems, multiscale-analysis approaches, control and biomedical applications. He was a visiting researcher at the University of Saskatchewan (2003), at the University of Manitoba in Canada (2010), and he was a visiting professor at Tohoku University (2011)	9cs.NE
Sequential Coding of Gauss–Markov Sources arXiv:1702.01779v3 [cs.IT] 3 Jun 2017 Anatoly Khina, Ashish Khisti, Victoria Kostina, and Babak Hassibi Abstract—We consider the problem of sequential transmission of Gauss–Markov sources. We show that in the limit of large spatial block lengths, greedy compression with respect to the squared error distortion is optimal; that is, there is no tension between optimizing the distortion of the source in the current time instant and that of future times. We then extend this result to the case where at time t a random compression rate Rt is allocated independently of the rate at other time instants. This, in turn, allows us to derive the optimal performance of sequential coding over packet-erasure channels with instantaneous feedback. For the case of packet erasures with delayed feedback, we connect the problem to that of compression with side information that is known at the encoder and may be known at the decoder — where the most recent packets serve as side information that may have been erased. We conclude the paper by demonstrating that the loss due to a delay by one time unit is rather small. Index Terms—Sequential coding of correlated sources, successive refinement, source streaming, packet erasures, source coding with side information. I. I NTRODUCTION Sequential coding of sources is increasingly finding applications, such as real-time video streaming, and cyberphysical and networked control. Such systems use compressed packetbased transmission and strive to achieve minimum distortion for the given compression rates. This setting was introduced and treated for the two-source case by Viswanathan and Berger [1] and for more users in [2]– [5]. For the special case of Gauss–Markov sources, an explicit expression for the achievable sum-rate for given distortions was derived in [2], [3] and extended for the (general) jointly Gaussian three-source case in [6]. In practice, however, packet-based protocols are prone to erasures and possible delays. The case of sequential coding in the presence of packet erasures was treated for various erasure models. The case when only the first packet is prone to an erasure was considered in [7]. A more general approach which trades between the performance given all previously sent packets and the performance given only the last packet was proposed in [8]. For random independent identically distributed (i.i.d.) packet erasures, a hybridation between pulse-code modulation (PCM) and differential PCM (DPCM), termed leaky DPCM, was proposed in [9] and analyzed for the case of very low erasure probability in [10]. The scenario in which the erasures occur in bursts was considered in [11], [12]. There, a sequence of source vectors sampled from a A. Khina, V. Kostina and B. Hassibi are with the Department of Electrical Engineering, California Institute of Technology, Pasadena, CA 91125, USA. E-mails: {khina, vkostina, hassibi}@caltech.edu A. Khisti is with the Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON M5S 3G4, Canada. E-mail: akhisti@comm.utoronto.ca Gauss–Markov process in the temporal dimension must be encoded sequentially and reconstructed with zero delay at the decoder. The channel introduces a burst of erasures of a certain maximum length and the decoder is not required to reconstruct the sequences that fall in the erasure period and a recovery window following it. All of these works assume no feedback is available at the encoder, namely that the encoder does not know whether a transmitted packet successfully arrives to the decoder or is erased in the process. In this paper, we first consider the problem of sequential coding of Gauss–Markov sources and determine the rate– distortion region for large frames. Specifically, we show that greedy quantization that optimizes the distortion for each time is also optimal for minimizing the distortion of future time instants. This insight allows us to extend the result to the case where the compression rate rt available for the transmission of the packet at time t is determined just prior to its transmission. The packet-erasure channel with instantaneous output feedback (ACK/NACK) can be viewed as a special case of the above noiseless channel with random rate allocation, with rt = 0 corresponding to a packet-erasure event [13]. The optimal rate–distortion region of sequential coding of Gauss–Markov sources in the presence of packet erasures and instantaneous output feedback thereby follows as a simple particularization of our more general result. We further tackle the more challenging delayed feedback setting, in which the encoder does not know whether the most recently transmitted packets arrived or not. Viewing these recent packets as side information (SI) that is available at the encoder and possibly at the decoder, and leveraging the results of Kaspi [14] along with their specialization for the Gaussian case by Perron et al. [15],1 we adapt our transmission scheme to the case of delayed feedback. We provide a detailed description of the proposed scheme for the case where the feedback is delayed by one time unit and demonstrate that the loss compared to the case of instantaneous feedback is small. II. P ROBLEM S TATEMENT We now present the model of the source, channel, and the admissible encoder and decoder both of which are required to be causal in this work; see Fig. 1. Throughout the paper, k·k denotes the Euclidean norm. Random variables are denoted by lower-case letters with 1 The scenario considered in [14], [15] can be also viewed as special case of the results of Heegard and Berger [16], where the SI is not available at the encoder, by adjusting the distortion measure and “augmenting” the source [17]. Interestingly, knowing the SI at the encoder allows one to improve the optimal performance of this scenario in the Gaussian case; see Rem. 9. wt st = αt st−1 + wt st Encoder N Rt bits Decoder ŝt We further define the steady-state distortion (assuming the limit exists): D∞ , lim Dt . Fig. 1: Sequential coding of a Gauss–Markov source setup. temporal subscripts (at ), and random vectors (“frames”) of length N by boldface possibly accented lower-case letters with ˆ t ). We denote temporal sequences by temporal subscripts (at , ã at , (a1 , . . . , at ). N is the set of natural numbers. All other notations represent deterministic scalars. We assume that the communication spans the time interval [1, T ], where T ∈ N. Source: Consider a Gauss–Markov source {st }, whose outcomes are vectors (“frames”) of length N with i.i.d. samples along the spatial dimension, that satisfy the temporal Markov relation: st = αt st−1 + wt , t = 1, . . . , T , (1) where {αt } are known process coefficients that satisfy \|αt \| < 1, and the outcomes of {wt } are i.i.d. along the spatial dimension, Gaussian and mutually independent across time of zero mean and variances {Wt }. We assume s0 = 0 for convenience. Denote by St , N1 E kst k2 the average power of the entries of vector st . Then, we obtain the following recursive relation: St = αt2 St−1 + Wt , t = 1, . . . , T , (2a) S0 = 0. (2b) Channel: At time t, a packet ft ∈ 1, 2, . . . , 2N Rt is sent over a noiseless channel of finite rate Rt . Causal encoder: Sees st at time t and applies a causal function Ft to the entire source sequence st to observed N Rt generate a packet ft ∈ 1, 2, . . . , 2 : t ft = Ft s . (3) Causal decoder: Applies a causal function Gt to the sequence of received packets f t to construct an estimate ŝt of st , at time t: ŝt = Gt f t . (4) Distortion: The mean-square error distortion at time t is defined as i 1 h 2 Dt , E kst − ŝt k . (5) N If we specialize the source process into that of fixed parameters, namely, αt ≡ α, t = 1, . . . , T , Wt ≡ W, then its power converges to S∞ = W . 1 − α2 (6) T →∞ Definition (Distortion-rate region). The distortion–rate region is the closure of all achievable distortion tuples DT , (D1 , . . . , DT ) for a rate tuple RT , (R1 , . . . , RT ), for any N , however large; its inverse is the rate–distortion region. III. D ISTORTION –R ATE R EGION OF S EQUENTIAL C ODING OF G AUSS –M ARKOV S OURCES The optimal achievable distortions for given rates for the model of Sec. II are provided in the following theorem. Theorem 1 (Distortion–rate region). The distortion–rate region of sequential coding for a rate tuple RT is given by all distortion tuples DT that satisfy Dt ≥ Dt∗ with ∗ Dt∗ = αt2 Dt−1 + Wt 2−2Rt , t = 1, . . . , T , (7a) D0∗ = 0. (7b) Remark 1. Th. 1 establishes the optimal rate–distortion region for the “causal encoder–causal decoder” setting of Ma and Ishwar [2] for the case of Gauss–Markov sources. We note that Ma and Ishwar [2] provide an explicit result only for the sumrate for the Gauss–Markov case [3]. Torbatian and Yang [6] extend the sum-rate result to the case of three jointly Gaussian sources (which do not necessarily constitute a Markov chain). Our work, on the other hand, fully characterizes the rate– distortion region for the case of Gauss–Markov sources. Remark 2. The results and proof (provided in the sequel) of Th. 1 imply that optimal greedy quantization at every step — which is achieved via Gaussian backward [18, Ch. 10.3] or forward [18, pp. 338–339] channels — becomes optimal when N is large. Moreover, it achieves the optimum for all t ∈ [1, T ] simultaneously, meaning that there is no tension between minimizing the current distortion and future distortions. To prove this theorem we first construct the optimal greedy scheme and determine its performance in Sec. III-A. We then show that it is in fact (globally) optimal when N goes to infinity, by constructing an outer bound for this scenario, in Sec. III-B. A. Achievable We construct an inner bound using the optimal greedy scheme. In this scheme all the quantizers are assumed to be minimum mean square error (MMSE) quantizers. We note that the quantized values of such quantizers are uncorrelated with the resulting quantization errors. Scheme. Encoder. At time t: • Generates the prediction error s̃t , st − αt ŝt−1 , (8) where ŝt−1 , defined in (4), is the previous source reconstruction at the decoder, and ŝ0 = 0. A linear recursive relation for ŝt is provided in the sequel in (9) .2 • Generates ˆ s̃t , the quantized reconstruction of the prediction error s̃t , by quantizing s̃t using the optimal MMSE quantizer of rate Rt and frame length N . • Sends ft = ˆ s̃t over the channel. Decoder. At time t: • Receives ft . • Recovers a reconstruction ˆ s̃t of the prediction error s̃t . • Generates an estimate ŝt of st : ŝt = αt ŝt−1 + ˆs̃t . (9) The optimal achievable distortions {Dt } of this scheme for long frame lengths N , are as follows. Assertion 1 (Inner bound). Let > 0, however small. Then, the expected distortion of the scheme at time t ∈ [1, T ] satisfies the recursion Dt ≤ αt2 Dt−1 + Wt 2−2Rt + , t = 1, . . . , T , (10a) D0 = 0, (10b) for a large enough N . Proof: First note that the error between st and ŝt , denoted by et , is equal to et , st − ŝt = (s̃t + αt ŝt−1 ) − αt ŝt−1 + ˆs̃t = s̃t − ˆs̃t Assertion 2 (Outer bound). Consider the setting of Sec. II. Then, the average achievable distortion Dt at time t ∈ [1, T ] is bounded from below by Dt ≥ Dt∗ , where Dt∗ satisfies (7) with equality. Proof: Let N ∈ N. We shall prove Dt ≥ 2−2Rt Efˇt−1 N st \|f t−1 = fˇt−1 ≥ (11c) s̃t , st − αt ŝt−1 = αt (st−1 − ŝt−1 ) + wt = αt et−1 + wt . Since wt is independent of et−1 , the average power of the entries of s̃t is equal to S̃t = αt2 Dt−1 + Wt . Using the property that the rate–distortion function under mean square error distortion of an i.i.d. source with given variance is upper bounded by that of a white Gaussian source with the same variance (see, e.g., [18, pp. 338–339]), we obtain the following recursion: Dt ≤ αt2 Dt−1 + Wt 2−2Rt , and hence (7) is achievable within an arbitrarily small > 0, for a sufficiently large N . = E st−1 f t−1 and αt ŝt−1 = E st f t−1 are the MMSE estimators of st−1 and st , respectively, given all the past channel outputs. Dt∗ , t = 1, . . . , T , (12a) (12b) by induction, where the sequence {Dt∗ } is defined in (7), 1 2 2 N h(st ) , N (st ) , 2π e 2 h s f k =fˇk 1 N t N st f k = fˇk , 2 2π e denote the entropy-power and conditional entropy-power of st given f k = fˇk , the expectation Efˇt−1 [·] is with respect to fˇt−1 , and the random vector fˇt is distributed the same as f t . Basic step (t = 1). First note that, since s0 = 0 and vector w1 consists of i.i.d. Gaussian entries of variance W1 , (12b) is satisfied with equality. To prove (12a), we use the fact that the optimal achievable distortion D1 for a Gaussian source (s1 = w1 ) with i.i.d. entries of power W1 and rate R1 is dictated by its rate–distortion function (RDF) [18, Ch. 10.3.2]: D1 ≥ W1 2−2R1 . (11b) where (11b) follows from (8) and (9). Thus, the distortion (5) is also the distortion in reconstructing s̃t . Using (1), (8) and (11), we express s̃t as t−1 We shall now construct an outer bound that coincides with the inner bound of Assert. 1 for large frame lengths N . (11a) 2 ŝ B. Impossible (Converse) Inductive step. Let k ≥ 2 and suppose (12) is true for t = k − 1. We shall now prove that it holds also for t = k. i 1 h 2 Dk = E ksk − ŝk k N ii 1 h h 2 = E E ksk − ŝk k f k−1 (13a) N h h ii 1 2 (13b) = Efˇk−1 E ksk − ŝk k f k−1 = fˇk−1 N (13c) ≥ Efˇk−1 N sk f k−1 = fˇk−1 2−2Rk −2R k−1 k−1 k ˇ = Efˇk−1 N αk sk−1 + wk \|f =f 2 (13d) n h i ≥ Efˇk−2 Efˇk−1 N αk sk−1 \|f k−1 = fˇk−1 fˇk−2 o + N (wk ) 2−2Rk (13e) n ≥ αk2 Efˇk−2 N sk−1 f k−2 = fˇk−2 , fk−1 o + Wk 2−2Rk (13f) n ≥ αk2 Efˇk−2 N sk−1 \|f k−2 = fˇk−2 2−2Rk−1 o + Wk 2−2Rk (13g) −2Rk 2 ∗ ≥2 αk Dk−1 + Wk (13h) = Dk∗ , (13i) where (13a) follows from the law of total expectation, (13b) holds since f k−1 and fˇk−1 have the same distribution, (13c) follows by bounding from below the inner expectation (conditional distortion) by the rate–distortion function and the Shannon lower bound [18, Ch. 10] — this also proves (12a), (13d) is due to (1), (13e) follows from the entropy-power inequality [18, Ch. 17], (13f) holds since wk is Gaussian, the scaling property of differential entropies and Jensen’s inequality: Efˇ 2 h s k−1 2N f k−1 =fˇk−1 k−1 2 ≥ 2N 2 ≡ 2N Efˇ k−1 h fˇk−2 f k−2 =fˇk−2 , fk−1 IV. R ANDOM -R ATE B UDGETS , Assertion 3 (Outer bound for non-Gaussian noise). Consider the setting of Sec. II with independent non-Gaussian noise entries {wt }. Then, the average achievable distortion Dt at time t ∈ [1, T ] is bounded from below by Dt ≥ Dt∗ , with Dt∗ given by the recursion ∗ Dt∗ = α2 Dt−1 + N (wt ) 2−2Rt , D0∗ = 0, is the entropy-power of wt . Proof: The proof is identical to that of Assert. 2 with Wt replaced by N (wt ).3 C. Steady State of Asymptotically Stationary Sources For the asymptotically stationary source in (6), the steadystate average distortion is as follows. Corollary 1 (Steady state). Let > 0, however small. Then, the minimum steady-state distortion is equal to ∗ D∞ = 0 ≶ Dt − D∞ ≶ Dt−1 − D∞ (13h) is by the induction hypothesis, and (13i) holds by the definition of {Dt∗ } as the sequence that satisfies (7) — which also proves (12b). This concludes the proof of (12b) as desired. 1 h(wt ) 2πe 2 Hence, if 0 ≶ Dt−1 − D∞ , then Remark 3. As is evident from the proof, the result of Corol. 1 remains true for any initial value D0 . (13g) follows from the following standard set of inequalities: N Rk−1 ≥ H fk−1 f k−2 = fˇk−2 ≥ I sk−1 ; fk−1 f k−2 = fˇk−2 = h sk−1 f k−2 = fˇk−2 − h sk−1 f k−2 = fˇk−2 , fk−1 , where N (wt ) = Dt − D∞ = α2 2−2R < 1. Dt−1 − D∞ and converges (exponentially fast) to D∞ . i h h sk−1 f k−1 =fˇk−1 sk−1 or equivalently W 2−2R + , 1 − α2 2−2R for a large enough N . (15) In this section we generalize the results of Sec. III to random rates {rt } that are independent of each other and of {wt }. Rate rt is revealed to the encoder just before the transmission at time t. Theorem 2 (Distortion–rate region). The distortion–rate region of sequential coding with independent rates rT is given by all distortion tuples DT that satisfy Dt ≥ Dt∗ withD0∗ = 0 and ∗ Dt∗ = αt2 Dt−1 + Wt E 2−2rt , t = 1, . . . , T . (16) Remark 4. An immediate consequence of this theorem and Jensen’s inequality is that using packets of a fixed rate that is equal to E [rt ] performs better than using random rates. Proof: Achievable. Since the achievability scheme in Th. 1 does not use the knowledge of future transmission rates to encode and decode the packet at time t, we have i 1 h 2 (17a) dt , E kst − ŝt k rT N h i 1 2 = E kst − ŝt k rt (17b) N 2 −2rt ≤ (αt dt−1 + Wt )2 + , (17c) Taking an expectation of (17c) with respect to rt and using the independence of rt−1 and rt , we achieve (16). Impossible. Revealing the rates to the encoder and the decoder prior to the start of transmission can only improve the distortion. Thus, the distortions {dt } conditioned on {rt } (17a) are bounded from below as in Th. 1; by taking the expectation with respect to {rt }, we attain the desired result. For the special case of an asymptotically stationary source (6), the steady-state distortion is given as follows. Corollary 2 (Steady state). Assume that the rates {rt } are Proof: Note that (15) is a fixed point of (7a) (up to ). i.i.d. Let > 0, however small. Then, the minimum steady−2R Now since α < 1 and 2 < 1, Dt converges to D∞ . state distortion is equal to This can be easily proved as follows. Assume Dt−1 6= D∞ BW (otherwise we are already at the fixed point). Then, D∞ = + (18) 2 −2R 2 −2R 1 − α2 B Dt − D∞ = α Dt−1 + W 2 − α D∞ + W 2 for a large enough N , where 2 −2R =α 2 (Dt−1 − D∞ ) , B , E 2−2rt . 3 Recall that in the Gaussian setting N (wt ) = Var (wt ) ≡ Wt . Proof: Note that (18) is a fixed point of (16). Since α < 1 and B < 1, E [Dt ] converges to D∞ . This can be easily proved as follows. Assume Dt−1 6= D∞ (otherwise we are already at the fixed point). Then, Dt − D∞ = α2 Dt−1 + W B − α2 D∞ + W B = α2 B (Dt−1 − D∞ ) , B. Multiple Packets Per Frame or equivalently Dt − D∞ = α2 B < 1. Dt−1 − D∞ Hence, if 0 ≶ Dt−1 − D∞ , then 0 ≶ Dt − D∞ ≶ Dt−1 − D∞ and converges exponentially fast to D∞ . V. PACKET E RASURES WITH I NSTANTANEOUS F EEDBACK A. One Packet Per Frame An important special case of the random-rate budget model of Sec. IV is that of packet erasures [13]. Since a packet erasure at time t can be viewed as rt = 0, and assuming that the encoder sends packets of fixed rate R and is cognizant of any packet erasures instantaneously, the packet erasure channel can be cast as the random rate channel of Sec. IV with rt = bt R ( R, bt = 1 = 0, bt = 0 (19a) (19b) where {bt } are the packet-erasure events, such that bt = 1 corresponds to a successful arrival of the packet ft at time t, and bt = 0 means it was erased. We further denote by gt , bt ft (20) the received output where gt = 0 corresponds to an erasure, and otherwise gt = ft . We assume that {bt } are i.i.d. according to a Ber(β) distribution for β ∈ [0, 1]. Remark 5. We shall concentrate on the case of packets of fixed rate R to simplify the subsequent discussion. This way the only randomness in rate comes from the packet-erasure effect. Nevertheless, all the results that follow can be easily extended to random/varying rate allocations to which the effect of packet erasures {bt } is added in the same manner as in (19). Corollary 3 (Distortion–rate region). The distortion–rate region of sequential coding with packet erasures and instantaneous feedback is given as in Th. 2 with E 2−2rt = 1 − β 1 − 2−2R . Proof: Computing the expectation, we obtain E 2−2rt = E 2−2bt R = β2−2R + (1 − β), as desired. Corollary 4 (Steady state). The steady-state distortion is given as in Corol. 2 with B , E 2−2rt = 1 − β 1 − 2−2R . In Sec. V-A we assumed one packet (ft ) was sent per each source frame (st ). Instead, one may choose to transmit multiple packets of lower rate per one source frame. Naı̈ve repetition of the same packet trades off diversity for multiplexing in this case [19] and can potentially improve the overall performance. An improvement over this naı̈ve repetition-based scheme was proposed in [20], where the repetitive transmission of a single compressed description was replaced by multiple descriptions compression [21]–[24]. If we assume the availability of a perfect instantaneous feedback after each packet, a further improvement can be achieved by noting that this scenario falls again in the randomrate budget framework of Sec. IV. Specifically, if we assume the use of K packets of equal rate R/K (and hence a total rate of R), the rate probability distribution amounts to bt rt = R K with bt denoting the number of successful packet arrivals at time t, corresponding to source frame st . Assuming that the erasure events of all packets are i.i.d. with probability 1 − β implies that {bt } are i.i.d. according to a Binomial distribution B (K, β). Interestingly, the optimal number of packets depends on the (total) rate R and packet successful arrival probability β, and is determined by the number that minimizes E [2−rt ]. This is demonstrated in Fig. 2. Remark 6. We only considered uniform rate allocations for all the packets. Clearly, one can generalize the same approach to non-uniform packet rates. Remark 7. In practice one might expect longer packets to be prone to higher erasure probability. This can be taken into account when deciding on the K that minimizes E 2−2rt . VI. PACKET E RASURES WITH D ELAYED F EEDBACK In this section we consider the case of i.i.d. packet erasures with a delayed-by-one output feedback, i.e., the case where at time t, the encoder does know whether the last packet arrived or not (does not know bt−1 ), but knows the erasure pattern of all preceding packets (knows bt−2 ). The encoder (3) and decoder (4) mappings can be written as [recall the definition of gt , bt ft in (20)]: ft = Ft st , g t−2 , ŝt = Gt g t . To that end, we recall the following result by Perron et al. [15, Th. 2], which is a specialization to the jointly Gaussian rate required to achieve these distortions is given by RKaspi (S, Z, D− , D+ )   0, D− ≥ S and D+ ≥ Z  1  S  D− < S and D+ kS ≥ D− kZ   2 log D− , Z 1 D+ < Z and D− ≥ D+ + S − Z = 2 log D+ , (    D− < S and D+ kS < D− kZ  1 S   log , − 2 2 D −∆ and D− < D+ + S − Z ab denotes the harmonic mean of a and b, and where akb , a+b p p (S − Z)(S − D− )D+ − (Z − D+ )(D− − D+ )S √ ∆, . Z (S − D+ ) (a) R = 1 (b) R = 5.5 Fig. 2: Evaluation of 2−rt for K = 1, 2 and 3 packets, all possible values of β ∈ [0, 1], and two (total) rates R = 1 and 5.5. case of the result by Kaspi [14, Th. 1], who established the rate–distortion region of lossy compression with two-sided SI where the SI may or may not be available at the decoder.4 Remark 8. Kaspi’s result [14, Th. 1] can also be viewed as a special case of [16] with some adjustments; see [17]. Theorem 3 ( [15]). Let s be an i.i.d. zero-mean Gaussian source of power S, which is jointly Gaussian with SI y, which is available at the encoder and satisfies s = y + z where z is an i.i.d. Gaussian noise of power Z that is independent of y. Denote by ŝ+ and ŝ− the reconstructions of s with and without the SI y, and by D+ and D− their mean squared error distortion requirements, respectively. Then, the smallest 4 We use a backward channel to represent the SI s = y + z, as opposed to the forward channel y = s + z used in [15], [16]. Remark 9. Surprisingly, as observed by Perron et al. [15], if the side-information signal y is not available at the encoder — corresponding to the case considered in [16] and [14, Th. 2] — the required rate can be strictly higher than that in Th. 3. This is in stark contrast to the case where the side-information is never available at the encoder and the case where the sideinformation is always available at the decoder studied by Wyner and Ziv [25], [26]. Knowing the SI at the encoder allows to (anti-)correlate the noise z with the quantization error — some thing that is not possible when the SI is not available at the encoder, as the two noises must be independent in that case. This allows for some improvement, though a modest one, as implied by the results for the dual channel problem [27, Prop. 1], [28]. In our case, at time t, the previous packet ft−1 will serve as the SI. Note that it is always available to the encoder; the decoder may or may not have access to it, depending whether the previous packet arrived or not. Since the feedback is delayed, during the transmission of the current packet ft the encoder does not know whether the previous packet was lost. The tradeoff between D+ and D− for a given rate R will be determined by the probability of a successful packet arrival β. Scheme (Kaspi-based). Encoder. At time t: • Generates the prediction error s̃t , st − αt2 ŝt−2 . • Generates ft by quantizing the prediction error s̃t as in Th. 3, where ft−1 is available as SI at the encoder and possibly at the decoder (depending on bt−1 ) using the optimal quantizer of rate R and frame length N that minimizes the averaged over bt−1 distortion: DtWeighted = βDt+ + (1 − β)Dt− ; (21) more precisely, since the encoder does not know (bt−1 , bt ) at time t: – Denote the reconstruction of s̃t at the decoder from ft and g t−1 — namely given that bt = 1 — by Qt (s̃t ), and the corresponding distortion by DtWeighted . • – Denote the reconstruction from ft and g t−2 — namely given that bt = 1 and bt−1 = 0 — by − Q− t (s̃t ), and the corresponding distortion by Dt . – Denote the reconstruction from (ft−1 , ft ) and g t−2 — namely given that bt = 1 and bt−1 = 1 — + by Q+ t (s̃t ), and the corresponding distortion by Dt . Then, the encoder sees αt Q+ t−1 (s̃t−1 ) as possible SI available at the decoder to minimize DtWeighted as in (21). Sends ft over the channel. Decoder. At time t: • Receives gt . • Generates a reconstruction ˆ s̃t of the prediction error s̃t :  +  Qt (s̃t ), bt = 1, bt−1 = 1 ˆs̃t = Q− (s̃t ), bt = 1, bt−1 = 0 (22) t   0, bt = 0 • Generates an estimate ŝt of st : ŝt = αt ŝt−1 + ˆs̃t . This scheme is the optimal greedy scheme whose performance is stated next, in the limit of large N . Theorem 4. Let > 0, however small. Then, for a large enough N , the expected distortion of the scheme at time t ∈ [2, T ] given (b1 , . . . , bt ) satisfies the recursion  +  bt = 1, bt−1 = 1 Dt + , − Dt = Dt + , bt = 1, bt−1 = 0   2 αt Dt−1 + W + , bt = 0 D1 = D1+ = D1− −b1 2R = Wt 2 + , Fig. 3: Distortions Dt as a function of the time t of the various schemes presented in this section, along with that of the instantaneous-feedback scheme of Sec. V, for α = 0.7, W = 1, β = 0.5 and R = 2. (we derive their performance for the special case of an asymptotically stationary source): • No prediction: A scheme that does not use prediction at all, as if the source samples were independent. This scheme achieves a distortion of Dt = βSt 2−2R + (1 − β)St , • where Dt+ and Dt− are the distortions that minimize DtWeighted = βDt+ + (1 − β)Dt− , − + RKaspi (αt Dt−1 + W, αt Dt−1 + W, Dt− , Dt+ ) = R. The proof is again the same as that of Ths. 1 and 2, with ˆs̃t generated as in (22). Remark 10. Here, in contrast to the case of instantaneous feedback, evaluating the average distortions {Dt } in explicit form (recall Corol. 3) is much more challenging. We do it numerically, instead. Somewhat surprisingly, the loss in performance of the Kaspi-based scheme due to the feedback delay is rather small compared to the scenario in Sec. V where the feedback is available instantaneously, for all values of β.5 This is demonstrated in Fig. 3, where the perfomances of these schemes are compared along with the performances of the following three simple schemes for αt ≡ 0.7, W ≡ 1, β = 0.5, R = 2 5 For β values close to 0 or 1, the loss becomes even smaller as in these cases using the scheme of Sec. V that assumes that the previous packet arrived or was erased, respectively, becomes optimal. where St is the power of the entries of st as given in (2). Assumes worst case (WC): Since at time t the encoder does not know bt−1 , a “safe” way would be to work as if bt−1 = 0. This achieves a distortion of Dt = α4 Dt−2 + (1 + α2 )W β2−2R + (1 − β)2 + β(1 − β)(α2 Dt−1 + W ), such that the rate of Th. 3 satisfies D0 = 0, • t = 1, . . . , T , D1 = W 2 −2R t = 2, . . . , T , . Assumes best case (BC): The optimistic counterpart of the previous scheme is that which always works as if bt−1 = 1. This scheme achieves a distortion of Dt = β α2 Dt−1\|t−2 2−2R + W β2−2R + (1 − β) + (1 − β) α2 Dt−1\|t−2 + W , t = 2, . . . , T , Dt−1\|t−2 , α2 Dt−2 + W, D0 = 0, t = 2, . . . , T , D1 = W 2−2R . VII. D ISCUSSION : F EEDBACK WITH L ARGER D ELAYS To extend the scheme of Sec. VI for larger delays, a generalization of Th. 3 is needed. Unfortunately, the optimal rate–distortion region for more than two decoders remains an open problem and is only known for the case when the source and the possible SIs form a Markov chain (“degraded”). Nonetheless, achievable regions for multiple decoders have been proposed in [16], which can be used for the construction of schemes that accommodate larger delays. ACKNOWLEDGMENT The authors thank Yu Su from Caltech for valuable discussions. R EFERENCES [1] H. Viswanathan and T. Berger, “Squential coding of correlated sources,” IEEE Trans. Inf. Theory, vol. 46, no. 1, pp. 236–246, Jan. 2000. [2] N. Ma and P. Ishwar, “On delayed sequential coding of correlated sources,” IEEE Trans. Inf. Theory, vol. 57, no. 6, pp. 3763–3782, 2011. [3] ——, “Erratum to “on delayed sequential coding of correlated sources”,” IEEE Trans. Inf. Theory, vol. 58, no. 6, p. 4074, June 2012. [4] E.-H. Yang, L. Zheng, and D.-K. He, “Rate distortion theory for causal video coding: Characterization, computation algorithm, and comparison,” IEEE Trans. Inf. Theory, vol. 57, no. 8, pp. 5258–5280, 2011. [5] ——, “On the information theoretic performance comparison of causal video coding and predictive video coding,” IEEE Trans. Inf. Theory, vol. 60, no. 3, pp. 1428–1446, Mar. 2014. [6] M. Torbatian and E.-H. Yang, “Causal coding of multiple jointly Gaussian sources,” in Proc. Annual Allerton Conf. on Comm., Control, and Comput., Monticello, IL, USA, Oct. 2012, pp. 2060–2067. [7] M. Eslamifar, “On causal video coding with possible loss of the first encoded frame,” Master’s thesis, University of Waterloo, Waterloo, Ontario, Canada, 2013. [8] L. Song, J. Chen, J. Wang, and T. Liu, “Gaussian robust sequential and predictive coding,” IEEE Trans. Inf. Theory, vol. 59, no. 6, pp. 3635– 3652, June 2013. [9] H. C. Huang, W. H. Peng, and T. Chiang, “Advances in the scalable amendment of H.264/AVC,” IEEE Comm. Magazine, vol. 45, no. 1, pp. 68–76, Jan. 2007. [10] Y.-Z. Huang, Y. Kochman, and G. W. Wornell, “Causal transmission of colored source frames over packet erasure channel,” in Proc. Data Comp. Conf. (DCC), Snowbird, UT, USA, Mar. 2010, pp. 129–138. [11] F. Etezadi, A. Khisti, and M. Trott, “Zero-delay sequential transmission of Markov sources over burst erasure channels,” IEEE Trans. Inf. Theory, vol. 60, no. 8, pp. 4584–4613, Aug. 2014. [12] F. Etezadi, A. Khisti, and J. Chen, “A truncated prediction framework for streaming over erasure channels,” IEEE Trans. Inf. Theory, Submitted Jul. 2014, Revised Oct. 2016. [13] P. Minero, M. Franceschetti, S. Dey, and G. N. Nair, “Data rate theorem for stabilization over time-varying feedback channels,” IEEE Trans. Auto. Control, vol. 54, no. 2, pp. 243–255, Feb. 2009. [14] A. H. Kaspi, “Rate–distortion when side-information may be present at the decoder,” IEEE Trans. Inf. Theory, vol. 40, no. 6, pp. 2031–2034, Nov. 1994. [15] E. Perron, S. Diggavi, and I. E. Telatar, “On the role of encoder sideinformation in source coding for multiple decoders,” in Proc. IEEE Int. Symp. on Inf. Theory (ISIT), Seattle, WA, USA, July 2006, pp. 331–335. [16] C. Heegard and T. Berger, “Rate–distortion when side information may be absent,” IEEE Trans. Inf. Theory, vol. 31, pp. 727–734, Nov. 1985. [17] A. Khina and U. Erez, “Source coding with composite side information at the decoder,” in Proc. IEEE Conv. Electrical and Electron. Engineers in Israel (IEEEI), Eilat, Israel, Nov. 2012. [18] T. M. Cover and J. A. Thomas, Elements of Information Theory, Second Edition. New York: Wiley, 2006. [19] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. U.K: Cambridge Univ. Press, 2005. [20] J. Ostergaard and D. Quevedo, “Multiple descriptions for packetized predictive control,” EURASIP J. Advances in Sig. Proc., vol. 2016, no. 45, Apr. 2016. [21] H. Witsenhausen, “On source networks with minimal breakdown degradation,” Bell Sys. Tech. Jour., vol. 59, pp. 1083–1087, July-Aug. 1980. [22] J. K. Wolf, A. D. Wyner, and J. Ziv, “Source coding for multiple descriptions,” Bell Sys. Tech. Jour., vol. 59, pp. 1417–1426, Oct. 1980. [23] L. H. Ozarow, “On the source coding problem with two channels and three receivers,” Bell Sys. Tech. Jour., vol. 59, pp. 1909–1922, 1980. [24] A. El Gamal and T. M. Cover, “Achievable rates for multiple descriptions,” IEEE Trans. Inf. Theory, vol. 28, no. 6, pp. 851–857, Nov. 1982. [25] A. D. Wyner and J. Ziv, “The rate–distortion function for source coding with side information at the decoder,” IEEE Trans. Inf. Theory, vol. 22, pp. 1–10, Jan. 1976. [26] A. D. Wyner, “The rate–distortion function for source coding with side information at the decoder — II: General sources,” Information and Control, vol. 38, pp. 60–80, 1978. [27] R. Zamir and U. Erez, “A Gaussian input is not too bad,” IEEE Trans. Inf. Theory, vol. 50, no. 6, pp. 1362–1367, Jun. 2004. [28] T. Philosof and R. Zamir, “The cost of uncorrelation and noncooperation in MIMO channels,” IEEE Trans. Inf. Theory, vol. 53, no. 11, pp. 3904– 3920, Nov. 2007.	7cs.IT
Reuse of Neural Modules for General Video Game Playing Alexander Braylan Mark Hollenbeck Elliot Meyerson Risto Miikkulainen arXiv:1512.01537v1 [cs.NE] 4 Dec 2015 Department of Computer Science, The University of Texas at Austin {braylan,mhollen,ekm,risto}@cs.utexas.edu Abstract A general approach to knowledge transfer is introduced in which an agent controlled by a neural network adapts how it reuses existing networks as it learns in a new domain. Networks trained for a new domain can improve their performance by routing activation selectively through previously learned neural structure, regardless of how or for what it was learned. A neuroevolution implementation of this approach is presented with application to high-dimensional sequential decision-making domains. This approach is more general than previous approaches to neural transfer for reinforcement learning. It is domain-agnostic and requires no prior assumptions about the nature of task relatedness or mappings. The method is analyzed in a stochastic version of the Arcade Learning Environment, demonstrating that it improves performance in some of the more complex Atari 2600 games, and that the success of transfer can be predicted based on a high-level characterization of game dynamics. Introduction The ability to apply available previously learned knowledge to new tasks is a hallmark of general intelligence. Transfer learning is the process of reusing knowledge from previously learned source tasks to bootstrap learning of target tasks. In long-range sequential control domains, such as robotics and video game-playing, transfer is particularly important, because previous experience can help agents explore new environments efficiently (Taylor and Stone 2009; Konidaris, Scheidwasser, and Barto 2012). Knowledge acquired during previous tasks also contains information about an agent’s domain-independent decision making and learning dynamics, and thus can be useful even if the domains seem unrelated. Existing approaches to transfer learning in such domains have demonstrated successful transfer of varying kinds of knowledge, but they make two fundamental assumptions that restrict their generality: (1) some sort of a priori humandefined understanding of how tasks are related, and (2) separability of knowledge extraction and target learning. The first assumption limits how well the approach can be applied by restricting its use only to cases where the agent Copyright c 2016, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. has been provided with this additional relational knowledge, or, if it can be learned (Talvitie and Singh 2007; Taylor, Kuhlmann, and Stone 2008; Ammar et al. 2015b), cases where task mappings are useful. The second assumption implies that it is known what knowledge will be useful and how it should be incorporated before learning on the target task begins, preventing the agent from adapting the way it uses source knowledge as it gains information about the target domain. General ReUse of Static Modules (GRUSM) is proposed in this paper as a general neural network approach to transfer learning that avoids both of these assumptions. GRUSM augments the learning process to allow learning networks to route through existing neural modules (source networks) selectively as they simultaneously develop new structure for the target task. Unlike previous work, which has dealt with mapping task variables between source and target, GRUSM is domain-independent, in that no knowledge about the structure of the source domain or even knowledge about where the network came from is required for it to be reused. Instead of using mappings between task-spaces to facilitate transfer, it searches directly for mappings in the solution space, that is, connections between existing source networks and the target network. This approach is motivated by studies that have shown in both naturally occurring complex networks (Milo et al. 2002) and in artificial neural networks (Swarup and Ray 2006) that certain network structures repeat and can be useful across domains, without any context for how exactly this structure should be used. This work is further motivated by the idea that neural resources in the human brain are reused for countless purposes in varying complex ways (Anderson 2010). In this paper, an implementation of GRUSM based on the Enforced Subpopulations (ESP) neuroevolution framework (Gomez and Miikkulainen 1997; 1999) is presented. The approach is validated on the stochastic Atari 2600 general game playing platform, finding that GRUSM-ESP improves learning for more complex target games, and that these improvements may be predicted based on domain complexity features. This result demonstrates that even without traditional transfer learning assumptions, successful knowledge transfer via general reuse of existing neural modules is possible and useful for long-range sequential control tasks. In principle, this approach scales naturally to transfer from an arbitrary number of source tasks, which suggests that in the future it may be possible to build GRUSM agents that accumulate and reuse knowledge throughout their lifetimes across a variety of diverse domains. Background Transfer learning encompasses machine learning techniques that involve reusing existing source knowledge in a different target task or domain. A domain is an environment in which learning takes place, characterized by the input and output space; a task is a particular function from input to output to be learned (Pan and Yang 2010). In sequentialdecision domains, a task is characterized by the values of sensory-action sequences corresponding to the pursuit of a given goal. A taxonomy of types of knowledge that may be transferred was also enumerated by Pan and Yang. Because the GRUSM approach reuses the structure of existing neural networks, it falls under feature representation transfer. Transfer Learning for RL Transfer learning for sequential decision-making domains has been studied extensively within the reinforcement learning (RL) paradigm (Taylor and Stone 2009). Reinforcement learning domains are often formulated as Markov decision processes (MDPs) in which the state space comprises all possible observations, and the probability of an observation depends only on the previous observation and action taken by a learning agent. However, many real world RL domains are non-Markovian, including many Atari 2600 games, for example, the velocity of a moving object cannot be determined by looking at a single frame. The Atari 2600 platform also supports a wide variety of games. Existing RL approaches to transfer differ on the types of differences allowed between source and target task. Some approaches that are general with respect to the kind of knowledge that can be transferred are restricted in that they require a consistent agent-space (Konidaris, Scheidwasser, and Barto 2012), or an a priori specification of inter-task mappings defining relationships between source and target state and action variables (Brys et al. 2015). Existing approaches to transfer learning that encode policies as neural networks require such a specification (Taylor, Whiteson, and Stone 2007; Verbancsics and Stanley 2010). On the other hand, existing modular neuroevolution approaches that are more general with respect to connectivity (Reisinger, Stanley, and Miikkulainen 2004; Khare et al. 2005) have not been applied to cross-domain transfer. Some of the most general existing approaches to transfer for RL automatically learn task mappings, so they need not be provided beforehand. These approaches are general enough to apply to any reinforcement learning domains, but initial approaches (Taylor, Kuhlmann, and Stone 2008; Talvitie and Singh 2007) were intractable for high dimensional state and action spaces due to combinatorial blowup in the number of possible mappings. However, recent approaches in policy gradient RL (Ammar et al. 2015b; 2015a) can both tractably learn mappings and be applied across diverse domains. These approaches have been successful in continuous control domains, but it is unclear how they would scale to domains with many discretely-valued features such as Atari. Also, the above approaches assume MDP environments, whereas GRUSM can use recurrent neural networks to extend to POMDPs. General Neural Structure Transfer There are existing algorithms similar to GRUSM in that they make it possible to reuse existing neural structure. They can apply to a wide range of domains and tasks in that they automatically select source knowledge and avoid inter-task mappings. For example, Shultz and Rivest (2001) developed a technique to build increasingly complex networks by inserting source networks chosen by how much they reduce error. This technique is only applicable to supervised learning, because the source selection depends heavily on an immediate error calculation. Also, connectivity between source and target networks is limited to the input and output layer of the source. As another example, Swarup and Ray (2006) introduced an approach that creates sparse networks out of primitives, or commonly used sub-networks, mined from a library of source networks. This subgraph mining approach depends on a computationally expensive graph mining algorithm, and tends to favor exploitation over innovation and small primitives rather than larger networks as sources. The GRUSM approach is more general in that it can be applied to unsupervised and reinforcement learning tasks, makes few a priori assumptions about what kind of sources and mappings should work best, and is able to develop memory via recurrent connections. Although an evolutionary approach is developed in this paper, GRUSM should be extensible to any neural network-based learning algorithm. Approach This section introduces the general idea behind GRUSM, provides an overview of the ESP neuroevolution framework, and describe the particular implementation: GRUSM-ESP. General ReUse of Static Modules (GRUSM) The underlying idea is that an agent learning a neural network for a target task can reuse knowledge selectively from existing neural modules (source networks) while simultaneously developing new structure unique to a target task. This approach attempts to balance reuse and innovation in an integrated architecture. Both source networks and new hidden nodes are termed recruits. Recruits are added to the target network during the learning process. Recruits are incorporated adaptively into the target network as it learns connection parameters from the target to the recruit and from the recruit to the target. All internal structure of source networks is frozen to allow learning of connection parameters to remain consistent across recruits. This mechanism forces the target network to transfer learned knowledge, rather than simply overwrite it. Connections to and from source networks can, in the most general case, connect to any nodes in the source and target, minimizing assumptions about what knowledge will be useful. A GRUSM network is a 3-tuple G = (M, S, T ) where M is a traditional neural network (feedforward or recurrent) containing the new nodes and connections unique to the target task, with input and output nodes corresponding to inputs and outputs defined by the target domain; S is a (possibly empty) set of pointers to recruited source networks S1 , ..., Sk ; and T is a set of weighted transfer connections between nodes in M and nodes in source networks, that is, for any connection ((u, v), w) ∈ T , (u ∈ M ∧v ∈ Si )∨(u ∈ Si ∧ v ∈ M ) for some 0 ≤ i ≤ k. This construction strictly extends traditional neural networks so that each Si can be a traditional neural network or a GRUSM network of its own. When G is evaluated, only the network induced by directed paths from inputs of M to outputs of M , including those which pass through some Si via connections in T is evaluated. During each evaluation of G, all recruited source network inputs are fixed at 0, since the agent is concerned only with performing the current target task. The parameters to be learned are the usual parameters of M , along with the contents of S and T . The internal parameters of each Si are frozen in that they cannot be rewritten through G. The motivation for this architecture is that if the solution to a source task contains any information relevant to solving a target task, then the neural network constructed for the source task will contain some structure (subnetwork or module) that will be useful for a target network. This has been previously observed in naturally occurring complex networks (Milo et al. 2002), as well as cross-domain artificial neural networks (Swarup and Ray 2006). Unlike the subgraph-mining approach to neural structure transfer (Swarup and Ray 2006), this general formalism makes no assumptions as to what subnetworks actually will be useful. One interpretation is that a lifelong learning agent maintains a system of interconnected neural modules that it can potentially reuse at any time for a new task. Even if existing modules are unlabeled, they may still be useful, due to the fact that they contain knowledge of how the agent can successfully learn. Furthermore, advances in reservoir computing (Lukoševičius and Jaeger 2009) have demonstrated the power of using large amounts of frozen neural structure to facilitate learning of complex and chaotic tasks. The above formalism is general enough to allow for an arbitrary number of source networks and arbitrary connectivity between source and target. In this paper, to validate the approach and simplify analysis, at most one source network is used at a time and only connections from target input to source hidden layer and source output layer to target output are permitted. By not allowing target input to connect to source input, this restriction avoids high-dimensional transformations between domain-specific sensor substrates, and more intuitively captures the domain-agnostic goals of the approach, differentiating the approach from previous methods that have used direct mappings between sensor spaces. This restriction is sufficient to show that the implementation can reuse hidden source features successfully, and it is possible to analyze the cases in which transfer is most useful. Future refinements are discussed in the Discussion and Future Work section. The current implementation, described below, is a neuroevolution approach based on ESP. Enforced Subpopulations (ESP) Enforced Sub-Populations (ESP; Gomez and Miikkulainen 1997; 1999) is a neuroevolution technique in which different components of a neural network are evolved in separate subpopulations rather than evolving the whole network in a single population. ESP has been shown to perform well in a variety of reinforcement learning domains, and has shown promise in extending to POMDP environments, in which use of recurrent connections for memory is critical (Gomez and Miikkulainen 1999; Gomez and Schmidhuber 2005; Schmidhuber et al. 2007). In traditional ESP, there is a single hidden layer, each neuron of which is evolved in its own subpopulation. Recombination occurs only between members of the same subpopulation, and mutants in a subpopulation derive only from members of that subpopulation. The genome of each individual in a subpopulation is a vector of weights corresponding to the weights of connections to and from that neuron, including node bias. In each generation, networks to be evaluated are randomly constructed by inserting one neuron from each subpopulation. Each individual that participated in the network receives the fitness achieved by that network. When fitness converges, i.e., does not improve over several consecutive generations, ESP makes use of burst phases. In initial burst phases each subpopulation is repopulated by mutations of the single best neuron ever occuring in that subpopulation, so that it reverts to searching a δ-neighborhood around the best solution found so far. If a second consecutive burst phase is reached, i.e., no improvements were made since the previous burst phase, a new neuron with a new subpopulation may be added (Gomez 2003). GRUSM-ESP The idea of enforced sub-populations is extended to transfer learning via GRUSM networks. For each reused source network Si , the transfer connections in T between Si and M evolve in a distinct subpopulation. At the same time new hidden nodes can be added to M ; they evolve within their own subpopulations in the manner of standard ESP. In this way, the integrated evolutionary process simultaneously searches the space for how to reuse each potential source network and how to innovate with each new node. The GRUSMESP architecture (Figure 1) is composed of the following elements: (1) A pool of potential source networks. In the experiments in this paper, each target network reuses at most one source at a time; (2) Transfer genomes encoding the weights of cross-network connections between source and target. Each potential source network in the pool has its own subpopulation for evolving transfer genomes between it and the target network. Each connection in T is contained in some transfer genome. In our experiments, the transfer connections included are those such that the target’s inputs are fully connected to the source’s hidden layer, and the source’s outputs are fully connected into the target’s outputs; (3) A burst mechanism that determines when innovation is necessary based on a recent history of performance improvement. New hidden recruits (source networks when available, and single nodes otherwise) added during the burst phase evolve within their own subpopulations as in standard ESP. Figure 1: The GRUSM-ESP architecture, showing the balance between reused and new structure. In this example, the target network has three recruits: one source network, and two single nodes. Each bold edge between target network nodes and source network recruit indicate connections to multiple source nodes. The genome in each subpopulation encodes weight information for the connections from and to the corresponding recruit. All hidden and output neurons use a hyperbolic tangent activation function. Networks include a single hidden layer, and include recurrent self loops on hidden nodes; they are otherwise feedforward. The details of the genetic algorithm in our implementation used to evolve each subpopulation mirror those described by Gomez (2003). This algorithm has been shown to work well within the ESP framework, though any suitable evolutionary algorithm could potentially be substituted in its place. (Preliminary experiments using this approach were discussed in Braylan et al. (2015b).) Experiments GRUSM-ESP was evaluated in a stochastic version of the Atari 2600 general video game-playing platform using the Arcade Learning Environment simulator (ALE; Bellemare et al. 2013). Atari 2600 is currently a very popular platform, because it challenges modern approaches, contains non-markovian games, and entertained a generation of human video game players, who would regularly reuse knowledge gained from previous games when playing new games. To make the simulator more closely resemble the human game-playing experience, the -repeat action approach as suggested by Hausknecht and Stone (2015) is used in this paper to make the environment stochastic; in this manner, like human players, the algorithm cannot as easily find loopholes in the deterministic nature of the simulator. The recommended = 0.251 is used. Note that the vast majority of previously published Atari 2600 results are in the deterministic setting; we are unaware of any existing scores that have been published in the -repeat setting. Agents were trained to play eight games: Pong, Breakout, Asterix, Bowling, Freeway, Boxing, Space Invaders, 1 https://github.com/mgbellemare/Arcade-Learning-Environment/tree/dev and Seaquest. Neuroevolution techniques are competitive in the Atari 2600 platform (Hausknecht et al. 2013), and ESP in particular has yielded state-of-the-art performance for several games (Braylan et al. 2015a). Three GRUSM-ESP conditions are evaluated: scratch, transfer, and random. In the scratch condition, networks are trained from scratch on a game using standard ESP (GRUSM-ESP with S = ∅). In the transfer condition, each scratch network is reused as a source network in training new GRUSM networks for different target games. In the random control condition, random networks are initialized and reused as source networks. Such networks contain on average the same number of parameters as fully-trained scratch networks. Each run lasted 200 generations with 100 evaluations per generation. Since the environment is stochastic, each evaluation consists of five independent trials of individual i playing game g, and the resulting score s(i, g) is the average of the scores across these trials. The score of an evolutionary run at a given generation is the highest s(i, g) achieved by an individual by that generation. A total of 333 runs were run split across all possible setups. Evolutionary parameters were selected based on their success with standard ESP. To interface with ALE, the output layer of each network consists of a 3x3 substrate representing the nine directional movements of the Atari joystick in addition to a single node representing the Fire button. The input layer consisted of a series of object representations manually generated as previously described by Hausknecht et al. (2013). The location of each object on the screen was represented in an 8 × 10 input substrate corresponding to the object’s class. The numbers of object classes varied between one and four. Although object representations were used in these experiments, pixel-level vision could also be learned from scratch below the neuroevolution process, e.g., via convolutional networks as was done by Koutník, Schmidhuber, and Gomez (2014). Domain Characterization Understanding when transfer will be useful is important for any transfer learning approach. In many cases, attempting transfer can impede learning, leading to negative transfer, when an approach is not able to successfully adapt knowledge from the source to the target domain. Negative transfer is a serious concern for many practitioners (Taylor and Stone 2009; Pan and Yang 2010). To help understand when GRUSM-ESP should be applied, it is useful to consider the diverse array of games within a unified descriptive framework. Biological neural reuse is generally thought to be most useful in transferring knowledge from simple behaviors to more complex, and the vast majority of previous computational approaches do exactly that. Thus, the characterization of games in this paper is grounded by a sense of relative complexity. Each game can be characterized by generic binary features that determine what successful game play requires: (1) horizontal movement (joystick left/right), (2) vertical movement (joystick up/down), (3) shooting (fire button); (4) delayed rewards; and (5) long-term planning. Intuitively, more complex games will include more of these features. A partial ordering of games by complexity defined by these v h s d p none pong breakout asterix bowling freeway boxing s. invaders seaquest v = vertical movement h = horizontal movement s = shooting d = delayed rewards p = long-term planning v h pong d breakout s h v bowling freeway asterix h h,s,p s,p s boxing space invaders v,d d seaquest Figure 2: (left) Feature representation for each game, and (right) games partially-ordered by feature inclusion: every path from none to g contains along its edges each complexity feature of g exactly once, showing how games are related across the feature space. The existence of such a hierarchy motivates the use of atari for transfer. features is shown in Figure 2. The assignment of features (1), (2) and (3) is completely defined based on game interface (Bellemare et al. 2013). Freeway and Seaquest are said to have delayed rewards because a high score can only be achieved by long sequences of rewardless behavior. Only Space Invaders and Seaquest were deemed to require long-term planning (Mnih et al. 2015), since the long-range dynamics of these games penalize reflexive strategies, and as such, agents in these games can perform well with a low frequency decision-making (Braylan et al. 2015a). In addition to being intuitive, these features are validated below based on how well they characterize games by complexity and how well they predict successful transfer. Analysis Methods There are many possible metrics for evaluating success of transfer, depending on what kind of transfer is desired or expected. Learning curves are irregular across different games, as illustrated in Figure 3, which makes it difficult to choose a single metric that makes sense across all source-target pairs. Thus, the analysis is focused on a broad notion of transfer effectiveness (TE), which aggregates metrics such as jumpstart and max overall score, with a weighted approximation of area under the curve (Taylor and Stone 2009). Success of a setup is defined as the sum of the average score of that setup at a series of nonuniformly-spaced generations: [1, 10, 50, 100, 200]. This series favors early performance over later performance, as in general, in the long run, training from transfer and scratch should converge, as scratch eventually relearns everything that was effectively transferred. Then, the TE of a setup is its success minus the success of the control on the target game, the difference normalized by the size of the range of max scores achieved across all runs for that game, in order to draw comparison across games. The first hypothesis is that transfer would outperform scratch in some setups, and that those setups could be predicted (i.e., they are not coincidental). However, any outperformance of transfer over scratch could be due to a Figure 3: Raw mean score learning curves by generation for each target game aggregated over transfer (solid), random (dashed), and scratch (dotted) setups. The diversity of these learning curves shows the difficulty in comparing performance across games. larger number of network parameters. Therefore, as a second hypothesis, random setups were used as a control for the number of parameters, to test how transfer could predictably outperform random. We postulated and tested several useful indicators for predicting the outperformance of transfer, i.e., TE: (1) feature similarity: count of features that are 1 for both source and target); (2) source feature complexity: feature count of source game; (3) target feature complexity: feature count of target game; (4) source training complexity: source game average time to threshold; (5) target training complexity: target game average time to threshold, where the threshold for each game is the minimum max score achieved across all scratch runs for that game, and time to threshold is the average number of generations to reach this threshold. To predict TE, a linear regression model was trained in a leave-one-out cross-validation analysis. For each possible source-target pair (s, t), the model was trained on all pairs (s0 , t0 6= t) with TE as the dependent variable and the five indicators as the independent variables. Subsequently, the trained model was used to predict the TE of (s, t). Correlation between the actual and predicted TE across all test pairs was used to gauge the predictability of TE. This experiment was conducted identically for both transfer versus scratch and transfer versus random conditions. Results For both hypotheses, the indicator-based model proved to be a statistically significant predictor of transfer effectiveness in the test data: correlation R = 0.40 and pvalue < 0.0025 for transfer versus scratch; correlation R = 0.53 and p-value < 10−7 for transfer versus random (Figure 4). The strongest indicators for transfer versus scratch were target feature complexity and target training complexity, and for transfer versus random the strongest indicator was target feature complexity. The fact that more complex games are more successful targets should not be surprising. As noted before, in most transfer learning scenarios, only simple-to-complex transfer is considered. The ability to predict when GRUSM-ESP will work is an important tool when applying this method to boxing breakout bowling asterix space invaders pong Figure 4: Predicted vs. actual transfer effectiveness with respect to scratch (left) and random (right). Both predictors have a significant correlation between predicated and actual transfer effectiveness. game pong breakout asterix bowling freeway boxing s. invad. seaquest scratch 0.0 31.0 2800 249.3 31.4 93.9 1438.0 466.0 random 21.0 35.0 3216.7 265.0 31.5 92.7 1407.5 460.0 transfer 10.0ast 30.3box 3355bow 265.0f r 32.2brk 95.0sea 1655.0po 975.0sp human 9.3 31.8 8503 154.8 29.6 4.3 1652 20182 DQN 18.9 401.2 6012 42.4 30.3 71.8 1976 5286 Table 1: For each game, average scores for scratch, random, and transfer from best source (subscripted). Interestingly, the best source for each target is unique. We also show human and DQN scores (Mnih et al. 2015). Note: DQN uses deterministic ALE, so the most apt external comparison here may be to humans, who cannot deterministically optimize trajectories at the frame level. larger problems, and it is encouraging that the predictive indicator coincides with the ‘common sense’ expectations of transfer effectiveness in the current experiments. TE for all source-target pairs is visualized in Figure 5. Also, although it is difficult to compare to the deterministic Atari 2600 domain, Table 1 provides a comparison of GRUSM-ESP to recent results in that domain for context (Mnih et al. 2015). Discussion and Future Work The results show that GRUSM-ESP (an evolutionary algorithm for general transfer of neural network structure) can improve learning in Atari game playing by reusing previously developed knowledge. They also make it possible to characterize the conditions under which transfer may be useful. More specifically, the improvement in learning performance in the target domain depends heavily on the complexity of the target domain. The effectiveness of transfer in complex games aligns with the common-sense notion of hierarchical knowledge representation, as argued previously in transfer learning (Konidaris, Scheidwasser, and Barto 2012) as well as in biology (Anderson 2010; Milo et al. 2002). It will be interesting to investigate whether the same principles extend to other general video game playing platforms, such as VGDL (Perez et al. 2015; Schaul 2013). Such work should help understand how subsymbolic knowledge can be recycled indefinitely across diverse domains. freeway seaquest bowling asterix breakout boxing pong space invaders seaquest freeway Figure 5: Transferability graphs over all pairs of tasks with respect to scratch (top) and random (bottom) illustrating the target-centric clustering of successful source-target pairs. Each graph includes a directed edge from g1 to g2 ⇐⇒ the TE (see Analysis) for g2 reusing g1 is positive. Transfer is likely inefficient in simpler games due to the effort involved in finding the necessary connections for reusing knowledge from a given source network effectively, in which case it is more efficient to relearn from scratch. For particular low-complexity games, it can also be seen that random consistently outperforms both scratch and transfer (e.g., pong). The initial flexibility of untrained parameters in the random condition may explain this result. Unfreezing reused networks, and allowing them to change with a low learning rate may help close this gap. Some transfer pairs do not consistently outperform training from scratch or random, indicating negative transfer. This highlights the importance of source and target selection in transfer learning. These results have taken a step towards answering the target-selection problem: What kinds of games make good targets for transfer? More data across many more games is required to answer the source-selection problem: For a given game, what sources should be used? A next step will involve pooling multiple candidate sources and testing GRUSM-ESP’s ability to exploit the most useful structure available. Despite negative transfer in some of the setups, the technique of training a classifier to predict transfer success is shown to be a useful approach for helping decide when to transfer: given some space of complex disparate domains, try transfer with a subset of source-target pairs, and use the results to build a classifier to inform when to attempt transfer in the future. In this paper, domain-characterization features were provided, but domain-agnostic features could be learned from analysis of the networks and/or learning process; this is an interesting avenue for future work. Another area of future work involves increasing the flexibility in the combined architecture by (1) relaxing the requirement for all transfer connections to be input-to-hidden and output-to-output, (2) allowing deeper architectures for the source and target networks, and (3) including multiple source networks with adaptive connectivity to each. These extensions will promote reuse of subnetworks of varying depth, along with flexible positioning and combination of modules. However, for GRUSM-ESP, as networks become large and plentiful, maintaining full connectivity between layers will become intractable, and it will be necessary to enforcing sparsity. GRUSM-ESP can also be extended to include LSTM units, e.g., as by Schmidhuber et al. (2007), when deep memory is a primary concern. Conclusion This paper introduced an approach for general transfer learning using neural networks. The approach minimizes a priori assumptions of task relatedness and enables a flexible approach to adaptive learning across many domains. In a stochastic version of the Atari 2600 general video gameplaying platform, a specific implementation developed in this paper as GRUSM-ESP can boost learning by reusing neural structure across disparate domains. The success of transfer is shown to correlate with intuitive notions of domain complexity. These results indicate the potential for general neural reuse to predictably assist agents in increasingly complex environments. Acknowledgments We would like to thank Ruohan Zhang for useful feedback. This research was supported in part by NSF grant DBI-0939454, NIH grant R01-GM105042, and an NPSC fellowship sponsored by NSA. References Ammar, H. B.; Eaton, E.; Luna, J. M.; and Ruvolo, P. 2015a. Autonomous cross-domain knowledge transfer in lifelong policy gradient reinforcement learning. Proc. of IJCAI. Ammar, H. B.; Eaton, E.; Ruvolo, P.; and Taylor, M. E. 2015b. Unsupervised cross-domain transfer in policy gradient reinforcement learning via manifold alignment. Proc. of AAAI. Anderson, M. L. 2010. Neural reuse: A fundamental organizational principle of the brain. Behavioral and Brain Sciences 33:245–266. Bellemare, M. G.; Naddaf, Y.; Veness, J.; and Bowling, M. 2013. The arcade learning environment: An evaluation platform for general agents. JAIR 47:253–279. Braylan, A.; Hollenbeck, M.; Meyerson, E.; and Miikkulainen, R. 2015a. Frame skip is a powerful parameter for learning to play atari. Workshops at AAAI-15. Braylan, A.; Hollenbeck, M.; Meyerson, E.; and Miikkulainen, R. 2015b. On the cross-domain reusability of neural modules for general video game playing. IJCAI-15 GIGA Workshop. Brys, T.; Harutyunyan, A.; Taylor, M. E.; and Nowé, A. 2015. Policy transfer using reward shaping. Proc. of AAMAS 181–188. Gomez, F. J., and Miikkulainen, R. 1997. Incremental evolution of complex general behavior. Adaptive Behavior 5(3-4):317–342. Gomez, F. J., and Miikkulainen, R. 1999. Solving non-markovian control tasks with neuroevolution. Proc. of IJCAI 1356–1361. Gomez, F. J., and Schmidhuber, J. 2005. Co-evolving recurrent neurons learn deep memory pomdps. Proc. of GECCO 491–498. Gomez, F. J. 2003. Robust non-linear control through neuroevolution. Technical report, UT Austin. Hausknecht, M., and Stone, P. 2015. The impact of determinism on learning atari 2600 games. Workshops at AAAI-15. Hausknecht, M.; Lehman, J.; Miikkulainen, R.; and Stone, P. 2013. A neuroevolution approach to general atari game playing. IEEE Trans. on Comp. Intelligence in AI in Games 6(4):355–366. Khare, V. R.; Yao, X.; Sendhoff, B.; Jin, Y.; and Wersing, H. 2005. Co-evolutionary modular neural networks for automatic problem decomposition. Proc. of CEC 2691–2698. Konidaris, G.; Scheidwasser, I.; and Barto, A. G. 2012. Transfer in reinforcement learning via shared features. JMLR 13(1):1333– 1371. Koutník, J.; Schmidhuber, J.; and Gomez, F. J. 2014. Evolving deep unsupervised convolutional networks for vision-based reinforcement learning. Proc. of GECCO 541–548. Lukoševičius, M., and Jaeger, H. 2009. Reservoir computing approaches to recurrent neural network training. Computer Science Review 3(3):127–149. Milo, R.; Shen-Orr, S.; Itzkovitz, S.; Kashtan, N.; Chklovskii, D.; and Alon, U. 2002. Network motifs: Simple building blocks of complex networks. Science 298(5594):824–827. Mnih, V.; Kavukcuoglu, K.; Silver, D.; Rusu, A. A.; Veness, J.; Bellemare, M. G.; et al. 2015. Human-level control through deep reinforcement learning. Nature 518(7540):529–533. Pan, S. J., and Yang, Q. 2010. A survey on transfer learning. IEEE Trans. on Knowledge and Data Engineering 22(10):1345–1359. Perez, D.; Samothrakis, S.; Togelius, J.; Schaul, T.; et al. 2015. The 2014 general video game playing competition. IEEE Trans. on Comp. Intel. and AI in Games (99). Reisinger, J.; Stanley, K. O.; and Miikkulainen, R. 2004. Evolving reusable neural modules. Proc. of GECCO 69–81. Schaul, T. 2013. A video game description language for modelbased or interactive learning. Proc. of CIG 1–8. Schmidhuber, J.; Wierstra, D.; Gagliolo, M.; and Gomez, F. J. 2007. Training recurrent networks by evolino. Neural Computation 19(3):757–779. Shultz, T. R., and Rivest, F. 2001. Knowledge-based cascadecorrelation: Using knowledge to speed learning. Connection Science 13(1):43–72. Swarup, S., and Ray, S. R. 2006. Cross-domain knowledge transfer using structured representations. Proc. of AAAI 506–511. Talvitie, E., and Singh, S. 2007. An experts algorithm for transfer learning. Proc. of IJCAI 1065–1070. Taylor, M. E., and Stone, P. 2009. Transfer learning for reinforcement learning domains: A survey. JMLR 1633–1685. Taylor, M. E.; Kuhlmann, G.; and Stone, P. 2008. Autonomous transfer for reinforcement learning. Proc. of AAMAS 283–290. Taylor, M. E.; Whiteson, S.; and Stone, P. 2007. Transfer via intertask mappings in policy search reinforcement learning. Proc. of AAMAS 156–163. Verbancsics, P., and Stanley, K. O. 2010. Evolving static representations for task transfer. JMLR 11:1737–1769.	9cs.NE
arXiv:1705.04910v2 [math.GR] 23 Aug 2017 Discretization of SU(2) and the orthogonal group using icosahedral symmetries and the golden numbers Robert V. Moody & Jun Morita August 25, 2017 Department of Mathematics and Statistics, University of Victoria, Victoria, BC., V8W 3R4 Canada Department of Mathematics, University of Tsukuba, Tsukuba, Ibaraki, 305-8571, Japan. E-mail: rvmoody@mac.com, morita@math.tsukuba.ac.jp Abstract The vertices of the four-dimensional 600-cell form a non-crystallographic root system whose corresponding symmetry group is the Coxeter group H4 . There are two special coordinate representations of this root system in which they and their corresponding Coxeter groups involve only rational numbers and the golden ratio τ . The two are related by the conjugation τ 7→ τ 0 = −1/τ . This paper investigates what happens when the two root systems are combined and the group generated by both versions of H4 is allowed to operate on them. The result is a new, but infinite, ‘root system’ Σ which itself turns out to have a natural structure of the unitary group SU (2, R) over the ring R = Z[ 21 , τ ] (called here golden numbers). Acting upon it is the naturally associated infinite reflection group H ∞ , which we prove is of index 2 in the orthogonal group O(4, R). The paper makes extensive use of the quaternions over R and leads to a highly structured discretized filtration of SU (2). We use this to offer a simple and effective way to approximate any element of SU (2) to any degree of accuracy required using the repeated actions of just five fixed reflections, a process that may find application in computational methods in quantum mechanics. Keywords: icosahedral symmetry, non-crystallographic root systems, infinite reflection groups, discretization of SU (2), golden numbers, quaternions 1 Introduction The symmetry group of the icosahedron (and the dodecahedron) is the icosahedral group, denoted H3 here, with 120 elements. It is a finite Coxeter group, that is to say, it is a finite group generated by reflections and Coxeter relations, and it is simply transitive on the simplicial cells into which the icosahedron is partitioned by the mirrors of the reflections in H3 . Apart from numerous dihedral groups, there are only two finite indecomposable non-crystallographic Coxeter groups: H3 and its big sister H4 , which is the symmetry group of the four-dimensional regular polytope called the 600-cell (and its dual, the 120-cell). This paper involves both these groups, but notably the latter. The 600-cell has 120 vertices and 600 three-dimensional faces, see Fig. 1. The 120 vertices form a root system, of type H4 . With this interpretation of the vertices as roots, the set of 60 reflections in these roots (opposite roots give the same reflection) is the entire set of reflections in the group 1 Figure 1: A two-dimensional projection of the 600-cell, whose vertices form a root system of type H4 . The dots make up the 120 vertices that constitute a root system of type H4 . H4 , and they generate it. The order of the group is \|H4 \| = 1202 = 14400. Notice that we use the symbol H4 both as an adjective signifying the type of root system involved and as a noun signifying the reflection group generated by the roots of that system.√We will do the same thing with H3 . 0 Both √ H3 and H4 involve the golden ratio τ = (1 + 5)/2 and its algebraic conjugate τ = (1 − 5)/2 in all sorts of significant ways. For instance, there is the well-known model of the vertices of the 600-cell [5] §20 and §27, or equivalently the roots of the root system of type H4 , as the set of points    R[0,0] (±1, 0, 0, 0) and all its permutations: 8 roots in all; ∆ = R[1,0] 12 (±1, ±1, ±1, ±1): 16 roots in all; (1)   1 0 R[1,1] 2 (0, ±1, ±τ , ±τ ) and all its even permutations: 96 roots in all. Here the first component of the label refers to the power of 2 appearing in the denominators of the components and the second component is used to distinguish the rational versus irrational nature of (some of) the components. One cannot help noticing the interesting fact that only even permutations are allowed in the third type of root. Half of the potential permutations are missing. The other half can be obtained by conjugating these roots by interchanging τ and τ 0 throughout. Of course doing so produces another model ∆0 of the root system of type H4 , and the reflections in these generate a second model H40 of the Coxeter group H4 . We shall also use the notation H4 = W (∆) and H40 = W (∆0 ), that suggests the origin of these groups as reflection groups. 2 It is tempting to look at the group generated by the reflections of ∆ ∪ ∆0 , but one quickly realizes that this group, let’s call it H ∞ (or H ∞ [4]) is infinite.1 Since all the points of the set Σ = Σ[4] := H ∞ (∆ ∪ ∆0 ) (2) lie on the sphere of radius 1 in R4 , which is compact, Σ is certainly not a discrete set. Up to now no one seems to have paid much attention to it. The objective of this paper is to get some idea of what Σ and H ∞ look like. In fact they have some very attractive features, as we shall see. Not surprisingly Σ is a dense set of points on the unit 3-sphere S in R4 , see Prop. 3.7. It is also a group under quaternion multiplication, and we shall show that it is an amalgamation of ∆ and ∆0 , each of which is itself a group. Similarly H ∞ is an amalgamation of the two groups H4 and H40 . What is especially interesting is that using Σ and H ∞ allows us to explicitly approximate elements of the two groups SU (2) and SO(3), the unitary and special orthogonal groups of C2 and R3 respectively, by elements with matrix coefficients of the form 1 2n Z[τ ] (we call them dyadic integers from the field Q[τ ]). In fact the matrices used all arise from the 96 reflections of ∆ ∪ ∆0 which are the ones involving τ and τ 0 , and even from just 5 reflections if one gets down to the level of root system bases. The approximation can be made as fine as one wishes (with increasing powers of two in the denominators) and there is a simple and efficient algorithm for doing it. The key to all of this is to interpret R4 as the standard division ring of quaternions, and use the fact that its unit sphere can be identified with SU (2) and that it can be made to act as SO(3) on the three-dimensional subspace of pure imaginary quaternions (those which change their sign under quaternionic conjugation). The paper is primarily concerned with the H4 picture, but we also require information from the corresponding, and essentially parallel situation based on the three-dimensional icosahedral root system H3 and its corresponding conjugate system H30 . The arguments are not identical but sufficiently parallel that we provide only sketches of the arguments for the three-dimensional case. The paper is organized around understanding the structure of the root system Σ and the group ∞ H that the reflections in its roots generate. The approximation results appear in the final section, though they can be read directly after §3. Icosahedral symmetry and the Coxeter groups H2 , H3 , H4 have continued to intrigue people from ancient times until the present day, where they are now familiar in such diverse places as carbon molecules, buckyballs, the capsids (outer shells) of viruses, Penrose tilings, and the aperiodic order of quasicrystals. There remains continuing interest in the mathematical world too, for instance, [6] which investigates the subgroup structure of H4 in a quaternionic context, and [3] which explores some of the ways of making affine extensions of the H4 root systems, following the success of affinization of the root systems of the simple Lie algebras. Affinization is accomplished by extending the Coxeter diagram, and might be thought of as based on using translations to extend root systems. In the present paper, although the root systems involved are defined intrinsically from the H3 and H4 root systems and the Galois conjugation τ 7→ τ 0 , effectively the result is to extend by using contractions. Indeed this contractive aspect is something seems worthy of fuller study. We might note here that there have been considerable advances in the understanding of Coxeter groups along with their associated geometrical meanings by using the Clifford algebras, see [4] and its associated references. Although we have not used these ideas here, and indeed we require only a 1 Throughout the paper we will see parallel structures pertaining to the three-dimensional setting around H3 and the four-dimensional setting around H4 . Generally we use the same symbol for the corresponding objects, but if the context does not make it clear then we shall make the distinction as in H ∞ [3], H ∞ [4], Σ[3], Σ[4], S[3], S[4], etc. 3 very limited number of tools, the Clifford algebra approach may offer new insights into the setting we are introducing. 2 H4 The set ∆ of roots of the root system of type H4 can be presented in the standard form of (1) as the union of three types of vectors in R4 . We let K = K0 ∪ K1 denote the set of roots of types R[0,0] ˙ for the set of roots of and R[1,0], which form a crystallographic root system of type D4 , and ∆ type R[1,1]. Of course these distinctions are not intrinsic to the root system itself, but only to our coordinatization of it. However this distinction plays a fundamental role in what is to follow. Along with ∆ we have its conjugate ∆0 which is the set of conjugates of ∆, and the corresponding ˙ 0 (it being irrelevant whether or not the dot operators are applied before or after sets K and ∆ conjugation). K = ∆ ∩ ∆0 . The reflections in the roots ∆ generate the group we call H4 . It is a Coxeter group of type H4 . Similarly we have H40 generated by the reflections in the roots of ∆0 , also of type H4 . We are primarily interested in the group H ∞ which is generated by H4 and H40 together. The reflections given by K, the roots in common to both systems, generate a subgroup D of type D4 For example (1, 0, 0, 0), 1 (−1, −1, −1, ±1), (0, 1, 0, 0), (0, 0, 1, 0), 2 (3) is a base (with either choice of the sign). We let S = S[4] denote the unit sphere in 4-space. Both ∆ and ∆0 are in S, and so too is the set Σ = Σ[4] := H ∞ (∆ ∪ ∆0 ). It is the smallest subset containing ∆ ∪ ∆0 and closed under its own reflections (i.e. if a ∈ Σ and ra is the reflection in a then ra (Σ) = Σ). We note the important fact that 1 Z[τ ]4 ∩ S = ∆ ∪ ∆0 . 2 (4) At the start we use only the algebraic consequences of reflections applied in the context of Z[τ ]4 and R4 , but later we will interpret everything in the real quaternions, where the elements that we are discussing can be interpreted as elements of SU(2). It is useful to keep in mind the basic facts about τ : τ + τ 0 = 1, τ τ 0 = −1, τ 2 = τ + 1, τ 02 = τ 0 + 1 . (5) The quotient ring F4 := Z[τ ]/2Z[τ ] is the Galois field of 4 elements. Let a 7→ a denote the natural homomorphism of Z[τ ] −→ F4 . The elements of F4 are 0, 1, τ , τ 0 . We extend the meaning of the map · to its component-wise version Z[τ ]4 −→ F44 . Inside the F4 -vector space F44 , define A to be the subspace spanned by the elements of 2∆ taken modulo 2Z[τ ]4 , i.e., 2∆. This is a 2-dimensional space with basis (1, 1, 1, 1) and (0, 1, τ 0 , τ ). We define ˙ Note the cardinalities: card (A) = 16, card (Å) = 15, and card (Ȧ) = 12. Å := A\{0} and Ȧ := 2∆. ˙ 0 . As we have already Similarly we have A0 using 2∆0 and the corresponding subset Ȧ0 = 2∆ mentioned, and this sentence suggests, and will be true throughout, everything we say will come in two versions, which are interchanged by conjugation. Henceforth we will usually only state and give proofs for one of the versions, understanding that the other will be equally true. 4 5 1 5 Figure 2: This is the Coxeter diagram for the pairwise relationships between the reflection generators of H ∞ [4]. The bonds between any two nodes indicate the orders of the product of the reflections corresponding to them, with the convention that unmarked bonds indicate order 3 and bonds between commuting reflections are omitted completely. The proof establishing the bond marked with ∞ is found in Cor. 3.6. We shall see that in fact H ∞ [4] is a proper factor of the Coxeter group of this diagram. The corresponding Coxeter diagram for the relationships between the generators of H ∞ [3] is obtained by deleting the first node and its attached bond. With the obvious dot product on F4 we find that A is totally isotropic. This is just another way ˙ 2α.2β ∈ 2Z[τ ]. In addition, if α and β are linearly dependent, of saying that for all α, β ∈ K1 ∪ ∆, then 2α.2β ∈ 4Z[τ ]. This simply follows from (0, 1, τ 0 , τ ).(0, 1, τ 0 , τ ) = 4 = (1, 1, 1, 1).(1, 1, 1, 1). We have parallel statements for A0 and Ȧ0 . However, for any a ∈ Ȧ and b ∈ Ȧ0 we always have a.b 6= 0. Proposition 2.1. (i) A.A = 0. In particular, for all x, y ∈ ∆, 2x.y ∈ Z[τ ]; ˙ and y ∈ ∆ ˙ 0 then 2x.y ∈ (ii) for all a ∈ Ȧ and b ∈ Ȧ0 , a.b 6= 0. In particular if x ∈ ∆ / Z[τ ]. 3 H ∞ as a reflection group The set of vectors 1 1 1 {a1 , a2 , a3 , a4 , a5 } = {(1, 0, 0, 0), − (1, 1, 1, 1), (0, 0, 1, 0), (0, −1, τ 0 , τ ), (0, 1, −τ, −τ 0 )} (6) 2 2 2 define reflections that correspond to the Coxeter diagram of Fig. 2 and these reflections generate H ∞ = H ∞ [4]. Indeed they contain a set of generators for both of the Coxeter groups H4 and H40 . The associated Gram matrix with entries ai .aj is positive semidefinite with null vector (3 − τ, 6 − 2τ, 9 − 3τ, −6 + 4τ, 2τ ). Certainly H ∞ is a factor of the corresponding Coxeter group, and later on we shall see that it is a proper factor, see Prop. 8.1. Proposition 3.1. If x = (x1 , x2 , x3 , x4 ) ∈ Z[τ ]4 satisfies x21 + x22 + x23 + x24 ≡ 0 mod 4Z[τ ] then one of the three following cases prevails:   all the xi are congruent to 0 modulo 2Z[τ ], all the xi are congruent modulo 2Z[τ ] to each other, but not to 0,   (x1 , x2 , x3 , x4 ) is congruent modulo 2Z[τ ]4 to a permutation of the components of (0, 1, τ 0 , τ ). (7) 5 In the language and notation introduced above, these three cases correspond to   x = (0, 0, 0, 0) x ∈ Å\(Ȧ ∪ Ȧ0 ) = Å0 \(Ȧ ∪ Ȧ0 )   x ∈ Ȧ or x ∈ Ȧ0 . (8) Proof: Modulo 2Z[τ ] each xi ≡ 0, 1, τ 0 , τ . Their squares are {0, 1, 1 + τ 0 , 1 + τ } modulo 4Z[τ ]. A quick check shows that the sum of four such elements can be zero modulo 4Z[τ ] if and only if one of the three conditions of the thesis is true. Examples of all three types occur when we double each of the roots, i.e. when we look at 2(∆ ∪ ∆0 ). Corollary 3.2. The three-dimensional version of Prop. 3.1 reads: if x = (x1 , x2 , x3 ) ∈ Z[τ ]3 satisfies x21 + x22 + x23 ≡ 0 mod 4Z[τ ] then one of the two following cases prevails: ( all the xi are congruent to 0 modulo 2Z[τ ], (x1 , x2 , x3 ) is congruent modulo 2Z[τ ]3 to a permutation of the components of (1, τ 0 , τ ). (9) For n = 1, 2, 3 . . . define the spherical sets 1 Sn := x ∈ n Z[τ ]4 : x ∈ S, and 2n x ∈ Å , 2 and Ṡn := 1 4 n x ∈ n Z[τ ] : x ∈ S, and 2 x ∈ Ȧ ; 2 and similarly the conjugated versions using A0 . We call n here the level of the spherical set. We define Kn = Kn0 := Sn ∩ Sn0 , so Sn = Ṡn ∪ Kn , and similarly Sn0 = Ṡn0 ∪ Kn . To these we add S0 = S00 := K0 , the set of roots of type R[0,0]. Here are two consequences of Prop. 3.1: ˙ and Ṡ1 = ∆. ˙ Proposition 3.3. S1 = K1 ∪ ∆, Proof: Let y ∈ S1 . Then 2y ∈ Å and 2y = (y1 , y2 , y3 , y4 ) ∈ Z[τ ]4 satisfies y12 + y22 + y32 + y42 = 4. Using Prop. 3.1, one can hunt through all solutions to see that only elements of ∆ ∪ ∆0 can satisfy ˙ this, and then restrict to Ȧ filter them out to ∆. 6 Proposition 3.4. For n ≥ 1, [ 1 0 4 (Sm ∪ Sm ). Z[τ ] ∩ S = 2n m≤n Furthermore, x ∈ 1 4 2n Z[τ ] ∩ S is in Sn ∪ Sn0 if and only if 2n x 6= {0, 0, 0, 0}. Proof: Let x ∈ 21n Z[τ ]4 ∩ S, and let m ≤ n be minimal so that 2m x ∈ Z[τ ]4 . If m = 0 then x ∈ K0 . 0 . If m ≥ 1 then 2m x.2m x = 4m and by Prop. 3.1, 2m x ∈ Å ∪ Å0 , see (8), from which x ∈ Sm ∪ Sm The reverse inclusion comes from the definitions, and the last line is clear. ˙ 0 , rα0 Ṡn ⊂ Ṡ 0 . Proposition 3.5. For all α0 ∈ ∆ n+1 Proof: Let x ∈ Ṡn , so by definition 2n x ∈ Z[τ ]4 and 2n x ∈ Ȧ. We have 2n rα0 x = 2n x − (2n x.2α0 )α0 . The elements 2n x and 2α0 are in Ȧ and Ȧ0 respectively, so in all cases (2n x.2α0 ) 6= 0. Since α0 is not in Z[τ ]4 , but only in (1/2)Z[τ ]4 , 2n rα0 x = 2n x − (2n x.2α0 )α0 ∈ / Z[τ ]4 . However 2n+1 rα0 x = n+1 n 0 0 4 0 2 x − (2 x.2α )2α ∈ Z[τ ] and furthermore 2n+1 rα0 x = (2n x.2α0 )2α0 ∈ Ȧ0 . Thus rα0 (x) ∈ Ṡn+1 , which is what we wished to prove. ˙ and for all β 0 ∈ ∆ ˙ 0 , rα rβ 0 has infinite order. Corollary 3.6. For all α ∈ ∆ Proof: Let x ∈ Ṡ1 . Then rβ 0 x ∈ Ṡ20 , rα rβ 0 x ∈ Ṡ3 , rβ 0 rα rβ 0 x ∈ Ṡ4 and so on. ∞ [4] denote the commutator subgroup of H ∞ [4]. Let H+ Proposition 3.7. ∞ [4] is dense in SO(4); (i) H+ (ii) H ∞ [4] is dense in O(4); (iii) Σ[4] is dense in S[4]. ˙ and β 0 := a5 = (1/2)(0, 1, −τ, −τ 0 ) ∈ ∆ ˙ 0 , rα rβ 0 is a Proof: With α := a4 = (1/2)(0, −1, τ 0 , τ ) ∈ ∆ 1 rotation in the plane P orthogonal to the two roots (1, 0, 0, 0) and 2 (1, 1, 1, 1), and by Cor. 3.6 it ∞ [4] be the infinite cyclic group generated by (r r 0 )2 = r r 0 r r 0 ∈ has infinite order. Let G1 ⊂ H+ α β α β α β ∞ H+ [4]. The orbit of α under G1 generated is infinite and hence is dense on the circle on the sphere S ∩ P . In particular G1 has rotations through angles as small as we please. ˙ 0 , so rα rγ 0 is a rotation of infinite order in a Now also γ 0 := ra3 (a5 ) = (1/2)(0, 1, τ, −τ 0 ) ∈ ∆ ∞ [4] with arbitrarily small different plane orthogonal to (1, 0, 0, 0) and we get a second group G2 ⊂ H+ ∞ [4] that is a dense subgroup of the group of rotations. These two groups generate a subgroup of H+ ∞ SO(3)1 of rotations of (0, R, R, R). Since H+ [4] can certainly map (0, R, R, R) to (R, R, R, 0), we see that it also contains a subgroup that is a dense subgroup of the SO(3)4 of rotations of this space. ∞ [4] contains a Since SO(3)1 and SO(3)4 generate the entire rotation group SO(4), we see that H+ dense subgroup of this space. Now (i),(ii),(iii) are all clear. Proposition 3.8. Let n ≥ 2. Then ˙ rα (Sn ) ⊂ Sn ∪ Ṡ 0 ; (i) for all α ∈ ∆, n−1 7 ˙ and for each x ∈ Sn \Ṡn there exists α ∈ ∆ ˙ for which rα (x) ∈ Ṡn ; (ii) rα (Sn \Ṡn ) ⊂ Sn for all α ∈ ∆, ˙ for which rα x ∈ Ṡ 0 . (iii) for each x ∈ Ṡn there are three reflections of the form rα where α ∈ ∆, n−1 ˙ Proof: (i). We start with the slightly weaker assumption that n ≥ 1. Let x ∈ Sn and α ∈ ∆. 2n rα x = 2n x − (2n x.2α)α . (10) Let 2n x = b ∈ Å. Since 2n x and 2α are elements of A, (2n x.2α) ∈ 2Z[τ ]4 . Let (2n x.2α)α = a ∈ A. As long as a 6= b, 2n rα x = b − a ∈ Å, so rα x ∈ Sn . If a = b then 2n rα x ∈ 2Z[τ ]4 , so 2n−1 rα x ∈ Z[τ ]4 . 0 for some m < n. By Prop. 3.4, rα x is either in Sm or Sm Now whatever the case is, if n ≥ 1 then we have shown that rα (x) is either in Sn or in some Sm 0 for m < n. Now assume that n ≥ 2. Then if m < n it is not possible that r (x) ∈ S , for or Sm α m either m ≥ 1 in which case by the same reasoning, x = rα rα (x) is on some sphere of level at most 0 , m, contrary to x ∈ Sn , or m = 0 in which case rα rα (x) has level 1. In the case that rα (x) ∈ Sm 0 0 0 then noting Prop. 3.5 we see that actually rα (x) ∈ Sn−1 . But in fact rα (x) cannot be in Sn−1 \Ṡn−1 , since these elements do not change level under the application of rα and so x = rα rα (x) would be 0 impossible. Thus rα (x) ∈ Ṡn−1 This proves part (i). This last case, where the level drops under reflection, is particularly interesting and we wish to look more carefully how it can happen in (iii) below. ˙ In the notation of part (i) above, if x ∈ Sn \Ṡn then b := 2n x ∈ F× (1, 1, 1, 1). (ii). Let α ∈ ∆. 4 Now if a := (2n x.2α)α 6= 0 then a ∈ Ȧ, and then also b−a ∈ Ȧ and the resulting rα x ∈ Ṡn . However, ˙ if a = 0 then b − a = b and rα (x) remains in ∈ Sn \Ṡn . We need to show that we can choose α ∈ ∆ for which the corresponding a 6= 0. For definiteness (the other cases work the same way) we assume that b = (1, 1, 1, 1) and write n 2 x = b + 2v, where v = (v1 , v2 , v3 , v4 ) ∈ Z[τ ]. We begin with α = 21 (0, 1, τ 0 , τ ), though ultimately we will cycle its last three components around. Then the question is what does it mean for a = (2n x.2α)α = 0 (11) to hold, or quivalently, what does it mean for 2n x.2α = 2 + 2v.2α ≡ 0 mod 4Z[τ ]? The latter is equivalent to 1 + v.2α ≡ 0 mod 2Z[τ ], or finally v2 + τ 0 v3 + τ v4 = 1 . This may happen, but if we cycle the last three terms of α around then we get the same equation but with the coefficients v2 , v3 , v4 likewise cycled around. Adding all three of the equations together and using the fact that 1 + τ 0 + τ = 0 we get the contradiction 0 = 1. Thus, for at least one of these three possibilities for α, equation (11) fails, and this version of α produces rα (x) ∈ Ṡn . (iii). Suppose that x ∈ Ṡn and let α ∈ Ȧ. What is required to have 2n rα x = 0 ? To see what happens here, it is easiest to fix one particular form that 2n x can take, and we choose (0, 1, τ 0 , τ ). All other forms are even permutations of this, and as will be seen, there is no loss in generality in ˙ 2α is some even permutation of (0, 1, τ 0 , τ ). Referring to assuming this form. Now for any α ∈ ∆, (10), the question becomes, how can (2n x.2α)α = a = (0, 1, τ 0 , τ ) ? Looking at the 0 in the first coordinate of 2n x, it is clear that α must be one of the elements, (0, ±1, ±τ 0 , ±τ ), (0, ±τ 0 , ±τ, ±1), or (0, ±τ, ±1, ±τ 0 ). Since an overall sign change in α makes no difference to the reflection defined by α nor to the equation we are trying to solve, we can replace the ±τ 0 in each of these by simply τ 0 . Doing this we find there is exactly one solution in each of the three sets. 8 For instance, taking the second type we have (2n x.2α)α =  mod 2   0(0, τ 0 , τ, 1) −→ (0, 0, 0, 0)    mod 2 1(0, τ 0 , −τ, 1) −→ (0, τ 0 , τ , 1) 1 (0, 1, τ 0 , τ ).(0, τ 0 , uτ, v1) (0, τ 0 , uτ, v1) = mod 2  2 −τ (0, τ 0 , τ, −1) −→ (0, 1, τ 0 , τ )     mod 2 τ 0 (0, τ 0 , −τ, −1) −→ (0, τ , 1, τ 0 ) (12) as u, v go through the various possibilities of sign, so u = 1, v = −1 gives the value we are looking for. In fact the other two solutions are just obtained by cyclically permuting the last three coordinates, so the solutions in this case are (0, τ 0 , τ, −1), (0, τ, −1, τ 0 ), (0, −1, τ 0 , τ ) . (13) The three solutions, and only these three, produce the desired effect that 2n rα (x) ∈ 2Z[τ ]4 , and 0 as a consequence, as we saw above, rα (x) ∈ Ṡn−1 . This concludes the proof of part (iii). It is worthwhile noting that the three vectors we have found in (13) are all in the plane orthogonal to the roots (1, 0, 0, 0) and 21 (1, 1, 1, 1) and they lie at angles of 2π/3 to each other. Let ∞ [ 1 a + bτ 1 R := Z[τ ] = : a, b ∈ Z, n ∈ Z≥0 = Z[ , τ ] . n n 2 2 2 (14) n=0 This is a ring which we shall call the ring of dyadic golden numbers. Proposition 3.9. Σ[4] = H ∞ (∆ ∞ [ (Sn ∪ Sn0 ) = S[4] ∩ R4 . n=0 ∪ ∆0 ). Proof: By definition (2), Σ = Let Σn := Sn ∪ Sn0 for all n = 0, 1, 2, . . . . We have 0 Σ0 ∪ Σ1 = ∆ ∪ ∆ ⊂ Σ. Now proceed by induction and assume that we have shown that Σn ⊂ Σ. 0 Then from Prop. 3.5 and Prop. 3.8(iii), Ṡn+1 ∪ Ṡn+1 ⊂ Σ. By Prop. 3.8(ii) we see that all of Σn+1 ⊂ Σ. Thus ∞ [ (Sn ∪ Sn0 ) ⊂ Σ . n=0 However Prop. 3.5 and Prop. 3.8 show that the left-hand side here is closed under the action of ˙ ∪∆ ˙ 0 , and these reflections generate all of H ∞ . This gives the reverse the reflections in the roots ∆ inclusion. For the final equality see Prop. 3.4. Proposition 3.10. For all n ≥ 1, card (Ṡn ) = card (Ṡn0 ) = 3.(2)4n+1 . ˙ and similarly for Ṡ 0 . Now proceed by induction. We Proof: Ṡ1 consists of the 96 roots of ∆, 1 ˙ 0 that will simply have to note that for each element of Ṡn there are 48 reflections arising from ∆ map it into 48 images in Ṡn+1 , see Prop. 3.5. However, according to Prop. 3.8(c) each of these ˙ 0 . Thus elements will be produced exactly three times as we use all the reflections available from ∆ 0 card Ṡn+1 = 16 card Ṡn , and likewise for card Ṡn+1 . This completes the induction step. The consequence of Prop. 3.8 and Prop. 3.5 is that to go deeper into the structure of Σ we ˙ and ∆ ˙ 0 . It is this path that leads to finer and finer need to apply reflections alternately from ∆ discretization of the unitary group. The situation that we have just described is pictorially represented in Fig. 3. 9 Ṡn+1 Kn+1 ˙0 1 0 Ṡn0 ˙ ˙0 Ṡn 0 ˙ Kn Ṡn 0 Ṡn+1 Kn 1 Ṡn0 1 0 Figure 3: The action of reflections from different types of roots is indicated diagramatically. The intersection Sn ∩ Sn0 =: Kn is the common ‘crystallographic’ part of the two spherical sets. The lateral actions of ∆ apply to Sn but not to Sn0 , and similarly with ∆0 . 4 H3 The four-dimensional theory that we have outlined so far has a parallel version in three dimensions around the Coxeter group H3 . Most of the features that we have described in the four-dimensional case appear again here, but now notationally distinguished with the additional suffix ‘[3]’. When we are dealing with both settings together, we shall also use the suffix ‘[4]’ for the four-dimensional situation in order to avoid confusion. The unit sphere in (1/2)Z[τ ]3 consists of 54 points, namely the 6 vectors (±1, 0, 0) and all permutations of its coordinates, and the 48 points (1/2)(±1, ±τ 0 , ±τ ) and all permutations of these coordinates. If in the second set of points we restrict them by allowing only even (= cyclic) permutations then they provide 24 points which together with the original 6, produce a set of 30 points which may be viewed as the non-crystallographic root system ∆[3] of type H3 . They are the vertices of the icosadodecahedron. The reflection group generated by the 15 distinct reflections in these points, now called roots, that is the reflections in the mirrors (planes) through the origin orthogonal to these roots, is the icosahedral group H3 of order 120. As a base for the root system of type H3 we can use the vectors (1/2)(1, τ 0 , τ ), (1/2)(1, −τ 0 , −τ ), (1/2)(τ 0 , τ, 1). Instead of the even permutations of (1/2)(±1, ±τ 0 , ±τ ) we can conjugate them, thus taking only all the even permutations of (1/2)(±1, ±τ, ±τ 0 ) (or equivalently we take all the odd permutations) and do the same thing. We then end up with ∆[3]0 , the set of conjugates of the elements of ∆[3]. Again we have a root system of type H3 and again the reflections in these roots generates a copy of the icosahedral group. ˙ ˙ 0. We write ∆[3] for the non-rational roots of ∆[3], and similarly for ∆[3] Let H ∞ [3] denote the group generated by the reflections in ∆[3] ∪ ∆[3]0 and let Σ[3] be the orbit 10 of ∆[3] ∪ ∆[3]0 under H ∞ [3]. We shall see that Σ[3] is a dense set of points on the unit 2-sphere S[3]. We put this into the 4-dimensional setting in which we have discussed H4 by considering the 3-dimensional space here as the subspace V := (0, R, R, R). Then it is clear that ∆[3] = ∆[4] ∩ V and ∆[3]0 = ∆[4]0 ∩ V , and indeed all of Σ[3], lies in V ∩ Σ[4]. In fact we shall see that Σ[3] = V ∩ Σ[4]. Later on we shall put everything into the setting of the algebra of quaternions H, at which point V will be come the subspace of pure quaternions. As before, let F4 := Z[τ ]/2Z[τ ] and let x 7→ x denote the natural homomorphism of Z[τ ] −→ F4 . In place of F44 we will now wish to use F34 . It is often convenient to identify this with the subspace (0, F4 , F4 , F4 ) of F44 , and we will do this when it is useful to do so. Using this we define the analogues of A[4] and A[4]0 , namely A[3] := A[4] ∩ V , A[3]0 := A[4]0 ∩ V (15) which will play the same role as A and A0 in the four-dimensional case. Written as elements of F34 A[3] = {(0, 0, 0), (1, τ 0 , τ ), (τ 0 , τ , 1), (τ , 1, τ 0 )} , and similarly for A[3]0 . Define Ȧ[3] := Ȧ[4] ∩ V = Å[4] ∩ V , and similarly Ȧ[3]0 . The dot product on F34 is defined by u.v = u.v for all u, v ∈ Z[τ ]3 , which matches the dot product it has as a subspace of F44 . From Prop. 2.1 we have: Proposition 4.1. (i) A[3].A[3] = 0. In particular, for all α, β ∈ ∆[3], 2α.β ∈ Z[τ ]; ˙ ˙ 0 then 2α.β ∈ (ii) for all a ∈ Ȧ[3] and b ∈ Ȧ0 [3], a.b 6= 0. In particular if α ∈ ∆[3] and β ∈ ∆[3] / Z[τ ]. We now obtain very similar results to those of §3. We mainly just list these, the proofs being basically the same. However, there is one significant difference that appears immediately—there are no longer two spherical sets to deal with. For n ≥ 1, we define 1 Sn [3] := x ∈ n Z[τ ]3 : x ∈ S[3], and 2n x ∈ Ȧ[3] , 2 and similarly Sn [3]0 using A0 . We call n here the level of the spherical set. The sets Kn and Kn0 have no relevant counterparts if n > 0 since Sn [3] ∩ Sn [3]0 = ∅ . To these we add K0 [3] := S0 [3] = S0 [3]0 = {±1, 0, 0), (0, ±1, 0), (0, 0, ±1)} . It is clear that S[3] = S[4] ∩ V and Sn [3] = Sn [4] ∩ V , and parallel statements for S 0 . As a consequence of Cor. 3.2: Proposition 4.2. 11 ˙ (i) S1 [3] = ∆[3]; (ii) for all n ≥ 0, [ 1 0 3 (Sm [3] ∪ Sm [3]) . Z[τ ] ∩ S[3] = 2n m≤n Furthermore, x ∈ 1 3 2n Z[τ ] ∩ S[3] is in Sn [3] ∪ Sn0 [3] if and only if 2n x 6= {0, 0, 0}. For each α ∈ ∆[3], we also have α ∈ ∆[4] and the reflection in α taken in V is just the restriction of the reflection rα taken in R4 restricted to V . We use the same symbol rα for both, with context distinguishing them if necessary. Corresponding primed notation will be used for ∆[3]0 . H3 , H30 denote the groups generated by the reflections of ∆[3] and ∆[3]0 respectively. We may think of these groups as subgroups of H4 and H40 . From Prop. 3.5 we have ˙ 0 , rα0 Sn [3] ⊂ S 0 [3]. Proposition 4.3. For all α0 ∈ ∆[3] n+1 ˙ ˙ 0 , rα rβ 0 has infinite order. Corollary 4.4. For all α ∈ ∆[3] and for all β 0 ∈ ∆[3] Following the arguments of Prop. 3.7 we have Proposition 4.5. ∞ [3] is dense in SO(3); (i) H+ (ii) H ∞ [3] is dense in O(3); (iii) Σ[3] is dense in S[3]. Define VR := V ∩ R4 . It is now rather easy to use the four-dimensional results of §3 to deduce corresponding results about the three-dimensional situation. We state these without further comment. Proposition 4.6. Σ[3] = ∞ [ (Sn [3] ∪ Sn0 [3]) = S[3] ∩ VR = S[4] ∩ VR = Σ[4] ∩ VR = Σ[4] ∩ V . n=0 Proposition 4.7. For all n ≥ 1, card (Sn [3]) = card (Sn0 [3]) = 3.(2)2n+1 . Proof: S1 [3] consists of the 24 roots of ∆[3], and similarly for S10 [3]. Now proceed by induction. We simply have to note that for each element of Sn [3] there are 12 reflections arising from ∆[3]0 that will map it into 12 images in Sn+1 [3], see Prop. 3.5. However, according to Prop. 3.8(iii) (which is actually set up so as to make the corresponding three dimensional case rather apparent) each of these elements will be produced exactly three times as we use all the reflections available from 0 ∆[3]0 . Thus card Sn+1 [3] = 4 card Sn [3], and likewise for card Sn+1 [3]. This completes the induction step. 12 5 How Σ[4] acts on Σ[3] Proposition 5.1. (i) H ∞ [4] acts transitively on Σ[4] ; (ii) H ∞ [3] acts transitively on Σ[3] ; (iii) For each α ∈ Σ[4] the set of roots (elements of Σ[4]) orthogonal to α is a translate of Σ[3] by an element of H ∞ [4] ; (iv) For each α ∈ Σ[4] the stabilizer of α in H ∞ [4] contains a subgroup conjugate to H ∞ [3]. Proof: (i) It suffices to show that H ∞ [4](1, 0, 0, 0) contains the basic roots of (6). This happens because in the Coxeter diagram of Fig. 2, (1, 0, 0, 0) is connected to the four others by chains of bonds that are all odd (namely labelled with 3 or 5). The argument is easily explained by doing it for the case of a bond labelled 3 between two roots α, β. Then rα rβ rα rβ rα = rβ . But the left side is also rrα rβ (α) , showing that rα rβ (α) = ±β. If the sign is wrong we can use one more reflection rβ . The proof of (ii) is similar. (iii) The stabilizing reflections ra of (1, 0, 0, 0), where a ∈ Σ[4], consists of all elements of Σ[4] ∩ V = Σ[3], see Prop. 4.6. Parts (iii) and (iv) follow directly from this. Instead of thinking of ∆[4] and ∆[4]0 simply as subsets of R4 , we can view them as subsets of the quaternion algebra H = R1 + Ri + Rj + Rk. We will identify the subspace V of R4 that we have been using with the subspace of pure quaternions: V = Ri + Rj + Rk. Then ∆[3] = ∆[4] ∩ V ⊂ H and similarly for ∆[3]0 . Generally we use this interpretation henceforth. We equip H with the usual conjugation u 7→ u e which changes the signs of the last three compo2 nents of each vector. The usual dot product of R4 is the standard one for H, and can be expressed in terms of the quaternionic multiplication by 1 x.y = (xe y + ye x) . 2 As is well known, the unit sphere S of H, that is the set of vectors a ∈ H satisfying e aa = 1, can be identified with SU (2). Explicitly S can be interpreted as the group SU (2) by the usual matrix representation of H using √ √ 1 0 −1 0 0 1 −1 0 √ 1⇔ , i⇔ , j⇔ , k⇔ √ . (16) 0 − −1 −1 0 −1 0 0 1 Returning to the ring (14), we define SU (2, R) to be the set of all unitary matrices with coefficients in R + iR. Using (16), SU (2, R) identifies with the set of all quaternions with coefficients in R with norm 1, i.e. with R4 ∩ S[4]. SU (2, R) is a subgroup of SU (2), and it is S[4] ∩ R4 = Σ[4] by Prop. 3.9. This tells us that Σ[4] has the natural structure of a group, namely SU (2, R). Furthermore, since Σ[4] is dense in S[4], SU (2, R) is dense in SU (2). We recall that for any a ∈ S and any x ∈ H we have ra (x) = x − (2x.a)a = x − (xã + ax̃)a = −ax̃a . 2 Usually the conjugation is designated by the , but that would cause confusion with our earlier notation. 13 (17) In particular this shows that reflection can be effected by multiplication. Since Σ[4] is by definition the smallest set containing ∆∪∆0 and closed under its own reflections, it follows that Σ[4] is generated as a group by ∆ ∪ ∆0 . Now we are viewing V as ordinary three-dimensional space, and since V = Ri + Rj + Rk, so x ∈ V if and only if x e = −x. Then for all a ∈ S and all x ∈ V , ^ ga = a(−x)e −1 = axe axa a = −axa−1 , showing that axa−1 ∈ V . This is the familiar way in which elements of SU (2) turn into elements of SO(3), so that a ∈ SU (2) maps to the conjugation γa : x 7→ axa−1 on V . This is a rotation, and in this way we obtain the well-known double cover mapping SU (2) −→ SO(3) . (18) However, there is more. If a ∈ Σ[4] then for all x ∈ Σ[3], then axã ∈ Σ[3], since if a ∈ 21n Z[τ ]4 and 1 x ∈ 21m Z[τ ]3 then axã ∈ 2m+2n Z[τ ]3 ∩ S[3] ⊂ Σ[3]. Thus the group of rotations on the sphere in 3-space induced by Σ[4] actually acts as a group of symmetries of Σ[3]. In particular they all lie in SO(3, R). Proposition 5.2. (i) Σ[4] = R4 ∩ S[4] is a group under quaternion multiplication and is generated by ∆ ∪ ∆0 ; (ii) Σ[4] = SU (2, R) and SU (2, R) is dense in SU (2); (iii) Σ[4] acts naturally as a group of isometries on Σ[3] via the mapping γa : x 7→ axã for all a ∈ Σ[4] = SU (2, R); (iv) Under the mapping γ in (iii), SU (2, R) maps onto a subgroup of index 2 in SO(3, R). Proof: There remains only to prove (iv). Let ρ := ( √12 , √12 , 0, 0). We begin by making the observation that γi ((x1 , x2 , x3 , x4 )) = (x1 , x2 , −x3 , −x4 ) γj (x1 , x2 , x3 , x4 )) = (x1 , −x2 , x3 , −x4 ) (19) γk ((x1 , x2 , x3 , x4 )) = (x1 , −x2 , −x3 , x4 ) γρ ((x1 , x2 , x3 , x4 )) = (x1 , x2 , −x4 , x3 ) . Since the mapping SU (2) −→ SO(3) is 2 : 1 and its kernel is ±1 ∈ SU (2), it is clear that the last of these rotations, γρ , while is is in SO(3, R), is not something that can arise from SU (2, R). We shall see that this is the only obstruction that we have to deal with. Since Σ[3] = S ∩ VR , SO(3, R) maps Σ[3] into itself. Let G := γSU (2,R) denote the image of SU (2, R) in SO(3, R) under the mapping SU (2) −→ SO(3, R). Note that if a ∈ Σ[3] and x ∈ V then x̃ = −x and (17) reads ra (x) = −ax̃a = axa = −axa−1 . (20) Now suppose that R ∈ SO(3, R). We wish to make R as simple as possible by multiplying R on the left by elements in G. Consider R(0, 1, 0, 0) ∈ S ∩ VR = Σ[3]. Since H ∞ [3] is transitive on Σ[3], 14 there exists w ∈ H ∞ [3] with wR = (0, 1, 0, 0). Writing w = rak . . . ra1 , (20) shows that (−1)k w ∈ G. This being the case we can assume at the outset that R(0, 1, 0, 0) = (0, ±1, 0, 0). However we see from (19) that we can alter this sign, so we can actually assume that R(0, 1, 0, 0) = (0, 1, 0, 0). Now, the only elements of Σ[3] that have the form (0, 0, ∗, ∗) are (0, 0, ±1, 0) and (0, 0, 0, ±1), and since R is an orthogonal transformation fixing (1, 0, 0, 0) and (0, 1, 0, 0), it can only permute these two pairs roots around. Of the four possible rotations that are possible here, two of them require interchanging the two coordinate positions and changing one sign, and that requires γρ . Thus, by composing R with elements of G we can reduce to one of the pair {1, γρ }; and indeed, γρ ∈ / G. This proves that [SU (2) : G] = 2. We note that SO(2, R) is not dense in SO(2), since the only R-points on the unit circle are (±1, 0) and (0, ±1). Likewise U (2, R) (the subgroup of all the elements of U (2) whose coefficients are in R+iR) is not dense in U (2)—and for basically the very same reason: det on U (2) is surjective onto the unit circle. Proposition 5.3. The stabilizer of (1, 0, 0, 0) in hH ∞ [4], ιi is hH ∞ [3], ιi where ι is the permutation operator on the last two coordinates. Here H ∞ [3] is in its obvious form of working on the last three coordinates, that is on the space V = Ri + Rj + Rk in the quaternions, and hH ∞ [3], ιi is the group generated by H ∞ [3] and ι. Proof: The first observation is that in H3 the rotation of the last three coordinates exists (as a rotation of order 3) and also the changes of signs of the individual coordinates exist (as the reflections in these coordinates). The same goes for H30 , of course. Now let w ∈ H ∞ [4] be in the stabilizer of z := (1, 0, 0, 0). Let y := (0, 1, 0, 0) and let x := w(y). Then x ∈ V since w is an isometry. Since H ∞ [3] is transitive on Σ[3], there is a v ∈ H ∞ [3] with v(x) = y = (0, 1, 0, 0). Now the element vw fixes both x and y. We can ask what vw does to Σ[4]. It must map the remaining two elements (0, 0, 1, 0) and (0, 0, 0, 1) into roots of the form (0, 0, ∗, ∗). However the general form of roots shows that the only possibilities are (0, 0, ±1, 0) and (0, 0, 0, ±1). The signs can be adjusted as we please by using an element from the Weyl group of K, but vw may still interchange the two points, in other words may be the involution ι. We shall show in Prop. 9.1 that H ∞ [4] does not contain ι. 6 ∆ and ∆0 as groups The two basic root systems we began with are ∆ and ∆0 . The rational parts of these sets is their intersection K = ∆ ∩ ∆0 . Proposition 6.1. K, ∆, and ∆0 are groups (under quaternion multiplication) of orders 24, 120, 120 respectively. This is a well-known fact, see for instance [5] §20. But for the convenience of the reader we prove it here. Proof: For the rational root system K it is clear that the product of any two elements of K is a unit vector in the rational space (1/2)Z4 , and so is still in K. Inversion is just quaternion conjugation, and it preserves K. Thus K is a group with order 24. We shall prove that ∆ is a group; the proof for ∆0 then follows immediately. The roots of ∆ are given in (1). The first set of these consists, in quaternion notation, of the elements ±1, ±i, ±j, ±k. Multiplication of a quaternion (x1 , x2 , x3 , x4 ) on the left by i turns it into (−x2 , x1 , −x4 , x3 ). There are sign changes, but the important thing to observe is that it performs an even permutation of 15 the coordinates, interchanging the first two coordinates and the last two coordinates. Likewise multiplication by j or k performs an even permutation of the coordinates (and makes two sign changes). Similar things happen if the multiplication by i, j, k is on the right. In particular, left and right multiplications by ±1, ±i, ±j, ±k map ∆ into itself. Now apart from the elements ±1, ±i, ±j, ±k in ∆, every other x ∈ ∆ has the property that 2x ∈ Z[τ ]4 and 2x 6= (0, 0, 0, 0) in F44 . In fact 2x is either (1, 1, 1, 1) or an even permutation of (0, 1, τ 0 , τ ). This property actually characterizes elements of norm 1 that are in ∆\{±1, ±i, ±j, ±k}. Similar remarks hold for ∆0 where the roles of τ and τ 0 are interchanged. Now let x, y ∈ ∆. We want to prove that xy ∈ ∆. By what we said above, we can assume neither is in the set {±1, ±i, ±j, ±k}. We begin by showing that 2xy ∈ Z[τ ]4 . By assumption 2x and 2y are in Z[τ ]4 . We know that even permutations and arbitrary sign changes on y will leave it in ∆. In particular ỹ ∈ ∆, and so we know that 2x.ỹ ∈ Z[τ ], by Prop. 2.1. Now x.ỹ is the first coordinate in the quaternion xy, so we know that that the first coordinate in 2xy is 2x.y ∈ Z[τ ]. If we look at the second coefficient of xy we can see that it is (x1 , x2 , x3 , x4 ).(y2 , y1 , y4 , −y3 ), so this shows that the second coefficient of 2xy is in Z[τ ]. Continuing this idea with the third and fourth coefficients we end up with 2xy ∈ Z[τ ]. At this point, since \|xy\| = \|x\| \|y\| = 1 we know that xy is an element of ∆ or ∆0 . However, it is ˙ 0 (that is to say in ∆0 \∆). We can see this in the following way. The first not possible for xy ∈ ∆ ˙ and v ∈ ∆ ˙ 0 then neither uv nor vu is in ∆∪∆0 . The reason is that 2u.v ∈ thing to note if u ∈ ∆ / Z[τ ], 0 −1 −1 ˙ see Prop. 2.1. So if we assume that xy ∈ ∆ then from y = x (xy) we have that x cannot be in ˙ But x and x−1 are in ∆, so this shows that x−1 , and hence also x, is in K. Similarly x = (xy)y −1 ∆. shows that y is in K. But then xy ∈ K since K is a group. This is contradicts the assumption that ˙ 0 , so xy ∈ ∆. This finishes the proof. xy ∈ ∆ We note in passing that ˙ ⊂ ∆; ˙ K∆ ˙0⊂∆ ˙ 0. K∆ (21) since ∆ (resp. ∆0 ) is a group and the products cannot be in K, for it is a subgroup. We now determine basic the structure of Σ as seen from its generation by the two groups ∆ and ∆0 . The following set C of elements is a set of representatives for the four non-trivial right cosets of K in ∆ : 1 1 1 1 0 0 0 0 (0, 1, τ , τ ), (0, −1, τ , τ ), (0, 1, −τ , τ ), (0, 1, τ , −τ ) . (22) 2 2 2 2 Similarly for C 0 and the right cosets of K in ∆0 . Now, any non-trivial element of Σ can be written as an alternating product of elements of ∆ and 0 ∆ . Using the coset representatives we can always write such a product in the form wam am−1 . . . a2 a1 (23) where the ai are alternately in C and C 0 and w ∈ K, since K is a subgroup of both ∆ and ∆0 . We are going to prove that this writing is unique: each element of Σ can be written uniquely in this form. The key to this is to introduce the set E ⊂ Z[τ ]4 /4Z[τ ]4 , defined as the set of all {(x1 + y1 τ, x2 + y2 τ, x3 + y3 τ, x4 + y4 τ ) where the xi and yi are integers and 16 mod 4Z[τ ]4 } (24) • all the xi are congruent to 1 or 3 modulo 4; • all the yi are congruent to 0 or 2 modulo 4; • the number of xi congruent to 1 is even ; • the number of yi congruent to 2 is odd. It is straightforward to see that this set has 64 elements. Using conjugation, we have the second set E 0 which is easily seen to have the same four properties except that the number of xi congruent to 1 is odd (we are retaining the x + yτ form). These two sets have remarkable properties: Lemma 6.2. (i) (1/2)EE = E and (1/2)E 0 E 0 = E 0 ; (ii) E ∩ E 0 = ∅; (iii) 4CC 0 ⊂ E and 4C 0 C ⊂ E 0 . Part (ii) is clear from the definitions. It is only necessary to verify the first parts of (i) and (iii). At the present time we have depended on brute force computation to prove this. For example, in looking at 4CC 0 in (iii), we have the following of these 16 products: √ √ √ √ √ √ (1, − √5, 5, 5) (3,√− 5,√−1, 1)√ (−1, 3, 5, (−1, −3, √−1) √1, 5) (3, − 5, √ 1, −1) (1, − 5, − 5,√− 5) (1, (1, −3,√ 5, −1) √ 3,√1, 5)√ (1, −3, −1, 5) (1, 5,√ 5, − 5) (−3, 5, 1,√1) (−1, −3, 5, √1) √ √− √ (−1, 3, −1, 5) (1, 3, 5, 1) (−3, − 5, −1, −1) (1, 5, − 5, 5) from which one can convert into x + yτ form to show 4CC 0 ⊂ E. Conjugating we see that 4C 0 C ⊂ E0. Using Lemma 6.2, let us show that is not possible for zam am−1 . . . a2 a1 of (23) to be equal to 1, except trivially. The ai alternate between C and C 0 , and, for definiteness assume that a1 ∈ C 0 . This analysis of this case uses E, while the case with a1 ∈ C uses E 0 . Now suppose that wam am−1 . . . a2 a1 = 1. Then am am−1 . . . a2 a1 = w−1 . For this to happen −1 m ≥ 2. If m is odd we can take the am to the other side to get am−1 . . . a2 a1 = a−1 m w . In either case the right hand side is of level at most 1, and now we can assume that m = 2n is even and n ≥ 1. We will show that the level of the left-hand side is n + 1, leading to a contradiction of levels. Starting with 4a2 a1 ∈ 4CC 0 ⊂ E, which shows that a2 a1 has level 2. Next we have 8a4 a3 a2 a1 = (1/2)(4a4 a3 )(4a2 a1 ) ⊂ (1/2)EE = E , showing that a4 a3 a2 a1 has level 3. If n > 2 we can continue in this way to show that a6 a5 a4 a3 a2 a1 has level 4, and so on. This shows that indeed the level of the left-hand side is n + 1. Of course the argument when a1 ∈ C can be obtained by conjugating the equation. Now to show the uniqueness of the standard form, suppose that wam am−1 . . . a2 a1 = zbp bp−1 . . . b2 b1 are two expressions in the form (23). 17 O(4) SU (2) dense dense SU (2, R) O(4, R) ⌃ H1 = 2 amalgamation of and 0 over K 0 H40 H4 K N (K) {1} {1} root systems amalgamation of H4 and H40 over N (K) groups Figure 4: How Σ and H ∞ appear as amalgamations. Supposing that the two expressions are different, we can assume that p + m is minimal for such ˙ an equality. We can push w to the other side, and so assume from the outset that w = 1. If a1 ∈ ∆ 0 ˙ and b1 ∈ ∆ , or vice-versa, then we can invert all the elements of the left hand side and get a new reduced expression that equals 1, and we know that this cannot happen. ˙ say, then b1 a−1 ∈ ∆ and we can rewrite the If on the other hand a1 and b1 are both in ∆, 1 right-hand side, to bring it into standard form again, while not increasing the number of terms on the right-hand side, and in so doing reduce the combined length p + m, contrary to our minimality assumption. So this case does not happen either. This proves the uniqueness of expression in (23) Putting this together we have shown: Proposition 6.3. Every element of Σ is uniquely expressible in the form (23). The group Σ is the amalgamation of the two groups ∆ and ∆0 over K, i.e. it is the free product of ∆ and ∆0 factored by the normal subgroup generated identifying the elements of a ∈ K appearing in ∆ and ∆0 . 7 W (K) and its normalizer We know that H ∞ [4] is generated by two copies of the Coxeter group of type H4 , namely H4 and H40 . In this section we shall call them W (∆) and W (∆0 ). Both of them contain the subgroup W (K) generated by the reflections defined by K, and this is a Coxeter group (of type D4 ). Using the root bases described above we see that together they create the Coxeter diagram (2). As we will see below, it appears that there is an additional relation beyond the Coxeter relations, see (26). We shall show the origin of this relation and how together with the obvious Coxeter relations, we obtain a presentation of H ∞ [4]. 18 We begin with looking at the group N (K) := {w ∈ W (∆) : w(K) = K} . (25) Similarly we have N (K)0 for the stabilizer of K in W (∆0 ). Since K is a root system of type D4 , its complete group of automorphisms is the semi-direct product of W (K) and the symmetric group S3 (which acts as diagram automorphisms). Lemma 7.1. (i) N (K) is the semi-direct product of W (K) and a cyclic group of order 3. (ii) N (K) is the normalizer of W (K) in W (∆) . (iii) The parallel results hold for W (∆0 ) and its normalizer N (K)0 . (iv) N (K) = N (K)0 = W (∆) ∩ W (∆0 ) = H4 ∩ H40 . Proof: The set K is a root system of type D4 and W (K) is its Weyl group. Thus N (K)/W (K) is a subgroup S3 of diagram automorphisms of D4 . Comparing the orders 26 .32 .52 of W (∆) and 26 .3 of W (K) ([1]), we see that the index [N (K) : W (K)] is either 1 or 3. We shall now see that it is 3. The Sylow 3-groups of W (∆) are of order 9. They are in fact the direct product of two cyclic groups of order 3. We can see this from the fact that there are subgroups of type A2 × A2 inside H4 . E.g. we have the root pairs {(1, 0, 0, 0), (1/2)(−1, 1, 1, 1)} and {u := (1/2)(0, −1, τ 0 , τ ), v := (1/2)(0, τ 0 , τ, −1)}, both of which are A2 bases. Let Γ be the associated A2 × A2 root system. Multiplying, we find that uv = (1/2, −1/2, −1/2, −1/2) ∈ K and ru rv (x1 , x2 , x3 , x4 ) = uv(x1 , x2 , x3 , x4 )vu = (x1 , x3 , x4 , x2 ) . This three cycle performs a diagram automorphism of the D4 root system K and together with W (K) it generates N (K). Although u, v ∈ ∆\K, their product is in K. Thus we have u0 , v 0 ∈ ∆0 \K and their product is the same element of K. They serve to produce the very same three cycle, and hence lie in the normalizer N (K)0 of W (K) in W (∆0 ). Thus N (K) = N (K)0 . Any element of w ∈ W (∆0 ) ∩ W (∆) maps K = ∆ ∩ ∆0 into itself and so stabilizes K. Thus it is in N (K) = N (K)0 . Let s := ru rv = ru0 rv0 . (26) Using (26) we can write down a presentation for the group H ∞ [4]. The presentation uses abstract generators Rα , Rα0 and relations that involve the reflections rα , rα0 . This is discussed in the next section. Comment: All of the Sylow 3-subgroups of W (∆) are conjugate by elements of W (∆). The ones that are in N (K) are Sylow 3-subgroups of it and so are conjugate by elements of N (K) to each other, and in particular to the one that we used above. These conjugations must conjugate Γ into other A2 × A2 root systems in ∆. Each of these root systems generates another relation, just like (26). Of course these are just ones that follow from the conjugation of (26). 19 8 A presentation of H ∞ The purpose of this section is to prove Proposition 8.1. H ∞ = H ∞ [4] is generated by the following generators and relations: (i) generators Rα , α ∈ ∆, and Rα0 , α0 ∈ ∆0 ; (ii) Rα2 = 1 = Rα2 0 for all α ∈ ∆ and for all α0 ∈ ∆0 ; (iii) relations Rα = Rα0 for all α ∈ K; (iv) relations Rα Rβ Rα−1 = Rrα (β) for all α, β ∈ ∆; and similar relations for ∆0 ; (v) with u, v ∈ ∆\K and u0 , v 0 ∈ ∆0 \K as given above, Ru Rv = Ru0 Rv0 . Notice that Rα = Rα Rα Rα−1 = R−α for all α ∈ ∆ and similarly for ∆0 . The proof depends on working explicitly with the group H ∞ and the reflections rα , rα0 as they appear in terms of the algebra of the quaternions, and showing that we can untangle any relation written in terms of these generators by only using the relations corresponding to those itemized in the statement of the Proposition. We begin by recalling that for a ∈ S, the quaternionic form of the reflection ra in a (17). The effect of a product of reflections rak · · · ra2 ra1 acting on x is ( −ak ag e a1 ae2 . . . ag k−1 . . . ae2 a1 x k−1 ak if k is odd x 7→ (27) aek ak−1 . . . a2 e a1 x e a1 a2 . . . ak−1 aek if k is even . When we write out products of reflections as two-sided products, as in (27), this involves alternately conjugating elements (this is quaternion conjugation, which we are designating by tilde), with the explicit form depending on whether there are an even or an odd number of reflections involved, and a possible overall sign change. This is rather messy to write down but quite trivial to do in practice. In order to avoid lots of notation we simply signify the whole conjugation process by putting a hat on each symbol, e.g. â, which is capable of being conjugated. Thus the symbol beneath the hat may or may not be conjugated according to the overall length being even or odd. Thus we write rak . . . ra1 (x) = (−1)k abk . . . ab1 x b ab1 . . . abk . (28) ˙ and ∆ ˙ 0. Notice that conjugation stabilizes each of K, ∆, Proof:(Prop. 8.1) Let rak . . . ra2 ra1 = 1 (29) be a non-trivial relation in H ∞ , written in terms of root reflections raj where the aj ∈ ∆ ∪ ∆0 . Our objective is to show that such a relation is a consequence of the relations (i)–(v). Equivalently we wish to show that the word on the left-hand side of (29) can be reduced to the empty word by using only these relations. We can rewrite the left-hand side in the form (28). To begin with consider any product rbp . . . rb1 with all bj ∈ ∆ and its action on H: rbp . . . rb1 (x) = (−1)p bbp . . . bb1 x b bb1 . . . bbp , 20 for all x ∈ H. Let bbp . . . bb1 =: BL and bb1 . . . bbp =: BR . Then BL , BR ∈ ∆ since ∆ is a group, and the mapping assumes the form x 7→ (−1)p BL x b BR . There seems to be no simple relationship between BL and BR , and it seems quite possible that one would belong to K while the other would not. However, if both BL , BR ∈ K then we can see that rbp . . . rb1 ∈ N (K). In fact, using the fact that the central inversion −1 ∈ W (K) we see that for all x ∈ K, rbp . . . rb1 (x) = (−1)p BL x b BR ∈ K, see (25). Notice that the fact that rbp . . . rb1 ∈ N (K) is a fact that takes place entirely inside the group W (∆), that is to say, it is determined entirely by the structure of W (∆) as a Coxeter group. In view of the structure of N (K), we can rewrite the product rbp . . . rb1 so as to assume that all the bj ∈ K with the possible exception of bp , bp−1 which may be u, v or v, u, see (26). This is all deducible from the Coxeter relations of W (∆). We now begin to parse the word on the left-hand side of (29), utilizing the form of (28), from right to left. For definiteness we will assume that a1 ∈ ∆. We now move to the left until we reach the first ap+1 ∈ / ∆. If all the roots involved are in ∆, so p = n, there is nothing to prove since the whole word is in W (∆) and the relation (29) is a consequence of the Coxeter relations of W (∆). Assuming (1) (1) that p 6= n, we rewrite rap . . . ra1 using the two blocks BL = abp . . . ab1 and BR = ab1 . . . abp which are both in ∆. In the case that rap . . . ra1 ∈ N (K) we rewrite with the word (using the Coxeter relations of W (∆)) so that all the aj involved are in K with the possible exception that ap , ap−1 are the pair u, v in some order. In that case we use (26) to replace this pair with u0 , v 0 in the corresponding order, thus shifting these last two letters into ∆0 and passing them on to the next block and using (1) (1) only ad p−2 . . . ab1 to make our blocks BL and BR . These two blocks are then elements of K. So (1) (1) in the case rap . . . ra1 ∈ N (K)\K, we do this little trick to make BL , BR ∈ K and pass along the non-K terms to subsequent blocks. We write q = p or q = p − 2 accordingly. We may assume that this rewriting and the passing along of pairs ru rv or rv ru , has been done ˙ in advance. That means that each block multiplies out to be either in K or in ∆. We now continue the parsing process from where we left off, this time producing a new set (2) (2) aq , . . . , ap+1 which are of elements in ∆0 whose left and right product produce blocks BL and BR , which are both in ∆0 . Again, if these two blocks are both in K then ras . . . raq+1 ∈ N (K)0 = N (K), and this is derivable from the Coxeter relations of W (∆0 ). (2) (2) There is a little bit of extra attention needed here. The elements aj appearing BL and BR have their conjugations determined by the parity of their positions in the sequence of reflections appearing (2) (2) in (29). But as an element of W (∆0 ), the action of ras . . . raq+1 might be either u 7→ ±BL û BR , or the conjugations may be interchanged so its action would be g (2) e g (2) u 7→ ±BR u b BL . (2) (2) In either case, if both BL and BR are in K, the ras . . . raq+1 maps K into K and so lies in N (K). We continue this process until we finally achieve a rewriting of left-hand side of (29) in the form (h) (1) (1) (h) (−1)n BL . . . BL (b x)BR . . . BR = x , (30) where n is the original length of the word we began with. This equation is valid for all x ∈ H4 . Of ˙ or ∆ ˙ 0. these, all blocks multiply to be in K or ∆ 21 Now if all the blocks here lie in K then, as we saw above, the entire effect on x is that of an element of N (K), and the entire relation can be determined, block by block, only using the Coxeter ˙ or ∆ ˙ 0 . Choose relations of W (∆) and W (∆0 ). So let us suppose that at least one block is either in ∆ a first such block to appear either to the left or right of x; call it B. The blocks, if any, that lie between x and B are in K. ˙ and suppose it is on the right. Choose x ∈ ∆ ˙ 0 . Then on For definiteness suppose B is of type ∆ the left-hand side we have a partial product in the set ˙ 0 [ possible blocks in K] ∆ ˙ ... ...∆ From this we see that every element in the left-hand side has level at least 2, see Prop. 6.3 and the discussion preceding it. The level can only increase if there are any other alternations of blocks ˙ and ∆ ˙ 0 in (30). The right hand side of (30) has level 1. This is a contradiction. This of types ∆ argument works in the same way if B is on the left. This contradiction shows that all the blocks must lie in K, and so the relations (i)-(v) Prop. 8.1 suffice to reduce the left-hand side of (29) to the empty word. This proves that these relations are a presentation of H ∞ . There is a more succinct way of saying this that is useful. Let C be a set of coset representatives for all the non-trivial right cosets of H4 mod N (K), and let D be a set of coset representatives for all the non-trivial right cosets of H40 mod N (K). That is [ H4 = N (K) ∪ N (K) y (disjoint union) y∈C and D does the same for H40 . A natural choice would be to relate C and D by using the conjugation ( )0 . However it is better to keep the freedom for other choices, as we shall see below. Proposition 8.2. (i) Let C (resp. D) be a set of coset representatives for all the non-trivial right cosets of H4 mod N (K) (resp. H40 mod N (K)). Then very element of H ∞ [4] is uniquely expressible in the form z yk . . . y1 (31) where z ∈ N (K), and where the yj alternate between C and D. (ii) H ∞ is the amalgamation of H4 = W (∆) and H40 = W (∆0 ) over their intersection N (K). Proposition 8.3. Let us choose C and D of Prop. 8.2 so that the representatives are actually in H3 (resp. H30 ) when the coset contains elements of these subgroups. Then (i) every element of H ∞ [3] ⊂ H ∞ [4] is uniquely expressible in the form z yk . . . y1 where z ∈ H3 ∩ H30 , and where the yj alternate between C and D ; (ii) H ∞ [3] is the amalgamation of H3 = W (∆[3]) and H30 = W (∆[3]0 ) over their intersection. 22 Proof: Notice that K ∩ V = {(0, ±1, 0, 0), (0, 0, ±1, 0), (0, 0, 0, ±1)} = ∆[3] ∩ ∆[3]0 , which is a root system of type A1 × A1 × A1 inside ∆[3]. The stabilizer of K ∩ V in H3 is the semi-direct product of the Weyl group of A1 × A1 × A1 and the cyclic group of order three arising by cycling the last three components of K ∩ V (or equivalently cycling the three A1 s). All of these elements are in N (K) and from this we see that the stabilizer of K ∩ V in H3 is N (K) ∩ H3 . In the same way the stabilizer of K ∩ V in H30 is N (K) ∩ H30 , and from this we see N (K) ∩ H3 = N (K) ∩ H30 = H3 ∩ H30 . Now H ∞ [3] is the subgroup of H ∞ [4] generated by H3 and H30 . According to Prop. 8.2 every element of H ∞ [4] is uniquely in the form (31). Amongst all the expressions (31), consider all those in which all the elements occurring are actually in H3 or H30 , and z ∈ H3 ∩ H30 ⊂ N (K). All of these expressions are obviously elements of H ∞ [3]. What’s more, every element of H ∞ [3] can evidently be written in this form. So H ∞ [3] is composed precisely of all of these products. Since the expression in this form is unique, we see that in fact it says that this is the amalgamation of H3 and H30 over their intersection. 9 The orthogonal group Define O(Σ) = O(Σ[4]) to be the group of all orthogonal transformations of R4 that stabilize the root system Σ. We already know a lot of transformations in O(Σ), namely all the elements of H ∞ . Along with O(Σ) we shall consider the orthogonal group O(4, R) of all orthogonal matrices with coefficients in R. Since Σ ⊂ R4 and it contains a standard orthogonal basis of R4 , we have O(Σ) ⊂ O(4, R). Notice that there is an orthogonal transformation of Σ that interchanges ∆ and ∆0 , namely the automorphism of Σ that interchanges the last two coordinates (x1 , x2 , x3 , x4 ) 7→ (x1 , x2 , x4 , x3 ). Call this transformation ι. (Of course it has numerous conjugates.) Evidently O(Σ) contains the subgroup hH ∞ [4], ιi generated by H ∞ and ι. Now we prove the reverse inclusion. Proposition 9.1. (i) O(Σ[4]) = O(4, R) ; (ii) O(Σ[4]) = hH ∞ [4], ιi = hιi n H ∞ [4] = hH4 , ιi ; (ii) [O(4, R)) : H ∞ [4]] = 2 . Proof: (i) Let g ∈ O(4, R). Then g stabilizes S ∩ R4 = Σ, and so g ∈ O(Σ). (ii) We know that K is a root system of type D4 , for which (3) is a root base. Let B be any D4 root base in Σ. We will try to use elements of H ∞ to bring B back to the basis (3). We will see that we can almost do this. To finish the job we may need to reverse the last two coordinates, hence use the transformation ι. Let V := (0, R, R, R), as before, and recall that V ∩ Σ[4] = Σ[3], see (4.6). This base has the traditional D4 Coxeter diagram with one central node and three nodes attached to it. Using H ∞ and simple transitivity, we can assume that one of the non-central nodes of the base elements of B is (1, 0, 0, 0). Then the other two non-central nodes are in V ∩ Σ[4] and so in Σ[3]. We take one of these and again using transitivity— this time of H ∞ [3]— to bring this node to (0, 1, 0, 0). The remaining non-central node is now of the form (0, 0, ∗, ∗). The only elements of Σ[3] 23 of this form are (0, 0, ±1, 0) and (0, 0, 0, ±1). We can use ι to assume that it is (0, 0, ±1, 0). Since the sign changes are in H ∞ [3], we can assume this element is (0, 0, 1, 0). There remains only the central node. Its scalar product with the three other nodes (that is, with (1, 0, 0, 0), (0, 1, 0, 0), (0, 0, 1, 0) is −1/2 and so it has the form (1/2)(−1, −1, −1, ∗). Given that it is a vector of length 1, the last term must be ±1. We again have the choice of sign. The upshot of this is that we have brought the base B to the standard basis of (3) using H ∞ and ι. Since any orthogonal transformation of R4 is determined entirely by its action on any base, we have proved that O(Σ) ⊂ hH ∞ , ιi. The rest of (ii) follows once we know that (iii) is true. (iii) We know that ι2 = 1 and ι stabilizes Σ and hence normalizes H ∞ (which is generated by ˙ and ∆ ˙ 0 (since it is an odd permutation the reflections in the elements of Σ). Also ι interchanges ∆ of the coordinates) while stabilizing K. It follows that conjugation by ι interchanges W (∆) = H4 and W (∆0 ) = H40 while stabilizing their intersection, which is N (K). This means that in choosing the coset representatives used in Prop. 8.2 we can choose D = ι(C)ι−1 . We will assume this. Now we prove that ι ∈ / H ∞ . If, on the contrary, it were in H ∞ then using Prop. 8.2 we could write ι uniquely in the form ι = zyk . . . y1 where z ∈ N (K) and the other elements are alternately in C and D. Then ι = ιιι−1 = ιzyk . . . y1 ι−1 = ιzι−1 ιyk ι−1 . . . ιy1 ι−1 . This gives us a second way of writing ι as an element of H ∞ , in the form described in Prop. 8.2 and accordingly they are identical. This obviously is not true if any of the coset terms are present, since C and D are reversed by the conjugation and y1 6= ιy1 ι−1 , etc. Thus we end up with ι = ιzι−1 and so ι = z ∈ N (K). Now certainly ι ∈ / W (K), for it effects a diagram automorphism of D4 , and we have already noted that [N (K) : W (K)] = 3. Thus ι ∈ / N (K), and this contradiction shows that ι ∈ / H ∞. There is a corresponding result for the three-dimensional case involving Σ[3] and O(3, R). Here we take O(3, R) to be the subgroup of O(4, R) which fixes the vector (1, 0, 0, 0) in H, and define O(Σ[3]) to be the subgroup of O(3, R) stabilizing Σ[3]. Proposition 9.2. (i) O(Σ[3]) = O(3, R) ; (ii) O(Σ[3]) = hH ∞ [3], ιi = hιi n H ∞ [3] = hH3 , ιi ; (ii) [O(3, R)) : H ∞ [3]] = 2 . Proof: The argument is essentially a repeat of what we saw in Prop. 9.1. Since Σ[3] = S[3] ∩ R3 , it is clear that O(3, R) stabilizes Σ[3], and this proves (i). We need only prove that O(3, R) ⊂ hH ∞ [3], ιi. Working in H, Σ[3] contains the standard basis vectors v1 := (0, 1, 0, 0), v2 := (0, 0, 1, 0), v3 := (0, 0, 0, 1). Let g ∈ O(3, R) and let g(v1 ) = x ∈ Σ[3]. Since H ∞ [3] is transitive on Σ[3], there is an element h ∈ H ∞ [3] so that hg(v1 ) = v1 . Then hg stabilizes the span of v2 and v3 as well as mapping them into elements of Σ[3]. However, the only elements of R3 of this form are (0, 0, ±1, ±1). Since the sign changes on individual coordinates are elements of H ∞ [3], we can assume that we have the positive signs here. Thus hg fixes both v2 and v3 , in which case hg = 1, or hg interchanges v2 and v3 , in which case hg = ι[3]. 24 10 Approximation in SU (2) The unitary group SU (2) plays a significant role in the theory of quantum computation as the group of admissible transformations of a quantum qbit. In this section we offer an algorithm by which any element of SU (2) can be approximated to any degree of accuracy by the repeated use of just a few very simple fixed elements in SU (2, R). In fact, this number can be reduced to five, these five being related to a root base of Σ[4]. Passing to SO(3) via (18) we can similarly approximate any rotation to any degree of accuracy by elements of SO(3, R). The idea is this. Inside the quaternion algebra H the unit sphere S is a group under quaternion multiplication and this group is SU (2) from the point of view of the matrix representation of H of (16). The full set of roots Σ[4] that we have generated out of ∆[4] and ∆0 [4] is in fact the subgroup of S generated by them and it is the orbit of ∆[4] ∪ ∆0 [4] under H ∞ [4]. Furthermore it is dense on the sphere S, Prop. 3.7. Hence appropriate sets of mirrors (reflecting hyperplanes) associated with the reflections in these elements partition S into convex regions of arbitrarily small diameter. We call such regions chambers. S A natural choice is a set Un := k≤n Ṡk ∪ Ṡk0 and we can use the associated hyperplanes of these points to define the set of chambers that we use. So we note that the chambers here are not absolute things, but rather the outcome of some selection of a finite number of root hyperplanes. Assuming that the chambers are selected and that one point p has been chosen in one of these chambers, say p ∈ C, then the idea is that any element x ∈ S can be reflected into C using only reflections in these mirrors, whereby it is a close approximation of p. Then reversing the operations we can take p to some point equally close to x. So we take some set Un and use its points to define a set of chambers. We should note that the situation with these chambers is not the same as in the customary case of root systems. The chambers are not all isometric copies of each other, nor are they necessarily simplices. The cellular decomposition comes about through the sets of reflecting hyperplanes that we use, but we never use all of them, only a finite number of them, and these finite sets are not invariants of H ∞ .3 For a ∈ S, the reflection ra in a has the quaternionic form ra x = x − 2(x.a)a = x − (xe a + ae x)a = −ae xa see (17). Suppose we wish to approximate x ∈ S using the reflections determined by the elements of Un . The idea is to reflect x a number of times so that it ends up as close as possible to 1 ∈ H. Choose a chamber C0 containing 1 (actually 1 will be at the vertex of such a chamber) and choose a point z close to 1 in the interior of C0 . For each element pair ±a ∈ Un we choose the one which satisfies a.z < 0 and let Un− denote this subset chosen in this way. So C0 = {u ∈ S : u.a ≤ 0 for all a ∈ Un− }. Now if there is a reflection hyperplane for which x and z lie on opposite sides, then a.x > 0 and we see that (ra x − z).(ra x − z) = (x − z).(x − z) + 4(a.x)(a.z) < (x − z).(x − z) , thus moving the point x closer to z. We can always do this so that we choose a for which a.x is as large as possible. 3 If the process is carried out to the end, allowing n → ∞, then every point y not on one of the reflecting hyperplanes is uniquely specified by the signs that it makes with each of the roots: y.a is either positive or negative. In this way it looks similar to a Dedekind cut where each real number is specified by the rationals that are less than it or greater than or equal to it. 25 Almost surely, this process has to stop in a finite number of steps with the point ending up in C0 . Here is why. Suppose that x = x0 , x1 , x2 , . . . were an infinite sequence of successive steps applied to the starting point x. It has to have limit points, and we choose one, say y, that is closest to z. Now a.z ≤ 0 for all a ∈ Un− , and it follows that y ∈ C0 . In fact every iteration brings us closer to z and if y∈ / C0 there is an a ∈ Un− with a.y > 0 and ra (y) is closer to z, a contradiction. Almost surely (in the sense of Lebesgue measure) x is not on the boundary of a mirror so neither is y. Thus y is in the interior of C0 . This means that almost surely there is a k for which xk is in the interior of C0 , and the iteration will stop as soon as it reaches such a k. From the point of view of approximation theory, we can always assume that x is not on a boundary. At the end of the process we will have achieved a relation of the form rak · · · ra2 ra1 (x) = p ∈ C0 . If d is the diameter of C0 , then since each reflection is an isometry of H relative to its usual norm, \|x − ra1 ra2 · · · rak 1\| = \|rak · · · ra2 ra1 (x) − 1\| ≤ d . This then gives us the desired approximation ( −a1 ae2 · · · ag k−1 ak ak ag k−1 · · · ae2 a1 if k is odd x' ae1 a2 · · · ak−1 e ak e ak ak−1 · · · a2 ae1 if k is even . (32) Treated as elements of SU (2), the matrix entries all come from the ring of dyadic golden numbers R. ˙ ˙ 0 we need to write each To write this fully in terms of the basic reflections in the roots of ∆∪K ∪∆ of the roots appearing in Un− in terms of these elements. Thus if a ∈ Un− is a = rbm rbm−1 · · · rb1 b0 ˙ ∪K ∪∆ ˙ 0 , then with all the bj ∈ ∆ ( e f bf m bm−1 · · · b1 b0 b1 b2 · · · bm−1 bm if m is odd a= e e ] −bm b] m−1 · · · b1 b0 b1 · · · bm−1 bm if m is even . Example: Here we give some sample results from a computer implementation of the approximation process. The set U3 of roots formed from the sets Ṡk , Ṡk0 is evaluated for k = 1, 2, 3, resulting in a total of 52416 = 2.3.(25 + 29 + 213 ) elements. A random element of SU (2) is computed, e.g. −0.244828 − 0.155561i 0.731977 − 0.616498i u= −0.731977 − 0.616498i −0.244828 + 0.155561i and the process then determines (in this case six) reflections from roots of R3 that bring this point on S[4] as close to the identity element 1 ∈ H as possible (using reflections from R3 ). These reflections ˙ ∪∆ ˙ 0 , giving rise to a total of fifteen. are rewritten as products of reflections from the original set ∆ The matrix 1 1 + τ i (1 − τ )i 2 (1 − τ )i 1 − τ i is a typical looking example. 26 After computing the resulting matrix product according to (32), we obtain the approximation −0.245712 − 0.15869i 0.730844 − 0.616694i , −0.730844 − 0.616694i −0.245712 + 0.15869i the difference being 0.000883332 + 0.00312853i 0.00113294 + 0.000195296i . −0.00113294 + 0.000195296i 0.000883332 − 0.00312853i This example is entirely typical both in the number of reflections required and in the degree of ˙ and ∆ ˙ 0 in terms of approximation. If we wished we could rewrite the reflections in the roots of ∆ the five generators of (6). Computing out more roots, e.g. U4 will result in more accuracy. We do not have any decent estimates on the diameters of the chambers arising from Un . Acknowledgements We thank the referee for his/her careful parsing of the manuscript and helpful historical suggestions. RVM was supported by Discovery Grant 461292003, Natural Sciences and Engineering Research Council of Canada. JM was supported by Grants-in-Aid for Scientific Research 26400005, 17K05158 (Monkasho Kakenhi, Japan). References [1] N. Bourbaki, Groupes et algèbres de Lie, Ch. IV-VI, Hermann, Paris, 1968. [2] L. Chen, R. V. Moody, J. Patera, Non-crystallographic root systems in Quasicrystals and Discrete Geometry, ed. J. Patera, Fields Institute Monographs, American Mathematical Society, 1997. [3] P-P. Dechant, C. Böhm, R. Twarock, Novel Kac-Moody-type affine extensions of noncrystallographic Coxeter groups, J. Phys. A: Math. Theor. 45(28)(2012), 285202. [4] P-P. Dechant,A Clifford Algebraic Framework for Coxeter group theoretical computations, Adv. Appl. Clifford Algebras, 24(2014), 89-104. [5] P. Du Val, Homographies, Quaternions, and Rotations, Oxford, 1964. [6] M. Koca, R. Koç, M. Al-Barwani, S. Al-Farsi, Maximal subgroups of the Coxeter group W (H4) and quaternions, Linear algebra and its applications, 412 (2006), pp.441-452. [7] R.V. Moody and J. Patera, Quasicrystals and Icosians, J. Phys. A: Math. Gen. 26 (1993), 28292853. 27	4math.GR
PREPRINT VERSION Improving files availability for BitTorrent using a diffusion model C. Napoli, G. Pappalardo, and E. Tramontana arXiv:1409.4565v1 [cs.NI] 16 Sep 2014 PUBLISHED ON: 2014 IEEE 23rd International WETICE Conference BIBITEX: @incollection{Napoli2014Improving year={2014}, isbn={978-1-4799-4249-7/14}, doi={10.1109/WETICE.2014.65}, booktitle={23rd International WETICE Conference}, title={Improving files availability for BitTorrent using a diffusion model}, publisher={IEEE}, author={Napoli, Christian and Pappalardo, Giuseppe and Tramontana Emiliano}, pages={191-196} } Published version copyright c 2014 IEEE UPLOADED UNDER SELF-ARCHIVING POLICIES NO COPYRIGHT INFRINGEMENT INTENDED Improving files availability for BitTorrent using a diffusion model Christian Napoli, Giuseppe Pappalardo, and Emiliano Tramontana Dipartimento di Matematica e Informatica, University of Catania Viale A. Doria 6, 95125 Catania, Italy Email: {napoli, pappalardo, tramontana}@dmi.unict.it Abstract—The BitTorrent mechanism effectively spreads file fragments by copying the rarest fragments first. We propose to apply a mathematical model for the diffusion of fragments on a P2P in order to take into account both the effects of peer distances and the changing availability of peers while time goes on. Moreover, we manage to provide a forecast on the availability of a torrent thanks to a neural network that models the behaviour of peers on the P2P system. The combination of the mathematical model and the neural network provides a solution for choosing file fragments that need to be copied first, in order to ensure their continuous availability, counteracting possible disconnections by some peers. Keywords-Dependability, distributed caching, P2P, neural networks, wavelet analysis. I. I NTRODUCTION In peer to peer (P2P) systems using BitTorrent, a shared content (named torrent) becomes split into fragments and the rarest fragments are automatically chosen to be sent first to users requesting the file. Fragments availability is given by the number of peers storing a fragment at a given moment, and periodically computed by a server storing peer ids, fragments held and requested [7]. For computing the priority of fragments to spread, at best availability has been freshly updated, however peers often leave the system hence file fragments availability quickly changes, possibly before the least available fragments have been spread [13]. This occurs so frequently that such a fundamental BitTorrent mechanism may become ineffective, and as a result some fragments can quickly become unavailable. Moreover, when choosing fragments to spread, communication latencies [8], [18] among peers are not considered, therefore fragments spreading will surely occur sooner on peers nearby the one holding the fragment to spread. As a result, the furthest peers could disconnect before receiving the whole fragment. This paper proposes a model for spreading file fragments that considers both a time-dependent priority for a fragment to spread and latencies among nodes. I.e. the priority of fragments to spread gradually changes according to the passing time and fragments availability. The priority variation is regulated in such a way that the availability of fragments is maximised while time goes. Fragments to spread are selected according to the results of our proposed diffusion model, developed by analogy to a diffusion model on a porous medium. Moreover, we propose to characterise the typical availability of a torrent observed on the P2P system, by using an appropriate neural network. Then, the selection of fragments to spread aims at counteracting their decreasing availability estimated for a later time. Therefore, the proposed work aims at supporting Quality of Service and dependability of P2P systems by attributing a priority on both the fragment to spread and the destination peer. This in turn increases availability and performances [9], [10], [11], as well as consistency [1]. The rest of the paper is organised as follow. Next section provides the mathematical representation for the proposed model. Section III develops the model for the diffusion of contents on a P2P system. Section IV introduces the neural network that predicts the user behavious. Section V describes some experiments based on our proposed model and neural network predictions, as well as the preliminary results. Section VI compares with related work, and finally conclusions are drawn in Section VII. II. M ATHEMATICAL R EPRESENTATION In order to put forward our analogy between BitTorrent and a physical system, some conventions must be chosen and some extrapolations are needed. We first describe a continuum system using a continuum metric, however later on we will single out a few interesting discrete points of the continuum. Due to the analogy to BitTorrent, we use a distance metric (named δ), which will be assimilated to the network latency among nodes, i.e. the hosts on a network holding seeds, peers or leeches. For the nodes we use notations ni or niα : the first indicates a generic i-esime node on the BitTorrent network, the second indicates the α-esime node as seen from the i-esime node. Of course, niα and njα could be different nodes if i 6= j. Double indexing is needed since when we use something like δ ij , it will be representing the distance of the j-esime node as measured by the i-esime node. Moreover, let us express Pkij as the probability of diffusion of the k-esime file fragment from the i-esime node to the j-esime node. Finally, we distinguish between time and time steps: the first will be used for a continuum measure of temporal intervals and we will use for it the latin letter t, the second will indicate computational time steps (e.g. the steps of an iterative cycle) and we will use for it the greek letter τ . Therefore, while δ ij (t) will represent the continuous evolution during time t of the network latency δ, which is measured by the i-esime node for the distance with the j-esime node, the notation δ ij (τ ) will represent the latency measured at the τ -esime step, i.e. the time taken by a ping from the i-esime node to the j-esime node, only for a specific time step τ . Finally, we will suppose that each node has the fragment zk of a file z and is interested in sharing or obtaining other portions of the same file, hence we will compute the probabilitylike function that expresses how easily the k-esime shared fragment is being copied from the i-esime node to the jesime node at a certain step τ and we will call it Pkij (τ ). Eventually, we are interested in an analytical computation for the urgency to share a fragment zk from ni to nj during a time set τ , and we will call it χi,j k (τ ). In the following sections we will distinguish between an actually measured value and a value predicted by a neural network using a tilde for predicted values as in x̃. III. F RAGMENTS DIFFUSION ON A P2P NETWORK In our work we compare the file fragments of a shared file to the diffusion of mass through a porous means. To embrace this view, it is mandatory to develop some mathematical tools, which we will explain in the following. A. Spaces and metrics Users in a P2P BitTorrent network can be represented as elements of a space where a metric could be given by the corresponding network communication latency. Therefore, for each node ni ∈ N , set of the nodes, it is possible to define a function δ : N × N → R / δ(ni , nj ) = δ ij ∀ ni , nj ∈ N (1) where δ ij is the amount of time taken to bring a minimum amount of data (e.g. as for a ping) from ni to nj . By using the given definition of distance, for each node ni , it is also possible to obtain an ordered list Ωi so that n o\|N \| Ωi = niα ∈ N : δ(ni , niα ) ≤ δ(ni , niα+1 ) (2) α=0 In such a way, the first item of the list will be ni0 = ni and the following items will be ordered according to their network latency as measured by ni . Using this complete ordering of peers, it is possible to introduce the concept of content permeability and diffusion. Let us consider the files shared by one user of a P2P system: each file consists of fragments that can be diffused. Then, the diffusion of a file fragment can be analysed in terms of Fick’s law. B. Fick’s law and its use for P2P Fick’s second law is commonly used in physics and chemistry to describe the change of concentration per time unit of some element diffusing into another [21]. This work proposes an analogy between a P2P system and a physical system. The key idea is to model the sharing file fragments as the diffusion of a substance into a porous means along one dimension. Different places of the porous means would represent different P2P nodes, whereas distances along such a one-dimension would be proportional to the network latencies. Then, P2P entities would be accommodated into the formalism of equations (1) and (2). Using both the First and Second Fick’s laws, the diffusion of a substance into a means is given as the solution of the following vectorial differential equation ∂Φ = ∇ · (D∇Φ) (3) ∂t where Φ is the concentration, t the time and D the permeability of the means to the diffusing matter. Since this is a separable equation and we use a 1–dimensional metric based on the distance δ, and assuming D as constant among the nodes, equation (3) can be written as a scalar differential equation ∂2Φ ∂Φ =D 2 (4) ∂t ∂δ This partial differential equation, once imposed the initial and boundaries conditions, admits at least Green’s Function as a particular solution [2]. Green’s Function lets us study the diffusion dynamics of a single substance and can be rewritten as solution of the equation (4) in the form: δ 1 Φ(δ, t) = Φ0 Γ √ (5) , Φ0 = √ 4Dt 4πDt where Γ is the complementary gaussian error function. Γ function should then be computed by means of a Taylor expansion. However, to avoid such a computationally difficult task, we use an approximation proposed in [6], where a pure exponential approximation for Γ(x) has been obtained, having an error on the order of 10−9 . Then, it is possible to have the following equation h i  2 2 4 1 1   Φ(δ, t) ≈ Φ0 6 e(Φ0 δ) + 2 e− 3 (Φ0 δ) (6)   − 12 Φ0 = (4πDt) for every node at a certain distance δ ∈ R+ at a time t ∈ R+ . C. From concentration to probability In equation (6) the scaling factor Φ0 is a function of the time t. On the other hand, the used formalism was developed mainly to focus on the distance δ and managing t merely as a parameter. The above mathematical formalism is valid as long as the distances δ(ni , nj ) remain time-invariant. The common practice considers the distance between nodes δ as time-invariant, however the actual network latencies vary (almost) continuously, with time, and a stationary Ωi ordered set is a very unlikely approximation for the network. In our solution, we make time-dependent the latency embedded into our model. In turn, this makes it possible to choose different fragments to be shared as time goes. For the P2P system, the equation (6) states that a certain file fragment zki in a node ni at a time t0 has a probability Pkij (t0 , t) to be given (or diffused) to node nj , at a distance δ ij (t0 ) from ni , within a time t, which is proportional to the Φ(δ, t) so that ij 2 ij 2 1 − 43 (pij 1 (pij k δ ) + k δ ) Pkij (t0 , t) = pij e e (7) k 6 2 ij where the function pij k = pk (t0 , t) carries both the diffusion factors and the temporal dynamics. And since we are interested in a simple proportion, not a direct equation, we can also neglect the factor 4π and then write pij k in the normalised form 1 1 1 (8) ·p ·√ pij k (t, t0 ) = √ t 4π Dk (t0 ) It is now important to have a proper redefinition of the coefficient D. Let us say that Tk is the number of users using file fragment zk (whether asking or offering it), Sk is the number of seeders for the file fragment and ρk is the mean share ratio of the file fragment among peers (and leeches), then it is possible to consider the ’urge’ to share the resource as an osmotic pressure which, during time, varies the coefficient of permeability of the network D. In order to take into account the mutable state in a P2P system, D should vary according to the amounts of available nodes and file fragments. We have chosen to define D as Dk (t0 ) , Tk (t0 ) Sk (t0 ) + [Tk (t0 ) − Sk (t0 )] ρk (t0 ) (9) by a formal substitution of D with Dk in Φ0 , we obtain the analytical form of the term pij k. D. Discrete time evolution on each node Indeed, the physical nature of the adopted law works in the entire variable space, however for the problem at hand discrete-time simplifications are needed. Let us suppose that for a given discrete time step τ = 0 node ni effectively measures the network latencies of a set of nodes {nj }, then an ordered set Ωi as in equation (2) is computed. Now, every node ni computes probability Pkij for each of its own file fragment zk and for every node nj . This probability corresponds to a statistical prevision of the possible file fragments spreading onto other nodes. Suppose that for a while no more measures for δ have been taken, and at a later discrete time step τ file fragment zki be copied to the first node to be served, which is chosen according to the minimum probability of diffusion, latencies and time since last measures were taken (see following subsection and equation (12)). Moreover, such a file fragment is reaching other nodes if the latency for such nodes is less than time ti , computed as tik (τ ) = τ X αk =0 δ(n i , niα ) (10) Index k is used in equation (10) to refer to file fragment zki . Indeed, since nodes need and offer their own file fragments, the ordered set of nodes referred by a given node depend on resource zk , i.e. Ωik = {niαk }. It is now possible to have a complete mapping of the probability of diffusion by reducing the time dependence from (t0 , t) to a single variable dependence from the discrete time-step τ . For each resource zk as Pkij (τ ) stated that it is possible to reduce Dk (t0 , t) to a one-variable function Dk (τ ) by assuming that at t0 we have τ = 0 and considering only the values of Dk (t0 , t) when t is the execution moment of a computational step τ . E. Assigning priorities and corrections Once all the Pkij (τ ) have been computed, and values stored to a proper data structure, it is actually simple to determine the most urgent file fragment to share, which is the resource that has the least probability to be spread, i.e. the k for which Pkij (τ ) is minimum. Furthermore, we consider that while time goes on an old measured δ differs from the actual value, hence the measure becomes less reliable. To take into account the staleness of δ values, we gradually consider less bound to δ the choice of a fragment, and this behaviour is provided by the negative exponential in equation (11). Given enough time, the choice will be based only on the number of available fragments. However, we consider that by that time a new value for δ would have been measured and incorporated again into the model choosing the fragment. Generally, for nodes having the highest latencies with respect to a given node ni , more time will be needed to receive a fragment from the node ni . We aim at compensating such a delay by incorporating into our model the inescapable latencies of a P2P network. Therefore, the node that will receive a fragment first will be chosen according to its distance. In order to model the fact that distant nodes, having the highest values for δ, will take more time to send or receive file fragments, we have chosen a decay law. Now it is possible to obtain a complete time-variant analytical form of the spreading of file fragments χij k (τ ) = e−cτ δ ij Pkij (τ ) (11) where the decay constant c can be chosen heuristically, without harming the said law, and tuned according to other parameters. If k indicates a file fragment and k ∗ the index of the most urgent file fragment to share, this latter is trivially found as the solution of a maximum problem so that n o ij k ∗ : χij (12) k∗ (τ ) = max χk (τ ) k Of course, all the priorities depend on the value of the bidimensional matrix of values of Pkij (we mark that the index An innovative neuro-wavelet method for reconstruction of photometric data 7 1.728 step of the series, a set of coefficients di and residuals ai , and so that x 10 Original signal Forward reconstruction Backward reconstruction Midpoint reconstruction 1.727 Ŵi [ai−1 (τ )] = [di (τ ), ai (τ )] 1.726 Topology of the WRNN Corrected Flux [ADU] 1.725 Figure 1. 7 ∀ i ∈ [1, M ] ∩ N (13) where we intend a0 (τ ) = x(τ ). The input set can then be represented as an N × (M + 1) matrix of N time steps of a M level wavelet decomposition, where the τ -esime row represents the τ -esime time step as the decomposition 1.724 Mid Point Forward NNET 1.723 1.722 Backward NNET 1.721 u(τ ) = [d1 (τ ), d2 (τ ), . . . , dM (τ ), aM (τ )] (14) i does not variate within the same node n ). Among these Each row of this dataset is given as input value to the values, there is no need to compute elements where j = i M input neurons of the proposed WRNN (Figure 1). The j and for those elements where the node n is not in the queue ij properties of this network make it possible, starting from for resource zk . In both these cases it is assumed Pk = 1. an input at a time step τn , (right) to predict the effective number Fig. 4. Neural networks structures (left), Forward and Backward reconstruction i Moreover, after n having completed to transfer zk to the of requests (or offers) at a time step τn+r . In this way the j node n , the element of indexes (j, k) is set to 1. In a similar WRNN acts like a functional fashion, each peer is able to identify a possible resource to ask for in order to maximize the diffusion of rare ones N̂ [u(τ (15) n )] = x̃(τ becauseones. wavelets do not verify some basic properties such as the absence of n+r ) instead this of common local minima, and does not provide by itself a suﬃciently response where r is graded the number of time[17]. steps of forecast in the future the existent rangeAND of possible transfer functions some particular andonly x̃ the predicted serie. classes IV.InWRNN PREDICTORS U SERS BEHAVIOR i 1.72 1.719 0 0.05 0.1 0.15 0.2 0.25 Time [Julian Days] 0.3 0.35 0.4 0.45 approximate the functional form of a wavelet. In this work the radial basis func- In order make thewere P2P chosen system able to properly reB. Predicted behaviour tionsto(radbas) as transfer functions, indeed this user particular kind of act to peaks of requests, as well as very fast changes by even equation (15), it is then possible to functions well describes in first approximation halfAsof described a wavelet, if these of fragments availability and/or share ratio, we propose obtain future prediction the number of requests for functions do not verify the properties shown by (1) and a(2). Anyway, afterofscalan innovative solution based on Wavelet Recurrent Neural a specific torrent, as well as its ing, shift and chosen activation it is possible to obtainavailability in the future. Networks (WRNN) to repetition characteriseofthethe user behaviour and function, + fact, by considering both the predicted x̃k (τn+r ) and In several mother forecast. wavelet For filters. Let ftorrent, : [−1;the 1] → R to be the choosen transfer producing a short-term a given the modeled χk (τn+r ), it is possible, at a time step τn , function, then wavelet analysis provides compression and denoising on to take counteracting actions and improve the probability the observed time series of the amount of!users prividing diffusion + f (2x + 1) x of ∈ [− 1 , 0 ]for a rare torrent. This is achieved, in practice, or requesting fragments; f˜(x) a=proper f˜(x +recurrent 2k) = neural network ∀k values ∈ Z for Dk(4) by using (τn+r ), which account for − f (2x − 1) x ∈ [ 0 , 1altered ] (RNN), trained with the said observed time series, provides the forecast of future time steps. Such modified values are well-timed estimations of future data. The ensamble of the computed by ourfor WRNN, e.g. in a computing node of a verifies all the properties of a wavelet function. it is possible the selected said wavelet analysis and RNN is called WRNN [5], [15], Socloud. Therefore, predicted values neural networks to for simulate a wavelet by that using the radbas function defined in for Tk (τn+r ), Sk (τn+r ) [16] and provides forecasts the number of users and ρk (τn+r ) are sent to each node acting as a peer. thea[−1; 1] realSeveral domain. It isnetworks indeed possible that will share fragment. neural have beento verify Each time a new torrent becomes shared on the P2P employed to find polaritons propagation " 2k+1 and metal thicknetwork, then a new WRNN is created and trained on ness correspondence [3]; to predict the behaviour a cloud system, to provide f˜(x) dx of = users 0 ∀ h <a kserver, ∈ Z e.g. requested from (5) requesting resources [17], to perform wavelet transform in 2h+1 predictions related to the availability and peers of the novel a recursive lifting procedure [4], [20]. set of shared fragments for that torrent. The predictions will It was shown that, in order to simulate a wavelet function, the chosen transfer be sent to the peers periodically, and allow peers to update functions wavelet. Thisupdate is thefrequency can be tuned in A. WRNN setup must be symmetrically periodical to emulate the valuesa of Dk (τ ). The reason for choosing a pair number of neurons in the aim to have the same number order to correctly match the dynamic of hosts. For this work, the initial datasets consists of a time series of positive weights in theorreconstruction layer. Theoretically, representing requestsand for negative a torrent,layer coming from peers, V. E XPERIMENTS if this happens, then the neuron of the exactly given declaration of availability for the torrentpairs coming fromsecond layer are emulating a reconstruction filter. Althoug this was a theoretical schema, there are both peers and seeds. Independently of the specificities of As shown in Figure 2, strong for the initial condition of the reasons for this theseries weights, thisτ experimental to havesome a non-zero sum, for a torrent happen to such data, let us call x(τ ), in where is the discrete setup, P2P system file fragments because neuralwith network beyond perform the inverse wavelet transform time step of data,the sampled intervals of onetohour. A be heterogeneously spread among peers (e.g. no one shares biorthogonal of the time series has fragment n. 2, and very few nodes have fragment n. 5). We must wavelet performdecomposition also the signal prediction. been computed to obtain the correct input set for the WRNN as required by the devised architecture shown in Figure 1. This decomposition can be achieved by applying the wavelet transform as a recursive couple of conjugate filters in such a way that the i-esime recursion Ŵi produces, for any time report a simulation comprising 11 peers and 5 file fragments, and an evolution in only 5 time steps. In the order, step after step, peer n0 selects file fragment z4 and sends this to peer n10 . Both the fragment to spread and the destination peer have been chosen according to equation (12). i In order: the initial condition (white cells represent the fragments), the normalised χij k (τ ) for a certain node n at different time steps τ VI. R ELATED W ORK Several studies have analysed the behaviour of BitTorrent systems from the point of view of fairness, i.e. how to have users contribute with contents that can be uploaded by other users, levelling the amount of downloads with that of upload. Fewer works have studied the problem of unavailability of contents in P2P BitTorrent networks. In [19], authors proposed to order peers according to their 1 0.8 1 ( 3 ) 2 ( 3 ) 3 ( 3 ) 0.6 ( ) Later on, as soon as possible, peer n0 selects another fragment, i.e. z3 to spread. Such a fragment could be send just after the transfer of the previous fragment has been completed, or concurrently to the first transmission. The shown evolution does not consider that the file fragment could have been passed, e.g., to node n10 , and so that for the next time step the value of χ for n10 would drop to zero. Note that the highest values of χ are an indication of the urgency of receiving a fragment. The described model and formula allow subsequent sharing activities, after the initial time steps, to be determined, in terms of which fragments should be sent. Figure 2 shows that after the first time steps it becomes more and more urgent for node n0 to obtain the missing fragments z2 and z5 . It is possible to see that the highest priority is for fragment z2 since its share ratio and the relative availability are very low with respect to fragment z5 . This was the expected behaviour of the developed model. For the simulation shown in Figure 2, all fragments, except z2 since it is actually unavailable, would be spread to peers in a very low number of time steps. Figure 3 shows the decay of several computed χ for different peers requiring fragment number 3. On the long run, this law will benefit nearby nodes, while on the short term, distant nodes are given the highest priority. j 3 Figure 2. 0.4 0.2 0 1 Figure 3. 2 3 4 5 6 7 8 9 10 Time decay of normalised χjk (τ ) for increasing time steps τ uploading bandwidth, hence when providing contents the selection of peers is performed accordingly. One of the mechanism proposed to increase files availability has been to use multi-torrent, i.e. for ensuring fairness, instead of forcing users stay longer, they provide their contribution to uploaders for fragments belonging to different files [12]. Similarly, in [13] authors show that by using multi-torrent availability can be easily increased, and confirm that fast replication of rare fragments is essential. Furthermore, bundling, i.e. the dissemination of a number of related files together, has been proposed to increase availability [14]. The above proposed mechanisms differ from our proposal, since we take into account several novel factors: the dynamic of data exchange between distant peers, a decay for the availability of peers, and the forecast of contents availability. Such factors have been related into a proposed model that manages to select the rarest content to be spread taking into account the future availability, and the peers that should provide and take such a content. VII. C ONCLUSIONS This paper proposed to improve availability of fragments on a P2P system by adopting a mathematical model and a neural network. The former describes the fragments diffusion and the urgency to share fragments, whereas the latter provides an estimation of the availability of peers, and hence fragments, at later time. By using the estimate of future availability into the diffusion model, we can select the fragments that need to be spread to counteract their disappearance due to users disconnections. ACKNOWLEDGMENT This work has been supported by project PRISMA PON04a2 A/F funded by the Italian Ministry of University and Research within PON 2007-2013 framework. R EFERENCES [1] F. Bannò, D. Marletta, G. Pappalardo, and E. Tramontana. Tackling consistency issues for runtime updating distributed systems. In Proceedings of International Symposium on Parallel & Distributed Processing, Workshops and Phd Forum (IPDPSW), pages 1–8. IEEE, 2010. DOI: 10.1109/IPDPSW.2010.5470863. [2] B. S. Bokshtein, M. I. Mendelev, and D. J. Srolovitz. Thermodynamics and kinetics in materials science: a short course. Oxford University Press Oxford, 2005. [3] F. Bonanno, G. Capizzi, G. Lo Sciuto, C. Napoli, G. Pappalardo, and E. Tramontana. A cascade neural network architecture investigating surface plasmon polaritons propagation for thin metals in openmp. In Proceedings of Artificial Intelligence and Soft Computing, volume 8467, pages 22–33. Springer, 2014. [4] G. Capizzi, C. Napoli, and F. Bonanno. Innovative secondgeneration wavelets construction with recurrent neural networks for solar radiation forecasting. IEEE Transactions on Neural Networks and Learning Systems, 23(11):1805–1815, 2012. [5] G. Capizzi, C. Napoli, and L. Paternò. An innovative hybrid neuro-wavelet method for reconstruction of missing data in astronomical photometric surveys. In Artificial Intelligence and Soft Computing, pages 21–29. Springer, 2012. [6] M. Chiani, D. Dardari, and M. K. Simon. New exponential bounds and approximations for the computation of error probability in fading channels. IEEE Transactions on Wireless Communications, 2(4):840–845, July 2003. [9] R. Giunta, F. Messina, G. Pappalardo, and E. Tramontana. Augmenting a web server with QoS by means of an aspectoriented architecture. In Proceedings of WETICE, pages 179– 184. IEEE, 2012. DOI: 10.1109/WETICE.2012.105. [10] R. Giunta, F. Messina, G. Pappalardo, and E. Tramontana. Providing qos strategies and cloud-integration to web servers by means of aspects. Concurrency and Computation: Practice and Experience, 2013. DOI:10.1002/cpe.3031. [11] R. Giunta, F. Messina, G. Pappalardo, and E. Tramontana. Kaqudai: a dependable web infrastructure made out of existing components. In Proceedings of WETICE, pages 146–151. IEEE, 2013. DOI:10.1109/WETICE.2013.47. [12] L. Guo, S. Chen, Z. Xiao, E. Tan, X. Ding, and X. Zhang. Measurements, analysis, and modeling of bittorrent-like systems. In Proceedings of ACM SIGCOMM conference on Internet Measurement. USENIX Association, 2005. [13] S. Kaune, R. C. Rumin, G. Tyson, A. Mauthe, C. Guerrero, and R. Steinmetz. Unraveling bittorrent’s file unavailability: Measurements and analysis. In Proceedings of P2P, pages 1–9. IEEE, 2010. [14] D. S. Menasche, A. A. Rocha, B. Li, D. Towsley, and A. Venkataramani. Content availability and bundling in swarming systems. In Proceedings of Co-NEXT, pages 121– 132. ACM, 2009. [15] C. Napoli, F. Bonanno, and G. Capizzi. Exploiting solar wind time series correlation with magnetospheric response by using an hybrid neuro-wavelet approach. In Advances in Plasma Astrophysics, number S274 in Proceedings of the International Astronomical Union, pages 250–252. Cambridge University Press, 2010. [16] C. Napoli, F. Bonanno, and G. Capizzi. An hybrid neurowavelet approach for long-term prediction of solar wind. In Advances in Plasma Astrophysics, number S274 in Proceedings of the International Astronomical Union, pages 247–249. Cambridge University Press, 2010. [17] C. Napoli, G. Pappalardo, and E. Tramontana. A hybrid neuro–wavelet predictor for qos control and stability. In AI*IA 2013: Advances in Artificial Intelligence, volume 8249, pages 527–538. Springer, 2013. [18] G. Novelli, G. Pappalardo, C. Santoro, and E. Tramontana. A grid-based infrastructure to support multimedia content distribution. In Proceedings of UPGRADE-CN, pages 57–64. ACM, 2007. DOI: 10.1145/1272980.1272983. [19] D. Qiu and R. Srikant. Modeling and performance analysis of bittorrent-like peer-to-peer networks. SIGCOMM Computer Communication Review, 34(4):367–378, 2004. [7] B. Cohen. Incentives build robustness in bittorrent. In Workshop on Economics of Peer-to-Peer systems, volume 6, pages 68–72, 2003. [20] W. Sweldens. The lifting scheme: A construction of second generation wavelets. Journal on Mathematical Analysis, 29(2):511–546, 1998. [8] R. Giunta, F. Messina, G. Pappalardo, L. Toscano, and E. Tramontana. Testing Replica Selection Policies in a PanEuropean Grid VO. In Proceedings of WETICE, pages 210– 215. IEEE, June 2008. DOI: 10.1109/WETICE.2008.48. [21] J. L. Vázquez. The Porous Medium Equation: Mathematical Theory: Mathematical Theory. Oxford University Press, 2006.	9cs.NE
IEEE TRANSACTIONS ON INFORMATION THEORY 1 Sum-networks from incidence structures: construction and capacity analysis arXiv:1611.01887v2 [cs.IT] 29 Jan 2018 Ardhendu Tripathy, Student Member, IEEE, and Aditya Ramamoorthy, Member, IEEE Abstract—A sum-network is an instance of a function computation problem over a directed acyclic network in which each terminal node wants to compute the sum over a finite field of the information observed at all the source nodes. Many characteristics of the well-studied multiple unicast network communication problem also hold for sum-networks due to a known reduction between the two problems. In this work, we describe an algorithm to construct families of sum-network instances using incidence structures. The computation capacity of several of these sumnetwork families is evaluated. Unlike the coding capacity of a multiple unicast problem, the computation capacity of sumnetworks depends on the characteristic of the finite field over which the sum is computed. This dependence is very strong; we show examples of sum-networks that have a rate-1 solution over one characteristic but a rate close to zero over a different characteristic. Additionally, a sum-network can have arbitrarily different computation capacities for different alphabets. Index Terms—network coding, function computation, sumnetworks, characteristic, incidence structures I. I NTRODUCTION Applications as diverse as parallel processing, distributed data analytics and sensor networks often deal with variants of the problem of distributed computation. This has motivated the study of various problems in the fields of computer science, automatic control and information theory. Broadly speaking, one can model this question in the following manner. Consider a directed acyclic network with its edges denoting communication links. A subset of the nodes observe certain information, these nodes are called sources. A different subset of nodes, called terminals, wish to compute functions of the observed information with a certain fidelity. The computation is carried out by the terminals with the aid of the information received over their incoming edges. The demand functions and the network topology are a part of the problem instance and can be arbitrary. This framework is very general and encompasses several problems that have received significant research attention. Prior work [1], [2], [3] concerning information theoretic issues in function computation worked under the setting of correlated information observed at the sources and simple network structures, which were simple in the sense that there were edges connecting the sources to the terminal without any intermediate nodes or relays. For instance, [2] characterizes This work was supported in part by the National Science Foundation (NSF) under grants CCF-1149860, CCF-1320416 and CCF-1718470. The material in this work has appeared in part at the 52nd Allerton Conference on Communication, Control and Computing, 2014 and the 2015 IEEE International Symposium on Information Theory. The authors are with the Department of Electrical and Computer Engineering, Iowa State University, Ames, IA, 50011 USA. e-mail: {ardhendu@iastate.edu, adityar@iastate.edu} the amount of information that a source must transmit so that a terminal with some correlated side-information can reliably compute a function of the message observed at the source and the side-information. Reference [3] considered distributed functional compression, in which two messages are separately encoded and given to a decoder that computes a function of the two messages with an arbitrarily small probability of error. With the advent of network coding [4], [5], the scope of the questions considered included the setting in which the information observed at the sources is independent and the network topology is more complex. Under this setting, information is sent from a source to a terminal over a path of edges in the directed acyclic network with one or more intermediate nodes in it, these relay nodes have no limit on their memory or computational power. The communication edges are abstracted into error-free, delay-free links with a certain capacity for information transfer and are sometimes referred to as bit-pipes. The messages are required to be recovered with zero distortion. The multicast scenario, in which the message observed at the only source in the network is demanded by all terminals in the network, is solved in [4], [5], [6]. A sufficient condition for solvability in the multicast scenario is that each terminal has a max-flow from the source that is at least the entropy rate of the message random process [4]. Reference [6] established that linear network codes over a sufficiently large alphabet can solve this problem and [5] provided necessary and sufficient conditions for solving a multicast problem instance in an algebraic framework. The work in [5] also gave a simple algorithm to construct a network code that satisfies it. Unlike the multicast problem, the multiple unicast problem does not admit such a clean solution. This scenario has multiple source-terminal pairs over a directed acyclic network of bit-pipes and each terminal wants to recover the message sent by its corresponding source with the help of the information transmitted on the network. To begin with, there are problem instances where more than one use of the network is required to solve it. To model this, each network edge is viewed as carrying a vector of n alphabet symbols, while each message is a vector of m alphabet symbols. A network code specifies the relationship between the vector transmitted on each edge of the network and the message vectors, and it solves a network coding problem instance if m = n. It is shown that linear network codes are in general not sufficient to solve this problem [7]. One can define the notion of coding capacity of a network as the supremum of the ratio m/n over all network codes that allow each terminal to recover its desired message; this ratio m/n for a particular network code is called IEEE TRANSACTIONS ON INFORMATION THEORY its rate. The coding capacity of a network is independent of the alphabet used [8]. While a network code with any rational rate less than the coding capacity exists by definition and zero-padding, a network code with rate equal to coding capacity does not exist for certain networks, even if the coding capacity is rational [9]. The multi-commodity flow solution to the multiple unicast problem is called a routing solution, as the different messages can be interpreted as distinct commodities routed through the intermediate nodes. It is well-known that in the case of multicast, network coding can provide a gain in rate over traditional routing of messages that scales with the size of the network [10]. However, evaluating the coding capacity for an arbitrary instance of the network coding problem is known to be hard in general [11], [12], [13], [14]. Expanding the scope of the demands of the terminals, [15] considered function computation over directed acyclic networks with only one terminal; the value to be recovered at the terminal was allowed to be a function of the messages as opposed to being a subset of the set of all messages. This computation is performed using information transmitted over the edges by a network code. Analogous to the coding capacity, a notion of computation capacity can be defined in this case. A rate-m/n network code that allows the terminal to compute its demand function has the interpretation that the function can be computed by the terminal m times in n uses of the network. Cut-set based upper bounds for the computation capacity of a directed acyclic network with one terminal were given in [15], [16]. A matching lower bound for function computation in tree-networks was given in [15] and the computation capacity of linear and non-linear network codes for different classes of demand functions was explored in [17]. A different flavor of the function computation problem, often called the sum-network problem, considers directed acyclic networks with multiple terminals, each of which demands the finite-field sum of all the messages observed at the sources [18], [19]. Reference [20] characterized the requirements that sum-networks with two or three sources or terminals must satisfy so that each terminal can recover the sum at unit rate. Similar to the network coding scenario, a sum-network whose terminals are satisfied by a rate-1 network code are called solvable sum-networks. Reference [19] established that deciding whether an arbitrary instance of a sum-network problem instance is solvable is at least as hard as deciding whether a suitably defined multiple unicast instance is solvable. As a result of this reduction the various characteristics of the solvability problem for network coding instances are also true for the solvability problem for sum-networks; this establishes the broadness of the class of sum-networks within all communication problems on directed acyclic networks. While solvable sum-networks and solvable network coding instances are intimately related, the results in this paper indicate that these classes of problems diverge when we focus on coding/computation capacity, which can be strictly less than one. In [8, Section VI], the coding capacity of networks is shown to be independent of the finite field chosen as the alphabet for the messages and the information transmitted over the edges. We show that an analogous statement is not true 2 for sum-networks by demonstrating infinite families of sumnetwork problem instances whose computation capacity vary depending on the finite field alphabet. Moreover, the gap in computation capacity on two different finite fields is shown to scale with the network size for certain classes of sumnetworks. For two alphabets F1 , F2 of different cardinality and a network N , the authors in [8, Theorem VI.5] described a procedure to simulate a rate-m2 /n2 network code on F2 for N using a rate-m1 /n1 network code on F1 for the same network, such that m2 /n2 ≥ (m1 /n1 ) − for any > 0. That procedure does not apply for sum-networks. Along the lines of the counterexample given in [20] regarding minimum maxflow connectivity required for solvability of sum-networks with three sources and terminals, we provide an infinite family of counterexamples that mandate certain value of max-flow connectivity to allow solvability (over some finite field) of a general sum-network with more than three sources and terminals. These sum-network problem instances are arrived at using a systematic construction procedure on combinatorial objects called incidence structures. Incidence structures are structured set systems and include, e.g., graphs and combinatorial designs [21]. We note here that combinatorial designs have recently been used to address issues such as the construction of distributed storage systems [22], [23] and coded caching systems [24]–[26]. This paper is organized as follows. Section II describes previous work related to the problem considered and summarizes the contributions. Section III describes the problem model formally and Section IV describes the construction procedure we use to obtain the sum-network problem instances considered in this work. Section V gives an upper bound on the computation capacity of these sum-networks and Section VI describes a method to obtain linear network codes that achieve the upper bound on rate for several families of the sum-networks constructed. Section VII interprets the results in this paper and outlines the key conclusions drawn in this paper. Section VIII concludes the paper and discusses avenues for future work. II. BACKGROUND , R ELATED W ORK AND S UMMARY OF C ONTRIBUTIONS The problem setting in which we will work is such that the information observed at the sources are independent and uniformly distributed over a finite field alphabet F. The network links are error-free and assumed to have unit-capacity. Each of the possibly many terminals wants to recover the finite field sum of all the messages with zero error. This problem was introduced in the work of [18]. Intuitively, it is reasonable to assume the network resources, i.e., the capacity of the network links and the network structure have an effect on whether the sum can be computed successfully by all the terminals in the network. Reference [20] characterized this notion for the class of sum-networks that have either two sources and/or two terminals. For this class of sum-networks it was shown that if the source messages had unit-entropy, a max-flow of one between each source-terminal pair was enough to solve the problem. It was shown by means of IEEE TRANSACTIONS ON INFORMATION THEORY a counterexample that a max-flow of one was not enough to solve a sum-network with three sources and terminals. However, it was also shown that a max-flow of two between each source-terminal pair was sufficient to solve any sumnetwork with three sources and three terminals. Reference [27] considered the computation capacity of the class of sumnetworks that have three sources and three or more terminals or vice versa. It was shown that for any integer k ≥ 2, there exist three-source, n-terminal sum-networks (where n ≥ 3) whose k . The work most closely related computation capacity is k+1 to this paper is [28], which gives a construction procedure that for any positive rational number p/q returns a sumnetwork whose computation capacity is p/q. Assuming that p and q are relatively prime, the procedure described in [28] constructs a sum-network that has 2q − 1 + 2q−1 sources and 2 2q−1 2q+ 2 terminals, which can be very large when q is large. The authors asked the question if there exist smaller sumnetworks (i.e., with fewer sources and terminals) that have the computation capacity as p/q. Our work in [29] answered it in the affirmative and proposed a general construction procedure that returned sum-networks with a prescribed computation capacity. The sum-networks in [28] could be obtained as special cases of this construction procedure. Some smaller instances of sum-networks for specific values were presented in [30]. Small sum-network instances can be useful in determining sufficiency conditions for larger networks. The scope of the construction procedure proposed in [29] was widened in [31], as a result of which, it was shown that there exist sum-network instances whose computation capacity depends rather strongly on the finite field alphabet. This work builds on the contributions in [29], [31]. In particular, we present a systematic algebraic technique that encompasses the prior results. We also include proofs of all results and discuss the implications of our results in depth. A. Summary of contributions In this work, we define several classes of sum-networks for which we can explicitly determine the computation capacity. These networks are constructed by using appropriately defined incidence structures. The main contributions of our work are as follows. • We demonstrate novel techniques for determining upper and lower bounds on the computation capacity of the constructed sum-networks. In most cases, these bounds match, thus resulting in a determination of the capacity of these sum-networks. • We demonstrate a strong dependence of the computation capacity on the characteristic of the finite field over which the computation is taking place. In particular, for the same network, the computation capacity changes based on the characteristic of the underlying field. This is unlike the coding capacity for the multiple unicast problem which is known to be independent of the network coding alphabet. • Consider the class of networks where every sourceterminal pair has a minimum cut of value at least α, where α is an arbitrary positive integer. We demonstrate that there exists a sum-network within this class (with a large 3 number of sources and terminals) whose computation capacity can be made arbitrarily small. This implies that the capacity of sum-networks cannot be characterized just by individual source-terminal minimum cuts. III. P ROBLEM FORMULATION AND P RELIMINARIES We consider communication over a directed acyclic graph (DAG) G = (V, E) where V is the set of nodes and E ⊆ V × V × Z+ are the edges denoting the delay-free communication links between them. The edges are given an additional index as the model allows for multiple edges between two distinct nodes. For instance, if there are two edges between nodes u and v, these will be represented as (u, v, 1) and (u, v, 2). Subset S ⊂ V denotes the source nodes and T ⊂ V denotes the terminal nodes. The source nodes have no incoming edges and the terminal nodes have no outgoing edges. Each source node si ∈ S observes an independent random process Xi , such that the sequence of random variables Xi1 , Xi2 , . . . indexed by time (denoted by a positive integer) are i.i.d. and each Xij takes values that are uniformly distributed over a finite alphabet F. The alphabet F is assumed to be a finite field with \|F\| = q and its characteristic denoted as ch(F). Each edge represents a communication channel of unit capacity, i.e., it can transmit one symbol from F per time slot. When referring to a communication link (or edge) without its third index, we will assume that it is the set of all edges between its two nodes. For such a set denoted by (u, v), we define its capacity cap(u, v) as the number of edges between u and v. We use the notation In(v) and In(e) to represent the set of incoming edges at node v ∈ V and edge e ∈ E. For the edge e = (u, v) let head(e) = v and tail(e) = u. Each terminal node t ∈ T demands thePsum (over F) of the individual source messages. Let Zj = {i:si ∈S} Xij for all j ∈ N (the set of natural numbers); then each t ∈ T wants to recover the sequence Z := (Z1 , Z2 , . . . ) from the information it receives on its incoming edges, i.e., the set In(t). A network code is an assignment of local encoding functions to each edge e ∈ E (denoted as φ̃e (·)) and a decoding function to each terminal t ∈ T (denoted as ψt (·)) such that all the terminals can compute Z. The local encoding function for an edge connected to a set of sources only has the messages observed at those particular source nodes as its input arguments. Likewise, the input arguments for the local encoding function of an edge that is not connected to any source are the values received on its incoming edges and the inputs for the decoding function of a terminal are the values received on its incoming edges. As we consider directed acyclic networks, it can be seen that there is a global encoding function that expresses the value transmitted on an edge in terms of the source messages in the set X := {Xi : si ∈ S}. The global encoding function for an edge e is denoted as φe (X). The following notation describes the domain and range of the local encoding and decoding functions using two natural numbers m and n for a general vector network code. m is the number of i.i.d. source values that are encoded simultaneously by the local encoding function of an edge that emanates from IEEE TRANSACTIONS ON INFORMATION THEORY 4 1 a source node. n is the number of symbols from F that are transmitted across an edge in the network. Thus for such an edge e whose tail(e) = s ∈ S, the local encoding function is φ̃e (Xs1 , Xs2 , . . . , Xsm ) ∈ F n . We will be using both row and column vectors in this paper and they will be explicitly mentioned while defining them. If u is a vector, the uT represents its transpose. • Local encoding function for edge e ∈ E. 2 7 6 φ̃e : F m → F n if tail(e) ∈ S, φ̃e : F n\| In(tail(e))\| → F n if tail(e) ∈ / S. • 3 Decoding function for the terminal t ∈ T . ψt : F n\| In(t)\| → F m . A network code is linear over the finite field F if all the local encoding and decoding functions are linear transformations over F. In this case the local encoding functions can be represented via matrix products where the matrix elements are from F. For example, for an edge e such that tail(e) ∈ / S, let c ∈ N be such that c = \| In(tail(e))\| and In(tail(e)) = {e1 , e2 , . . . , ec }. Also, let each φei (X) ∈ F n be denoted as a column vector of size n whose elements are from F. Then the value transmitted on e can be evaluated as φe (X) = φ̃e (φe1 (X), φe2 (X), . . . , φec (X)), T = Me φe1 (X)T φe2 (X)T . . . φec (X)T , where Me ∈ F n×nc is a matrix indicating the local encoding function for edge e. For the sum-networks that we consider, a valid (m, n) fractional network code solution over F has the interpretation that the component-wise sum over F of m i.i.d. source symbols can be communicated to all the terminals in n time slots. Definition 1: The rate of a (m, n) network code is defined to be the ratio m/n. A sum-network is solvable if it has a (m, m) network coding solution for some m ∈ N. Definition 2: The computation capacity of a sum-network is defined as m there is a valid (m, n) network code : sup for the given sum-network. n We use different types of incidence structures for constructing sum-networks throughout this paper. We now formally define and present some examples of incidence structures. Definition 3: Incidence Structure. Let P be a set of elements called points, and B be a set of elements called blocks, where each block is a subset of P. The incidence structure I is defined as the pair (P, B). If p ∈ P, B ∈ B such that p ∈ B, then we say that point p is incident to block B. In general B can be a multiset, i.e., it can contain repeated elements, but we will not be considering them in our work. Thus for any two distinct blocks B1 , B2 there is at least one point which is incident to one of B1 and B2 and not the other. We denote the cardinalities of the sets P and B by the constants v and b respectively. Thus the set of points and blocks can be indexed by a subscript, and we have that P = {p1 , p2 , . . . , pv }, and B = {B1 , B2 , . . . , Bb }. 4 5 Fig. 1. A pictorial depiction of the Fano plane. The point set P = {1, . . . , 7}. The blocks are indicated by a straight line joining their constituent points. The points 2, 4 and 6 lying on the circle also depict a block. Definition 4: Incidence matrix. The incidence matrix associated with the incidence structure I is a (0, 1)-matrix of dimension v × b defined as follows. AI (i, j) := 1 0 if pi ∈ Bj , otherwise. Thus, incidence matrices can be viewed as general set systems. For example, a simple undirected graph can be viewed as an incidence structure where the vertices are the points and edges are the blocks (each block is of size two). Combinatorial designs [21] form another large and well-investigated class of incidence structures. In this work we will use the properties of t-designs which are defined next. Definition 5: t-design. An incidence structure I = (P, B) is a t-(v, k, λ) design, if • • • it has v points, i.e., \|P\| = v, each block B ∈ B is a k-subset of the point set P, and P and B satisfy the t-design property, i.e., any t-subset of P is present in exactly λ blocks. A t-(v, k, λ) design is called simple if there are no repeated blocks. These designs have been the subject of much investigation when t = 2; in this case they are also called balanced incomplete block designs (BIBDs). Example 1: A famous example of a 2-design with λ = 1 is the Fano plane I = (P, B) shown in Figure 1. Letting numerals denote points and alphabets denote blocks for this design, we can write: P = {1, 2, 3, 4, 5, 6, 7}, B = {A, B, C, D, E, F, G}, where A = {1, 2, 3}, B = {3, 4, 5}, C = {1, 5, 6}, D = {1, 4, 7}, E = {2, 5, 7}, F = {3, 6, 7}, G = {2, 4, 6}. The corresponding incidence matrix AI , with rows and columns arranged in numerical and alphabetical order, is IEEE TRANSACTIONS ON INFORMATION THEORY 5 shown below.  1 1  1  AI =  0 0  0 0 0 0 1 1 1 0 0 1 0 0 0 1 1 0 1 0 0 1 0 0 1 0 1 0 0 1 0 1 0 0 1 0 0 1 1  0 1  0  1 . 0  1 0 (1) It can be verified that every pair of points in P appears in exactly one block in B. There are some well-known conditions that the parameters of a t-(v, k, λ) design satisfy (see [21]). • For integer i ≤ t the number of blocks incident to any i-subset of P is the same. We let bi denote that constant. Then, v−i k−i bi = λ / , ∀i ∈ {0, 1, 2, . . . , t}. (2) t−i t−i • We note that b0 is simply the total number of blocks denoted by b. Likewise, b1 represents the number of blocks that each point is incident to; we use the symbol ρ to represent it. Furthermore, bt = λ. It follows that a necessary condition for the existence of v−i a t-(v, k, λ) design is that k−i divides λ for all t−i t−i i = 1, 2, . . . , t. Counting the number of ones in the point-block incidence matrix for a particular design in two different ways, we arrive at the equation bk = vρ. IV. C ONSTRUCTION OF A FAMILY OF SUM - NETWORKS Let [t] := {1, 2, . . . , t} for any t ∈ N. Our construction takes as input a (0, 1)-matrix A of dimension r × c. Definition 6: Notation for row and column of A. Let pi denote the i-th row vector of A for i ∈ [r] and Bj denote the j-th column vector of A for j ∈ [c] 1 . It turns out that the constructed sum-networks have interesting properties when the matrix A is the incidence matrix of appropriately chosen incidence structures. The construction algorithm is presented in Algorithm 1. The various steps in the algorithm that construct components of the sum-network G = (V, E) are described below. 1) Source node set S and terminal node set T : S and T both contain r+c nodes, one for each row and column of A. The source nodes are denoted at line 4 as spi , sBj if they correspond to the i-th row, j-th column respectively. The terminal nodes are also denoted in a similar manner at line 5. They are added to the vertex set V of the sumnetwork at line 6. 2) Bottleneck edges: We add r unit-capacity edges indexed as ei for i ∈ [r] in line 2 to the edge set E. Each edge ei corresponds to a row of the matrix A. We also add the required tail and head vertices of these edges to V . 1 A justification for this notation is that later when we use the incidence matrix (AI ) of an incidence structure I to construct a sum-network, the rows and columns of the incidence matrix will correspond to the points and blocks of I respectively. 3) Edges between S ∪ T and the bottleneck edges: For every i ∈ [r], we connect tail(ei ) to the source node corresponding to the row pi and to the source nodes that correspond to all columns of A with a 1 in the i-th row. This is described in line 8 of the algorithm. Line 9 describes a similar operation used to connect each head(ei ) to certain terminal nodes. 4) Direct edges between S and T : For each terminal in T , these edges directly connect it to source nodes that do not have a path to that particular terminal through the bottleneck edges. Using the notation for rows and columns of the matrix A, they can be characterized as in lines 12 and 15. Algorithm 1 SUM-NET-CONS Input: A. Output: G = (V, E). 1: Initialize V, E, S, T ← φ. 2: E ← {ei : i ∈ [r]}. 3: V ← {head(ei ), tail(ei ) : i ∈ [r]}. 4: S ← {spi : i ∈ [r]} ∪ {sBj : j ∈ [c]}. 5: T ← {tpi : i ∈ [r]} ∪ {tBj : j ∈ [c]}. 6: V ← V ∪ S ∪ T . 7: for all i ∈ [r] do 8: E ← E ∪ {(sBj , tail(ei )) : A(i, j) = 1; j ∈ [c]} ∪ {(spi , tail(ei ))}. 9: E ← E ∪ {(head(ei ), tBj ) : A(i, j) = 1; j ∈ [c]} ∪ {(head(ei ), tpi )}. 10: end for 11: for all i ∈ [r] do 12: E ← E ∪ {(spj , tpi ) : i 6= j; j ∈ [r]} ∪ {(sBj , tpi ) : A(i, j) = 0; j ∈ [c]}. 13: end for 14: for all j ∈ [c] do 15: E ← E ∪ {(spi , tBj ) : A(i, j) = 0; i ∈ [r]} ∪ {(sBj0 , tBj ) : BjT Bj 0 = 0; j 0 ∈ [c]}. 16: end for 17: return G ← (V, E). For an incidence structure I, let AI represent its incidence matrix. The sum-networks constructed in the paper are such that the matrix A used in the SUM-NET-CONS algorithm is either equal to AI or ATI for some incidence structure I. When A = AI , we call the sum-network constructed as the normal sum-network for I. Otherwise when A = ATI , we call the sum-network constructed as the transpose sum-network for I. The following definitions are useful for analysis. For every p ∈ P, we denote the set of blocks that contain the point p as hpi := {B ∈ B : p ∈ B}, (3) and for every B ∈ B, the collection of blocks that have a non-empty intersection with B is denoted by the set hBi := {B 0 ∈ B : B 0 ∩ B 6= φ} 0 T 0 = {B ∈ B : B B 6= 0}, (4) (5) where boldface B indicates the column of AI corresponding to block B ∈ B. IEEE TRANSACTIONS ON INFORMATION THEORY s1 6 sB s2 s{1,2} sA e1 s1 e2 sC s2 s3 e2 t1 t{1,2} t2 s{1,2} s2 e1 e3 (a) s1 t1 e{1,2} t2 t3 tA tC tB t1 t{1,2} t2 (b) Fig. 2. Two sum-networks obtained from the line graph on two vertices described in Example 2. The source set S and the terminal set T contain three nodes each. All edges are unit-capacity and point downward. The edges with the arrowheads are the bottleneck edges constructed in step 2 of the construction procedure. (a) Normal sum-network, and (b) transposed sumnetwork. The inner product above is computed over the reals. In the sequel, we will occasionally need to perform operations similar to the inner product over a finite field. This shall be explicitly pointed out. We now present some examples of sum-networks constructed using the above technique. Example 2: Let I be the unique simple line graph on two vertices, with points corresponding to the vertices and blocks corresponding to the edges of the graph. Denoting the points as natural numbers, we get that P = {1, 2} and B = {{1, 2}}. Then the associated incidence matrices are as follows. AI = 1 , and ATI = 1 1 1 . Following the SUM-NET-CONS algorithm the two sumnetworks obtained are as shown in the Figure 2. Example 3: In this example we construct a sum-network using a simple t-design. Let I denote the 2-(3, 2, 1) design with its points denoted by the numbers {1, 2, 3} and its blocks denoted by the letters {A, B, C}. For this design we have that A = {1, 2}, B = {1, 3}, C = {2, 3} and its associated incidence matrix under row and column permutations can be Fig. 3. The normal sum-network obtained for the incidence structure I described in Example 3. All edges are unit-capacity and directed downward. The edges with the arrowheads are the bottleneck edges, and the edges denoted by dashed lines correspond to the direct edges introduced in step 4 of the construction procedure. For this case, the normal and the transposed sumnetwork are identical. written as follows.  1 AI = 1 0 1 0 1  0 1 1 Note that AI = ATI . Hence the normal sum-network and the transposed sum-network are identical in this case. Following the SUM-NET-CONS algorithm, we obtain the sum-network shown in Figure 3. Remark 1: Note that each edge added in the SUM-NETCONS algorithm has unit capacity. Proposition 6 in Section VII modifies the SUM-NET-CONS algorithm so that each edge e in the sum-network has cap(e) = α > 1, α ∈ N. V. U PPER BOUND ON THE COMPUTATION CAPACITY In this section, we describe an upper bound on the computation capacity of a sum-network obtained from a (0, 1)-matrix A of dimension r × c. We assume that there exists a (m, n) fractional network code assignment, i.e., φ̃e for e ∈ E (and corresponding global encoding functions φe (X)) and decoding functions ψt for t ∈ T so that all the terminals in T can recover the sum of all the independent sources. For convenience of presentation, we will change notation slightly and let the messages observed at the source nodes corresponding to the rows of A as Xpi for i ∈ [r] and those corresponding to the columns of A as XBj for j ∈ [c]. Each of the messages is a column vector of length m over F. The set of all source messages is represented by X. We let φe (X) denote the n-length column vector of symbols from F that are IEEE TRANSACTIONS ON INFORMATION THEORY 7 transmitted by the edge e ∈ E, as it is the value returned by the global encoding function φe for edge e on the set of source messages denoted by X. As is apparent, non-trivial encoding functions can only be employed on the bottleneck edges, i.e., ei for i ∈ [r] as these are the only edges that have more than one input. For brevity, we denote φi (X) = φei (X). We define the following set of global encoding functions. φIn(v) (X) := {φe (X) : e ∈ In(v)}, ∀v ∈ V. Let H(Y ) be the entropy function for a random variable Y . We let {Yi }l1 denote the set {Y1 , Y2 , . . . , Yl } for any l > 1. The following lemma demonstrates that certain partial sums can be computed by observing subsets of the bottleneck edges. Lemma 1: If a network code allows each terminal 0 to compute P the demanded sum, then the value Xpi := Xpi + j:A(i,j)=1 XBj can be computed from φi (X), i.e., H Xp0 i \|φi (X) = P 0 for all i ∈ [r]. Similarly for any j ∈ [c] P 0 := the value XB X + 0 p i i:A(i,j)=1 j :Bj 0 ∈hBj i XBj 0 can j be computed from the set of values {φi (X) : for i ∈ [r], A(i, j) = 1}. Proof: We let for any i ∈ [r] X X X Z1 = Xpi0 , Z2 = XBj and Z3 = XBj , i0 6=i j:A(i,j)=1 Pr Xp0 i = x0pi , {Xpi0 = xpi0 : i0 6= i}, {XBj = xBj : A(i, j) = 0}\|φi (X) = φi (x) (a) = Pr(Xp0 i = x0pi , φi (X) = φi (x))/ Pr(φi (X) = φi (x)) · Pr({Xpi0 = xpi0 : i0 6= i}, {XBj = xBj : A(i, j) = 0}) (b) = Pr Xp0 i = x0pi \|φi (X) = φi (x) · Pr {Xpi0 = xpi0 : i0 6= i}, {XBj = xBj : A(i, j) = 0} \|φi (X) = φi (x) , where equalities (a) and (b) are due to the fact that the source messages are independent and φi (x) is only a function of xpi and the set {xBj : A(i, j) = 1}. Since terminal tpi can compute Z, H Z\|φIn(tp ) (X) = 0 i and we get from eq. (6) that H(Xpi + Z2 \|φi (X)) = 0. For the second part of the lemma, we argue similarly as follows. We let for any j ∈ [c] X X Z1 = Xpi , Z2 = Xpi , i:A(i,j)=1 Z3 = j:A(i,j)=0 such that the sum Z = Xpi +Z1 +Z2 +Z3 and Xp0 i = Xpi +Z2 . By our assumption that each terminal can recover the demanded sum, we know that Z can be evaluated from φIn(tp ) (X) for all i ∈ [r], i.e., H Z\|φIn(tpi ) (X) = 0 for i all i ∈ [r]. Since {Xpi0 : i0 6= i} and {XBj : A(i, j) = 0} determine the value of Z1 and Z3 respectively and also determine the values transmitted on each of the direct edges that connect a source node to tpi , we get that H Z\|φIn(tp ) (X) i = H Z\|φi (X), {φ(sp 0 ,tpi ) (X) : i0 6= i}, i {φ(sBj ,tpi ) : A(i, j) = 0} (a) bold lowercase symbols represent specific realizations of the random variables. 0 ≥ H Xpi + Z1 + Z2 + Z3 \|φi (X), {Xpi0 : i 6= i}, {XBj : A(i, j) = 0} = H Xp0 i \|φi (X), {Xpi0 : i0 6= i}, {XBj : A(i, j) = 0} = H Xp0 i , {Xpi0 : i0 6= i}, {XBj : A(i, j) = 0}\|φi (X) − H {Xpi0 : i0 6= i}, {XBj : A(i, j) = 0}\|φi (X) = H Xp0 i \|φi (X) + H {Xpi0 : i0 6= i}, {XBj : A(i, j) = 0}\|Xp0 i , φi (X) − H {Xpi0 : i0 6= i}, {XBj : A(i, j) = 0}\|φi (X) (b) (6) = H Xp0 i \|φi (X) , where inequality (a) follows from the fact that φ(sp 0 ,tpi ) (X) i is a function of Xpi0 for i0 6= i and φ(sBj ,tpi ) (X) is a function of {XBj : A(i, j) = 0} and equality (b) is due to the fact that Xp0 i is conditionally independent of both {Xpi0 : i0 6= i} and {XBj : A(i, j) = 0} given φi (X). This conditional independence can be checked as follows. Let X B∈hBj i i:A(i,j)=0 XB , Z4 = X XB B ∈hB / ji 0 = Z1 +Z3 . By our such that Z = Z1 +Z2 +Z3 +Z4 and XB j assumption, for all j ∈ [c], H Z\|φIn(tB ) (X) = 0. The sets j {Xp : p ∈ / Bj } and {XB : B ∈ / hBj i} determine the value of Z2 and Z4 respectively and also the values transmitted on each of the direct edges that connect a source node to the terminal tBj . Let Φ denote the set {φi (X) : A(i, j) = 1}. Then, H Z\|φIn(tB ) (X) j = H Z1 + Z2 + Z3 + Z4 \|Φ, {φ(spi ,tBj ) (X) : A(i, j) = 0}, {φ(sB ,tBj ) : B ∈ / hBj i} (a) ≥ H (Z1 + Z2 + Z3 + Z4 \|Φ, {Xpi : A(i, j) = 0}, {XB : B ∈ / hBj i}) 0 = H XB \|Φ, {Xpi : A(i, j) = 0}, {XB : B ∈ / hBj i} j 0 = H XB , {X : A(i, j) = 0}, {X : B ∈ / hB i}\|Φ p B j i j = (b) − H ({Xpi : A(i, j) = 0}, {XB : B ∈ / hBj i}\|Φ) 0 H(XB \|Φ) j −H({Xpi : A(i, j) = 0}, {XB : B ∈ / hBj i}\|Φ) 0 + H({Xpi : A(i, j) = 0}, {XB : B ∈ / hBj i}\|XB , Φ) j 0 = H(XB \|Φ). j Inequality (a) is due to the fact that φ(spi ,tBj ) (X) is a function of Xpi and similarly for φ(sB ,tBj ) (X). Equality (b) follows from the fact that Z1 +Z3 is conditionally independent of both {Xpi : A(i, j) = 0} and {XBj0 : B ∈ / hBj i} given the set of random variables {φi (X) : A(i, j) = 1}. This can be verified in a manner similar to as was done previously. This gives us 0 the result that H(XB \|{φi (X) : A(i, j) = 1}) = 0. j IEEE TRANSACTIONS ON INFORMATION THEORY Next, we show the fact that the messages observed at the source nodes are independent and uniformly distributed over F m imply that the random variables Xp0 i for all i ∈ [r] are also uniform i.i.d. over F m . To do that, we introduce some notation. For a matrix N ∈ F r×c , for any two index sets R ⊆ [r], C ⊆ [c], we define the submatrix of N containing the rows indexed by R and the columns indexed by C as N [R, C]. Consider two (0, 1)-matrices N1 , N2 of dimensions r1 × t and t×c2 respectively. Here 1 and 0 indicate the multiplicative and additive identities of the finite field F respectively. The i-th row of N1 is denoted by the row submatrix N1 [i, [t]] ∈ {0, 1}t and the j-th column of N2 be denoted by the column submatrix N2 [[t], j] ∈ {0, 1}t . Then we define a matrix function on N1 N2 that returns a r1 × c2 matrix (N1 N2 )# as follows.  1, if the product N1 [i, [t]] N2 [[t], j] over Z is positive, (N1 N2 )# (i, j) = 0, otherwise. For an incidence structure I = (P, B) with r × c incidence matrix A, let Xp , ∀p ∈ P and XB , ∀B ∈ B be m-length vectors with each component i.i.d. uniformly distributed over F. We collect all the independent source random variables in a column vector X having m(r + c) elements from F as follows T T T T · · · XB XB . X := XpT1 XpT2 · · · XpTr XB c 2 1 Recall that pi denotes the i-th row and Bj denotes the j-th column of the matrix A. For all i ∈ [r] let ei ∈ F r denote the column vector with 1 in its i-th component and zero elsewhere. 0 as defined in lemma 1, one can check that Then for Xp0 i , XB j (⊗ indicates the Kronecker product of two matrices) 0 Xpi = eTi pi ⊗ Im X, for all i ∈ [r] and (7) T 0 T T XBj = Bj (Bj B1 )# . . . (Bj Bc )# ⊗ Im X, (8) for all j ∈ [c] where Im is the identity matrix of size m. By stacking these values in the correct order, we can get the following matrix equation. 0T 0 0 0 T Xp1 · · · XpTr XBT1 · · · XBTc = (MA ⊗ Im )X (9) where the matrix MA ∈ F (r+c)×(r+c) is defined as Ir A MA := . (10) AT (AT A)# Note that the first r rows of MA are linearly independent. There is a natural correspondence between the rows of MA and the points and blocks of I of which A is the incidence matrix. If 1 ≤ i ≤ r, then the i-th row MA [i, [r + c]] corresponds to the point pi ∈ P and if r + 1 ≤ j ≤ r + c, then the j-th row MA [j, [r + c]] corresponds to the block Bj ∈ B. 0 Lemma 2: For a (0, 1)-matrix A of size r×c, let Xp0 i , XB ∈ j m F be as defined in Eqs. (7), (8) and matrix MA be as defined in eq. (10). Let r + t := rankF MA for some non-negative integer t and index set S 0 ⊆ {r + 1, r + 2, . . . , r + c} be such that rankF MA [[r] ∪ S 0 , [r + c]] = r + t. Let BS 0 := {BS10 , BS20 , . . . , BSt0 } ⊆ B be the set of blocks that correspond 8 to the rows of MA indexed by S 0 in increasing order. Then we have 0 0 0 0 Pr Xp0 1 = x01 , . . . , Xp0 r = x0r , XB = y , . . . , X = y 1 BS 0 t S0 t 1 = q −m(r+t) , and (11) Pr Xp0 i = x0i = Pr 0 XB S0 j = yj0 = q −m , ∀i ∈ [r], j ∈ [t]. Proof: The quantities in the statement of the lemma satisfy the following system of equations T T T · · · XB (M [[r] ∪ S 0, [r + c]] ⊗ Im ) XpT1 · · · XpTr XB 1 c h 0 iT 0 0 0 T T T T = Xp1 · · · Xpr XBS 0 · · · XBS 0 . 1 t T T T · · · XB The vector XpT1 · · · XpTr XB is uniform 1 c over F m(r+c) . Since the matrix M [[r] ∪ S 0 , [r + c]] ⊗ Im has full row rank equal to m(r + t), the R.H.S. of the above equation is uniformly distributed over F m(r+t) , giving the first statement. The second statement is true by marginalization. Theorem 1: The computation capacity of any sum-network constructed by the SUM-NET-CONS algorithm is at most 1. Proof: By the construction procedure, there is a terminal tpi which is connected to the sources spi and {sBj : A(i, j) = 1} through the edge ei . By lemmas 1 and 2 we have that H(φi (X)) ≥ m log2 q bits. From the definition of n the maximum amount of information transmitted on ei is n log2 q bits and that implies that m ≤ n. Next, we show that the upper bound on the computation capacity exhibits a strong dependence on the characteristic of the field (denoted ch(F)) over which the computation takes place. Theorem 2: Let A be a (0, 1)-matrix of dimension r × c and suppose that we construct a sum-network corresponding to A using the SUM-NET-CONS algorithm. The matrix MA is as defined in eq. (10). If rankF MA = r + t, the upper bound on computation capacity of the sum-network is r/(r + t). Proof: Let BS 0 ⊆ B be as defined in lemma 2. Then from 0 \|φ (X) = 0, ∀i ∈ [r] and lemmas 1 and 2, we have H X i p i 0 H XB \|{φi (X) : A(i, j) = 1} S0 j H({φi (X)}r1 ) = 0, ∀j ∈ [t]. Hence we have that ≥ m(r + t) log q. From the definition of n, we get nr log q ≥ H({φi (X)}r1 ) ≥ m(r + t) log q =⇒ m/n ≤ r/(r + t). Proposition 1: We have that rankF MA = r + t if and only if rankF (AT A)# − AT A = t. Furthermore, rankF MA = r + c if and only if ch(F) - detZ MA , where detZ indicates the determinant of the matrix with its elements interpreted as 0 or 1 in Z. Proof: From eq. (10), we have that I A MA = rT A (AT A)# I 0 Ir 0 Ir A = rT , (12) A Ic 0 (AT A)# − AT A 0 Ic which gives us the rank condition. Since MA is a (0, 1)matrix, if it has full rank, then its determinant is some non-zero element of F, where F is the base subfield of F having prime order. We could also interpret the elements of MA as integers IEEE TRANSACTIONS ON INFORMATION THEORY 9 and evaluate its determinant detZ MA . Then if MA has full rank, we have that ch(F) - detZ MA . Example 4: Consider the normal sum-network obtained from using the Fano plane for which the incidence matrix AI is as defined in eq. (1), so that r = c = 7. It can be verified that rankGF (2) MAI = 7. Hence theorem 2 gives an upper bound of 1 for the computation capacity. In fact, there is a rate-1 network code that satisfies all terminals in the normal sum-network obtained using the Fano plane as described later in proposition 4. We can obtain a different upper bound on the computation capacity by considering submatrices of MA that do not necessarily contain all the initial r rows. To do this we define a 00 new index set S based on an index set S ⊆ [r] as follows. 00 S ⊆ {r + 1, r + 2, . . . , r + c} such that 00 ∀i ∈ S , AT [i − r, [r]] ∈ Span{Ir [j, [r]] : j ∈ S}. (13) Here Span indicates the subspace spanned by the vectors in a set. The submatrix of MA that contains all the rows indexed 00 00 by numbers in S ∪ S is M [S ∪ S , [r + c]]. Theorem 3: Let A be a (0, 1)-matrix of dimension r × c and suppose that we construct a sum-network corresponding to A using the SUM-NET-CONS algorithm. For any (m, n)network code that enables all the terminals to compute the sum, we must have that \|S\| m ≤ min , n xS S⊆[r] 00 00 where xS := rankF MA [S ∪ S , [r + c]] and S is as defined in eq. (13). Proof: Note that for the choice S = [r], the index set 00 S is the same as the index set S 0 defined in lemma 2 and xS = rankF MA , thus recovering the r/ rankF MA upper bound on the computation capacity from theorem 2. For S = 00 {S1 , . . . , S\|S\| } ⊂ [r], we have an index set T ⊆ S such that 00 xS = rankF MA [S ∪ S , [r + c]], = rankF MA [S ∪ T , [r + c]] = \|S\| + \|T \|. We collect the blocks indexed in increasing order by T in the set BT = {BT1 , . . . , BTy } ⊆ B, where y := \|T \|. Then one can recover the L.H.S. of the following equation from the set of messages {φi (X) : i ∈ S} h 0 XpTS1 0 0 0 · · · XpTS XBTT · · · XBTTy 1 \|S\| MA [S, [r + c]] = ⊗ Im X. MA [T , [r + c]] iT Hence we have that q n\|S\| ≥ q m(\|S\|+y) =⇒ m/n ≤ \|S\|/xS . The same reasoning is valid for any choice of S ⊆ [r] and that gives us the result. Example 5: Consider the transposed sum-network corresponding to the undirected graph G shown in Figure 4. One can check that the matrix MATG when the rows and 3 2 B A 1 C 5 D 6 4 Fig. 4. A simple undirected graph G with two connected components. It has 6 vertices and 4 edges. columns of the incidence matrix ATG are arranged in increasing alphabetical and numeric order is as follows.   1 0 0 0 1 1 0 0 0 0  0 1 0 0 1 0 1 0 0 0     0 0 1 0 1 0 0 1 0 0     0 0 0 1 0 0 0 0 1 1     1 1 1 0 1 1 1 1 0 0    MATG =    1 0 0 0 1 1 0 0 0 0   0 1 0 0 1 0 1 0 0 0     0 0 1 0 1 0 0 1 0 0     0 0 0 1 0 0 0 0 1 1  0 0 0 1 0 0 0 0 1 1 We choose our finite field alphabet to be GF (3) in this example. Then rankGF (3) MATG = 5 and theorem 2 gives that the computation capacity is at most 4/5. However, theorem 3 gives a tighter upper bound in this case. Specifically, if S = {1, 2, 3} 00 00 then S = {5, 6, 7, 8} and rankGF (3) MATG [S ∪ S , [10]] = 4. Hence theorem 3 states that the computation capacity of the transposed sum-network for the graph G is at most 3/4. We apply the above theorems to obtain characteristic dependent upper bounds on the computation capacity of some infinite families of sum-networks constructed using the given procedure. Corollary 1: Let I = (P, B) be an incidence structure obtained from a simple undirected graph where P denotes the set of vertices and B consists of the 2-subsets of P corresponding to the edges. Let deg(p) ∈ Z represent the degree of vertex p ∈ P. The incidence matrix AI has dimension \|P\| × \|B\|. The computation capacity of the normal \|P\| sum-network constructed using AI is at most \|P\|+\|B\| for any finite field F. Let F be the finite field alphabet of operation and define P 0 ⊆ P as P 0 := {p : ch(F) - (deg(p) − 1), p ∈ P}. Consider the set of edges B 0 := ∪p∈P 0 hpi. The computation capacity of 0 \| the transposed sum-network is at most \|B0\|B \|+\|P 0 \| . Proof: Recall that BiT is the i-th row of ATI for all i ∈ [\|B\|]. Then the inner product over F between two rows is   2 (mod ch(F)), if i = j,   if edges indexed by i and T Bi Bj = 1, j have a common vertex,    0, otherwise. IEEE TRANSACTIONS ON INFORMATION THEORY 10 It can be observed that the matrix of interest, i.e., (ATI AI )# − ATI AI = −I\|B\| has full rank over every finite field. The transposed sum-network for I is obtained by applying the SUM-NET-CONS algorithm on the \|B\|×\|P\| matrix ATI , so that the parameters r = \|B\|, c = \|P\|. We apply theorem 3 by choosing the index set S ⊆ [\|B\|] such that S = {j : Bj ∈ B 0 }. 00 Defined this way, \|S\| = \|B 0 \| and S is obtained from S using eq. (13). We collect all the points corresponding to the rows in 00 the submatrix MATI [S , [r + c]] in a set PS 00 ⊆ P. Note that PS 00 depends on the set of edges B 0 . By definitions of B 0 and 00 S , we have that P 0 ⊆ PS 00 . This is true because B 0 consists of all the edges that are incident to at least one point in P 0 00 while indices in the set S correspond to all points that are not incident to any edge outside B 0 . For instance, in Example 5 above, as F = GF (3), P 0 = {1}. Then B 0 = {A, B, C} and PS 00 = {1, 2, 3, 4}. 00 We now show that rankF MA [S ∪ S ] = \|B 0 \| + \|P 0 \| and that gives us the result using theorem 3. Recall that pi denotes the i-th row of AI , which corresponds to the vertex pi for all i ∈ [\|P\|]. It follows that the inner product between pi , pj over F is   deg(pi ) (mod ch(F)), if i = j, T pi pj = 1, if {i, j} ∈ B,   0, otherwise. together in exactly one block. Thus, we have the inner product over F as ( v−1 (mod ch(F)), if j = i, T pi pj = k−1 1, otherwise. h i v−1 This implies that AI ATI − (AI ATI )# = − 1 Iv = k−1 i h v−k k−1 Iv and setting its determinant non-zero gives the result. For the normal sum-network, we argue as follows. Note that BiT Bi = k (mod ch(F)) for any i. Since any two points determine a unique block, two blocks can either have one point or none in common. Hence, for i 6= j, the inner product over F is ( 1, if Bi ∩ Bj 6= ∅, T Bi Bj = 0, otherwise. Because of the above equation, all the off-diagonal terms in the matrix (AI ATI )# − AI ATI are equal to zero. We focus 00 on the submatrix M [S ∪ S , [r + c]] obtained from eq. (12), 00 00 letting S\|B\| = {j − \|B\| : j ∈ S } we get that ch(F) - [ρ − b2 + v(b2 − 1)](ρ − b2 )v−1 . # 0 h i 00 00 00 M [S ∪ S , [r + c]] = AI [S\|B\| , S] I\|P\| S\|B\| , S\|B\| I ATI · Λ · \|B\| , 0 I\|P\| " 00 I\|B\| [S, S] Proof: By definition, we have that the inner product over F between two rows is ( ρ (mod ch(F)), if j = i, T pi pj = b2 (mod ch(F)), otherwise. It follows that AI ATI − (AI ATI )# has the value (ρ − 1) on the diagonal and (b2 − 1) elsewhere. That is, AI ATI − (AI ATI )# where Λ := Then ATI AI − (ATI AI )# = [(k − 1)] Ib and setting its determinant as non-zero gives the result. Corollary 3: Let I = (P, B) be a t-(v, k, λ) design,for t ≥ k−1 2. From eq. (2), each point is present in ρ := λ v−1 t−1 / t−1 blocks and the number of blocks incident to any pair of points k−2 is given by b2 := λ v−2 / t−2 t−2 . Consider the transposed sumnetwork constructed using the incidence matrix ATI which has dimension \|B\| × \|P\|. The computation capacity of the \|B\| transposed sum-network is at most \|B\|+\|P\| if " I\|B\| [S, [\|B\|]] 0 # 0 (AI ATI )# − AI ATI h i 00 . S\|B\| , [\|P\|] By definition of P 0 the points in the set PS 00 \ P 0 are such that deg(pi ) − 1 ≡ 0 (mod ch(F)), i.e., the diagonal entry corresponding to those points in (AI ATI )# − AI ATI in the matrix Λ is zero. Thus, Λ has exactly \|B 0 \| + \|P 0 \| rows which are not equal to the all-zero row vector. The first and third matrices are invertible, and hence we get that rankF MA [S ∪ 00 S , [r + c]] = \|B 0 \| + \|P 0 \|. Corollary 2: Let I = (P, B) be a 2-(v, k, 1) design. For the normal sum-network constructed using the \|P\| × \|B\| incidence \|P\| matrix AI , the computation capacity is at most \|P\|+\|B\| if ch(F) - (k − 1). For the transposed sum-network constructed \|B\| using ATI , the computation capacity is at most \|P\|+\|B\| if v−k ch(F) - k−1 . Proof: We first describe the case of the transposed sumnetwork. From eq. (2) each point in a 2-(v, k, 1) design is v−1 incident to ρ = k−1 blocks. Moreover any two points occur = [(ρ−b2 ) (mod ch(F))] Iv + [(b2 −1) (mod ch(F))] Jv , where Jv denotes the square all ones matrix of dimension v. Then by elementary row and columns operations, det AI ATI − (AI ATI )# can be evaluated to be equal to [ρ − b2 + v(b2 − 1)](ρ − b2 )v−1 (mod ch(F)). Corollary 4: Let D = (P, B) be a t-(v, t + 1, λ) design with λ 6= 1 and incidence matrix AD . We define a higher incidence matrix AD0 of dimension \|P\| t × \|B\| such that each row corresponds to a distinct t-subset of P and each column corresponds to ablock in B. AD0 is a (0, 1)-matrix such that for any i ∈ vt , j ∈ [\|B\|], its entry AD0 (i, j) = 1 if each of the points in the t-subset corresponding to the i-th row is incident to the block Bj ∈ B and zero otherwise. The computation capacity of the normal sum-network constructed using (vt) t+1 = λ+t+1 if ch(F) - t. The computation AD0 is at most v +\|B\| (t) capacity of the transposed sum-network constructed using ATD0 \|B\| λ is at most \|B\|+ = λ+t+1 if ch(F) - (λ − 1). (vt) Proof: The incidence matrix AD0 is a (0, 1) matrix of λ v dimension vt × t+1 t . Let pi , Bu denote the i-th row and IEEE TRANSACTIONS ON INFORMATION THEORY 11 h i λ v u-th column respectively of AD0 for i ∈ vt , u ∈ t+1 t . Each row of AD0 corresponds to a distinct t-subset of P. By tdesign criterion, any set of t points belongs to exactly λ blocks. Since the columns have a one-to-one correspondence with the blocks in B, each row of AD0 has exactly λ 1’s. Two rows will have a 1 in the same column if the block corresponding to the column is incident to both the t-subsets corresponding to the two rows. Since each block has t + 1 points, there cannot be more than one block incident to two different t-subsets. T Hence, for the inner product over F, we vhave that pi pi = λ (mod ch(F)) and for all i 6= j; i, j, ∈ t ,  1, if the union of the t-subsets corresponding to the i-th and j-th rows is a block in B, pi pTj = 0, otherwise. most 2/3. We describe a network code with m = 2, n = 3. The global encoding functions for the two bottleneck edges are shown in Table I. Using the values transmitted, all three Then AD0 ATD0 − (AD0 ATD0 )# = [(λ − 1) (mod ch(F))] I(v) t and that gives the result for the transposed sum-network. For the normal sum-network, we look at the columns of AD0 in a similar manner. Each column of AD0 corresponds to a block in B. Since the size of each block is t+1, each column has exactly t+1 = t + 1 elements as 1. Also, two different t blocks can have at most t points in common, and only when that happens, will the two columns have a 1 in the same row. Hence, for the inner product over F, we have that BuT Bu = v (t + 1) (mod ch(F)) and for all u 6= v; u, v ∈ t ,  1, if the u-th and v-th blocks have t points in common, BuT Bv = 0, otherwise. terminals can recover the sum in the following manner. t1 receives the value of X2 from the direct edge (s2 , t1 ) while t2 receives the value of X1 from the direct edge (s1 , t2 ). Then t1 recovers the sum using the first two components of φ1 (X) while t2 recovers the sum using the first two components of φ2 (X). Additionally, t{1,2} receives both φ1 (X), φ2 (X) and can carry out the operation (X1 + X{1,2} ) + (X2 + X{1,2} ) − X{1,2} . Thus, each terminal is satisfied. Then ATD0 AD0 − (ATD0 AD0 )# = t (mod ch(F))I λ (v) and t+1 t theorem 2 gives the result. VI. L INEAR NETWORK CODES FOR CONSTRUCTED SUM - NETWORKS In this section, we propose linear network codes for the sum-networks constructed using the SUM-NET-CONS algorithm. Recall that the algorithm takes a (0, 1)-matrix A that has r rows and c columns as its input. In Section V, we demonstrated that the incidence matrix of certain incidence structures result in sum-networks whose capacity can be upper bounded (cf. Corollaries 1, 2, 4). We now demonstrate that under certain conditions, we can obtain network codes whose rate matches the corresponding upper bound. Thus, we are able to characterize the capacity of a large family of sum-networks. We emphasize that random linear network codes that have been used widely in the literature for multicast code constructions are not applicable in our context. In particular, it is not too hard to argue that a random linear network code would result in each terminal obtaining a different linear function or subspace. Thus, constructing codes for these sum-networks requires newer ideas. We outline the key ideas by means of the following example. Example 6: Consider the sum-network shown in Figure 2(a). The matrix AI used in its construction is of dimension r × c where r = 2, c = 1 and is described in Example 2. It can be observed that ATI AI − ATI AI # = 1. Then theorem 2 states that the computation capacity of this sum-network is at TABLE I T HE FUNCTION VALUES TRANSMITTED ACROSS e1 , e2 IN F IGURE 2( A ) FOR A NETWORK CODE WITH RATE = 2/3. E ACH MESSAGE X1 , X2 , X{1,2} IS A VECTOR WITH 2 COMPONENTS , AND φ1 (X), φ2 (X) ARE VECTORS WITH 3 COMPONENTS EACH . A NUMBER WITHIN SQUARE BRACKETS ADJOINING A VECTOR INDICATES A PARTICULAR COMPONENT OF THE VECTOR . Component 1 2 3 φ1 (X) X1 [1] + X{1,2} [1] X1 [2] + X{1,2} [2] X{1,2} [1] φ2 (X) X2 [1] + X{1,2} [1] X2 [2] + X{1,2} [2] X{1,2} [2] The network code in the example has the following structure. For each bottleneck edge, the first r components of the global encoding vector are the sum of all messages that are incident to that bottleneck. The remaining c components of the encoding vectors transmit certain components of messages observed at source nodes that correspond to columns in the matrix AI . In the example, t{1,2} received the first component of X{1,2} from φ1 (X) and the second component from φ2 (X). Thus it was able to recover the value of X{1,2} , which it used in computing the demanded sum. Our construction of network codes for sum-networks will have this structure, i.e., the first r components on a bottleneck edge will be used to transmit a partial sum of the messages observed at the sources that are connected to that bottleneck edge and the remaining c components will transmit portions of certain sources in an uncoded manner. For a given incidence matrix A, our first step is to identify (if possible) a corresponding non-negative integral matrix D of the same dimensions with the following properties. • D(i, j) = 0 if A(i, j) = 0. • Each row in D sums to r. • Each column in D sums to c. Under certain conditions on the incidence matrix A, we will show that D can be used to construct suitable network codes for the sum-networks under consideration. The existence of our proposed network codes are thus intimately related to the existence of non-negative integral matrices that satisfy certain constraints. The following theorem [32, Corollary 1.4.2] is a special case of a more general theorem in [33] that gives the necessary and sufficient conditions for the existence of non-negative integral matrices with constraints on their row and column sums. We give the proof here since we use some ideas from it in the eventual network code assignment. IEEE TRANSACTIONS ON INFORMATION THEORY Theorem 4: Let R = (r1 , r2 , . . . , rm ) and S = (s1 , s2 , . . . , sn ) be non-negative integral vectors satisfying r1 +. . .+rm = s1 +. . .+sn . There exists an m×n nonnegative integral matrix D such that 0 ≤ D(i, j) ≤ cij , ∀i ∈ [m], ∀j ∈ [n], n X D(i, j) = ri , ∀i ∈ [m], and 12 Theorem 5: Let I = (P, B) be an incidence structure and let AI denote the corresponding incidence matrix of dimension v × b. Suppose that the following conditions are satisfied. T T • AI AI − (AI AI )# = diag(µ1 , . . . , µb ) (mod ch(F)), where µi is a non-zero element of F ∀i ∈ {1, 2, . . . , b}. • There exists a matrix DI with integer elements of the same dimension as AI whose entries satisfy j=1 m X i=1 v X D(i, j) = sj , ∀j ∈ [n] if and only if for all I ⊆ [m] and J ⊆ [n], we have that XX X X cij ≥ sj − ri . (14) i∈I j∈J j∈J i∈I / Proof: Consider a capacity-limited flow-network modelled using a bipartite graph on m + n nodes. The left part has m nodes denoted as xi , ∀i ∈ [m] and the right part has n nodes denoted as yj , ∀j ∈ [n]. For all i, j there is a directed edge (xi , yj ) with capacity cij . There are two additional nodes in the flow-network, the source node S ∗ and terminal node T ∗ . There are directed edges (S ∗ , xi ) with capacity ri for all i ∈ [m] and directed edges (yj , T ∗ ) with capacity sj for all j ∈ [n]. Let xI be the set of all nodes in the left part whose indices are in I and let yJ¯ be the set of all nodes in the right part whose indices are not in J. Consider a cut separating nodes in {S ∗ } ∪ xI ∪ yJ¯ from its complement. Let f ∗ be the value of the maximum S ∗ -T ∗ flow in this network. Then we must have that for all possible choice of subsets I ⊆ [m], J ⊆ [n], X X X ri + cij + sj ≥ f ∗ . (15) i∈I / (i,j):i∈I,j∈J ∗ j ∈J / j∈J (17) DI (i, j) = v, and (18) DI (i, j) = b. (19) j=1 Then the computation capacity of the sum-network constructed v . This rate using AI via the SUM-NET-CONS algorithm is v+b can be achieved by a linear network code. Proof: Note that ATI AI − (ATI AI )# has full rank by assumption, theorem 2 states that the computation capacity of the sum-network is at most v/(v + b). We construct a (m, n) linear network code with m = v, n = v + b using the matrix DI . Since m = v, each message vector has v components. For a vector t ∈ F v , the notation t[l1 : l2 ] for two positive integers l1 , l2 ∈ [v] denotes a (l2 −l1 +1) length vector that contains the components of t with indices in the set {l1 , l1 + 1, . . . , l2 } in order. We need to specify the global encoding vectors φi (X) only for the bottleneck edges ei , i ∈ [v] as all the other edges in the network act as repeaters. The linear network code is such that the first v components of the vector transmitted along ei ∀i ∈ [v] is X φi (X)[1 : v] = Xpi + XBj . j:AI (i,j)=1 P In particular, suppose that f = j∈[n] sj in the flownetwork. Substituting this in eq. (15), we get the condition that for all possible subsets I ⊆ [m], J ⊆ [n], XX X X cij ≥ sj − ri . (16) i∈I j∈J i=1 b X DI (i, j) = 0, if AI (i, j) = 0, i∈I / Note that by choosing all possible subsets I, J, we are considering every possible S ∗ -T ∗ cut in the network. Then by the maxflow-mincut theorem, the set of conditions of the form of eq. (16) for all I, J are not only necessaryPbut also sufficient for the existence of a flow of value f ∗ = j∈[n] sj in the network. A feasible flow with this value can be used to arrive at the matrix D as follows. We set the value of element D(i, j) in the matrix to be equal to the value of the feasible flow on the edge (xi , yj ) for all i ∈ [m], j ∈ [n]. It is easy to verify that the matrix D satisfies the required conditions. Using the existence theorem for nonnegative integral matrices, we can obtain network codes for sum-networks constructed from certain incidence structures. The following theorem describes a set of sufficient conditions that, if satisfied by an incidence structure, allow us to construct a linear network code that has the same rate as the computation capacity of that sumnetwork. The proof of the theorem is constructive and results in an explicit network code. By construction, each tpi ∀i ∈ [v] is connected to the source nodes in {spi0 : i0 6= i}∪{sBj : AI (i, j) = 0} by direct edges. tpi can then compute the following value from the information received on the direct edges. X X Xpi + XBj . i0 6=i j:AI (i,j)=0 Adding the above value to φi (X)[1 : v] enables tpi to compute the required sum. In what follows, we focus on terminals of the form tBj ∀j ∈ [b]. Since n = v+b, each vector φi (X) ∈ F n has b components that haven’t been specified yet. We describe a particular assignment for the b components on every φi (X), i ∈ [v] using the matrix DI that enables each tBj ∀j ∈ [b] to compute the sum. Recall the bipartite flow network constructed in the proof of theorem 4. The nodes in the left part are denoted as pi ∀i ∈ [v] and the nodes in the right part are denoted as Bj ∀j ∈ [b]. There is an edge (pi , Bj ) if and only if AI (i, j) = 1. The flow on the edge (pi , Bj ) is denoted as f (pi , Bj ) and its value is determined by DI (i, j), i.e., f (pi , Bj ) := DI (i, j). By constraints on the row and column sums of DI , we conclude that the value of the flow through any pi ∀i ∈ [v] is b and the value of the flow through any Bj ∀j ∈ [b] is v. Without IEEE TRANSACTIONS ON INFORMATION THEORY 13 loss of generality, assume that Bj = {p1 , p2 , · · · , p\|Bj \| }. We can partition the v components of message vector XBj into \|Bj \| parts such that the i-th partition contains f (pi , Bj ) distinct components of XBj . Such a partitioning can be done for all message vectors XBj , j ∈ [b]. Then the flow f (pi , Bj ) indicates that the vector φi (X)[v+1 : v+b] includes f (pi , Bj ) uncoded components of XBj . Assigning such an interpretation to every edge in the flow-network is possible as the total number of components available in each φi (X) is b and that is also equal to the flow through the point pi . By construction, terminal tBj is connected to all bottleneck edges in the set {ei : AI (i, j) = 1}. From the assignment based on the flow, tBj receives f (pi , Bj ) distinct components of XBj from φi (X) for all {i : AI (i, j) = 1}. Since P v i=1 f (pi , Bj ) = v, it can recover all v components of XBj in a piecewise fashion. By adding the first v components transmitted on all the bottleneck edges that are connected to tBj , it can recover X φi (X)[1 : v] 1 2 = Xpi + i:AI (i,j)=1 X = X Xpi + sE sB sA e1 sD e2 e3 tB tE BjT Bl XBl . Bl ∈hBj i Bl ∈hBj i\Bj By the flow-based assignment, each tBj obtains the value of XBj in a piecewise manner. It can then carry out the following X φi (X)[1 : v] − µj XBj Xpi + (µj + 1)XBj + = p∈Bj Xp + X XBl − µj XBj , Bl ∈hBj i\Bj i:AI (i,j)=1 X tC tD (b) A XA [1 : 2] XA [3 : 4] 1 XBl . Bl ∈hBj i The messages not present in this partial sum, i.e., {Xp : p ∈ / Bj } ∪ {XB : B ∈ / hBj i} are available at tBj through direct edges by construction. Hence, terminals that correspond to a column of AI are also able to compute the required sum. We illustrate the linear network code proposed above by means of the following example. Example 7: Consider the normal sum-network obtained from the undirected simple graph G shown in Figure 5(a). A part of the sum-network is shown in Figure 5(b). The 4 × 5 incidence matrix AG satisfies the condition of theorem 4 and therefore has an associated matrix DG with row-sum as 5 and column-sum 4 as shown below. The rows and columns of AG are arranged in increasing numeric and alphabetical order.     1 0 0 1 1 2 0 0 2 1 1 1 0 0 0 2 3 0 0 0    AG =  0 1 1 0 1 , DG = 0 1 1 0 3 . 0 0 1 1 0 0 0 3 2 0 B 2 XD [1 : 2] C 3 XE [1] D 4 E i:AI (i,j)=1 X e4 XBl , Because of the condition that ATI AI − (ATI AI )# = diag(µ1 , µ2 , . . . , µb ), one can verify that X X BjT Bl XBl = (µj + 1)XBj + XBl . = sC Bl ∈hBj i i:AI (i,j)=1 X 3 (a) i:AI (i,j)=1 l:AI (i,l)=1 X C B tA X 4 E A i:AI (i,j)=1 X D (c) Fig. 5. (a) Undirected graph considered in Example 7. (b) Part of the corresponding normal sum-network constructed for the undirected graph in (a). The full normal sum-network has nine nodes each in the source set S and the terminal set T . However, for clarity, only the five sources and terminals that correspond to the columns of the incidence matrix of the graph are shown. Also, the direct edges constructed in Step 4 of the construction procedure are not shown. All edges are unit-capacity and point downward. The edges with the arrowheads are the bottleneck edges constructed in step 2 of the construction procedure. (c) Bipartite flow network as constructed in the proof of theorem 4 for this sum-network. The message values corresponding to the flow on the solid lines are also shown. Using the matrix DG , one can construct a structured linear network code with rate = v/(v + b) = 4/9 as shown in Table II. One can check that it enables all the terminals to compute the required sum. The flow-network corresponding to DG is shown in Figure 5(c), and the messages corresponding to the flow on the solid edges are shown alongside the respective edge. We can also consider the transposed sum-network for the same graph G. Corollary 1 gives an upper bound on the computation capacity that depends on F. If F = GF (2), IEEE TRANSACTIONS ON INFORMATION THEORY 14 then the subset of points P 0 = {2, 4} and the upper bound is 4/6. Note that theorem 5 is not applicable here as the matrix ATG AG − (ATG AG )# does not have all its diagonal elements as non-zero over GF (2). Proposition 3 gives a condition for the existence of a network code for transposed sum-networks obtained using irregular graphs. We apply that condition to the transposed sum-network of the graph G considered here in Example 8. In the following proposition we show that certain infinite families of incidence structures satisfy the requirements stated in theorem 5. In particular, the incidence structures considered in Corollaries 1, 2 and 4 satisfy the conditions and hence the computation capacity of the associated sum-networks can be calculated. Proposition 2: The following incidence structures and their transposes satisfy condition (ii) in theorem 5, i.e., if their incidence matrix of dimension v × b is denoted by AI , there exists a corresponding non-negative integral matrix DI that satisfies the conditions in equations (17) – (19). 1) Incidence structures derived from a regular graph or a biregular bipartite graph. 2) t-(v, k, λ) designs with λ = 1. 3) The higher incidence structure of a t-(n, t + 1, λ) design with λ 6= 1 obtained using the procedure described in corollary 4. Proof: The existence of DI with row-sums as v and column-sums b is the same as the existence of DIT with rowsums as b and column-sums v. Thus, it suffices to argue for DI . To check the validity of the condition we first choose the bounds on the elements of the matrix DI . We set ri = b and sj = v for all i ∈ [v], j ∈ [b] and 0, if AI (i, j) = 0, cij = ∞, if AI (i, j) = 1. By this choice the condition in inequality (14) is trivially satisfied whenever I, J are chosen such that there is a point in I which is incident to some block in J, i.e., there exist i ∈ I, j ∈ J such that AI (i, j) = 1. Hence we restrict our attention to choices of I and J such that none of the points in I are incident to any block in J. Under this restriction, the L.H.S. of inequality (14) is 0 and the condition is equivalent to (v − \|I\|)b ≥ \|J\|v. We will assume that ∃I ⊆ [v], J ⊆ [b] such that (20) AI (i, j) = 0 ∀i ∈ I, j ∈ J, and (v − \|I\|)b < \|J\|v, and show that it leads to a contradiction for each of the three incidence structures considered. If I corresponds to a d-regular simple graph, then b = dv/2. Consider the point-block incidence matrix AI , which is a (0, 1)-matrix of size v × b. For the chosen I in eq. (20), we look at the submatrix AI [I, [b]] of size \|I\| × b that consists of the rows of AI indexed by the points in I and all the columns. Let l1 be the number of columns with a single 1 in AI [I, [b]] and l2 be the number of columns with two 1s in AI [I, [b]]. By counting the total number of 1s in AI [I, [b]] in two ways, we get that d\|I\| . d\|I\| = l1 + 2l2 ≤ 2(l1 + l2 ) =⇒ l1 + l2 ≥ 2 Since the number of edges incident to at least one point in I is l1 + l2 , any subset J of the edges that has no incidence with any point in I satisfies \|J\| ≤ b − d\|I\|/2. Using these in eq. (20) we get that dv d\|I\| dv < − v, (v − \|I\|)b < \|J\|v =⇒ (v − \|I\|) 2 2 2 which is a contradiction. Now suppose that I corresponds to a biregular bipartite graph, with L vertices having degree dL in the left part and R vertices having degree dR in the right part. Then b = LdL = RdR . Consider a subset IL ∪ IR of its vertices. Let EL (resp. ER ) be the set of edges which are incident to some vertex in IL (resp. IR ) but not incident to any vertex in IR (resp. IL ). The number of edges that are not incident to any vertex in IL ∪ IR is equal to (L−\|IL \|)dL −\|ER \| = (R−\|IR \|)dR −\|EL \|. Suppose there is a choice of I in eq. (20) is such that I = IL ∪ IR for some IL , IR . Then we have that (v − \|I\|)b < \|J\|v, LdL + RdR 2 (L − \|IL \|)dL − ER + (R − \|IR \|)dR − \|EL \| (L + R), < 2 \|Il \|dL + \|IR \|dR + \|EL \| + \|ER \| \|IL \| + \|IR \| =⇒ , < LdL + RdR L+R =⇒ (L+R)(\|EL \|+\|ER \|) < (L−R)\|IL \|dL +(R−L)\|IR \|dR , =⇒ (L + R − (\|IL \| + \|IR \|)) =⇒ (L + R)(\|EL \| + \|ER \|) < (L − R)(\|EL \| − \|ER \|). If L > R or \|EL \| > \|ER \|, then we have a contradiction. That leaves the case when L < R and \|EL \| < \|ER \|, which implies (L + R)(\|EL \| + \|ER \|) < (R − L)(\|ER \| − \|EL \|) and that is also a contradiction. Next, consider a t-(v, k, 1) design with b blocks such that repetition degree of each point is ρ and we have that bk = vρ. With the I of eq. (20), we employ a similar procedure as for the case of the d-regular graph. We choose the submatrix AI [I, [b]] of size \|I\| × b that corresponds to the rows indexed by the points in I and let li , ∀i ∈ [k] denote the number of columns with exactly i 1s in AI [I, [b]]. We count the total number of 1s in AI [I, [b]] in two ways, yielding ρ\|I\| = l1 + 2l2 + · · · + (k − 1)lk−1 + klk ≤ k =⇒ k X i=1 li ≥ k X li , i=1 ρ\|I\| b\|I\| = . k v The number of that are incident to at least one point in Pblocks k I is equal to i=1 li . Hence any subset J of blocks that has no incidence with any point in I satisfies \|J\| ≤ b − \|I\|b/v. Using this in eq. (20) we get that \|I\|b (v − \|I\|)b < \|J\|v =⇒ (v − \|I\|)b < b − v, v which is a contradiction. If I = (P, B) is the higher incidence structure obtained from a t-(n, t + 1, λ) design as describedin corollary 4, then λ n we have that \|P\| = nt and \|B\| = t+1 t . By definition of t IEEE TRANSACTIONS ON INFORMATION THEORY 15 TABLE II T HE FUNCTION VALUES TRANSMITTED ACROSS e1 , e2 , e3 , e4 IN F IGURE 5( B ) FOR A NETWORK CODE WITH RATE = 4/9. E ACH MESSAGE XA , XB , XC , XD , XE IS A VECTOR WITH 4 COMPONENTS , AND φ1 (X), φ2 (X), φ3 (X), φ4 (X) ARE VECTORS WITH 9 COMPONENTS EACH . T HE NUMBER INSIDE SQUARE BRACKETS ADJOINING A VECTOR INDICATES A PARTICULAR COMPONENT OF THE VECTOR . Component 1 to 4 5 6 7 8 9 φ1 (X) X1 + XA + XD + XE XA [1] XA [2] XD [1] XD [2] XE [1] φ2 (X) X2 + XA + XB XA [3] XA [4] XB [1] XB [2] XB [3] for the original design, we have that each of the points in P are incident to exactly λ blocks. Also, each block in B consists of t+1 = t + 1 points. For the submatrix AI [I, [b]] whose t rows correspond to the points in I from Condition 20, we let li , ∀i ∈ [t+1] denote the number of columns that have exactly i 1s in them. By counting the total number of 1s in AI [I, [b]] in two ways we get that t+1 X t+1 X t+1 X λ\|I\| . λ\|I\| = ili ≤ (t + 1) li =⇒ li ≥ t+1 i=1 i=1 i=1 The number of blocks incident to at least one point in I Ptotal t+1 is i=1 li . Then the number of blocks \|J\| that are not incident to any point in I satisfy \|J\| ≤ \|B\| − \|I\|λ/(t + 1). Using these we get that (v − \|I\|)b < \|J\|v, n λ n n n λ < − \|I\| , =⇒ − \|I\| t+1 t t+1 t t t which is a contradiction. Thus in all the three kinds of incidence structures considered, we have shown that they admit the existence of the associated matrix DI under the stated qualifying conditions. This enables us to apply theorem 5 and obtain a lower bound on the computation capacity of these sum-networks. For an undirected graph I = (P, B) that is not regular, proposition 2 is not applicable. Theorem 5 describes a sufficient condition for the existence of a linear network code that achieves the upper bound on the computation capacity of normal sum-networks constructed from undirected graphs that are not necessarily regular. The upper bound on the capacity of the transposed sum-network constructed using the incidence \|B\| matrix ATI however can be different from \|B\|+\|P\| depending on the finite field F (cf. corollary 1) and theorem 5 needs to be modified to be applicable in that case. The following example illustrates this. Example 8: Consider the transposed sum-network for the irregular graph G described in Example 7. Corollary 1 gives an upper bound of 4/6 on the computation capacity when F = GF (2), as for that case P 0 = {2, 4} and B 0 = {A, B, C, D}. We show the submatrix ATG [B 0 , P 0 ] in the equation below and also an associated matrix DG whose support is the same as that of ATG [B 0 , P 0 ] and whose row-sum = 6 − 4 = 2 and column- φ3 (X) X3 + XB + XC + XE XB [4] XC [1] XE [2] XE [3] XE [4] φ4 (X) X4 + XC + XD XC [2] XC [3] XC [4] XD [3] XD [4] sum = 4. The rows and columns are arranged in increasing alphabetical and numeric order.     2 0 1 0 2 0 1 0    ATG [B 0 , P 0 ] =  0 1 , DG = 0 2 . 0 2 0 1 Using DG we can construct a rate-4/6 linear network code, shown in Table III, that achieves the computation capacity for F = GF (2) of the transposed sum-network constructed using the irregular graph G shown in Figure 5(a). In particular, terminals t1 , t3 don’t need any information other than the partial sums obtained over their respective bottleneck edges to compute the sum. Terminals t2 , t4 need the value X2 , X4 respectively, and that is transmitted in a piecewise fashion according to the matrix DG over the bottleneck edges. For an undirected graph I = (P, B) that is not regular, let P 0 , B 0 be the set of points and edges as chosen in the statement of corollary 1. We describe a condition on the submatrix ATI [B 0 , P 0 ] which consists of the rows and columns of ATI corresponding to the blocks and points in the sets B 0 , P 0 respectively. This condition allows us to construct a capacityachieving linear network code for the transposed sum-network. Proposition 3: For an undirected graph I = (P, B), let \|P 0 \| = v 0 , \|B 0 \| = b0 , where P 0 , B 0 are subsets of points and blocks as defined in corollary 1 and let ATI [B 0 , P 0 ](i, j) indicate an element of the submatrix for indices i ∈ [b0 ], j ∈ [v 0 ]. Suppose there is a matrix DI of dimension b0 × v 0 such that 0 b X i=1 DI (i, j) = 0, if ATI [B 0 , P 0 ](i, j) = 0, DI (i, j) = b0 , for all j ∈ [v 0 ], and 0 v X j=1 DI (i, j) = v 0 , for all i ∈ [b0 ]. 0 b Then there is linear network code of rate b0 +v 0 that allows each terminal in the transposed sum-network constructed using I to compute the required sum. Proof: We describe a rate-b0 /(b0 + v 0 ) network code that enables each terminal to compute the sum. Then by corollary 1 we know that this is a capacity-achieving code. Since this is a transposed sum-network, the bottleneck edges in the sumnetwork correspond to the blocks in the undirected graph I. IEEE TRANSACTIONS ON INFORMATION THEORY 16 TABLE III T HE FUNCTION VALUES TRANSMITTED ACROSS THE BOTTLENECK EDGES OF THE TRANSPOSED SUM - NETWORK CORRESPONDING TO THE GRAPH SHOWN IN F IGURE 5( A ) FOR A RATE -4/6 NETWORK OVER GF (2). E ACH MESSAGE X2 , X4 IS A VECTOR WITH 4 COMPONENTS , AND φA (X), φB (X), φC (X), φD (X), φE (X) ARE VECTORS WITH 6 COMPONENTS EACH . T HE NUMBER INSIDE SQUARE BRACKETS ADJOINING A VECTOR INDICATES A PARTICULAR COMPONENT OF THE VECTOR . A DASH INDICATES THAT THE VALUE TRANSMITTED ON THAT COMPONENT IS NOT USED IN DECODING BY ANY TERMINAL . Component 1 to 4 5 6 φA (X) X1 + X2 + XA X2 [1] X2 [2] φB (X) X2 + X3 + XB X2 [3] X2 [4] The first b0 components transmitted over each bottleneck is obtained by the following equation. X φi (X)[1 : b0 ] = XBi + Xpj , for all Bi ∈ B. j:pj ∈Bi We show that this partial sum satisfies all the terminals in the set {tBi : Bi ∈ B} ∪ {tpj : pj ∈ / P 0 }. Terminals in {tBi : Bi ∈ B} can recover the sum as all messages not present in the partial sum are available to tBi through direct edges. For terminals in the set {tp : p ∈ / P 0 }, they carry out the following operation as a part of their decoding procedure.   X X X XBi + φi (X)[1 : b0 ] = Xpj  (21) i:Bi ∈hpi = X i:Bi ∈hpi i:Bi ∈hpi XBi + X ppTj Xpj j:pj ∈Bi + deg(p)Xp . (22) j:{p,pj }∈B For pj 6= p, we have that ppTj = 1 if {p, pj } ∈ B. Also by condition on the points that are not in P 0 , we have that deg(p) ≡ 1 (mod ch(F)), and hence all the coefficients in the above partial sum are 1. The messages in the set {XB : B ∈ / hpi} ∪ {Xpj : {pj , p} ∈ / B} are available to tp through direct edges and hence it can recover the sum. The remaining v 0 components available on the bottleneck edges {ei : Bi ∈ B0 } are used to transmit information that enable the terminals in the set {tp : p ∈ P 0 } to compute the sum. Specifically, we construct a flow on a bipartite graph whose one part corresponds to the points in P 0 and the other part corresponds to the blocks in B 0 , with incidence being determined by the submatrix ATI [B 0 , P 0 ]. Since there exists a matrix DI with specified row and column sums, we can use it to construct a flow on the bipartite graph such that the messages in the set {Xpi : pi ∈ P 0 } are transmitted in a piecewise fashion over the bottleneck edges {ej : Bj ∈ B 0 } in a manner similar to the proof of theorem 5. Arguing in the same way, one can show that the network code based on the flow solution allows each tp ∀p ∈ P 0 to obtain the value of Xp from the information transmitted over the bottleneck edges in the set {ei : Bi ∈ hpi}. Terminal tp computes the sum in eq. (21) as a part of its decoding procedure. Since deg(p) 6≡ 1 (mod ch(F)), every term in the RHS of eq. (22) except Xp has its coefficient as 1. But since tp knows the value of Xp it can subtract a multiple of it and recover the relevant partial sum. The messages not present in this partial sum are available to tp through direct edges and hence it can also compute the value of the sum. φC (X) X3 + X4 + XC X4 [1] X4 [2] φD (X) X1 + X4 + XD X4 [3] X4 [4] φE (X) X1 + X3 + XE – – Proposition 2 describes families of incidence structures for which the sum-networks constructed admit capacity-achieving linear network codes. The upper bound on the computation capacity of these sum-networks is obtained from Corollaries 1, 2 and 4. We now describe a rate-1 linear network code for the sum-networks when their corresponding incidence structures do not satisfy the qualifying conditions for the upper bounds. By theorem 1, the computation capacity of any sum-network obtained using the SUM-NET-CONS algorithm is at most 1. Proposition 4: For an incidence structure I = (P, B) and a finite field F, there exists a rate-1 linear network code that satisfies the following listed sum-networks. If • I is a 2-(v, k, 1) design: – the normal sum-network with ch(F) \| k − 1, – the transpose sum-network with ch(F) \| v−k k−1 , • I is a t-(v, t + 1, λ) design: – the normal sum-network obtained using the higher incidence matrix with ch(F) \| t, – the transpose sum-network obtained using the higher incidence matrix with ch(F) \| λ − 1. Proof: Suppose we construct a sum-network using the SUM-NET-CONS algorithm on a (0, 1)-matrix A of dimension r × c. If AT A = (AT A)# , the following rate-1 linear network code X φi (X) = Xpi + XBj , ∀ i ∈ [r], j:Bj ∈hpi i satisfies every terminal in the sum-network in the following manner. A terminal tpi , ∀i ∈ [r] receives all the messages not present in the partial sum transmitted along ei through direct edges, and hence it can compute the sum. A terminal tB , ∀B ∈ B can carry out the following operation. X X X X φi (X)= Xpi + XBj i:pi ∈Bj pi ∈B = X pi ∈B pi ∈B Bj ∈hpi i Xpi + X BlT Bj XBl . l:Bl ∈hBj i Since AT A P = (AT A)# , all the coefficients in the above sum are 1 and i:pi ∈Bj φi (X) is equal to the sum of all the messages in the set {Xpi : pi ∈ Bj } ∪ {XB : B ∈ hBj i}. All the messages that are not present in this set are available to tBj through direct edges. Such a rate-1 linear network code gives us our proposition in v−1 the following manner. Let AI be the v × k−1 incidence matrix 0 for a 2-(v, k, 1) design and let AI be the higher incidence IEEE TRANSACTIONS ON INFORMATION THEORY 17 matrix as defined in corollary 2 for a t-(v, t + 1, λ) design with λ 6= 1. Then, we have (from proofs of Corollaries 2, 4) S6 S14 S10 ATI AI − (ATI AI )# = (k − 1)I, v−k AI ATI − (AI ATI )# = I, k−1 0 0 AIT A0I − (AIT A0I )# = tI, a b c 0 0 A0I AIT − (A0I AIT )# = (λ − 1)I. Thus, whenever any of the above matrices is a zero matrix, we have a scalar linear network code that achieves the computation capacity of the associated sum-network. VII. D ISCUSSION AND COMPARISON WITH PRIOR WORK The discussion in Sections V and VI establishes the computation capacity for sum-networks derived from several classes of incidence structures. We now discuss the broader implications of these results by appealing to existence results for these incidence structures. BIBDs have been the subject of much investigation in the literature on combinatorial designs. In particular, the following two theorems are well-known. Theorem 6: [21, Theorem 6.17] There exists a (v, 3, 1)BIBD (also known as a Steiner triple system) if and only if v ≡ 1, 3 (mod 6); v ≥ 7. Theorem 7: [21, Theorem 7.31] There exists a (v, 4, 1)BIBD if and only if v ≡ 1, 4 (mod 12); v ≥ 13. In particular, these results show that there are an infinite family of Steiner triple systems and BIBDs with block size 4 and λ = 1. Since k = 3 for any Steiner triple system, we can demonstrate the existence of sum-networks whose computation capacity is greatly affected by the choice of the finite field F used for communication. Proposition 5: Consider the normal sum-network constructed using a 2-(v, 3, 1) design. If ch(F) = 2, then the computation capacity of the sum-network is 1. For odd ch(F), 6 the computation capacity is 5+v . For the normal sum-network constructed using a (v, 4, 1)-BIBD, the computation capacity 12 is 1 if ch(F) = 3 and 11+v otherwise. Proof: The number of blocks in a 2-(v, 3, 1) design is equal to v(v−1)/6. From corollary 2, if ch(F) is odd, then the computation capacity of the sum-network constructed using a 6 v Steiner triple system is at most v+v(v−1)/6 = 5+v . Moreover by proposition 2, we can construct a linear network code with rate equal to the upper bound. On the other hand, if ch(F) = 2, then the computation capacity of the same sum-network is 1 by proposition 4. The number of blocks in a 2-(v, 4, 1) design is v(v − 1)/12. We can recover the result for the computation capacity of a normal sum-network constructed using it in a manner similar to the previous case. Thus, this result shows that for the same network, computing the sum over even characteristic has capacity 1, while the capacity goes to zero as O(1/v) for odd characteristic. Moreover, this dichotomy is not unique to the prime number 2. Similar results hold for sum-networks derived from higher incidence structures (cf. corollary 4). Fig. 6. The schematic shown represents an undirected graph with three components: S6 , S14 and S10 . St denotes the star graph on t + 1 vertices, with only one vertex having degree t while the rest have degree 1. The vertices with the maximum degree in the three star graphs are a, b, c respectively. In addition, a is connected to b and b is connected to c, such that deg(a) = 7, deg(b) = 16, deg(c) = 11. Theorem 8: [34] For two integers t, v such that v ≥ t+1 > 0 and v ≡ t (mod (t + 1)!2t+1 ), a t-(v, t + 1, (t + 1)!2t+1 ) design with no repeated blocks exists. The number of blocks in a t-(v, t + 1, (t + 1)!2t+1 ) design 2t+1 . We consider the normal can be evaluated to be vt (t+1)! t+1 sum-network obtained using the higher incidence matrix of this t-design. If ch(F) - t, then by corollary 4 and proposition 2, we have that the computation capacity of this sum-network is v 1 t = . 2 v v (t+1)!2t+1 1 + t! (t + 1)!2t−1 + t t t+1 On the other hand, if ch(F) is a divisor of t, then by theorem 1 and proposition 4 we have that the computation capacity of the normal sum-network constructed using the higher incidence matrix is 1. Thus for the same sum-network, computing the sum over a field whose characteristic divides the parameter t can be done at rate = 1. However, if the field characteristic does not divide t, zero-error computation ofthe sumcan only −t2 be done at a rate which goes to zero as O et . Theorem 6 describes an infinite family of BIBDs with k = 3 and λ = 1. There are further existence results for BIBDs with λ = 1 and k 6= 3. In particular, for λ = 1, k ≤ 9 there exist BIBDs with value of v as given in Table 3.3 in [35, Section II.3.1]. As an example, if k = 5, then there exists a 2(v, 5, 1) design whenever v ≡ 1, 5 (mod 2)0. For any choice of a BIBD from this infinite family, we can construct a corresponding normal sum-network, whose computation capacity for a particular finite field can be found using corollary 2 and proposition 2. Even though theorem 8 states the existence of t-designs for v, t that satisfy the qualifying conditions, explicit constructions of such t-designs with t ≥ 6 are very rare. For a transposed sum-network obtained from an undirected graph that is not regular, the computation capacity can show a more involved dependence on the finite field alphabet as the following example demonstrates. Example 9: Consider the transposed sum-network obtained by applying the SUM-NET-CONS algorithm on the undirected graph I shown in Figure 6. Corollary 1 gives us an upper bound on the computation capacity of the transposed sumnetwork based on the finite field alphabet F. The upper bound for three different choices of F is as follows. IEEE TRANSACTIONS ON INFORMATION THEORY F = GF (2): Then P 0 = {b}, so the upper bound is 16/(16 + 1) = 16/17. 0 • F = GF (3): Then P = {c}, so the upper bound is 11/(11 + 1) = 11/12. 0 • F = GF (5): Then P = {a}, so the upper bound is 7/(7 + 1) = 7/8. We use proposition 3 to check if we can construct a linear network code in each case that has the same rate as the respective upper bound. To do that, we focus on the appropriate submatrix of AI for each case and see if it satisfies the required condition on row and column sums. The rows of AI corresponding to the vertices a, b, c (in order) are shown below.   16 1 0 ··· 0 06 1 114 1 010  , 0 ··· 0 1 110 • where 1, 0 indicate all-one and all-zero row vectors of size specified by their subscripts. Using this, one can verify that the appropriate submatrix for each of the three choices of F satisfies the conditions of proposition 3 and hence we can construct a capacity-achieving linear network code in each case. Thus, as the previous example demonstrates, the computation capacity of a particular sum-network need not take just one of two possible values, and can have a range of different values based on the finite field chosen. We can generalize the example to obtain sum-networks that have arbitrary different possible values for their computation capacity. Our constructed sum-networks have a unit maximum flow between any source and any terminal. We can modify our construction so that each edge in the network has a capacity of α > 1. Specifically, the following result can be shown. Proposition 6: Let N denote the sum-network obtained by applying the SUM-NET-CONS algorithm on a matrix A of dimension r × c. For an integer α > 1, let Nα denote the sum-network obtained by modifying the SUM-NET-CONS algorithm such that Nα has the same structure as N but each edge eα in Nα has cap(eα ) = α > 1. Then, if A satisfies the qualifying conditions in Theorems 2 and 5, the computation αr . capacity of Nα is r+c Proof: Since A satisfies the conditions in theorem 5, there exists a (m, n) vector linear network code with m = r, n = r + c. For every unit-capacity edge in N , we have α unitcapacity edges between the same tail and head in Nα . At the tail of every edge in Nα , we can apply the same network code except now we have α distinct edges on which we can transmit the encoded value. Thus we need transmit only r+c α symbols on each of those edges. If r+c α is not an integer, one can appropriately multiply both m, n with a constant. This αr . Since A also satisfies modified network code has rate = r+c the conditions in theorem 2, we have that an upper bound on the computation capacity of N is r/(r + c). Applying the same argument on Nα , we get that an upper bound on the αr computation capacity of Nα is r+c . This matches the rate of the modified vector linear network code described above. This result can be interpreted as follows. Consider the class of sum-networks where the maximum flow between any 18 source-terminal pair is at least α. Our results indicate, that for any α, we can always demonstrate the existence of a sumnetwork, where the computation capacity is strictly smaller than 1. Once again, this indicates the crucial role of the network topology in function computation. A. Comparison with prior work The work of Rai and Das [28] is closest in spirit to our work. In [28], the authors gave a construction procedure to obtain a sum-network with computation capacity equal to p/q, where p, q are any two co-prime natural numbers. The procedure involved first constructing a sum-network whose capacity was 1/q. Each edge in this sum-network had unitcapacity. By inflating the capacity of each edge in the sumnetwork to p > 1, the modified sum-network was shown to have computation capacity as p/q. Our work is a significant generalization of their work. In particular, their sum-network with capacity 1/q can be obtained by applying the SUM-NET-CONS algorithm to the incidence matrix of a complete graph on 2q − 1 vertices [29]. We provide a systematic procedure for constructing these sumnetworks for much larger classes of incidence structures. In [28], the authors also posed the question if smaller sumnetworks (with lesser sources and terminals) with capacity as p/q existed. Using the procedure described in this paper, we can answer that question in the affirmative. Example 10: The normal sum-network for the undirected graph in Figure 5(a) has computation capacity = 4/9 and has nine sources and terminals. To obtain a sum-network with the same computation capacity using the method described in [28] would involve constructing the normal sum-network for a complete graph on 17 vertices, and such a sum-network would have 153 source nodes and terminal nodes each. In [20], it was shown by a counter-example that for the class of sum-networks with \|S\| = \|T \| = 3, a maximum flow of 1 between each source-terminal pair was not enough to guarantee solvability (i.e., no network code of rate 1 exists for the counterexample). It can be observed that their counter-example is the sum-network shown in Figure 2(a). Our characterization of computation capacity for a family of sum-networks provides significantly more general impossibility results in a similar vein. In particular, note that for the α-capacity edge version of a sum-network, the maximum flow between any sourceterminal pair is at least α. Then suppose we consider the class of sum-networks with \|S\| = \|T \| = x = β(β + 1)/2 for some β ∈ N. Consider a complete graph Kβ = (V, E) on β vertices; then \|V \| + \|E\| = x. Consider the sum-network obtained by applying the procedure on Kβ , with each edge added having capacity as α. Then the computation capacity of this sumnetwork is αβ/x, which is less than 1 if α < (β + 1)/2. This implies that a max-flow of (β + 1)/2 between each sourceterminal pair is a necessary condition for ensuring all sumnetworks with \|S\| = \|T \| = x are solvable. When x cannot be written as β(β + 1)/2 for some β, a similar argument can be made by finding an undirected graph G = (V, E) (whose incidence matrix AG satisfies the condition in theorem 5) such that \|V \| is minimal and \|V \| + \|E\| = x. IEEE TRANSACTIONS ON INFORMATION THEORY VIII. C ONCLUSIONS AND F UTURE W ORK Sum-networks are a large class of function computation problems over directed acyclic networks. The notion of computation capacity is central in function computation problems, and various counterexamples and problem instances have been used by different authors to obtain a better understanding of solvability and computation capacity of general networks. We provide an algorithm to systematically construct sumnetwork instances using combinatorial objects called incidence structures. We propose novel upper bounds on their computation capacity, and in most cases, give matching achievable schemes that leverage results on the existence of non-negative integer matrices with prescribed row and column sums. We demonstrate that the dependence of computation capacity on the underlying field characteristic can be rather strong. There are several opportunities for future work. Our proposed linear network codes for the constructed sum-networks require the corresponding incidence structures to have a specific property. In particular, our techniques only work in the case when AT A − (AT A)# is a diagonal matrix. It would be interesting to find capacity achieving network codes in cases when AT A − (AT A)# is not diagonal. More generally, it would be interesting to obtain achievability schemes and upper bounds for sum-networks with more general topologies. R EFERENCES [1] J. Korner and K. Marton, “How to encode the modulo-2 sum of binary sources,” IEEE Trans. on Info. Th., vol. 25, no. 2, pp. 219–221, 1979. [2] A. Orlitsky and J. Roche, “Coding for computing,” IEEE Trans. on Info. Th., vol. 47, no. 3, pp. 903–917, Mar 2001. [3] V. Doshi, D. Shah, M. Medard, and M. Effros, “Functional compression through graph coloring,” IEEE Trans. on Info. Th., vol. 56, no. 8, pp. 3901–3917, Aug 2010. [4] R. Ahlswede, N. Cai, S.-Y. Li, and R. W. Yeung, “Network Information Flow,” IEEE Trans. on Info. Th., vol. 46, no. 4, pp. 1204–1216, 2000. [5] R. Koetter and M. Medard, “An algebraic approach to network coding,” IEEE/ACM Transactions on Networking, vol. 11, no. 5, pp. 782–795, Oct 2003. [6] S. Y. R. Li, R. W. Yeung, and N. Cai, “Linear network coding,” IEEE Trans. on Info. Th., vol. 49, no. 2, pp. 371–381, Feb 2003. [7] R. Dougherty, C. Freiling, and K. Zeger, “Insufficiency of linear coding in network information flow,” IEEE Trans. on Info. Th., vol. 51, no. 8, pp. 2745–2759, Aug. 2005. [8] J. Cannons, R. Dougherty, C. Freiling, and K. Zeger, “Network routing capacity,” IEEE Trans. on Info. Th., vol. 52, no. 3, pp. 777–788, Mar. 2006. [9] R. Dougherty, C. Freiling, and K. Zeger, “Unachievability of network coding capacity,” IEEE Trans. on Info. Th., vol. 52, no. 6, pp. 2365– 2372, June 2006. [10] S. Jaggi, P. Sanders, P. A. Chou, M. Effros, S. Egner, K. Jain, and L. M. G. M. Tolhuizen, “Polynomial time algorithms for multicast network code construction,” IEEE Trans. on Info. Th., vol. 51, no. 6, pp. 1973– 1982, June 2005. [11] S. Huang and A. Ramamoorthy, “An achievable region for the double unicast problem based on a minimum cut analysis,” IEEE Trans. on Comm., vol. 61, no. 7, pp. 2890–2899, 2013. [12] ——, “On the multiple unicast capacity of 3-source, 3-terminal directed acyclic networks,” IEEE/ACM Trans. on Networking, vol. 22, no.1, pp. 285–299, 2014. [13] A. R. Lehman and E. Lehman, “Complexity classification of network information flow problems,” in Proceedings of the Fifteenth Annual ACM-SIAM Symposium on Discrete Algorithms, Jan 2004, pp. 142–150. [14] S. Kamath, D. N. C. Tse, and C. C. Wang, “Two-unicast is hard,” in IEEE Intl. Symposium on Info. Th., June 2014, pp. 2147–2151. [15] R. Appuswamy, M. Franceschetti, N. Karamchandani, and K. Zeger, “Network coding for computing: Cut-set bounds,” IEEE Trans. on Info. Th., vol. 57, no. 2, pp. 1015–1030, Feb 2011. 19 [16] C. Huang, Z. Tan, and S. Yang, “Upper bound on function computation in directed acyclic networks,” in IEEE Info. Th. Workshop, April 2015, pp. 1–5. [17] R. Appuswamy, M. Franceschetti, N. Karamchandani, and K. Zeger, “Linear codes, target function classes, and network computing capacity,” IEEE Trans. on Info. Th., vol. 59, no. 9, pp. 5741–5753, Sept 2013. [18] A. Ramamoorthy, “Communicating the sum of sources over a network,” in IEEE Intl. Symposium on Info. Th., July 2008, pp. 1646–1650. [19] B. K. Rai and B. K. Dey, “On network coding for sum-networks,” IEEE Trans. on Info. Th., vol. 58, no. 1, pp. 50 –63, 2012. [20] A. Ramamoorthy and M. Langberg, “Communicating the sum of sources over a network,” IEEE J. Select. Areas Comm., vol. 31, no. 4, pp. 655– 665, 2013. [21] D. R. Stinson, Combinatorial Designs: Construction and Analysis. Springer, 2003. [22] O. Olmez and A. Ramamoorthy, “Fractional repetition codes with flexible repair from combinatorial designs,” IEEE Trans. on Info. Th., vol. 62, no. 4, pp. 1565 –1591, 2016. [23] S. E. Rouayheb and K. Ramchandran, “Fractional repetition codes for repair in distributed storage systems,” in 48th Annual Allerton Conference on Comm., Control & Computing, Sept 2010, pp. 1510 – 1517. [24] L. Tang and A. Ramamoorthy, “Coded caching with low subpacketization levels,” in Intl. Symp. on Network Coding (NetCod), Dec 2016, pp. 1–6. [25] ——, “Coded caching for networks with the resolvability property,” in IEEE Intl. Symposium on Info. Th., 2016. [26] ——, “Low Subpacketization Schemes for Coded Caching,” 2017 [Online] Available: https://arxiv.org/abs/1706.00101. [27] B. K. Rai and N. Das, “On the capacity of ms/3t and 3s/nt sumnetworks,” in IEEE Info. Th. Workshop, Sep. 2013, pp. 1–5. [28] B. Rai and N. Das, “On the capacity of sum-networks,” in 51st Annual Allerton Conference on Comm., Control & Computing, Oct 2013, pp. 1545–1552. [29] A. Tripathy and A. Ramamoorthy, “Sum-networks from undirected graphs: Construction and capacity analysis,” in 52nd Annual Allerton Conference on Comm., Control & Computing, Sept 2014, pp. 651–658. [30] N. Das and B. K. Rai, “On the number of sources and terminals of sum-networks with capacity p/q,” in 21st National Conference on Communications (NCC), Feb 2015, pp. 1–6. [31] A. Tripathy and A. Ramamoorthy, “Capacity of sum-networks for different message alphabets,” in IEEE Intl. Symposium on Info. Th., June 2015, pp. 606–610. [32] R. A. Brualdi, Combinatorial matrix classes. Cambridge University Press, 2006, vol. 13. [33] L. Mirsky, “Combinatorial theorems and integral matrices,” Journal of Combinatorial Theory, vol. 5, no. 1, pp. 30–44, 1968. [34] L. Teirlinck, “Non-trivial t-designs without repeated blocks exist for all t,” Discrete Mathematics, vol. 65, no. 3, pp. 301–311, 1987. [35] C. J. Colbourn and J. H. Dinitz, Handbook of combinatorial designs. CRC press, 2006. Ardhendu Tripathy (S’15) is a Ph.D. student in the Department of Electrical and Computer Engineering at Iowa State University. He obtained his B.Tech. degree in Electrical Engineering from the Indian Institute of Technology, Kanpur in 2012. His research interests are in the areas of information theory, machine learning and signal processing. Aditya Ramamoorthy (M’05) received the B.Tech. degree in electrical engineering from the Indian Institute of Technology, Delhi, in 1999, and the M.S. and Ph.D. degrees from the University of California, Los Angeles (UCLA), in 2002 and 2005, respectively. He was a systems engineer with Biomorphic VLSI Inc. until 2001. From 2005 to 2006, he was with the Data Storage Signal Processing Group of Marvell Semiconductor Inc. Since fall 2006, he has been with the Electrical and Computer Engineering Department at Iowa State University, Ames, IA 50011, USA. His research interests are in the areas of network information theory, channel coding and signal processing for bioinformatics and nanotechnology. Dr. Ramamoorthy served as an editor for the IEEE Transactions on Communications from 2011 – 2015. He is currently serving as an associate editor for the IEEE Transactions on Information Theory. He is the recipient of the 2012 Early Career Engineering Faculty Research Award from Iowa State University, the 2012 NSF CAREER award, and the Harpole-Pentair professorship in 2009 and 2010.	7cs.IT
FROM ISOLATED SUBGROUPS TO GENERIC PERMUTATION REPRESENTATIONS arXiv:1601.07538v1 [math.GR] 27 Jan 2016 YAIR GLASNER, DANIEL KITROSER, AND JULIEN MELLERAY Abstract. Let G be a countable group, Sub(G) the (compact, metric) space of all subgroups of G with the Chabauty topology and Is(G) ⊆ Sub(G) the collection of isolated points. We denote by X! the (Polish) group of all permutations of a countable set X. Then the following properties are equivalent: (i) Is(G) is dense in Sub(G), (ii) G admits a “generic permutation representation”. Namely there exists some τ ∗ ∈ Hom(G, X!) such that the collection of permutation representations {ϕ ∈ Hom(G, X!) \| ϕ is permutation isomorphic to τ ∗ } is co-meager in Hom(G, X!). We call groups satisfying these properties solitary. Examples of solitary groups include finitely generated LERF groups and groups with countably many subgroups. 1. Introduction Let G be a countable group and Sub(G), the space of all subgroups of G endowed with the Chabauty topology, which makes it into a compact metrizable totally disconnected space. The easiest way to define this topology is to embed Sub(G) ⊆ {0, 1}G as a closed subset and induce the Tychonoff topology on {0, 1}G . The group G acts on Sub(G) continuously by conjugation g · ∆ = g∆g −1 . One is naturally led to the question of how the structure of the topological space Sub(G), or more generally the topological dynamical system (G, Sub(G)), is reflected in the algebraic structure of G. The Cantor-Bendixon structure theory of compact spaces leads us to consider the decomposition Sub(G) = Is(G) ⊔ Sub(G)′ . Here Is(G) is the collection of isolated points, or isolated subgroups as we shall refer to them, and Sub(G)′ is its complement. Isolated subgroups are special from the algebraic point of view! Clearly there are only countably many of them. One can think of isolated subgroups in algorithmic terms as subgroups that are detectable or recognizable via a finite algorithmic procedure. A subgroup ∆ ∈ Sub(G) is isolated if and only if it can be identified by making a finite number of membership and non-membership tests for specific elements. It is sometimes convenient to think of this in terms of Schreier graphs. Let S be a symmetric generating set for G. A subgroup ∆ ∈ Sub(G) is isolated if one can find a finite algorithm that would recognize the Schreier graph Sch(G, ∆, S) out of all Schreier graphs of the group G. Note that S might very well be infinite, and consequently the Schreier graphs in question may fail to be locally finite. Still the algorithm is allowed to look only at finitely many edges. From these characterizations it is easy to see that isolated groups are always finitely generated. In the special case where G itself is finitely generated every finite index subgroup is isolated and we obtain inclusions Subfi (G) ⊆ Is(G) ⊆ Subfg (G) where Subfi (G), Subfg (G) stand for finite index and finitely generated subgroups respectively. 2010 Mathematics Subject Classification. Primary 20E26; Secondary 0B07, 03E15. Key words and phrases. Isolated subgroups, Solitary groups, LERF groups, amenable groups, ample generics, amenable actions. 1 It is clear that Is(G) is always a discrete countable open subset of Sub(G). Our main new definition is the following: Definition 1.1. A group G is called solitary if the isolated points Is(G) are dense in Sub(G). Let X be a countable set and X! the group of all permutations of X. The topology of pointwise convergence makes X! into a Polish group: separable, metrizable and complete. The space Hom(G, X!) is the space of all permutation representations of G and is clearly also a Polish space. There is a natural action X! × Hom(G, X!) → Hom(G, X!) (1) (α, ϕ) 7→ α · ϕ : g 7→ αϕ(g)α−1 Two permutation representations that are in the same orbit are said to be isomorphic as permutation representations. We will be interested in Baire generic properties of permutation representations and in particular in the existence of a generic permutation representation in the sense of the following: Definition 1.2. The group G is said to have a generic permutation representation if there is a permutation representation τ ∗ ∈ Hom(G, X!) whose orbit X!(τ ∗ ) = {α · τ ∗ \| α ∈ X!} = {ϕ ∈ Hom(G, X!) \| ϕ is a permutation isomorphic to τ ∗ } is co-meager in Hom(G, X!). It turns out that the existence of a generic permutation representation is captured by the structure of the topological space Sub(G). Theorem 1.3. (Main theorem) A countable group G admits a generic permutation representation if and only if it is solitary. Definition 1.4. A group G is called subgroup separable or locally extended residually finite (LERF for short), if every finitely generated subgroup of G is the intersection of finite index subgroups. Or equivalently if any finitely generated subgroup is closed in the profinite topology on G. Examples of LERF groups include finitely generated abelian groups, free groups [Hal49], surface groups [Sco85] and more generally, limit groups [Wil08], the Grigorchuk group [GW03], many lamplighter groups [GK14]. Recently the LERF property attracted a lot of attention as Agol’s proof of the LERF property for the fundamental group of a closed hyperbolic 3-manifold [Ago13] was a central ingredient in his solution to Thruston’s virtual Haken conjecture. The following theorem is analogous to our main theorem above. It shows in particular that finitely generated LERF groups are solitary. Theorem 1.5. Let G be a finitely generated group, then the following conditions are equivalent: (1) G is LERF. (2) The collection Subfi (G) of finite index subgroups is dense in Sub(G). (3) G has a generic permutation representation, all of whose orbits are finite. 2 Remark 1.6. The fact that the first and third condition above are equivalent also follows from earlier work of Rosendal: Proposition 8(B) in [Ros11a] proves that G is LERF if and only if a generic permutation representation has only finite orbits, while Theorem 11 of [Ros11b] shows that a finitely generated LERF group admits a generic permutation representation. Note that Theorem 11 of [Ros11b] is only formulated for groups acting by isometries on the rational Urysohn space, but see the remark in the last paragraph of [Ros11b], where it is pointed out that the proofs adapt to other metric spaces, notably the Urysohn space with distances {0, 1}, that is, an infinite countable set. In the paper we state and prove a more detailed version of this theorem that holds also for countable groups. Once we leave the realm of finitely generated groups isolated subgroups and solitary groups no longer generalize finite index subgroups and LERF groups respectively. Each of these theories goes in its own way. Our impression is that in some settings the choice of isolated subgroups and solitary groups is the more natural one. The following theorem summarizes some examples and structural results we have about solitary groups. Theorem 1.7. Some properties of solitary groups. (1) Finitely generated LERF groups are solitary. (2) If Sub(G) is countable then G is solitary. (3) Let 1 → N → H → G → 1 be a short exact sequence of countable groups such that N is finitely generated, as an abstract group. If H is solitary then so is G. (4) The free product of two countable groups G ∗ H is solitary if and only if one of the following two options hold: • both G and H are LERF and finitely generated. • G is solitary and H is trivial, or vice-versa. The situation considered in condition (3) above is identical to the one appearing in the famous Rips construction. This naturally leads to the following: Question 1.8. Is it true that every finitely generated solitary group G can be placed in a short exact sequence 1 → N → H → G → 1 where H is solitary and hyperbolic and N is finitely generated as an abstract group. Theorem 1.3 is tightly connected to the notion of ample generics in Polish groups. We adopt the notation of [MT13] ; [Ros11b] was the first paper to express this property in term of generic orbits on “presentation varieties”. Definition 1.9. We say that a Polish group P has ample generics, if Hom(Fn , P ) admits a generic P orbit for every n ∈ N. The notion of ample generics was first introduced in [HHLS93] in order to study the small index property in Polish groups. Namely that every subgroup of index < 2ℵ0 is open. Indeed this and additional consequences such as automatic continuity of abstract homomorphisms into any separable group were subsequently established for all groups with ample generics in [KR07]. We refer the readers to all of the above mentioned papers (see also the survey [GW08]) and the references therein. With this terminology in place Theorem 1.5 shows that the following two well known facts: • X! has ample generics, • finitely generated free groups Fn are LERF, 3 are in fact two different realizations of the same phenomenon. In both cases one seeks a generic P -orbit in Hom(G, P ). But in the study of groups with ample generics, one fixes G (or more precisely lets G range over all finitely generated free groups) and lets the Polish group P vary; whereas in the study of solitary groups we fix the Polish group P = X! and consider the class of all the countable groups G that give rise to a generic orbit. In view of the very natural characterization that arises from Theorem 1.5 the following question seems natural: Question 1.10. Given a Polish group P , describe the class of all finitely generated groups G for which Hom(G, P ) has a generic P -orbit. In particular, the answer should contain all finitely generated free groups whenever P has ample generics. The group X! is probably the simplest example of a Polish group with ample generics, but there are many others. We now turn to generalizing Theorem 1.5, and the notion of LERF groups, in a different direction. A subgroup H ≤ G is called co-amenable if there is a G-invariant mean on G/H (see also Definitions 6.1, 6.2). Co-amenable subgroups generalize finite index subgroups in much the same way that amenable groups generalize finite groups. In view of that and of Theorems 1.5 and 1.3 we can generalize the notion of LERF groups as follows: Definition 1.11. A group G is amenably separable, or A-separable for short, if the set of co-amenable subgroups of G is dense in Sub(G). In view of the Theorem 1.5, every LERF group is A-separable. Another obvious source for examples is the class of all amenable groups. These are A-separable since all of their subgroups are co-amenable. In chapter 6 of this work, we initiate the study of A-separable groups. Our hope is that the notion of A-separability will prove to be a useful generalization of the, a-priori very different, properties of LERF and amenability. This situation is perhaps reminiscent of the way in which sofic groups simultaneously generalize the notions of residual finiteness and amenability. In these terms the analogue of Theorem 1.3 is the following Theorem 1.12. A countable group G is A-separable if and only if for a generic action of G on a countable set, the action on every orbit is amenable. Here are some properties of A-separable groups. Theorem 1.13. The following properties hold for the class of A-separable groups: • LERF groups and amenable groups are A-separable. • The class of A-separable groups is closed under free products. • There exist A-separable groups which are neither LERF nor amenable. • A group with property (T) is A-separable if and only if it is LERF. • Higher rank lattices in non-compact simple Lie groups which satisfy the congruence subgroup property are never A-separable. The paper is arranged as follows. Section 2 is dedicated to a systematic investigation of the topological spaces Sub(G), Hom(G, X!) and the standard stabilizer map Hom(G, X!) → Sub(G) between them. In Section 3 we prove Theorem 1.5. Chapter 4 is dedicated to solitary groups and there we prove Theorems 1.3 and 1.7. Finally, in chapter 6 we prove Theorems 1.12 and 1.13 and give examples of non amenable, non LERF groups that are A-separable. The results in this work also to appear as part of the Ph.D. dissertation of the second author [Kit15]. 4 2. Dense and generic properties of actions and subgroups 2.1. The space of permutation representations Hom(G, X!). Let X be a countable set and X! the full symmetric group of all bijections of X onto itself. We endow X! with the topology of pointwise convergence which makes it into a Polish topological group. In other words a topological group that is separable, and admits a complete metric. The latter fact is important for us because it shows that X! is a Baire space; though we will never consider any specific metric. An explicit basis for the topology can be given by the sets U(α, A) := β ∈ X! \| β A = α A (α ∈ X!, A ⊆ X finite). On X!n we will always put the product topology, which is still Polish for every n ∈ N ∪ {∞}. Let G be a countable group with a given presentation G = hS \| Ri where S = {s1 , s2 , . . . }. Then, we can identify Hom(G, X!) with a closed subset of X!S via the following embedding: Hom(G, X!) → {α ∈ X!S \| ∀w ∈ R : w(α) = 1X } ⊆ X!S ρ 7→ {ρ(si )}si∈S Thus Hom(G, X!) is a closed subspace of X!S and the induced topology makes it into a Polish space (note that this topology does not depend on the choice of presentation). A basis for the topology on Hom(G, X!) is given by: O(ρ, T, A) := {σ ∈ Hom(G, X!) \| ∀t ∈ T : σ(t) A = ρ(t) A } (ρ ∈ Hom(G, X!), T ⊆ G and A ⊆ X both finite). If S itself happens to be finite, then the sets O(ρ, S, A) form a basis for Hom(G, X!). As mentioned in the introduction (see Equation 1) the group X! acts, from the left, on Hom(G, X!) and the orbits of this action are exactly the standard isomorphism classes of permutation representations. It is well known that two permutation representations are isomorphic if and only if they contain the same transitive components, appearing with the same multiplicity. The transitive components, in turn, are isomorphic to quasiregular actions of the form G y G/H for some H ∈ Sub(G). If {Hi } ⊆ Sub(G) is a countable or finite collection of subgroups and if di ∈ N ∪ {∞} we will denote the (isomorphism class of) the permutation representation that has exactly di transitive components isomorphic to ηi G y G/Hi by G G di · (G/Hi ) = di · ηi . i i Some care is due with this notation. It is not always possible to identify such an action with an element of Hom(G, X!) because if the sum is finite and Hi are all of finite index then the underlying set is finite. When this is not the case we can identify such an action with an element of Hom(G, X!) via an arbitrary bijection between ⊔i G/Hi ∼ = X. Different choices of this bijection will yield all the different points in the corresponding X! orbit. We will make frequent use of the following: Definition 2.1. Let G y X, x ∈ X, g1 , . . . , gn ∈ G and let w = wk wk−1 · · · w1 be a word over {g1±1, . . . , gn±1}. The trace of x under w is the set: tracew (x) = {x, w1 x, . . . , wk−1 · · · w1 x, wx}. 5 2.2. The space of subgroups. Let G be a countable group and consider the space {0, 1}G of subsets of G, equipped with the product topology. This is a compact, metrizable space. Let Sub(G) denote the set of all subgroups of G. It is easy to verify that Sub(G) is closed in {0, 1}G and so it is a compact, metrizable space. The induced topology on Sub(G) is called the Chabauty topology and a basis for this topology can be given by the sets W(H, Ω) = {K ∈ Sub(G) \| K ∩ Ω = H ∩ Ω} (H ∈ Sub(G), Ω ⊆ G finite). If H ∈ Sub(G) we denote by Env(H) = {K ∈ Sub(G) \| K ≥ H} the envelope of H. Both subsets Sub(H) and Env(H) are closed in Sub(G). If H is finitely generated then Env(H) is also open. Denoting by Subfg (G) ⊆ Sub(G) the collection of finitely generated subgroups, it is easy to check that the collection [ {Env(H) \| H ∈ Subfg (G)} {Sub(G) r Env(H) \| H ∈ Subfg (G)} forms another basis for the topology of Sub(G). 2.3. Isolated subgroups. Let Is(G) and Occ(G) denote the isolated points of Sub(G) and the subgroups of G with open conjugacy classes respectively. Note that a subgroup H is in Occ(G) if and only if there is an open neighborhood of H, consisting of only conjugates of H. Both subsets are open and conjugation invariant. Proposition 2.2. Here are some basic properties of these subgroups: (1) Is(G) = Occ(G). (2) Every H ∈ Is(G) is finitely generated. (3) if H ∈ Subfg (G) then Env(H) is an open neighborhood of H. (4) When G itself is finitely generated then every finite index subgroup is isolated. (5) If H ∈ Subfg (G) and \|Env(H)\| < ∞ then H is isolated. In particular every finitely generated maximal subgroup is isolated. Proof. It is clear that Is(G) ⊆ Occ(G) and that both sets are open. The opposite inclusion follows from Baire’s theorem: Let K ∈ Occ(G) and let [K] = {gKg −1 \| g ∈ G} be its conjugacy class, which is open by definition. Since G is countable, [K] is a countable (or finite) union of closed points so by Baire’s theorem one of them has to be open. Since G is transitive on [K] all of these points are open and in particular K ∈ Is(G). This proves (1). If H is not finitely generated then we can find a sequence of finitely generated subgroups H1 < H2 < H3 < . . . with ∪i Hi = H. Clearly Hi → H in the topology of Sub(G) but none of these subgroups is equal to H because H is not finitely generated. (3) is clear and (5) follows directly from (3). Finally if G itself is finitely generated then so is every finite index subgroup and (4) follows directly from (5). Thus, for a finitely generated group G isolated subgroups form a class of subgroups that sits between the finitely generated subgroups and the subgroups of finite index, namely: Subfg (G) ⊆ Is(G) ⊆ Subfg (G). We find it very useful to think of isolated subgroups as generalizations of finite index subgroups. 6 2.4. Generic properties. A subset A ⊆ Y in a Polish space is called generic (or alternatively residual or co-meager) if it contains a countable intersection of dense open sets. By Baire’s category theorem generic sets are always dense. We say that the property (P) is generic in Y or that a generic element of Y has the property (P) if the set {y ∈ Y \| y has the property (P)} is generic in Y . In this paper we will be interested in generic properties of permutation representations Hom(G, X!). The simplest example is Hom(Z, X!) ∼ = X!. The following, well known proposition, summarizes the generic properties of this space. Its proof is an exercise in Baire’s category theorem which we leave to the readers. We chose to mention it here because our main theorem, and its proof, are basically far reaching generalizations of this fact. Proposition 2.3. X! has a residual conjugacy class. This conjugacy class can be described explicitly: G τ∗ = ∞ · (Z/nZ) n∈N In terms of the definition below, the above proposition just says that G = Z admits a generic permutation representation. Definition 2.4. We say that G admits a generic permutation representation if there exists a permutation representation τ ∗ ∈ Hom(G, X!) whose orbit under the action X! y Hom(G, X!) is residual in Hom(G, X!). 2.5. Properties of the stabilizer map. Given a permutation representation σ ∈ Hom(G, X!) and a point x ∈ X, we denote by Gx (σ) = {g ∈ G \| σ(g)x = x} the stabilizer of this point. Fixing x this gives rise to a stabilizer map Gx : Hom(G, X!) → Sub(G) σ 7→ Gx (σ) (2) Lemma 2.5. (Main lemma) For every x ∈ X, the stabilizer map Gx : Hom(G, X!) → Sub(G) is continuous, surjective and open. Proof. It is clear that this map is surjective. Let ϕ ∈ Hom(G, X!) and let Ω ⊆ G be finite. If σ ∈ O(ϕ, Ω, {x}) then ∀g ∈ Ω : g ∈ Gx (σ) ⇔ x = σ(g)x = ϕ(g)x ⇔ g ∈ Gx (ϕ) i.e Gx (σ) ∈ W(Gx (ϕ), Ω). This proves that the map Gx is continuous. To prove that this map is open let τ ∈ O = O(ϕ, S, A) ⊆ Hom(G, X!) be a basic open neighborhood in Hom(G, X!) and a point therein. By extending S we may assume that this set contains the identity and is symmetric, i.e. that S = S −1 . We have to exhibit an open neighborhood W ⊆ Sub(G) such that Gx (τ ) ∈ W ⊆ Gx (O(ϕ, S, A)). Let Y = τ (G)x ⊆ X be the orbit of x under τ and L = Gx (τ ) be the stabilizer. We can identify Y = G/L under the orbit map gL 7→ τ (g)x. Set AY := A ∩ Y . Let Ω ⊆ G be a finite symmetric set such that [ τ (s)AY ⊆ τ (Ω)x = {τ (ω)x \| ω ∈ Ω}. s∈S 7 We claim that the basic open set W = W(L, ΩSΩ) satisfies all our requirements. Where ΩSΩ = {ω1 sω2 \| ωi ∈ Ω, s ∈ S} is the set product. Fixing a group K ∈ W, we will complete our proof by finding η ∈ O such that K = Gx (η). Consider the finite subset ΩK = {ωK ∈ G/K \| ω ∈ Ω} and define a partial map: f : ΩK → Y ωK 7→ τ (ω)x, ∀ω ∈ Ω. This map is well-defined and injective on its domain, since by the definition of the open set W we have ω1 K = ω2 K ⇐⇒ ω1 L1 = ω2 L1 ⇐⇒ ω1 L = ω2 L ⇐⇒ τ (ω1 )x = τ (ω2 )x for every ω1 , ω2 ∈ Ω. For a similar reason this map partially respects the G action on both sides in the sense that sf (ωK) = f (sωK) ∀ω ∈ Ω, s ∈ S. Define a bijection f˜ : G/K ⊔ X → X satisfying the following conditions: • f˜ extends f , namely f˜(ωK) = f (ωK), ∀ω ∈ Ω. • f˜ is the identity on ∪s∈S τ (s)(A) \ Y . Now let us define an action σ1 of G on G/K ⊔ X where G acts on G/K by the quasi-regular action on on X by τ and let σ =∈ Hom(G, X!) be defined by σ(g)x = f˜σ1 (g)f˜−1 (x). It is easy to verify that K = Gx (σ). Also for every s ∈ S and y ∈ A we have σ(s)y = τ (s)y. Indeed if y ∈ AY then by our choice of Ω, we have y = τ ωx for some ω ∈ Ω and hence: ( feσ1 (s)fe−1 y = f sf −1 τ (ω)x = f sωK = τ (sω)x = τ (s)y, y ∈ AY σ(s)y = feσ1 (s)fe−1 y = feτ (s)fe−1 y = τ (s)y y 6∈ AY This concludes the proof that Gx is open. Corollary 2.6. For a subset D ⊆ Sub(G) denote e := {ϕ ∈ Hom(G, X!) \| Gx (ϕ) ∈ D ∀x} = D \ (Gx )−1 (D). x∈X Then e is dense in Hom(G, X!). (1) If D is conjugation-invariant and dense in Sub(G) then D e is Gδ in Hom(G, X!). (2) If D is Gδ in Sub(G) then D e is generic whenever D is. In particular, D Proof. Suppose D is conjugation-invariant and dense in Sub(G), let ρ ∈ Hom(G, X!) and let S ⊆ G, A ⊆ X be finite. Let x ∈ A. By Lemma 2.5, the set G−1 x (D) is dense in Hom(G, X!) and so there exists ϕ1 ∈ Hom(G, X!) such that ϕ1 ∈ O(ρ, S, A) and Gx (ϕ1 ) ∈ D. Denote Y1 = ϕ1 (G)x and note that since D is conjugation-invariant we have Gy (ϕ1 ) ∈ D for all y ∈ Y1 . Now, if z ∈ A r Y1 , we can apply the same argument and get a permutation representation ϕ2 of G on X r Y1 such that ϕ2 (s) agrees with ϕ1 (s) on A ∩ (X r Y1 ) for all s ∈ S and such that all the stabilizers of points belonging to Y2 = ϕ2 (G)z are in D. We get that the action ϕ ∈ Hom(G, X!) defined by ( ϕ1 (g)x, x ∈ Y1 ∀g ∈ G, x ∈ X : ϕ(g)x = ϕ2 (g)x, x ∈ X r Y1 8 belongs to O(ρ, S, A) and every stabilizer of a point belonging to the ϕ-invariant set Y1 ∪ Y2 is in D. By repeating the process described above, we get after finitely many steps an action ψ ∈ O(ρ, S, A) and a ψ-invariant set Y ⊆ X such that A ⊆ Y and such that Gy (ψ) ∈ D for ψ every y ∈ Y . Finally, extend the action G y Y to an action ψe ∈ Hom(G, X!) in such a way e and since A ⊆ Y we have that all stabilizers of points in X r Y belong to D. Thus, ψe ∈ D e ψ(s) = ψ(s) A = ρ(s) A for all s ∈ S, i.e. ψe ∈ T O(ρ, S, A). This proves part 1. A Now, assume D is Gδ so we can write D =T n Dn , where Dn is open for every n ∈ N. −1 Then, for every x ∈ X we have G−1 x (D) = n Gx (Dn ). Since Dn is open we get from −1 Lemma 2.5 that Gx (Dn ) is open for every x ∈ X and n ∈ N and so G−1 x (D) is Gδ . Since X T −1 e is countable this means that D = x∈X (Gx ) (D) is also Gδ and part 2 is proven. Lemma 2.7. Let U ⊆ Sub(G) be open and non-empty. Then for a generic permutation representation ϕ ∈ Hom(G, X!) we have \|{x ∈ X \| Gx (ϕ) ∈ U}\| = ∞. Proof. Let Λm = {ϕ ∈ Hom(G, X!) \| \|{x ∈ X \| Gx (ϕ) ∈ U}\| > m} . \ Λ= Λm . m∈N We claim that Λ is generic, and by Baire’s theorem it would be enough to show that Λm is open and dense for each m ∈ N. The fact that Λm is open follows directly from the continuity of the map Gx . To prove density we just add m new orbits with a stabilizer from U far away. This is made explicit in the following way. Given a basic open set O(ϕ, S, A) ⊆ Hom(G, X!) we want to find an element η ∈ O(ϕ, S, A) ∩ Λm . Fix any subgroup H ∈ U and consider the set Y = X ⊔ (G/H)m endowed with the diagonal G action η ′ ∈ Hom(G, Y !) given by η ′ (g)(x, g1H, . . . , gm H) = (ϕ(g)x, gg1H, . . . , ggm H). Now, let ι : X → Y be the identity map from X toSthe copy of X contained in Y , ι\|A : A → Y be its restriction to the finite set A=A∪ s∈S∪S −1 ϕ(s)A and let I : X → Y be an extension of ι\|A to a bijection between X and Y . One easily checks that η = I −1 ◦ η ′ ◦ I ∈ O(ϕ, S, A) ∩ Λm is as required. Lemma 2.8. To any given ρ ∈ Hom(G, X!) there is an arbitrarily close action, with infinitely many fixed points. Proof. Given finite sets S ⊆ G, A ⊆ X we seek an action ρ′ ∈ O(ρ, S, A) with infinitely many (ρ,1) fixed points. Consider the action G y X ⊔ N, obtained from ρ by adding countably many fixed points. The desired action ρ′ = ϕ−1 (ρ, 1)ϕ is obtained by intertwining this action via any bijection ϕ :X → X ⊔ N with the property that ϕ is the identity when restricted to S A∪ ρ(s)A . s∈S 3. The LERF property In this section we prove Theorem 1.5. In fact, as promised in the introduction, we will prove the following slightly more general version of the theorem that for arbitrary countable groups (not necessarily finitely generated). 9 Theorem 3.1. Let G be a finitely generated group, then the following conditions are equivalent: (1) G is LERF. (2) The collection of finite index subgroups of G is dense in Sub(G). (3) The collection of permutation representations all of whose orbits are finite is dense in Hom(G, X!). (4) The collection of permutation representations all of whose orbits are finite is generic in Hom(G, X!). (5) G has a generic permutation representation, all of whose orbits are finite. When G is countable, but not necessarily finitely generated, then only the first three conditions are equivalent. It is easy to verify that this theorem implies Theorem 1.5 when G is finitely generated. The apparent complications when G fails to be finitely generated, and lack thereof in the proof of Theorem 1.3, emphasize one of our main points: Isolated subgroups are more natural than finite index subgroups in this setting. We mention again that the equivalence of the first, third and fourth conditions is already present in Rosendal’s works [Ros11a] and [Ros11b] (see the discussion in the introduction). Proof of Theorem 3.1. In a countable group G every subgroup is an ascending union of finitely generated subgroups. Hence Subfg is dense in Sub(G). The LERF property implies that Subfg ⊆ Subfi as every finitely generated subgroup is a descending intersection of finite index subgroups. This shows (1) =⇒ (2). Now (2) =⇒ (3) follows directly from Corollary (2.6). To prove (3) =⇒ (1), assume we are given a finitely generated infinite index subgroup L = hSi and g ∈ G r L. Let ϕ ∈ Hom(G, X!) be any permutation representation which is isomorphic to the quasiregular action G y (G/L) in such a way that x ∈ X is identified with the trivial coset eL. If ψ ∈ O(ϕ, S ∪ {g}, {x}) is an action with finite orbits then [L : Gx (ψ)] < ∞, L < Gx (ψ) but g 6∈ Gx (ψ) proving the LERF property. Assume now that G is finitely generated. The implications (5) =⇒ (4) =⇒ (3) are obvious, so it is enough to prove the implication (2) =⇒ (5). We start by describing the generic permutation representation τ ∗ ∈ Hom(G, X!). Let Subfg (G) = {H1 , H2 , . . . } be an enumeration of the finite index subgroups of G, let ϕn be the quasi-regular representation of G on G/Hn and define: G τ∗ = ∞ · ϕn . Namely, take countably many copies of each representation in the list, and let G act naturally on the disjoint union of the corresponding sets. It follows from Corollary 2.6, applied to the open dense set D = Subfi (G) ⊆ Sub(G), that the collection of permutation representations all of whose orbits are finite is generic. It follows from Lemma 2.7 applied to the open set {Hi } ⊆ Sub(G) that the collection of permutation representations in which ρi appears countably many times as a transitive component is also generic. By Baire’s category theorem a generic permutation representation has only finite orbits and each ρi appears in it countably many times. But such a permutation representation must be permutation isomorphic to τ ∗. Remark 3.2. As the examples below demonstrate if G is an infinitely generated LERF group it is no longer true that a generic permutation representation has only finite orbits. It 10 is still true, however, that the restriction of such a generic action to every finitely generated subgroup H < G has only finite orbits. It is even true that the restriction of a generic permutation to H admits a well defined isomorphism type (up to isomorphism of permutation representations of H). The details of the proof are quite similar to our proof above and we leave them to the reader. In order to demonstrate the use of Theorem 1.5 above and give some basic examples we analyze the situation in free groups. Providing a short proof to Hall’s theorem that free groups are LERF. Proposition 3.3. Let Fn , 1 ≤ n ≤ ∞ be a free group. (1) Fn is LERF. (2) A generic permutation representation ϕ ∈ Hom(F∞ , X!) is transitive. (3) If Γ is a countable LERF group that is not finitely generated then Sub(Γ) is perfect. Proof. First notice that Hom(Fn , X!) = X!n . Let X!(f ) < X! be subgroup of the dense n (f ) n (f ) finitely supported permutations. Clearly Hom(Fn , X! ) = X! ⊆ X! = Hom(Fn , X!) is a dense set of permutation representations all of whose orbits are finite. This proves (1) by establishing Theorem 1.5(3). To prove (2) it is enough, by Baire’s theorem, to show that the set Θ(x, y) = {ϕ ∈ Hom(F∞ , X!) \| y ∈ ϕ(F∞ )x} is open and dense. Openness is obvious. For the density, fix a free generating set F∞ = hx1 , x2 , . . .i. Given a basic open set O(ϕ, S, A) ⊆ Hom(F∞ , X!) the finite set S ⊆ F∞ contains words that involve only finitely many of the generators, say S ⊆ hx1 , x2 , . . . , xr i. We can find σ ∈ O(ϕ, S, A) ∩ Θ(x, y) by setting σ(xi ) = ϕ(xi ) for every 1 ≤ i ≤ r and then defining σ(xr+1 ) in such a way that σ(xr+1 )x = y. Finally if Γ fails to be finitely generated then so do her finite index subgroups. So by Proposition 2.2(2) none of these are isolated. If Γ is also LERF then the finite index subgroups are dense and in particular there can be no isolated subgroups at all. This proves (3). 4. Solitary groups This section is dedicated to the proof of Theorem 1.3. Proof. Assume that Is(G) is dense in Sub(G). Let Is(G) = {∆1 , ∆2 , ∆3 , . . . , Σ1 , Σ2 , . . .} be representatives for the conjugacy classes of these isolated subgroups of G. Where we made a distinction between the groups ∆i that are of finite index in their normalizer and the groups δi σi Σi that are not. We denote by G y G/∆i , G y G/Σi the corresponding quasiregular actions. It is important to note that in G/∆i there are finitely many points whose stabilizer is ∆i under the δi action. In G/Σi there are infinitely many similar points. With this terminology in place we can describe the generic permutation representation. It will have countably many orbits isomorphic to each δi and one orbit isomorphic to each σi : ! ! G G ∗ τ = ℵ0 δi ⊔ σj . i j Applying Corollary 2.6 to the open dense subset Is(G) ⊆ Sub(G) we conclude that a generic permutation representation has all of its stabilizers in Is(G). In other words a generic (⊔i di δi ) ⊔ (⊔i si σi ) for some di , si ∈ N ∪ {∞}. By Lemma 2.7, applied to the open set {∆i } 11 we know that a generic representation has infinitely many points whose stabilizer is ∆i , which immediately implies that di = ∞, ∀i. Note that an identical argument tells us that a generic representation has countably many points whose stabilizer is Σi , but even one orbit isomorphic to σi is enough to ensure that so this does not add any information about the coefficients si . All that is left to prove is that a generic representation has only one orbit isomorphic to σi for every i. In order to simplify the notation we will hence fix the index i and denote σ = σi , Σ = Σi . Let us denote by [Σ] = {gΣg −1 \| g ∈ G} ⊆ Sub(G) the conjugacy class of Σ. The bad event is the existence of two different orbits with stabilizers in [Σ]: [ Θ = Θx,y x,y∈X Θx,y = {ϕ ∈ Hom(G, X!) \| Gx (ϕ) = Gy (ϕ) = Σ; but ϕ(G)x 6= ϕ(G)y} By continuity of the stabilizer map (Lemma 2.5) the sets Gx−1 ({Σ}), G−1 y ({Σ}) and hence also Θx,y are closed. So by Baire’s theorem it suffices to prove that Θx,y is nowhere dense. Assume to the contrary that O = O(ϕ, S, A) ⊆ Θx,y for some basic open set. Replacing −1 if necessary O by a smaller basic open set, we may assume S that S = S and that O ⊆ −1 −1 Gx ({Σ}) ∩ Gy ({Σ}). Let Ax = ϕ(G)x ∩ A and Ωx = s∈S ϕ(s)Ax . We define Ay , Ωy similarly. Consider the quasiregular action σ : G y G/Σ, if g ∈ NG (Σ) then there is a unique G invariant isomorphism ηgΣ : ϕ(G)x → G/Σ ϕ(h)x 7→ hgΣ. Since [NG (Σ) : Σ] = ∞ by assumption there are infinitely many possible choices of points gΣ that would work for x and similarly for y. Let gx Σ, gy Σ be two such choices satisfying the additional property ηgx (Ωx ) ∩ ηgy (Ωy ) = ∅. Let α : ϕ(G)x ⊔ ϕ(G)y → G/Σ be any bijection such that α(z) = ηgx (z), ∀z ∈ Ωx and α(z) = ηgy (z), ∀z ∈ Ωy . We define a new action ψ ∈ O by the following formula: −1 α (gα(x)) if x ∈ ϕ(G)x ⊔ ϕ(G)y ψ(g)(x) = ϕ(g)x otherwise It is easy to verify that ψ ∈ O, that Gx (ψ), Gy (ψ) = Σ and that x, y are in the same ψ(G)-orbit. Which completes the proof of the first implication. Assume now that there exists a generic permutation representation τ ∗ ∈ Hom(G, X!), by assumption its isomorphism class Φ = {aτ ∗ a−1 \| a ∈ X!} is co-meager in Hom(G, X!). The collection of subgroups appearing as point stabilizers of τ ∗ are given by O = {Gx (τ ∗ ) \| x ∈ X} = Gx0 (Φ) ⊆ Sub(G), where x0 is an arbitrary basepoint. Since, by Lemma 2.5, the map Gx0 is surjective and continuous O is dense in Sub(G). In particular O ⊃ Is(G). We will show that O ⊆ Is(G) thus showing that the latter is dense and completing the proof. Let Σ ∈ O and [Σ] = {gΣg −1 \| g ∈ G} its conjugacy class. By Proposition 2.2(1) it is enough to show that [Σ] is open. If [Σ] fails to be open it must have an empty interior, 12 because G acts transitively on [Σ]. Since [Σ] is countable it follows from Baire’s theorem that Sub(G) \ [Σ] is a dense Gδ set. By Corollary 2.6 ^\ [Σ] = {σ ∈ Hom(G, X!) \| Gx (σ) 6∈ [Σ], ∀x ∈ X} Sub(G) is also a dense Gδ set. But this contradicts the fact that Φ is dense Gδ as the intersection of these two sets is empty. We conclude this section by proving Theorem 1.7. Proof. (of Theorem 1.7) In view of the fact that in a finitely generated group G finite index subgroups are isolated (1) follows directly from the comparison of Theorems 1.3 and 1.5. Density of isolated points (statement 2) is a general fact about countable Baire spaces. Indeed the set Sub(G) \ Is(G) is nowhere dense since it is a countable union of closed, nowhere dense points. Consider a short exact sequence as in statement (3). Since N is finitely generated Env(N) is clopen. It is easy to verify that the correspondence principle, between subgroups of G and subgroups of H containing N, gives rise to a homeomorphism Env(N) → Sub(G) Θ 7→ ϕ(Θ) Claim (3) follows immediately. It is well known that the free product of two LERF groups is LERF [Rom69, Bur71]. If one of the groups, say H is trivial then G ∗ hei ∼ = G and the situation is clear. Thus to establish (4) we have to show that if neither group is trivial and G fails to be LERF then G ∗ H cannot be solitary. Let Σ ∈ Subfg (G) \ Subfi be some finitely generated subgroup that cannot be approximated by finite index subgroups. Let η : G ∗ H → G be the map that is the identity on G and trivial on H. We will show that Ξ := η −1 (Σ) = Σ ker(η) ∈ Sub(G ∗ H) is a subgroup that cannot be approximated by isolated subgroups. Indeed let Ω ⊆ G be some finite set such that the neighborhood WSub(G) (Σ, Ω) does not contain any finite index subgroup. We can use the same Ω to define an open neighborhood in G ∗ H and it is clear that [G : ∆ ∩ G] = ∞ for every ∆ ∈ SubSub(G∗H) (Ξ, Ω). Thus our theorem is proved in view of the following lemma, which seems very useful in its own right. Lemma 4.1. Let G, H be two countable groups with G infinite and H non-trivial. If ∆ ∈ Is(G ∗ H) then [G : G ∩ ∆] < ∞. Proof. It will be more convenient to argue at the level of actions and Schreier graphs. Note that Hom(G ∗ H, X!) ∼ = Hom(G, X!) × Hom(H, X!). We will denote this isomorphism by ϕ ∗ ψ 7→ (ϕ, ψ), namely ϕ ∗ ψ is the unique action of G ∗ H whose restriction to G is ϕ and to H is ψ. Let ∆ ∈ Sub(G ∗ H) and assume that [G : G ∩ ∆] = ∞. Let ϕ ∗ ψ ∈ Hom(G ∗ H, X!) be any action which is isomorphic to the quasiregular action G ∗ H y (G ∗ H)/∆ with x ∈ X identified with the trivial coset e∆. Note that, while the action of G ∗ H is transitive, ϕ, ψ themselves need not be transitive. Still, by our assumption we know that the orbit Y := ϕ(G)x ⊆ X is infinite. The argument now is simple enough : we obtain approximating n actions of the form ϕ ∗ ψn → ϕ ∗ ψ by carrying out small perturbations on the action of H. Since the orbit Y is infinite we can do arbitrarily small such perturbations on the action while still affecting the stabilizer of the point x. We elaborate below but this is basically a complete proof. 13 Let x ∈ A0 ⊆ A1 ⊆ A2 . . . be finite sets ascending to the whole of X, Tn ⊆ H finite sets ascending to a generating set of H. We can assume that 1 ∈ Tn = Tn−1 . Of course if H is finitely generated we can just take Tn = T, ∀n ∈ N to be some fixed symmetric set of generators. Set Bn = ∪t∈Tn ψ(t)(An ) and ξn ∈ Y \ Bn . Since X is infinite we can find bijections fn : X → X × {0, 1}, with the additional properties that fn (x) = (x, 0), ∀x ∈ An and fn (ξn ) = (ξn , 1). Let η ∈ hom(H, X!) be any fixed action of H on X. Using all this data we construct a sequence of actions ψ η ∈ Hom(H, (X × {0, 1})!) as follows: (ψ(h)x, 0), i = 0 η ψ (h)(x, i) = (η(h)x, 1), i = 1 and let ψnη (h) = (fn )−1 ◦ ψ η ◦ fn . It is clear from the definitions that ψnη ∈ O(ψ, Tn , An ) and in particular ψnη → ψ as n → ∞. Consequently of course ϕ ∗ Ψηn → ϕ ∗ ψ in Hom(G ∗ H, X!). Let 1, λ ∈ Hom(H, X!) be the trivial and the regular left action of H on X. The latter is defined via an arbitrarily chosen identification of X with H which will not play a role in the discussion. If H is finite we replace the regular left action by countably many copies of the same action, just to make sure λ is an action on an infinite set. We use these to obtain two convergent sequences of actions and hence two convergent sequences of subgroups ϕ ∗ ψn1 → ϕ ∗ ψ ϕ ∗ ψnλ → ϕ ∗ ψ (G ∗ H)x (ϕ ∗ ψn1 ) → ∆ (G ∗ H)x (ϕ ∗ ψnλ ) → ∆ These sequences are different because if gn ∈ Gn is any element such that ϕ(gn )x = ξn then for any 1 6= h ∈ H we have gn−1hgn ∈ (G ∗ H)x (ϕ ∗ ψn1 ) \ (G ∗ H)x (ϕ ∗ ψnλ ). Thus at least one of these sequences is not eventually constant, proving that the limit point ∆ is not isolated. 5. Sketch of another proof of Theorem 1.3 We briefly sketch another proof of Theorem 1.3, which is more along the lines of the arguments in [KR07]. First, we note that for any countable group G, the action of X! on Hom(G, X!) is topologically transitive, that is, given any two nonempty open subsets U, V of Hom(G, X!), there always exists α ∈ X! such that α · U ∩ V 6= ∅ (equivalently, there exists elements in Hom(G, X!) which have a dense conjugacy class). This is true simply because any two actions π1 , π2 of G on an infinite countable set embed in a third one π3 (for instance, obtained by considering a disjoint union of two infinite countable sets, with G acting as π1 on the first copy, and π2 on the second copy). Then the closure of the conjugacy class of π3 contains both π1 and π2 , proving the desired result. This brings us to the setting of the following lemma; the equivalence between (1) and (3) below is the criterion we will be using, and is due to C. Rosendal. The equivalence of these conditions with (2) appears to be new, and seems potentially useful so we are including it here even though it will not be needed. Lemma 5.1. Assume that P is a Polish group acting continuously and topologically transitively on a Polish space Z. Then the following conditions are equivalent: (1) There exists a comeager orbit. 14 (2) For any open identity neighborhood 1 ∈ V ⊆ P the collection of points {z ∈ Z \| V (z) is somewhere dense} = {z ∈ Z \| Int(V (z)) 6= ∅} is dense in Z. (3) For any open identity neighborhood 1 ∈ V ⊆ P and any nonempty open subset U of Z, there exists a nonempty open U ′ ⊆ U such that, for any nonempty open W1 , W2 ⊆ U ′ , one has V W1 ∩ W2 6= ∅. Proof. Fix an open identity neighborhood V and a sequence of group elements such that ∪n αn V = P . If there exists a comeager orbit then by Baire’s category theorem for every z in this orbit there exists an n such that αn V (z) is somewhere dense. Translating by αn we deduce that V (z) itself is somewhere dense. This proves that (1) implies (2). Assume that (2) holds; fix an open identity neighborhood 1 ∈ V ⊆ P and a nonempty open subset U of Z. Using our assumption, and continuity of the action, we may find a symmetric open identity neighborhood V ′ such that V ′ V ′ ⊆ V , an open U ′ ⊆ U and x ∈ U such that the closure of V ′ (x) contains U ′ . Then, for any W1 , W2 nonempty open and contained in U ′ , we have some v1 , v2 ∈ V ′ such that v1 x ∈ W1 , v2 x ∈ W2 . Thus v2 v1−1 W1 ∩ W2 6= ∅ and (3) holds. Finally, assume that (1) is false; since there exist dense orbits, any orbit must be meager or comeager, by the 0–1 topological law [RR74]. So in this case all orbits are meager. Given z S ∈ Z, we then have an family of closed subsets Fn with empty interior such that P (z) ⊆ Fn ; some {g ∈ P : gz ∈ Fn } must have nonempty interior, proving that there exists some open neighborhood V of 1 such that V (z) is nowhere dense. Thus Z is the union of sets of the form {z ∈ Z : V (z) is nowhere dense}, where V ranges over a countable basis of neighborhoods of 1; one of these sets must be nonmeagre, hence (since these sets are Borel) comeagre in some nonempty open U. Assume that (3) holds, and pick U ′ ⊆ U witnessing it. The assumption of (3) amounts to saying that {z ∈ U ′ : V (z) ∩ W 6= ∅} is dense open in U ′ for any nonempty open W ⊆ U ′ ; this implies that {z ∈ U ′ : V (z) is dense in U ′ } is comeagre in U ′ . This is a contradiction with the fact that V (z) must be nowhere dense for a generic element of U, hence also of U ′ . Now, we need to understand when the above criterion is satisfied, for G a countable group, P = X! and Z = Hom(G, X!). Given an open set U = O(ρ, S, A), let i(ρ) denote the number of distinct G-orbits of elements of A; we may pick ρ such that i(ρ) number is minimal among elements of U. Then, enlarging S and shrinking A as needed, we can reduce to the situation where ∀π ∈ U ∀a 6= a′ ∈ A ∀g ∈ G π(g)(a) 6= a′ . We are now in a situation where orbits of elements of A cannot interfere with each other; this enables us to reduce to the case where A = {a} is a singleton and we are working inside the Polish space Z ′ of transitive G-actions. Consider an open set U = O(ρ, S, a) ∩ Z ′ ; let Ga denote the stabilizer of a for this action, and let V be the group of permutations fixing a finite set F . Enlarging S if needed, we assume that F ⊆ Sa. Then, it is readily checked that two elements ρ1 , ρ2 of U belong to the same V -orbit iff the stabilizers of a for ρ1 and ρ2 are the same. Let S1 = {g2 g1−1 : g1 , g2 ∈ S and g1 (a) = g2 (a)} , S2 = {g ∈ S : g(a) 6= a} . 15 The discussion above shows that the criterion (3) of Lemma 5.1 is satisfied iff there exists an open set W = O(ρ̃, S̃, a) ∩ Z ′ contained in U such that the stabilizer of a is the same for any two elements of W ; that is, if and only if there exist finite sets S̃1 , S̃2 ⊆ G, with S̃1 ⊇ S1 and S̃2 ⊇ S2 , and a subgroup G′a such that: ∀H ∈ Sub(G) (∀g ∈ S1 g ∈ H and ∀g ∈ S2 g 6∈ H) ⇔ (H = G′a ) . Thus H is an isolated point in Sub(G); since Ga was an arbitrary subgroup of G, and S1 , S2 encode an arbitrary open neighborhood of Ga , we just established that there exists a generic action in Hom(G, X!) if, and only if, G is solitary. One can try to use the same approach as above to understand when there exists generic conjugacy classes in Hom(G, P ) for other Polish groups P . But, as the structure of P becomes more complicated, the above analysis is harder to carry out (in particular, the reduction to transitive actions no longer works). One case when one can do it is when P = Aut(R) is the automorphism group of the random graph. Then, reasoning in much the same way as above, one obtains the following criterion, which probably can be further simplified. Proposition 5.2. Given G a countable group, there exists a generic element in Hom(G, Aut(R)) iff the following condition is satisfied: For any finitely-generated subgroups H1 , . . . , Hn of G, and any finite Ki,j ⊆ G with Hi Hj ∩ Ki,j = ∅ for all i ≤ j, there exists finitely generated subgroups H1′ , . . . , Hn′ such that: • For all i Hi′ contains Hi . • For all i, j the double coset space Hi′\G/Hj′ is finite. • For all i ≤ j one has Hi′Hj′ ∩ Ki,j = ∅. The above property is hard to grasp; it does imply that G is finitely generated, and that any finitely generated subgroup of G is an intersection of finitely generated subgroups with finite bi-index. Must a group satisfying the previous conditions be LERF? 6. A-separability Definition 6.1. An action G y X of a discrete, countable group G is called amenable if it satisfies any one of the following equivalent conditions: • For every ε > 0 and Ω ⊆ G finite, X admits an (ε, Ω)-Følner subset, that is, a finite \|gF ∆F \| set F ⊆ X such that < ε for all g ∈ Ω. \|F \| • There exists a finitely additive G-invariant probability measure on X. When the action is transitive, of the form G y G/K, these conditions are further equivalent to the following: • If G acts continuously on a compact space and K admits an invariant Borel measure, then so does G. In the transitive case it is sometimes convenient to adopt group theoretic terminology as follows: Definition 6.2. A subgroup K of a group G is called co-amenable if the quasiregular action G y G/K is amenable. 16 The equivalence of these three conditions is classical. By definition, an amenable action always admits a Følner-sequence. This is a sequence of finite subsets Fn ⊆ X such that for \|gFn △Fn \| all g ∈ G we have lim = 0. We recall the following: n→∞ \|Fn \| Remark 6.3. A Følner-sequence can be chosen to be increasing (with respect to inclusion). As defined in the introduction, a group G is A-separable if the set of co-amenable subgroups is dense in Sub(G). We now prove Theorem 1.12, giving a characterization of A-separability in the language of generic actions: Proof of Theorem 1.12. Denote by CoAm(G) the set of all co-amenable subgroups of G. Note For x ∈ X denote σ Σ(x) = {σ ∈ Hom(G, X!) \| G y (G)x is amenable} = {σ ∈ Hom(G, X!) \| Gx (σ) ∈ CoAm(G)}. and Σ = ∩x∈X Σ(x). If Σ is generic then Σ(x) is dense in Hom(G, X!) and by Lemma 2.5 the image of this set {Gx (σ) \| σ ∈ Σ(x)} is a dense subset of Sub(G) consisting of co-amenable subgroups. Conversely, assume that CoAm(G) is dense in Sub(G) and we wish to prove that the set Σ is generic in Hom(G, X!). It is enough to show that Σ(x) is generic in Hom(G, X!) for every x ∈ X. The density of Σ(x) is assured by the hypothesis, the fact that Σ(x) = G−1 x (CoAm(G)) and Lemma 2.5. To show that Σ(x) is Gδ , it is enough to show that the condition that a specific finite set F ⊆ σ(G)x is (ε, Ω)-Følner is open, where ε > 0 and Ω ⊆ G is finite. Assume this is the case for some σ ∈ Hom(G, X!) we seek an open neighborhood σ ∈ O ⊆ Hom(G, X!) such that F is still contained in the orbit, and is still Følner for every ϕ ∈ O. For every f ∈ F pick a group element gf ∈ G such that f = σ(gf )x. Let F ′ = F ∪{x} and Ω′ = Ω ∪ {gf \| f ∈ F } - the desired neighborhood is given by O = O(σ, F ′ , Ω′ ). As mentioned in the introduction, LERF groups and amenable groups are A-separable but they are not the only examples of A-separable groups. In order to give an example of an Aseparable group which is neither LERF nor amenable, we will first prove that A-separability is closed under taking free products: Theorem 6.4. Let G and K be countable groups. If G and K are A-separable then so is G ∗ K. Proof. Every element of Hom(G ∗ K, X!) is of theform ϕ ∗ ψ for ϕ ∈ Hom(G, X!), ψ ∈ Hom(K, X!); where ϕ ∗ ψ is defined by setting ϕ ∗ ψ (g) = ϕ(g) and ϕ ∗ ψ (k) = ψ(k), ∀g ∈ G, k ∈ K and expanding the definition to the free product. For every x ∈ X, ε > 0 and finite subsets S ⊆ G, T ⊆ K let Σ(x, ε, S, T ) = {σ ∗ τ ∈ Hom(G ∗ K, X!) \| the (σ ∗ τ )-orbit of x contains an (ε, S ∪ T )-Følner set}. We want to prove that Σ = Σ(x, n1 , S, T ) x ∈ X, n ∈ N, S ⊆ G, T ⊆ K finite is generic in Hom(G ∗ K, X!). Since X, G and K are countable, it is enough to show that the sets Σ(x, ε, S, T ) are open and dense for every x ∈ X, ε > 0 and finite subsets S ⊆ G, T ⊆ K. The argument that shows that Σ(x, ε, S, T ) is open was given in the proof of Theorem 1.12. T 17 Fix x, ε, S and T as above. We prove that Σ(x, ε, S, T ) is dense in Hom(G ∗ K, X!). Let ϕ ∗ ψ ∈ Hom(G ∗ K, X!) and let A ⊆ X be finite. We will find ϕ′ ∈ Hom(G, X!) and ψ ′ ∈ Hom(K, X!) such that ϕ′ (s)a = ϕ(s)a, ψ ′ (t)a = ψ(t)a for all s ∈ S, t ∈ T, a ∈ A and such that ϕ′ ∗ ψ ′ ∈ Σ(x, ε, S, T ). We can assume that x ∈ A. By A-separability, there exist σ ∈ Hom(G, X!) and τ ∈ Hom(K, X!) such that σ(s)a = ϕ(s)a, τ (t)a = ψ(t)a for all σ τ s ∈ S, t ∈ T, a ∈ A and such that the actions G y X and K y X are amenable on every orbit. Let L := σ ∗ τ (G ∗ K) = hσ(G), τ (K)i. Case 1: all the σ and τ orbits which are contained S in Lx are finite. Let B ⊆ Lx be a finite, σ-invariant set containing A∩Lx and let C = b∈B τ (K)b. We define a representation ϕ′ ∈ Hom(G, X!) by declaring every c ∈ C r B and every element in the σ-orbit of c to be a fixed point for ϕ′ and on every other element of X, ϕ′ (g) identifies with σ(g) for all g ∈ G. Notice that since B is σ-invariant, ϕ′ (g) is well defined and acts the same as σ(g) on B for all g ∈ G. In particular, ϕ′ (g) agrees with ϕ(g) on A. We have that C is finite, invariant under both ϕ′ and τ and contains x. Setting ψ ′ = τ , the (ϕ′ ∗ ψ ′ )-orbit of x is finite so the orbit itself is an (ε, S ∪ T )-Følner set for ϕ′ ∗ ψ ′ . Case 2: Lx contains either an infinite σ-orbit or an infinite τ -orbit. Assume, S without loss of generality, that Lx contains an infinite τ -orbit Y . Denote B = A ∪ s∈S σ(s)A and let Fn be an increasing Følner-sequence in Y for the τ -action. Since the sets Fn are finite, none of them is τ -invariant and so the Følner-sequence does not stabilize. This implies that \|Fn \| → ∞ and in particular, Y contains an (ε, T )-Følner set F such that 2(\|B\| + 1) \|F \| > . Now, let z ∈ G ∗ K be such that (σ ∗ τ )(z)x ∈ F and such that z is of ε minimal length with respect to the canonical presentation: z = gn kn gn−1kn−1 · · · g1 k1 (gi ∈ G, kj ∈ K, g1, . . . , gn−1 , k2 , . . . , kn 6= 1). Denote y = (σ ∗ τ )(z)x. By Lemma 2.8, we can assume that σ has infinitely many fixed points. In particular, there exits a set C ⊆ X on which σ(G) acts trivially, such that \|C\| = \|F r (B ∪ {y})\| and such that C does not intersect the finite set B ∪ F ∪ trace(σ∗τ )(z) (x) where we think of (σ ∗ τ )(z) as the word over X! corresponding to the given presentation of z. Denote D = F r (B ∪ {y}) and let ξ ∈ X! be a permutation of order 2 that takes C bijectively onto D and acts trivially on X r(C ∪D). We define an action ϕ′ ∈ Hom(G, X!) by ϕ′ (g) = ξ −1 σ(g)ξ for all g ∈ G. Since ξ acts trivially on B we have that ∀s ∈ S, ∀a ∈ A : ϕ′ (s)a = σ(s)a = ϕ(s)a and that every element of D is fixed under ϕ′ (s) for all s ∈ S. Hence: ∀s ∈ S : 2\|F r D\| 2(\|B\| + 1) \|ϕ′ (s)F ∆F \| ≤ ≤ < ε. \|F \| \|F \| \|F \| Thus F is (ε, S)-Følner for ϕ′ and (ε, T )-Følner for ψ ′ := τ and thus F is (ε, S ∪T )-Følner for ϕ′ ∗ψ ′. Notice that by the minimality of the length of z we have that trace(σ∗τ )(z) (x)∩F = {y} and so ξ acts trivially on trace(σ∗τ )(z) (x). This means that (ϕ′ ∗ ψ ′ )(z)x = y ∈ F and since F is contained in a τ -orbit this implies that F is contained in the (ϕ′ ∗ ψ ′ )-orbit of x, as required. Recall that the (m, n) Baumslag-Solitar group is the group BS(m, n) = hs, t \| t−1 sm t = sn i. It is well known that BS(m, n) is solvable (hence amenable) if and only if m = 1. Proposition 6.5. For every n, the group BS(1, n) is not LERF. 18 Proof. Write: BS(1, n) = hs, t \| t−1 st = sn i and notice that t−1 hsit = hsn i $ hsi. Thus, an element of hsi r t−1 hsit cannot be separated from t−1 hsit by a homomorphism into a finite group. Corollary 6.6. There exist non-LERF, non-amenable A-separable groups. Proof. Let G = BS(1, n) for some n. G is amenable hence A-separable and so, by Proposition 6.4, G ∗ G is A-separable. On the other hand, G ∗ G is not LERF since G is not LERF and the LERF property passes to subgroups. G ∗ G is also not amenable since it contains a free subgroup on two generators. In order to complete the proofs of all the statements promised in the introduction we prove the following: Proposition 6.7. A group G with Kazhdan property (T) is A-separable if and only if it is LERF. In particular the following groups are never A-separable: • Groups with property (T) that are not residually finite, and in particular any simple group with property (T). • Irreducible lattices in higher rank semi-simple Lie groups with no compact factors that satisfy the congruence subgroup property. Proof. This follows directly from the fact that a transitive action G y G/H is amenable if and only if G/H is finite. The argument for that follows from property (T). If this action is amenable and F ⊆ G/H is an (K, ε) Følner set then 1F ∈ ℓ2 (G/H) is a (K, ε)-almost invariant vector. Taking (K, ε) to be Kazhdan constants for G we can deduce the existence of a non-zero invariant vector f ∈ ℓ2 (G/H). Since the action of G on G/H is transitive f must be a constant function. But a non-zero constant function is in ℓ2 if and only if G/H is finite. Now if Γ is a lattice as in the statement of the theorem it cannot be LERF because by the strong approximation theorem [LS03, Window 9],[Nik11] every Zariski dense subgroup has a finite index closure in the pro-congruence topology; which coincides with the profinite topology by assumption. By Kazhdan’s theorem such a lattice has property (T) and the statement follows. Note that conjecturally the congruence subgroup property automatically holds for such higher rank lattices, and this is indeed proven in many different cases. In particular the groups SLn (Z), n ≥ 3 are good examples for generically finite groups that is not A-separable. We conjecture further that the LERF property can never occur in a nontrivial way for property (T) groups namely: Conjecture 6.8. A countable group G with Kazhdan property (T) is LERF if and only if it is finite. It was pointed out to us by Matthew Stover that A similar question was already asked by Long and Reid in [LR08, Question 4.5]. We are grateful to some very constructive comments on the previous version of this paper by Miklós Abért, Eli Glasner, and Matthew Stover. J.M wishes to thank Romain Tessera for interesting conversations and suggestions about solitary groups. This work was written while the first author was on sabbatical at the University of Utah. I am grateful to the math department there for their hospitality. Y.G. acknowledges support 19 from U.S. National Science Foundation grants DMS 1107452, 1107263, 1107367 “RNMS: Geometric structures And Representation varieties” (the GEAR Network) that enabled this visit. The first and second authors were both partially supported by the Israel Science Foundation grant ISF 2095/15, and the third author was partially supported by Agence Nationale de la Recherche grant Grupoloco (ANR-11-JS01-0008). References [Ago13] Ian Agol. The virtual Haken conjecture. Doc. Math., 18:1045–1087, 2013. With an appendix by Agol, Daniel Groves, and Jason Manning. [Bur71] R. G. Burns. On finitely generated subgroups of free products. J. Austral. Math. Soc., 12:358–364, 1971. [GK14] R. Grigorchuk and R. Kravchenko. On the lattice of subgroups of the lamplighter group. Internat. J. Algebra Comput., 24(6):837–877, 2014. [GW03] R. I. Grigorchuk and J. S. Wilson. A structural property concerning abstract commensurability of subgroups. J. London Math. Soc. (2), 68(3):671–682, 2003. [GW08] Eli Glasner and Benjamin Weiss. Topological groups with Rokhlin properties. Colloq. Math., 110(1):51–80, 2008. [Hal49] Marshall Hall, Jr. Subgroups of finite index in free groups. Canadian J. Math., 1:187–190, 1949. [HHLS93] Wilfrid Hodges, Ian Hodkinson, Daniel Lascar, and Saharon Shelah. The small index property for ω-stable ω-categorical structures and for the random graph. J. London Math. Soc. (2), 48(2):204– 218, 1993. [Kit15] Daniel Kitroser. Generic Actions of Countable Groups. PhD thesis, Ben-Gurion university of the Negev, 2015. [KR07] Alexander S. Kechris and Christian Rosendal. Turbulence, amalgamation, and generic automorphisms of homogeneous structures. Proc. Lond. Math. Soc. (3), 94(2):302–350, 2007. [LR08] D. D. Long and A. W. Reid. Subgroup separability and virtual retractions of groups. Topology, 47(3):137–159, 2008. [LS03] Alexander Lubotzky and Dan Segal. Subgroup growth, volume 212 of Progress in Mathematics. Birkhäuser Verlag, Basel, 2003. [MT13] Julien Melleray and Todor Tsankov. Generic representations of abelian groups and extreme amenability. Israel J. Math., 198(1):129–167, 2013. [Nik11] Nikolay Nikolov. Strong approximation methods. In Lectures on profinite topics in group theory, volume 77 of London Math. Soc. Stud. Texts, pages 63–97. Cambridge Univ. Press, Cambridge, 2011. [Rom69] N. S. Romanovskiı̆. On the residual finiteness of free products with respect to subgroups. Izv. Akad. Nauk SSSR Ser. Mat., 33:1324–1329, 1969. [Ros11a] Christian Rosendal. Finitely approximable groups and actions part I: The Ribes-Zalesskiı̆ property. J. Symbolic Logic, 76(4):1297–1306, 2011. [Ros11b] Christian Rosendal. Finitely approximable groups and actions part II: Generic representations. J. Symbolic Logic, 76(4):1307–1321, 2011. [RR74] M. Bhaskara Rao and K. P. S. Bhaskara Rao. A category analogue of the Hewitt-Savage zero-one law. Proc. Amer. Math. Soc., 44:497–499, 1974. [Sco85] Peter Scott. Correction to: “Subgroups of surface groups are almost geometric” [J. London Math. Soc. (2) 17 (1978), no. 3, 555–565; MR0494062 (58 #12996)]. J. London Math. Soc. (2), 32(2):217– 220, 1985. [Wil08] Henry Wilton. Hall’s theorem for limit groups. Geom. Funct. Anal., 18(1):271–303, 2008. Yair Glasner. Department of Mathematics. Ben-Gurion University of the Negev. P.O.B. 653, Be’er Sheva 84105, Israel. yairgl@math.bgu.ac.il Daniel Kitroser. Department of Mathematics. Ben-Gurion University of the Negev. P.O.B. 653, Be’er Sheva 84105, Israel. kitrosar@post.bgu.ac.il 20 Julien Melleray. Université Claude Bernard – Lyon 1 . Institut Camille Jordan, CNRS UMR 5208 . 43 boulevard du 11 novembre 1918, 69622 Villeurbanne Cedex, melleray@math.univ-lyon1.fr 21	4math.GR
arXiv:1802.05662v1 [cs.DS] 15 Feb 2018 List Heaps Andrew Frohmader∗ February 2018 Abstract This paper presents a simple extension of the binary heap, the List Heap. We use List Heaps to demonstrate the idea of adaptive heaps: heaps whose performance is a function of both the size of the problem instance and the disorder of the problem instance. We focus on the presortedness of the input sequence as a measure of disorder for the problem instance. A number of practical applications that rely on heaps deal with input that is not random. Even random input contains presorted subsequences. Devising heaps that exploit this structure may provide a means for improving practical performance. We present some basic empirical tests to support this claim. Additionally, adaptive heaps may provide an interesting direction for theoretical investigation. 1 Introduction A heap is a data structure which holds a finite set of items. Each item is associated with a key drawn from a totally ordered set. Heaps support the following operations: make heap (h): insert (h, x, k): find min (h): delete min (h): decrease key (h, x, k): delete (h, x): meld (h1 , h2 ): Create and return a new, empty heap h Insert item x with key k into heap h and return a reference to where x is stored in h Return a reference to where the item with the minimum key is stored in heap h Delete the item with the minimum key from heap h and return it Decrease the key of item x in heap h to k Delete item x from heap h Return a heap formed by taking the union of heaps h1 and h2 The binary heap was introduced by Williams in 1964 [14]. Its simplicity and speed have made it and its generalization, the d-array heap, a popular choice in ∗ andrewfrohmader@gmail.com 1 practice. It supports insert, delete min, and decrease key in O(log n) time. It can be used to sort n items in O(n log n), which matches the worst-case lower bound for a comparison sort. Vuillemin’s introduction of the binomial queue in 1978 [13], added meld to the list of operations supported in O(log n). In 1984, Fibonacci heaps [8], an extension of the binomial queue, achieved O(1) amortized time for insert, decrease key, and meld. The decrease key result was particularly important in that it improved the worst-case bounds for a number of well-known graph algorithms. More recently, a few structures have achieved worst-case O(1) time for decrease key and meld, see [2] or [4]. While this work has produced interesting and important theoretical results, it has failed to yield a structure that consistently outperforms the original binary heap and its variants in practice [9]. In this paper, we return to the binary heap and develop a simple extension, the List Heap. This straightforward extension can be given adaptive operations: operations whose performance depends not only on the problem size, but also on the level of presortedness (disorder) in the problem instance. A bit of work has gone into developing the theory of adaptive sorting algorithms, see [7], but to our knowledge, this work has not migrated into the related work on heap data structures, [3]. We believe that adaptive heaps may provide an interesting angle for theoretical investigation. Additionally, they may provide a means of improving the empirical performance of current heap variants. The List Heap is a first step in this direction. List Heaps support decrease key, insert, and delete min in O(log k), where k is the number of lists in the List Heap. As we will show, the number of lists in a List Heap is a function of both the size of the problem instance and the disorder of the problem instance. We returned to the binary heap because of its simplicity and ubiquity, but this was not without costs. List Heaps lose the O(1) insert, decrease key, and meld of more sophisticated structures. 1.1 Preliminaries Here we present notational conventions and definitions used through the remainder of this paper. Let X = hx1 , ..., xn i be a sequence of n distinct elements xi from some totally ordered set. If x1 < x2 < ... < xn , X is monotonically increasing or just increasing. If x1 > x2 > ... > xn , X is monotonically decreasing or just decreasing. A sequence is monotonic if it is either increasing or decreasing. The head of a sequence X is x1 , the tail is xn . If A is a set, then \|\|A\|\| is its cardinality. If X is a sequence, then \|X\| is its length. For two sequences X = hx1 , ..., xn i and Y = hy1 , ..., ym i, their concatenation XY is the sequence hx1 , ..., xn , y1 , ..., ym i. If the sequence X contains no elements, we write X = ∅. A sequence obtained by deleting zero or more elements from X is called a subsequence of X. A subsequence Y = hxi , ..., xj i of X is consecutive if the indices i, ..., j are consecutive integers. Let Y = hxi , ..., xj i and Z = hxk , ..., xl i be subsequences of X. The intersection of Y and Z, Y ∩ Z, is the subsequence of X obtained by deleting from X all xh not in both Y and Z for 1 ≤ h ≤ n. Similarly, the union of Y and Z, 2 Y ∪ Z, is the subsequence of X obtained by deleting from X all xh not in either Y or Z for 1 ≤ h ≤ n. Y and Z are disjoint if Y ∩ Z = ∅. Let P = {X1 , ..., Xk } be a set of disjoint subsequences of X, if the union of all subsequences in P equals X, then P is a partition of X. 1.2 Adaptive Sorting This section gives a very brief review of adaptive sorting. Heaps solve a generalized sorting problem, so adaptive sorting provides some intuition for why adaptive heaps might be useful. For a more detailed survey of adaptive sorting, see [7] or [11]. Consider the sorting problem: take as input some arbitrary sequence X = hx1 , ..., xn i of elements from a totally ordered set and return a permutation of the sequence that is in increasing sorted order. Comparison based sorting has a well-know worst-case lower bound of Ω(n log n) [5]. However, it is clear that this lower bound must not always hold. What if our input sequence is already sorted? What if only one element is out of place? What if it is the concatenation of two sorted subsequences? The lower bound can be refined if we account for the disorder in the input sequence. The main achievements of the adaptive sorting literature are: proposing a variety of measures of disorder, proving new lower bounds with respect to these measures, developing sorting algorithms whose performance matches these new lower bounds, and developing a partial order on the set of measures. We stop here and again direct the reader to [7] or [11] for more information. 1.3 Outline of Paper The remainder of the paper is organized as follows. Section 2 discusses why adaptive heaps might be worth developing. Section 3 presents List Heaps their structure and operations. Section 4 presents the results of a series of brief empirical tests suggesting List Heaps may have promise in practice. Section 5 summarizes results obtained. 2 Adaptive Heaps This section presents a few reasons for developing adaptive heaps. We use the term adaptive heap loosely throughout this paper to refer to a heap whose performance is some function of the level of presortedness (disorder) of the input sequence. There are clearly complications we are glossing over, the largest of which is how to deal with decrease key. Further, heap problem instances can have disorder of different types, not just related to the presortedness of the input sequence. A starting point to formalize these notions is the adaptive sorting literature. It is fairly easy to extend the results of adaptive sorting to arbitrary sequences of insert and delete min; decrease key may prove more challenging. 3 Why build a heap whose performance adapts to the presortedness of the input sequence? Just as the lower bound for sorting can be restated as a function of the presortedness of the input sequence, the lower bound for an arbitrary heap problem can be restated as a function of the disorder of the problem instance. As delete min is the only operation for which O(1) performance is not available, the disorder of the problem directly impacts the bound on delete min. Thus, we can restate the bound on delete min as O(log k) where k is some measure of disorder less than or equal to the number of elements in the heap. Then, intuitively, the bound on delete min ranges from O(1) for sorted input to O(log n) for random input. The bounds on algorithms which rely on heaps, such as Dijkstra’s shortest path algorithm [15], can similarly be adjusted to reflect disorder. For example, we can restate the Dijkstra bound as O(n log k + m) where n is the number of vertices, m is the number of edges, and k is a measure of disorder with k ≤ n. The discussion above suggests that heap-based applications that have input sequences with some level of presortedness could benefit from adaptivity. What is less obvious is that even heap-based applications that have random input sequence can benefit from adaptivity. Here, the benefit is not in asymptotic performance, but in constant factors. Let X be a random sequence of n elements. As n → ∞, the minimum number of √ increasing subsequences into which X can be partitioned is approximately 2 n subsequences [1]. As we will show in Section 3.3, we can create an extension of the binary heap that is adaptive to the minimum number of increasing subsequences. Thus, where a binary heap constructed from X performs roughly 2 log n comparisons to delete min in the worst-case, an adaptive heap constructed from X can perform delete min in worst-case log n comparisons - cutting the constant factor in half. 3 List Heaps This section introduces the List Heap. The List Heap structure closely mirrors that of the binary heap. The hope is that this similarity makes the changes required to add adaptivity to heaps clear. Additionally, we believe adaptive variants of the binary heap have the greatest potential to be immediately useful in practice. However, binary heaps are flawed as a choice in that decrease key and insert take O(log n) time while O(1) implementations of these operations are clearly possible. Thus the List Heap cannot be optimal on all problem instances. We tolerate this flaw and show that we can construct a heap whose performance is a function of a number of measures of disorder. In particular, we will focus on developing heaps that are adaptive to runs, SU S, and Enc (defined later). These measures of disorder partition the input into monotonic subsequences which are particularly simple for a heap to use. This paper focuses only on insert, delete min, and decrease key and assumes all keys are unique. We first consider the structure of the heap. Next, we outline operations that are adaptive with respect to runs. Finally, we present operations that are adaptive with respect to SU S and Enc. 4 3.1 Structure A List Heap h is an array of circular doubly linked lists of nodes, h = hl1 , ..., lk i. Each node has a unique key associated with it. Throughout this paper we will refer to the node x and its key interchangeably. We use the notation x(i,j) to refer to the jth node in list li . List Heaps must maintain two invariants. 1. For any list li in heap h, the nodes of li are in increasing sorted order by key value, li = hx(i,1) , ..., x(i,j) i, with x(i,1) < x(i,2) < ... < x(i,j) . 2. The lists are arranged in heap order in the array, that is l⌊i/2⌋ < li where we define li < lj if x(i,1) < x(j,1) . With the invariants above, we can view a List Heap as a standard binary heap of elements: each binary heap element is a list and the key for each list is the key of its head node. With this structure, the minimum node is x(1,1) - the head node of list l1 . We will refer to l1 as the root list of the heap. 0 1 l1 1 3 2 l2 4 14 15 3 l3 2 8 9 4 l4 5 13 5 l5 6 10 6 l6 11 7 l7 7 12 16 Figure 1: The List Heap Structure The sorted linked lists are the key feature of the List Heap. They enable us to capture some of the existing order in the input sequence. As a result, List Heaps can be given operations that are adaptive with respect to some measures of disorder. 3.2 A Runs Adaptive List Heap In this section, we present operations for a List Heap that is adaptive with respect to runs, abbreviated RA List Heap. We show that the performance of each heap operation is a function of the number of lists in the heap and that the insert operation partitions the sequence of inserted elements into L lists where L is less than or equal to the number of runs in the sequence. 5 Definition 1. A run is a consecutive decreasing subsequence of maximum length. A subsequence Xi = hxj , ..., xk i of X is a run if and only if j, ..., k are consecutive integers and xj−1 < xj > ... > xk < xk+1 . The total number of runs in a sequence gives a measure of disorder. For example, consider the following random sequence, h3, 15, 14, 4, 9, 13, 5, 12, 10, 6, 1, 11, 8, 16, 2, 7i. (1) This sequence of sixteen elements can be partitioned into eight runs, h3i, h15, 14, 4i, h9i, h13, 5i, h12, 10, 6, 1i, h11, 8i, h16, 2i, h7i. (2) A decreasing sequence will consist of just one run, an increasing sequence of n elements will contain n runs, each with only one element. We now outline operations for the RA List Heap. 3.2.1 Insert To insert a new item x into heap h = hl1 , ..., lk i, do the following. If x is less than the head node of lk , (recall that we denote this x < lk ), set i = k. Otherwise, set i = k + 1. Then, while x < l⌊i/2⌋ , set i = ⌊i/2⌋ and repeat. Once the loop terminates, append x to the front of list li . Note that i might equal k + 1 in which case a new empty list lk+1 will need to be created before x is appended. Figure 1 shows the RA List Heap generated by inserting the random sequence (1) into an empty heap. 3.2.2 Delete Min To perform delete min on heap h, remove the head node x(1,1) from the root list l1 and set x(1,1) to the return value of the function. If l1 is now empty, replace it with the last list in the heap lk . If l1 is not empty, there is no need to swap it with lk . Either way, at this point l1 might be out of heap order. Restore heap order by calling heapify down. This is identical to the heapify operation of the binary heap except that it manipulates entire lists instead of individual nodes. To heapify down on a list li , compare li with its two children, l2i and l2i+1 . If li is less than both children, do nothing. Otherwise, swap li with its smallest child and repeat. 3.2.3 Decrease Key To perform decrease key on a node x(i,j) , set x(i,j) ’s key to the new value. Now there are two cases. 1. If x(i,j) is a head node, that is if x(i,j) = x(i,1) , decreasing x(i,1) ’s key maintains the sorted order of the list li but may destroy the heap order of the array. Call heapify up to restore the heap order. To heapify up a 6 list li , compare li to its parent l⌊i/2⌋ . If li < l⌊i/2⌋ swap the lists. Recurse until heap order is restored. 2. If x(i,j) is not a head node, li may no longer be sorted. If x(i,j) is less than its left sibling x(i,j−1) , then the list li is out of sorted order. Remove x(i,j) from li and reinsert x(i,j) into the List Heap using the insert routine. 3.2.4 Analysis It is clear from their descriptions that each of these functions run in O(log k) time, where k is the number of lists in the RA List Heap. Theorem 1. A sequence of consecutive insert operations partitions the inserted elements into L lists where L is less than or equal to the minimum number of runs in the input sequence. Proof. Say the consecutive sequence of elements inserted into heap h = hl1 , ...lk i is X = hx1 , ..., xi , xi+1 , ...i. We must show that if xi > xi+1 , then xi+1 is appended to an existing list. There are a couple of cases to consider. 1. If xi < lk , xi is appended to lk or one of lk ’s ancestors. Now, xi+1 < xi so xi+1 < lk and xi+1 is also appended to lk or one of lk ’s ancestors. 2. If xi > lk , xi is appended to lk+1 or one of its ancestors. If xi is appended to lk+1 , the new list lk+1 must be created. Now, xi+1 is compared first to lk+1 . Since xi+1 < xi , xi+1 < lk+1 , thus xi+1 is appended to lk+1 or one of lk+1 ’s ancestors. If instead xi was appended to an ancestor of lk+1 , lk+1 was not created. Then xi+1 is compared first to lk . If xi+1 > lk , xi+1 is compared to the ancestors of lk+1 . Since xi was appended to an ancestor of lk+1 , xi+1 must be less than one of the ancestors, so xi+1 is appended to an existing list. 3.3 An Enc Adaptive List Heap In this section we present operations for a List Heap that is adaptive with respect to runs, SU S, and Enc. We will abbreviate this EA List Heap. SU S is due to Levcopoulos and Petersson [10], Enc was proposed by Skiena [12]. Definition 2. SU S (Shuffled UpSubsequences) is the minimum number of increasing subsequences into which a sequence X can be partitioned. This differs from runs in that the subsequences are increasing and they are not required to be consecutive. For example, the sequence (1) can be partitioned into the following seven increasing subsequences, h3, 15, 16i, h14i, h4, 9, 13i, h5, 12i, h10, 11i, h6, 8i, h1, 2, 7i. (3) This is, in fact, an optimal partition, so SU S for the sequence (1) is seven. 7 Definition 3. An encroaching set is an ordered set of increasing sequences hE1 , ..., Ek i such that the head(Ei ) < head(Ei+1 ) and tail(Ei ) > tail(Ei+1 ) for 1 ≤ i < k. Thus, the increasing sequences nest or encroach upon one another. Skiena describes an encroaching set by an algorithm, melsort, that builds them. Given an input sequence Y = hy1 , ..., yn i that is a permutation of an ordered set, build the encroaching set E = hE1 , ..., Ek i as follows. If the item yi fits on either end of one of the increasing sequences Ej , put it there. Otherwise, form a new sequence Ek+1 . Place each item on the oldest Ej upon which it fits. We note that this strategy bears a lot of similarities to Patience Sorting, see [1]. The encroaching set for our example sequence (1) is, h1, 3, 15, 16i, h2, 4, 14i, h5, 9, 13i, h6, 10, 12i, h7, 8, 11i. (4) Thus, Enc for the sequence (1) is five. For adaptive sorting, an Enc optimal algorithm is also SU S optimal and runs optimal [11]. Therefore, we focus on developing a heap which is adaptive to Enc, thereby also achieving adaptivity with respect to SU S and runs. Again, we emphasize that this paper treats the notation of adaptivity informally. In order to achieve some level of adaptivity with respect to Enc, we only need to alter our insert function. The decrease key and delete min functions can remain the same as above. 3.3.1 Insert The insert function attempts to build an encroaching set. We adapt Skiena’s strategy to the dynamic context of heaps where there is not a fixed input sequence. Let y be a new item to insert and h = hl1 , ..., lk i be the heap. In this context, lj is roughly equal to Ej of the encroaching list set. There are three cases. 1. Try to insert y at the front of a list in h. Use binary search on the lists to find the list that has y < li and minimizes \|y − li \|. Note, the lists are maintained in heap order, not sorted order. As a result, this binary search may only be an approximate heuristic. Moreover, the binary search may have failed to consider li ’s direct parent. Appending y to the front of li without checking li ’s parent could destroy heap order. Thus, we must also compare y to li ’s ancestors. If y < l⌊i/2⌋ , set i = ⌊i/2⌋ and repeat. If we find an li , we are done. Otherwise, we move on to Case 2. 2. Try to insert y at the tail of a list in h. Use binary search on the tail nodes to find the list li that has tail(li ) < y and minimizes \|y − tail(li )\|. If no li is found, go to Case 3. 3. Create a new list lk+1 with the new node as its only element and insert the list after lk in the array. 8 3.3.2 Analysis It is easy to see that the insert routine described above runs in O(log k), where k is the number of lists in the EA List Heap. Theorem 2. A sequence of n consecutive inserts to an empty EA List Heap h partitions the input into an optimal encroaching set. Proof. Simply observe that the insert function reduces to the list creation phase of melsort if the heap starts empty. The only algorithmic difference between the insert routine above and Skiena’s melsort is the addition of the heap order check. Corollary 2.1. Given an empty EA List Heap h, perform n consecutive insert operations of elements with uniformly random keys. Then, as n grows large, the expected worst-case number of comparisons to delete the minimum item from h is less than or equal to half the the number of comparisons required by a binary heap. Proof. By Theorem 2, h = hl1 , ..., lk i where the lists li form an optimal encroaching set. As mentioned above, an Enc optimal partition is also SU √ S optimal, [11], so k is less than or equal to SU S. Now, as n√→ ∞, SU S → 2 n. [1] Thus, k approaches something less than or equal to 2 n. The worst-case number of comparisons for delete min on a standard binary heap is 2⌈log n − 1⌉. For the Enc Adaptive List Heap, the worst-case number of comparisons is less than or √ equal to 2⌈log k − 1⌉ ≤ 2⌈log (2 n) − 1⌉ = ⌈log n⌉. As n → ∞, the binary heap bound approaches 2⌈log n⌉. 4 Empirical Results We implemented both the RA List Heap and the EA List Heap for testing purposes, but made no attempts to optimize the code. The results of a series of brief empirical tests against a similarly unoptimized binary heap are presented below. The tests were performed using a codebase written in C and workloads from the 5th DIMACS challenge [6] (with modifications) and a simple sorting routine. The results presented below are raw wallclock times divided by the minimum time attained by any heap. Thus, 1.00 is the minimum wallclock time and k is k times the minimum. RA List Heap EA List Heap Binary Heap Sorted Sorting 1.00 1.15 12.62 Random Sorting 2.29 1.00 2.52 Random Dijkstra 1.00 1.10 1.22 Table 1: Normalized Wallclock Times Sorted Sorting refers to the task of inserting an n-element decreasing sorted sequence into the heap, followed by n consecutive calls to delete min. Random 9 sorting refers to the task of inserting an n-element random sequence into the heap, followed by n consecutive calls to delete min. The results presented for sorting are for n = 10 million. Random Dijkstra refers to Dijkstra workloads from DIMACS. The results presented for Dijkstra are from a strongly connected randomly generated network with 2 million nodes and 8 million edges. These test results suggest that some variant of the List Heap may be useful in practice. We stress that the results presented above were derived from very simplistic testing. They are merely suggestive that List Heaps may have potential in practice. They are in no way the final word on the empirical performance of the List Heap. 5 Conclusion This paper introduced adaptive heaps - heaps whose performance is a function of both the size of the problem instance and the disorder of the problem instance. We introduced the List Heap as a generic structure that can be endowed with adaptive operations. Finally, we presented operations for the List Heap that are adaptive with respect to a number of measures of disorder on the input sequence. The discussion within this paper has been relatively informal and the List Heap introduced is far from theoretically optimal. If there is interest in this topic, there are a number of directions from here. Additional empirical testing of the List Heap is needed. We presented two insert functions for the List Heap, but there are clearly more (d-array, increasing/decreasing runs, etc.). The optimal choice likely depends on the intended application. If decrease key is not needed, List Heaps could be implemented entirely with arrays. On the theory side, the List Heap and discussion of adaptivity included in this paper leaves much to be desired. We are working on formalizing some notions of adaptivity and have developed a variant of the Fibonacci Heap that is closer to optimal from a theoretical perspective. Modification to other existing heap variants may provide even better results. References [1] David Aldous and Persi Diaconis. Longest increasing subsequences: From patience sorting to the baik-deift-johansson theorem. Bull. Amer. Math. Soc, 36:413–432, 1999. [2] Gerth Stølting Brodal. Worst-case efficient priority queues. In Proceedings of the Seventh Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’96, pages 52–58, Philadelphia, PA, USA, 1996. Society for Industrial and Applied Mathematics. [3] Gerth Stølting Brodal. A Survey on Priority Queues, pages 150–163. Springer Berlin Heidelberg, Berlin, Heidelberg, 2013. 10 [4] Gerth Stølting Brodal, George Lagogiannis, and Robert E. Tarjan. Strict fibonacci heaps. In Proceedings of the Forty-fourth Annual ACM Symposium on Theory of Computing, STOC ’12, pages 1177–1184, New York, NY, USA, 2012. ACM. [5] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Introduction to Algorithms, Third Edition. The MIT Press, 3rd edition, 2009. [6] DIMACS. Dimacs 5th challenge: https://app.cs.amherst.edu/ ccmcgeoch/challenge5/. Priority queues. [7] Vladmir Estivill-Castro and Derick Wood. A survey of adaptive sorting algorithms. ACM Comput. Surv., 24(4):441–476, December 1992. [8] Michael L. Fredman and Robert Endre Tarjan. Fibonacci heaps and their uses in improved network optimization algorithms. J. ACM, 34(3):596–615, July 1987. [9] Daniel H. Larkin, Siddhartha Sen, and Robert E. Tarjan. A back-to-basics empirical study of priority queues. In Proceedings of the Meeting on Algorithm Engineering & Expermiments, pages 61–72, Philadelphia, PA, USA, 2014. Society for Industrial and Applied Mathematics. [10] C. Levcopoulos and O. Petersson. Sorting shuffled monotone sequences. Information and Computation, 112(1):37 – 50, 1994. [11] Ola Petersson and Alistair Moffat. A framework for adaptive sorting. Discrete Applied Mathematics, 59(2):153 – 179, 1995. [12] Steven S. Skiena. Encroaching lists as a measure of presortedness. BIT Numerical Mathematics, 28(4):775–784, Dec 1988. [13] Jean Vuillemin. A data structure for manipulating priority queues. Commun. ACM, 21(4):309–315, April 1978. [14] John William Joseph Williams. Algorithm 232 - heapsort. Commun. ACM, 7(6):347–349, June 1964. [15] Edsger Wybe. Dijkstra. A note on two problems in connexion with graphs. 1:269–271, 12 1959. 11	8cs.DS
Balancing Explicability and Explanations Emergent Behaviors in Human-Aware Planning Tathagata Chakraborti, Sarath Sreedharan, Subbarao Kambhampati∗ School of Computing, Informatics, and Decision Systems Engineering Arizona State University, Tempe, AZ 85281 USA tchakra2,ssreedh3,rao@asu.edu arXiv:1708.00543v2 [cs.AI] 3 Feb 2018 ABSTRACT Human aware planning requires an agent to be aware of the intentions, capabilities and mental model of the human in the loop during its decision process. This can involve generating plans that are explicable to a human observer as well as the ability to provide explanations when such plans cannot be generated. In this paper, we bring these two concepts together and show how an agent can account for both these needs and achieve a trade-off during the plan generation process itself by means of a model-space search method MEGA. This in effect provides a comprehensive perspective of what it means for a decision-making agent to be “human-aware” by bringing together existing principles of planning under the umbrella of a single plan generation process. We situate our discussion in the context of recent work on explicable planning and explanation generation, and illustrate these concepts in modified versions of two well-known planning domains, as well as in a demonstration of a robot involved in a typical search and reconnaissance task with an external supervisor. Human factor studies in the latter highlight the usefulness of the proposed approaches. CCS CONCEPTS • Computing methodologies → Artificial intelligence; Planning and scheduling; Cognitive robotics; • Human-centered computing; • Computer systems organization → Robotics; KEYWORDS Human-Aware Planning, Explicable Planning, Plan Explanations, Explanation as Model Reconciliation, Minimal Explanations. 1 INTRODUCTION Mh first two authors contributed equally. Mh the robot instead attempts to update the human’s mental model to cR in which the robot’s original plan is an intermediate model M h equivalent (with respect to a metric such as cost or similarity) to ∗ ∗ the optimal and hence explicable, i.e. π M R ≡ π cR . Mh Until now, these two processes of plan explanations and explicability have remained separate in so far as their role in an agent’s deliberative process is considered - i.e. a planner either generates an explicable plan to the best of its ability or it produces explanations of its plans where they required. However, there may be situations where a combination of both provide a much better course of action – if the expected human plan is too costly in the planner’s model (e.g. the human might not be aware of some safety constraints) or the cost of communication overhead for explanations is too high (e.g. limited communication bandwidth). Consider, for example, a human working with a robot that has just received a software update allowing it to perform new complex maneuvers. Instead of directly trying to conceive all sorts of new interactions right away that might end up spooking the user, the robot could instead reveal only certain parts of the new model while still using its older model (even though suboptimal) for the rest of the interactions so as to slowly reconcile the drifted model of the user. This is the focus of the current paper where we try to attain the sweet spot between plan explanations and explicability. 1.1 It is often useful for a planning agent while interacting with a human in the loop to use, in the process of its deliberation, not only the model M R of the task it has on its own, but also the model MhR that the human thinks it has (refer to Figure 1a). This mental model of the human [9] is in addition to the physical model of the human. This is, in essence, the fundamental thesis of the recent works on plan explanations [10] and explicable planning [35], summarized under the umbrella of multi-model planning, and is in addition to the originally studied human-aware planning (HAP) problems where actions of the human (and hence the actual human model and the robot’s belief of it) are also involved in the planning process. The need for explicable planning or plan explanations in fact occur when these two models – M R and MhR – diverge. This means that ∗ ∗ the optimal plans in the respective models – π M R and π R – may ∗ The not be the same and hence optimal behavior of the robot in its own model is inexplicable to the human in the loop. In the explicable planning process, the robot produces a plan πb that is closer to the human’s expected plan, i.e. πb ≈ π ∗ R . In the explanation process, Related Work As AI agents become pervasive in our daily lives, the need for such agents to be cognizant of the beliefs and expectations of the humans in their environment has been well documented [21]. From the perspective of task planning, depending on the extent of involvement of the human in the life cycle of a plan, work in this direction has ranged on a spectrum of “human-aware planning” [2, 3, 8, 11–13, 22, 33] where a robot passively tries to account for the plans of humans cohabiting its workspace, to “explicable planning” [14, 23, 34, 35] where a robot generates plans that are explicable or predictable to a human observer, to “plan explanations” [10, 17, 24] where the agent uses explanations to bring the human (who may have a different understanding of the agent’s abilities) on to the same page, to “human-in-the-loop planning” [1, 4, 16, 25, 30] in general where humans and planners are participating in the plan generation and/or execution process together. Technical Report, 2018, ASU Tathagata Chakraborti, Sarath Sreedharan, Subbarao Kambhampati (a) The evolving scope of Human-Aware Planning (HAP) (b) A Subsumption Architecture for HAP Figure 1: The expanding scope of human-aware planning (HAP) acknowledging the need to account for the mental model of the human in the loop in the deliberative process of an autonomous agent. The planner can, for example, choose to bring the human’s model closer to the ground truth using explanations via a process called model reconciliation (MRP) so that an otherwise inexplicable plan makes sense in the human’s updated model or it can compute explicable plans which are closer to the human’s expectation. These capabilities can be stacked to realize more and more complex behavior – in this paper we will concentrate on the explicability versus explanation trade-off as a form of argumentation during human-aware planning. 1.1.1 The Evolving Scope of Human-Aware Planning. The ongoing efforts to make planning more “human-aware” is illustrated in Figure 1a – initial work on this topic had largely focused on incorporating an agent’s understanding of the human model M H into its decision making process. Since then the importance of considering the human’s understanding MhR of the agent’s actual model M R in the planning process has also been acknowledged, sometimes implicitly [3] and later explicitly [10, 35]. These considerations engender interesting behaviors both in the space of plans and models. For example, in the model space, the modifications to the human mental model MhR is used for explanations in [10] while reasoning over the actual model M H can reveal interesting behavior by affecting the belief state of the human, such as in planning for serendipity [8]. In the plan space, a human-aware agent can use M H and MhR to compute joint plans for teamwork [32] or generate behavior that conforms to the human’s preferences [11, 22] and expectations [23, 34, 35]. From the point of view of the planner, this is, in a sense, an asymmetric epistemic setting with single-level nested beliefs over its models. Indeed, existing literature on epistemic reasoning [19, 27, 28] can also provide interesting insights in the planning process of an agent in these settings. 1.1.2 A Subsumption Architecture for HAP. These different forms of behavior can be composed to form more and more sophisticated forms of human-aware behavior. This hierarchical composition of behaviors can be viewed in the form of a subsumption architecture for human-aware planning, similar in motivation to [7]. This is illustrated in Figure 1b. The basic reasoning engines are the Plan and MRP (Model Reconciliation) modules. The former accepts model(s) of planning problems and produces a plan, the latter accepts the same and an produces a new model. The former operates in plan space and gives rise to classical, joint and explicable planning depending on the models it is operating on, while the latter operates in model space to produce explanations and belief shaping behavior. These are then composed to form argumentation modules for trading of explanations and explicability (which is the topic of the current paper) and human-aware planning in general. 1.1.3 The Explicability-Explanation Trade-off. From the perspective of design of autonomy, this trade-off has two important implications – (1) the agent can now not only explain but also plan in the multi-model setting with the trade-off between compromise on its optimality and possible explanations in mind; and (2) the argumentation process is known to be a crucial function of the reasoning capabilities of humans [26], and now by extension of autonomous agents as well, as a result of algorithms we develop here to incorporate the explanation generation process into the agent’s decision making process itself. General argumentation frameworks Balancing Explicability and Explanations for resolving disputes over plans have indeed been explored before [6, 15]. Our work can be seen as the specific case where the argumentation process is over a set of constraints that prove the correctness and quality of plans by considering the cost of the argument specifically as it relate to the trade-off in plan quality and the cost of explaining that plan. This is the first of its kind algorithm that can achieve this in the scope of plan explanations and explicable in presence of model differences with the human. HUMAN-AWARE PLANNING REVISITED The problem formulation closely follows that introduced in [10], reproduced here for clarity of methods built on the same definitions. A Classical Planning Problem. is a tuple M = ⟨D, I, G⟩ 1 with domain D = ⟨F , A⟩ - where F is a set of fluents that define a state s ⊆ F , and A is a set of actions - and initial and goal states I, G ⊆ F . Action a ∈ A is a tuple ⟨c a , pre(a), eff± (a)⟩ where c a is the cost, and pre(a), eff± (a) ⊆ F are the preconditions and add/delete effects, i.e. δ M (s, a) \|= ⊥ if s ̸ \|= pre(a); else δ M (s, a) \|= s ∪ eff+ (a) \ eff− (a) where δ M (·) is the transition function. The cumulative transition function is δ M (s, ⟨a 1 , a 2 , . . . , an ⟩) = δ M (δ M (s, a 1 ), ⟨a 2 , . . . , an ⟩). The solution to the planning problem is a sequence of actions or a (satisficing) plan π = ⟨a 1 , a 2 , . . . , an ⟩ such that δ M (I, π ) \|= G. Í The cost of a plan π is C(π , M) = a ∈π c a if δ M (I, π ) \|= G; ∞ otherwise. The cheapest plan π ∗ = arg minπ C(π , M) is the (cost) optimal plan. We refer to this cost as M as C ∗M . A Human-Aware Planning (HAP) Problem. is given by the tuple Ψ = ⟨M R , M H , MhR ⟩ where M R = ⟨D R , I R , G R ⟩ is the planner’s model of a planning problem, while M H = ⟨D H , I H , G H ⟩ and MhR = ⟨DhR , IhR , GhR ⟩ are respectively the planner’s estimate of the human’s model and the human’s understanding of its own model. The solution to the human-aware planning problem is a joint plan [8] π = ⟨a 1 , a 2 , . . . , an ⟩; ai ∈ {D R ∪ D rH } such that δ Ψ (I R ∪ IrH , π ) \|= G R ∪ GrH . The robot’s component in the plan is referred to as π (R) = ⟨ai \| ai ∈ π ∧ D R ⟩. For the purposes of this paper, we ignore the robot’s belief of the human model, i.e. MrH = ⟨{}, {}, {}⟩ – in effect, making the human an observer only or a passive consumer of the plan – and focus instead on the challenges involves in planning with the human’s model of the planner. Planning with the human model has indeed been studied extensively in the literature, as noted above, and this assumption does not change in any way the relevance of the work here. Specifically, the following concepts are built on top of the joint planning problem – e.g. an explicable plan in this paper would, in the general sense, correspond to the robot’s component in the joint plan being explicable to the human in the loop. Thus, for the purposes of this paper, we have π (R) ≡ π ; without loss of generality, we focus on the simplified setting with only the model of the planner and the human’s approximation of it. Explicable Planning In “explicable planning" a solution to the human-aware planning problem is a plan π such that (1) it is executable (but may no longer Technical Report, 2018, ASU be optimal) in the robot’s model but is (2) “closer” to the expected plan in the human’s model, given a particular planning problem – (1) δ M R (I R , π ) \|= G R ; and (2) C(π , MhR ) ≈ C ∗ R . Mh “Closeness” or distance to the expected plan is modeled here in terms of cost optimality, but in general this can be any preference metric like plan similarity. In existing literature [23, 34, 35] this has been usually achieved by modifying the search process so that the heuristic that guides the search is driven by the robot’s knowledge of the human’s mental model. Such a heuristic can be either derived directly [23] from the human’s model (if it is known) or learned [35] through interactions in the form of affinity functions between plans and their purported goals. Plan Explanations The other approach would be to (1) compute optimal plans in the planner’s model as usual, but also provide an explanation (2) in the form of a model update to the human so that (3) the same plan is now also optimal in the human’s updated model of the problem. Thus, a solution involves a plan π and an explanation E such that – (1) C(π , M R ) = C ∗M R ; cR ←− M R + E; and (2) M h Mh Note that here a model update, as indicated by the + operator may include a correction to the belief (goals or state information) as well as information pertaining to the action model itself. In [10] the authors explored various ways of generating such solutions – including methods to minimize the lengths of the explanations given as a result. However, this was done in an after-the-fact fashion, i.e. the optimal plan was already generated and it was just a matter of finding the best explanation for it. This not only ignores the possibility of finding better plans (that are equally optimal) with smaller explanations, but also avenues of compromise in a manner we discussed previously whereby the planner sacrifices its optimality to further reduce overhead in the explanation process. MEGA We bring the notions of explicability and explanations together in a novel planning technique MEGA (Multi-model Explanation Generation Algorithm) that trades off the relative cost of explicability to providing explanations during the plan generation process itself2 . The output of MEGA is a plan π and an explanation E such that (1) π is executable in the robot’s model, and with the explanation (2) in the form of model updates it is (3) optimal in the updated human model while (4) the cost (length) of the explanations, and the cost of deviation from optimality in its own model to be explicable to the human, is traded off according to a constant α – (1) δ M R (I R , π ) \|= G R ; cR ←− M R + E; (2) M h h cR ) = C ∗ ; and (3) C(π , M h 2 As 1 Note that the “model of a planning problem” includes the action model as well as the initial and goal states of an agent. h cR ) = C ∗ . (3) C(π , M h cR cR M h in [10] we assume that the human mental model is known and has the same computation power ([10] also suggests possible ways to address these issues, the same discussions apply here as well). Also refer to the discussion on model learning later. Technical Report, 2018, ASU Tathagata Chakraborti, Sarath Sreedharan, Subbarao Kambhampati (4) π = arg minπ { \|E \| + α × \| C(π , M R ) − C ∗M R \| }. Clearly, with higher values of α the planner will produce plans that require more explanation, with lower α it will generate more explicable plans. Thus, with the help of this hyperparameter α, an autonomous agent can deliberate over the trade-off in the costs it incurs in being explicable to the human (second minimizing term in (4)) versus explaining its decisions (first minimizing term in (4)). Note that this trade-off is irrespective of the cognitive burden of those decisions on the human in the loop. For example, for a robot in a collapsed building during a search and rescue task, or the rover on Mars, may have limited bandwidth for communication and hence prefer to be explicable instead instead. We employ a model space A∗ search (Algorithm 1) to compute the expected plan and explanations for a given value of α. Similar to [10] we define a state representation over planning problems with a mapping function Γ : a M 7→ F which represents a planning problem by transforming every condition in it into a predicate. The set Λ of actions contains unit model change actions λ : F → F which make a single change to a domain at a time. We start by initializing the min node tuple (N ) with the human mental model and an empty explanation. For each new possible model we come across during our model space search, we test if the objective value of the new node is smaller than the current min node. We stop the search once we identify a model that is capable of producing a plan that is also optimal in the robot’s own model. This is different from the stopping condition used by the original MCE-search3 in [10], where the authors are just trying to identify the first node where the given plan is optimal. Algorithm 1 MEGA 1: 2: 3: 4: 5: 6: 7: 8: 9: Mh This is beyond what is offered by the model reconciliation search in [10], which only computes the smallest explanation given as a plan that is optimal in the planner’s model. Property 2. α = \| M R ∆ MhR \| yields the most optimal plan in the planner’s model along with the minimal explanation possible given a human-aware planning problem. This is easy to see, since with ∀E, \|E \| ≤ \| M R ∆ MhR \|, the latter being the total model difference, the penalty for departure from explicable plans is high enough that the planner must choose from possible explanations only (note that the explicability penalty is cR ) = always positive until the search hits the nodes with C(π , M h ∗ C cR , at which point onwards the penalty is exactly zero). In general Mh this works for any α ≥ \|MCE\| but since an MCE will only be known retrospectively after the search is complete, the above condition suffices since the entire model difference is known up front and is the largest possible explanation in the worst case. 3 An MCE or a minimally complete explanation is the shortest model update so that a given plan optimal in the robot model is also optimal in the updated human model. ▷ Closed list ▷ Node with minimum objective value ▷ Optimal plan being explained ▷ Plan expected by human Nmin ← ⟨MhR , { }⟩ π R∗ ← π ∗ πhR ← π s.t. C(π, MhR ) = C ∗ R M h 10: fringe.push(⟨M R , { }⟩, priority = 0) 11: while True do h 12: c E ⟩, c ← fringe.pop( M) c ⟨ M, 13: c E ⟩) ≤ OBJ_VAL(Nmin ) then if OBJ_VAL(⟨ M, 14: 15: 16: 17: 18: 19: 20: 21: 22: ▷ Update min node c E⟩ Nmin ← ⟨ M, if C(π ∗ , M R ) = C ∗ R then c M M ⟨Mmin , Emin ⟩ ← Nmin , Emin ⟩ min ▷ If π ∗c is optimal in M R return ⟨π M M else c c_list ← c_list ∪ M c \ Γ(M R ) do for f ∈ Γ( M) c}, { }, {f }⟩ λ ← ⟨1, { M c λ) < c_list then if δ R (Γ( M), M , MR ▷ Models that satisfy condition 1 c ▷ Removes f from M h 23: fringe.push(⟨δ 24: c do for f ∈ Γ(M R ) \ Γ( M) 25: 26: MR , MR h c λ), E ∪ λ ⟩, c + 1) (Γ( M), c}, {f }, { }⟩ λ ← ⟨1, { M c λ) < c_list then if δ R (Γ( M), M , MR ▷ Models that satisfy condition 2 c ▷ Adds f to M h 27: 28: 29: Property 1. MEGA yields the smallest possible explanation for a given human-aware planning problem. This means that with a high enough α (see below) the algorithm is guaranteed to compute the best possible plan for the planner as well as the smallest explanation associated with it. This is by construction of the search process itself, i.e. the search only terminates after cR ) = C ∗ have been exhausted. the all the nodes that allow C(π , M h cR procedure MEGA-Search Input: HAP Ψ = ⟨M R , M R ⟩ , α h Output: Plan π and Explanation E Procedure: fringe ← Priority_Queue() c_list ← {} fringe.push(⟨δ MR , MR h c λ), E ∪ λ ⟩, c + 1) (Γ( M), c E ⟩) procedure OBJ_VAL( ⟨ M, return \| E \| + α × \| C(π ∗ , M R ) − C ∗ R \| c M M Property 3. α = 0 yields the most explicable plan. Under this condition, the planner has to minimize the cost of explanations only. Of course, at this point it will produce the plan that requires the shortest explanation, and hence the most explicable plan. Note that this is distinct from just computing the optimal plan in the human’s model, since such a plan may not be executable in the planner’s model so that some explanations are required even in the worst case. This is also a welcome additions to the explicability only view of plan generation introduced in [23, 34, 35], where the human model only also guides the plan generation process instead of doing so directly, though none of these works provided any insight into how to make the remainder of the model reconciliation possible in such cases, as done here with the explanations associated with the generated plans. Property 4. MEGA-search is required only once per problem, and is independent of α. Algorithm 1 terminates only after all the nodes containing a minimally complete explanation3 have been explored. This means that for different values of α, the agent only needs to post-process the nodes with the new objective function in mind. Thus, a large part of the reasoning process for a particular problem can be pre-computed. 2 EVALUATIONS We will now provide internal evaluations of MEGA in modified versions of two well-known IPC domains Rover and Barman [20] Balancing Explicability and Explanations Domain Name Problem ∆=2 Technical Report, 2018, ASU ∆=7 ∆ = 10 \|E \| Time (secs) \|E \| Time (secs) \|E \| Time (secs) Rover p1 p2 p3 0 1 0 1.22 1.79 8.35 1 5 2 5.83 125.64 10.46 3 6 3 143.84 1061.82 53.22 Barman p1 p2 p3 2 2 2 18.70 2.43 45.32 6 4 5 163.94 57.83 4183.55 6 6 6 5576.06 953.47 5061.50 Table 1: Computation time for human-aware plans in Rover and Barman domains along with the length of explanations. demonstrating the trade-off in the cost and computation time of plans with respect to varying size of the model difference and the hyper-parameter α and follow it up with a demonstration of MEGA in action on a robot in a search and reconnaissance domain. Finally, we will report on human factor studies on how this trade-off is received by users. The code and the domain models will be available after the double-blind review process is over. 2.1 (a) The Rover (Meets a Martian) Domain Empirical Results: Cost Trade-off The value of α determines how much an agent is willing to sacrifice its own optimality versus the cost of explaining a (perceived) suboptimal plan to the human. In the following, we illustrate this trade-off on modified versions of two well-known IPC domains. 2.1.1 The Rover (Meets a Martian) Domain. Here the Mars Rover has a model as described in the IPC domain, but has gone an update whereby it can carry all the rock and soil samples needed for a mission at the same time. This means that it does not need to empty the store before collecting new rock and soil samples anymore so that the new action definitions for sample_soil and sample_rock no longer contain the precondition (empty ?s). During its mission it runs across a Martian who is unaware of the robot’s expanded storage capacity, and has an older, extremely cautious, model of the rover it has learned while spying on it from its cave. It believes that any time we collect a rock sample, we also need to collect a soil sample and need to communicate this information to the lander. It also believes that before the rover can perform take_image action, it needs to send the soil data and rock data of the waypoint from where it is taking the image. Clearly, if the rover was to follow this model in order not to spook the Martians, it will end up spending a lot of time performing unnecessary actions (like dropping old samples and collecting unnecessary samples). For example, if the rover is to communicate an image of an objective objective2, all it needs to do is move to a waypoint (waypoint3) from where objective2 is visible and perform the action – (take_image waypoint3 objective2 camera0 high_res) If the rover was to produce a plan that better represents the Martian’s expectations, it would look like – (sample_soil store waypoint3) (communicate_soil_data general waypoint3 waypoint3 waypoint0) (drop_off store) (sample_rock store waypoint3) (communicate_rock_data general waypoint3 waypoint3 waypoint0) (take_image waypoint3 objective1 camera0 high_res) (b) The Barman (in a Bar) Domain Figure 2: Trade-off between explicability versus explanation cost for plans produced at different values of α. Now if the rover chose to directly use an MCE it could end up explaining up to six different model differences based on the problem and the plan under execution. In some case, this may be acceptable, but in others, it may make more sense for the rover to bear the extra cost rather than laboriously walking through all the updates with an impatient Martian. MEGA lets us naturally model these scenarios through the use of the α parameter – the rover would choose to execute the Martian’s expected optimal plan when the α parameter is set to zero (which means the rover does not care about the extra cost it needs to incur to ensure that the plan makes sense to the Martian with the least explaining involved). Figure 2 shows how the explicability cost and explanation cost varies for three typical problem instances in this domain. The algorithm starts converging to the smallest possible MCE, when α is set to one. For smaller α, MEGA chooses to save explanation costs by choosing more expensive (and explicable) plans. 2.1.2 The Barman (in a Bar) Domain. Here, the brand new twohanded Barman robot is wowing onlookers with its single-handed Technical Report, 2018, ASU Tathagata Chakraborti, Sarath Sreedharan, Subbarao Kambhampati skills, even as its admirers who may be unsure of its capabilities expect, much like in the original IPC domain, that it is required to have one hand free to perform actions like fill-shot, refill-shot, shake etc. This means that to make a single shot of a cocktail with two shots of the same ingredient with three shots and one shaker, the human expects the robot to execute the following plan – (fill-shot shot2 ingredient2 left right dispenser2) (pour-shot-to-used-shaker shot2 ingredient3 shaker1 left l1 l2) (refill-shot shot2 ingredient3 left right dispenser3) (pour-shot-to-used-shaker shot2 ingredient3 shaker1 left l1 l2) (leave left shot2) (grasp left shaker1) The robot can, however, directly start by picking both the shot and the shaker and does not need to put either of them down while making the cocktail. Similar to the Rover domain, we again illustrate on three typical problems from the barman domain (Figure 2) how at lower values of α the robot choose to perform plans that require less explanation. As α increases the algorithm produces plans that require larger explanations with the explanations finally converging at the smallest MCE required for that problem. 2.2 Empirical Results: Computation Time Contrary to classical notions of planning that occurs in state or plan space, we are now planning in the model space, i.e. every node in the search tree is a new planning problem. As seen in Table 1 this becomes quite time consuming with increasing number of model differences between the human and the robot, even as there are significant gains to be had in terms of minimality of explanations, and the reduction in cost of explicable plans as a result of it. This motivates the need for developing approximations and heuristics [10] for the search for multi-model explanations. 2.3 Demonstration: The USAR Domain We first demonstrate MEGA on a robot performing an Urban Search And Reconnaissance (USAR) task – here a remote robot is put into disaster response operation often controlled partly or fully by an external human commander. This is a typical USAR setup [5], where the robot’s job is to infiltrate areas that may be otherwise harmful to humans, and report on its surroundings as and when required / instructed by the external. The external usually has a map of the environment, but this map is no longer accurate in a disaster setting - e.g. new paths may have opened up, or older paths may no longer be available, due to rubble from collapsed structures like walls and doors. The robot (internal) however may not need to inform the external of all these changes so as not to cause information overload of the commander who may be otherwise engaged in orchestrating the entire operation. This calls for an instantiation of the MEGA algorithm where the model differences are contributed to by changes in the map, i.e. the initial state of the planning problem (the human model has the original unaffected model of the world). Figure 3 shows a relevant section of the map of the environment where this whole scenario plays out. The orange marks indicate rubble that has blocked a passage, while the green marks indicate collapsed walls. The robot (Fetch), currently located at the position marked with a blue O, is tasked with taking a picture at location marked with an orange O in the figure. The external commander’s Figure 3: A typical search and reconnaissance scenario with an internal semi-autonomous agent (robot) and an external supervisor (human) – a video demonstration can be accessed at https://www.youtube.com/watch?v=u_t1TQotzo4. expects the robot to take the path shown in red, which is no longer possible. The robot armed with MEGA has two choices – it can either follow the green path and explain the revealed passageway due to the collapse, or compromise on its optimal path, clear the rubble and proceed along the blue path. A video demonstration of the scenario can be viewed at https:// www.youtube.com/ watch?v=u_t1TQotzo4. The first part of the video demonstrates the plan generated by MEGA for low α values. As expected, it chooses the blue path that requires the least amount of explanation, and is thus the most explicable plan. In fact, the robot only needs to explain a single initial state change to make this plan optimal, namely – Explanation >> remove-has-initial-state-clear_path p1 p8 This is also an instance where the plan closest to the human expectation, i.e. the most explicable plan, still requires an explanation, Balancing Explicability and Explanations Technical Report, 2018, ASU which previous approaches in the literature cannot provide. Moreover, in order to follow this plan, the robot must perform the costly clear_passage p2 p3 action to traverse the corridor between p2 and p3, which it could have avoided in its optimal plan (shown in green on the map). Indeed, MEGA switches to the robot’s optimal plan for higher values of α along with the following explanation – Explanation >> add-has-initial-state-clear_path p6 p7 Explanation >> add-has-initial-state-clear_path p7 p5 Explanation >> remove-has-initial-state-clear_path p1 p8 By providing this explanation, the robot is able to convey to the human the optimality of the current plan as well as the infeasibility of the human’s expected plan (shown in red). 2.4 Human Factors Evaluations Finally, we will now use the above search and reconnaissance domain to analyze how humans respond to the explicability versus explanations trade-off. This is done by exposing the external commander’s interface to participants who get to analyze plans in a mock USAR scenario. The participants were incentivized to make sure that the explanation does indeed help them understand the optimality of the plans in question by formulating the interaction in the form of a game. This is to make sure that participants were sufficiently invested in the outcome as well as mimic the high-stakes nature of USAR settings to accurately evaluate the explanations. Figure 4 shows a screenshot of the interface which displays to each participant an initial map (which they are told may differ from the robot’s actual map), the starting point and the goal. A plan is illustrated in the form of a series of paths through various waypoints highlighted on the map. The participant has to identify if the plan shown is optimal. If the player is unsure, they can ask for an explanation. The explanation is provided to the participant in the form of a set of model changes in the player’s map. The scoring scheme for the game is as follows. Each player is awarded 50 points for correctly identifying the plan as either optimal or satisficing. Incorrectly identification costs them 20 points. Every request for explanation further costs them 5 points, while skipping a map does not result in any penalty. The participants were additionally told that selecting an inexecutable plan as either feasible or optimal would result in a penalty of 400 points. Even though there were no actual incorrect plans in the dataset, this information was provided to deter participants from taking chances with plans they did not understand well. Each participant was paid $10 dollars and received additional bonuses based on the following payment scheme – - Scores higher than or equal to 540 were paid $10. Scores higher than 540 and 440 were paid $7. Scores higher than 440 and 340 were paid $5. Scores higher than 340 and 240 were paid $3. Scores below 240 received no bonuses. The scoring systems for the game was designed to make sure – • Participants should only ask for an explanation when they are unsure about the quality of the plan (due to small negative points on explanations). Figure 4: Interface to the external commander in a mock search and reconnaissance study. Figure 5: Responses to explicable plans versus balanced or robot optimal plans with explanations. Figure 6: Click-through rates for explanations. • Participants are incentivized to identify the feasibility and optimality of the given plan correctly (large reward and penalty on doing this wrongly). Each participant was shown a total of 12 maps. For 6 of the 12 maps, the participant was assigned the optimal robot plan, and when they asked for an explanation, they were randomly shown different types of explanations as introduced in [10]. For the rest of the maps, in place of the robot’s optimal plan, participants could potentially be assigned a plan that is optimal in the human model (i.e. an explicable plan) with no explanation or somewhere in between (i.e. the balanced plan) with a shorter explanation. Note that out of the 6 maps, only 3 had both balanced plans as well as explicable plans, the other 3 either had a balanced plan or the optimal human plan. In total, we had 27 participants for the study, including 4 female and 22 male participants between the age range of 19-31 (1 participant did not reveal their demographic). Technical Report, 2018, ASU Optimal Plan \| E \| C(π, M R ) 2.5 5.5 Balanced Plan \| E \| C(π, M R ) 1 8.5 Tathagata Chakraborti, Sarath Sreedharan, Subbarao Kambhampati Explicable Plan \| E \| C(π, M R ) 16 Table 2: Statistics of explicability versus explanation tradeoff with respect to explanation length and plan cost. Figure 5 shows how people responded to the different kinds of explanations / plans. These results are from 382 problem instances that required explanations, and 25 and 40 instances that contained balanced and explicable plans respectively. From the perspective of the human, the balanced plan and the robot optimal plan do not make any difference since both of them appear suboptimal. This is evident from the fact that the click-through rate for explanations in these two conditions are similar. However, the rate of explanations is significantly less in case of explicable plans as desired. Table 2 shows the statistics of the explanations / plans. These results are from 124 problem instances that required minimal explanations as per [10], and 25 and 40 instances that contained balanced and explicable plans respectively, as before. As desired, the robot gains in length of explanations but loses out in cost of plans produced as it progresses along the spectrum of optimal to explicable plans. Thus, while Table 2 demonstrates the cost of explanation versus explicability trade-off from the robot’s point of view, Figure 5 shows how this trade-off is perceived from the human’s perspective. It is interesting to see that in Figure 5 about a third of the time participants still asked for explanations even when the plan was explicable, and thus optimal in their map. This is an artifact of the risk-averse behavior incentivized by the gamification of the explanation process and indicative of the cognitive burden on the humans who are not (cost) optimal planners. Thus, going forward, the objective function should incorporate the cost or difficulty of analyzing the plans and explanations from the point of view of the human in addition to the current costs in equation MEGA(4) and Table 2 modeled from the perspective of the robot model. Finally, in Figure 6, we show how the participants responded to inexplicable plans, in terms of their click-through rate on the explanation request button. Such information can be used to model the α parameter to situate the explicability versus explanation trade-off according to preferences of individual users. It is interesting to see that the distribution of participants (right inset) seem to be bimodal indicating that there are people who are particularly skewed towards risk-averse behavior and others who are not, rather than a normal distribution of response to the explanation-explicability trade-off. This further motivates the need for learning α interactively with the particular human in the loop. 3 DISCUSSION AND FUTURE WORK In the following section, we will elaborate on some of the exciting avenues of future research borne out of this work. 3.1 Model learning and picking the right α We assumed that the hyper-parameter α is set by the designer in determining how much to trade-off the costs of explicability versus explanations on the part of the autonomous agent. However, the design of α itself can be more adaptive and “human-aware” in the sense that the parameter can be learned in course of interactions with the human in the loop to determine what kind of plans are preferred (as seen in Figure 6) and how much information can be transmitted. This is also relevant in cases where the human mental model is not known precisely or if there is uncertainty towards what the new model is after an update or explanation. This is a topic of future work; existing literature on iterative model learning [18, 29] can provide useful guidance towards the same. Authors in [9] discuss a few useful representations for learning such models for the purposes of task planning at various levels of granularity. Note that search with uncertainty over a learned human (mental) model can often times be compiled to the same planning process as described in [31] by using annotated models, so the same techniques as introduced in this paper still apply. 3.2 Cost of explanations and cognitive load Currently, we only considered the cost of explanations and explicability from the point of view of the robot. However, there might be additional (cognitive) burden on the human – measured in terms of the complexity of interpreting an explanation and how far away the final plan is from the optimal plan in the human’s mental model. This again ties back to the assumptions on the cognitive abilities (i.e. optimality) of the human in the loop, and needs calibration [18, 29] based on repeated interactions (as seen in Figure 5). 4 CONCLUSION We saw how an agent can achieve human-aware behavior while at the same time keeping in mind the cost of departure from its own optimality which could otherwise have been explained away if given the opportunity. This raises several intriguing challenges in the plan generation process, most notably in finding better heuristics in speeding up the model space search process as well as dealing with model uncertainty and identifying the sweet spot of the algorithm in explicability-explanations trade-off. Indeed, the revised human-aware planning paradigm opens up exciting new avenues of research such as learning human mental models, providing explanations at different levels of abstractions, and so on. REFERENCES [1] Mitchell Ai-Chang, John Bresina, Len Charest, Adam Chase, JC-J Hsu, Ari Jonsson, Bob Kanefsky, Paul Morris, Kanna Rajan, Jeffrey Yglesias, et al. 2004. MAPGEN: Mixed-Initiative Planning and Scheduling for the Mars Exploration Rover Mission. IEEE Intelligent Systems (2004). [2] Rachid Alami, Aurélie Clodic, Vincent Montreuil, Emrah Akin Sisbot, and Raja Chatila. 2006. Toward Human-Aware Robot Task Planning. In AAAI Spring Symposium: To Boldly Go Where No Human-Robot Team Has Gone Before. [3] Rachid Alami, Mamoun Gharbi, Benjamin Vadant, Raphaël Lallement, and Adolfo Suarez. 2014. On human-aware task and motion planning abilities for a teammate robot. In Human-Robot Collaboration for Industrial Manufacturing Workshop, RSS. [4] James F Allen. 1994. Mixed initiative planning: Position paper. In ARPA/Rome Labs Planning Initiative Workshop. [5] Cade Earl Bartlett. 2015. Communication between Teammates in Urban Search and Rescue. Thesis (2015). [6] Alexandros Belesiotis, Michael Rovatsos, and Iyad Rahwan. 2010. Agreeing on plans through iterated disputes. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems: volume 1-Volume 1. International Foundation for Autonomous Agents and Multiagent Systems, 765–772. [7] Rodney Brooks. 1986. A robust layered control system for a mobile robot. IEEE journal on robotics and automation 2, 1 (1986), 14–23. [8] T. Chakraborti, G. Briggs, K. Talamadupula, Yu Zhang, M. Scheutz, D. Smith, and S. Kambhampati. 2015. Planning for serendipity. In IROS. Balancing Explicability and Explanations [9] Tathagata Chakraborti, Subbarao Kambhampati, Matthias Scheutz, and Yu Zhang. 2017. AI Challenges in Human-Robot Cognitive Teaming. arXiv preprint arXiv:1707.04775 (2017). [10] Tathagata Chakraborti, Sarath Sreedharan, Yu Zhang, and Subbarao Kambhampati. 2017. Plan Explanations as Model Reconciliation: Moving Beyond Explanation as Soliloquy. In IJCAI. [11] Tathagata Chakraborti, Yu Zhang, David Smith, and Subbarao Kambhampati. 2016. Planning with Resource Conflicts in Human-Robot Cohabitation. In AAMAS. [12] Marcello Cirillo. 2010. Planning in inhabited environments: human-aware task planning and activity recognition. Ph.D. Dissertation. Örebro university. [13] Marcello Cirillo, Lars Karlsson, and Alessandro Saffiotti. 2010. Human-aware Task Planning: An Application to Mobile Robots. ACM Transactions on Intelligent Systems and Technology (2010). [14] Anca Dragan, Kenton Lee, and Siddhartha Srinivasa. 2013. Legibility and Predictability of Robot Motion. In Human-Robot Interaction. [15] Chukwuemeka D Emele, Timothy J Norman, and Simon Parsons. 2011. Argumentation strategies for plan resourcing. In The 10th International Conference on Autonomous Agents and Multiagent Systems-Volume 3. International Foundation for Autonomous Agents and Multiagent Systems, 913–920. [16] George Ferguson, James F Allen, Bradford W Miller, et al. 1996. TRAINS-95: Towards a Mixed-Initiative Planning Assistant. In AIPS. [17] Maria Fox, Derek Long, and Daniele Magazzeni. 2017. Explainable Planning. In IJCAI XAI Workshop. [18] Dylan Hadfield-Menell, Stuart J Russell, Pieter Abbeel, and Anca Dragan. 2016. Cooperative inverse reinforcement learning. In NIPS. [19] Marc Hanheide, Moritz Göbelbecker, Graham S Horn, Andrzej Pronobis, Kristoffer Sjöö, Alper Aydemir, Patric Jensfelt, Charles Gretton, Richard Dearden, Miroslav Janicek, et al. 2015. Robot task planning and explanation in open and uncertain worlds. Artificial Intelligence (2015). [20] International Planning Competition. 2011. IPC Competition Domains. https: //goo.gl/i35bxc. (2011). [21] Subbarao Kambhampati and Kartik Talamadupula. 2014. Human-in-the-Loop Planning and Decision Support. http://rakaposhi.eas.asu.edu/hilp-tutorial/. (2014). [22] Uwe Koeckemann, Federico Pecora, and Lars Karlsson. 2014. Grandpa Hates Robots - Interaction Constraints for Planning in Inhabited Environments. In AAAI. [23] Anagha Kulkarni, Tathagata Chakraborti, Yantian Zha, Satya Gautam Vadlamudi, Yu Zhang, and Subbarao Kambhampati. 2016. Explicable Robot Planning as Minimizing Distance from Expected Behavior. CoRR abs/1611.05497 (2016). [24] Pat Langley, Ben Meadows, Mohan Sridharan, and Dongkyu Choi. 2017. Explainable Agency for Intelligent Autonomous Systems. In AAAI/IAAI. [25] Lydia Manikonda, Tathagata Chakraborti, Kartik Talamadupula, and Subbarao Kambhampati. 2017. Herding the Crowd: Using Automated Planning for Better Crowdsourced Planning. Journal of Human Computation (2017). [26] Hugo Mercier and Dan Sperber. 2010. Why Do Humans Reason? Arguments for an Argumentative Theory. Behavioral and Brain Sciences (2010). [27] Tim Miller, Jens Pfau, Liz Sonenberg, and Yoshihisa Kashima. 2017. Logics of common ground. Journal of Artificial Intelligence Research (2017). [28] Christian J Muise, Paolo Felli, Tim Miller, Adrian R Pearce, and Liz Sonenberg. 2016. Planning for a Single Agent in a Multi-Agent Environment Using FOND.. In IJCAI. [29] Stefanos Nikolaidis, Przemyslaw Lasota, Ramya Ramakrishnan, and Julie Shah. 2015. Improved human–robot team performance through cross-training, an approach inspired by human team training practices. International Journal of Robotics Research (2015). [30] Sailik Sengupta, Tathagata Chakraborti, Sarath Sreedharan, and Subbarao Kambhampati. 2017. RADAR - A Proactive Decision Support System for Human-inthe-Loop Planning. In AAAI Fall Symposium on Human-Agent Groups. [31] S. Sreedharan, T. Chakraborti, and S. Kambhampati. 2017. Explanations as Model Reconciliation - A Mutli-Agent Perspective. In AAAI Fall Symposium on HumanAgent Groups. [32] Kartik Talamadupula, Gordon Briggs, Tathagata Chakraborti, Matthias Scheutz, and Subbarao Kambhampati. 2014. Coordination in human-robot teams using mental modeling and plan recognition. In Intelligent Robots and Systems (IROS 2014), 2014 IEEE/RSJ International Conference on. IEEE, 2957–2962. [33] Stevan Tomic, Federico Pecora, and Alessandro Saffiotti. 2014. Too Cool for School??? Adding Social Constraints in Human Aware Planning. In Workshop on Cognitive Robotics (CogRob). [34] Yu Zhang, Sarath Sreedharan, Anagha Kulkarni, Tathagata Chakraborti, Hankz Hankui Zhuo, and Subbarao Kambhampati. 2016. Plan Explicability for Robot Task Planning. In RSS Workshop on Planning for Human-Robot Interaction. [35] Yu Zhang, Sarath Sreedharan, Anagha Kulkarni, Tathagata Chakraborti, Hankz Hankui Zhuo, and Subbarao Kambhampati. 2017. Plan Explicability and Predictability for Robot Task Planning. In ICRA. Technical Report, 2018, ASU	2cs.AI
Action and perception for spatiotemporal patterns Martin Biehl1,2 and Daniel Polani 2 1 2 Araya, Tokyo, 151-0001 University of Hertfordshire, Hatfield, AL10 9AB arXiv:1706.03576v1 [cs.AI] 12 Jun 2017 martin@araya.org Abstract This is a contribution to the formalization of the concept of agents in multivariate Markov chains. Agents are commonly defined as entities that act, perceive, and are goal-directed. In a multivariate Markov chain (e.g. a cellular automaton) the transition matrix completely determines the dynamics. This seems to contradict the possibility of acting entities within such a system. Here we present definitions of actions and perceptions within multivariate Markov chains based on entitysets. Entity-sets represent a largely independent choice of a set of spatiotemporal patterns that are considered as all the entities within the Markov chain. For example, the entityset can be chosen according to operational closure conditions or complete specific integration. Importantly, the perceptionaction loop also induces an entity-set and is a multivariate Markov chain. We then show that our definition of actions leads to non-heteronomy and that of perceptions specialize to the usual concept of perception in the perception-action loop. 1 Introduction The perception-action loop (PA-loop) has been used to formalize, in mostly information theoretic terms, various properties associated to agents. These include empowerment Klyubin et al. (2005), autonomy (Bertschinger et al., 2008), decisions (Tishby and Polani, 2011), and embodiment (Zahedi and Ay, 2013). In the literature agents are usually seen as entities that act, perceive, and are in some way goal-directed (cmp. Barandiaran et al., 2009). The PA-loop assumes that the entities that make up agents as well as their environments can be captured by interacting stochastic processes. This is a convenient assumption since actions and perceptions can then be easily identified with the interactions (see Section 3). Further requirements (e.g. autonomy) can then be introduced to distinguish stochastic processes that actually constitute agents. It has not been formally established whether the assumption that the set of entities that contains agents can be represented by stochastic processes is justified. We have argued in previous work, that (naively) using stochastic processes to capture entities within a given multivariate Markov chain (for example cellular automata like the Game of Life) fails to account for essential proporties of agents irrespective of chosen additional conditions. Instead, we have argued for the use of spatiotemporal patterns (STPs, previously employed by Beer, 2014b,a) to represent entities. The immediate advantage is that STPs are a superset of structures like gliders in the game of life, spots in reaction diffusion systems (Virgo, 2011; Froese et al., 2014; Bartlett and Bullock, 2015), and particle based systems exhibiting individuation into multi-particle “cells” (Schmickl et al., 2016). Formally capturing these structures then becomes a matter of selecting the according subsets of STPs i.e. the entity-sets (see below). A disadvantage of the STP based entities is that they lack the formal construction/interpretation of actions and perceptions that the PA-loop provides. As far as we know, no formal definitions of actions and perception exists for STP based entities. The first contribution of this paper are proposals of such formal definitions called entity action (Section 4) and entity perception (Section 5). The second contribution is a formal connection between the PA-loop and our STP-based entity actions and perceptions (Section 6). This connection is achieved via the notion of entity-sets. This is just the set of those STPs in a multivariate Markov chain that are considered as entities according to an independently specified criterion (e.g. organizational closure (Beer, 2014a) or complete local integration (Biehl et al., 2017, 2016)). Importantly, we can quite naturally identify an entity-set for the PA-loop and use our definitions of entity action and entity perception for this entity-set. The result is that our entity perception coincides with the standard notion of perceptions in the PA-loop and that entity actions are a necessary and sufficient condition for nonheteronomy. Non-heteronomy (i.e. not being determined by the environment) was proposed in Bertschinger et al. (2008) as part of an information theoretic measure of autonomy. The most closely related work is that of Beer (2014b). Apart from a generalization to stochastic systems our set of entity perceptions seems to be a straightforward (but surprisingingly tedious) formalization of the cognitive domain of STPs described for the glider in game of life. We use “perceptions” instead of “cognitive domain” only because our motivation came more from the PA-loop and unlike Beer (2014b) not directly from autopoiesis. Concerning the entity actions we deviate from Beer (2014b) by requiring more from an action than just the continuation of the entity. We note that Ikegami and Taiji (1998) propose to use possible/compatible counterfactual trajectories of game players as signs of autonomy. We construct the capability to act from the counterfactual trajectories and find that they imply non-heteronomy which is a related to autonomy. E0 E1 E2 M0 M1 M2 Figure 1: First timesteps of the PA-loop BN. The processes represent environment tEt utPT , and agent memory tMt utPT . 2 Notation We restrict ourselves to finite, time-discrete, multivariate Markov chains. These are unrolled in time and can then be formally described as Bayesian networks (BNs). We index the random variables in the BN via the index set V “ J ˆ T where T is the set of all timesteps and J is the set of (spatial) degrees of freedom. If i P V is an index we also write i “ pj, tq P J ˆ T where convenient. The BN is then a set tXi uiPV of random variables together with a set of edges determining the parents papiq “ papj, tq of each node i P V , and associated mechanisms ppj,tq pxj,t \|xpapj,tq q. We assume that the parents of any node are a subset of the nodes at the previous timestep papj, tq Ď pJ, tq :“ tpj, tq\|j P Ju. We write XA :“ pXi qiPA for the joint random variable consisting of random variables indexed by elements of A Ď V . We also sometimes write At :“ tpj, tq\|j P Au for the elements in A that correspond to indices at timestep t. We refer to At as the time-slice of A at t. The state space of random variable Xi is denoted as Xi and the specific values are denoted by lower case letters xi , yś i , x̂i , ... P Xi . For joint random variables we write XA :“ iPA Xi and xA , yA , x̂A , ... P XA . A spatiotemporal pattern (STP) is a value xA P XA of a joint random variable XA with A Ď V . Since A Ď V is an arbitrary subset of V a STP can specify the values of random variables at multiple timesteps Ť and multiple spatial locations. The set of all STPs is: AĎV XA “ txA P XA \|A Ď V u . It is important to envision the difference between the set of all STPs and the set of all subsets of random variables of tXi uiPV . The latter Ť is isomorphic to the power set of V and can be written as AĎV tXA u “ tXA \|A Ď V u and is a set of random variables not a set of values of random variables. A trajectory is a STP xV that occupies all random variables in the BN tXi uiPV . We then say that a STP yA P XA occurs within trajectory xV if yA “ xA . An entity-set EptXi uiPV q is a subset of all STPs. One choice would be to use the entire set of STPs as the entity set. Other choices include using organizational closure conditions like in Beer (2014b) or the complete specific integration criterion (Biehl et al., 2016, 2017). The following definitions and theorems all assume that tXi uiPV is a multivariate Markov chain with V “ J ˆ T and E is a given entity-set. 3 Perception-action loop Given two interacting stochastic processes (e.g. Fig. 1) we can always extract random stochastic process that explicitly represent the interactions. If we see one of the processes as the agent’s memory process tMt utPT and the other as the environment process tEt utPT then these extracted random variables can be seen as the perceptions and actions of the agent. Perceptions tSt utPT then capture exactly all the influence of the environment on the agent and actions tAt utPT capture the influence of the agent on the environment. This means we introduce another BN containing two more processes, the action process tAt utPT and the sensor process tSt utPT . The result of this extraction is an extended BN (Fig. 2) with identical joint probability distribution over the two initial stochastic processes tMt utPT and tEt utPT . This perception action loop is used for example in Bertschinger et al. (2008). The idea behind the extraction of perceptions tSt utPT (and conversely actions tAt utPT ) is to partition the state space Et of the environment at t into blocks having identical influence on the next memory state Mt`1 . These blocks are then the possible perceptions i.e. the states of St . Formally: Definition 1. For each time t P T and êt , ēt P Et let êt ”ǫt ēt ô @mt`1 P Mt`1 , mt P Mt : pMt`1 pmt`1 \|mt , êt q “ pMt`1 pmt`1 \|mt , ēt q. (1) Then: (i) The sensor partition1 ǫt is then defined as the set of equivalence classes of the equivalence relation ”ǫt . (ii) The set of sensor values is defined as St :“ ǫt and an element st P St (which is also a block in ǫt is called a perception of a sensor value. In the symmetrical way we define actions via a partition of Mt and arrive at the extended PA-loop of Fig. 2. It is then straightforward to prove the following theorem: 1 The construction of the sensor partition is not new. It is also used for example in Balduzzi (2011) to obtain coarser states (alphabet) of joint random variables. The authors thank Benjamin Heuer for originally pointing them to this construction. E0 M0 E1 E2 S0 S1 A0 A1 M1 M2 Figure 2: First time-steps of the BN of the extended PA-loop. The processes tAt utPT and tSt utPT mediate all interactions between tMt utPT and tEt utPT without changing the probability distributions over the latter (see Theorem 1). Theorem 1 (Invariant extension theorem). Given a perception action loop tXi uiPV “ tMt , Et utPT and its extended PA-loop tXi uiPW “ tMt , At , St , Et utPN` . Let pV “ pMT ,ET be the probability distribution over the entire perception action loop tXi uiPV and let pw MT ,ET be the marginal probability distribution over the memory and environment process obtained from the probability distribution pw W over the entire extended PA-loop. Then pMT ,ET “ pw MT ,ET . (2) Proof. This is probably fairly well known but see Biehl (2017) for an explicit proof. This shows that the introduction of action and sensor process in the above way only makes the interactions between agent and environment processes explicit and does not introduce any additional dynamics. The theorem also shows that the sensor process (and conversely the actions) captures all influences from the environment on the agent. Else the dynamics of the original processes could not remain identical. In Section 5 we want to capture all influences of the environment on a set of STPs / entities instead of on a stochastic process like tMt utPT . This will require a generalization of the perception extraction procedure in Definition 1. 4 Entity action We now define a concept of actions for a given entity-set in a multivariate Markov chain. First we briefly sketch the main ideas behind the definition. Due to our setting of a given multivariate Markov chain that actions have to occur within our concept of actions differs from other approaches. Paraphrasing Wilson and Shpall (2012) only slightly, what distinguishes actions among events is that they do not merely happen to individuals but rather that they are made to happen by the individuals. This is problematic in our setting where STPs (as entities) take the role of individuals. What “happens” in a multivariate Markov chain are the trajectories and the STPs occurring within them. The Markov chain’s dynamics are determined by its mechanisms pj,t with j P J, t P T . These in turn determine (possibly stochastically) what is going to happen at all times anywhere within the chain. Therefore, it is impossible that a multivariate Markov chain contains an STP or entity that can make something happen beyond what happens anyway due to the mechanisms. This means we have to explain and define actions in a different way. We should also note that unlike other accounts of actions (Wilson and Shpall, 2012) we do not require actions to be necessarily purposeful or goal-directed in any way. The way we conceive agency an entity with actions will be considered goal-directed if its actions are goal-directed in some sense. To get to our own definition of actions we note that events called actions are usually attributed to a limited or bounded region or part of the universe e.g. the body of a living organism or sometimes just its brain if it has one. These parts usually contain mechanisms or configurations of matter that are either aq not directly observable to a human observer e.g. hidden in an opaque container, or bq not well understood by the human observer, or cq both. These factors inevitably lead to unpredictability of such events. In other words, events that are attributed to well understood and therefore predictable mechanisms, e.g. sunrises, are not considered actions. From this point of view actions are not, beyond their possibly complex and unobserved origin, special events but may appear as special to observers that lack the sensory and computational capacity to resolve or understand them. In our approach we construct actions as events in such a way that they are fundamentally unpredictable by any observer within the system. As we will see this can be done without the need for a definition of an observer. Note that this approach remains compatible with observer-dependent notions of apparent or as-if actions (cmp. McGregor, 2017). The “fundamental” actions are apparent actions for every possible observer while other events are actions for some observers and plain, predictable events for others. Up to now we have ignored randomness. True randomness (in the sense of stochastic independence of the event from any other event in the universe), if it exists in a system, can never be explained, predicted, or understood. Combined with our reasoning above this suggests that all random events are actions and fundamentally so. This is against our intuition, random events should not be seen as actions of agents. However, we place the burden of ruling out random events from being interpreted as actions of agents on the entity-set and/or goal-directedness. We expect useful notions of entities not to consist of (completely) independent events which would prevent such events from becoming actions in our approach. Furthermore independent events seem not to be useful in order to achieve a particular goal. However, the usefulness of random number generators might be seen as a counterexample. As of now we have no formal definition of goal-directedness such that answering these questions is future work. When we want to define actions for entities the first issue we run into is that entities are already fixed STPs. Therefore each entity xA P E already consists of the consequences of whatever actions (if any) it took. It is in this sense the result of its own actions. There is no freedom left. In order to investigate the actions we therefore have to deconstruct the entity and see what other actions it could have taken time-slice by time-slice. For this we look for counterfactual entities yB whose time-slice at t can occur in exactly the same environment xVt zAt and has a different immediate future. Requiring the co-occurrence with the same environment makes the different futures unpredictable for anything part of that environment. No observer can therefore predict these futures either. Even an observer that distinguished the two entities in the past is in the same state at t independent of the entity it is faced with. Therefore it must have forgotten their difference. In this way we get a definition of actions as unpredictable events for observers without needing to formally define observers. For the formal definition we first define: Definition 2 (Environment of an STP). Let xA be a STP. Then the environment of xA at time t is the spatial pattern xVt zAt . We now state the definition of an action of an entity at a time t in a particular trajectory formally. Definition 3 (Action and co-action of an entity). Let xV P XV with pV pxV q ą 0. Also let xA be an entity with nonempty time-slices at t, t ` 1. Then xA performs an action xAt`1 at time t in trajectory xV if there exists an entity yB with non-empty time-slices at t, t ` 1 such that (i) yB occurs in yV ‰ xV with pV pyV q ą 0, (ii) at t the entities xA and yB occupy the same random variables: Bt “ At , (iii) at t the trajectories xV and yV are otherwise identical: xVt zAt “ yVt zAt , (iv) at t ` 1 the entities are different: xAt`1 ‰ yBt`1 . We also call yB a co-action entity, yV a co-action trajectory, and yBt`1 a co-action. Note that all requirements are symmetric. Therefore, if xA performs an action then its co-action also performs an action. Also, the notion of co-action entities can easily be extended to more than one co-action entity. We only have to make sure that all entities in a set of co-action entities are mutually different at t ` 1. Furthermore, it is easy to generalise the definition of actions to situations where xA and yB must occupy the same variables for an interval of time rt ´ m : ts before the action. In that case, the environment xVrt´m:ts zArt´m:ts must be identical during this interval. Finally, note that the condition that the two acting entities differ at time t ` 1 can be fulfilled in two ways. If At`1 ‰ Bt`1 then we call these actions extent actions. Else, if the actions differ only in value i.e. we have At`1 “ Bt`1 so that xAt`1 ‰ yAt`1 then we call these actions value actions. The difference between value actions and extent actions is made possible due to our definition of entities as STPs. An intriguing question for the future is whether the capabilities of agents to act both in value and extent are truly superior to agents that only act in value such as those modelled by PA-loops. As we will see in Section 6 probabilistic and information theoretic expressions are easy to formulate for actions in value only. However, for actions in extent this has not been done yet. 5 Entity perception In this section we formally define perception for (STP-) entities. We make no distinction here between perception, experience, and sensory input. In the tradition of modelling agent-environment systems using dynamical systems or their probabilistic generalisations as stochastic processes in PA-loops we define perception as all effects that the environment has on an individual/agent (Beer, 1995). We run into a similar problem as with the actions. An entity is already a fixed STP that contains all influence that it may have been subjected to. It is in this sense the result of influence (or no influences) from its surroundings. In order to investigate these influences we therefore have to deconstruct the entity and see how it was “formed” by external influences / perceptions time-slice by time-slice. The idea is to fix the past of the entity up to t and use the set of counterfactual entities with the same past as the alternative futures. We can then partition the possible environments of these counterfactual entities according to their influence on the probability distribution over the entities’ futures. This is done in basically the same way as we defined perception in the PA-loop in Definition 1. However, there are some technical issues to overcome. The set of entities with identical pasts up to time t can be interpreted as the set of entities that are the most like xA up to t. These are different entities but they only differ in the future. Their futures (including their next time-slices) are therefore a close analogue to the next states mt`1 of agent memories in the PA-loop. To make sure however that the entities have a next time-slice we also require that they have non-empty next time-slice. These requirements together define the notion of the co-perception entities of an entity xA at time t. These are entities that also perceive something (maybe the same thing) at t (in their trajectories) if xA perceives something at t. Definition 4 (Co-perception entities of an entity at t). Let xA P E be an entity with non-empty time-slices at t and t ` 1. The set of co-perception entities SpxA , tq of entity xA at t is the set of entities with non-empty time-slices at t and t ` 1, and that are identical up to t: SpxA , tq :“ tyB P E : Bt , Bt`1 ‰ H, yBĺt “ xAĺt u. (3) Next we want to define a conditional probability distribution over these co-perception entities similar to pMt`1 pmt`1 \|et , mt q. For this we need a random variable that ranges over all the possible futures of the co-perception entities. The naivest way to do this would be to use for xkAk P SpxA , tq: pS pxkAk \|xVt zAt , xAĺt q :“ ąt pAk ,Vt zAt pxkAk , xVt zAt , xAĺt q ąt pVt zAt ,Aĺt pxVt zAt , xAĺt q . (4) However, this has two problems. The first is that in general the denominator may vanish for some environments xVt zAt P XVt zAt . The second is that it is not a conditional probability since the sum over all co-perception indices is not necessarily one. These problems can be solved by restricting the set of environments and by introducing a mutual-exclusivity condition on the futures of the coperception entities. To restrict the environments we therefore define the coperception environments in the following way. Definition 5 (Co-perception environments). Let xA P E be an entity with non-empty time-slices at t and t ` 1 and SpxA , tq its co-perception entities. Then define the associated co-perception environments XVSt zAt Ď XVt zAt by XVSt zAt :“ tx̄Vt zAt : DyB P SpxA , tq, pB,Vt zAt pyB , x̄Vt zAt q ą 0u. (5) The co-perception environments of a co-perception set SpxA , tq are then the spatial patterns XVt zAt at t that can co-occur with at least one co-perception environment. By definition of the co-perception environments the denominator of Eq. (4) cannot vanish anymore if we require that the construction only allows co-perception environments. However, in order for the sum over all co-perception entities to equal one (for all co-perception environments) we need to have ÿ pAk ,XVt zAt pxkAk , xVt zAt , xAĺt q ąt k (6) “ pVt zAt ,Aĺt pxkAk , xVt zAt , xAĺt q. ąt This is the case in general if for all xVt zAt P XVSt bsAt and all co-perception entities’ futures are mutually exclusive, i.e. for all xkAk ‰ xlAl P SpxA , tq we have PrpxkAk , xlAl , xVt zAt , xAĺt q “ 0. (7) This condition can be guaranteed if we require a form of non-interpenetration. This condition on entity-sets states that there cannot be two different entities which are identical up to some point in time t and then, in the same single trajectory (with positive probability), at some point “reveal” their difference. If entities with identical pasts ever reveal their difference they must be in different trajectories i.e. they must be mutually exclusive. Definition Ť 6 (Non-interpenetration). An entity-set E Ď BĎV XB satisfies non-interpenetration or is non-interpenetrating if for all yB , zC P E we have Dt P T : yBĺt “ zCĺt and yBtă ‰ zCtă ñ PrpXBtă “ yBtă , XBtă “ zCtă \|yBĺt q “ 0. (8) Non-interpenetration implies that co-perception entities are mutually exclusive: Theorem 2. Let xA P E be an entity with non-empty timeslices at t and t ` 1 and SpxA , tq its co-perception entities. If E satisfies non-interpenetration then SpxA , tq is mutually exclusive. Proof. Let yB , zC P SpxA , tq with yB ‰ zC . Then they have identical pasts and so we have yBĺt “ zCĺt . From non-interpenetration we then get PrpXB “ yB , XC “ zC q “ 0. (9) This is stronger than Eq. (7). This means that under non-interpenetration Eq. (4) is a well defined conditional probability distribution . However, this conditional probability distribution is still quite different from pMt`1 pmt`1 \|et , mt q since it ranges over the entire futures xkAk of the co-perception entities and not just next ąt timesteps. At each transition from time-step t to t ` 1 the coperception entities SpxA , tq split up into sets of entities that are identical up to t`1 (we will call these sets the branches). Only one of these sets is the set SpxA , t ` 1q. For example an entity yB P SpxA , tq with the same past up to t but with a different time-slice at t ` 1 i.e. yBt`1 ‰ xAt`1 is part of a different branch. In that case this branch is SpyB , t ` 1q and we have SpyB , t ` 1q X SpxA , t ` 1q “ H. In summary then the dynamics of the system split up the co-perception entities of xA up to t into disjoint sets (the branches) of entities with identical pasts up to t ` 1. We can then interpret the branches at the time t ` 1 as the distinctions among the co-perception entities that are revealed at time t ` 1. Further distinctions among the co-perception entities are only revealed at later times. This also means that these are all differences that could possibly be due to the influence of the environment at t and that show their effect at t`1 (not later). In this way the perceptions at t should also be defined with respect to these branches. We call the partition that is defined via the identification of entities in SpxA , tq that are identical up to t ` 1 the branching partition. Definition 7 (Branching partition). Let xA P E be an entity with non-empty time-slices at t and t ` 1 and SpxA , tq its co-perception entities. Then define the branching partition ηpxA , tq of SpxA , tq as the partition induced by the equivalence classes of the equivalence relation yB „zC ô yBt`1 “ zCt`1 , (10) where yB , zC P SpxA , tq. We note that the definition of the branching partition can easily be generalised to more than one time-step into the future. Instead of requiring equality at t ` 1 we can require equality for the next r time-steps. Given the branching partition ηpxA , tq for a noninterpenetrating entity set we can then define a conditional probability distribution over the branches by just summing up the probabilities of all entities in each branch (remember that they are all mutually exclusive) to get the probability of a branch. This gives us the branch-morph defined below.2 Definition 8 (Branch-morph). Let tXi uiPV be a multivariate Markov chain with index set V “ J ˆ T and entity set E. Let xA P E be an entity with non-empty time-slices at t and t`1 and SpxA , tq its co-perception entities and ηpxA , tq the branching partition. Furthermore, let XVSt zAt Ď XVt zAt be the associated co-perception environments. Also write for every block b P ηpxA , tq: ÿ ppb\|x̂Vt zAt , xAĺt q :“ pBtă ,Vt zAt pyBtă \|x̂Vt zAt , xAĺt q. yB Pb (11) Then for each x̂Vt zAt P XVSt zAt we define the branchmorph over ηpxA , tq as the probability distribution pηpxA ,tq p.\|x̂Vt zAt , xAĺt q : ηpxA , tq Ñ r0, 1s with pηpxA ,tq pb\|x̂Vt zAt , xAĺt q :“ ř ppb\|x̂Vt zAt , xAĺt q , cPηpxA ,tq ppc\|x̂Vt zAt , xAĺt q (12) for all b P ηpxA , tq. With the branch-morph we can then define, as expected, the perceptions as equivalence classes of the co-perception environments with respect to the associated branch-morph. First we define a partition of the co-perception environments called the co-perception environment partition. The perceptions are then the blocks of this partition. Definition 9. Let tXi uiPV be a multivariate Markov chain with index set V “ J ˆ T and entity set E. Let xA P E be an entity with non-empty time-slices at t and t`1 and SpxA , tq its co-perception entities and ηpxA , tq the branching partition. Furthermore, let XVSt zAt Ď XVt zAt be the associated co-perception environments. Then define the co-perception environment partition π S pxA , tq of XVSt zAt as the partition induced by the equivalence classes of the equivalence relation x̂Vt zAt „ x̄Vt zAt ô @b P ηpxA , tq : pηpxA ,tq pb\|x̂Vt zAt , xAĺt q “ pηpxA ,tq pb\|x̄Vt zAt , xAĺt q. (13) This means all associated co-perception environments in the same block of π S pxA , tq have the same branch-morph. In other words they lead to the same branch of entity futures (i.e. the same future branch) with the same probabilities. Then all elements of these environment blocks have identical effects on the future branches and these branches cannot distinguish between environments within the blocks. Definition 10 (Perceptions). Let tXi uiPV be a multivariate Markov chain with index set V “ J ˆ T and entity set E. Let xA P E be an entity with non-empty time-slices at t and t ` 1 and SpxA , tq its co-perception entities. Furthermore, let XVSt zAt Ď XVt zAt be the associated co-perception environments and π S pxA , tq its co-perception environment partition. Then the blocks of π S pxA , tq are called the perceptions of xA at t. 6 Entity action and perception in the PA-loop We now show that agent-environment systems as modelled by the PA-loop are multivariate Markov chains containing a specific choice of entity sets. In the PA-loop each trajectory xV is considered to consist of a time-evolution mT of the agent and a time-evolution of the environment eT . The agent therefore occurs in every trajectory and occupies the same degree of freedom in every trajectory. Each of the time-evolutions mT is a STP in the PA-loop. Since for us entities are STPs we define the entityset EP A of a PA-loop as the set of time-evolutions of the agent process i.e. ź EP A :“ tmT P Mt u. (14) tPT Similarly, we can define entities for the environments and add them to EP A . 2 We note here that for entity-sets that exhibit interpenetration we can still define branch-morphs for mutually-exclusive subsets of the co-perception entities. Since the choice of these subsets is arbitrary however this does not lead to a uniquely defined notion of perception. For more details see Biehl (2017). Entity actions in the PA-loop We can write every trajectory as a pair pmT , eT q where mT is an entity. The entity mT then performs an entity action at time t in trajectory pmT , eT q with pMT ,ET pmT , eT q ą 0 if there is an entity m̄t such that • m̄T occurs in pm̄T , ēT q ‰ pmT , eT q with pMT ,ET pm̄T , ēT q ą 0, • at t entities mT and m̄T occupy the same random variables, which is the case for all entities in EP A , • at t environments of mT and m̄T are identical: et “ ēt , • at t ` 1 the entities are different: mt`1 ‰ m̄t`1 . Since all entities occupy the same random variables we can only have value actions in the PA-loop. If we assume that these conditions are fulfilled at some time t for two entities mT , m̄T we can derive that the conditional entropy HpMt`1 \|Et q of the next agent state given the current environment state is greater than zero. To see this note that from pMT ,ET pmT , eT q ą 0 and pMT ,ET pm̄T , ēT q ą 0 it directly follows that pMt`1 pmt`1 \|et q ą 0, pMt`1 pm̄t`1 \|et q ą 0 and pEt pet q ą 0. Plugging this into the definition we get HpMt`1 \|Et q ą 0. It can also be seen that the more different co-action entities there are for a time t the higher the conditional entropy HpMt`1 \|Et q can get. The final value of HpMt`1 \|Et q depends on the actual probabilities but the maximum value for n co-actions is log n. Also note that if there are no actions at t i.e. no co-action entity in no co-action trajectory at t then HpMt`1 \|Et q “ 0. Entity actions of entities in EP A are therefore necessary and sufficient for HpMt`1 \|Et q ą 0. The conditional entropy HpMt`1 \|Et q measures the uncertainty about the next agent state when the current environment state is known. It has been proposed as part of an autonomy measure as a measure of non-heteronomy in Bertschinger et al. (2008). Non-heteronomy means that the agent is not determined by the history of the environment. In this terminology entity actions are a necessary and sufficient condition for non-heteronomy. Entity perception in the PA-loop We now look at how entity perception as defined in Section 5 specialises to the case of the PA-loop. This argument in effect constitutes a proof that our Definition 8 of the branch-morph is a generalisation of the conditional probability distributions pMt`1 p.\|mt , et q : Mt`1 Ñ r0, 1s to non-interpenetrating arbitrary sets of co-perception entities This result is not surprising since we set out to do just this but it is also instructive to work through the recovery of the original expression of the conditional probability distribution starting from the general branch-morph. We pick an entity mT from the entity set EP A and consider its perceptions at an arbitrary time-step t P T . In order to get the perceptions at t we need 1. the co-perception entities SpmT , tq of mT at t, 2. the branching partition ηpmT , tq with its branches, 3. the co-perception environments, 4. the branch-morphs for each environment, 5. and the co-perception environment partition π S pxA , tq with its blocks, the perceptions. These can be identified in the following way. 1. The co-perception entities SpmT , tq are the entities in EP A that have non-empty time-slices at t, t ` 1, and that are identical to mT up to t. All entities in EP A have non-empty time slices at all times. So we have: SpmT , tq “ tm̄T P EP A : m̄ĺt “ mĺt u (15) PA 2. First note that the entity set E satisfies noninterpenetration since they all occupy the same set tMt utPT of random variables. The branching partition ηpmT , tq is composed out of blocks (the branches) of co-perception entities that are identical up to t ` 1 i.e. m̂T „m̄T ô m̂t`1 “ m̄t`1 . (16) We can therefore identify the blocks of ηpmT , tq i.e. the future branches by the values that the entities take at t ` 1. Define the branch bpm̄t`1 q associated to m̄t`1 P Mt`1 via bpm̄t`1 q :“ tm̂T P SpmT , tq : m̂t`1 “ m̄t`1 u. (17) The branching partition is then: ηpmT , tq “ tbpm̄t`1 q Ď SpmT , tq : m̄t`1 P Mt`1 u. (18) 3. The co-perception environments are the STPs xVt zAt compatible with at least one co-perception entity. For the PA-loop and entity mT at t we have XVt zAt “ Et and therefore XVSt zAt “ EtS . Where EtS is EtS “ tet P Et : Dm̄T P SpmT , tq, pMT ,Et pm̄T , et q ą 0u. (19) If we marginalize over Mąt we can see that this is equivalent to (20) EtS “ tet P Et : pMt ,Et pm̄t , et q ą 0u. 4. The branch-morphs are the probability distributions pηpmT ,tq p.\|et , mĺt q : ηpmT , tq Ñ r0, 1s over the branches for each co-perception environment et P EtS . These are defined using Eq. (11) which for the perception-loop becomes ÿ ppbpm̄t`1 q, et \|mĺt q :“ pMtă ,Et pm̂tă , et \|mĺt q. m̂T Pbpm̄t`1 q (21) We can rewrite the sum on the right hand side using Eq. (17) for bpm̄t`1 q and then SpmT , tqtă “ Mtă : ppbpm̄t`1 q, et \|mĺt q “ pMt`1 ,Et pm̄t`1 , et \|mĺt q. (22) The definition of the branch-morph for the PA-loop is ppbpm̄t`1 q, et \|mĺt q pηpmT ,tq pbpm̄t`1 q\|et , mĺt q : “ ř bPηpmT ,tq ppb, et \|mĺt q (23) which can be rewritten (with some work) as pηpmT ,tq pbpm̄t`1 q\|et , mĺt q “ pMt`1 pm̄t`1 \|et , mĺt q (24) “ pMt`1 pm̄t`1 \|et , mt q (25) where we used the BN of the PA-loop. 5.The co-perception environment partition π S pmT , tq of S Et “ is induced by Eq. (13) which, using the PA-loop and Eq. (24) becomes êt „ ēt ô @mt`1 P Mt`1 : (26) pMt`1 pmt`1 \|êt , mt q “ pMt`1 pmt`1 \|ēt , mt q which is just the equivalence relation of Eq. (1) used to extract the sensor-values in Section 3. So we have seen that our definitions of Section 5 specialise in the case of the PA-loop to the same concept of perception as in Section 3. 7 Conclusion We have defined actions and perceptions for entity-sets and therefore for sets of spatiotemporal patterns. This provides a formally defined way to associate gliders and similar spatiotemporal patterns in reaction-diffusion systems with actions and perceptions. This is a step towards a foramization of agency of such patterns. Here a notion of goaldirectedness is still missing and future work. We have also shown how our definitions specialize to a necessary and sufficient condition for non-heteronomy and the standard notion of perceptions of the agent process in the PAloop. For future research it is interesting to note that the branch-morphs are generalisations of the conditional probability distribution pMt`1 p.\|et , mt q. These conditional probability distributions play a role in various information theoretic concepts formulated for the PA-loop. This suggests we might be able to translate bakc and forth between PA-loop concepts and those for spatiotemporal patterns in the future. We noted that a unique definition of entity perception is dependent on the condition of non-interpenetration of the entity-set. Beer, R. D. (1995). A dynamical systems perspective on agentenvironment interaction. Artificial Intelligence, 72(1-2):173– 215. Beer, R. D. (2014a). Characterizing autopoiesis in the game of life. Artificial Life, 21(1):1–19. Beer, R. D. (2014b). The cognitive domain of a glider in the game of life. Artificial Life, 20(2):183–206. Bertschinger, N., Olbrich, E., Ay, N., and Jost, J. (2008). Autonomy: An information theoretic perspective. Biosystems, 91(2):331–345. Biehl, M., Ikegami, T., and Polani, D. (2016). Towards information based spatiotemporal patterns as a foundation for agent representation in dynamical systems. In Proceedings of the Artificial Life Conference 2016, pages 722–729. The MIT Press. Biehl, M., Ikegami, T., and Polani, D. (2017). Specific and Complete Local Integration of Patterns in Bayesian Networks. Entropy, 19(5):230. Biehl, M. A. (2017). Formal Approaches to a Definition of Agents. PhD thesis, University of Hertfordshire, Hatfield. Froese, T., Virgo, N., and Ikegami, T. (2014). Motility at the origin of life: Its characterization and a model. Artificial Life, 20(1):55–76. Ikegami, T. and Taiji, M. (1998). Uncertainty, possible worlds and coupled dynamical recognizers. http://sacral.c.u-tokyo.ac.jp/pdf/ikegami_polytechn Klyubin, A., Polani, D., and Nehaniv, C. (2005). Empowerment: a universal agent-centric measure of control. In The 2005 IEEE Congress on Evolutionary Computation, 2005, volume 1, pages 128–135 Vol.1. McGregor, S. (2017). The Bayesian stance: Equations for ‘as-if’ sensorimotor agency. Adaptive Behavior, page 105971231770050. Schmickl, T., Stefanec, M., and Crailsheim, K. (2016). How a lifelike system emerges from a simplistic particle motion law. Scientific Reports, 6:37969. Tishby, N. and Polani, D. (2011). Information Theory of Decisions and Actions. In Cutsuridis, V., Hussain, A., and Taylor, J. G., editors, Perception-Action Cycle, Springer Series in Cognitive and Neural Systems, pages 601–636. Springer New York. DOI: 10.1007/978-1-4419-1452-1 19. References Virgo, N. (2011). Thermodynamics and the Structure of Living Systems. University of Sussex. Unpublished PhD thesis. Balduzzi, D. (2011). Detecting emergent processes in cellular automata with excess information. Advances in Artificial Life, ECAL, abs/1105.0158. Wilson, G. and Shpall, S. (2012). Action. In Zalta, E. N., editor, The Stanford Encyclopedia of Philosophy. Summer 2012 edition. Barandiaran, X. E., Paolo, E. D., and Rohde, M. (2009). Defining agency: Individuality, normativity, asymmetry, and spatiotemporality in action. Adaptive Behavior, 17(5):367–386. Zahedi, K. and Ay, N. (2013). Quantifying morphological computation. Entropy, 15(5):1887–1915. arXiv:1301.6975 [cs, math]. Bartlett, S. and Bullock, S. (2015). Emergence of Competition between Different Dissipative Structures for the Same Free Energy Source. In Proceedings of the European Conference on Artificial Life, pages 415–422. The MIT Press.	2cs.AI
GATED RECURRENT NETWORKS FOR SEIZURE DETECTION M. Golmohammadi, S. Ziyabari, V. Shah, E. Von Weltin, C. Campbell, I. Obeid and J. Picone Neural Engineering Data Consortium, Temple University, Philadelphia, Pennsylvania, USA {meysam, saeedeh, vinitshah, eva.vonweltin, christopher.campbell, obeid, picone}@temple.edu Abstract— Recurrent Neural Networks (RNNs) with sophisticated units that implement a gating mechanism have emerged as powerful technique for modeling sequential signals such as speech or electroencephalography (EEG). The latter is the focus on this paper. A significant big data resource, known as the TUH EEG Corpus (TUEEG), has recently become available for EEG research, creating a unique opportunity to evaluate these recurrent units on the task of seizure detection. In this study, we compare two types of recurrent units: long short-term memory units (LSTM) and gated recurrent units (GRU). These are evaluated using a state of the art hybrid architecture that integrates Convolutional Neural Networks (CNNs) with RNNs. We also investigate a variety of initialization methods and show that initialization is crucial since poorly initialized networks cannot be trained. Furthermore, we explore regularization of these convolutional gated recurrent networks to address the problem of overfitting. Our experiments revealed that convolutional LSTM networks can achieve significantly better performance than convolutional GRU networks. The convolutional LSTM architecture with proper initialization and regularization delivers 30% sensitivity at 6 false alarms per 24 hours. I. INTRODUCTION Diagnosis of clinical conditions such as epilepsy are dependent on electroencephalography (EEG), the recording of the brain’s electrical activity through electrodes placed on the scalp. Delivering a conclusive diagnosis of a brain-related illness without an EEG is often unfeasible [1]. The large amounts of time required by specialized neurologists to interpret these records, has created a workflow bottleneck – neurologists are overwhelmed with the amount of data that needs to be manually reviewed [2]. There is a great need for partial or complete automation of the EEG analysis process, and automated technology is slowly emerging to fill this void [3][4]. Automatic analysis of EEG scans reduces time to diagnosis, reduces error and enhances a neurologist’s ability to administer medications. The ability to search EEG records symbolically greatly accelerates the review process. In this paper, we focus specifically on the problem of seizure detection. Many algorithms have been applied to this problem including time–frequency analysis methods [5][6], nonlinear statistical models [7] and more modern machine learning approaches such as neural networks and support vector machines [8]. Despite much progress, current EEG analysis methodologies are far from perfect with many being considered impractical due to high false detection rates [9]-[11]. The machine learning challenges of this application are described extensively in [1]. A significant big data resource, known as the TUH EEG Corpus (TUEEG) [12], has become available for EEG interpretation creating a unique opportunity to advance technology. Using a subset of this data that has been manually annotated for seizure events [13], a novel deep structure has been recently introduced which achieves a low false alarm rate on EEG signals [14]. This system integrates convolutional neural networks (CNNs) with recurrent neural networks (RNNs) to deliver state of the art performance. In this paper, our goal is to investigate the use of RNNs using the high-performance architecture described in [13]. We also explore improved initialization methods and regularization approaches. II. RECURRENT NEURAL NETWORKS A recurrent neural network (RNN) is an extension of a conventional feedforward neural network which can handle a variable-length input. The RNN handles the variable-length sequence by having a recurrent hidden state whose activation at each time is dependent on that of the previous time. Standard RNNs are hard to train due to the well-known vanishing or exploding gradient problems [15][16]. To address the vanishing gradient problem, the gated recurrent network architectures such as long short-term memory (LSTM) [17] unit and gated recurrent unit (GRU) were proposed [18]. LSTM was presented in [17]. The most commonly used architecture was described in [19], and is formulated as: +𝑏 ), = 𝜎(𝑈 𝑥 + 𝑊 𝑠 + 𝑝 ∘ 𝑐 + 𝑏 ) , = 𝑓 ∘ 𝑐 + 𝑖 ∘ 𝑔(𝑈 𝑥 + 𝑊 𝑠 + 𝑏 ) , = 𝜎(𝑈 𝑥 + 𝑊 𝑠 + 𝑝 ∘ 𝑐 + 𝑏 ) , = 𝑜 ∘ 𝑔(𝑐 ) . 𝑖 = 𝜎(𝑈 𝑥 + 𝑊 𝑠 𝑓 𝑐 𝑜 𝑠 +𝑝 ∘𝑐 (1) (2) (3) (4) (5) where 𝑖 , 𝑓 , 𝑐 , 𝑜 , and 𝑠 are the input gate, forget gate, cell state, output gate and block output at time instance t, respectively; 𝑥 is the input at time t; 𝑈 ∗ , and 𝑊 ∗ are the weight matrices applied on input and recurrent hidden units, respectively; 𝜎(. ) and 𝑔(. )are the sigmoid and tangent activation functions, respectively; 𝑝∗ and 𝑏 ∗ are the peep-hole connections and biases, respectively; and ∘ means element-wise product. As an alternative to the LSTM, the Gated Recurrent Unit (GRU) architecture was proposed in [20]. In [18], a GRU architecture was found to achieve better performance than LSTM on some tasks. A GRU is formulated as: 𝑟 𝑧 𝑠̃ 𝑠 = = = = 𝜎(𝑈 𝑥 𝜎(𝑈 𝑥 𝑔(𝑈 𝑥 𝑧 ∘𝑠 +𝑊 𝑠 +𝑏 ), +𝑊 𝑠 +𝑏 ), + 𝑟 ∘ (𝑊 𝑠 ) + 𝑏 , + (1 − 𝑧 ) ∘ 𝑠̃ . (6) (7) (8) (9) As one can see, a GRU architecture is similar to LSTM but without a separate memory cell. Unlike LSTM, a GRU does not include output activation functions and peep-hole connections. It also integrates the input and forget gates into an update gate, 𝑧 , to balance between the previous activation, 𝑠 , and the candidate activation, 𝑠̃ . The reset gate, 𝑟 , allows it to forget the previous state [21]. III. EXPERIMENTAL DESIGN Our basic architecture, which employs a convolutional recurrent neural network, is presented in Figure 1. In this architecture, we integrate 2D CNNs, 1-D CNNs and LSTM networks to better exploit long-term dependencies. This structure currently uses LSTMs. However, we can easily replace LSTMs with GRUs. Feature extraction is performed using a fairly standard linear frequency cepstral coefficient-based feature extraction approach (LFCCs) popularized in applications such as speech recognition [22][23]. We also use first and second derivatives of the features since these provide a small improvement in performance [22]. Drawing on a video classification analogy, input data to the first layer of CNNs is composed of frames distributed in time where each frame is an image of width (W) equal to the length of a feature vector, the height (H) equals the number of EEG channels, and the number of image channels (N) equals one. Input data consists of T frames where T is equal to the window length multiplied by the number of samples per second. In our optimized system with a window duration of 21 seconds, the first 2D convolutional layer filters 210 frames (T = 21 × 10) of EEGs distributed in time with a size of 26 × 22 × 1 (W=26, H=22, N=1) using 16 kernels of size 3 × 3 and with a stride of 1. The first 2D max pooling layer takes as input a vector which is 260 frames distributed in time with a size of 26 × 22 × 16 and applies a pooling size of 2 × 2. This process is repeated two times with two 2D convolutional layers with 32 and 64 kernels of size 3 × 3 respectively and two 2D max pooling layers with a pooling size 2 × 2. The output of the third max pooling layer is flattened to 210 frames with size of 384 × 1. Then a 1D convolutional layer filters the output of the flattening layer using 16 kernels of size 3 which decreases the dimensionality in space to 210 × 16. Then we apply a 1D maxpooling layer with a size of 8 to decrease the dimensionality to 26 × 16. This is the input to a deep bidirectional LSTM network where the dimensionality of the output space is 128 and 256. The output of the last bidirectional LSTM layer is fed to a 2-way sigmoid function which produces a final classification of an epoch. Epochs are typically 1 sec in duration. Figure 1: A deep recurrent convolutional architecture for two-dimensional decoding of EEG signals that integrates 2D CNNs, 1-D CNNs and LSTM networks is shown. In this structure, LSTMs can be easily replaced by GRUs. To overcome the problem of overfitting and force the system to learn more robust features, regularizations are used in first two layers of CNNs. To increase nonlinearity, Exponential Linear Units (ELU) are used [24]. Adam is used in the optimization process along with a mean squared error loss function [14]. System CNN/GRU CNN/LSTM Sensitivity 30.83% 30.83% Specificity 91.49% 97.10% FA/24 Hrs. 21 6 Table 2. Recognition results for convolutional recurrent neural networks using GRU and LSTM architectures, for sensitivity in range of 30%. IV. RESULTS The lack of big data resources that can be used to train sophisticated statistical models compounds a major problem in automatic seizure detection. Inter-rater agreement for this task is low, especially when considering short seizure events [1]. Manual annotation of a large amount of data by a team of certified neurologists is extremely expensive and time consuming. It is difficult to employ large numbers of board-certified neurologists to perform this task. In this study, we are reporting results on the TUH EEG Seizure Corpus (TUSZ) [13]. This dataset, which is publicly available, is a subset of the TUH EEG Corpus that focuses on the problem of seizure detection. A summary of the corpus (v1.1.1) is shown in Table 1. A comparison of the performance of the convolutional recurrent neural networks using GRU and LSTM architectures, for sensitivity in range of 30%, are shown in Table 2 The related DET curve is illustrated in Figure 2 These systems were evaluated using a method of scoring popular in the EEG research community known as the overlap method [25]. True positives (TP) are defined as the number of epochs identified as a seizure in the reference annotations and correctly labeled as a seizure by the system. True negatives (TN) are defined as the number of epochs correctly identified as non-seizures. False positives (FP) are defined as the number of epochs incorrectly labeled as seizure while false negatives (FN) are defined as the number of epochs incorrectly labeled as non-seizure. Sensitivity shown in Table 2 is computed as TP/(TP+FN). Specificity is computed as TN/(TN+FP). The false alarm rate is the number of FPs per 24 hours. By comparing the results of CNN/LSTM with CNN/GRU, demonstrated in Figure 2, we find that in lower false positive rates, CNN/LSTM has significantly better performance from CNN/GRU, due to the fact that while GRU unit controls the flow of information like the LSTM unit, but it does not have a memory unit. LSTMs Description Patients Sessions Files Seizure (secs) Non-Seizure (secs) Total (secs) Train 196 456 1505 51,140 877,821 928,962 Eval 50 230 984 53,930 547,728 601,659 Table 1. An overview of the TUH EEG Seizure Corpus (v1.1.1) Figure 2. DET Curves for convolutional recurrent neural networks using GRU and LSTM architectures. can remember longer sequences better than GRUs and outperform it in this task, since seizure detection requires modeling long distance relationships. Additionally, the training time of CNN/GRU was 10% less than CNN/LSTM. Hence, the training time of these two systems is comparable, since most of the cycles are used for training of convolutional layers. In neural networks, determining the proper initialization strategy for the parameters in the model is part of the difficulty in training. Hence, we investigated a variety of initialization methods using the structure introduced in Figure 1. These results are presented in Table 4. The related DET curve is illustrated in Figure 4. In our experiments, we observed that the proper initialization of weights in a convolutional recurrent neural network is critical to convergence. For example, initialization with zeros or ones methods completely stalled the convergence process. Also, as we can see in Table 4, the performance of the system for the same sensitivity of 30% can change from 7 to 40, for different initialization methods. This decrease in performance and deceleration of convergence arises because some initializations can result in the deeper layers receiving inputs with small variances, which in turn slows down back propagation, and retards the overall convergence process. The best performance is achieved using orthogonal initialization. This method is a simple yet effective way of combatting exploding and vanishing gradients. Orthogonal matrices preserve the norm of a vector and their eigenvalues have absolute value of one. This means that, no matter how many times we perform repeated matrix multiplication, the resulting matrix doesn't explode or vanish [26]. Also in orthogonal matrices, columns and rows are all orthonormal to one another, which helps the weights to learn different input features. Initialization Orthogonal [26] Lecun Uniform [27] Glorot Uniform [28] Glorot Normal [28] Variance Scaling [26] Lecun Normal [27] He Normal [29] Random Uniform [26] Truncated Normal [26] He Uniform [29] Sensitivity Specificity FAs 30.8% 96.9% 7 30.3% 96.5% 8 31.0% 94.2% 13 29.5% 92.4% 18 31.8% 92.1% 19 31.8% 92.1% 19 31.3% 91.1% 22 30.2% 90.0% 25 31.6% 87.8% 31 29.2% 85.1% 40 Initialization L1/L2 Dropout Gaussian L2 L1 Sensitivity 30.8% 30.8% 30.8% 30.2% 30.0% Specificity FAs 97.1% 6 96.9% 7 95.8% 9 95.6% 10 43.7% 276 Table 3. Recognition results for convolutional LSTM architecture, for sensitivity in range of 30%, using different regularizations. Table 4. Results for CNN/LSTM, for sensitivity in the range of 30%, using different initialization methods. Figure 3. DET curves for the CNN/LSTM architecture using different regularizations. SUMMARY Figure 4. DET curves for the CNN/LSTM architecture using different initialization methods. Overfitting is a serious problem in deep neural nets with many parameters. In this study, we used five popular regularization methods to address this problem. By using L1, L2 and L1/L2 techniques, we apply penalties on layer parameters during optimization [30]. These penalties are incorporated in the loss function that the network optimizes. In an alternative approach, we used dropout to prevents units from co-adapting too much by randomly dropping units and their connections from the neural network during training [30]. Also, we studied the impact of introducing zero-centered Gaussian noise to the network [30]. The results of these experiments are presented in Table 3 along with a DET curve in Figure 3. While generally L1/L2 has the best performance, as we move towards a low FA rate, dropout delivers a lower FA rate. Additionally, we found that the primary error modalities observed were false alarms generated during brief delta range slowing patterns such as intermittent rhythmic delta activity. Our closed-loop experiments showed us that all the regularizing methods presented in Table 3 are playing an important role in increasing the false alarms of slowing patterns. Even though dropout is effective in CNNs, when dropout is placed over kernels it leads to diminished results. To solve this problem, in our future work, an efficient Bayesian convolutional neural network is being explored that places a probability distribution over the CNN’s kernels [31]. This approach offers better robustness to overfitting on small data and show improve the robustness of our training process. In this paper, we investigated two deep learning architectures (LSTM and GRU) for automatic classification of EEGs using CNNs. LSTMs outperformed GRUs. We also studied initialization and regularizations of these networks. In future research, we are designing a more powerful architecture based on reinforcement learning concepts. We are also optimizing regularization and initialization algorithms for these approaches. Our goal is to approach human performance which is in the range of 75% sensitivity with a false alarm rate of 1 per 24 hours [11]. Robust training procedures are needed to make this technology relevant to a wide range of healthcare applications. ACKNOWLEDGEMENTS Research reported in this publication was most recently supported by the National Human Genome Research Institute of the National Institutes of Health under award number U01HG008468. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. This material is also based in part upon work supported by the National Science Foundation under Grant No. IIP-1622765. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. The TUH EEG Corpus work was funded by (1) the Defense Advanced Research Projects Agency (DARPA) MTO under the auspices of Dr. Doug Weber through the Contract No. D13AP00065, (2) Temple University’s College of Engineering and (3) Temple University’s Office of the Senior Vice-Provost for Research. REFERENCES [1] I. Obeid and J. Picone, “Machine Learning Approaches to Automatic Interpretation of EEGs,” in Biomedical Signal Processing in Big Data, 1st ed., E. Sejdik and T. Falk, Eds. Boca Raton, Florida, USA: CRC Press, 2017, p. N/A. [2] Lopez, S., Suarez, G., Jungries, D., Obeid, I., & Picone, J. (2015). Automated Identification of Abnormal EEGs. In IEEE Signal Processing in Medicine and Biology Symposium (pp. 1–4). Philadelphia, Pennsylvania, USA. B. Direito, C. Teixeira, B. Ribeiro, M. Castelo-Branco, F. Sales, and A. Dourado, “Modeling epileptic brain states using EEG spectral analysis and topographic mapping,” J. Neurosci. Methods, vol. 210, no. 2, pp. 220–229, 2012. [3] [4] [5] A. Temko, E. Thomas, W. Marnane, G. Lightbody, and G. Boylan, “EEG-based neonatal seizure detection with Support Vector Machines,” Clin. Neurophysiol., vol. 122, no. 3, pp. 464– 473, 2011. [20] [21] [22] [23] [24] J. Gotman, “Automatic recognition of epileptic seizures in the EEG,” Electroencephalogr. Clin. Neurophysiol., vol. 54, no. 5, pp. 530–540, Nov. 1982. [25] [6] J. Gotman, “Automatic detection of seizures and spikes.,” J. Clin. Neurophysiol., vol. 16, no. 2, pp. 130–140, 1999. [26] [7] C. J. Stam, “Nonlinear dynamical analysis of EEG and MEG: Review of an emerging field,” Clinical Neurophysiology, vol. 116, no. 10. pp. 2266–2301, 2005. [8] T. Alotaiby, S. Alshebeili, T. Alshawi, I. Ahmad, and F. Abd ElSamie, “EEG seizure detection and prediction algorithms: a survey,” EURASIP J. Adv. Signal Process., vol. 2014, no. 1, pp. 1–21, 2014. S. Ramgopal, “Seizure detection, seizure prediction, and closedloop warning systems in epilepsy,” Epilepsy Behav., vol. 37, pp. 291–307, 2014. A. Varsavsky and I. Mareels, “Patient un-specific detection of epileptic seizures through changes in variance,” in Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2006, pp. 3747–3750. C. B. Swisher, C. R. White, B. E. Mace, and K. E. Dombrowski, “Diagnostic Accuracy of Electrographic Seizure Detection by Neurophysiologists and Non-Neurophysiologists in the Adult ICU Using a Panel of Quantitative EEG Trends,” J. Clin. Neurophysiol., vol. 32, no. 4, pp. 324–330, Jan. 2015. I. Obeid and J. Picone, “The Temple University Hospital EEG Data Corpus,” Front. Neurosci. Sect. Neural Technol., vol. 10, p. 196, 2016. M. Golmohammadi et al., “The TUH EEG Seizure Corpus,” in Proceedings of the American Clinical Neurophysiology Society Annual Meeting, 2017, p. 1. M. Golmohammadi, S. Ziyabari, V. Shah, I. Obeid, and J. Picone, “Deep Architectures for Automated Seizure Detection in Scalp EEGs,” in Proceedings of the AAAI Conference on Artifical Intelligence, 2018, pp. 1–8. Y. Bengio, P. Simard, and P. Frasconi, “Learning Long-Term Dependencies with Gradient Descent is Difficult,” IEEE Trans. Neural Networks, vol. 5, no. 2, pp. 157–166, 1994. R. Pascanu, T. Mikolov, and Y. Bengio, “On the difficulty of training recurrent neural networks,” in International Conference on Machine Learning, 2013, no. 2, pp. 1310–1318. S. Hochreiter, S. Hochreiter, J. Schmidhuber, and J. Schmidhuber, “Long short-term memory.,” Neural Comput., vol. 9, no. 8, pp. 1735–80, 1997. J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling,” arXiv Prepr. arXiv1412.3555, pp. 1–9, 2014. A. Graves and J. Schmidhuber, “Framewise phoneme classification with bidirectional LSTM networks,” in Proceedings [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [27] [28] [29] [30] [31] of the International Joint Conference on Neural Networks, 2005, vol. 4, pp. 2047–2052. K. Cho et al., “Learning phrase representations using rnn encoderdecoder for statistical machine translation,” arXiv Prepr. arXiv1406.1078, 2014. Z. Wu and S. King, “Investigating gated recurrent networks for speech synthesis,” in International Conference on Acoustics, Speech and Signal Processing, 2016. A. Harati, M. Golmohammadi, S. Lopez, I. Obeid, and J. Picone, “Improved EEG Event Classification Using Differential Energy,” in Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium, 2015, pp. 1–4. S. Lopez, M. Golammadi, I. Obeid, and J. Picone, “An Analysis of Two Common Reference Points for EEGs,” in Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium, 2016, pp. 1–4. D. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs),” arXiv Prepr. arXiv1511.07289, pp. 1–14, 2016. S. B. Wilson, M. L. Scheuer, C. Plummer, B. Young, and S. Pacia, “Seizure detection: correlation of human experts,” Clin. Neurophysiol., vol. 114, no. 11, pp. 2156–2164, Nov. 2003. A. M. Saxe, J. L. McClelland, and S. Ganguli, “Exact solutions to the nonlinear dynamics of learning in deep linear neural networks,” in International Conference on Learning Representations, 2014, pp. 1–22. Y. A. LeCun, L. Bottou, G. B. Orr, and K. R. Mulller, “Efficient backprop,” Lect. Notes Comput. Sci, 2012, pp. 1–44. X. Glorot and Y. Bengio, “Understanding the difficulty of training deep feedforward neural networks,” Proc. 13th Int. Conf. Artif. Intell. Stat., pp. 1–8, 2010. K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: Surpassing human-level performance on imagenet classification,” in Proceedings of the IEEE International Conference on Computer Vision, pp. 1–11, 2015. N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: A Simple Way to Prevent Neural Networks from Overfitting,” J. Mach. Learn. Res., vol. 15, pp. 1929–1958, 2014. Y. Gal and Z. Ghahramani, “Bayesian convolutional neural networks with bernoulli approximate variational inference,” arXiv Prepr. arXiv:1506.02158, pp. 1–12, 2016.	2cs.AI
NodeTrix Planarity Testing with Small Clusters Emilio Di Giacomo1 , Giuseppe Liotta1 , Maurizio Patrignani2 , Alessandra Tappini1 arXiv:1708.09281v2 [cs.DS] 4 Sep 2017 1 Università degli Studi di Perugia, Italy {emilio.digiacomo,giuseppe.liotta}@unipg.it alessandra.tappini@studenti.unipg.it 2 Roma Tre University, Italy patrigna@dia.uniroma3.it Abstract. We study the NodeTrix planarity testing problem for flat clustered graphs when the maximum size of each cluster is bounded by a constant k. We consider both the case when the sides of the matrices to which the edges are incident are fixed and the case when they can be arbitrarily chosen. We show that NodeTrix planarity testing with fixed 3 sides can be solved in O(k3k+ 2 n3 ) time for every flat clustered graph that can be reduced to a partial 2-tree by collapsing its clusters into single vertices. In the general case, NodeTrix planarity testing with fixed sides can be solved in O(n3 ) time for k = 2, but it is NP-complete for any k ≥ 3. NodeTrix planarity testing remains NP-complete also in the free side model when k > 4. 1 Introduction Motivated by the need of visually exploring non-planar graphs, hybrid planarity is one of the emerging topics in graph drawing (see, e.g., [1,2,3,8]). A hybrid planar drawing of a non-planar graph suitably represents in restricted geometric regions those dense subgraphs for which a classical node-link representation paradigm would not be visually effective. These regions are connected by edges that do not cross each other. Different representation paradigms for the dense subgraphs give rise to different types of hybrid planar drawings. Angelini et al. [1] consider hybrid planar drawings where dense portions of the graph are represented as intersection graphs of sets of rectangles and study the complexity of testing whether a non-planar graph admits such a representation. In the context of social network analysis, Henry et al. [8] introduce NodeTrix representations, where the dense subgraphs are represented as adjacency matrices. Batagelj et al. [2] study the question of minimizing the size of the matrices in a NodeTrix representation of a graph while guaranteeing the planarity of the edges that connect different matrices. While Batagelj et al. can choose the subgraphs to be represented as matrices, Da Lozzo et al. [3] consider the problem of testing whether a flat clustered graph (i.e. a graph with clusters and no subclusters) admits a NodeTrix planar representation. In the paper of Da Lozzo et al. each cluster must be represented by a different adjacency matrix and the inter-cluster edges are represented as non-intersecting simple Jordan arcs. They prove that NodeTrix planarity testing for flat clustered graphs is NP-hard even in the constrained case where for each matrix it is specified which inter-cluster edges must be incident on the top, or on the left, or on the bottom, or on the right side. Motivated by these hardness results, in this paper we study whether NodeTrix planarity testing can be efficiently solved when the size of the clusters is not “too big”. More precisely, we consider flat clustered graphs whose clusters have size bounded by a fixed parameter k and we want to understand whether the NodeTrix planarity testing problem is fixed parameter tractable, i.e. it can be solved in time f (k)T (n), where T (n) is a polynomial in n and f (k) is a function that depends only on k. Our main results can be listed as follows: 3 – We describe an O(k 3k+ 2 ·n3 )-time algorithm to test NodeTrix planarity with fixed sides for flat clustered graphs that are partial 2-trees. Informally, a flat clustered graph G is a partial 2-tree if the graph obtained by collapsing every cluster of G into a single vertex is a partial 2-tree. – When the flat clustered graph is not a partial 2-tree, NodeTrix planarity testing with fixed sides can still be solved in O(n3 ) time for k = 2, but it becomes NP-complete for any larger value of k. – Finally, we extend the above hardness result to the free sides model and show that NodeTrix planarity testing remains NP-complete when the maximum cluster dimension is larger than four. This is done by proving that NAE3SAT is NP-complete even for triconnected Boolean formulas, which may be a result of independent interest. Our polynomial-time solution solves a special type of the planarity testing problem where the order of the edges around each vertex is suitably constrained to take into account the fact that a vertex of a matrix M has four copies along the four sides of M . It may be worth recalling that Gutwenger et al. [6] have considered an apparently similar problem. Namely, they studied planarity testing where the order of the edges around the vertices may not be arbitrarily permuted. Unfortunately, not only our problem does not fall in any of the cases addressed by Gutwenger et al., but it does not seem solvable by introducing a gadget of polynomial size that models the embedding constraints at each vertex. This characteristic associates NodeTrix planarity testing with other known variants of planarity testing, including clustered planarity, where the use of gadgets of polynomial size has been so far an elusive goal. The rest of the paper is organized as follows. Preliminary definitions are in Section 2. Sections 3 and 4 describe a polynomial time algorithm for clustered 2-trees with bounded cluster-size. In Section 5 we show that for general flat clustered graphs and fixed sides NodeTrix planarity testing can be solved in polynomial time for k = 2 but it is NP-complete for k ≥ 3. In the same section we extend this completeness result to NodeTrix planarity testing of flat clustered graphs with free sides. Finally open problems can be found in Section 6. Some proofs can be found in the appendix. 2 Preliminaries We assume familiarity with basic definitions of graph theory and graph drawing and in particular with the notions of block-cut-vertex tree and of SPQR-tree (see, e.g., [4,7]). A flat clustered graph G = (V, E, C) is a simple graph with vertex set V , edge set E, and a partition C of V into sets V1 , . . . , Vh , called clusters. An edge (u, v) ∈ E with u ∈ Vi and v ∈ Vj is an intra-cluster edge if i = j and it is an inter-cluster edge if i 6= j. A NodeTrix representation of a flat clustered graph G is such that: (i) Each cluster Vi with \|Vi \| = 1 (called trivial cluster) is represented as a distinct point in the plane. (ii) Each cluster Vi with \|Vi \| > 1 (called non-trivial cluster) is represented by a symmetric adjacency matrix Mi (with \|Vi \| rows and columns), where Mi is drawn in the plane so that its boundary is a square with sides parallel to the coordinate axes. (iii) There is no intersection between two distinct matrices or between a point representing a vertex and a matrix. (iv) Each intracluster edge of a cluster Vi is represented by the adjacency matrix Mi . (v) Each inter-cluster edge (u, v) with u ∈ Vi and v ∈ Vj is represented by a simple Jordan arc connecting a point on the boundary of matrix Mi with a point on the boudary of matrix Mj , where the point on Mi (on Mj ) belongs to the column or to the row of Mi (resp. of Mj ) associated with u (resp. with v). A NodeTrix representation of a flat clustered graph G is planar if there is no intersection between any two inter-cluster edges (except possibly at common end-points) nor an intersection between an inter-cluster edge and a matrix. A flat clustered graph is NodeTrix planar if it admits a planar NodeTrix representation. Fig. 1(a) is an example of a NodeTrix planar representation. A formal definition of the problem investigated in the paper is as follows. Let G = (V, E, C) be a flat clustered graph with n vertices and let k be the maximum cardinality of a cluster in C. Clustered graph G is NodeTrix planar with fixed sides if it has a NodeTrix planar representation where for each inter-cluster edge, the sides of matrices it attaches to is specified as part of the input; G is NodeTrix planar with free sides if the sides of the matrices to which inter-cluster edges attach can be arbitrarily chosen. Let Mi be the matrix representing cluster Vi in a NodeTrix representation of G; let v be a vertex of Vi and let (u, v) be an inter-cluster edge. Edge (u, v) can intersect the boundary of Mi in four points pv,t , pv,b , pv,l , and pv,r since the row and column that represent v in Mi intersect the four sides of the boundary of Mi . We call such points the top copy, bottom copy, left copy, and right copy of v in Mi , respectively. A side assignment for Vi ∈ C specifies for each inter-cluster edge whether the edge must attach to the matrix Mi representing Vi in its top, S left, right, or bottom side. More precisely, a side assignment is a mapping φi : j6=i Ei,j → {t, b, l, r}, where Ei,j is the set of inter-cluster edges between the clusters Vi and Vj (Vi and Vj are adjacent if Ei,j 6= ∅). A side assignment for C is a set Φ of side assignments for each Vi ∈ C. We denote as G = (V, E, C, Φ) a flat clustered graph G = (V, E, C) with a given side assignment Φ = {φ1 , φ2 , . . . , φ\|C\| }. Let Γ be a NodeTrix representation of G such that, for every inter-cluster edge e = (u, v) ∈ E with u ∈ Vi and v ∈ Vj , the incidence points of e with the matrices Mi and Mj representing Vi and Vj in Γ are exactly the points pu,φi (e) and pv,φj (e) , respectively. We call Γ a NodeTrix representation of G consistent with Φ. We say that G = (V, E, C, Φ) is NodeTrix planar if it admits a NodeTrix planar representation consistent with Φ. An inter-cluster edge is heavy if both its end-vertices belong to non-trivial clusters. It is light otherwise. A flat clustered graph is light if every intercluster edge is light. A 1-subdivision of a heavy edge e = (u, v) of a flat clustered graph G = (V, E, C) replaces e with a path u0 = u, u1 , u2 = v and defines a new flat clustered graph G0 = (V 0 , E 0 , C 0 ), where V 0 = V ∪ {u1 }, E 0 = E/e ∪ {(u0 , u1 ), (u1 , u2 )}, and C 0 = C ∪ {u1 }. The light reduction of G is the flat clustered graph G0 obtained by performing a 1-subdivision of every heavy inter-cluster edge of G. A consequence of Theorem 1 in [5] about the edge density of NodeTrix planar graphs, is that the light reduction G0 of a NodeTrix planar flat clustered graph G has O(\|V \|) vertices and O(\|V \|) inter-cluster edges. Property 1. A flat clustered graph G is NodeTrix planar if and only if its light reduction G0 is NodeTrix planar. Based on Property 1, in the remainder we shall assume that flat clustered graphs are always light and we call them clustered graphs, for short. The frame of a clustered graph G = (V, E, C) is the graph F obtained by collapsing each cluster Vi ∈ C, with \|Vi \| > 1, into a single vertex ci of F , called the representative vertex of Vi in F . Let ci and cj be the two representative vertices of Vi and Vj in F , respectively. For every inter-cluster edge connecting a vertex of Vi to a vertex of Vj in G there is an edge in F connecting ci and cj . Observe that the frame graph F of G is in general a multigraph; however, F is simple when G is light. Since the NodeTrix planarity of a clustered graph implies the planarity of its frame graph, we will test NodeTrix planarity only on those clustered graphs that have a planar frame. A 2-tree is a graph recursively defined as follows: (i) an edge is a 2-tree; (ii) the graph obtained by adding a vertex v to a 2-tree G and by connecting v to two adjacent vertices of G is a 2-tree. A (planar) graph is a partial 2-tree if it is a subgraph of a (planar) 2-tree. A biconnected partial 2-tree is a series-parallel graph. A clustered graph is a partial 2-tree if its frame is a partial 2-tree. We will sometimes talk about series-parallel clustered graphs when their frames are series parallel. 3 NodeTrix Representations and Wheel Reductions The polynomial-time algorithms described in Sections 4 and 5 are based on decomposing the planar frame F of a clustered graph G = (V, E, C, Φ) into its biconnected components and storing them into a block-cut-vertex tree. We process each block of F by using an SPQR decomposition tree that is rooted at a reference edge and visited from the leaves to the root. For each visited node µ of the decomposition tree of a block of F we test whether the subgraph of G whose frame is the pertinent graph of µ satisfies the planar constraints imposed by the side assignment on the inter-cluster edges. A key ingredient to efficiently perform the test at µ is the notion of wheel replacement. Let G = (V, E, C, Φ) be a clustered graph with side assignment Φ and let Vi ∈ C be a cluster with k > 1 vertices. Vi admits k! permutations of its vertices and we associate a suitable graph to each such permutation. Let πi = v0 , v1 , . . . , vk−1 be a permutation of the vertices of Vi . The wheel of Vi consistent with πi is the wheel graph consisting of a vertex v of degree 4k adjacent to the vertices of an oriented cycle v0,t , v1,t , . . . , vk−1,t , v0,r , v1,r , . . . , vk−1,r , vk−1,b , vk−2,b , . . . , v0,b , vk−1,l , vk−2,l , . . . , v0,l where each edge of the cycle is oriented forward. Intuitively, this oriented cycle will be embedded clockwise to encode the constraints induced by a matrix Mi representing Vi when its left-to-right order of columns is πi . More precisely, a wheel replacement of cluster Vi consistent with πi is the clustered graph obtained as follows: (i) remove Vi and all the inter-cluster edges incident to Vi ; (ii) insert the wheel Wi of Vi consistent with πi ; and (iii) for each inter-cluster edge e = (u, vj ), with vj ∈ Vi , insert edge (u, vj,φi (e) ) incident to Wi . We call edge (u, vj,φi (e) ) the image of edge e = (u, vj ). Let G = (V, E, C, Φ, Π) be a clustered graph with side assignment Φ where Π is a set of permutations {π1 , π2 , . . . , π\|C\| }, one for each cluster Vi (with i = 1, . . . , \|C\|). We call Π the permutation assignment of G and we say that G is NodeTrix planar with side assignment Φ and permutation assignment Π if G admits a NodeTrix planar representation with side assignment Φ where for each matrix Mi the permutation of its columns is πi . The wheel reduction of G consistent with Π is the graph obtained by performing a wheel replacement of Vi ∈ C consistent with πi for each i = 1, . . . , \|C\|. Theorem 1. Let G = (V, E, C, Φ, Π) be a clustered graph with side assignment Φ and permutation assignment Π. G is NodeTrix planar if and only if the planar wheel reduction of G admits a planar embedding where the external oriented cycle of each wheel Wi is embedded clockwise. Fig. 1(a) and Fig. 1(b) show respectively a NodeTrix planar representation of a clustered graph G and the corresponding wheel reduction with its planar embedding. Based on Theorem 1, we can test the graph G = (V, E, C, Φ) for NodeTrix planarity by exploring the space of the possible permutation sets Π and corresponding wheel reductions in search of a NodeTrix planar G = (V, E, C, Φ, Π). Note that, if the maximum size of a cluster is given as a parameter k, every cluster Vi can be replaced by k! wheel graphs, one for each possible permutation of the vertices of Vi . In order to test planarity, for any such wheel replacement Wi , the cyclic order of the inter-cluster edges incident to the same vertex of Wi can be arbitrarily permuted. While each wheel reduction yields an instance of constrained planarity testing that can be solved with the linear-time algorithm u2 v2,L u0,L v1,R u1,L v2,R u2,L vo id vo id vo id ,T ,T u2 ,T v0,R u1 u0 ,T ,T ,T v2 v v1,L u1 v2 Wtµ v0,L u0 v1 v1 v0 v0 u u0,R void u1,R void u2,R void Wtµ void v ext int-ext id vo t in id vo ,B ,B u2 ,B u1 u0 ,B ,B ,B v2 v1 v0 2, T 1, T w w w 0, T µ w0,L w0 w0,R w w1,L w1 w2 w2,L w2,R ,B ,B ,B w2 w1 w0 (a) w1,R Ws µ (b) (c) Fig. 1. (a) A NodeTrix planar representation of a clustered graph. (b) The planar embedding of the corresponding wheel reduction. (c) Labeling of the vertices of Wtµ ; the complete internal and external sequences are highlighted. described in [6], a brute-force approach that repeats this algorithm on each possible wheel reduction may lead to testing planarity on \|C\|k! different instances. Instead, for each visited node µ of the decomposition tree T we compute a succinct description of the possible NodeTrix planar representations of the subgraph Gµ of G represented by the subtree of T rooted at µ. This is done by storing for the poles of µ those pairs of wheel graphs that are compatible with a NodeTrix planar representation of Gµ . How to efficiently compute such a succinct description will be the subject of the next sections. 4 Testing NodeTrix Planarity for Partial 2-Trees In this section we prove that NodeTrix planarity testing with fixed sides can be solved in polynomial time for a clustered graph G = (V, E, C, Φ) when the maximum size of any cluster of C is bounded by a constant and the frame graph is a partial 2-tree. This contrasts with the NP-hardness of NodeTrix planarity testing with fixed sides proved in [3] in the case where the size of the clusters is unbounded. We first study the case of a clustered graph whose frame graph is a seriesparallel graph, i.e., it is biconnected and its SPQR decomposition tree only has Q-, P-, and S-nodes. We then consider the case of partial 2-trees, i.e., graphs whose biconnected components are series-parallel. 4.1 Series-Parallel Frame Graphs In this section we prove that NodeTrix planarity testing with fixed sides can 3 be solved in O(k 3k+ 2 · n2 ) time for clustered graphs whose frame graphs are series-parallel and have cluster size at most k. Let G = (V, E, C, Φ) be a series-parallel clustered graph with side assignment Φ and let F be its frame graph. Let T be the SPQ decomposition tree of F rooted at any Q-node. To simplify the description and without loss of generality, we assume that every S-node of T has exactly two children. Let µ be a node of T , and let sµ , tµ be the poles of µ. Consider the pertinent graph Fµ represented by the subtree of T rooted at µ and let vµ be a pole of µ (vµ ∈ {sµ , tµ }). Pole vµ in the frame graph F may correspond to a non-trivial cluster Vi of C. In this case, we call vµ a non-trivial pole of µ and cluster Vi the pertinent cluster of vµ . The edges of Fµ incident to vµ are the intra-component edges of vµ . The other edges of F incident to vµ are the extra-component edges of vµ . Each intracomponent (extra-component) edge of vµ corresponds to an inter-cluster edge e0 of G incident to one vertex of the pertinent cluster Vµ of vµ . We call e0 an intracomponent edge (extra-component edge) of Vµ . We associate k! wheel graphs to each non-trivial pole vµ of µ. Each of them is a wheel replacement of the pertinent cluster of vµ , consistent with one of the k! permutations of its vertices. Let vµ be a non-trivial pole of µ, let Vµ be the pertinent cluster of vµ , let πµ be a permutation of the vertices of Vµ , and let Wµ be the wheel replacement of Vµ consistent with πµ . Every edge e incident to Wµ such that e is the image of an inter-cluster edge e0 of G is labeled either int or ext, depending on whether e0 is an intra-component or an extra-component edge of Vµ . A vertex w of the external cycle of Wµ is assigned one label of the set {void, int, ext, int-ext} as follows. Vertex w is labeled void if no edge incident to w is the image of an inter-cluster edge. Vertex w is labeled int (resp. ext) if we have a label int (resp. ext) on every edge e incident to w such that e is the image of an inter-cluster edge. Otherwise, vertex w is labeled int-ext. See Fig. 1(c) for an example concerning the wheel Wtµ of Fig. 1(b); the dashed curve of Fig. 1(b) shows the subgraph of the wheel reduction corresponding to Fµ . A clockwise sequence v0 , v1 , . . . , vj of vertices of the external cycle of Wµ is an external sequence of pole vµ consistent with πµ if v0 and vj are labeled either ext or int-ext and all the other vertices of the sequence are labeled either void or ext. An external clockwise sequence of pole vµ is complete if it contains all the vertices of Wµ that are labeled ext and int-ext. Note that a complete external sequence may contain many void vertices but no int vertex. Internal and complete internal sequences of pole vµ are defined analogously. Observe that a complete internal sequence and a complete external sequence of vµ may not exist when vertices labeled int and vertices labeled ext alternate more than twice when traversing clockwise the external cycle of Wµ , or when three vertices are labeled int-ext. A special case is when Wµ has exactly two vertices w1 and w2 labeled int-ext and all other vertices are void. In this case, the clockwise sequence from w1 to w2 and the clockwise sequence from w2 to w1 are both complete internal and complete external sequences. In order to test G = (V, E, C, Φ) for NodeTrix planarity, we implicitly take into account all possible permutation assignments Π by considering, for each non-trivial pole wµ of each node µ of T , its k! possible wheels and by computing their complete internal and complete external sequences. We visit the SP Q decomposition tree T from the leaves to the root and equip each node µ of T with information regarding the complete internal and complete external sequences of its non-trivial poles. Let µ be an internal node of T , let vµ be a non-trivial pole of µ, let πvµ be a permutation of the pertinent cluster Vµ of vµ , and let Wµ be the wheel of Vµ consistent with πvµ . We denote as ISeq(µ, vµ , πvµ ) the complete internal sequence of vµ consistent with πvµ in pole µ and as ESeq(µ, vµ , πvµ ) the complete external sequence of vµ consistent with πvµ in pole µ. We distinguish between the different types of nodes of T . Node µ is a Q-node. Since G is light, at most one of its poles is non-trivial. Let e be an edge of F that is the pertinent graph of µ. One end-vertex of e is the representative vertex in F of the pertinent cluster of the non-trivial pole vµ . In fact, edge e corresponds to an edge e0 = (u, z) of G such that u ∈ Vµ and z is a trivial cluster. The side assignment φvµ defines whether e is incident to the top, bottom, left, or right copy uW of u in the wheel Wµ of Vµ . For any possible permutation πvµ we have ISeq(µ, vµ , πvµ ) = uW . If uW is labeled int-ext, then ESeq(µ, vµ , πvµ ) is the external cycle of Wµ starting at uW and ending at uW . Otherwise, traverse the external cycle of Wµ starting at uW and following the direction of the edges; ESeq(µ, vµ , πvµ ) consists of all the encountered vertices from the first labeled ext to the last labeled ext. Node µ is a P-node. Let ν0 , ν1 , . . . , νh−1 be the children of µ. Observe that vµ is a non-trivial pole also for the children ν0 , ν1 , . . . , νh−1 of µ. We consider every permutation πvµ such that ν0 , ν1 , . . . , νh−1 have both a complete internal sequence and a complete external sequence compatible with πvµ . The complete internal sequence of vµ consistent with πvµ is the union of the complete internal sequences of the children ν0 , ν1 , . . . , νh−1 , that is ISeq(µ, vµ , πvµ ) = ∪h−1 i=0 ISeq(νi , vµ , πvµ ). To determine the complete external sequence of vµ consistent with πvµ we consider the intersection of the complete external sequences of the children of µ. If this intersection consists of exactly one sequence of consecutive vertices, h−1 then ESeq(µ, vµ , πvµ ) = ∩i=0 ESeq(νi , vµ , πvµ ). Otherwise (i.e., the intersection is empty or it consists of more than one sequence of consecutive vertices), vµ does not have a complete external sequence consistent with πvµ . Node µ is an S-node. Let ν be the child of µ that shares the pole vµ with µ. We consider every permutation πvµ such that ν has both ISeq(ν, vµ , πvµ ) and ESeq(ν, vµ , πvµ ). The complete internal (external) sequence of vµ consistent with πvµ is ISeq(µ, vµ , πvµ ) = ISeq(ν, vµ , πvµ ) (ESeq(µ, vµ , πvµ ) = ESeq(ν, vµ , πvµ )). To test G for NodeTrix planarity we execute a bottom-up traversal of T and, for each node µ with poles sµ and tµ , we check whether each possible pair (πsµ , πtµ ) induces complete internal and external sequences for sµ and tµ that are ‘compatible’ with a planar embedding of the wheel reduction of G. If this is the case, by Theorem 1, G is NodeTrix planar, otherwise we reject G. More formally, let πsµ (πtµ , respectively) be a permutation such that sµ (tµ , respectively) has both a complete internal sequence and a complete external sequence compatible with πsµ (πtµ , respectively). We say that (πsµ , πtµ ) is a compatible pair of permutations for µ if either one of the poles is a trivial pole or one of the following cases applies. Node µ is a Q-node. In this case all k! possible pairs of permutations for sµ or tµ (recall that only one of them is non-trivial) are compatible for µ. Node µ is a P-node. Let ν0 , ν1 , . . . , νh−1 be the children of µ. Consider a pair of permutations (πsµ , πtµ ); we recall that, for i = 0, . . . , h−1, each νi has poles sµ and tµ . A first condition for pair (πsµ , πtµ ) to be a compatible pair for µ is that (πsµ , πtµ ) is also a compatible pair for νi , with i = 0, . . . , h−1. A second condition asks that the pair (πsµ , πtµ ) defines opposite orders on the poles of µ. Namely, let Wµs (resp., Wµt ) be the wheel of Vsµ (resp., Vtµ ) consistent with πsµ (resp., πtµ ). Traversing clockwise the external cycle of Wµs starting from the first vertex of ESeq(µ, sµ , πsµ ), let ISeq(ν0 , sµ , πsµ ), ISeq(ν1 , sµ , πsµ ), . . . , ISeq(νh−1 , sµ , πsµ ) be the order by which the internal sequences are encountered. Pair (πsµ , πtµ ) defines opposite orders on the poles of µ if, traversing clockwise the external cycle of Wµt starting from the first vertex of ESeq(µ, tµ , πsµ ), the order by which we encounter the internal sequences of ν0 , ν1 , . . . , νh−1 is the opposite one, i.e., the order is ISeq(νh−1 , tµ , πtµ ), ISeq(νh−2 , tµ , πtµ ), . . . , ISeq(ν0 , tµ , πtµ ). Node µ is an S-node. Let ν0 and ν1 be the children of µ such that sν0 = sµ , tν0 = sν1 , and tν1 = tµ . A pair (πsµ , πtµ ) is a compatible pair for µ if there exists a permutation πtν0 such that the pair (πsµ , πtν0 ) is compatible for ν0 and the pair (πtν0 , πtµ ) is compatible for ν1 . Fig. 2 suggests that a NodeTrix planar representation of a clustered graph G defines a permutation assignment Π such that, for every node µ of T , pair (πsµ , πtµ ) is a compatible pair for µ. Lemma 1. Let G = (V, E, C, Φ) be a clustered graph with side assignment Φ and let T be the SPQ decomposition tree of the frame graph of G. Graph G is NodeTrix planar if and only if there exists a permutation assignment Π such that, for every node µ of T with poles sµ and tµ , we have that permutation πsµ ∈ Π and permutation πtµ ∈ Π form a compatible pair of permutations for µ. Lemma 2. Let G = (V, E, C, Φ) be a series-parallel clustered graph with side assignment Φ. Let k be the maximum size of any cluster in C and let n be the 3 cardinality of V . There exists an O(k 3k+ 2 · n2 )-time algorithm that tests whether G is NodeTrix planar with side assignment Φ and if so, it computes a NodeTrix planar representation of G consistent with Φ. Proof. Let F be the frame graph of G; for any possible choice of an edge e of F we repeat the following procedure. We construct the SPQ decomposition tree of G rooted at the Q-node whose pertinent graph is e. We visit T from the leaves to the root and test whether G has a permutation assignment Π such that G = (V, E, C, Φ, Π) is NodeTrix planar. We first equip each non-trivial pole vµ of every node µ of T with its possible complete internal and complete external sequences. The maximum number of complete internal sequences of vµ is k!. The same is true for the complete external sequences. If each complete (internal or Wtµ µ Wsµ (a) (b) Fig. 2. (a) A NodeTrix planar representation Γ of G = (V, E, C, Φ). (b) Γ induces a permutation assignment and a planar embedding of a wheel reduction of G; the complete internal and external sequences for a pair of poles are also highlighted. external) sequence of pole vµ is encoded by means of its first and last vertex in the clockwise order around Wvµ , then each complete internal or external sequence needs constant space. It follows that the intersection or the union of two complete internal or external sequences can be computed in constant time. Therefore, all complete internal and external sequences for each non-trivial pole of T can be computed in O(k!) time. Hence, the whole bottom-up traversal to equip all non-trivial poles with every possible complete internal/external sequence can be executed in O(k! · n) time. We now test whether there exists a permutation assignment Π such that any node µ of T has a compatible pair of permutations. To this aim, we look at the complete internal and external sequences for the pair of poles of the children of µ. For each pair (πsµ , πtµ ) of permutations of the poles of µ we equip µ with the information about whether such pair is compatible for µ. This requires O(k!2 ) space. If µ is a Q-node, every pair of permutations (πsµ , πtµ ) is compatible for µ. It follows that all compatible pairs for µ can be computed in O(k!) time (recall that one between sµ and tµ is non-trivial) and, hence, in O(k! · n) time for all the Q-nodes of T . If µ is a P-node with children ν0 , ν1 , . . . , νh−1 , πsµ is one of the permutations that equip sµ , and πtµ is one of the permutations that equip tµ , testing whether the pair (πsµ , πtµ ) is a compatible pair for µ can be executed in O(h) time. It follows that all compatible pairs for µ can be computed in O(k!2 · h) time and, hence, in O(k!2 · n) time for all P-nodes of T . If µ is an S-node with children ν0 and ν1 , πsµ is one of the permutations that equip sµ , and πtµ is one of the permutations that equip tµ , testing whether the pair (πsµ , πtµ ) is a compatible pair for µ can be executed in O(k!) time, corresponding to choosing all possible permutations for the pole shared between ν0 and ν1 . It follows that all compatible pairs for µ can be computed in O(k!3 ) time and, hence, in O(k!3 · n) time for all S-nodes of T . In conclusion, the time complexity of a bottom-up visit of T rooted at e is O(k!3 · n). By rooting T at all possible Q-nodes, we obtain√an overall time complexity of O(k!3 · n2 ). By Stirling’s approximation, k! ∼ 2πk( ke )k and thus a series-parallel clustered graph G with n vertices, side assignment Φ, and maxi3 mum cluster size k can be tested for NodeTrix planarity in O(k 3k+ 2 · n2 ) time. Note that the compatible pair of permutations stored at each node µ of T implicitly define a planar embedding of a wheel reduction of G. It can be shown that it is possible to construct a NodeTrix planar representation of G in time proportional to the number of edges of G, which is O(n · k) [5]. The statement of the lemma follows. 4.2 Partial 2-Trees We now consider clustered graphs whose cluster size is at most k and such that their frame graph is a partial 2-tree, i.e. it is a planar graph whose biconnected components are series-parallel. We handle this case by decomposing the frame graph into its blocks and we store them into a block-cut-vertex tree. The following theorem generalizes the result of Lemma 2. Theorem 2. Let G = (V, E, C, Φ) be a partial 2-tree clustered graph with side assignment Φ. Let k be the maximum size of any cluster in C and let n be the 3 cardinality of V . There exists an O(k 3k+ 2 · n3 )-time algorithm that tests whether G is NodeTrix planar with side assignment Φ and if so, it computes a NodeTrix planar representation of G consistent with Φ. 5 General Planar Frame Graphs In this section we study the problem of extending Theorem 2 to planar frame graphs that may not be partial 2-trees. We prove that NodeTrix planarity testing with fixed sides can be solved in polynomial time for maximum cluster size k = 2. However, the problem becomes NP-complete with fixed sides for k ≥ 3 and it remains NP-complete even in the free sides scenario for k ≥ 5. Every block of the frame graph can be decomposed into its triconnected components by means of an SPQR decomposition tree. For each block, we adopt the same approach as for series-parallel graphs and look for a permutation assignment Π such that, for every pair of poles sµ and tµ , (πsµ , πtµ ) forms a compatible pair for µ when µ is either a Q-node, a P-node, or an S-node. We extend the definition of compatible pairs of permutations for an R-node as follows. Let G = (V, E, C, Φ) be a clustered graph with side assignment Φ, let F be the frame graph of G, and let T be the SPQR decomposition tree of F . Let µ be an R-node of T with poles sµ and tµ . A pair of permutations (πsµ , πtµ ) forms a compatible pair for µ if there exists a planar embedding of the skeleton skel(µ) of µ for which the following conditions hold: (i) For each vertex v of skel(µ), let e0 , e1 , . . . eh−1 be the virtual edges of skel(µ) incident to v in clockwise order around v. Each such edge ei is associated with a child νi of µ. There exists a permutation πv such that the complete internal sequences ISeq(ν0 , v, πv ), ISeq(ν1 , v, πv ), . . . , ISeq(νh−1 , v, πv ) appear in this clockwise order around v. (ii) Every vertex v of skel(µ) can be assigned a permutation πv such that: πv = πsµ if v = sµ and πv = πsµ if v = tµ , and for each virtual edge e = (u, v) in skel(µ) that corresponds to a child ν of µ, the permutation pair (πu , πv ) is compatible for ν. Observe that, in the case of maximum cluster size k = 2, the possible permutations of the induced cluster Vv of a vertex v of skel(µ) are exactly two, denoted by πv+ and πv− . In order to test whether (πsµ , πtµ ) forms a compatible pair for µ, we perform a traversal of skel(µ) starting at sµ . Permutation πsµ and the clockwise order of the edges incident to sµ can impose to choose only one of + − the two permutations πw o πw available for each vertex w adjacent to sµ and corresponding to a non-trivial cluster of G. Each such w and its incident edges, in turn, propagate constraints on the possible permutations to their neighbors, till tµ is reached. Therefore, testing whether πsµ and πtµ form a compatible pair for µ can be reduced to a suitable problem of labeling the edges and vertices of skel(µ) and verifying that at the end sµ and tµ are labeled with πtµ and πsµ . Theorem 3. Let G = (V, E, C, Φ) be an n-vertex clustered graph with side assignment Φ such that the maximum size of any cluster in C is two. There exists an O(n3 )-time algorithm that tests whether G is NodeTrix planar with the given side assignment and if so, computes a NodeTrix planar representation of G consistent with Φ. The proof of the following theorem is based on a reduction from (non-planar) NAE3SAT. Theorem 4. NodeTrix planarity testing with fixed sides and cluster size at most k is NP-complete for any k ≥ 3. Now, we extend the above hardness result to the free sides model and show that NodeTrix planarity testing remains NP-complete when the maximum cluster dimension is larger than four. This is done by proving that NAE3SAT is NP-complete even for triconnected Boolean formulas, which may be a result of independent interest. Theorem 5. NAE3SAT is NP-complete for triconnected Boolean formulas. Theorem 6. NodeTrix planarity testing with free sides and cluster size at most k is NP-complete for any k ≥ 5. 6 Open Problems We conclude the paper by listing some open problems that, in our opinion, are worth investigating. (i) Study the complexity of NodeTrix planarity testing in the free sides scenario for values of k between 2 and 5. (ii) Study families of clustered graphs for which NodeTrix planarity testing is fixed parameter tractable in the free sides scenario. (iii) Determine whether the time complexity of the algorithms in Theorems 2 and 3 can be improved. References 1. Angelini, P., Da Lozzo, G., Di Battista, G., Frati, F., Patrignani, M., Rutter, I.: Intersection-link representations of graphs. In: Di Giacomo, E., Lubiw, A. (eds.) Proceedings of 23rd International Symposium on Graph Drawing and Network Visualization (GD ’15). Lecture Notes in Computer Science, vol. 9411, pp. 217– 230 (2015) 2. Batagelj, V., Brandenburg, F., Didimo, W., Liotta, G., Palladino, P., Patrignani, M.: Visual analysis of large graphs using (X,Y)-clustering and hybrid visualizations. IEEE Trans. Vis. Comput. Graph. 17(11), 1587–1598 (2011) 3. Da Lozzo, G., Di Battista, G., Frati, F., Patrignani, M.: Computing NodeTrix representations of clustered graphs. In: Nöllenburg, M., Hu, Y. (eds.) Graph Drawing and Network Visualization (GD ’16). Lecture Notes in Computer Science, vol. 9801, pp. 107–120 (2016) 4. Di Battista, G., Eades, P., Tamassia, R., Tollis, I.G.: Graph Drawing. Prentice Hall, Upper Saddle River, NJ (1999) 5. Di Giacomo, E., Liotta, G., Patrignani, M., Tappini, A.: Planar k-NodeTrix Graphs - a new family of beyond planar graphs. In: Frati, F., Ma, K. (eds.) Graph Drawing and Network Visualization (GD 2017). LNCS, Springer, to appear 6. Gutwenger, C., Klein, K., Mutzel, P.: Planarity testing and optimal edge insertion with embedding constraints. J. Graph Algorithms Appl. 12(1), 73–95 (2008) 7. Harary, F.: Graph Theory. Addison-Wesley Series in Mathematics, Addison Wesley (1969) 8. Henry, N., Fekete, J., McGuffin, M.J.: NodeTrix: a hybrid visualization of social networks. IEEE Trans. Vis. Comput. Graph. 13(6), 1302–1309 (2007) 9. Kratochvı́l, J.: A special planar satisfiability problem and a consequence of its NP-completeness. Discrete Applied Mathematics 52(3), 233 – 252 (1994) 10. Moret, B.M.E.: Planar NAE3SAT is in P. SIGACT News 19(2), 51–54 (1988) 11. Schaefer, T.J.: The complexity of satisfiability problems. In: Proceedings of the 10th Annual ACM Symposium on Theory of Computing. pp. 216–226 (1978) A A.1 Omitted Proofs Proof of Theorem 1 Theorem 1. Let G = (V, E, C, Φ, Π) be a clustered graph with side assignment Φ and permutation assignment Π. G is NodeTrix planar if and only if the planar wheel reduction of G admits a planar embedding where the external oriented cycle of each wheel Wi is embedded clockwise. Proof. If G is NodeTrix planar, we construct a planar embedding of a wheel reduction of G where the external oriented cycle of each wheel is embedded clockwise as follows. Let Γ be a NodeTrix planar representation of G. We replace each matrix Mi representing a cluster Vi ∈ C with the wheel Wi of Vi consistent with the permutation πi ∈ Π of Vi . Also Wi is embedded in such a way that a forward traversal of its external cycle is a clockwise traversal of the cycle. Every inter-cluster edge e = (u, vj ), with vj ∈ Vi , is incident to the vertex vj,φi (e) of the wheel Wi . Also, for all j = 0, 1, . . . , k − 1 and for all x ∈ {t, b, l, r}, the cyclic order of the inter-cluster edges incident to vj,x in Wi is the same as the cyclic order of the inter-cluster edges incident to pvj ,x in Mi . It is immediate to see that, since no two inter-cluster edges cross in Γ , no two edges cross in the constructed embedding of the wheel reduction of G. Conversely, suppose that we are given a planar embedding of the wheel reduction of G where the external oriented cycle of each wheel is embedded clockwise. We show how to construct a NodeTrix planar representation of G. For each wheel Wi we remove the center vertex of the wheel and insert a matrix Mi inside the created face. We now morph every vertex vj,φi (e) of the external cycle of Wi to point pvj ,φi (e) in Mi and maintain around pvj ,φi (e) the cyclic order of the inter-cluster edges incident to vj,φi (e) in the planar embedding of the wheel reduction. A.2 Proof of Lemma 1 Lemma 1. Let G = (V, E, C, Φ) be a clustered graph with side assignment Φ and let T be the SPQ decomposition tree of the frame graph of G. Graph G is NodeTrix planar if and only if there exists a permutation assignment Π such that, for every node µ of T with poles sµ and tµ , we have that permutation πsµ ∈ Π and permutation πtµ ∈ Π form a compatible pair of permutations for µ. Proof. We prove first that, if G = (V, E, C, Φ) is NodeTrix planar, then there exists a permutation assignment Π such that, for every node µ of T with poles sµ and tµ , the pair (πsµ , πtµ ) is compatible for µ. Let Γ be a NodeTrix planar representation of G with side assignment Φ and let M0 , M1 , . . . , Mh−1 be the matrices representing the non-trivial clusters of G. For each matrix Mi (i = 0, . . . , h − 1) of Γ , let πi = v0 , v1 , . . . , vk−1 be the left to right order of the columns of Mi . We replace Mi with a wheel Wi consisting of a vertex wi of degree 4k adjacent to all vertices of a cycle v0,t , v1,t , . . . , vk−1,t , v0,r , v1,r , . . . , vk−1,r , vk−1,b , vk−2,b , . . . , v0,b , vk−1,l , vk−2,l , . . . , v0,l . For all j = 0, 1, . . . , k − 1 and for all x ∈ {t, b, l, r}, vertex vj,x is drawn at the point pvj ,x , that represents the attachment of the inter-cluster edges incident to vertex vj on the side x of matrix Mi . The edges of the external cycle of Wi are drawn along the external boundary of Mi . Every inter-cluster edge e = (u, vj ), with vj ∈ Mi , is incident to the vertex vj,φi (e) of the wheel Wi . Also, for all j = 0, 1, . . . , k − 1 and for all x ∈ {t, b, l, r}, the cyclic order of the inter-cluster edges incident to vj,x in Wi is the same as the cyclic order of the inter-cluster edges incident to pvj ,x in Mi . It is straightforward to verify that the computed drawing defines a planar embedding for the wheel reduction of G = (V, E, C, Φ) consistent with Π = {π1 , π2 , . . . , π\|C\| }. From the planarity of the wheel reduction of G it follows that each non-trivial pole vµ of the frame graph F has a complete internal and a complete external sequence consistent with πvµ and that for every node µ of the SPQ decomposition tree of F having poles sµ and tµ , the pair (πsµ , πtµ ), with πsµ , πtµ ∈ Π, is compatible for µ. An example of the above described procedure is illustrated in Fig. 2 We now show that, if there exists a permutation assignment Π such that for every node µ of T with poles sµ and tµ we have that permutation pair (πsµ , πtµ ) is compatible for µ, then G = (V, E, C, Φ) is NodeTrix planar with side assignment Φ. We construct a planar embedding of the wheel reduction of G consistent with Π such that all external cycles of the wheels are embedded clockwise which, by Theorem 1, implies that G is NodeTrix planar. Let Wsµ and Wtµ be the two wheels consistent with πsµ and πtµ of sµ and tµ , respectively. We visit T from the leaves to the root and incrementally construct the desired planar embedding of the wheel reduction of G. If the visited node µ is a Q-node, at most one of its poles is non-trivial because G is light; assume, without loss of generality, that the non-trivial pole of µ is sµ and let Vsµ be the cluster of G represented by sµ in the frame graph of G. We embed the wheel Wsµ of Vsµ consistent with πsµ ∈ Π such that, when traversing the edges of the external cycle of Wsµ in the forward direction, the cycle is traversed clockwise. We embed tµ in the external face of Wsµ and planarly connect the top, bottom, left, or right copy of its end-vertex on Wsµ as specified by Φ. Suppose now µ is an S-node and let ν0 and ν1 be the children of µ such that sν0 = sµ , tν0 = sν1 , and tν1 = tµ . The planar embedding of the wheel reduction at node µ is obtained by composing the planar embedding of the wheel reduction at node ν0 with the planar embedding of the wheel reduction at ν1 . This is done by identifying the planar embedding of the wheel Wtν0 consistent with πtν0 with the planar embedding of the wheel Wsν1 consistent with πsν1 . Note that this is possible because πtν0 is the same as πsν1 , (πsµ , πtν0 ) is a compatible pair for ν0 , and (πsν1 , πtµ ) is a compatible pair for ν1 . Finally, assume µ is a P-node and let ν0 , ν1 , . . . , νh−1 be the children of µ. Similar to the case of the S-node, the planar embedding of the wheel reduction at node µ is obtained by composing the planar embeddings of the wheel reductions at nodes ν0 , ν1 , . . . , νh−1 . Since pair (πsµ , πtµ ) is compatible for µ, it defines opposite orders on the poles of µ. These opposite circular orders correspond to a planar embedding of the wheel reduction at µ obtained by combining the planarly embedded wheel reductions at its children ν0 , ν1 , . . . νh−1 . It follows that G = (V, E, C, Φ) is NodeTrix planar with permutation assignment Π. A.3 Proof of Theorem 2 Theorem 2. Let G = (V, E, C, Φ) be a partial 2-tree clustered graph with side assignment Φ. Let k be the maximum 3 size of any cluster in C and let n be the cardinality of V . There exists an O(k 3k+ 2 · n3 )-time algorithm that tests whether G is NodeTrix planar with side assignment Φ and if so, it computes a NodeTrix planar representation of G consistent with Φ. Proof. We compute a block-cut-vertex tree Tbcv of the frame graph of G, we root it at a block Broot and visit Tbcv bottom-up. Let Bi be the currently visited block and let c be the parent cut-vertex of Bi in Tbcv . We execute the testing algorithm of Lemma 2 by rooting the SPQ decomposition tree of Bi only at those Q-nodes that have c as one of their poles. If the test fails at any block of Tbcv , we conclude that G is not NodeTrix planar with the given side assignment for the chosen root Broot of Tbcv and we repeat the test by rooting Tbcv at a different block. Otherwise, we test whether, among the permutation assignments computed for the blocks B0 , B1 , . . . , Bh−1 that are children of a same cut-vertex c, there exists a set {Π0 , Π1 , . . . , Πh−1 } such that: (i) π0,c = π1,c = · · · = πh−1,c with πj,c ∈ Πj , for j = 0, 1, . . . , h − 1, and (ii) the complete internal sequence of c for the block Bj and for the permutation assignment πj,c does not overlap with the complete internal sequence of c for the block Bi and the permutation assignment πi,c , with i 6= j, i, j = 0, . . . , h − 1. We equip c with all the permutations that pass this test. Let B 0 be the block that is the parent of c in Tbcv . When testing B 0 for NodeTrix planarity, we consider for c only the permutations that have been computed when processing blocks B0 , B1 , . . . , Bh−1 and check that the complete internal sequences of c in Bj , j = 0, . . . , h − 1, do not intersect with the complete internal sequences of c in B 0 . Let nB be the number of vertices of a block B. By using Lemma 2, the procedure described above can be executed in 3 3 time O(k 3k+ 2 · n2B ) per block. Therefore, a bottom-up visit of Tbcv can be computed in O(k 3k+ 2 · n2 ) time for each 3 root block. Since there are O(n) possible roots for Tbcv , it follows that the overall time complexity is O(k 3k+ 2 · n3 ). A.4 Proof of Theorem 3 The lemmas and the discussion in this section prove Theorem 3. For completeness we report the statement of the theorem. Theorem 3. Let G = (V, E, C, Φ) be an n-vertex clustered graph with side assignment Φ such that the maximum size of any cluster in C is two. There exists an O(n3 )-time algorithm that tests whether G is NodeTrix planar with the given side assignment and if so, computes a NodeTrix planar representation of G consistent with Φ. Let G = (V, E, C, Φ) be a clustered graph with side assignment Φ and let F be its frame graph. Assume first that the frame graph F of G is biconnected. Let T be the SPQR decomposition tree of F . Analogously to the case of series-parallel frame graph described in Section 4.1, we traverse T from the leaves to the root and compute, for each node µ of T with poles sµ and tµ , all the pairs of permutations (πsµ , πtµ ) compatible for µ (the formal definitions of pair of permutations compatible for a Q-node, an S-node, or a P-node can be found in Section 4.1; the formal definition of pair of permutations compatible for an R-node is in Section 5). Differently from the algorithm for series-parallel frame graphs, we have to handle the case of an R-node µ, where the skeleton skel(µ) is a triconnected graph. In fact, the cases of Q-node, S-node and P-node can be handled exactly as described in Section 4.1, with an overall time complexity O(n2 ) (Lemma 2). Let µ be an R-node of T with poles sµ and tµ and with children ν0 , ν1 , . . . , νh−1 . Given a non-trivial vertex w of skel(µ), since the maximum cluster size is two, the possible permutations of Vw are exactly two, that we denote by + − πw and πw . For the sake of simplicity, we assume that each vertex w of skel(µ) is labeled with one or two labels in + − + − the set {πw , πw } corresponding to the two possible permutations πw and πw of the pertinent cluster Vw of w. Also, we assume that each edge ei = (u, v) of skel(µ), corresponding to child νi of µ, is assigned one or more labels in the set {(πu+ , πv+ ), (πu+ , πv− ), (πu− , πv+ ), (πu− , πv− )} that are initialized based on the pairs of permutations compatible with child νi . In the first phase of the algorithm for node µ we discard the edge and vertex labels that do not give rise to a planar embedding for skel(µ). In particular, we consider each one of the two possible planar embeddings for µ, one after the other. Let Eµ be the chosen embedding of µ. We perform a Rotation Coherence Check, that + has the purpose of enforcing the compatibility with respect to the embedding Eµ of the two permutations πw − and πw of each vertex w of skel(µ). Namely, consider a vertex w of skel(µ) and one of its possible permutations ∗ + − πw ∈ {πw , πw }. Let e0 , e1 , . . . eh−1 be the virtual edges of skel(µ) as they appear in clockwise order around w in the embedding Eµ of skel(µ) and let ν0 , ν1 , . . . , νh−1 be the corresponding nodes of T , respectively. If the complete ∗ internal sequences ISeq(νi , w, πw ), with i = 0, 2, . . . , h−1, appear in a different clockwise order around w with respect to the corresponding virtual edges e0 , e1 , . . . , eh−1 , then G is not NodeTrix planar with permutation assignment Π ∗ ∗ ∗ when πw ∈ Π. Hence, label πw can be discarded for w (and, correspondingly, permutation πw can be discarded for cluster Vw ). If at the end of this phase a vertex or an edge remains without any label, then G is not NodeTrix planar with the given embedding Eµ . Now, we want to decide whether there are pairs of permutations that are compatible for µ and, in the affirmative case, we want to compute them. The problem of deciding whether there are permutations pairs compatible with µ is equivalent to the following Coherent-Labeling-Problem: Given a plane graph G where each vertex w has at + − least a label in {πw , πw } and each edge (u, v) has at least one label in {(πu+ , πv+ ), (πu+ , πv− ), (πu− , πv+ ), (πu− , πv− )}, can ∗ a label πw be selected for each vertex w in such a way that each edge (u, v) has the label (πu∗ , πv∗ )? We call a plane graph G with the described labeling for its vertices and edges an instance of the Coherent-Labeling-Problem and a solution for G a coherent labeling for G. Observe that the vertex and edge labels can be further filtered off based on the following obvious properties: − + ∗ } cannot be selected for vertex w if there , πw ∈ {πw Vertex Coherence: In any coherent labeling for G, label πw ∗ exists a virtual edge ei = (w, v) of skel(µ) that does not have any label (πw , πv× ), with πv× ∈ {πv+ , πv− }. × ∗ Edge Coherence: In any coherent labeling for G, label (πw , πv ) cannot be selected for the virtual edge e = (w, v) ∗ or v does not admit label πv× . of skel(µ) if w does not admit label πw Hence, we can discard vertex and edge labels that do not satisfy the above properties. An instance of the Coherent-Labeling-Problem where each vertex label satisfies Property Vertex Coherence and each edge label satisfies Property Edge Coherence is said to be succinct. Observe that, each time a vertex label is discarded because of Property Vertex Coherence at vertex w, the labels of each virtual edge incident to w have to be checked against Property Edge Coherence. Also, each time an edge label is discarded because of Property Edge Coherence, Property Vertex Coherence has to be checked for the labels of one of its incident vertices. Due to the fact that at most one label per vertex and at most three labels per edge can be discarded (otherwise the instance is not NodeTrix planar), an instance of Coherent-Labeling-Problem can be reduced to a succinct instance in time linear in the number of edges of skel(µ). A succinct instance G of the Coherent-Labeling-Problem is called reduced if each vertex of G has exactly two labels and each edge of G has exactly three labels. We use the following lemma to efficiently solve the problem. Lemma 3. Let G be an instance of the Coherent-Labeling-Problem. There exists a reduced instance G∗ such that G admits a coherent labeling if and only if G∗ admits a coherent labeling. Proof. Graph G∗ can be found through a sequence of transformations starting from the original instance G. The first transformation is described in the proof of the following claim. Claim 1 Let G be a succinct instance of the Coherent-Labeling-Problem. There exists a succinct instance G∗ + − whose vertices have both the labels πw and πw and such that G admits a coherent labeling if and only if G∗ admits a coherent labeling. Proof of claim. Let w be a vertex of G that has only one label and let G−w be the graph obtained from G by removing w and all its incident edges. We show that G−w admits a coherent labeling if and only if G admits one. One direction is easy to prove: Every solution for G is also a solution for G−w , since all vertices and edges of G−w + are also vertices and edges of G. We now prove the opposite direction. Suppose vertex w has only label πw (the − case of πw being analogous) and consider any edge (v, w) incident to w. By Property Edge Coherence, (v, w) cannot − − + + ). It follows that edge (v, w) may have: (1) label (πv+ , πw ); (2) label (πv− , πw ); or have labels (πv+ , πw ) and (πv− , πw + + − + (3) both labels (πv , πw ) and (πv , πw ). In turn, by Property Vertex Coherence for the labels of vertex v, we have the following three cases, respectively: (1) vertex v has only label πv+ ; (2) vertex v has only label πv− ; and (3) vertex v has both labels πv+ and πv− . Suppose G−w admits a solution. You can easily obtain a solution for the original instance G by reinserting w + and all its incident edges and by selecting their labels as follows. (a) Select label πw for w. (b) Select, for each edge + − + + + ), respectively. Depending e = (v, w) incident to w, depending on the label πv or πv of v, label (πv , πw ) or (πv− , πw on the three cases above, we prove that G admits a coherent labeling. + + 1. Since any solution of G−w selects label πv+ for vertex v, selecting label πw for vertex w and label (πv+ , πw ) for edge (v, w) yields a solution for G. + + 2. Since any solution of G−w selects label πv− for vertex v, selecting label πw for vertex w and label (πv− , πw ) for edge (v, w) yields a solution for G. + 3. A solution of G−w may select either label πv+ (Case A) or label πv− (Case B) for vertex v. Selecting label πw for − + − + vertex w and label (πv , πw ) in Case A or (πv , πw ) in Case B for edge (v, w) yields a solution for G. By repeating this procedure for every vertex of G that has only one label, an equivalent instance G∗ such that all vertices have both labels is obtained. Finally, observe that, since the removal of edges from an instance of Coherent-Labeling-Problem does not disrupt the properties of vertex coherence and edge coherence of its labels, when applying the transformation described above we obtain a succinct instance G∗ whenever the original instance G was succinct. Claim 1 has several important consequences. In order to explicit them, we need some further notation. We say that an edge (u, v) has label set x, with x ∈ {1, 2, . . . , 16}, if (u, v) has the following sets of labels: Label set (πu+ , πv+ ) (πu+ , πv− ) (πu− , πv+ ) (πu− , πv− ) 1 2 × 3 × 4 × 5 × 6 × × 7 × × 8 × × 9 × × 10 × × 11 × × 12 × × × 13 × × × 14 × × × 15 × × × 16 × × × × An immediate consequence of Claim 1 is the following property. + − Claim 2 If each vertex w of an instance G of the Coherent-Labeling-Problem has both the labels πw and πw , then edges have exclusively label sets in {8, 9, 12, 13, 14, 15, 16}. Proof of claim. The proof is based on the observation that the remaining label sets either miss πu+ (label sets 1, 4, 5, and 11), or miss πu− (among others, label sets 2, 3, and 6), or miss πv+ (among others, label set 10), or πv− (among others, label set 7). However, by Property Vertex Coherence, u and v cannot have both labels if there is some edge that does not allow both of them. Claim 3 Let G be a succinct instance of the Coherent-Labeling-Problem. There exists a succinct instance G∗ such that no edge has label set 16 and such that G∗ admits a coherent labeling if and only if G admits one. Further, a coherent labeling for G can be obtained from a coherent labeling for G∗ in constant time. Proof of claim. If an edge (u, v) has all four labels (πu+ , πv+ ), (πu+ , πv− ), (πu− , πv+ ), and (πu− , πv− ), one could remove (u, v), compute a coherent labeling of the obtained instance G∗ , and then insert back again (u, v) with a suitable label depending on the pair of labels selected for its end-vertices. Claim 4 Let G be a succinct instance of the Coherent-Labeling-Problem. There exists a succinct instance G∗ such that no edge has label set 8 and such that G∗ admits a coherent labeling if and only if G admits one. Further, a coherent labeling for G can be obtained from a coherent labeling for G∗ in constant time. Proof of claim. Suppose edge (u, v) only has labels (πu+ , πv+ ) and (πu− , πv− ). Contract edge (u, v) and merge vertices u and v into a new vertex w. Suppose that the obtained graph G∗ admits a coherent labeling. Expand the vertex + w back into u and v. If w has label πw , select label πu+ for u, label πv+ for v, and label (πu+ , πv+ ) for edge (u, v). Otherwise, select label πu− for u, label πv− for v, and label (πu− , πv− ) for (u, v). It is easy to verify that the obtained selection of labels is a coherent labeling for G. Claim 5 Let G be a succinct instance of the Coherent-Labeling-Problem. There exists a succinct instance G∗ such that no edge has label set 9 and such that G∗ admits a coherent labeling if and only if G admits one. Further, a coherent labeling for G can be obtained from a coherent labeling for G∗ in constant time. Proof of claim. Suppose edge (u, v) only has labels (πu+ , πv− ) and (πu− , πv+ ). Replace πv− with πv+ and πv+ with πv− for the labels of v and for all the labels of the edges incident to v. Observe that the instance where the labels have been renamed admits a coherent labeling if and only if instance G admits one. In particular, edges with label set 9 (12, 13, 14, 15, respectively) have now label set 8 (13, 12, 15, 14, respectively). Hence, for edge (u, v) Claim 4 applies, and the edge can be contracted and reinserted after a coherent labeling for the resulting instance G∗ is found. Claims 1, 2, 3, 4, and 5, imply that we can assume that every vertex has both labels and every edge has a label set in {12, 13, 14, 15}. Hence, the obtained instance G∗ is a reduced instance of the Coherent-Labeling-Problem equivalent to instance G, which concludes the proof of Lemma 3. − + } and every , πw Based on Lemma 3, we can assume that every vertex w of G has exactly two labels in {πw edge (u, v) of G has exactly three labels in {(πu+ , πv+ ), (πu+ , πv− ), (πu− , πv+ ), (πu− , πv− )}. Consider an edge (u, v) whose missing label is (πu+ , πv+ ). This implies that, if the instance G of the Coherent-Labeling-Problem has a coherent labeling in which label πu+ is selected for vertex u, then the label selected for vertex v in this solution must be πv− . This consideration can be generalized as follows: Property 2. Let G be a reduced instance of the Coherent-Labeling-Problem and let e = (u, v) be an edge of G missing label {πu∗ , πv∗ }, with πu∗ ∈ {πu+ , πu− } and πv∗ ∈ {πv+ , πv− }. In any coherent labeling of G, if the label selected for u is πu∗ , then the label selected for v is the label different from πv∗ . Analogously, if the label selected for v is πv∗ , then the label selected for u is the label different from πu∗ . Property 2 shows that selecting a label for a vertex may have consequences on the label that have to be selected for the neighbor vertices, and a local choice may propagate in the graph. However, Property 2 also implies the following property. Property 3. If the missing label of edge (u, v) is such that the selection of a label πu∗ for vertex u, with πu∗ ∈ {πu+ , πu− }, has no implication on the selection of the label for v, then any selection of a label πv∗ for vertex v, with πv∗ ∈ {πv+ , πv− }, is compatible with the selection of label πu∗ for vertex u. This gives rise to the following procedure to compute all pairs of permutations (πs∗µ , πt∗µ ) compatible for R-node µ. Since skel(µ) has two possible planar embeddings, the procedure will be repeated twice, once for each embedding. Let Eµ be the current embedding of skel(µ). – Build an instance G of Coherent-Labeling-Problem (if a vertex or edge has zero labels, then µ does not admit any compatible pair when Eµ is the embedding chosen for skel(µ)). – Obtain a reduced instance G∗ of the Coherent-Labeling-Problem that admits a coherent labeling if and only if G admits one. Let s∗ and t∗ be the two vertices of G∗ corresponding to the vertices sµ and tµ of the original instance G. – Consider one by one the four possible permutation pairs of µ. To decide whether a permutation pair {πs∗µ , πt∗µ } is compatible for µ, select for s∗ the label πs∗∗ corresponding to πs∗µ and propagate the selection through G + according to Property 2. If a vertex w is reached two times, once by imposing the selection of label πw and − ∗ ∗ once by imposing the selection of label πw , then the pair {πsµ , πtµ } is not compatible for µ. Suppose this first propagation phase stops without any such contradiction. If a label for vertex t∗ has not been selected yet, start a second propagation by selecting for t∗ the label πt∗∗ corresponding to πt∗µ . While there are vertices for which no label has been selected, start a new propagation phase selecting an arbitrary label for one of them. Property 3 guarantees that a vertex traversed during a propagation phase cannot be traversed in a successive propagation phase. Hence, the labels selected during one propagation phase cannot cause a contradiction with the labels selected in successive propagation phases. If no contradiction is found and if πt∗∗ is the selected label for t∗ , permutation pair {πs∗µ , πt∗µ } is compatible for µ. The proof of Theorem 3 is concluded by observing that each of the transformations described in the proof of Lemma 3 requires constant time and therefore the overall time complexity to compute all pairs of permutations (πs∗µ , πt∗µ ) is linear in the number of edges of skel(µ). This observation, together with Lemma 2, implies that NodeTrix planarity with fixed sides can be tested in O(n2 ) time on a clustered graph whose maximum cluster size is two and such that the frame graph is biconnected. For the connected case we construct the block-cut-vertex tree Tbcv of the frame graph of G and apply the same strategy described in the proof of Theorem 2. Namely, for every node µ that is either a Q-, or an S-, or a P-node, we test whether there exists a compatible pair for µ by means of the procedure in Lemma 2; if µ is an R-node we use the strategy described above. This gives rise to an O(n2 )-time procedure for each possible choice of the root of Tbcv . Since there are O(n) such possible choices, we have an overall time complexity of O(n3 ) if the frame graph is not biconnected. The statement of Theorem 3 follows. A.5 Proof of Theorem 4 Theorem 4. NodeTrix planarity testing with fixed sides and cluster size at most k is NP-complete for any k ≥ 3. Proof. NodeTrix planarity testing with fixed sides is trivially in NP. In fact, given a clustered graph G = (V, E, C, Φ), all possible permutations assignments Π could be non-deterministically computed and the problem of deciding whether a clustered graph G = (V, E, C, Φ, Π) admits a NodeTrix planar representation with side assignment Φ and permutation assignment Π is solvable in linear time [3]. In order to prove its NP-hardness, we reduce NAE3SAT to it. An instance of NAE3SAT consists of a collection of clauses on a set of Boolean variables, where each clause consists of exactly three literals. The problem asks whether there exists a truth assignment to the variables so that each clause has at least one true literal and at least one false literal. This problem has been shown to be NP-complete by Thomas J. Schaefer [11]. However, it is known to be polynomial when the graph of the adjacencies of the variables and clauses is planar [10]. Starting from a (non-planar) instance of NAE3SAT with variables x1 , x2 , . . . , xn and clauses C1 , C2 , . . . , Cm , we construct an instance G = (V, E, C, Φ) of NodeTrix planarity testing with fixed sides as follows. First, we obtain a (non-planar) drawing Γ of the graph of its variables and clauses like the one in Fig. 3: The clause vertices are vertically aligned, the variable vertices are horizontally aligned, and the edges are drawn with L-shapes. Clearly this drawing can be computed in polynomial time and has a polynomial number of crossings. − C1 + + + C2 − − + C3 + + + C4 − − − C5 − − x1 x2 x3 x4 Fig. 3. A non-planar drawing of an instance of NAE3SAT. The small circled plus signs and circled minus signs represent direct and negated occurrences of the variables in the clauses, respectively. Then, we replace each vertex representing a variable with a variable gadget. The variable gadget for a variable of degree h is composed by h size-two clusters connected together as depicted in Fig. 4(a) and 4(b). Namely, let Vv,1 , Vv,2 , . . . , Vv,h be the h clusters of C composing the variable gadget of variable v and let {ui,1 , ui,2 } be the nodes of Vv,i , with i = 1, . . . , h. We may encode a truth value with each one of the two possible representations of cluster Vi : if in the matrix Mi representing Vi the column corresponding to the vertex ui,1 precedes the column corresponding to the vertex ui,2 , we say that Mi is true. Otherwise, we say that Mi is false. Correspondingly, we say that πi = ui,1 , ui,2 is the true permutation of cluster Vi and that πi = ui,2 , ui,1 is its false permutation. In order to connect the clusters composing the variable gadget for v, we add to E, for i = 1, . . . , h − 1, the inter-cluster edges ei,1 = (ui,1 , ui+1,1 ) and ei,2 = (ui,2 , ui+1,2 ) and set φi (ei,1 ) = r, φi+1 (ei,1 ) = b, φi (ei,2 ) = r, φi+1 (ei,2 ) = b. It is immediate that in any NodeTrix planar representation of G, all Mi , i = 1, . . . , h, are either simultaneously true or simultaneously false. Correspondingly, we say that the variable gadget is true or false. Fig. 4(a) and 4(b) show an example of a true and of a false drawing of a variable gadget. FALSE TRUE 2 1 1 2 1 2 1 2 2 1 2 1 TRUE (a) (b) FALSE (c) (d) Fig. 4. (a) The true configuration of a variable gadget for a variable of degree four. (b) The false configuration. (c) The not gadget transforming an encoded true value into a false value. (d) The not gadget transforming an encoded false value into a true value. Each edge e attaching to a variable v in the drawing Γ (refer to Fig. 3) corresponds to two ‘parallel’ inter-cluster edges e1 and e2 attached to one of the clusters composing the variable gadget of v. Let Vj be such a cluster. We set φj (e1 ) = t and φj (e2 ) = t. Observe that the order in which e1 and e2 exit Mj depends on the truth value encoded by Mj and, hence, on the truth value encoded by the variable gadget for v. Fig. 5 depicts the gadget we use to replace crossings in Γ , consisting of a cluster Vx of size three. From the figure it is apparent that, in any representation of Vx , the left-to-right order of the edges entering Mx from the bottom side is the same as the left-to-right order of the edges exiting Mx from the top side. An analogous consideration holds for the top-to-bottom order of the edges entering Mx from the right side and the top-to-bottom order of the edges exiting Mx from the left side. This implies that the truth value encoded by the edges entering Mx is the same as the truth value encoded by the edges exiting it. FALSE TRUE 1 2 3 2 1 (a) 3 3 1 2 TRUE FALSE (b) (c) TRUE FALSE 3 2 1 FALSE TRUE FALSE TRUE FALSE FALSE TRUE TRUE FALSE TRUE TRUE TRUE 1 3 FALSE FALSE (d) 2 2 3 TRUE FALSE TRUE (e) 1 FALSE (f) Fig. 5. The six possible configurations of a crossing gadget. We now describe the clause gadget. We assume that three pairs of edges, encoding the truth value of the variables occurring in the clause, arrive to the clause gadget. Let v1 , v2 , and v3 be the three variables whose literals l1 , l2 , and l3 occur in clause C. Before entering the clause gadget, if literal li is a negated literal of variable v (i.e., if li = v i ) then we attach the edges coming from vi to a not gadget, depicted in Fig. 4(c) and 4(d), and use the edges exiting the not gadget instead of the edge coming directly from variable vi . This has the effect that all the three pairs of edges entering the clause gadget encode a truth value that is true if the literal is true and false if the literal is false. In the following, therefore, we will consider the truth values of the literals, rather than the truth values of the variables. The clause gadget, depicted in Fig. 6, is composed by three clusters V1 , V2 , and V3 of size three, each having vertices {ui,a , ui,b , ui,c , }, i = 1, 2, 3. The three clusters are connected together in such a way that, in any NodeTrix planar representation of G, their permutations πi , i = 1, 2, 3, always present the same sequence of the labels a, b, c. For example, if π1 = u1,c , u1,a , u1,b , then also π2 = u2,c , u2,a , u2,b and π3 = u3,c , u3,a , u3,b . For i = 1, . . . , 3, the edges encoding the truth value of literal li attach to cluster Vi , where the prescribed side is the right side of matrix Mi . The two edges e1,1 and e1,2 encoding the truth value of the literal l1 attach to u1,a and u1,b , respectively. The two edges e2,1 and e2,2 encoding the truth value of the literal l2 attach to u2,b and u2,c , respectively. Finally, the two edges e3,1 and e3,2 encoding the truth value of the literal l3 attach to u3,c and u3,a , respectively. Hence, if literal l1 is true, matrix M1 must have a permutation of its columns such that column a precedes column b, while if literal l1 is false, matrix M1 must have a permutation of its columns such that column a follows column b. Analogously, the truth value of literal l2 determines whether in matrix M2 column b precedes or follows column c, and the truth value of literal l3 determines whether in matrix M3 column c precedes or follows column a. TRUE if literal TRUE TRUE if literal TRUE TRUE if literal TRUE Fig. 6. The clause gadget. It follows that, if all three literals are true, then they induce unsatisfiable constraints on the ordering of the columns of the matrices, since column a should precede column b, b should precede column c, and c should precede a. The same holds if all three literals are false. It can be easily checked that, for any other combination of truth values of the literals, there exists an ordering of the columns of matrices M1 , M2 , and M3 that makes a planar drawing of the edges possible. Therefore, the constructed instance of NodeTrix planarity with fixed sides admits a planar NodeTrix representation if and only if the original instance of NAE3SAT admits a solution. A.6 Proof of Theorem 5 Theorem 5. NAE3SAT is NP-complete for triconnected Boolean formulas. Proof. We use a reduction strategy that is similar to the one used in [9] to prove the hardness of Planar Triconnected 3SAT. Here, however, we do not have to worry about planarity, since we are searching for a non-planar triconnected instance of NAE3SAT. Let ϕ be an instance of NAE3SAT such that the graph G that represents the occurrences of the variables into the clauses of ϕ is not triconnected and, possibly, not biconnected and not connected. We show how to construct an instance ϕ∗ such that its graph G∗ is triconnected. The strategy consists of adding to ϕ a suitable number of variables and clauses such that the obtained instance ϕ∗ is triconnected and has a solution if and only if ϕ has one. Consider an edge e = (x, C) in the graph G (refer to Fig. 7(a)). Suppose that variable x occurs in clause C with its direct literal (negated literal, respectively). We remove e (that is, we remove the occurrence of x into clause C) and replace it with the occurrence of the direct literal (negated literal, respectively) of a new variable x0 . Also we add a new variable y and the two clauses (x ∨ x0 ∨ y) and (x ∨ x0 ∨ y) (see Fig. 7(b)). It can be shown that, in any assigment of truth values to the variables such that all clauses are satisfied in the NAE3SAT sense, the truth value of x has to be the same as the truth value of x0 . In fact, whatever is the truth x0 x − C 00 x + C C0 C1 C2 y + + x1 (a) + C1 0 + (b) C2 0 − − + x2 0 − x2 − y + − + + − + − + C2 x1 0 C + C1 x1 x2 (c) − + C1 00 + + C2 00 (d) Fig. 7. (a) A variable x occurring in a clause C of an instance of NAE3SAT. (b) The gadget that replaces edge (x, C). (c) A pair of edges (x1 , C1 ) and (x2 , C2 ) in the NAE3SAT graph. (c) The gadget that replaces edges (x1 , C1 ) and (x2 , C2 ). value of variable y, literals x and x0 occur in a clause with a negative literal and a positive literal. Hence, they cannot be both positive or both negative. This implies x = x0 . It follows that the obtained instance of NAE3SAT admits a solution if and only if the original instance admits one. Now, consider two edges e1 = (x1 , C1 ) and e2 = (x1 , C2 ) of G (see Fig. 7(c)). By replacing both edges with the above described gadget and by identifying the y variables of the two gadgets, one obtains a bridge gadget for e1 and e2 . The instance of NAE3SAT where e1 and e2 have been replaced by their bridge gadget is equivalent to the original instance and it has a path going from C1 to C2 and passing through y. Hence, suppose you have a non-connected graph G and let e1 and e2 be two edges belonging to two different connected components of G. Replacing e1 and e2 with their bridge gadget, the number of connected components of G is decreased. Assume now that G is connected. Let Tbcv be the block-cut-vertex tree of G and let e1 and e2 be two edges belonging to two different blocks of Tbcv . Replacing e1 and e2 with their bridge gadget, the number of blocks of G is decreased (see for example Fig. 8). The same holds if we replace two edges e1 and e2 belonging to two triconnected components of G. B1 B2 B4 B1 B6 B5 B3 B7 (a) B2 B8 B4 B6 B5 B3 B7 B8 (b) Fig. 8. (a) The block-cut-vertex tree Tbcv of a graph (green vertices are the leaves of Tbcv ). (b) A bridge gadget added to link two leaf blocks of Tbcv . Therefore, we obtain a NAE3SAT instance ϕ∗ such that the graph that represents the occurrences of the variables into the clauses is triconnected and such that ϕ∗ admits a solution if and only if ϕ admits one. Furthermore, the size of the gadgets used in the replacement is constant and therefore the size of ϕ∗ is polynomial in the size of ϕ. This concludes the proof. A.7 Proof of Theorem 6 Theorem 6. NodeTrix planarity testing with free sides and cluster size at most k is NP-complete for any k ≥ 5. Proof. The proof is based on a reduction from triconnected NAE3SAT. Starting from a triconnected instance ϕ of NAE3SAT, we first build the instance Gfix of NodeTrix planarity with fixed sides, analogously to the proof of Theorem 4, but using the drawing strategy shown in Fig. 9 instead of that shown in Fig. 3. Observe that, since ϕ is triconnected and because of the special planarization strategy we used, the frame graph of Gfix is also triconnected. In fact, inserting crossing gadgets is equivalent to planarizing the drawing of Fig. 9. x1 − C1 + + + C2 − − + C3 + + + C4 − − − − C5 − x2 x3 x4 Fig. 9. A non-planar drawing of an instance of NAE3SAT such that, by replacing crossings with dummy nodes, you would obtain a planar and triconnected graph. (a) (b) Fig. 10. (a) A cluster Vi of size three of the instance of NodeTrix planarity with fixed sides. (b) The gadget used to replace Vi in the instance of NodeTrix planarity with free sides. Second, we obtain an instance Gfree of NodeTrix planarity with free sides by replacing each cluster of maximum size three, with the gadget depicted in Fig. 10(b) that uses exclusively clusters of size 5. Such a gadget consists of nine clusters such that the corresponding nodes in the frame graph of Gfree form a wheel graph with an external cycle of eight nodes all connected to a central one. Since the wheel has only one embedding (up to a flip), the gadget admits a NodeTrix planar representation only if the hub of the wheel is drawn inside the cycle formed by the other eight clusters. Also, the edges that in the instance Gfix are constrained to attach to a specific side of a matrix due to the side assignment Φ, are now all incident to the same cluster of the wheel. Since Gfree has a triconnected frame graph, the embedding of the frame graph is fixed, and a NodeTrix planar representation with fixed sides of Gfix can be immediately obtained from a NodeTrix planar representation with free sides of Gfree .	8cs.DS
TC0 circuits for algorithmic problems in nilpotent groups Alexei Myasnikov1 and Armin Weiß2 1 2 Stevens Institute of Technology, Hoboken, NJ, USA Universität Stuttgart, Germany arXiv:1702.06616v2 [math.GR] 26 Jul 2017 Abstract Recently, Macdonald et. al. showed that many algorithmic problems for finitely generated nilpotent groups including computation of normal forms, the subgroup membership problem, the conjugacy problem, and computation of subgroup presentations can be done in LOGSPACE. Here we follow their approach and show that all these problems are complete for the uniform circuit class TC0 – uniformly for all r-generated nilpotent groups of class at most c for fixed r and c. In order to solve these problems in TC0 , we show that the unary version of the extended gcd problem (compute greatest common divisors and express them as linear combinations) is in TC0 . Moreover, if we allow a certain binary representation of the inputs, then the word problem and computation of normal forms is still in uniform TC0 , while all the other problems we examine are shown to be TC0 -Turing reducible to the binary extended gcd problem. Keywords and phrases nilpotent groups, TC0 , abelian groups, word problem, conjugacy problem, subgroup membership problem, greatest common divisors Contents 1 Introduction 2 2 Preliminaries 2.1 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Nilpotent groups and Mal’cev coordinates . . . . . . . . . . . . . . . . . . . . 3 3 5 3 Presentation of subgroups 3.1 Quotient presentations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 7 4 Word problem and computation of Mal’cev coordinates 8 5 The extended gcd problem 9 6 Matrix reduction and subgroup membership problem 14 6.1 Subgroup membership problem . . . . . . . . . . . . . . . . . . . . . . . . . . 16 6.2 Subgroup presentations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 7 More algorithmic problems 18 7.1 Homorphisms and kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 7.2 Centralizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 7.3 The conjugacy problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 8 Computing quotient presentations 20 9 Power problem and conjugacy in wreath products of nilpotent groups 21 10 Conclusion and Open Problem 22 © Alexei Myasnikov, Armin Weiß; licensed under Creative Commons License CC-BY Leibniz International Proceedings in Informatics Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany 2 A. Myasnikov, A. Weiß 1 Introduction The word problem (given a word over the generators, does it represent the identity?) is one of the fundamental algorithmic problems in group theory introduced by Dehn in 1911 [3]. While for general finitely presented groups all these problems are undecidable [23, 2], for many particular classes of groups decidability results have been established – not just for the word problem but also for a wide range of other problems. Finitely generated nilpotent groups are a class where many algorithmic problems are (efficiently) decidable (with some exceptions like the problem of solving equations – see e. g. [6]). In 1958, Mal’cev [18] established decidability of the word and subgroup membership problem by investigating finite approximations of nilpotent groups. In 1965, Blackburn [1] showed decidability of the conjugacy problem. However, these methods did not allow any efficient (e. g. polynomial time) algorithms. Nevertheless, in 1966 Mostowski provided “practical” algorithms for the word problem and several other problems [20]. In terms of complexity, a major step was the result by Lipton and Zalcstein [15] that the word problem of linear groups is in LOGSPACE. Together with the fact that finitely generated nilpotent groups are linear (see e. g. [7, 10]) this gives a LOGSPACE solution to the word problem of nilpotent groups, which was later improved to uniform TC0 by Robinson [24]. A typical algorithmic approach to nilpotent groups is using so-called Mal’cev (or Hall– Mal’cev) bases (see e. g. [7, 10]), which allow to carry out group operations by evaluating polynomials (see Lemma 3). This approach was systematically used in [11] and [20] or – in the more general setting of polycyclic presentations – in [25] for solving (among others) the subgroup membership and conjugacy problem of polycyclic groups. Recently in [21, 22] polynomial time bounds for the equalizer and subgroup membership problems in nilpotent groups have been given. Finally, in [16] the following problems were shown to be in LOGSPACE using the Mal’cev basis approach. Here, Nc,r denotes the class of nilpotent groups of nilpotency class at most c generated by at most r elements. The word problem: given G ∈ Nc,r and g ∈ G, is g = 1 in G? Given G ∈ Nc,r and g ∈ G, compute the (Mal’cev) normal form of g. The subgroup membership problem: Given G ∈ Nc,r and g, h1 , . . . , hn ∈ G, decide whether g ∈ hh1 , . . . , hn i and, if so, express g as a word over the subgroup generators h1 , . . . , hn (in [16] only the decision version was shown to be in LOGSPACE – for expressing g as a word over the original subgroup generators a polynomial time bound was given). Given G, H ∈ Nc,r and K = hg1 , . . . , gn i ≤ G, together with a homomorphism ϕ : K → H specified by ϕ(gi ) = hi , and some h ∈ Im(ϕ), compute a generating set for ker(ϕ) and find g ∈ G such that ϕ(g) = h. Given G ∈ Nc,r and K = hg1 , . . . , gn i ≤ G, compute a presentation for K. Given G ∈ Nc,r and g ∈ G, compute a generating set for the centralizer of g. The conjugacy problem: Given G ∈ Nc,r and g, h ∈ G, decide whether or not there exists u ∈ G such that u−1 gu = h and, if so, find such an element u. These problems are not only of interest in themselves, but also might serve as building blocks for solving the same problems in polycyclic groups – which are of particular interest because of their possible application in non-commutative cryptography [4]. In this work we follow [16] and extend these results in several ways: We give a complexity bound of uniform TC0 for all the above problems. In order to derive this bound, we show that the extended gcd problem (given a1 , . . . , an ∈ Z, P compute x1 , . . . , xn ∈ Z with gcd(a1 , . . . , an ) = i ai xi ) with input and output in unary is in uniform TC0 . TC0 circuits for algorithmic problems in nilpotent groups Our description of circuits is for the uniform setting where G ∈ Nc,r is part of the input (in [16] the uniform setting is also considered; however, only in some short remarks). Since nilpotent groups have polynomial growth, it is natural to allow compressed inputs: we give a uniform TC0 solution for the word problem allowing words with binary exponents as input – this contrasts with the situation with straight-line programs (i. e., contextfree grammars which produces precisely one word – another method of exponential compression) as input: then the word problem is hard for C= L [12]. Thus, the difficulty of the word problem with straight-line programs is not due to their compression but rather due to the difficulty of evaluating a straight-line program. We show that the other of the above problems are uniform-TC0 -Turing-reducible to the (binary) extended gcd problem when the inputs (both the ambient group and the subgroup etc.) are given as words with binary exponents. We show how to solve the power problem in nilpotent groups. This allows us to apply a result from [19] in order to show that iterated wreath products of nilpotent groups have conjugacy problem in uniform TC0 . Thus, in the unary case we settle the complexity of the above problems completely. Moreover, it also seems rather difficult to solve the subgroup membership problem without computing gcds – in this case our results on binary inputs would be also optimal. Altogether, our results mean that many algorithmic problems are no more complicated in nilpotent groups than in abelian groups. Notice that while in [16] explicit length bounds on the outputs for all these problems are proven, we obtain polynomial length bounds simply by the fact that everything can be computed in uniform TC0 (for which in the following we only write TC0 ). Throughout the paper we follow the outline of [16]. For a concise presentation, we copy many definitions from [16]. Most of our theorems involve two statements: one for unary encoded inputs and one for binary encoded inputs. In order to have a concise presentation, we always put them in one result. We only consider finitely generated nilpotent groups without mentioning that further. Outline. We start with basic definitions on complexity as well as on nilpotent groups. In Section 3 we describe how subgroups of nilpotent groups can be represented and develop a “nice” presentation for all groups in Nc,r . Section 4 deals with the word problem and computation of normal forms. After that we solve the unary extended gcd problem in TC0 and introduce the so-called matrix reduction in order to solve the subgroup membership problem. In Section 7 we present our result for the remaining of the above problems, in Section 8 we explain how to compute “nice” presentations, and in Section 9 we apply the results of [19] in order to show that the conjugacy problem of iterated wreath products of nilpotent groups is in TC0 . Finally, we conclude with some open questions. 2 2.1 Preliminaries Complexity For a finite alphabet Σ, the set of words over Σ is denoted by Σ∗ . Computation or decision problems are given by functions f : ∆∗ → Σ∗ for some finite alphabets ∆ and Σ. A decision problem (= formal language) L is identified with its characteristic function χL : ∆∗ → {0, 1} with χL (x) = 1 if, and only if, x ∈ L. (In particular, the word and conjugacy problems can be seen as functions Σ∗ → {0, 1}.) We use circuit complexity as described in [26]. Circuit Classes. The class TC0 is defined as the class of functions computed by families of circuits of constant depth and polynomial size with unbounded fan-in Boolean gates (and, 3 4 A. Myasnikov, A. Weiß or, not) and majority gates. A majority gate (denoted by Maj) returns 1 if the number of 1s in its input is greater or equal to the number of 0s. In the following we always assume that the alphabets ∆ and Σ are encoded over the binary alphabet {0, 1} such that each letter uses the same number of bits. We say a function f is TC0 -computable if f ∈ TC0 . In the following, we only consider Dlogtime-uniform circuit families and we simply write TC0 as shorthand for Dlogtime-uniform TC0 . Dlogtime-uniform means that there is a deterministic Turing machine which decides in time O(log n) on input of two gate numbers (given in binary) and the string 1n whether there is a wire between the two gates in the n-input circuit and also computes of which type some gates is. Note that the binary encoding of the gate numbers requires only O(log n) bits – thus, the Turing machine is allowed to use time linear in the length of the encodings of the gates. For more details on these definitions we refer to [26]. We have the following inclusions (note that even TC0 ⊆ P is not known to be strict): AC0 $ TC0 ⊆ LOGSPACE ⊆ P. Reductions. A function f is TC0 -Turing-reducible to a function g if there is a Dlogtimeuniform family of TC0 circuits computing f which, in addition to the Boolean and majority gates, also may use oracle gates for g (i. e., gates which on input x output g(x)). This is expressed by f ∈ TC0 (g). Note that if f1 , . . . , fk are in TC0 , then TC0 (f1 , . . . , fk ) = TC0 . In particular, if f and g are TC0 -computable functions, then also the composition g ◦ f is TC0 -computable. We will extensively make use of this observation – which will also guarantee the polynomial size bound on the outputs of our circuits without additional calculations. We will also use another fact frequently without giving further reference: on input of two alphabets Σ and ∆ (coded over the binary alphabet), a list of pairs (a, va ) with a ∈ Σ and va ∈ ∆∗ such that each a ∈ Σ occurs in precisely one pair, and a word w ∈ Σ∗ , the image ϕ(w) under the homomorphism ϕ defined by ϕ(a) = va can be computed in TC0 [13]. Encoding numbers: unary vs. binary. There are essentially two ways of representing integer numbers: the usual way as a binary number where a string a0 · · · an with ai ∈ {0, 1} represents P ai 2n−i , and as a unary number where k ∈ N is represented by 1k = 11 · · · 1} (respectively \| {z k by 0n−k 1k if n is the number of input bits). We will state most results in this paper with both representations. The unary representation corresponds to group elements given as words over the generators, whereas the binary encoding will be used if inputs are given in a compressed form. I Example 1. The following problem Count is in TC0 : given a bit-string u of length n and a number j < n (we assume that it is given in unary as 0n−j 1j ), decide whether the number of ones \|u\|1 in u is exactly j. We have \|u\|1 ≥ j if, and only if, u0j 1n−j 1 ≥ n. Thus, Count(u, j) = Maj(u0j 1n−j ) ∧ ¬Maj(u0j 1n−j ) . In particular, the word problem of Z when 1 is encoded as 1 and −1 as 0, which is simply the question whether \|u\|1 = n/2 and n even, is in TC0 . Arithmetic in TC0 . Iterated Addition (resp. Iterated Multiplication) are the following computation problems: On input of n binary integers a1 , . . . , an each having n bits Pn (i. e., the input length is N = n2 ), compute the binary representation of the sum i=1 ai Qn (resp. product i=1 ai ). For Integer Division the input are two binary n-bit integers a, b; the binary representation of the integer c = ba/bc has to be computed. The first statement of Theorem 2 is a standard fact, see [26]; the other statements are due to Hesse, [8, 9]. TC0 circuits for algorithmic problems in nilpotent groups 5 I Theorem 2 ([8, 9, 26]). The problems Iterated Addition, Iterated Multiplication, Integer Division are all in TC0 no matter whether inputs are given in unary or binary. Note that if the numbers a and b are encoded in unary (as strings 1a and 1b ), division can be seen to be in TC0 very easily: just try for all 0 ≤ c ≤ a whether 0 ≤ a − bc < b. Representing groups for algorithmic problems. We consider finitely generated groups G together with finite generating sets A. Group elements are represented as words over the generators and their inverses (i. e., as elements of (A ∪ A−1 )∗ ). We make no distinction between words and the group elements they represent. Whenever it might be unclear whether we mean equality of words or of group elements, we write “g = h in G” for equality in G. Words over the generators ±1 of Z correspond to unary representation of integers. As a generalization of binary encoded integers, we introduce the following notion: a word with binary exponents is a sequence w1 , . . . , wn where the wi are from a fixed generating set of the group together with a sequence of exponents x1 , . . . , xn where the xi ∈ Z are encoded in binary. The word with binary exponents represents the word (or group element) w = w1x1 · · · wnxn . Note that in a fixed nilpotent group every word of length n can be rewritten as a word with binary exponents using O(log n) bits (this fact is well-known and also a consequence of Theorem 6 below); thus, words with binary exponents are a natural way of representing inputs for algorithmic problems in nilpotent groups. 2.2 Nilpotent groups and Mal’cev coordinates Let G be a group. For x, y ∈ G we write xy = y −1 xy (x conjugated by y) and [x, y] = x−1 y −1 xy (commutator of x and y). For subgroups H1 , H2 ≤ G, we have [H1 , H2 ] = h{[h1 , h2 ] \| h1 ∈ H1 , h2 ∈ H2 }i. A group G is called nilpotent if it has central series, i.e. G = G1 ≥ G2 ≥ · · · ≥ Gc ≥ Gc+1 = 1 (1) such that [G, Gi ] ≤ Gi+1 for all i = 1, . . . , c. If G is finitely generated, so are the abelian quotients Gi /Gi+1 , 1 ≤ i ≤ c. Let ai1 , . . . , aimi be a basis of Gi /Gi+1 , i.e. a generating set such e that Gi /Gi+1 has a presentation ai1 , . . . , aimi aijij , [aik , ai` ], for j ∈ Ti , k, ` ∈ {1, . . . , mi } , where Ti ⊆ {1, . . . , mi } (here T stands for torsion) and eij ∈ Z>0 (be aware that we explicitly allow eij = 1, which is necessary for our definition of quotient presentations in Section 3). Formally put eij = ∞ for j ∈ / Ti . Note that A = (a11 , a12 , . . . , acmc ) is a so-called polycyclic generating sequence for G, and we call A a Mal’cev basis associated to the central series (1). Sometimes we use A interchangeably also for the set A = {a11 , a12 , . . . , acmc }. For convenience, we will also use a simplified notation, in which the generators aij and P exponents eij are renumbered by replacing each subscript ij with j + m` , so the generating `<j sequence A can be written as A = (a1 , . . . , am ). We allow the expression ij to stand for P j+ m` in other notations as well. We also denote `<j T = {i \| ei < ∞}. By the choice of {a1 , . . . , am }, every element g ∈ G may be written uniquely in the form αm 1 g = aα 1 · · · am , 6 A. Myasnikov, A. Weiß where αi ∈ Z and 0 ≤ αi < ei whenever i ∈ T . The m-tuple (α1 , . . . , αm ) is called the coordinate vector or Mal’cev coordinates of g and is denoted Coord(g), and the expression αm 1 aα 1 · · · am is called the (Mal’cev) normal form of g. We also denote αi = Coordi (g). To a Mal’cev basis A we associate a presentation of G as follows. For each 1 ≤ i ≤ m, let ni be such that ai ∈ Gni Gni +1 . If i ∈ T , then aei i ∈ Gni +1 , hence a relation aiei = aµ` i` · · · aµmim (2) holds in G for µij ∈ Z and ` > i such that a` , . . . , am ∈ Gni +1 . Let 1 ≤ i < j ≤ m. Since the series (1) is central, relations of the form α ijm aj ai = ai aj a` ij` · · · aα m (3) βij` ai a−1 j a` (4) a−1 j ai = · · · aβmijm hold in G for αijk , βijk ∈ Z and l > j such that a` , . . . , am ∈ Gnj +1 . Now, G is the group with generators {a1 , . . . , am } subject to the relation of the the form (2)–(4). A presentation with relations of the form (2)–(4) for all i resp. i and j is called a nilpotent presentation. Indeed, any presentation of this form will define a nilpotent group. It is called consistent if the order of ai modulo hai+1 , . . . , am i is precisely ei for all i. While presentations of this form need not, in general, be consistent, those derived from a central series of a group G as above are consistent. Given a consistent nilpotent presentation, there is an easy way to solve the word problem: simply apply the rules of the form (3) and (4) to move all occurrences of a±1 1 in the input word to the left, then apply the power relations (2) to reduce their number modulo e1 ; finally, continue with a2 and so on. Multiplication functions. An crucial feature of the coordinate vectors for nilpotent groups β1 αm βm 1 is that the coordinates of a product (aα 1 · · · am )(a1 · · · am ) may be computed as a “nice” function (polynomial if T = ∅) of the integers α1 , . . . , αm , β1 , . . . , βm . I Lemma 3 ([7, 10]). Let G be a nilpotent group with Mal’cev basis a1 , . . . , am and T = ∅. There exist p1 , . . . , pm ∈ Z[x1 , . . . , xm , y1 , . . . , ym ] and q1 , . . . , qm ∈ Z[x1 , . . . , xm , z] such that for g, h ∈ G with Coord(g) = (γ1 , . . . , γm ) and Coord(h) = (δ1 , . . . , δm ) and l ∈ Z we have (i) Coordi (gh) = pi (γ1 , . . . , γm , δ1 , . . . , δm ), (ii) Coordi (g l ) = qi (γ1 , . . . , γm , l), (iii) Coord1 (gh) = γ1 + δ1 and Coord1 (g l ) = lγ1 . Notice that an explicit algorithm to construct the polynomials pi , qi is given in [14]. For further background on nilpotent groups we refer to [7, 10]. 3 Presentation of subgroups Before we start with algorithmic problems, we introduce a canonical way how to represent subgroups of nilpotent groups. This is important for two reasons: first, of course we need it to solve the subgroup membership problem, and, second, for the uniform setting it allows us to represent nilpotent groups as free nilpotent group modulo a kernel which is represented as αim i1 a subgroup. Let h1 , . . . , hn be elements of G given in normal form by hi = aα 1 · · · am , for i = 1, . . . , n, and let H = hh1 , . . . , hn i. We associate the matrix of coordinates   α11 · · · α1m  ..  , .. A =  ... (5) . .  αn1 ··· αnm TC0 circuits for algorithmic problems in nilpotent groups to the tuple (h1 , . . . , hn ) and conversely, to any n × m integer matrix, we associate an n-tuple of elements of G, whose Mal’cev coordinates are given as the rows of the matrix, and the subgroup H generated by the tuple. For each i = 1, . . . , n where row i is non-zero, let πi be the column of the first non-zero entry (‘pivot’) in row i. The sequence (h1 , . . . , hn ) is said to be in standard form if the matrix of coordinates A is in row-echelon form and its pivot columns are maximally reduced (similar to the Hermite normal form), more specifically, if A satisfies the following properties: (i) all rows of A are non-zero (i.e. no hi is trivial), (ii) π1 < π2 < · · · < πs (where s is the number of pivots), (iii) αiπi > 0, for all i = 1, . . . , n, (iv) 0 ≤ αkπi < αiπi , for all 1 ≤ k < i ≤ s (v) if πi ∈ T , then αiπi divides eπi , for i = 1, . . . , s. The sequence (resp. matrix) is called full if in addition (vi) H ∩ hai , ai+1 , . . . , am i is generated by {hj \| πj ≥ i}, for all 1 ≤ i ≤ m. Note that {hj \| πj ≥ i} consists of those elements having 0 in their first i − 1 coordinates. It is an easy exercise (see also [16]) to show that (vi) holds for a given i if, and only if, −1 for all 1 ≤ k < j ≤ s with πk < i, h−1 k hj hk and hk hj hk are elements of h hl \| l > k i, and eπ /αkπ k for all 1 ≤ k ≤ s with πk < i and πk ∈ T , hk k ∈ h hl \| l > k i. We will use full sequences and the associated matrices in full form interchangeably without mentioning it explicitly. For simplicity we assume that the inputs of algorithms are given as matrices. The importance of full sequences is described in the following lemma – a proof can be found in [25] Propositions 9.5.2 and 9.5.3. I Lemma 4 ([16, Lem. 3.1]). Let H ≤ G. There is a unique full sequence U = (h1 , . . . , hs ) that generates H. We have s ≤ m and H = {hβ1 1 · · · hβs s \| βi ∈ Z and 0 ≤ βi < eπi if πi ∈ T }. Thus, computing a full sequence will be the essential tool for solving the subgroup membership problem. Before we focus on subgroup membership, we will first solve the word problem and introduce how the nilpotent group can be part of the input. 3.1 Quotient presentations Let c, r ∈ N be fixed. The free nilpotent group Fc,r of class c and rank r is defined as Fc,r = h a1 , . . . , ar \| [x1 , . . . , xc+1 ] = 1 for x1 , . . . , xc+1 ∈ Fc,r i where [x1 , . . . , xc+1 ] = [[x1 , . . . , xc ], xc+1 ], i. e., Fc,r is the r-generated group only subject to the relations that weight c + 1 commutators are trivial. Throughout, we fix a Mal’cev basis A = (a1 , . . . , am ) (which we call the standard Mal’cev basis) associated to the lower central series of Fc,r such that the associated nilpotent presentation consists only of relations of the form (3) and (4) (i. e., T = ∅ – such a presentation exists since Fc,r is torsion-free), a1 , . . . , ar generates Fc,r , and all other Mal’cev generators are iterated commutators of a1 , . . . , ar . Denote by Nc,r the set of r-generated nilpotent groups of class at most c. Every group G ∈ Nc,r is a quotient of the free nilpotent group Fc,r , i. e., G = Fc,r /N for some normal subgroup N ≤ Fc,r . Assume that T = (h1 , . . . , hs ) is a full sequence generating N . Adding T to the set of relators of the free nilpotent group yields a new nilpotent presentation. This presentation will be called quotient presentation of G. For inputs of algorithms, we assume that a quotient presentation is always given as its matrix of coordinates in full form. Depending whether the entries of the matrix are encoded in unary or binary, we call the quotient presentation be given in unary or binary. 7 8 A. Myasnikov, A. Weiß I Lemma 5 ([16, Prop. 5.1]). Let c and r be fixed integers and let A = (a1 , . . . , am ) be the standard Mal’cev basis of Fc,r . Moreover, denote by S the set of relators of Fc,r with respect to A. Let G ∈ Nc,r with G = Fc,r /N and let T be the full-form sequence for the subgroup N of Fc,r . Then, h A \| S ∪ T i is a consistent nilpotent presentation of G. Proof. Clearly, we have G ' hA \| S ∪ T i. Since hA \| Si is a nilpotent presentation and the elements of T add relators of the form (2), the presentation is nilpotent. To prove that it is consistent, suppose some ai ∈ A has order αi modulo hai+1 , . . . , am i in hA \| S ∪ T i. Since αm i αi+1 the order is infinite in F , there must be element of the form aα i ai+1 · · · am in N . But then, 0 α0 α α0 i+1 by Lemma 4, T must contain an element ai i ai+1 · · · amm where αi0 divides αi . Hence αi cannot be smaller than αi0 and so the presentation is consistent. J For the following we always assume that a quotient presentation is part of the input, but c and r are fixed. Later, we will show how to compute quotient presentations from an arbitrary presentation. I Remark. Lemma 5 ensures that each group element has a unique normal form with respect to the quotient presentation; thus, it guarantees that all our manipulations of Mal’cev coordinates are well-defined. 4 Word problem and computation of Mal’cev coordinates In this section we deal with the word problem of nilpotent groups, which is well-known to be in TC0 [24]. Here, we generalize this result by allowing words with binary exponents (recall that word with binary exponents is a sequence w = w1x1 · · · wnxn where wi ∈ {a1 , . . . , am } and the xi ∈ Z). By using words with binary exponents the input can be compressed exponentially – making the word problem, a priori, harder to solve. Nevertheless, it turns out that the word problem still can be solved in TC0 when allowing the input to be given as a word with binary exponents. Note that this contrasts with the situation where the input is given as straight-line program (which like words with binary exponents allow an exponential compression) – then the word problem is complete for the counting class C= L [12]. I Theorem 6. Let c, r ≥ 1 be fixed and let (a1 , . . . , am ) be the standard Mal’cev basis of Fc,r . The following problem is TC0 -complete: on input of G ∈ Nc,r given as a binary encoded quotient presentation and a word with binary exponents w = w1x1 · · · wnxn , compute integers y1 , . . . , ym (in binary) such that w = ay11 · · · aymm in G and 0 ≤ yi < ei for i ∈ T . Moreover, if the input is given in unary (both G and w), then the output is in unary. Note that the statement for unary inputs is essentially the one of [24]. Be aware that in the formulation of the theorem, T and ei for i ∈ T depend on the input group G. These parameters can be read from the full matrix (αij )i,j of coordinates representing G (recall that πi denotes the column index of the i-th pivot and here s is the number of rows of the matrix): T = {πi \| i ∈ {1, . . . , s}} (all columns which have a pivot) and ei = αji if πj = i. As an immediate consequence of Theorem 6, we obtain: I Corollary 7. Let c, r ≥ 1 be fixed. The uniform, binary version of the word problem for groups in Nc,r is TC0 -complete (where the input is given as in Theorem 6). TC0 circuits for algorithmic problems in nilpotent groups The proof of Theorem 6 follows the outline given in Section 2.2; however, we cannot apply the rules (2)–(4) one by one. Instead we make only two steps for each generator: first apply all possible rules (3) and (4) in one step and then apply the rules (2) in one step. Proof of Theorem 6. The hardness part is clear since already the word problem of Z is TC0 -complete. For describing a TC0 circuit, we proceed by induction along the standard Mal’cev basis (a1 , . . . , am ) of the free nilpotent group Fc,r . If w does not contain any letter a1 , we have y1 = 0 and we can compute yi for i > 1 by induction. Otherwise, we rewrite w as ay11 uv (with 0 ≤ y1 < e1 if 1 ∈ T ) such that u and v are words with binary exponents not containing any a1 s. Once this is completed, uv can be rewritten as ay22 · · · aymm by induction. For computing y1 , u and v, we proceed in two steps: P First, we rewrite w as aỹ11 v with ỹ1 = wi =a1 xi (this is possible by Lemma 3 (iii)). The exponent ỹ1 can be computed by iterated addition, which by Theorem 2 is in TC0 (in the unary case ỹ1 can be written down in unary). Now, v consists of what remains from w after a1 has been “eliminated”: for every position i in w with wi = 6 a1 , we compute P zi = xj using iterated addition. Let wi = ak . By Lemma 3 (i) there are fixed j>i wj =a1 polynomials pk,k+1 , . . . , pk,m ∈ Z[x, y] such that in the free nilpotent group holds p k,k+1 axk ay1 = ay1 axk ak+1 (x,y) p · · · amk,m (x,y) for all x, y ∈ Z. Hence, in order to obtain w̃, it remains to replace every wixi with wi = a1 by the empty word pk,k+1 (xi ,zi ) p (x ,z ) and every wixi with wi = ak 6= a1 by axki ak+1 · · · amk,m i i , which is a word with binary exponents (resp. as a word of polynomial length in the unary case), for k = 2, . . . , m. The exponents can be computed in TC0 by Theorem 2. Since the pk,i are bounded by pk,k+1 (xi ,zi ) p (x ,z ) polynomials, in the unary case, axki ak+1 · · · amk,m i i can be written as a word without exponents. The second step is only applied if 1 ∈ T (as explained above, this can be decided and ei can be read directly from the quotient presentation by checking whether there is a pivot in the first column) – otherwise y1 = ỹ1 and u is the empty word. We rewrite aỹ11 to ay11 u with y1 = ỹ1 mod e1 and a word with binary exponents u not containing any a1 . Again y1 can be computed in TC0 by Theorem 2. Let ae11 = aµ2 12 · · · aµm1m be the power relation for a1 (which can be read from the quotient presentation – it is just the row where the pivot is in the first column) and write ỹ1 = s · e1 + y1 . Now, u should be equal to (aµ2 12 · · · aµm1m )s in Fc,r . We use the fixed polynomials qi ∈ Z[x1 , . . . , xm , z] from Lemma 3 (ii) for Fc,r yielding q (0,µ12 ,...,µ1m ,s) u = a22 · · · aqmm (0,µ12 ,...,µ1m ,s) (which, in the binary setting, is a word with binary exponents, and in the unary setting a word without exponents of polynomial length). Now, we have w = ay11 uv in G as desired. J 5 The extended gcd problem Computing greatest common divisors and expressing them as a linear combination is an essential step for solving the subgroup membership problem. Indeed, consider the nilpotent group Z and let a, b, c ∈ Z. Then c ∈ ha, bi if, and only if, gcd(a, b) \| c. Binary gcds. The (binary) extended gcd problem (ExtGCD) is as follows: on input of binary encoded numbers a1 , . . . , an ∈ Z, compute x1 , . . . , xn ∈ Z such that x1 a1 + · · · + xn an = gcd(a1 , . . . , an ). 9 10 A. Myasnikov, A. Weiß Clearly this can be done in P using the Euclidean algorithm, but it is not known whether it is actually in NC. Since we need to compute greatest common divisors, we will reduce the subgroup membership problem to the computation of gcds. Unary gcds. Computing the gcd of numbers encoded in unary is straightforward in TC0 by an exhaustive search; yet, it is not obvious how to express gcd(a1 , . . . , an ) as x1 a1 +· · ·+xn an in TC0 . By [17] such xi with \|xi \| ≤ 12 max{\|a1 \|, . . . , \|an \|} can be computed in LOGSPACE. However, that algorithm uses a logarithmic number of rounds each depending on the outcome of the previous one – so it does not work in TC0 . Note that for n = 2 the problem is easy: I Example 8. Let a, b ∈ Z. Then, there are x, y ∈ Z with \|x\| , \|y\| ≤ max {\|a\| , \|b\|} such that ax + by = gcd(a, b). This is easy to see: assume a, b > 0 (the other cases are similar) and we are given x, y with ax + by = gcd(a, b) and x ≥ b, then we can replace x with x − b and y with y + a. This does not change the sum and by iterating this step, we can assure that 0 ≤ x < b. Then we have y = − ax−gcd(a,b) ; hence, −a < y ≤ 1. b If a and b are given in unary, the coefficients x, y can be computed in TC0 by simply checking all (polynomially many) values for x and y with \|x\| , \|y\| ≤ max {\|a\| , \|b\|}. However, if we want to express the gcd of unboundedly many numbers ai as a linear combination, we cannot check all possible values for x1 , . . . , xn in TC0 because there are max{\|a1 \|n , . . . , \|an \|n } (i. e., exponentially) many. Expressing the gcd as a linear combination can be viewed as a linear equation with integral coefficients. Recently, in [5, Thm. 3.14] it has been shown that, if all the coefficients are given in unary, it can be decided in TC0 whether such an equation or a system of a fixed number of equations has a solution. Since from the proof of [5, Thm. 3.14] it is not obvious how to find an actual solution, we prove the following result: I Theorem 9. The following problem is in TC0 : Given integers a1 , . . . , an as unary numbers, compute x1 , . . . , xn ∈ Z (either in unary or binary) such that x1 a1 + · · · + xn an = gcd(a1 , . . . , an ) 2 with \|xi \| ≤ (n + 1) (max{\|a1 \|, . . . , \|an \|}) . Proof. Let A = max{\|a1 \|, . . . , \|an \|}, which clearly can be computed in TC0 . W. l. o. g. we assume that all the ai are positive. We assume that all numbers which appear as intermediate results are encoded in binary (indeed, these numbers will grow too fast to encode them in unary). First observe that gcd(a1 , . . . , ai ) can be computed in TC0 for all i ∈ {1, . . . , n}. The reason is simply that there are only linearly many numbers less than each ai . In fact, for computing gcd(a1 , . . . , an ), the circuit just checks for all d ≤ A whether for every i there is some ci ≤ ai with dci = ai . If for some d there are such ci for all i, we have found a common divisor. The gcd is simply the largest one. Thus, it remains to compute the coefficients xi . Since we can compute gcd(a1 , . . . , an ) in 0 TC , we can divide all numbers ai by the gcd and henceforth assume that gcd(a1 , . . . , an ) = 1 (note that this does not change the coefficients xi ). The first step for computing the xi s, is to compute di = gcd(a1 , . . . , ai ) for i = 1, . . . , n and d0 = 0 (note that by our assumption, dn = 1). We have di = gcd(a1 , . . . , ai ) = gcd(gcd(a1 , . . . , ai−1 ), ai ) = gcd(di−1 , ai ). TC0 circuits for algorithmic problems in nilpotent groups 11 Using this observation, the next step computes for each i integers yi and zi such that di = yi di−1 + zi ai . For all i this can be done in parallel in TC0 by simply trying all possible values with \|yi \| , \|zi \| ≤ A as in Example 8. We set n Y xi = zi yj . j=i+1 These xi can be computed in TC0 using iterated multiplication [8] – see Theorem 2. Moreover, an easy induction shows that x1 a1 + · · · + xn an = gcd(a1 , . . . , an ). There is only one problem with the numbers xi : in general, they do not meet the bounds \|xi \| ≤ (n + 1)A2 . So, the next step will be to modify these xi in such a way that they meet the desired bound. The idea is to apply a sequence of operations as in Example 8 to make the coefficients small. The difficulty here is to find out where exactly to add/subtract a multiple of which ai . Let P = {i ∈ {1, . . . , n} \| xi > 0} and N = {i ∈ {1, . . . , n} \| xi < 0}. Note that P ∩N = ∅ and w. l. o. g. we can assume that P ∪ N = {1, . . . , n}. For all i = 1, . . . n, we set j x a k −xi ai i i 0 p0i = max 0, , n = max 0, . (6) i A2 A2 Obviously, we have p0i = 0 for i ∈ N and n0i = 0 for i ∈ P. The non-zero p0i correspond to those indices which have a too large positive xi and the non-zero n0i to those indices which have a too small negative xi (this is because we assumed the ai to be positive). Moreover, xi should be decreased (resp. increased) by A2 p0i /ai (resp. A2 n0i /ai ) in order to make it reasonably small. We will not be able to reach this aim completely, but with a sufficiently small error. Pi Pi Next, we set Pi0 = j=1 p0j and Ni0 = j=1 n0j . All the p0i , n0i , Pi0 , Ni0 and P and N can be computed in TC0 using iterated addition and division – see Theorem 2. I Lemma 10. Pn0 − Nn0 ≤ \|N \| and Nn0 − Pn0 ≤ \|P\| Proof. For i ∈ P, we have 0 ≤ xi ai − p0i A2 < A2 by definition of p0i . Likewise, we have 0 ≥ xi ai + n0i A2 > −A2 for i ∈ N . Since P ∩ N = ∅ and P ∪ N = {1, . . . , n}, we obtain (Pn0 − Nn0 )A2 = X p0i A2 − i∈P X n0i A2 < i∈N X xi ai + i∈P X (xi ai + A2 ) i∈N 2 = x1 a1 + · · · + xn an + \|N \| A = 1 + \|N \| A2 meaning that Pn0 − Nn0 ≤ \|N \|. The same argument yields (Pn0 − Nn0 )A2 > 1 − \|P\| A2 , and thus Nn0 − Pn0 < \|P\|. J Let D = Nn0 − Pn0 . For i ∈ {1, . . . , n}, we set pi = ( p0i + 1 p0i if i ∈ P and i ≤ D, otherwise, ni = ( n0i + 1 n0i if i ∈ N and i ≤ −D, otherwise, (7) 12 A. Myasnikov, A. Weiß Pi Pi and Pi = j=1 pj and Ni = j=1 nj for i ∈ {0, . . . , n}. Because of Lemma 10, we have Nn = Pn . Clearly, the pi , ni , Pi , Ni can be computed in TC0 and from now on we will work with these numbers. Also, as an immediate consequence of (6) and (7), we have −A2 ≤ xi ai − pi A2 ≤ A2 2 2 −A ≤ xi ai + ni A ≤ A for i ∈ P, 2 for i ∈ N . (8) Now, for i ∈ P and j ∈ N , we define pj,i =   pi       Nj − Pi−1 if Nj−1 ≤ Pi−1 < Pi ≤ Nj if Nj−1 ≤ Pi−1 < Nj ≤ Pi Pi − Nj−1     nj    0 if Pi−1 ≤ Nj−1 < Pi ≤ Nj if Pi−1 ≤ Nj−1 < Nj ≤ Pi otherwise. Note that the cases overlap. However, then the different definitions of pj,i agree. For i ∈ N and j ∈ P, we set pj,i = pi,j and for i, j ∈ P or i, j ∈ N we set pj,i = 0. X X I Lemma 11. We have pj,i = pi and pj,i = nj . j i P Proof. We only show j pj,i = pi ; the other statement follows by symmetry. First, assume that pi = pi,j for some j. Then pi,j 0 = 0 for all j 0 6= j; hence, the lemma holds. Now, let pi 6= pi,j for any j. We define αi = min {j ∈ {1, . . . , n} \| Pi−1 < Nj } , βi = max {j ∈ {1, . . . , n} \| Nj−1 < Pi } . In particular, we have pj,i = 0 for j < αi or j > βi . Notice that αi and βi exist for all i ∈ P (since Nn = Pn ). Also αi < βi because αi = βi = j implies Nj−1 ≤ Pi−1 < Nj and Nj−1 < Pi ≤ Nj ; thus, pj,i = pi . Moreover, we have pαi ,i = Nαi −Pi−1 and pβi ,i = Pi −Nβi −1 and pj,i = nj for αi < j < βi . Since Pi − Pi−1 = i X pi − j=0 Nβi −1 − Nαi − i−1 X pi = pi and j=0 βX i −1 j=αi +1 nj = βX i −1 nj − j=1 αi X nj − j=1 βX i −1 nj = 0, j=αi +1 we obtain X pj,i = Nαi − Pi−1 + Pi − Nβi −1 + j βX i −1 nj = p i . j=αi +1 J We set yj,i = j k 2 pj,i A ai aj for i, j = 1, . . . , n. Notice that, since ai aj ≤ A2 , this means that (pj,i − 1)A2 < yj,i ai aj ≤ pj,i A2 . (9) TC0 circuits for algorithmic problems in nilpotent groups Finally, we define our new  P  xi − j yj,i aj  P x̃i = xi + j yi,j aj   x i 13 coefficients x̃i as follows: if i ∈ P, if i ∈ N , otherwise. It remains to show the following: (i) the numbers x̃i can be computed in TC0 , (ii) x̃1 a1 + · · · + x̃n an = 1, (iii) \|x̃i \| ≤ (n + 1)A2 for all i. The first point is straightforward: we already remarked that the pi , ni , Pi , Ni and P and N can be computed in TC0 . Hence, also the pj,i can be computed in TC0 (as simple Boolean combination resp. addition of the previous numbers). Now, the yj,i can be computed using division [8]. Finally, the computation of the x̃i is simply another application of iterated addition. For the second point observe that X X x̃1 a1 + · · · + x̃n an = x̃i ai + x̃i ai i∈P i∈N  = X xi − = n X i=1 n X xi ai − xi ai − n X  X  xi + j XX yj,i aj ai + i∈N XX i∈P j∈N yj,i aj ai + X yi,j aj  ai j i∈N XX i∈P i=1 = yj,i aj  ai + j i∈P =  X yi,j aj ai j XX yi,j aj ai i∈N j∈P xi ai i=1 The last equality is due to the fact that yj,i = yi,j for all i, j and that yi,j = 0 if i and j are both in P or both in N . For the third point, let i ∈ P. Then, X X x̃i ai = xi ai − yj,i aj ai ≥ xi ai − pj,i A2 (by (9)) j = xi ai − pi A ≥ −A j 2 (by Lemma 11) 2 (by (8)) and x̃i ai = xi ai − X yj,i aj ai ≤ xi ai − j = xi ai − A2 pi + nA2 ≤ (n + 1)A X (pj,i − 1)A2 (by (9)) j (by Lemma 11) 2 The case i ∈ N is completely symmetric. This concludes the proof of Theorem 9. (by (8)) J Notice that it is straightforward to improve the bounds of Theorem 9 further (e. g. getting rid of the factor n + 1). However, since there is no need for that in order to perform the matrix reduction, we do not do this additional effort. Also we could not find a TC0 circuit which yields the bound xi ≤ 12 A (which is achievable in LOGSPACE by [17]). 14 A. Myasnikov, A. Weiß 6 Matrix reduction and subgroup membership problem In [16], the so-called matrix reduction procedure converts an arbitrary matrix of coordinates into its full form and, thus, is an essential step for solving the subgroup membership problem and several other problems. It was first described in [25] – however, without a precise complexity estimate. In this section, we repeat the presentation from [16] and show that for fixed c and r, it can be actually computed uniformly for groups in Nc,r in TC0 – in the case that the inputs are given in unary (as words). If the inputs are represented as words with binary exponents, then we still can show that it is TC0 -Turing-reducible to ExtGCD. In Section 3, we defined the matrix representation of subgroups of nilpotent groups. We adopt all notation from Section 3. As before, let c, r ∈ N be fixed and let (a1 , . . . , am ) be the standard Mal’cev basis of Fc,r . Let G ∈ Nc,r be given as quotient presentation, i. e., as a matrix in full form (either with unary or binary coefficients). We define the following operations on tuples (h1 , . . . , hn ) (our subgroup generators) of elements of G and the corresponding operations on the associated matrix, with the goal of converting (h1 , . . . , hn ) to a sequence in full form generating the same subgroup H = hh1 , . . . , hn i: (1) Swap hi with hj . This corresponds to swapping row i with row j. (2) Replace hi by hi hlj (i 6= j, l ∈ Z). This corresponds to replacing row i by Coord(hi hlj ). (3) Add or remove a trivial element from the tuple. This corresponds to adding or removing a row of zeros; or (3’) a row of the form (0 . . . 0 ei αi+1 . . . αm ), where i ∈ T and αi+1 i m a−e = ai+1 · · · aα m . i −1 (4) Replace hi with h−1 i . This corresponds to replacing row i by Coord(hi ). (5) Append an arbitrary product hli11 · · · hlikk with i1 , . . . , ik ∈ {1, . . . , n} and l1 , . . . , lk ∈ Z to the tuple: add a new row with Coord(hli11 · · · hlikk ). Clearly, all these operations preserve H. I Lemma 12. On input of a quotient presentation of G ∈ Nc,r in unary (resp. binary) and a matrix of coordinates A given in unary (resp. binary), operations (1)–(5) can be done in TC0 . The output matrix will be also encoded in unary (resp. binary). For operations (2) and (5), we require that the exponents l, l1 , . . . , lk are given in unary (resp. binary). Moreover, as long as the rows in the matrix which are changed are pairwise distinct, a polynomial number of such steps can be done in parallel in TC0 . Proof. Operations (1) and (3), clearly can be done in TC0 . Notice that operation (3’) means simply that a row of the quotient presentation of G is appended to the matrix. In the unary case, it follows directly from Theorem 6 that operations (2), (4), and (5) are in TC0 because, since l, l1 , . . . , lk are given in unary, the respective group elements can be written down as words. In the case of binary inputs, (5) works as follows ((2) and (4) analogously): by Lemma 3 (ii), there are functions q1 , . . . , qm ∈ Z[x1 , . . . , xm , z] such that for every h ∈ Fc,r with Coord(h) = (γ1 , . . . , γm ) anda l ∈ Z, we have Coordi (hl ) = qi (γ1 , . . . , γm , l) in Fc,r . These l functions can be used to compute Coord(hijj ) for j = 1, . . . , k. After that, hli11 · · · hlikk can be written down as word with binary exponents and Theorem 6 can be applied. J Using the row operations defined above, in [16] it is shown how to reduce any coordinate matrix to its unique full form. Let us repeat these steps: Let A0 be a matrix of coordinates, as in (5) in Section 3. Recall that πk denotes the column index of the k-th pivot (of the full form of A0 ). We produce matrices A1 , . . . , As , where s is the number of pivots in the full form of A0 , such that for every k = 1, . . . , s the TC0 circuits for algorithmic problems in nilpotent groups 15 first πk columns of Ak form a matrix satisfying conditions (ii)-(v) of being a full sequence, condition (vi) is satisfied for all i < πk+1 , and As is the full form of A0 . Here we formally denote πs+1 = m + 1. Set π0 = 0 and assume that Ak−1 has been constructed for some k ≥ 1. In the steps below we construct Ak . We let n and m denote the number of rows and columns, respectively, of Ak−1 . At all times during the computation, hi denotes the group element corresponding to row i of Ak and αij denotes the (i, j)-entry of Ak , which is Coordj (hi ). These may change after every operation. Step 1. Locate the column πk of the next pivot, which is the minimum integer πk−1 < πk ≤ m such that αiπk 6= 0 for at least one k ≤ i ≤ n. If no such integer exists, then k − 1 = s and As is already constructed. Otherwise, set Ak to be a copy of Ak−1 and denote π = πk . Compute a linear expression of d = gcd(αkπ , . . . , αnπ ) = lk αkπ + · · · + ln αnπ . Let hn+1 = hlkk · · · hlnn and note that hn+1 has coordinates of the form Coord(hn+1 ) = (0, . . . , 0, d, . . .) with d occurring in position π. Perform operation (5) to append hn+1 as row n + 1 of Ak . −α /d Step 2. For each i = k, . . . , n, perform operation (2) to replace row i by Coord(hi · hn+1iπ ). −bα /dc and for each i = 1, . . . , k − 1, use (2) to replace row i by Coord(hi · hn+1iπ ). After that, swap row k with row n + 1 using (1). At this point, properties (ii)-(iv) hold on the first k columns of Ak . Step 3. If π ∈ T , we additionally ensure condition (v) as follows. Perform row operation (3’), with respect to π, to append a trivial element hn+2 with Coord(hn+2 ) = (0, . . . , 0, eπ , . . .) to Ak . Let δ = gcd(d, eπ ) and compute the linear expression δ = n1 d + n2 eπ , with 2 \|n1 \|, \|n2 \| ≤ max{d, eπ }. Let hn+3 = hnk 1 hnn+2 and append this row to Ak , as row n + 3. Note that Coord(hn+3 ) = (0, . . . , 0, δ, . . .), with δ in position π. Replace row k by −eπ /δ −d/δ Coord(hk · hn+3 ) and row n + 2 by Coord(hn+2 · hn+3 ), producing zeros in column π in these rows. Swap row k with row n + 3. At this point, (ii), (iii), and (v) hold (for the first πk columns) but (iv) need not, since the pivot entry is now δ instead of d. For each −bα /δc j = 1, . . . , k − 1, replace row j by Coord(hj · hk jπ ), ensuring (iv). Step 4. Identify the next pivot πk+1 (like in Step 1). If πk is the last pivot, we set πk+1 = m + 1. We now ensure condition (vi) for i < πk+1 . Observe that Steps 1-3 preserve h hj \| πj ≥ i i for all i < πk . Hence (vi) holds in Ak for i < πk since it holds in Ak−1 for the same range. Now consider i in the range πk ≤ i < πk+1 . It suffices to establish (vi.i) for all j > k and (vi.ii) for πk only. To obtain (vi.i), notice that −1 ±1 h−1 k hj hk , hk hj hk ∈ h h` \| ` > k i if, and only if, [hj , hk ] ∈ h h` \| ` > k i. Further, note that the subgroup generated by Sj = {1, hj , [hj , hk ], . . . , [hj , hk , . . . , hk ]}, where hk appears m − πk times in the last commutator, is closed under commutation with hk since if hk appears more than m−πk times then the commutator is trivial. An inductive ` argument shows that the subgroup hSj i coincides with hh−` k hj hk \| 0 ≤ ` ≤ m − πk i. −1 Similar observations can be made for conjugation by hk . Therefore, appending via ` operation (5) rows Coord(h−` k hj hk ) for all 1 ≤ \|`\| ≤ m − πk and all k < j ≤ n + 3 delivers (vi.i) for all j > k. Note that (vi.i) remains true for i < πk . ek /αkπk To obtain (vi.ii), in the case πk ∈ T , we add row Coord(hk commutes with hk and therefore (vi.i) is preserved. ). Note that this element 16 A. Myasnikov, A. Weiß Step 5. Using operation (3), eliminate all zero rows. The matrix Ak is now constructed. We have to show that each step can be performed in TC0 given that all Mal’cev coordinates are encoded in unary (resp. in TC0 (ExtGCD) if Mal’cev coordinates are encoded in binary). Since the total number of steps is constant (only depending on the nilpotency class and number of generators), this gives a TC0 (resp. TC0 (ExtGCD)) circuit for computing the full form of a given subgroup. Step 1. The next pivot can be found in TC0 since it is simply the next column in the matrix with a non-zero entry, which can be found as a simple Boolean combination of test whether the entries are zero. In the unary case, by Theorem 9, d = gcd(αkπ , . . . , αnπ ) can computed in TC0 together with lk , . . . , ln encoded in unary such that d = lk αkπ + · · · + ln αnπ . Now, by Lemma 12, Step 1 can be done in TC0 . In the binary case, d and lk , . . . , ln can be computed using ExtGCD. Hence, by Lemma 12, Step 1 can be done in TC0 (ExtGCD). Step 2. The numbers bαiπ /dc (either in unary or binary) can be computed in TC0 for all i in parallel by Theorem 2. After that one operation (2) is applied to each row of the matrix. By Lemma 12, this can be done in parallel for all rows in TC0 . Finally, swapping rows k and n + 1 can be done in TC0 . Step 3. As explained in Section 4, T and ei for i ∈ T can be read directly from the quotient presentation. Thus, it can be decided in TC0 whether Step 3 has to be executed. Appending a new row is in TC0 . Computing gcd(d, eπ ) = d = n1 dn2 eπ is in TC0 by Example 8 (in the unary case) and in TC0 (ExtGCD) in the binary case. After that one operation (5) is followed by two operations (2), one operation (1), and, finally, k − 1 times operation (2), which all can be done in TC0 again by Lemma 12. Step 4. The next pivot can be found in TC0 as outlined in Step 1. After that, Step 4 consists of an application of a constant number (only depending on the nilpotency class and number of generators) of operations (5) and thus, by Lemma 12, is in TC0 . Step 5. Clearly that is in TC0 . Thus, we have completed the proof of our main result: I Theorem 13. Let c, r ∈ N be fixed. The following problem is in TC0 : given a unary encoded quotient presentation of G ∈ Nc,r and h1 , . . . , hn ∈ G, compute the full form of the associated matrix of coordinates encoded in unary and hence the unique full-form sequence (g1 , . . . , gs ) generating hh1 , . . . , hn i. Moreover, if the G and h1 , . . . , hn are given in binary, then the full-form sequence with binary coefficients can be computed in TC0 (ExtGCD). 6.1 Subgroup membership problem We can now apply the matrix reduction algorithm to solve the subgroup membership problem in TC0 . I Theorem 14. Let c, r ∈ N be fixed. The following problem is in TC0 (resp. TC0 (ExtGCD) for binary inputs): given a quotient presentation of G ∈ Nc,r , elements h1 , . . . , hn ∈ G and h ∈ G, decide whether or not h is an element of the subgroup H = hh1 , . . . , hn i. Moreover, if h ∈ H, the circuit computes the unique expression h = g1γ1 · · · gsγs where (g1 , . . . , gs ) is the full-form sequence for H with the γi encoded in unary (resp. binary). Alternatively, for unary inputs, the output can be given as word h = hi11 · · · hitt where ij ∈ {1, . . . , n} and j = ±1. Note that we do not know whether there is an analog of the second type of output for binary inputs. A possible way of expressing the output would be as a word with binary exponents TC0 circuits for algorithmic problems in nilpotent groups over h1 , . . . , hn . However, simply applying the same procedure as for unary inputs will not lead to a word with binary exponents. Proof. The circuit works as follows: first, the the full form A of the coordinate matrix corresponding to H and the standard-form sequence (g1 , . . . , gs ) are computed in TC0 (resp. TC0 (ExtGCD)) using Theorem 13. As before, denote by αij the (i, j)-entry of A and by π1 , . . . , πs its pivots. By Lemma 4, any element of H can be written as g1γ1 · · · gsγs . We show how to find these (j) (j) exponents. Denote h(1) = h and Coord(h(j) ) = (β1 , . . . , βm ), with h(j) being defined below. (j) For j = 1, . . . , s, do the following. If βl 6= 0 for any 1 ≤ l < πj , then h ∈ / H. Otherwise, (j) check whether αjπj divides βπj . If not, then h ∈ / H. If yes, let γj = βπ(j) /αjπj j and −γj h(j+1) = gj h(j) . If j < s, continue to j + 1. If j = s, then h = g1γ1 · · · gsγs ∈ H if h(s+1) = 1 and h ∈ / H otherwise. Since s is bounded by a constant, there are only a constant number of steps. Each step can be done in TC0 by Theorem 2 (division) and Theorem 6 (computation of Mal’cev coordinates). For the second type of output in the unary case, while performing the matrix reduction, we store for every row of the matrix also how that row can be expressed as a word over the subgroup generators h1 , . . . , hn (here, we need the unary inputs, as otherwise the group elements cannot be expressed as words in polynomial space). In every operation on the matrix these words are updated correspondingly, which clearly can be done in TC0 . In the end after writing h = g1γ1 · · · gsγs , every gi can be substituted by the respective word. J Since abelian groups are nilpotent, we obtain: I Corollary 15. Let r be fixed. The following problem is in TC0 : Given a list h1 , . . . , hn ∈ Zr and g ∈ Zr (all as words over the generators), decide whether g ∈ hh1 , . . . , hn i. Moreover, in the case of a positive answer, compute x1 , . . . , xn ∈ Z in unary such that g = x1 h1 +· · ·+xn hn . In other words: for fixed r, given a unary encoded system of linear equations (A, b) with A ∈ Zr×n and b ∈ Zr , a unary encoded solution x ∈ Zn with Ax = b can be computed in TC0 . 6.2 Subgroup presentations The full-form sequence associated to a subgroup H forms a Mal’cev basis for H. This allows us to compute a consistent nilpotent presentation for H. Note, however, that the resulting presentation is not a quotient presentation (although it can be transformed into one, see Proposition 20) – partly this is due to the fact that, in general, H ∈ / Nc,r . The following is the extended version of [16, Thm. 3.11]: I Theorem 16. Let c, r ∈ N be fixed. The following is in TC0 for unary inputs and in TC0 (ExtGCD) for binary inputs: Input: a quotient presentation for G ∈ Nc,r and elements h1 , . . . , hn ∈ G. Output: a consistent nilpotent presentation for H = hh1 , . . . , hn i given by a list of generators (g1 , . . . , gs ) and numbers µij , αijk , βijk ∈ Z encoded in unary (resp. binary) for 1 ≤ i < j < k ≤ s representing the relations (2)-(4). 17 18 A. Myasnikov, A. Weiß Proof. First, the full sequence (g1 , . . . , gs ) for H is computed in TC0 (resp. TC0 (ExtGCD)) according to Theorem 13. Let Hi = hgi , gi+1 , . . . , gs i. In the proof of [16, Thm. 3.11], it is shown that (g1 , . . . , gs ) is a Mal’cev basis for H. Hence, it remains to compute the relators (2)-(4) in order to give a consistent nilpotent presentation of H. The order e0i of gi modulo Hi+1 is simply ei /Coordπi (gi ) (as before T and ei for i ∈ T can be read from the quotient presentation). Each relation (2) can be computed using the TC0 (resp. TC0 (ExtGCD)) e0 e0 circuit of Theorem 14 with input gi i and Hi+1 = hgi+1 , . . . , gs i. Since gi i ∈ Hi+1 and (gi+1 , . . . , gs ) is the unique full sequence for Hi+1 , the membership algorithm returns the expression on the right side of (2). Relations (3) and (4) are established using the same method. Note that there are only a constant number of relations to establish – so everything can be done in TC0 (resp. TC0 (ExtGCD)). J 7 More algorithmic problems 7.1 Homorphisms and kernels Given nilpotent groups G and H and a subgroup K ≤ G and a generating set g1 , . . . , gn of K, a homomorphism ϕ : K → H can be specified by a list of elements h1 , . . . , hn where ϕ(gi ) = hi for i = 1, . . . , n. For a homomorphism, we consider the problem of finding a generating set for its kernel, and given h ∈ ϕ(K) finding g ∈ G such that ϕ(g) = h. Following [16], both problems are solved using matrix reduction in the group H × G. I Theorem 17 (Kernels and preimages). Let c, r ∈ N be fixed. The following is in TC0 for unary inputs and in TC0 (ExtGCD) for binary inputs: On input of G, H ∈ Nc,r given as quotient presentations, a subgroup K = hg1 , . . . , gn i ≤ G, a list of elements h1 , . . . , hn defining a homomorphism ϕ : K → H via ϕ(gi ) = hi , and optionally, an element h ∈ H guaranteed to be in the image of ϕ, compute (i) a generating set X for the kernel of ϕ, and (ii) an element g ∈ G such that ϕ(g) = h. In case of unary inputs, X and g will be returned as words, and for binary inputs, as words with binary exponents. Proof. Let (a1 , . . . , am ) be the standard Mal’cev basis of Fc,r and (b1 , . . . , bm0 ) the standard Mal’cev basis of Fc,2r We have two embeddings of ϕH , ϕG : Fc,r → Fc,2r with ϕH (ai ) = bi and ϕG (ai ) = br+i for i = i, . . . , r. We can assume that the Mal’cev basis of Fc,2r is chosen in such a way that these embeddings send all Mal’cev generators of Fc,r to Mal’cev generators of Fc,2r . Note that we have ϕH (Fc,r ) ∩ ϕG (Fc,r ) = {1}. Thus, we can read all relators of H and G in Fc,2r via the embeddings ϕH and ϕG , respectively. To obtain a quotient presentation of H × G, we simply need to add the relations that H and G commute – that is we need to introduce additional relations bi = 1 for all Mal’cev generators which are not in the image of ϕG or ϕH . As the new quotient presentation is basically a copy of those of H and G, it can be computed in TC0 . From now on we work only in the direct product H × G ∈ Nc,2r and identify G and H with their images under ϕG and ϕH . Let Q = hhi gi \| 1 ≤ i ≤ ni and let W = (v1 u1 , . . . , vs us ) be the sequence in full form for the subgroup Q, where ui ∈ G and vi ∈ H. Let 0 ≤ r ≤ s be the greatest integer such that vr = 6 1 (with r = 0 if all vi are 1). Set X = (ur+1 , . . . , un ) and Y = (v1 , . . . , vr ). In [16, Thm. TC0 circuits for algorithmic problems in nilpotent groups 4.1] it is shown that X is the full-form sequence for the kernel of ϕ and Y is the full-form sequence for the image. Now, to solve (i), it suffices to compute W using Theorem 13 and return the corresponding X as defined above. For (ii), apply Theorem 14 to express h as h = v1β1 · · · vrβr – then return g = uβ1 1 · · · uβr r . J 7.2 Centralizers Before we focus on the conjugacy problem, we need one more preliminary result: the problem of computing centralizers. I Theorem 18 (Centralizers). Let c, r ∈ N be fixed. The following is in TC0 for unary inputs and in TC0 (ExtGCD) for binary inputs: On input of some G ∈ Nc,r given as quotient presentation and an element g ∈ G, compute a generating set X for the centralizer of g in G (in case of binary inputs, the generating set will be given as set of words with binary exponents). Proof. Let Fc,r = Γ0 ≥ Γ1 ≥ · · · ≥ Γc+1 = 1 be the lower central series of Fc,r . Clearly this central series projects onto a central series of G and we simply write Γi for its projection in G. Denote with A = (a1 , . . . am ) the standard Mal’cev basis of Fc,r , which is associated to the lower central series – in particular a1 , . . . , ar is a generating set for Fc,r . We proceed by induction on c. If c = 1, then Fc,r and G are abelian and C(g) = G so the output is {a1 , . . . , ar }. Assume that the theorem holds for groups in Nc−1,r – in particular, for G/Γc (we obtain a quotient presentation of G/Γc by simply forgetting about the Mal’cev generators in Γc ). A generating set K = {k1 Γc , . . . , kn Γc } for the centralizer of gΓc in G/Γc can be computed in TC0 (resp. TC0 (ExtGCD)) by induction. Let J = hk1 , . . . , kn , am0 , . . . , am i , where {am0 , . . . , am } = A ∩ Γc . Then J is the preimage of hKi under the homomorphism G → G/Γc . Define f : J → G by f (u) = [g, u]. Since u ∈ J, u commutes with g modulo Γc ; hence, [g, u] ∈ Γc and so Im(f ) ⊆ Γc . Moreover, f is a homomorphism: we have f (g, u1 u2 ) = [g, u1 u2 ] = [g, u2 ][g, u1 ][[g, u1 ], u2 ], and [g, u1 ] ∈ Γc ; therefore, [[g, u1 ], u2 ] ∈ Γc+1 = 1, and [g, u1 ] and [g, u2 ] commute, both being elements of the abelian group Γc . If h commutes with g, then hΓc ∈ hKi, i. e., h ∈ J. Thus, the centralizer of g is precisely the kernel of f : J → Γc . A generating set can be computed in TC0 (resp. TC0 (ExtGCD)) using Theorem 17. J 7.3 The conjugacy problem Now, we can combine the previous theorems to solve the conjugacy problem in TC0 following [16, Thm. 4.6]. I Theorem 19 (Conjugacy Problem). Let c, r ∈ N be fixed. The following is in TC0 for unary inputs and in TC0 (ExtGCD) for binary inputs: On input of some G ∈ Nc,r given as quotient presentation and elements g, h ∈ G, either 19 20 A. Myasnikov, A. Weiß produce some u ∈ G such that g = hu , or determine that no such element u exists. In case of unary inputs, u will be returned as a word, for binary inputs, as a word with binary exponents. Proof. Again we proceed by induction on c. If c = 1, then G is abelian and g is conjugate to h if and only if g = h. If so, we return u = 1. Now let us assume c > 1 and that the theorem holds for any nilpotent group of class c − 1 – in particular, for G/Γc . We use the notation as in the proof of Theorem 18. The first step of the circuit is to check conjugacy of gΓc and hΓc in G/Γc which can be done in TC0 by induction. If these elements are not conjugate, then g and h are not conjugate and the overall answer is ‘No’. Otherwise, we obtain some vΓc ∈ G/Γc such that gΓc = hv Γc . Let ϕ : G → G/Γc be the canonical homomorphism, J = ϕ−1 (C(gΓc )) (where C(gΓc ) denotes the centralizer of gΓc ), and define f : J → Γc by f (x) = [g, x]. As in the proof of Theorem 18, the image of f is indeed in Γc and f is a homomorphism. We claim that g and h are conjugate if and only if g −1 hv ∈ f (J). Indeed, if there exists w ∈ G such that g = hvw , then 1 · Γc = g −1 w−1 hv w · Γc = [g, w] · Γc , hence w ∈ J, so w−1 ∈ J as well. Then g −1 hv = [g, w−1 ] ∈ f (J), as required. The converse is immediate. So it suffices to express, if possible, g −1 hv as [g, w] with w ∈ J, in which case the conjugator is u = vw−1 . Now, the circuit computes a generating set {w1 Γc , . . . , wn Γc } for C(gΓc ) using Theorem 18. Then J is generated by {w1 , . . . , wn , am0 , . . . , am }, where again {am0 , . . . , am } = A∩Γc . After that, Coord(g −1 hv ) is computed and Theorem 14 used to determine whether g −1 hv ∈ f (J). If so, Theorem 17 is applied to find some w ∈ G such that g −1 hv = f (w). Finally, u = vw−1 is returned in case all previous tests succeed. Since we only concatenate a fixed constant number of TC0 (resp. TC0 (ExtGCD)) computations, the whole computation is in TC0 (resp. TC0 (ExtGCD)) again. J I Remark. We want to outline briefly how in the unary case the bounds of [16, Thm. 4.6] can be used to directly solve the conjugacy problem of nilpotent groups in TC0 . Since [16, Thm. 4.6] is for a non-uniform setting, we fix a nilpotent group G with generating set A. Let g, h be words over A±1 as inputs for the conjugacy problem with of total length n. By [16, Thm. 4.6], the length of conjugators is polynomial in n. By using binary exponents, the conjugators can be written with respect to a Mal’cev basis of G using only C log n bits for some constant C which only depends on G (this is a well-known fact – see e. g. [16, Thm. 2.3]). In particular, for all possible conjugators u which have bit-length at most C log n, it can be checked in parallel by a uniform family of TC0 circuits whether g = hu in G by using the circuits for the word problem [24] (note that for this purpose each u can be written down in unary since it is of length at most nC ). 8 Computing quotient presentations The results in the previous sections always required that the group is given as a quotient presentation. However, we can use Theorem 13 to transform an arbitrary presentation with at most r generators of a group in Nc,r into a quotient presentation. TC0 circuits for algorithmic problems in nilpotent groups I Proposition 20. Let c and r be fixed integers. The following is in TC0 : given an arbitrary finite presentation with generators a1 , . . . , ar of a group G ∈ Nc,r (as a list of relators given ±1 as words over {a1 , . . . , ar } ), compute a quotient presentation of G (encoded in unary) and an explicit isomorphism. Moreover, if the relators are given as words with binary exponents, then the binary encoded quotient presentation can be computed in TC0 (ExtGCD). Proof. Let A = {a1 , . . . , ar } and let R be the set of relators, i. e., G is presented as G = h A \| R i. Let F = Fc,r = ha1 , . . . , ar i be the free nilpotent group of class c on generators A. Let B = {b1 , . . . , bm } be the standard Mal’cev basis of F such that bi = ai for i = 1, . . . , r and let S denote the set of relations such that h B \| S i is a consistent nilpotent presentation for F . Consider the natural surjection ϕ : F → G and let N = ker(ϕ), which is the normal closure of R in F . Denoting R = {r1 , . . . , rk }, N is generated by iterated commutators [. . . [[ri , x1 ], x2 ], . . . , xj ], where i = 1, . . . , k, j ≤ c, and x1 , . . . , xj ∈ A ∪ A−1 . The total length of these generators is linear since c and r are constant. Using Theorem 13 in the group F , we can produce the full-form sequence T for N in TC0 (resp. in TC0 (ExtGCD) for binary inputs). Now G ' hB \| S ∪ T i and by Lemma 5 this is a (consistent) quotient presentation. J I Remark. Because of Proposition 20, in all theorems above where the input is a quotient presentation, we can also take an arbitrary r-generated presentation of a group in Nc,r as input. However, be aware that for the word problem (Theorem 6 and Corollary 7) the complexity changes from TC0 to TC0 (ExtGCD) in the binary case. 9 Power problem and conjugacy in wreath products of nilpotent groups In [19], the conjugacy problem in iterated wreath products of abelian is shown to be in TC0 (for a definition of iterated wreath products we refer to [19]). The crucial step there is the transfer result that the conjugacy problem in a wreath product A o B is TC0 -Turing-reducible to the conjugacy problems of A and B and the so-called power problem of B. The power problem of G is defined as follows: on input of g, h ∈ G (as words over the generators) decide whether h is a power of g that is whether there is some k ∈ Z such that g k = h in G. In the “yes” case compute this k in binary representation. If g has finite order in G, the computed k has to be the smallest non-negative such k. By [19], also the power problem of A o B is TC0 -Turing-reducible to the power problems of A and B given that torsion elements of B have uniformly bounded order. The latter condition is also preserved by wreath products. Thus, in the light of [19], it remains to show that the power problem of nilpotent groups is in TC0 and that the order of torsion elements is uniformly bounded, in order to establish the following theorem (note that [19] is only for fixed groups; therefore, we formulate also the following results in a non-uniform setting): I Theorem 21. Let A and B be finitely generated nilpotent groups and let d ≥ 1, then the conjugacy problem of the d-fold iterated wreath products A od B as well as A d o B is in TC0 . Proof. The following two lemmas together with a repeated application of Theorem 3, Lemma 5, and Theorem 5 of [19]. J I Lemma 22. Every finitely generated nilpotent group has a uniform bound on the order of torsion elements. 21 22 A. Myasnikov, A. Weiß Proof. We proceed by induction along a Mal’cev basis (a1 , . . . , am ) of G. If a1 has infinite order, we are done by induction. Otherwise, let k be the order of a1 and M be such that g M = 1 for all torsion elements g ∈ ha2 , . . . , am i. Consider a torsion element h ∈ ha1 , . . . , am i. Then hk ∈ ha2 , . . . , am i. Thus, hkM = 1. Therefore, kM is an upper bound on the order of torsion elements in G. J I Lemma 23. For every finitely generated nilpotent group G, the power problem of G is in uniform TC0 . Proof. We show a slightly more general statement by induction along a Mal’cev basis (a1 , . . . , am ) of G: for every fixed arithmetic progression α + βZ, the power problem restricted to α + βZ is in TC0 , i. e., given g, h ∈ G it can be decided in TC0 whether there is some n ∈ α + βZ with g n = h in G and, if so, that n can be computed in TC0 . We consider the input words g and h in the quotient G/ {a2 = · · · = am = 1}. Let g = ak1 and h = a`1 in this quotient. If k = ` = 0, it remains to solve the power problem in the subgroup ha2 , . . . , am i, which can be done by induction. Next, we distinguish the two cases that a1 has infinite order and that it has finite order (in G/ {a2 = · · · = am = 1}). In the case of infinite order, the only possible value for n can be computed as `/k (in TC0 by Theorem 2). If this is not an integer or not contained in the arithmetic progression (i. e., `/k 6≡ α mod β), then h is not a power of g. Otherwise, one simply checks whether g `/k = h in G (i. e., solving the word problem). As ` is bounded by the input length by Lemma 3, this can be done in TC0 by Theorem 6. In the case of finite order, let d denote the order of a1 . It can be checked for all 0 ≤ i < d in parallel whether ki = ` mod d. In case that there is such an i, the answer to the power problem is the same as the answer to the power problem in the subgroup ha2 , . . . , am i restricted to the arithmetic progression i + dZ ∩ α + βZ (the intersection can be hard-wired since there are only finitely many possibilities for a fixed group since the modulo is bounded by the least common multiple of the orders of finite order elements of the Mal’cev basis) – if there is no such i, the answer is “no”. J 10 Conclusion and Open Problem We have seen that most problems which in [16] were shown to be in LOGSPACE indeed are in TC0 even in the uniform setting where the number of generators and nilpotency class is fixed. Moreover, their binary versions are in TC0 (ExtGCD) meaning that nilpotent groups are no more complicated than abelian groups in many algorithmic aspects. This contrasts with the slightly larger class of polycyclic groups: while the word problem is still in TC0 [24, 12], the conjugacy problem is not even known to be in NP. We conclude with some possible generalizations of our results: I Question 24. Does a uniform version of Theorem 6 hold (i. e., is the uniform word problem still in TC0 ) for fixed nilpotency class but an arbitrary number of generators? What happens to the complexity if also the nilpotency class is part of the input? Note that in that case it is even not clear whether the word problem is still in polynomial time. I Question 25. Is there a way to solve the conjugacy problem for nilpotent groups with binary exponents in TC0 ? Notice that we needed to compute greatest common divisors for solving the subgroup membership problem. However, there might be a way of solving the conjugacy problem using another method. I Question 26. What is the complexity of the uniform conjugacy problem when the nilpotency class is not fixed? TC0 circuits for algorithmic problems in nilpotent groups On the way for proving that the subgroup membership problem of nilpotent groups is in TC0 , we established that the extended gcd problem with unary inputs and outputs is in TC0 . However, the computed solution is not as small as the one computed by the LOGSPACE algorithm from [17]: I Question 27. Is the following problem in TC0 : given unary encoded numbers a1 , . . . , an ∈ Z, compute x1 , . . . , xn ∈ Z with \|xi \| ≤ 12 max {\|a1 \| , . . . , \|an \|} such that x1 a1 + · · · + xn an = gcd(a1 , . . . , an )? References 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 N. Blackburn. Conjugacy in nilpotent groups. Proceedings of the American Mathematical Society, 16(1):143–148, 1965. W. W. Boone. The Word Problem. Ann. of Math., 70(2):207–265, 1959. M. Dehn. Über unendliche diskontinuierliche Gruppen. Math. Ann., 71(1):116–144, 1911. B. Eick and D. Kahrobaei. Polycyclic groups: A new platform for cryptology? ArXiv Mathematics e-prints, 2004. M. Elberfeld, A. Jakoby, and T. Tantau. Algorithmic meta theorems for circuit classes of constant and logarithmic depth. Electronic Colloquium on Computational Complexity (ECCC), 18:128, 2011. A. Garreta, A. Miasnikov, and D. Ovchinnikov. Properties of random nilpotent groups. ArXiv e-prints, Dec. 2016. P. Hall. The Edmonton notes on nilpotent groups. Queen Mary College Mathematics Notes. Mathematics Department, Queen Mary College, London, 1969. W. Hesse. Division is in uniform TC0 . In F. Orejas, P. G. Spirakis, and J. van Leeuwen, editors, ICALP, volume 2076 of Lecture Notes in Computer Science, pages 104–114. Springer, 2001. W. Hesse, E. Allender, and D. A. M. Barrington. Uniform constant-depth threshold circuits for division and iterated multiplication. J. Comput. Syst. Sci., 65:695–716, 2002. M. I. Kargapolov and J. I. Merzljakov. Fundamentals of the theory of groups, volume 62 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1979. Translated from the second Russian edition by Robert G. Burns. M. I. Kargapolov, V. N. Remeslennikov, N. S. Romanovskii, V. A. Roman’kov, and V. A. Čurkin. Algorithmic questions for σ-powered groups. Algebra i Logika, 8:643–659, 1969. D. König and M. Lohrey. Evaluating matrix circuits. In Computing and combinatorics, volume 9198 of Lecture Notes in Comput. Sci., pages 235–248. Springer, Cham, 2015. K. Lange and P. McKenzie. On the complexity of free monoid morphisms. In K. Chwa and O. H. Ibarra, editors, Algorithms and Computation, 9th International Symposium, ISAAC ’98, Taejon, Korea, December 14-16, 1998, Proceedings, volume 1533 of Lecture Notes in Computer Science, pages 247–256. Springer, 1998. C. R. Leedham-Green and L. H. Soicher. Symbolic collection using Deep Thought. LMS J. Comput. Math., 1:9–24 (electronic), 1998. R. J. Lipton and Y. Zalcstein. Word problems solvable in logspace. J. ACM, 24(3):522–526, July 1977. J. MacDonald, A. G. Myasnikov, A. Nikolaev, and S. Vassileva. Logspace and compressedword computations in nilpotent groups. CoRR, abs/1503.03888, 2015. B. S. Majewski and G. Havas. The complexity of greatest common divisor computations. In Algorithmic number theory (Ithaca, NY, 1994), volume 877 of Lecture Notes in Comput. Sci., pages 184–193. Springer, Berlin, 1994. A. Mal’cev. On homomorphisms onto finite groups. Transl., Ser. 2, Am. Math. Soc., 119:67–79, 1983. Translation from Uch. Zap. Ivanov. Gos. Pedagog Inst. 18, 49-60 (1958). 23 24 A. Myasnikov, A. Weiß 19 20 21 22 23 24 25 26 A. Miasnikov, S. Vassileva, and A. Weiß. The conjugacy problem in free solvable groups and wreath product of abelian groups is in TC0 . In P. Weil, editor, Computer Science - Theory and Applications - 12th International Computer Science Symposium in Russia, CSR 2017, Kazan, Russia, June 8-12, 2017, Proceedings, volume 10304 of Lecture Notes in Computer Science, pages 217–231. Springer, 2017. A. Mostowski. Computational algorithms for deciding some problems for nilpotent groups. Fundamenta Mathematicae, 59(2):137–152, 1966. A. Myasnikov, A. Nikolaev, and A. Ushakov. The Post correspondence problem in groups. J. Group Theory, 17(6):991–1008, 2014. A. Myasnikov, A. Nikolaev, and A. Ushakov. Non-commutative lattice problems. J. Group Theory, 19(3):455–475, 2016. P. S. Novikov. On the algorithmic unsolvability of the word problem in group theory. Trudy Mat. Inst. Steklov, pages 1–143, 1955. In Russian. D. Robinson. Parallel Algorithms for Group Word Problems. PhD thesis, University of California, San Diego, 1993. C. C. Sims. Computation with finitely presented groups, volume 48 of Encyclopedia of Mathematics and its Applications. Cambridge University Press, Cambridge, 1994. H. Vollmer. Introduction to Circuit Complexity. Springer, Berlin, 1999.	4math.GR
arXiv:1802.01730v1 [cs.MA] 5 Feb 2018 Local Wealth Redistribution Promotes Cooperation in Multiagent Systems Flávio L. Pinheiro Fernando P. Santos Collective Learning Group, The MIT Media Lab Massachusetts Institute of Technology, 22 Ames Street, Cambridge, Massachusetts flaviopp@mit.edu INESC-ID and Instituto Superior Técnico, Universidade de Lisboa 2744-016 Porto Salvo, Portugal fernando.pedro@tecnico.ulisboa.pt ABSTRACT Designing mechanisms that leverage cooperation between agents has been a long-lasting goal in Multiagent Systems. The task is especially challenging when agents are selfish, lack common goals and face social dilemmas, i.e., situations in which individual interest conflicts with social welfare. Past works explored mechanisms that explain cooperation in biological and social systems, providing important clues for the aim of designing cooperative artificial societies. In particular, several works show that cooperation is able to emerge when specific network structures underlie agents’ interactions. Notwithstanding, social dilemmas in which defection is highly tempting still pose challenges concerning the effective sustainability of cooperation. Here we propose a new redistribution mechanism that can be applied in structured populations of agents. Importantly, we show that, when implemented locally (i.e., agents share a fraction of their wealth surplus with their nearest neighbors), redistribution excels in promoting cooperation under regimes where, before, only defection prevailed. CCS CONCEPTS • Computing methodologies → Multi-agent systems; Cooperation and coordination; KEYWORDS Emergent behaviour; Social networks; Social simulation; Simulation of complex systems; Cooperation 1 INTRODUCTION Explaining cooperation among selfish and unrelated individuals has been a central topic in evolutionary biology and social sciences [23]. Simultaneously, the challenge of designing cooperative Multiagent Systems (MAS) has been a long standing goal of researchers in artificial intelligence (AI) [10, 19]. More than thirty years ago it was already clear that ”Intelligent agents will inevitably need to interact flexibly with other entities. The existence of conflicting goals will Proc. of the 17th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2018), M. Dastani, G. Sukthankar, E. Andre, S. Koenig (eds.), July 2018, Stockholm, Sweden © 2018 International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. This is the author’s version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in Proceedings of Proc. of the 17th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2018), M. Dastani, G. Sukthankar, E. Andre, S. Koenig (eds.), July 2018, https://doi.org/doi. need to be handled by these automated agents, just as it is routinely handled by humans.“ [10]. In Cooperative multiagent interactions, agents need to collaborate towards common goals, which introduces challenges associated with coordination, communication and teamwork modeling [19, 28]. Self-interested interactions, in contrast, require the design of indirect incentive schemes that motivate selfish agents to cooperate in a sustainable way [7, 19]. Cooperation is often framed as an altruistic act that requires an agent to pay a cost (c) in order to generate a benefit (b) to another. Refusing to incur in such a cost is associated with an act of defection and results in no benefits generated. Whenever the benefit exceeds the cost (b > c) and plays occur simultaneously, agents face the Prisoner’s Dilemma, a decisionmaking challenge that embodies a fundamental social dilemma within MAS [21]: rational agents pursuing their self-interests are expected to defect, while the optimal collective outcome requires cooperation. If defection is the likely decision of rational agents, however, how can we justify the ubiquity of cooperation in the real world? Evolutionary biology has pursued this fundamental question by searching for additional evolutionary mechanisms that might help to explain the emergence of cooperative behavior [22, 23]. Some of these mechanisms allowed to develop solutions that found applications in computer science, such as informing about ways of incentivizing cooperation in p2p networks [9, 11], wireless sensor networks [2], robotics [47] or resource allocation and distributed work systems [43] – to name a few. Network reciprocity is one of the most popular mechanisms to explain the evolution of cooperation in social and biological systems [24, 26, 30–32, 38]. In this context, populations are structured and interactions among agents are constrained. These constraints are often modelled by means of a complex network of interactions. Applications of this mechanism have been explored in the design of MAS that reach high levels of cooperation [1, 17, 29, 34]. Despite these advances, cooperation on structured populations is still hard to achieve when considering social dilemmas with high levels of temptation to defect. Additional complementary mechanisms are required. Here we consider that agents contribute a percentage of their surplus (defined below), which is later divided among a Beneficiary Set of other agents. In this context, we aim at answering the following questions: • Does redistribution of wealth promote the evolution of cooperation? • How should Beneficiary Sets be selected? • What are the potential disadvantages of such a mechanism? AAMAS’18, July 2018, Stockholm, Sweden Using methods from Evolutionary Game Theory (EGT) [44] and resorting to computer simulations, we explore how wealth redistribution impacts the evolution of cooperation on a population of agents without memory (i.e. unable to recall past interactions) and rationally bounded (i.e. lacking full information on payoff structure of the game they are engaging). We assume that agents resort to social learning through peer imitation, which proves to be a predominant adaptation scheme employed by humans [36]. Also, we consider that strategies are binary – Cooperate and Defect – opting to focus our attention on the complexity provided by 1) heterogeneous populations, 2) the redistribution mechanism and 3) the self-organizing process of agents when adapting over time. The role of larger strategy spaces (such as in [29, 34, 41]) lies outside the scope of the present work. With redistribution, we show that cooperation emerges in a parameter region where previously it was absent. Moreover, we show that the optimal choice of redistributing groups consists in picking the nearest neighbors (local redistribution). This result fits with a local and polycentric view of incentive mechanisms [27, 46] in MAS, which may not only be easier to implement but, as we show, establish an optimal scale of interaction in terms of eliciting cooperation. 2 RELATED WORK The problem of Cooperation is a broad and intrinsically multidisciplinary topic, which has been part of the MAS research agenda for a long time [10, 19]. In the realm of evolutionary biology, several mechanisms were proposed to explain the evolution of cooperation [22]. Kin selection [13], direct reciprocity [45], indirect reciprocity [25, 42] and network reciprocity [26, 38] constitute some of the most important mechanisms proposed. Remarkably, these mechanisms have been applied in AI in order to design MAS in which cooperation emerges. For example, Waibel et al. associated kin selection with evolutionary robotics [47]; Griffiths employed indirect reciprocity to promote cooperation in p2p networks while Ho et al. investigated the social norms that, through a system of reputations and indirect reciprocity, promote cooperation in crowdsourcing markets [12, 16]. Similarly, Peleteiro et al. combined indirect reciprocity with complex networks to design a MAS where, again, cooperation is able to emerge [29]. On top of that, Han applied EGT – as performed in our study – in order to investigate the role of punishment and commitments in multiagent cooperation, both in pairwise [14] and group interactions [15]. Regarding alternative agent-oriented approaches to sustain cooperation in MAS, we shall underline the role of electronic institutions [4, 8] whereby agents’ actions are explicitly constrained so that desirable collective behaviors can be engineered. The role of population structure and network reciprocity is, in this context, a prolific area of research. In [31] it was shown that complex networks are able to fundamentally change the dilemma at stake, depending on the particular topology considered [18, 31]; Ranjbar-Sahraei et al. applied tools from control theory in order to study the role of complex networks on the evolution of cooperation [34]. Importantly, the role of dynamic networks – i.e., agents are able to rewire their links – was also shown to significantly improve the levels of cooperation, especially in networks with a high average Flávio L. Pinheiro and Fernando P. Santos degree of connectivity [32, 39]. A survey on the topic of complex networks and the emergence of cooperation in MAS can be accessed in [17]. Previous works found that cooperation in structured population substantially decreases when the temptation to defect increases (see Model for a proper definition of Temptation). Thereby, here we contribute with an additional mechanism of cooperation on structured populations. We consider a mechanism of redistribution, inspired in the wealth redistribution mechanisms that prevail in modern economic/political systems, mainly through taxation. We are particularly interested in understanding how to sample redistribution groups in an effective way. In this context, we shall underline the works of Salazar et al. and Burguillo-Rial, in which a system of taxes and coalitions was shown to promote cooperation on complex networks [37] and regular grids [5]. While [37] and [5] do an excellent job showing how coalitions – leaded by a single agent – emerge, here we consider a simpler/decentralized model (e.g. no leaders are considered and taxes are redistributed rather than centralized in a single entity) and focus our analysis on showing that local redistribution sets are optimal. Our approach does not require additional means of reciprocity, memory, leadership, punishment or knowledge about features of the network. We cover a wide range of dilemma strengths and explicitly show when the local redistribution promotes cooperation by itself. Notwithstanding, the analysis performed in [37] and [5] surely provides important insights to address in future works, on how to explicitly model the adherence to beneficiary sets and guarantee their stability. Also, while here we assume an egalitarian redistribution over each individual in the Beneficiary Set, we shall note that different redistribution heuristics may imply different levels of allocation fairness [33]. In this context, a recent work introduces the concept of Distributed Distributed Justice [20] and shows that local interactions may provide a reliable basis to build trust and reputation between agents, which can be used to regulate, in a decentralized way, the levels of justice in agents’ actions. This way, it is rewarding to note that local interactions not only constitute an optimal scale to form cooperative Beneficiary Sets (as we show, see below), but also provide the convenient interaction environment to allow justice in contributions to be sustained. 3 MODEL 3.1 Three Stage Redistribution Game Here we propose a sequential game dynamics made of three stages. Focusing on an arbitrary agent i, these stages can be described as follows: (1) Agent i participates in a one-shot game (here a Prisoner’s Dilemma) with all his/her neighbors j. From each interaction j, he/she obtains a payoff πi, j . After all interactions, agent i Í accumulates a total payoff Πi = j πi, j ; (2) Next, agent i contributes a fraction α of his/her payoff surplus (Πi − θ ) to be redistributed. The group that benefits from agent i contribution is called Beneficiary Set i (Bi ). (3) Finally, agent i receives his/her share from each Beneficiary Set that he/she is part of. We refer to α as the level of taxation, as it defines the fraction of the surplus that agents contribute, while θ is the threshold level of Local Wealth Redistribution Promotes Cooperation in Multiagent Systems critical level of taxation, ⍺* 1.0 0.8 𝜃=0 .99 AAMAS’18, July 2018, Stockholm, Sweden a) .90 𝜃=0 𝜃=0 0.6 .75 .50 𝜃=0 5 .2 𝜃=0 0 𝜃 = 0.0 0.4 0.2 b) homogeneous network c) 1.1 1.2 1.3 1.4 1.5 1.6 1.7 Temptation parameter, T 1.8 1.9 2.0 Figure 1: Solutions for the two-person game with wealth redistribution. Each curve indicates the critical taxation levels (α ∗ ) above which the nature of the social dilemma changes, for different payoff thresholds (θ ) and as a function of the Temptation parameter (T ). payoff that defines the surplus. By definition, agents with negative payoff cannot contribute (i.e., θ > 0); they might, however, receive benefits from the Beneficiary Sets. Each agent i contributes only to one Beneficiary Set Bi from which they cannot be part of, that is, agents do not receive from the Beneficiary Set they contribute to. A central question of this work is how to select Bi for each i. As we show, this decision has a profound and non-trivial impact on the overall cooperation levels in the system. 3.2 single peak over hzi fraction of nodes 1.0 fraction of nodes 0.0 Heterogeneous Network d) degree (z) D(z) ⇡ z degree (z) Figure 2: Graphical depiction of the specific structures used in this work. a) Homogeneous Networks correspond to a structure in which all nodes have the same degree. b) Heterogeneous Networks are characterized by a high variance among the degree of nodes. The color of each node indicates its degree: blue tones represent lower degree and red tones higher degree. Panel c) and d) show, respectively, the degree distributions of the Homogeneous and Heterogeneous networks under analysis. In particular, we use scale-free networks as representatives of heterogeneous structures; these have a degree distribution that decays as a power law. The Prisoner’s Dilemma Game In general, all the possible outcomes of a two-strategy two-player game, in which two agents engage in a one-shot interaction that requires them to decide – independently and simultaneously – whether they wish to Cooperate (C) or to Defect (D), can be summarized in a payoff matrix, such as C D C R S D T P which reads as the payoff obtained by playing the row strategy when facing an opponent with the column strategy. Here, R represents the Reward payoff for mutual cooperation and P the Punishment for mutual defection. When one of the individuals Defects and the other Cooperates, the first receives the Temptation payoff (T ) while the second obtains the Sucker’s payoff (S). In this manuscript we consider that agents interact according to the Prisoner’s Dilemma (PD). Agents are said to face a PD whenever the relationship between the payoffs is such that T > R > P > S [44]. In such a scenario, rational agents seeking to optimize their self-returns are expected to always Defect. However, since the best aggregated outcome would have both players cooperating (2R > 2P), agents are said to face a social dilemma: optimizing self-returns clashes with optimizing the social outcome. In this sense, mutual cooperation is Pareto Optimal and contributes to increase both average payoff (over mutual defection) and egalitarian social welfare (over unilateral cooperation) [6]. It is noteworthy to mention that other situations – with different optimal rational responses – arise when the parameters take a different relationship [21]: the Stag Hunt game when R > T > P > S; the Snowdrift Game when T > R > S > P; the Harmony Game when R > T > S > P; or the Deadlock Game when T > P > R > S, to name a few. Notwithstanding, the PD is by far the most popular metaphor of social dilemmas [44] and the one that presents the biggest challenge for cooperation to emerge. For these reasons, PD shall be the main focus of study in this manuscript. We further simplify the parameter space by considering that R = 1, P = 0, S = 1 − T and 1 < T ≤ 2 with the game being fully determined by the Temptation value (T). In that sense, higher temptation creates more stringent conditions for the emergence of cooperation. 3.3 Prisoner’s Dilemma with Wealth Redistribution As an introductory example, let us start by analyzing the particular case of two interacting agents (i and j) in a one-shot event. In this case, the Beneficiary Sets of each agent (Bi and B j ) are composed only by the opponent. Wealth/payoff redistribution can thus be analyzed by considering a slightly modified payoff matrix, that takes into account the second and third stages. The resulting payoff matrix becomes C D C 1 1 − T + α(T − θ ) D T − α(T − θ ) 0 AAMAS’18, July 2018, Stockholm, Sweden Flávio L. Pinheiro and Fernando P. Santos (2) Hence, depending on the choice of θ and for a given T , α ∗ is the minimum level of taxation required to observe a change in the nature of the game faced by agents. It is straightforward to notice that the nature of the game changes from a Prisoner’s Dilemma to an Harmony Game as the relationship moves from T > R > P > S to R > T > S > P. Figure 1 shows α ∗ for different values of T and θ . Clearly, in well-mixed populations and under the simple scenario of a MAS composed by two agents, the redistribution mechanism has the simple effect of reshaping the payoff matrix, trivially changing the nature of the dilemma. Such a trivial conclusion cannot be drawn with large populations playing on networks, where we will show that different ways of assigning the Beneficiary Sets have a profound impact on the ensuing levels of cooperation. 3.4 1.00 Structured Populations Let us consider a population of Z agents in which agents correspond to the nodes/vertices of a complex network, while links dictate who interacts with whom. The structure reflects the existence of constraints that limit interactions between agents. These constraints can arise from spatial or communication limitations. The number of interactions that each agent i participates in defines his/her degree zi . The distribution of degrees, D(z), describes the fraction of agents that has degree z. In this work we consider two structures: Homogeneous Random Graphs [40, 41] and ScaleFree Barabási Networks [3]. Homogeneous Random Graphs are generated by successively randomizing the ends of pairs of links from an initially regular graph (e.g. Lattice or Ring). The resulting structure has a random interaction structure but all nodes in the network have the same degree. Figure 2a) depicts graphically an example of such structures and Figure 2c) the corresponding Degree distribution. Scale-free networks are generated by an algorithm of growth and preferential attachment [3]. This algorithm is as follows: 1) start from three fully connected nodes; 2) add, sequentially, each of the Z − m remaining nodes; 3) each time a new node is added, it connects to m pre-existing nodes, selecting preferentially nodes level of cooperation 0.8 0.7 0.6 0.5 0.4 0.3 𝜃 = 1.00 0.2 0.1 0.0 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 0.75 0.50 0.25 0.00 1.8 1.9 2.0 Temptation Parameter, T b) Heterogeneous 1.0 0.9 1.00 0.8 level of cooperation T −1 α > (T − θ ) 1.0 (1) which results in the critical values of α, ∗ Homogeneous 0.9 level of taxation, ⍺ T − α(T − θ ) < 1 a) level of taxation, ⍺ where θ is the payoff threshold and α is the level of taxation. The rationale to arrive at this payoff structure is the following: whenever both players choose to act the same way the payoff remains the same as their contributions (from taxes) and benefits (from receiving the contributions of their opponent) cancel out. A Defector playing against a Cooperator sees his payoff of T subtracted by an amount α(T − θ ) while not receiving any benefit, since the Cooperator has negative payoff and does not contribute. Likewise, the Cooperator is exempt from contributing but receives an additional contribution of α(T − θ ), which represents the amount taxed to the Defector. To inspect whether wealth redistribution changes the nature of the social dilemma (i.e. from a Prisoner’s Dilemma to another type of game) we have to inspect whether there is a difference in the relationship between the payoffs R and T or P and S. This sums up to solving a single inequality, 0.7 0.6 0.5 0.4 0.3 𝜃 = 1.00 0.2 0.1 0.0 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 0.75 0.50 0.25 0.00 1.8 1.9 2.0 Temptation Parameter, T Figure 3: Level of Cooperation on Homogeneous Random Networks (a) and Heterogeneous (Scale-free) Networks (b). Each plot shows the level of cooperation under a different combination of taxation level, α, and Temptation, T . In all cases the fitness threshold is fixed at θ = R = 1.0. Blue indicates regions where Cooperation dominates, Red delimits regions dominated by Defectors. Top bars above each panel indicate the level of cooperation in the absence of wealth redistribution, as a function of the Temptation payoff parameter. The level of cooperation is computed by estimating the expected fraction of cooperators when the population reaches a stationary state. To that end we run 104 independent simulations that start with 50% cooperators and 50% defectors. Population size of Z = 103 and intensity of selection β = 1.0. with higher degree. Here we have used m = 3 The resulting network is characterized by a heterogeneous degree distribution (one which decays as a power law), in which the majority of the nodes have few connections while a few have many. Figure 2b) shows a graphical example of such structure and Figure 2d) the degree distribution. In the following we explore the case of networks with Z = 103 Í nodes and average degree of ⟨z⟩ = z zD(z) = 4. During the Local Wealth Redistribution Promotes Cooperation in Multiagent Systems Heterogeneous a) AAMAS’18, July 2018, Stockholm, Sweden [𝜃 = 0.80, ⍺ = 0.50] 1.0 1.0 0.8 0.8 level of cooperation Level of Cooperation Heterogeneous a) [⍺=0.5] 0.6 0.4 No Wealth Redistribution 0.2 Nearest Neighbors (d = 1) 0.6 Random Group 0.4 0.2 0.0 0.0 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 1.0 2.0 1.1 1.2 1.3 Homogeneous b) 1.4 1.5 1.6 1.7 1.8 1.9 2.0 1.9 2.0 Temptation parameter, T Temptation Parameter, T Homogeneous [𝜃 = 1.00, ⍺ = 0.90] b) [⍺=0.8] 1.0 1.0 0.6 0.4 Nearest Neighbors (d = 1) 0.8 𝜃 = 0.0 𝜃 = 0.4 𝜃 = 0.8 𝜃 = 1.2 𝜃 = 1.6 𝜃 = 2.0 level of cooperation Level of Cooperation Legend 0.8 0.6 0.4 Random Group 0.2 0.2 0.0 0.0 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 1.0 2.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 Temptation parameter, T Temptation Parameter, T Figure 4: Level of cooperation on Heterogeneous (a) and Homogeneous (b) populations for different values of the payoff threshold (θ ) as a function of the Temptation payoff parameter (T ). Gray Dashed line shows the results obtained in the absence of a wealth redistribution scheme. Population size of Z = 103 and intensity of selection β = 1.0. Figure 5: Comparison between the effects of assigning the nearest neighbors of an agent i to the corresponding Beneficiary Set Bi (dark blue line) and when agents are assigned at random to Bi (light blue), on the level of cooperation in the domain of the Temptation payoff parameter, T. Panel a) shows the results on Heterogeneous populations and panel b) the impact on Homogeneous populations. Population size of Z = 103 and intensity of selection β = 1.0. simulations we make use of 20 independently generated networks of each type. 3.5 Games on Networks We study the expected level of cooperation attained by the population. We estimate this quantity through computer simulations. The level of cooperation corresponds to the expected fraction of cooperators in a population that evolved after 2.5 × 106 iterations. We estimate this quantity by averaging the observed fraction of cooperators at the final of each simulation, over 104 independent simulations. Each simulation starts from a population with an equal composition of Cooperators and Defectors, which are randomly placed along the nodes of the network. In between each update round, each agent i plays once with all his/her zi nearest neighbors (i.e., agents they are directly connected with). The accumulated payoff over all interactions an agent i participates can be computed as D C Π i = nC i T − σi (1 − T )(ni + ni ) (3) where niD (nC i ) is the number of neighbors of i that Defect (Cooperate) and σi is equal to 1 if i is a Cooperator and 0 otherwise. From the accumulated payoff, agents contribute to a pool a fraction α of the surplus Π − θ . The fitness fi of an agent i results from subtracting from his/her accumulated payoff his/her contributions plus the share he/she obtains from each of the Beneficiary Sets j he/she participates in. We shall underline that, while T is the same for all agents (that is, the dilemma is the same for everyone in the population), heterogeneous populations introduce an additional complexity layer by implying that different agents may vary in the maximum values of accumulated payoff that they are able to earn. This can be formalized as fi = (1 − α)(Πi − θ ) + Z Õ j δi, j α(Π j − θ ) \|B j \| (4) where δi, j is equal to one if i is part of the Beneficiary Set towards which j contributes and zero otherwise, while \|B j \| denotes the size of set B j . Evolution in the frequency of strategies adopted in the population happens through a process of imitation or social learning. At each iteration a random agent, say i, compares his fitness with the AAMAS’18, July 2018, Stockholm, Sweden Flávio L. Pinheiro and Fernando P. Santos fitness of a neighbor, say j. Depending on the fitness difference, i adopts the strategy of j with probability Heterogeneous [𝜃 = 0.8, ⍺ = 0.50] a) p= 1 1 + Exp(−β(f j − fi )) level of cooperation 1.0 (5) The meaning of this sigmoid function can be understood as follows: if j is performing much better than i, then i updates his/her strategy, adopting the strategy of j. Conversely, if j is performing much worse, i does not update the strategy. The parameter β, often called the intensity of selection and akin to a learning rate, dictates how sharp is the transition between these two regimes, as f j − fi approaches zero. Large β means that individuals act in a more deterministic way, updating strategies at the minimum difference; small β means that individuals are prone to make imitation mistakes. Nearest Neighbors (d = 1) d=2 0.6 d=4 0.4 Legend d=1 d=2 d=3 d=4 0.2 0.0 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 Temptation parameter, T Homogeneous [𝜃 = 1.0, ⍺ = 0.90] b) level of cooperation 1.0 4 RESULTS 4.1 Wealth Redistribution and the Level of Cooperation in Structured Populations In this section we start by analyzing the scenario in which the Beneficiary Set of each agent i corresponds to his/her nearest neighbors. Hence, the size of the Beneficiary set of i is \|Bi \| = zi . These are also the agents from whom he/she interacts with and obtains a payoff from. Figure 3 shows the achieved levels of cooperation when the payoff threshold is set to θ = R = 1.0, as a function of the Temptation payoff (T ) and the level of taxation (α). Figure 3a shows the results on Homogeneous networks, and Figure 3b on Heterogeneous. We find that, for a fixed payoff threshold (θ ), increasing the level of taxation results in an increase in the levels of cooperation. This effect diminishes with an increase in the Temptation (T ). That is, when increasing T the minimum value of α necessary to promote cooperation increases as well. The same behavior is observed in both structures. However, there is a larger degree of cooperation on Heterogeneous networks, where there is always a level of taxation for a given Temptation that guarantees a 100% level of cooperation. Hence, in order for cooperation to be evolutionary viable on homogeneous networks, more stringent conditions are necessary, e.g. higher tax levels. Figure 4 shows how the level of cooperation depends on variations of the fitness threshold (0 ≤ θ ≤ 2.0, in intervals of 0.4) while keeping a fixed level of taxation (α = 0.5) under different levels of the Temptation payoff (T ). Figure 4a shows the results obtained for Heterogeneous networks and panel b) the results on Homogeneous structures. For a constant level of taxation, α, decreasing the payoff threshold, θ , increases the range of Temptation, T , under which cooperation can possibly evolve. This is the case in both types of structures. However, once again, the effect is more limited in homogeneous populations. Both Figure 3 and 4 highlight the positive impact of a local wealth redistribution mechanism in the enhancement of cooperation. It also puts in evidence that the success of such mechanism depends on the volume of payoff that is redistributed. Ultimately, this can be done by either increasing the level of taxation, α or decreasing the payoff threshold, θ , that defines the taxable payoff. 0.8 0.8 Nearest Neighbors (d = 1) d=2 0.6 Legend d=3 0.4 d=1 d=2 d=3 d=4 0.2 0.0 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 Temptation parameter, T Figure 6: Panel a) compares how extending beneficiary sets, from the nearest neighbors (d = 1) to nodes at a distance up to d = 4 links away, impacts the level of cooperation on Heterogeneous networks. Panel b) shows how extended beneficiary sets impact the level of cooperation on Homogeneous networks. In both cases extending the set of beneficiaries has a negative a negative impact in the levels of cooperation. Population size of Z = 103 and intensity of selection β = 1.0. 4.2 Randomized Beneficiary Set Next we explore to which extent the results obtained depend on the way agents are being assigned to each Beneficiary Set. To that end, we compare two cases: i) nearest set assignment – the Beneficiary Set of each agent corresponds to her/his nearest neighbors, as above; and ii) random set assignment – agents are assigned at random to each Beneficiary Set. The number of agents assigned to each set is equal to the degree of the contributing agent, in both cases, which guarantees that the collected payoffs from each agent are distributed among the same number of individuals in both i) and ii). Figure 5a and b show the results obtained, respectively, on Heterogeneous and Homogeneous networks. We consider θ = 0.5, α = 0.9 and explore the domain 1.0 ≤ T ≤ 2. Dark blue curves show the results obtained under the nearest set assignment and light blue curves the results obtained under a random set assignment. The results show that the ability of a wealth redistribution mechanism lies in the redistribution of the taxed payoff among the agents that are spatially related. A random assignment of agents Local Wealth Redistribution Promotes Cooperation in Multiagent Systems a) 1.0 Homogeneous 0.9 number of generations level of taxation, α 0.8 0.7 0.6 0.5 0.4 0.3 0.2 =1.0 0.1 1000 316 100 32 10 0.0 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 Temptation Parameter, T b) 1.0 Heterogeneous 0.9 level of taxation, α 0.7 0.6 0.5 0.4 0.3 0.2 = 1.0 0.1 number of generations 316 0.8 100 32 10 0.0 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 Temptation Parameter, T Figure 7: Panel a) shows the fixation times (in generations) on homogeneous networks. Panel b) shows the fixation times (in generations) in heterogeneous networks. A generation corresponds to Z iteration steps and the fixation times indicate the expected time that the population takes to arrive to a state dominated by Cooperators or Defectors when starting from a state with equal abundance of both strategies. Population size of Z = 103 and intensity of selection β = 1.0. drastically decreases the levels of cooperation obtained in both networks. But to which extent do the Beneficiary Sets need to be constrained spatially? 4.3 Extended Beneficiary Set To answer the previous question, we explore the case in which all nodes (up to a distance of d links) are assigned to the Beneficiary Set of a focal agent i; when d = 1 the previous results are thereby obtained. Figure 6a and b show the results up to d = 4 on Heterogeneous and Homogeneous networks respectively. In both cases, we see that an expansion in the size of the Beneficiary Set leads to a decrease in the levels of cooperation. This result further reinforces the AAMAS’18, July 2018, Stockholm, Sweden conclusion that wealth redistribution is only efficient when agents return, in form of taxes, a share of the accumulated payoffs to the agents they have engaged with. We shall underline that here both distance and size of Bi play a role on the obtained results, while in the previous section the size of Bi was kept constant for each i across the different treatments, thus disambiguating the effect of Bi size and distance on the resulting cooperation levels. 4.4 What is the cost of wealth redistribution? Figure 7a and b shows the fixation times of populations when θ = 1.0 along the domain bounded by 0.0 ≤ α ≤ 1.0 and 1.0 ≤ T ≤ 2.0. The fixation times correspond to the expected number of generations (i.e., sets of Z potential imitation steps) for the population to reach a state in which only one strategy is present in the population. These plots map directly into Figure 3a and b, allowing to compare the relative fixation times of regions with high/low levels of cooperation. We observe that the evolution of cooperation is associated with an increase in the fixation times. This increase can be in some situations an order of magnitude higher. The regions that exhibit larger fixation times lie in the critical boundary that divides areas of defectors and cooperators dominance (Figure 3). Hence, promoting cooperation by redistributing wealth also requires a longer waiting time for the population to reach a state of full cooperation. However, setting higher taxation values than the bare minimum necessary for the emergence of cooperation allows populations to reach fixation quicker. 4.5 Multiple Contribution Brackets In the real world, taxes are unlikely to be defined by a single threshold (θ ) that separates agents who contribute from those that do not. In reality taxes are progressive, in the sense that taxation levels (α) increase with increasing level of income (in this case accumulated payoff). In this section we implement a similar approach and inspect the impact of increasing the number of taxation brackets. Let us consider that, instead of a single threshold we now have B taxation brackets divided by B − 1 threshold levels. For each bracket we define αb as the effective tax and θb as the bottom threshold of bracket b, where b ∈ {0, 1, 2, ..., B − 1, B}. By definition B = 0 corresponds to the case in which no taxes are collected, and the redistribution of wealth is absent. Moreover, B = 1, implies the existence of a single bracket were all individuals would contribute, a case that we do not explore in this manuscript. B = 2, corresponds to the case in which there are two brackets, which is the scenario that we have explored until now. We consider the case in which taxation increases linearly with increasing brackets. Let us define θb = bθ /B. Individuals in bracket b have their payoff surplus taxed by αb = (b − 1)α/B when their accumulated payoff falls into θb < Π ≤ θb+1 for b < B − 1. For b = B the tax level is αb = α and affects all individuals with Π > θ . As an example, for B = 3 each bracket would be characterized by the following tax levels b b b b = 0) = 1) = 2) = 3) αb αb αb αb = 0 for all individuals with Π ≤ θ /3; = α/3 for all individuals with θ /3 < Π ≤ 2θ /3; = 2α/3 for all individuals with 2θ /3 < Π ≤ θ ; = α for all individuals with Π > θ . AAMAS’18, July 2018, Stockholm, Sweden Flávio L. Pinheiro and Fernando P. Santos 4.0 5 3.5 To sum up, we show that wealth redistribution embodies an effective mechanism that significantly helps cooperation to evolve. It works by fundamentally changing the nature of the dilemma at stake: by appropriately choosing the level of taxation (α) and payoff threshold (θ ) it is possible to shift from a Defector dominance to a Cooperator dominance dynamics. Moreover, we find that in Heterogeneous populations allow us to ease the redistribution mechanism – that is, imposing lower taxation rates and/or lower taxable surplus values when compared with Homogeneously structured populations. Additionally, we show, for the first time, that different assignments of Beneficiary Sets significantly impact the ensuing levels of cooperation. Local Beneficiary Sets, where agents receive the contributions from their direct neighbors, constitute a judicious choice when compared with Beneficiary Sets that are formed by 1) agents randomly picked from the population or 2) by including agents at higher distances. Naturally, a Local wealth redistribution scheme may not only prove optimal in terms of achieved cooperation levels, but also reveal much simpler to implement, by exempting the need of central redistribution entities and by minimizing the number of peers that agents need to interact with. We shall highlight, however, that promoting cooperation through a wealth redistribution mechanism bears longer fixation times, in terms of the number of iterations required to achieve overall cooperation. Here we assume that the redistribution mechanism is externally imposed. Agents are not able to opt out from the taxation scheme. Given that this mechanism increases the overall cooperation and average payoff in the system, an argument for its acceptance - by rational agents - can be formulated based on the infamous veil of ignorance proposed by John Rawls [35]: Agents should decide the kind of society they would like to live in without knowing their social position. Agents would, this way, prefer a cooperative society where redistribution exists, provided that here average payoff is maximized. Notwithstanding, future research shall analyze the role of more complex strategies that give opportunity of agents to voluntarily engage (or not) in the proposed redistribution scheme. Alongside, effective mechanisms that discourage the second order free riding problem (i.e., free riding by not contributing to the redistribution pot, while expecting others to do so) shall be examined. Future works shall also evaluate whether alternative taxation schemes are prone to be more efficient than the one proposed here. In all these cases, an evolutionary game theoretic framework – such as the one here developed – constitutes a promising toolkit to employ. �� re as in g 2.0 Varf / Varp in eq ua lit y �� 2.5 1.5 de c payoff threshold, 𝜃 3.0 �� 1.0 0.5 2.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 level of taxation, ⍺ 0.8 0.9 1.0 Figure 8: Relative wealth inequality after the redistribution step, in a heterogeneous population dominated by cooperators and for different combinations of taxation level (α) and threshold (θ ). We quantify the relative wealth inequality after the redistribution step as the ratio between the variance of the fitness distribution (V ar f , i.e. variance in gains across the population after redistribution) and the variance of the accumulated payoff distribution (V arp , i.e. variance in gains before redistribution). Population size of Z = 103 and intensity of selection β = 1.0. In this way we use θ and α as the upper level bound and only parameters in this condition. We find that variations in the number of taxation brackets (B=3,4,5) have only a marginal impact in the overall levels of cooperation observed when compared with the scenarios studied so far (B=2). 4.6 Wealth Inequality Finally, we discuss the effect of wealth redistribution on fitness inequality. First, it is important to highlight that the observed levels of inequality depend, by default, on the distribution of strategies and network degree. In homogeneous structures, if every agent adopts the same strategy – either Defectors or Cooperators – everyone obtains the same fitness. In heterogeneous structures, a Cooperation dominance scenario bounds the feasible equality levels, given the degree distribution of the population. In fact, some agents engage in more interactions than others and Beneficiary Sets have different sizes, depending on the particular connectivity of agents. We shall focus on this scenario. We compare the variance of fitness (i.e. gains after the redistribution step) and the variance of accumulated payoff (i.e. gains before the redistribution step) in order to quantify the relative inequality after we apply the proposed redistribution mechanism. In particular, we use the ratio between the variance of fitness and the variance of accumulated payoff as a metric of resulting wealth inequality. Figure 8 shows how higher levels of θ and α reduce the resulting inequality. In fact, while increasing payoff threshold limits taxation to the richer agents, increasing level of taxation increases the flow of fitness from rich agents to their Beneficiary Sets. In the most strict case – high θ and α – the variance of the fitness distribution is reduced to as low as 7% of the accumulated payoff distribution. 6 CONCLUSION ACKNOWLEDGMENTS The authors acknowledge the useful discussions with Francisco C. Santos, Jorge M. Pacheco and Aamena Alshamsi. F.L.P. is thankful to the Media Lab Consortium for financial support. F.P.S. acknowledges the financial support of Fundação para a Ciência e Tecnologia (FCT) through PhD scholarship SFRH/BD/94736/2013, multi-annual funding of INESC-ID (UID/CEC/50021/2013) and grants PTDC/EEISII/5081/2014, PTDC/MAT/STA/3358/2014. Local Wealth Redistribution Promotes Cooperation in Multiagent Systems REFERENCES [1] Stéphane Airiau, Sandip Sen, and Daniel Villatoro. 2014. Emergence of conventions through social learning. Autonomous Agents and Multi-Agent Systems 28, 5 (2014), 779–804. [2] Ian F Akyildiz, Weilian Su, Yogesh Sankarasubramaniam, and Erdal Cayirci. 2002. Wireless sensor networks: a survey. Computer Networks 38, 4 (2002), 393–422. [3] Réka Albert and Albert-László Barabási. 2002. Statistical mechanics of complex networks. Reviews of Modern Physics 74, 1 (2002), 47. [4] Josep Ll Arcos, Marc Esteva, Pablo Noriega, Juan A Rodríguez-Aguilar, and Carles Sierra. 2005. Engineering open environments with electronic institutions. Engineering Applications of Artificial Intelligence 18, 2 (2005), 191–204. [5] Juan C Burguillo-Rial. 2009. A memetic framework for describing and simulating spatial prisoner’s dilemma with coalition formation. In Proceedings of AAAI’09. AAAI Press, 441–448. [6] Ulle Endriss and Nicolas Maudet. 2003. Welfare engineering in multiagent systems. In International Workshop on Engineering Societies in the Agents World. Springer, 93–106. [7] Eithan Ephrati and Jeffrey S Rosenschein. 1996. Deriving consensus in multiagent systems. Artificial Intelligence 87, 1-2 (1996), 21–74. [8] Marc Esteva, Bruno Rosell, Juan A Rodriguez-Aguilar, and Josep Ll Arcos. 2004. AMELI: An agent-based middleware for electronic institutions. In Proceedings of the Third International Joint Conference on Autonomous Agents and Multiagent Systems. IEEE Computer Society, 236–243. [9] Michal Feldman and John Chuang. 2005. Overcoming free-riding behavior in peer-to-peer systems. ACM SIGecom Exchanges 5, 4 (2005), 41–50. [10] Michael R Genesereth, Matthew L Ginsberg, and Jeffrey S Rosenschein. 1986. Cooperation without communication. In Proceedings of AAAI’86. AAAI Press. [11] Philippe Golle, Kevin Leyton-Brown, Ilya Mironov, and Mark Lillibridge. 2001. Incentives for sharing in peer-to-peer networks. In Electronic Commerce. Springer, 75–87. [12] Nathan Griffiths. 2008. Tags and image scoring for robust cooperation. In Proceedings of the 2008 International Conference on Autonomous Agents and Multi-agent Systems. International Foundation for Autonomous Agents and Multiagent Systems, 575–582. [13] William D Hamilton. 1964. The genetical evolution of social behaviour. Journal of Theoretical Biology 7, 1 (1964), 17–52. [14] TA Han. 2016. Emergence of Social Punishment and Cooperation through Prior Commitments. In Proceedings of AAAI’16. AAAI Press, 2494–2500. [15] TA Han, Luís Moniz Pereira, Luis A Martinez-Vaquero, and Tom Lenaerts. 2017. Centralized vs. Personalized Commitments and their influence on Cooperation in Group Interactions. In Proceedings of AAAI’17. AAAI Press. [16] Chien-Ju Ho, Yu Zhang, Jennifer Vaughan, and Mihaela Van Der Schaar. 2012. Towards social norm design for crowdsourcing markets. In AAAI’12 Technical Report WS-12-08. AAAI Press. [17] Lisa-Maria Hofmann, Nilanjan Chakraborty, and Katia Sycara. 2011. The evolution of cooperation in self-interested agent societies: a critical study. In Proceedings of the 2011 International Conference on Autonomous Agents and Multi-agent Systems. International Foundation for Autonomous Agents and Multiagent Systems, 685–692. [18] Genki Ichinose, Yoshiki Satotani, and Hiroki Sayama. 2017. How mutation alters fitness of cooperation in networked evolutionary games. arXiv preprint arXiv:1706.03013 (2017). [19] Nicholas R Jennings, Katia Sycara, and Michael Wooldridge. 1998. A roadmap of agent research and development. Autonomous Agents and Multi-agent Systems 1, 1 (1998), 7–38. [20] David Burth Kurka and Jeremy Pitt. 2016. Distributed distributive justice. In Self-Adaptive and Self-Organizing Systems (SASO), 2016 IEEE 10th International Conference on. IEEE, 80–89. [21] Michael W Macy and Andreas Flache. 2002. Learning dynamics in social dilemmas. Proceedings of the National Academy of Sciences 99 (2002), 7229–7236. [22] Martin A Nowak. 2006. Five rules for the evolution of cooperation. Science 314, 5805 (2006), 1560–1563. [23] Martin A Nowak. 2012. Evolving cooperation. Journal of Theoretical Biology 299 (2012), 1–8. [24] Martin A Nowak and Robert M May. 1992. Evolutionary games and spatial chaos. Nature 359, 6398 (1992), 826–829. [25] Martin A Nowak and Karl Sigmund. 2005. Evolution of indirect reciprocity. Nature (2005). [26] Hisashi Ohtsuki, Christoph Hauert, Erez Lieberman, and Martin A Nowak. 2006. A simple rule for the evolution of cooperation on graphs. Nature 441, 7092 (2006), 502. [27] Elinor Ostrom. 2015. Governing the commons. Cambridge University Press. [28] Liviu Panait and Sean Luke. 2005. Cooperative multi-agent learning: The state of the art. Autonomous Agents and Multi-agent Systems 11, 3 (2005), 387–434. [29] Ana Peleteiro, Juan C Burguillo, and Siang Yew Chong. 2014. Exploring indirect reciprocity in complex networks using coalitions and rewiring. In Proceedings of the 2014 International Conference on Autonomous Agents and Multi-agent Systems. [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] AAMAS’18, July 2018, Stockholm, Sweden International Foundation for Autonomous Agents and Multiagent Systems, 669– 676. Flávio L. Pinheiro and Dominik Hartmann. 2017. Intermediate Levels of Network Heterogeneity Provide the Best Evolutionary Outcomes. Scientific Reports 7, 1 (2017), 15242. Flavio L Pinheiro, Jorge M Pacheco, and Francisco C Santos. 2012. From local to global dilemmas in social networks. PloS ONE 7, 2 (2012), e32114. Flávio L Pinheiro, Francisco C Santos, and Jorge M Pacheco. 2016. Linking individual and collective behavior in adaptive social networks. Physical Review Letters 116, 12 (2016), 128702. Jeremy Pitt, Julia Schaumeier, Didac Busquets, and Sam Macbeth. 2012. Selforganising common-pool resource allocation and canons of distributive justice. In Self-Adaptive and Self-Organizing Systems (SASO), 2012 IEEE Sixth International Conference on. IEEE, 119–128. Bijan Ranjbar-Sahraei, Haitham Bou Ammar, Daan Bloembergen, Karl Tuyls, and Gerhard Weiss. 2014. Theory of cooperation in complex social networks. In Proceedings of AAAI’14. AAAI Press. John Rawls. 2009. A theory of justice. Harvard University Press. Luke Rendell, Robert Boyd, Daniel Cownden, Marquist Enquist, Kimmo Eriksson, Marc W Feldman, Laurel Fogarty, Stefano Ghirlanda, Timothy Lillicrap, and Kevin N Laland. 2010. Why copy others? Insights from the social learning strategies tournament. Science 328, 5975 (2010), 208–213. Norman Salazar, Juan A Rodriguez-Aguilar, Josep Ll Arcos, Ana Peleteiro, and Juan C Burguillo-Rial. 2011. Emerging cooperation on complex networks. In Proceedings of the 2011 International Conference on Autonomous Agents and Multiagent Systems. International Foundation for Autonomous Agents and Multiagent Systems, 669–676. Francisco C Santos and Jorge M Pacheco. 2005. Scale-free networks provide a unifying framework for the emergence of cooperation. Physical Review Letters 95, 9 (2005), 098104. Francisco C Santos, Jorge M Pacheco, and Tom Lenaerts. 2006. Cooperation prevails when individuals adjust their social ties. PLoS Computational Biology 2, 10 (2006), e140. Francisco C Santos, JF Rodrigues, and Jorge M Pacheco. 2005. Epidemic spreading and cooperation dynamics on homogeneous small-world networks. Physical Review E 72, 5 (2005), 056128. Fernando P Santos, Jorge M Pacheco, Ana Paiva, and Francisco C Santos. 2017. Structural power and the evolution of collective fairness in social networks. PloS ONE 12, 4 (2017), e0175687. Fernando P. Santos, Jorge M. Pacheco, and Francisco C. Santos. 2018. Social norms of cooperation with costly reputation building. In AAAI’18. AAAI Press. Sven Seuken, Jie Tang, and David C Parkes. 2010. Accounting Mechanisms for Distributed Work Systems. In AAAI’10. AAAI Press. Karl Sigmund. 2010. The calculus of selfishness. Princeton University Press. Robert L Trivers. 1971. The evolution of reciprocal altruism. The Quarterly Review of Biology 46, 1 (1971), 35–57. Vítor V Vasconcelos, Francisco C Santos, and Jorge M Pacheco. 2015. Cooperation dynamics of polycentric climate governance. Mathematical Models and Methods in Applied Sciences 25, 13 (2015), 2503–2517. Markus Waibel, Dario Floreano, and Laurent Keller. 2011. A quantitative test of Hamilton’s rule for the evolution of altruism. PLoS Biology 9, 5 (2011), e1000615.	2cs.AI
I/O-Efficient Generation of Massive Graphs Following the LFR Benchmark∗ Michael Hamann† , Ulrich Meyer‡ , Manuel Penschuck‡ , Hung Tran‡ , and Dorothea Wagner† † Institute of Theoretical Informatics, Karlsruhe Institute of Technology, Am Fasanengarten 5, 76131 Karlsruhe, Germany arXiv:1604.08738v3 [cs.DS] 14 Jun 2017 {michael.hamann, dorothea.wagner}@kit.edu ‡ Institute for Computer Science, Goethe-University Frankfurt, Robert-Mayer-Straße 11-15, 60325 Frankfurt am Main, Germany {umeyer, htran, mpenschuck}@ae.cs.uni-frankfurt.de June 15, 2017 Abstract LFR is a popular benchmark graph generator used to evaluate community detection algorithms. We present EM-LFR, the first external memory algorithm able to generate massive complex networks following the LFR benchmark. Its most expensive component is the generation of random graphs with prescribed degree sequences which can be divided into two steps: the graphs are first materialized deterministically using the Havel-Hakimi algorithm, and then randomized. Our main contributions are EM-HH and EM-ES, two I/Oefficient external memory algorithms for these two steps. We also propose EM-CM/ES, an alternative sampling scheme using the Configuration Model and rewiring steps to obtain a random simple graph. In an experimental evaluation we demonstrate their performance; our implementation is able to handle graphs with more than 37 billion edges on a single machine, is competitive with a massive parallel distributed algorithm, and is faster than a state-of-theart internal memory implementation even on instances fitting in main memory. EM-LFR’s implementation is capable of generating large graph instances orders of magnitude faster than the original implementation. We give evidence that both implementations yield graphs with matching properties by applying clustering algorithms to generated instances. Similarly, we analyse the evolution of graph properties as EM-ES is executed on networks obtained with EM-CM/ES and find that the alternative approach can accelerate the sampling process. ∗ This work was partially supported by the DFG under grants ME 2088/3-2, WA 654/22-2. Parts of this paper were published as [21]. 1 1 Introduction Complex networks, such as web graphs or social networks, are usually composed of communities, also called clusters, that are internally dense but externally sparsely connected. Finding these clusters, which can be disjoint or overlapping, is a common task in network analysis. A large number of algorithms trying to find meaningful clusters have been proposed (see [16, 22, 17] for an overview). Commonly synthetic benchmarks are used to evaluate and compare these clustering algorithms, since for most real-world networks it is unknown which communities they contain and which of them are actually detectable through structure [4, 17]. In the last years, the LFR benchmark [28, 27] has become a standard benchmark for such experimental studies, both for disjoint and for overlapping communities [14]. With the emergence of massive networks that cannot be handled in the main memory of a single computer, new clustering schemes have been proposed for advanced models of computation [9, 49]. Since such algorithms typically use hierarchical input representations, quality results of small benchmarks may not be generalizable to larger instances. To produce such large instances exceeding main memory, we propose a generator in the external memory (EM) model of computation that follows the LFR benchmark. The distributed CKB benchmark [11] is a step in a similar direction, however, it considers only overlapping clusters and uses a different model of communities. In contrast, our approach is a direct realization of the established LFR benchmark and supports both disjoint and overlapping clusters. 1.1 Random Graphs from a prescribed Degree Sequence In preliminary experiments, we identified the generation of random graphs with prescribed degree sequence as the main issue when transferring the LFR benchmark into an EM setting — both in terms of algorithmic complexity and runtime. To do so, the LFR benchmark uses the fixed degree sequence model (FDSM), also known as edge-switching Markov-chain algorithm (e.g. [35]). It consists of a) generating a deterministic graph from a prescribed degree sequence and b) randomizing this graph using random edge switches. For each edge switch, two edges are chosen uniformly at random and two of the endpoints are swapped if the resulting graph is still simple (for details, see section 5). Each edge switch can be seen as a transition in a Markov chain. This Markov chain is irreducible [13], symmetric and aperiodic [18] and therefore converges to the uniform distribution. It also has been shown to converge in polynomial time if the maximum degree is not too large compared to the number of edges [19]. However, the analytical bound of the mixing time is impractically high even for comparably small graphs as it contains the sum of all degrees to the power of nine. Experimental results on the occurrence of certain motifs in networks [35] suggest that 100m steps should be more than enough where m is the number of edges. Further results for random connected graphs [18] suggest that the average and maximum path length and link load converge between 2m and 8m swaps. More recently, further theoretical arguments and experiments showed that 10m to 30m steps are enough [39]. A faster way to realize a given degree sequence is the Configuration Model. The problem here is that multi-edges and loops may be generated. In the Erased Configuration Model these illegal edges are deleted. However, doing so alters the graph properties since skewed degree distributions, necessary for the LFR benchmark, are not properly realized [43]. In this context the question arises whether edge switches starting from the Configuration Model can be used to uniformly sample simple graphs at random. 2 1.2 Our Contribution Our main contributions are the first external memory versions of the LFR benchmark and the FDSM. After defining our notation, we introduce the LFR benchmark in more detail and then focus on the FDSM. We describe the realization of the two steps of the classic FDSM, namely a) generating a deterministic graph from a prescribed degree sequence [EM-HH, section 4] and b) randomizing this graph using random edge switches [EM-ES, section 5]. These steps form a pipeline moving data from one algorithm to the next. In section 6, we describe the alternative approach EM-CM/ES generating uniform random simple graphs using the Configuration Model and edge rewiring. Sections 7, 8 and 9 describe algorithms for the remaining steps of the external memory LFR benchmark, EM-LFR. We conclude with an experimental evaluation of our algorithms and demonstrate that our EM version of the FDSM is faster than an existing internal memory implementation, scales well to large instances, and can compete with a distributed parallel algorithm. Further, we compare EM-LFR to the original LFR implementation and show that EM-LFR is significantly faster while producing equivalent networks in terms of community detection algorithm performance and graph properties. We also investigate the mixing time of EM-ES and EM-CM/ES, give evidence that our alternative sampling scheme quickly yields uniform samples and that the number of swaps suggested by the original LFR implementation can be kept for EM-LFR. 2 Preliminaries and Notation Define [k] := {1, . . . , k} for k ∈ N>0 . A graph G = (V, E) has n = \|V \| sequentially numbered nodes V = {v1 , . . . , vn } and m = \|E\| edges. Unless stated differently, graphs are assumed to be undirected and unweighted. It is called simple if it contains neither multi-edges nor self-loops. To obtain a unique representation of an undirected edge {u, v} ∈ E, we write [u, v] ∈ E where u ≤ v; in contrast to a directed edge, the ordering shall be used algorithmically but does not carry any meaning for the application. D = (d1 , . . . , dn ) is a degree sequence of graph G iff ∀vi {∈}V : deg(vi ) = di . We denote an integer powerlaw distribution with exponent −γ ∈ R for γ ≥ 1 and values between the limits a, b ∈ N>0 with a < b as Pld ([a, b), γ). Let X be an integer random variable drawn from Pld ([a, b), γ) then P[X=k] ∝ k −γ (proportional to) if a ≤ k < b and P[X=k] = 0 otherwise. A statement depending on some number x > 0 is said to hold with high probability if it is satisfied with probability at least 1 − 1/xc for some constant c ≥ 1. Also refer to Table 2 (Appendix) which contains a summary of commonly used definitions. 2.1 External-Memory Model We use the commonly accepted external memory model by Aggarwal and Vitter [1]. It features a two-level memory hierarchy with fast internal memory (IM) which may hold up to M data items, and a slow disk of unbounded size. The input and output of an algorithm are stored in EM while computation is only possible on values in IM. The measure of an algorithm’s performance is the number of I/Os required. Each I/O transfers a block of B consecutive items between memory levels. Reading or writing n contiguous items from or to disk requires scan(n) = Θ(n/B) I/Os. Sorting n consecutive items triggers sort(n) = Θ((n/B) · logM/B (n/B)) I/Os. For all realistic values of n, B and M , scan(n) < sort(n) n. Sorting complexity constitutes a lower bound for most intuitively non-trivial EM tasks [1, 34]. 3 Degrees, community sizes and memberships Sample intra- and inter-community edges Intra-community edge Inter-community edge Remove (rewire) illegal edges Community Figure 1: Left: Sample node degrees and community sizes from two powerlaw distributions. The mixing parameter µ determines the fraction of the inter-community edges. Then, assign each node to sufficiently large communities. Center: Sample intra-community graphs and inter-community edges. Right: Lastly, remove illegal inter-community edges respective to the global graph. 2.2 Time Forward Processing Let A be an algorithm performing discrete events over time (e.g., iterations of a loop) that produce values which are reused by following events. The data dependencies of A can be modeled using a directed acyclic graph G=(V, E) where every node v ∈ V corresponds to an event [31]. The edge (u, v) ∈ E indicates that the value produced by u will be required by v. When computing a solution, the algorithm traverses G in some topological order. For simplicity, we assume G to be already ordered, i.e. ∀ (u, v) ∈ E : u < v. Then, the Time Forward Processing (TFP) technique uses a minimum priority queue (PQ) to provide the means to transport data as implied by G: iterate over the events in increasing order and receive for each u the messages sent to it by claiming and removing all items with priority u from the PQ. Inductively, these messages have minimal priority amongst all items stored in the PQ. The event then computes its result xu and sends it to every successor v by inserting xu into the PQ with priority v. Using a suited EM PQ [3, 42], TFP incurs O(sort(k)) I/Os, where k is the number of messages sent. 3 The LFR Benchmark The LFR benchmark [28] describes a generator for random graphs featuring a planted community structure, a powerlaw degree distribution, and a powerlaw community size distribution. A revised version [27] also introduces weighted and directed graphs with overlapping communities. We consider the most commonly used versions with unweighted, undirected graphs and possibly overlapping communities. All its parameters are listed in Table 1 and are fully supported by EM-LFR. The revised generator [27] changes the original algorithm [28] even for the initial scenario of unweighted, undirected graphs and non-overlapping communities. Here, we describe the more recent approach which is also used in the author’s implementation: initially, the degrees D=(d1 , . . . , dn ), the number (ν1 , . . . , νn ) of each node, and community sizes PC of memberships Pn S = (s1 , . . . , sC ) with ξ=1 sξ = i=1 νi are randomly sampled according to the supplied parameters. Observe that the number of communities C follows endogenously. For our analysis, we assume that nodes are members in ν=O(1) communities which implies C=O(n).1 Depending on the mixing parameter 0 < µ < 1, every node vi ∈ V features dext = µdi interi community edges and din = (1−µ)·d edges within its communities. The algorithm assigns every i i node vi to either νi =1 or νi = ν communities at random such that the requested community sizes and number of communities per node are realized. Further, the desired internal degree din i has to be strictly smaller than the size sξ of its community ξ. In the case of overlapping communities, the internal degree is evenly split among all communities the node is part of. Both the computation of din i and the splitting into several communities use non-deterministic rounding to avoid biases. 1 If the maximal community size grows with Θ(n ) for > 0, the number of communities is governed by C=o(n). 4 Parameter n Pld ([dmin , dmax ), γ) 0≤O≤n, ν≥1 Pld ([smin , smax ), β) 0<µ<1 Definition Number of nodes to be produced Degree distribution of nodes, typically γ = 2 O random nodes belong to ν communities; remainder has one membership Size distribution of communities, typically β=1 Mixing parameter: fraction of neighbors of every node u that shall not share a community with u Table 1: Parameters of overlapping LFR. The typical values follow suggestions by [27]. As illustrated in Fig. 1, the LFR benchmark then generates the inter-community graph using ext FDSM on the degree sequence (dext 1 , . . . , dn ). In order not to violate the mixing parameter µ, rewiring steps are applied to the global inter-community graph to replace edges between two nodes sharing a community. Analogously, an intra-community network is sampled for each community. In the overlapping case, rewiring steps are used to remove edges that exist in multiple communities and would result in duplicate edges in the final graph. While for realistic parameters most intra-community graphs fit into main memory, we cannot assume the same for the global graph. For the global graph and large communities, an EM variant of the FDSM is applicable, which we implement using EM-HH and EM-ES described in sections 4 and 5. 4 EM-HH : Deterministic Edges from a Degree Sequence In this section, we introduce an EM-variant of the well-known Havel-Hakimi scheme that takes a positive non-decreasing degree sequence2 D=(d1 , . . . , dn ) and, if possible, outputs a graph GD which realizes these degrees. A sequence D is called graphical if a matching simple graph GD exists. Havel [23] and Hakimi [20] gave inductive characterizations of graphical sequences which directly lead to a graph generator: given D, connect the first node v1 with degree d1 (minimal among all nodes) to d1 -many high-degree vertices by emitting edges to nodes vn − (d1 −1) , . . . , vn . Then remove d1 from D and decrement the remaining degree of every new neighbor which yields an updated sequence D0 .3 Subsequently, remove zero-entries and sort D0 while keeping track of the original positions to be able to output the correct node indices. Finally, recurse until no positive entries remain. After every iteration, the size of D is reduced by at least one resulting in O(n) rounds. For an implementation, it is non-trivial to keep the sequence ordered after decrementing the neighbors’ degrees. Internal memory solutions typically employ priority queues optimized for integer keys, e.g., bucket-lists [44, 47]. This approach incurs Θ(sort(n + m)) I/Os using a naı̈ve EM PQ since every edge triggers an update to the pending degree of at least one endpoint. We propose the Havel-Hakimi variant EM-HH which emits a stream of edges in lexicographical order. It can be fed to any single-pass streaming algorithm without a round-trip to disk. Additionally, EM-HH may be used in time O(n) to test whether a degree sequence D is graphical or to drop problematic edges yielding a graphical sequence (cf. section 6). Thus, we consider only internal I/Os and emphasize that storing the output – if necessary by the application – requires O(m) time and O(scan(m)) I/Os where m is the number of edges produced. 2 Within our pipeline, we generate a monotonic degree sequence by first sampling a monotonic uniform sequence online based on the ideas of [6, 48]. Applying the inverse sampling technique (carrying over the monotonicity) yields the required distribution. Thus, no additional sorting steps are necessary for the inter-community graph. 3 This variant is due to [20]; in [23], the node of maximal degree is picked and connected. 5 4.1 Data structure Instead of maintaining the degree of every node in D individually, EM-HH compacts nodes with equal degrees into a group, yielding DD := {di : 1≤i≤n} groups. Since D is monotonic, such nodes have consecutive ids and the compaction can be performed in a streaming fashion4 . The sequence is then stored as a doubly linked list L = [gj ]1≤j≤DD where group gj = (bj , nj , δj ) assigns degree δj to nodes vbj , . . . , vbj +(nj −1) . The algorithm is built around the following invariants holding at the begin of every iteration: (I1) δj < δj+1 ∀1 ≤ j < DD , i.e. the groups represent strictly monotonic degrees (I2) bj + nj = bj+1 ∀1 ≤ j < DD , i.e. there are no gaps in the node ids These invariants allow us to bound the memory P Dfootprint in two steps: first observe that a list L of size DD describes a graph with at least D i=1 i/2 edges due to (I1). Thus, graphs from an arbitrary L filling the whole IM have Ω(M 2 ) edges. Even under pessimistic assumptions this amounts to an edge list of more than 1 PB of size on realistic machines.5 Therefore, even in the worst case the whole data structure can be kept in IM for all practical scenarios. On top of this, a probabilistic argument applies: While there exist graphs with DD =Θ(n) (cf. Fig. 2), Lemma 1 gives a sub-linear bound on DD if D is sampled from a powerlaw distribution (refer also to section 10.1 for experimental results). Lemma 1. Let D be a degree sequence with n nodes sampled from Pld ([1, n), γ). Then, the number of unique degrees DD = {di : 1 ≤ i ≤ n} is bounded by O(n1/γ ) with high probability. Proof. Consider random variables (X1 , . . . , Xn ) sampled i.i.d. from Pld ([1, n), γ) as an unordered degree sequence. Fix an index 1≤j≤n. Due to the powerlaw distribution, Xj is likely to have a small degree. Even if all degrees 1, . . . , n1/γ were realized, their occurrences would be covered by the claim. Thus, it suffices to bound the number of realized degrees larger than n1/γ . We first show that their total probability mass is small. Then we can argue that DD is asymptotically unaffected by their rare occurrences: Pn−1 Pn−1 R n−1 −γ n−1 X i−γ (i) i−γ (ii) dx i=n1/γ +1 i=n1/γ +1 1/γ n1/γRx P∞ −γ ≤ P[Xj >n ] = P[Xj =i] = Pn−1 = ∞ −γ −γ ζ(γ) − i=n i ζ(γ) − n x dx i=1 i i=n1/γ +1 1 1−γ − n1/γ /n n1/γ /n − (n − 1)1−γ 1−γ (n−1) = = = O(n1/γ /n), 1 (γ − 1)ζ(γ) − n1−γ ζ(γ) + 1−γ n1−γ P −γ is the Riemann zeta function which satisfies ζ(γ) ≥ 1 for all γ∈R, γ≥1. where (i) ζ(γ) = ∞ i=1 i R b+1 In (ii) we exploit the sum’s monotonicity to bound it between the two integrals a x−γ dx ≤ R Pb −γ ≤ b x−γ dx. i=a i a−1 In order to bound the number of occurrences, define Boolean indicator variables Yi with Yi =1 iff Xi >n1/γ . Observe that they modelPBernoulli trials Yi ∈B(p) with p=O(n1/γ /n). Thus, P the expected number of high degrees is E[ ni=1 Yi ] = ni=1 P[Xi >n1/γ ] = O(n1/γ ). Chernoff’s inequality gives an exponentially decreasing bound on the tail distribution of the sum which thus holds with high probability. Due to Lemma 1, a graph sampled from a powerlaw distribution with m = O(M 2γ ) can be computed in IM with high probability. 4 While direct sampling of the group’s multinomial distribution is not beneficial in LFR, it may be used to omit the compaction phase for other applications. 5 A single item of L can be represented by its three values and two pointers, i.e. a total of 5·8=40 bytes per item (assuming 64 bit integers and pointers). Just 2 GB of IM suffice for storing 5 × 107 items, which result in at least 6.25 × 1014 edges, i.e. storing just two bytes per edge would require more than one Petabyte. Further, standard tricks (e.g., exploiting the redundancy due to (I2)) can be used to reduce the memory footprint of L. 6 v1 v3 v5 1 2 3 v2k−1 ··· k ··· k k−1 k−2 v2k v2(k−1) v2(k−2) ··· 1 v2 Figure 2: Materialization of the degree sequence Dk = (1, 1, 2, 2, . . . , k, k) with DDk = k = Θ(n) which maximizes EM-HH’s memory consumption asymptotically. A node’s label corresponds to the vertex’ degree. 4.2 Algorithm EM-HH works in n rounds, where every iteration corresponds to a recursion step of the original formulation. Each time it extracts vertex vb1 with the smallest available id and with minimal degree δ1 . The extraction is achieved by incrementing the lowest node id (b01 ← b1 +1) of group g1 and decreasing its size (n01 ← n1 −1). If the group becomes empty (n01 = 0), it is removed from L at the end of the iteration. We now connect node vb1 to δ1 nodes from the end of L. Let gj be the group of smallest index to which vb1 connects to. Then there are two cases: (C1) If node vb1 connects to all nodes in gj , we directly emit the edges [u, x] \| n−δ1 < x ≤ n and decrement the degrees of all groups gj , . . . , g\|L\| accordingly. Since degree δj−1 remains unchanged, it may now match the decremented δj . This violation of (I1) is resolved by merging both groups. Due to (I2), the union of gj−1 and gj contains consecutive ids and it suffices to grow nj−1 ← nj−1 +nj and to delete group gj . (C2) If vb1 connects only to a number a < nj of nodes in group gj , we split gj into two groups gj0 and gj00 of sizes a and nj −a respectively. We then connect vertex u to all a nodes in the first fragment gj0 and hence need to decrease its degree. Thus, a merge analogous to (C1) may be required (see Fig. 3). Now, groups gj+1 , . . . , g\|L\| are consumed wholly as in (C1). P\|L\| If the requested degree δ1 cannot be met (i.e., δ1 > k=1 nk ), the input is not graphical [20]. Since the vast majority of nodes have low degrees, a sufficiently large random powerlaw degree sequence D contains at most very few nodes that cannot be materialized as requested. Therefore, we do not explicitly ensure that the sampled degree sequence is graphical and rather correct the negligible inconsistencies later on by ignoring the unsatisfiable requests. 4.3 Improving the I/O-complexity In the current formulation of EM-HH we perform constant work per edge which is already optimal. However, we introduce a simple optimization which improves constant factors and gives I/O-efficient accesses. It also allows EM-HH to test whether D is graphical in time O(n). Observe that only groups in the vicinity of gj can be split or merge; we call these the active frontier. In contrast, the so-called stable groups gj+1 , . . . , gDD keep their relative degree differences as the pending degrees of all their nodes are decremented by one in each iteration. Further, they will become neighbors to all subsequently extracted nodes until group gj+1 eventually becomes an active merge candidate. Thus, we do not have to update the stable degrees in every round, but rather maintain a single global iteration counter I and count how many iterations a group remained stable: when a group gk becomes stable in iteration I0 , we annotate it with I0 by adding δk ← δk +I0 . If gk has to be activated again in iteration I > I0 , its updated degree follows as δk ← δk −I. The degree δk remains positive since (I1) enforces a timely activation. 7 List [(bi , ni , δi )]i (1, 2, 1), (3, 2, 2), (5, 2, 3) (2, 1, 1), (3, 3, 2), (6, 1, 3) (3, 4, 2) (4, 2, 1), (6, 1, 2) (5, 2, 1) Uncompressed Degree Sequence D merge split 1 1 2 2 ⇔ 3 2 3 extract edge-to merge 1 2 2 2 ⇔ 3 2 extract edge-to split 2 2 1 2 1 2 extract edge-to edge-to merge 1 1 ⇔ 2 1 extract edge-to 1 1 0 extract edge-to Initial situation group gj group gi d−1 Splitting at gi (front) group gj group gi d−1 d−1 d Splitting at gj (back) group gi d−1 d−1 d Before gi can be split, the degrees of groups gj with j > i are decreased group gj d−2 d−1 d−1 Figure 3: Left: EM-HH on D=(1, 1, 2, 2, 3, 3). Values of L and D in row i correspond to the state at the beginning of the i-th iteration. Groups are visualized directly after extraction of the head vertex. The number next to an edge-to symbol indicates the new degree. After these updates, splitting and merging takes place. Right: Consider two adjacent groups gi , gj with degrees d−1 and d. A split of gi (left) or gj (right) directly triggers a merge, so the number of groups remains the same. Lemma 2. The optimized variant of EM-HH requires O(scan(DD )) I/Os if L is stored in an external memory list. Proof. An external-memory list requires O(scan(k)) I/Os to execute any sequence of k sequential read, insertion, and deletion requests to adjacent positions (i.e. if no seeking is necessary) [38]. We will argue that EM-HH scans L roughly twice, starting simultaneously from the front and back. Every iteration starts by extracting a node of minimal degree. Doing so corresponds to accessing and eventually deleting the list’s first element gi . If the list’s head block is cached, we only incur an I/O after deleting Θ(B) head groups, yielding O(scan(DD )) I/Os during the whole execution. The same is true for accesses to the back of the list: the minimal degree increases monotonically during the algorithm’s execution until the extracted node has to be connected to all remaining vertices. In a graphical sequence, this implies that only one group remains and we can ignore the simple base case asymptotically. Neglecting splitting and merging, the distance between the list’s head and the active frontier decreases monotonically triggering O(scan(DD )) I/Os. Merging. As described before, it may be necessary to reactivate stable groups, i.e. to reload the group behind the active frontier (towards L’s end). Thus, we not only keep the block F containing the frontier cached, but also block G behind it. It does not incur additional I/O, since we are scanning backwards through L and already read G before F . The reactivation of stable groups hence only incurs an I/O when the whole block G is consumed and deleted. Since this does not happen before Ω(B) merges take place, reactivations may trigger O(scan(DD )) I/Os in total. Splitting. Observe that L at most doubles in size as splitting a group with degree d, which has a neighbor of degree d±1, directly triggers another merge (cf. Fig. 3). Since a split replaces one group by two adjacent fragments which differ in their degree by exactly one, a second split to one of the fragments does not increase the size of the list. 5 EM-ES: I/O-efficient Edge Switching EM-ES is a central building block of our pipeline and is used to randomize and rewire existing graphs. It applies a sequence S=[σs ]1≤s≤k of edge swaps σs to a simple graph G=(V, E), where typically k = c\|E\| for a constant c ∈ [1, 100]. The graph is represented by a lexicographically ordered edge list EL =[ei ]1≤i≤m which only stores the pair (u, v) and omits (v, u) for every ordered edge [u, v] ∈ E. As illustrated in Fig. 4, a swap σ(ha, bi, d) is encoded by a direction bit d and the edge ids a and b (i.e. the position in EL ) of the edges supposed to be swapped. The switched edges are denoted as eσa and eσb and are given by (eσa , eσb ) := σ(ha, bi, d) as defined in Fig. 4. 8 1 a a d c input a b 1 2 c b 2 d σ(h1, 2i, false) or b 2 c 1 d σ(h1, 2i, true) ( ({α1 , β1 }, {α2 , β2 }) if d = false σ(ha, bi, d) = , ({α1 , β2 }, {α2 , β1 }) if d = true where [α1 , α2 ] = ea and [β1 , β2 ] = eb are the edges at positions a and b in the edge list EL . illegal (creates multi-edge) Figure 4: A swap consists of two edge ids and a direction flag. The edge ids describe an induced subgraph (left); the flag indicates how the incident nodes are shuffled. We assume that the swap’s constituents are drawn independently and uniformly at random. Thus, the sequence can contain illegal swaps that would introduce multi-edges or self-loops if executed. Such illegal swaps are simply skipped. In order to do so, the following tasks have to be addressed for each σ(ha, bi, d): (T1) (T2) (T3) (T4) Gather the nodes incident to edges ea and eb . Compute eσa and eσb and skip if a self-loop arises. Verify that the graph remains simple, i.e. skip if edge eσa or eσb already exist in EL . Update the graph representation EL . If the whole graph fits in IM, a hash set per node storing all neighbors can be used for adjacency queries and updates in expected constant time (e.g., VL-ES [47]). Then, (T3) and (T4) can be executed for each swap in expected time O(1). However, in the EM model this approach incurs Ω(1) I/Os per swap with high probability for a graph with m ≥ cM and any constant c > 1. To improve the situation, we split S into smaller runs of r = Θ(\|E\|) swaps which are then batchwise processed. Note that two swaps within a run can depend on each other. If an edge is contained in more than one swap, the nodes incident to the edge may change after the first swap has been executed. We call this a source edge dependency. Since the resulting graph has to remain simple, there further is a dependency between two swaps σi , σj through target edges if executing σi creates or removes an edge created by σj . We model both types of dependencies explicitly and forward information between dependent swaps using Time Forward Processing. As illustrated in Fig. 5, EM-ES executes several phases for each run. They roughly correspond to the four tasks outlined above. For simplicity’s sake, we first assume that all swaps are independent, i.e. that there are no two swaps that share any source edge id or target edge. We will then explain how dependencies are handled. 5.1 EM-ES for Independent Swaps During the request nodes phase, EM-ES requests for each swap σ(ha, bi, ·) the endpoints of the two edges at positions a and b in EL . These requests are then executed in the load nodes phase. In combination, both implement task (T1). Subsequently, the step simulate swaps computes eσa and eσb which corresponds to (T2). In the fourth step, load existence, we check for each of these target edges whether it already exists. Then, the step perform swaps executes swaps iff the graph remains simple; this corresponds to (T3). To implement (T4), the state of all involved edges is materialized in the update graph phase. The communication between the different phases is mostly realized via external memory sorters6 . Independent swaps require only the communication shown at the top of Fig. 5. 6 The term sorter refers to a container with two modes of operation: in the first phase, items are pushed into the write-only sorter in an arbitrary order by some algorithm. After an explicit switch, the filled data structure becomes read-only and the elements are provided as a lexicographically non-decreasing stream which can be rewound at any time. While a sorter is functionally equivalent to filling, sorting and reading back an EM vector, the restricted access model reduces constant factors in the implementation’s runtime and I/O-complexity [5]. 9 markers for edges that receive updates (InvalidEdge) edge state to first swap (EdgeMsg) request nodes incident to edge id (EdgeReq) basic edge switching dependency handling Sorter Stream request nodes EL (swap id) Priority Queue edge existence request (ExistReq) load nodes simulate swaps (edge id) (swap id) inform about successor swap (IdSucc) edge state updates to successor EL edge exist. info to first swap (ExistMsg) edges after processed swaps (EdgeUpdates) load existence perform swaps (edge) (swap id) inform about successor swap (ExistSucc) EL update graph EL (edge) edge state and existence updates to successor Figure 5: Data flow during an EM-ES run. Communication between phases is implemented via EM sorters, self-loops use a PQ-based TFP. Brackets within a phase represent the type of elements iterated over. If multiple input streams are used, they are joined with this key. 5.1.1 Request nodes and load nodes The goal of these two phases is to load every referenced edge. We iterate over the sequence S of swaps. For the s-th swap σ(ha, bi, d), we push the two messages edge req(a, s, 0) and edge req(b, s, 1) into the sorter EdgeReq. A message’s third entry encodes whether the request is issued for the first or second edge of a swap. This information only becomes relevant when we allow dependencies. EM-ES then scans in parallel through the edge list EL and the requests EdgeReq, which are now sorted by edge ids. If there is a request edge req(i, s, p) for an edge ei =[u, v], the edge’s node pair is sent to the requesting swap by pushing a message edge msg(s, p, (u, v)) into the sorter EdgeMsg. Additionally, for every edge we push a bit into the sequence InvalidEdge, which is asserted iff an edge received a request. These edges are considered invalid and will be deleted when updating the graph. Since both phases produce only a constant amount of data per input element, we obtain an I/O complexity of O(sort(r) + scan(m)). 5.1.2 Simulate swaps and load existence The two phases gather all information required to decide whether a swap is legal. EM-ES scans through the sequence S of swaps and EdgeMsg in parallel: For the s-th swap σ(ha, bi, d), there are exactly two messages edge msg(s, 0, ea ) and edge msg(s, 1, eb ) in EdgeMsg. This information suffices to compute the switched edges eσa and eσb , but not to test for multi-edges. To avoid these, it remains to check whether the switched edges already exist. Thus, we push existence requests exist req(eσa , s) and exist req(eσb , s) in the sorter ExistReq (in contrast to request nodes we use the node pairs rather than edge ids, which are not well defined here). Afterwards, a parallel scan through the edge list EL and ExistReq is performed to answer the requests. Only if an edge e requested by swap id s is found, the message exist msg(s, e) is pushed into the sorter ExistMsg. Both phases hence incur a total of O(sort(r) + scan(m)) I/Os. 5.1.3 Perform swaps We rewind the EdgeMsg sorter and jointly scan through the sequence of swaps S and the sorters EdgeMsg and ExistMsg. As described in the simulation phase, EM-ES computes the switched edges eσa and eσb from the original state ea and eb . The swap is marked illegal if a switched edge is a self-loop or if an existence info is received via ExistMsg. If σ is legal we push the switched edges eσa and eσb into the sorter EdgeUpdates, otherwise we propagate the unaltered source edges ea and eb . This phase requires O(sort(r)) I/Os. 5.1.4 Update edge list The new edge list EL0 is obtained by merging the original edge list EL and the updated edges EdgeUpdates, triggering O(scan(m)) I/Os. During this process, we skip all edges in EL that are flagged invalid in the bit stream InvalidEdge. 10 5.2 Inter-Swap Dependencies In contrast to the earlier over-simplification, swaps may share source ids or target edges. In this case, EM-ES produces the same result as a sequential processing of S. Two swaps containing the same source edges are detected during the load nodes phase. In such a case there arrive multiple requests for the same edge id. We record these dependencies as an explicit dependency chain (see below for details). In the simulation phase we do not know yet whether a swap can be executed. Therefore we need to consider both cases (i.e. a swap has been executed or an existing edge prevented its execution) and dynamically forward all possible edge states using a priority queue. In the load existence phase, we detect whether several swaps might produce the same outcome; in this case both issue an existence request for the same edge during simulation. Again, an explicit dependency chain is computed. During the perform swaps phase, EM-ES forwards the source edge states and existence updates to successor swaps using information from both dependency chains. 5.2.1 Target edge dependencies Consider the case where a swap σs1 (ha, bi, d) changes the state of edges ea and eb to eσa 1 and σ1 eb respectively. Later, a second swap σ2 inquires about the existence of either of the four edges which has obviously changed compared to the initial state. We extend the simulation phase in order to track such edge modifications. Here, we not only push messages exist req(eσa 1 , s1 ) and exist req(eσb 1 , s1 ) into sorter EdgeReq, but also report that the original edges may change. This is achieved by using messages exist req(ea , s1 , may change) and exist req(eb , s1 , may change) that are pushed into the same sorter. If there are dependencies, multiple messages are received for the same edge e during the load existence phase. In this case, only the request of the first swap involved is answered as before. Also, every swap σs1 is informed about its direct successor σs2 (if any) by pushing the message exist succ(s1 , e, s2 ) into the sorter ExistSucc, yielding the aforementioned dependency chain. As an optimization, may change requests at the end of a chain are discarded since no recipient exists. During the perform swaps phase, EM-ES executes the same steps as described earlier. The swap may receive a successor for every edge it sent an existence request to, and it informs each successor about the state of the appropriate edge after the swap is processed. 5.2.2 Source edge dependencies Consider two swaps σs1 (ha1 , b1 i, d1 ) and σs2 (ha2 , b2 i, d2 ) with s1 <s2 which share a source edge id, i.e. {a1 , b1 } ∩ {a2 , b2 } is non-empty. This dependency is detected during the load nodes phase, where requests edge req(ei , s1 , p1 ) and edge req(ei , s2 , p2 ) arrive for the same edge id ei . In this case, we answer only the request of s1 and build a dependency chain as described before using messages id succ(s1 , p1 , s2 , p2 ) pushed into the sorter IdSucc. During the simulation phase, EM-ES cannot decide whether a swap is legal. Therefore, s1 sends for every conflicting edge its original state e as well as the updated state eσ1 to the p2 -th slot of s2 using a PQ. If a swap receives multiple edge states per slot, it simulates the swap for every possible combination. During the perform swaps phase, EM-ES operates as described in the independent case: it computes the swapped edges and determines whether the swap has to be skipped. If a successor exists, the new state is not pushed into the EdgeUpdates sorter but rather forwarded to the successor in a TFP fashion. This way, every invalidated edge id receives exactly one update in EdgeUpdates and the merging remains correct. Due to the second modification, EM-ES’s complexity increases with the number of swaps that target the same edge id. This number is quite low in case r = O(m). Let Xi be a random variable expressing the number of swaps that reference edge ei . Since every swap constitutes two independent Bernoulli trials towards ei , the indicator Xi is binomially distributed with p = 1/m, yielding an expected chain length of 2r/m. 11 Input Degree sequence D = (1, 1, 2, 2, 2, 4) Shuffled sequence [6, 6, 4, 5, 6, 1, 3, 2, 3, 6] 1 4 5 [3, 6] 6 3 2 Matched edges Materialized multi-set [1, 2, 3, 3, 4, 4, 5, 5, 6, 6, 6, 6] 4, 5, Resulting graph [6, 6] [4, 5] [4, 5] [1, 6] [2, 3] Figure 6: A Configuration Model run on degree sequence D = (1, 1, 2, 2, 2, 4). Also, for r = m/2 swaps, max1≤i≤n (Xi ) = O(ln(m)/ ln ln(m)) holds with high probability based on a balls-into-bins argument [36]. Thus, we can bound the largest number of edge states simulated with high probability by O(polylog(m)), assuming non-overlapping dependency chains. Further observe that Xi converges towards an independent Poisson distribution for large m. Then the expected state space per edge is O(1). Experiments suggest that this bound also holds for overlapping dependency chains (cf. section 10.2). In order to keep the dependency chains short, EM-ES splits the sequence of swaps S into runs of equal size. Our experimental results show that a run size of r = m/8 is a suitable choice. For every run, the algorithm executes the six phases as described before. Each time the graph is updated, the mapping between an edge and its id may change. The switching probabilities, however, remain unaltered due to the initial assumption of uniformly distributed swaps. Thus EM-ES triggers O(k/m sort(m)) I/Os in total with high probability. 6 EM-CM/ES: Sampling of random graphs from prescribed degree sequence In this section, we propose an alternative approach to generate a graph from a prescribed degree sequence. In contrast to EM-HH which generates a highly biased but simple graph, we use the Configuration Model to sample a random but non-simple graph. Thus, the resulting graph may contain self-loops and multi-edges which we then remove to obtain a simple graph. 6.1 Configuration Model Let D = [di ]1≤i≤n be a degree sequence with n nodes. The Configuration Model builds a multiset of node ids which can be thought of as half-edges (or stubs). It produces a total of di half-edges labelled vi for each node vi . The algorithm then chooses two half-edges uniformly at random and creates an edge according to their labels. It repeats the last step with the remaining half-edges until all are matched. In this naı̈ve implementation, the procedure requires Ω(m) I/Os with high probability for m ≥ cM and any constant c > 1. It is therefore impractical in the fully external setting. As illustrated in Fig. 6 and similar to [26], we instead materialize the multiset as a sequence in which each node appears di times. Subsequently, the sequence is shuffled to obtain a random permutation with O(sort(m)) I/Os by sorting the sequence by a uniform variate drawn for each half-edge7 . Finally, we scan over the shuffled sequence and match pairs of adjacent half-edges. We now give upper bounds for theP number of self-loops and P multi-edges introduced by the Configuration Model. Define hDi := v dv /n and hD2 i := v d2v /n to be the mean and the second moment of the sequence D. In [2] and [37] the expected number of self-loops and multi-edges has already been studied. The results are stated in the following two lemmata. 7 The random permutation can be obtained with O(scan(m)) I/Os in case M > which does not affect the complexity of our total pipeline. 12 √ mB(1 + o(1)) + O(B) [41] Lemma 3. Let D be a degree sequence with n nodes. The expected number of self-loops is given by hD2 i − hDi hD2 i − hDi −→ for n → ∞. 2(hDi − 1/n) 2hDi Lemma 4. Let D be a degree sequence with n nodes. The expected number of multi-edges is bounded by 2 (hD2 i − hDi)2 1 hD2 i − hDi 1 −→ for n → ∞. 2 (D − 1/n)(D − 3/n) 2 hDi Let D be a degree sequence drawn from the powerlaw distribution Pld ([a, b), γ). For fixed γ = 2, we bound the number of illegal edges as a function of a and b. Since each entry in D is independently drawn, it suffices to give bounds for the expected value and the second moment of the underlying distribution. P P For general γ, the expected value and second moment are given by bi=a i−γ+1 and bi=a i−γ+2 R b+1 −p Pb respectively. Both expressions are bound between the two integrals a x dx ≤ i=a i−p ≤ Rb −p dx where p = −γ + 1 or p = −γ + 2 respectively. In the case of γ = 2, the seca−1 x P ond moment is hD2 i = bi=a 1 = b−a+1. Then, by using this identity and the lower bound R b+1 −1 x dx ≤ hDi, we obtain the two following lemmata: a Lemma 5. Let D be drawn from Pld ([a, b), 2). The expected number of self-loops is bounded by 1 b−a+1 . 2 ln(b + 1) − ln(a) Lemma 6. Let D be drawn from Pld ([a, b), 2). The expected number of multi-edges is bounded by 2 1 b−a+1 . 2 ln(b + 1) − ln(a) 6.2 Edge rewiring for non-simple graphs As a consequence of lemmata 5 and 6, graphs generated using the Configuration Model may contain multi-edges and self-loops. In order to detect them, we first sort the edge list lexicographically. Then both types of illegal edges can be detected in a single scan. For each self-loop we issue a swap with a randomly selected partner edge. Similarly, for each group of f > 1 of parallel edges, we generate f −1 swaps with random partner edges. Subsequently, we execute the provisioned swaps using a variant of EM-ES (see below). The process is repeated until all illegal edges have been removed. To accelerate the endgame, we double the number of swaps for each remaining illegal edge in every iteration. Since EM-ES is employed to remove parallel edges based on targeted swaps, it needs to process non-simple graphs. Analogous to the initial formulation, we forbid swaps that introduce multi-edges even if they would reduce the multiplicity of another edge (cf. [50]). Nevertheless, EM-ES requires slight modifications for non-simple graphs. Consider the case where the existence of a multi-edge is inquired several times. Since EL is sorted, the initial edge multiplicities can be counted while scanning EL during the load existence phase. In order to correctly process the dependency chain, we have to forward the (possibly updated) multiplicity information to successor swaps. We annotate the existence tokens exist msg(s, e, #(e)) with these counters where #(e) is the multiplicity of edge e. More precisely, during the perform swaps phase, swap σ1 = σ(ha, bi, d) is informed (amongst others) of multiplicities of edges ea , eb , eσa 1 and eσb 1 by incoming existence messages. If σ1 is legal, we send requested edges and multiplicities of the swapped state to any successor σ2 of σ1 provided in ExistSucc. 13 The swapped state consists of the same edges where multiplicities for eσa 1 and eσb 1 are incremented and decremented for ea and eb . Otherwise, we forward the edges and multiplicities of the unchanged initial state. As an optimization, edges which have been removed (i.e. have multiplicity zero) are omitted. 7 EM-CA: Community Assignment For the sake of simplicity, we first restrict the EM community assignment EM-CA to the nonoverlapping case, in which every node belongs P to exactly one community. Consider a sequence of community sizes S=[sξ ]1≤ξ≤C with n= C ξ=1 sξ and a sequence of intra-community degrees in D=[dv ]1≤v≤n . Let S and D be non-increasing and positive. The task is to find a random surjective assignment χ : V →[C] with: (R1) Every community ξ is assigned sξ nodes as requested, with sξ := {v \| v ∈ V ∧ χ(v)=ξ} . (R2) Every node v ∈ V becomes member of a sufficiently large community χ(v) with sχ(v) > din v . 7.1 Ignoring constraint on community size (R2) Without constraint (R2), the bipartite assignment graph8 χ can be sampled in the spirit of the Configuration Model (cf. section 6): Draw a permutation π of nodes uniformly at random P and assign nodes {vπ(xξ +1) , . . . , vπ(xξ +sξ ) } to community ξ where xξ := ξ−1 i=1 si . To ease later modifications, we prefer an equivalent iterative formulation: while there exists a yet unassigned node u, draw a community X with probability proportional to the number of its remaining free slots (i.e. P[X=ξ] ∝ sξ ). Assign u to X, reduce the community’s probability mass by updating sX ← sX − 1 and repeat. By construction, the first scheme is unbiased and the equivalence of both approaches follows as a special case of Lemma 7. We implement the random selection process efficiently based on a binary tree where each community corresponds to a leaf with a weight equal to the number of free slots in the community. Inner nodes store the total weight of their left subtree. In order Pto draw a community, we sample an integer Y ∈ [0, WC ) uniformly at random where WC := C ξ=1 sξ is the tree’s total weight. Following the tree according to Y yields the leaf corresponding to community X. An I/Oefficient data structure [32] based on lazy evaluation for such dynamic probability distributions enables a fully external algorithm with O(n/B · logM/B (C/B)) = O(sort(n)) I/Os. However, since C < M , we can store the tree in IM, allowing a semi-external algorithm which only needs to scan through D, triggering O(scan(n)) I/Os. 7.2 Enforcing constraint on community size (R2) To enforce (R2), we exploit the monotonicity of S and D. Define pv := max{ξ \| sξ > din v } as the index of the smallest community node v may be assigned to. Since [pv ]v is therefore monotonic itself, it can be computed online with O(1) additional IM and O(scan(n)) I/Os in the fully external setting by scanning through S and D in parallel. In order to restrict the random sampling to the communities {1,P . . . , pv }, we reduce the aforementioned random interval pv −1 to [0, Wv ) where the partial sum Wv := ξ=1 sξ is available while computing pv . We generalize the notation of uniformity to assignments subject to (R2) as follows: Lemma 7. Given S={s1 , . . . , sC } and D, let u, v ∈ V be two nodes with the same constraints (pu = pv ) and let c be an arbitrary community. Further, let χ be an assignment generated by EM-CA. Then, P[χ(u)=c] = P[χ(v)=c]. Proof. Without loss of generality, assume that pu = p1 , i.e. u is one of the nodes with the tightest constraints. If this is not the case, we just execute EM-CA until we reach a node u0 which has the same constraints as u does (i.e. pu0 = pu ), and apply the Lemma inductively. 8 Consider a bipartite graph where the partition classes are given by the communities [C] and nodes [n] respectively. Each edge then corresponds to an assignment. 14 This is legal since EM-CA streams through D in a single pass and is oblivious to any future values. In case c > p1 , neither u nor v can become a member of c. Therefore, P[χ(u)=c] = P[χ(v)=c] = 0 and the claim follows trivially. (i) Now consider the case c ≤ p1 . Let sc be the number of free slots in community c at the P (i) beginning of round i ≥ 1 and W (i) = C j=1 sj their sum at that time. By definition, EM-CA (u) assigns node u to community c with probability P[χ(u)=c] = sc /W (u) . Further, the algorithm has to update the number of free slots. Thus, for iteration i ∈ [n] it holds that s(i) c =   sc if i = 1 (i−1) sc   (i−1) sc −1 if i−1 was assigned to c . otherwise The number of free slots is reduced by one in each step:   C X W (i) =  Sj  − (i − 1) j=1 (u) (1) It remains to show P[χ(u)=c] = sc /W (u) = sc /W (1) and the claim follows by transitivity. For u=1 it is true by definition. Now, consider the induction step for u>1: (u) P[χ(u)=c] = s(u) c /W (u−1) = P[χ(u−1)=c] (u−1) = 7.3 sc sc −1 W (u) (u−1) (u−1) · W (u−1) − sc W (u−1) · W (u) (u−1) (u−1) −1 sc sc sc + P[χ(u−1)6=c] (u) = (u−1) + (u) W W W (u−1) = (u−1) (W (u−1) − 1) sc sc = (u−1) (u−1) W · (W (u−1) − 1) W Ind. Hyp. = (u−1) sc 1 − (u−1) W ! (u−1) sc W (u) (1) sc W (1) Assignment with overlapping communities In the overlapping case, the weight of S increases to account for nodes with multiple memberships. There is further an additional input sequence [νv ]1≤v≤n corresponding to the number of memberships a node v shall have, each of which has din v intra-community neighbors. We then sample not only one community per node v, but νv different ones. Since the number of memberships νv M is small, a duplication check during the repeated sampling is easy in the semi-external case and does not change the I/O complexity. However, it is possible that near the end of the execution there are less free communities than memberships requested. We address this issue by switching to an offline strategy for the last Θ(M ) assignments and keep them in IM. As ν = O(1), there are Ω(ν) communities with free slots for the last Θ(M ) vertices and a legal assignment exists with high probability. The offline strategy proceeds as before until it is unable to find ν different communities for a node. In that case, it randomly picks earlier assignments until swapping the communities is possible. In the fully external setting, the I/O complexity grows linearly in the number of samples taken and is thus bounded by O(ν sort(n)). However, the community memberships are obtained lazily and out-of-order which may assign a node several times to the same community. This corresponds to a multi-edge in the bipartite assignment graph. It can be removed using the rewiring technique detailed in section 6.2. 15 8 8.1 Merging and repairing the intra- and inter-community graphs Global Edge Rewiring The global graph is materialized without taking the community structure into account. It therefore can contain edges between nodes that share a community. Those edges have to be removed as they increase the mixing parameter µ. In accordance with LFR, we use rewiring steps to do so and perform an edge swap for each forbidden edge with a randomly selected partner. Since it is unlikely that such a random swap introduces another illegal edge (if sufficiently many communities exist), the probabilistic approach effectively removes forbidden edges. We apply this idea iteratively and perform multiple rounds until no forbidden edges remain. The community assignment step outputs a lexicographically ordered sequence χ of (v, ξ)pairs containing the community ξ for each node v. For nodes that join multiple communities several such pairs exist. Based on this, we annotate every edge with the communities of both incident vertices by scanning through the edge list twice: once sorted by source nodes, once by target nodes. For each forbidden edge, a swap is generated by drawing a random partner edge id and a swap direction. Subsequently, all swaps are executed using EM-ES which now also emits the set of edges involved. It suffices to restrict the scan for illegal edges to this set since all edges not contained have to be legal. Complexity. Each round needs O(sort(m)) I/Os for selecting the edges and executing the swaps. The number of rounds is usually small but depends on the community size distribution: the smaller the communities, the less likely are edges inside them. 8.2 Community Edge Rewiring In the case of overlapping communities, an edge can be generated as part of multiple clusters. Similarly to section 6.2, we iteratively apply semi-random swaps to remove those parallel edges. Here, however, the selection of random partners is more involved. In order not to violate the community size distribution, both edges of a swap have to belong to the same community. While it is easy to achieve this by considering all communities independently, we need to consider the whole merged graph to detect forbidden edges. In order to do so, we annotate each edge with its community id and merge all communities together into one graph that possibly contains parallel edges. During a scan through all sorted edges, we select from each set of parallel edges all but one as candidates for rewiring. Then, a random partner from the same community is drawn for each of these edges. For this, we sort all edges and the selected candidates by community. By counting the edges per community, we can sample random partners, load them in a second scan, randomize their order, and assign them to the candidates of their community. For the execution, multi-edges need to be considered, i.e. we do not only need to know if an edge exists, but also how often, and update that information accordingly. Together with all loaded edges we also need to store community ids such that we can uniquely identify them and update the correct information. While these steps are possible in external memory, we exploit the fact that there are significantly less communities than nodes. Hence, storing some information per community in internal memory is possible. We assume further that all candidates can be stored in internal memory. If there were too many candidates, we would simply consider only some of them in each round. We can avoid the expensive step of sorting all edges by community for every round using the following observation: When scanning the edges, we can keep track of how many edges of each community we have seen so far. We sort the edge ids to be loaded for every community and keep a pointer on the current position in the list for every community. This allows us to load specific edges of all communities without the need to sort all edges by community. Complexity. The fully external rewiring requires O(sort(m)) I/Os for the initial step and each following round. The semi-external variant triggers only O(scan(m)) I/Os per round. The number of rounds is usually small and the overall runtime spend on this step is insignificant. 16 Nevertheless, the described scheme is a Las Vegas algorithm and there exist (unlikely) instances on which it will fail.9 To mitigate this issue, we allow a small fraction of edges (e.g., 10−3 ) to be removed if we detect a slow convergence. To speed up the endgame, we also draw additional swaps uniformly at random from communities which contain a multi-edge. 9 Implementation We implemented the proposed algorithms in C++ based on the STXXL library [12], providing implementations of EM data structures, a parallel EM sorter, and an EM priority queue. Among others, we applied the following optimizations for EM-ES: • Most message types contain both a swap id and a flag indicating which of the swap’s edges is targeted. We encoded both of them in a single integer by using all but the least significant bit for the swap id and store the flag in there. This significantly reduces the memory volume and yields a simpler comparison operator since the standard integer comparison already ensures the correct lexicographic order. • Instead of storing and reading the sequence of swaps several times, we exploit the implementation’s pipeline structure and directly issue edge id requests for every arriving swap. Since this is the only time edge ids are read from a swap, only the remaining direction flag is stored in an efficient EM vector, which uses one bit per flag and supports I/O-efficient writing and reading. Both steps can be overlapped with an ongoing EM-ES run. • Instead of storing each edge in the sorted external edge list as a pair of nodes, we only store each source node once and then list all targets of that vertex. This still supports sequential scan and merge operations which are the only operations we need. This almost halves the I/O volume of scanning or updating the edge list. • During the execution of several runs we can delay the updating of the edge list and combine it with the load nodes phase of the next run. This reduces the number of scans per additional run from three to two. • We use asynchronous stream adapters for tasks such as streaming from sorters or the generation of random numbers. These adapters run in parallel in the background to preprocess and buffer portions of the stream in advance and hand them over to the main thread. Besides parallel sorting and asynchronous pipeline stage, the current EM-LFR implementation facilitates parallelism only during the generation and randomization of intra-community graphs which can be computed pleasingly parallel. 10 Experimental Results The number of repetitions per data point (with different random seeds) is denoted with S. Errorbars correspond to the unbiased estimation of the standard deviation. For LFR we perform experiments based on two different scenarios: • (lin) — In one setting, the maximal degrees and community sizes scale linearly as a function of n. For a n and ν the parameters are chosen as: µ ∈ {0.2, 0.4, 0.6}, dmin =10ν, dmax =nν/20, γ=2, smin =20, smax =n/10, β=1, O=n. • (const) — In the second setting, we keep the community sizes and the degrees constant and consider only non-overlapping communities. The parameters are chosen as: dmin =50, dmax =10,000, γ=2, smin =50, smax =12,000, β=1, O=n. Real-world networks have been shown to have increasing average degrees as they become larger [30]. Increasing the maximum degree as in our first setting (lin) increases the average 9 Consider a node which is a member of two communities in which it is connected to all other nodes. If only one of its neighbors also appears in both communities, the multi-edge cannot be rewired. 17 100 γ = 1 : 0.379 · x1/1 γ = 2 : 1.380 · x1/2 5 10 γ = 3 : 1.270 · x1/3 Relative frequency Number of unique elements 106 104 103 102 Run size 0.05n Run size 0.125n Run size 0.5n 10−1 10−2 10−3 10−4 10−5 10−6 101 103 104 105 106 107 108 10 Number n of samples 100 Number of edge configuration received by swap Figure 7: Left: Number of distinct elements in n samples (i.e. node degrees in a degree sequence) taken from Pld ([1, n), γ). Right: Overhead induced by tracing inter-swap dependencies. Fraction of swaps as function of the number of edge configurations they receive during the simulation phase. degree. Having a maximum community size of n/10 means, however, that a significant proportion of the nodes belongs to such huge communities which are not very tightly knit due to the large number of nodes of low degree. While a more limited growth is probably more realistic, the exact parameters depend on the network model. Our second parameter set (const) shows an example of much smaller maximum degrees and community sizes. We chose the parameters such that they approximate the degree distribution of the Facebook network in May 2011 when it consisted of 721 million active users as reported in [46]. Note however that strict powerlaw models are unable to accurately mimic Facebook’s degree distribution [46]. Further, they show that the degree distribution of the U.S. users (removing connections to non-U.S. users) is very similar to the one of the Facebook users of the whole world, supporting our use of just one parameter set for different graph sizes. The actual minimum degree of the Facebook network is 1, but the smaller degrees are significantly less prevalent than a power law degree sequence would suggest, which is why we chose a larger value of 50. Our maximum degree of 10,000 is larger than the one reported for Facebook (5000), but the latter was also an arbitrarily enforced limit by Facebook. The expected average degree of this degree sequence is 264, which is slightly higher than the reported 190 (world) or 214 (U.S. only). Our parameters are chosen such that the median degree is approximately 99, which matches the worldwide Facebook network. Similar to the first parameter set, we chose the maximum community size slightly larger than the maximum degree of 12,000 nodes. 10.1 EM-HH ’s state size In Lemma 1, we bound the internal memory consumption of EM-HH by showing that a sequence of n numbers randomly sampled from Pld ([1, n), γ) contains only O(n1/γ ) distinct values with high probability. In order to support Lemma 1 and to estimate the hidden constants, samples of varying size between 103 and 108 are taken from distributions with exponents γ ∈ {1, 2, 3}. Each time, the number of unique elements is computed and averaged over S = 9 runs with identical configurations but different random seeds. The results illustrated in Fig. 7 support the predictions √ with small constants. For the commonly used exponent 2, we find 1.38 n distinct elements in a sequence of length n. 10.2 Inter-Swap Dependencies Whenever multiple swaps target the same edge, EM-ES simulates all possible states to be able to retrieve conflicting edges. We argued that the number of dependencies (and thus the state size) remains manageable if the sequence of swaps is split into sufficiently short runs. We found that for m edges and k swaps, 8k/m runs minimize the runtime for large instances of (lin). As indicated in Fig. 7, in this setting 78.7 % of swaps do not receive additional edge configurations during the simulation phase and less than 0.4 % have to consider more than four additional states. Similarly, 78.6 % of existence requests remain without dependencies. 18 106 104 105 Runtime [s] Runtime [s] 105 103 102 101 100 104 106 107 103 ¯ 100 VL-ES, d= ¯ 1000 VL-ES, d= ¯ 100 EM-ES, d= ¯ 1000 EM-ES, d= 102 Original LFR EM-LFR 105 104 101 108 107 Number n of node 108 109 1010 Number m of edges Figure 8: Left: Runtime on (SysA) of the original LFR implementation and EM-LFR for µ=0.2. Right: Runtime on (SysB) of IM VL-ES and EM-ES on a graph with m of edges and average degree d¯ executing k=10m swaps. 10.3 Performance benchmarks Runtime measurements were conducted on the following systems: • (SysA) — inexpensive compute server : Intel Xeon E5-2630 v3 (8 cores/16 threads, 2.40GHz), 64 GB RAM, 3× Samsung 850 PRO SATA SSD (1 TB). • (SysB) — commodity hardware: Intel Core i7 CPU 970 (6 cores/12 threads, 3.2GHz), 12 GB RAM, 1× Samsung 850 PRO SATA SSD (1 TB). Since edge switching scales linearly in the number of swaps (in case of EM-ES in the number of runs), some of the measurements beyond 3 h runtime are extrapolated from the progress until then. We verified that errors stay within the indicated margin using reference measurements without extrapolation. 10.4 Performance of EM-HH Our implementation of EM-HH produces 180 ± 5 million edges per second on (SysA) up to at least 2 × 1010 edges. Here, we include the computation of the input degree sequence, EM-HH ’s compaction step, as well as the writing of the output to external memory. 10.5 Performance of EM-ES Figure 8 presents the runtime required on (SysB) to process k=10m swaps in an input graph with m edges and an average degree d¯ ∈ {100, 1000}. For reference, the performance of the existing internal memory edge swap algorithm VL-ES based on the authors’ implementation [47] is included. Here we report only on the edge swapping process excluding any precomputation. To achieve comparability, we removed connectivity tests, fixed memory management issues, and adopted the number of swaps. Further, we extended counters for edge ids and accumulated degrees to 64 bit integers in order to support experiments with more than 230 edges. VL-ES slows down by a factor of 25 if the data structure exceeds the available internal memory by less than 10 %. We observe an analogous behavior on machines with larger RAM. EM-ES is faster than VL-ES for all instances with m > 2.5 × 108 edges; those graphs still fit into main memory. FDSM has applications beyond synthetic graphs, and is for instance used on real data to assess the statistical significance of observations [43]. In that spirit, we execute EM-ES on an undirected version of the crawled ClueWeb12 graph’s core [45] which we obtain by deleting all nodes corresponding to uncrawled URLs.10 Performing k = m swaps on this graph with n ≈ 9.8 × 108 nodes and m ≈ 3.7 × 1010 edges is feasible in less than 19.1 h on (SysB). 10 We consider such vertices unrealistically (simple) as they have only degree 1 and account for ≈84 % of nodes in the original graph. 19 7.5 7.0 6.5 6.0 5.5 Nodes: 105 , Mixing: 1.0 ×109 Mean after 5m swaps EM-ES, S = 1138 EM-CMES, S = 600 Degree assortativity Number of triangles 3.05 3.00 2.95 2.90 2.85 ×108 0 1 2 3 Number of swaps per edge k/m 4 5 Nodes: 107 , Mixing: 0.2 1.00 0.98 0.96 0.94 0.92 0.90 Mean after 5m swaps EM-ES, S = 40 EM-CMES, S = 40 0.50 0.48 0.46 0.44 0.42 0 1 2 3 Number of swaps per edge k/m 4 5 Figure 9: Left: Number of triangles on (const) with n = 1 × 105 and µ = 1.0. Right: Degree assortativity on (const) with n = 1 × 107 and µ = 0.2. In order to factor in the increased runtime of EM-CM/ES compared to EM-HH, plots of EM-CM/ES are shifted by the runtime of this phase relative to the execution of EM-ES. As EM-CM/ES is a Las-Vegas algorithm, this incurs an additional error along the x-axis. Bhuiyan et al. propose a distributed edge switching algorithm and evaluate it on a compute cluster with 64 nodes each equipped with two Intel Xeon E5-2670 2.60GHz 8-core processors and 64GB RAM [7]. The authors report to perform k=1.15 × 1011 swaps on a graph with m=1010 generated in a preferential attachment process in less than 3 h. We generate a preferential attachment graph using an EM generator [32] matching the aforementioned properties and carried out edge swaps using EM-ES on (SysA). We observe a slow down of only 8.3 on a machine with 1/128 the number of comparable cores and 1/64 of internal memory. 10.6 Performance of EM-CM/ES and qualitative comparison with EM-ES In section 6, we describe an alternative graph sampling method. Instead of seeding EMES with a highly biased graph using EM-HH , we employ the Configuration Model to quickly generate a non-simple random graph. In order to obtain a simple graph, we then carry out several EM-ES runs in a Las-Vegas fashion. Since EM-ES scans through the edge list in each iteration, runs with very few swaps are inefficient. For this reason, we start the subsequent Markov Chain early: First identify all multi-edges and self-loops and generate swaps with random partners. In a second step, we then introduce additional random swaps until the run contains at least m/10 operations11 . For an experimental comparison between EM-ES and EM-CM/ES, we consider the runtime until both yield a sufficiently uniform random sample. Of course, the uniformity is hard to quantify; similarly to related studies (cf. section 1.1), we estimate the mixing times of both approaches as follows. Starting from a common seed graph G(0) , we generate an ensemble (k) (k) {G1 , . . . , GS } of S 1 instances by applying independent random sequences of k m swaps (jm) each. During this process, we regularly export snapshots Gi of the intermediate instances j ∈ [k/m] of graph Gi . For EM-CM/ES, we start from the same seed graph, apply the algorithm and then carry out k swaps as described above. For each snapshot, we compute a number of metrics, such as the average local clustering coefficient (ACC), the number of triangles, and degree assortativity12 . We then investigate how the distribution of these measures evolves within the ensemble as we carry out an increasing number of swaps. We omit results for ACC since they are less sensitive compared to the other measures (cf. section 10.7). As illustrated in Fig. 9 and Appendix C, all proxy measures converge within 5m swaps with a very small variance. No statistically significant change can be observed compared to a Markov 11 We chose this number as it yields execution times similar to the m/8-setting of EM-ES on simple graphs In preliminary experiments, we also included spectral properties (such as extremal eigenvalues of the adjacency/laplacian matrix) and the closeness centrality of fixed nodes. As these measurement are more expensive to compute and yield qualitatively similar results, we decided not to include them in the larger trials. 12 20 5.0 Local cluster coeff. Degree assortativity Triangle count 4.5 4.0 Covergence after k/m steps Covergence after k/m steps 5.0 3.5 3.0 2.5 2.0 1.5 1.0 5 10 106 Number n of nodes 4.0 3.5 3.0 2.5 2.0 1.5 1.0 5 10 107 Local cluster coeff. Degree assortativity Triangle count 4.5 106 Number n of nodes 107 Figure 10: Number of swaps per edge after which ensembles of graphs with (const), 1 × 105 ≤ n ≤ 1 × 107 and µ = 0.4 (left) and µ = 0.6 (right) converge. Due to computation costs, the ensemble size is reduced from S > 100 to S > 10 for large graphs. chain with 30m operations (which was only computed for a subset of each ensemble). EMHH generates biased instances with special properties, such as a high number of triangles and correlated node degrees, while the features of EM-CM/ES’s output nearly match the converged ensemble. This suggests that the number of swaps to obtain a sufficiently uniform sample can be reduced for EM-CM/ES. Due to computational costs, the study was carried out on multiple machines executing several tasks in parallel. Hence, absolute running times are not meaningful, and we rather measure the computational costs in units of time required to carry out 1m swaps by the same process. This accounts for the offset of EM-CM/ES’s first data point. The number of rounds required to obtain a simple graph depends on the degree distribution. For (const) with n = 1 × 105 and µ = 1, a fraction of 5.1 % of the edges produced by the Configuration Model are illegal. EM-ES requires 18 ± 2 rewiring runs in case a single swap is used per round to rewire an illegal edge. In the default mode of operation, 5.0 ± 0.0 rounds suffice as the number of rewiring swaps per illegal edge is doubled in each round. For larger graphs with n = 1 × 107 , only 0.07 % of edges are illegal and need 2.25 ± 0.40 rewiring runs. 10.7 Convergence of EM-ES In a similar spirit to the previous section, we indirectly investigate the Markov chain’s mixing time as a function of the number of nodes n. To do so, we generate ensembles as before with 1 × 105 ≤ n ≤ 1 × 107 and compute the same graph metrics. For each group and measure, we then search for the first snapshot p in which the measure’s mean is within an interval of half the standard deviation of the final values and subsequently remains there for at least three phases. We then interpret p as a proxy for the mixing time. As depicted in Fig. 10, no measure shows a systematic increase over the two orders of magnitude considered. It hence seems plausible not to increase the number of swaps performed by EM-LFR compared to the original implementation. 10.8 Performance of EM-LFR Figure 8 reports the runtime of the original LFR implementation and EM-LFR as a function of the number of nodes n and ν = 1. EM-LFR is faster for graphs with n ≥ 2.5 × 104 nodes which feature approximately 5 × 105 edges and are well in the IM domain. Further, the implementation is capable of producing graphs with more than 1 × 1010 edges in 17 h.13 Using the same time budget, the original implementation generates graphs more than two orders of magnitude smaller. 13 Roughly 1.5 h are spend in the end-game of the Global Rewiring (at that point less than one edge out of 106 is invalid). In this situation, an algorithm using random I/Os may yield a speed-up. Alternatively, we could simply discard the few remaining invalid edges since they only constitute an insignificant fraction. 21 Mixing µ = 0.6, Cluster: Louvain 1 1 0.8 0.8 0.6 0.6 0.4 0.2 0 AR AR Mixing µ = 0.6, Cluster: Infomap 0.4 Orig NetworKit EM 103 0.2 104 0 105 Orig NetworKit EM 103 104 Number n of nodes 105 Number n of nodes Figure 11: Adjusted rand measure of Infomap/Louvain and ground truth at µ = 0.6 with disjoint clusters, smin = 10, smax = n/20. Mixing: µ = 0.4, Cluster: OSLOM, Overlap: ν = 2 Mixing: µ = 0.4, Cluster: OSLOM, Overlap: ν = 4 Orig EM 1 0.8 NMI NMI 0.8 0.6 0.4 0.2 0 Orig EM 1 0.6 0.4 0.2 103 104 0 105 Number n of nodes 103 104 105 Number n of nodes Figure 12: NMI of OSLOM and ground truth at µ = 0.4 with 2/4 overlapping clusters per node. 10.9 Qualitative Comparison of EM-LFR When designing EM-LFR, we made sure that it closely follows the LFR benchmark such that we can expect it to produce graphs following the same distribution as the original LFR generator. In order to show experimentally that we achieved this goal, we generated graphs with identical parameters using the original LFR implementation and EM-LFR. For disjoint clusters we also compare it with the implementation that is part of NetworKit [44]. Using NetworKit, we evaluate the results of Infomap [40], Louvain [8] and OSLOM [29], three state-of-the-art clustering algorithms [14, 9, 17], and compare them using the adjusted rand measure [24] and NMI [15]. Further, we examine the average local clustering coefficient, a measure for the percentage of closed triangles which shows the presence of locally denser areas as expected in communities [25]. We report these measures for graphs ranging from 103 to 106 nodes. In Fig. 11, 13 and 12 we present a selection of results; all of them can be found in Appendix B. There are only small differences within the range of random noise between the graphs generated by EMLFR and the other two implementations. Note that due to the computational costs above 105 edges, there is only one sample for the original implementation which explains the outliers in Fig. 11. Similar to the results in [14], we also observe that the performance of clustering algorithms drops significantly as the graph’s size grows. This might be due to less clearly defined community structures since the parameters are scaled, and also due to limits of current clustering algorithms. Such behavior clearly demonstrates the necessity of EM-LFR for being able to study this phenomenon on even larger graphs and develop algorithms that are able to handle such instances. 11 Outlook and Conclusion We propose the first I/O-efficient graph generator for the LFR benchmark and the FDSM, which is the most challenging step involved: EM-HH materializes a graph based on a prescribed degree distribution without I/O for virtually all realistic parameters. Including the generation of a powerlaw degree sequence and the writing of the output to disk, our implementation generates 22 Avg. Local Clustering Coeff. Mixing: µ = 0.6 Orig NetworKit EM 1 0.8 0.6 0.4 0.2 0 103 104 105 Number n of nodes Figure 13: Average local clustering coefficient at µ = 0.6 with disjoint clusters. 1.8 × 108 edges per second for graphs exceeding main memory. EM-ES perturbs existing graphs with m edges based on k edge switches using O(k/m · sort(m)) I/Os for k = Ω(m). We demonstrate that EM-ES is faster than the internal memory implementation [47] even for large instances still fitting in main memory and scales well beyond the limited main memory. Compared to the distributed approach by [7] on a cluster with 128 CPUs, EM-ES exhibits a slow-down of only 8.3 on one CPU and hence poses a viable and cost-efficient alternative. Our EM-LFR implementation is orders of magnitude faster than the original LFR implementation for large instances and scales well to graphs exceeding main memory while the generated graphs are equivalent. We further gave evidence indicating that commonly accepted parameters to derive the length of the Edge Switching Markov Chain remain valid for graph sizes approaching the external memory domain and that EM-CM/ES can be used to accelerate the process. Currently, EM-ES does not yet fully exploit the parallelism offered by modern machines. While dependencies between swaps make a parallelization challenging, preliminary experiments indicate that an extension is possible: each run can be split further into smaller batches which can be parallelized in the spirit of [33]. Another possibility for a speedup could be to use the recently proposed Curveball sampling algorithm for graphs with fixed degree sequence [10]. Further studies are necessary to establish whether this really leads to a faster sampling in practice. While the underlying Markov chain seems to require less steps to converge in practice, each step is more expensive. Also, a combination of Curveball with EM-CM/ES seems possible. Further, this is just the starting point for clustering large graphs that exceed main memory using external memory. Not only are new clustering algorithms needed, but also the evaluation of the results needs to be investigated since existing evaluation measures might not be easily computable in external memory. Acknowledgment We thank Hannes Seiwert and Mark Ortmann for valuable discussions on EM-HH. 23 References [1] Alok Aggarwal and Jeffrey S. Vitter. The input/output complexity of sorting and related problems. Communications of the ACM, 31(9), pages 1116–1127, 1988. [2] Omer Angel, Remco van der Hofstad, and Cecilia Holmgren. Limit laws for self-loops and multiple edges in the configuration model. CoRR, abs/1603.07172, 2017. URL: https: //arxiv.org/abs/1603.07172. [3] Lars Arge. The buffer tree: A new technique for optimal I/O-algorithms (extended abstract). In Algorithms and Data Structures, 4th International Workshop, WADS ’95, Kingston, Ontario, Canada, August 16-18, 1995, Proceedings, pages 334–345, 1995. URL: http://dx.doi.org/10.1007/3-540-60220-8_74, doi:10.1007/3-540-60220-8_74. [4] David A. Bader, Henning Meyerhenke, Peter Sanders, Christian Schulz, Andrea Kappes, and Dorothea Wagner. Encyclopedia of Social Network Analysis and Mining, chapter Benchmarking for Graph Clustering and Partitioning, pages 73–82. Springer New York, 2014. URL: http://dx.doi.org/10.1007/978-1-4614-6170-8_23, doi:10.1007/ 978-1-4614-6170-8_23. [5] Andreas Beckmann, Roman Dementiev, and Johannes Singler. Building a parallel pipelined external memory algorithm library. In 23rd IEEE International Symposium on Parallel and Distributed Processing, IPDPS 2009, Rome, Italy, May 23-29, 2009, pages 1–10, 2009. URL: http://dx.doi.org/10.1109/IPDPS.2009.5161001, doi:10.1109/IPDPS. 2009.5161001. [6] Jon Louis Bentley and James B. Saxe. Generating sorted lists of random numbers. ACM Trans. Math. Softw., 6(3):359–364, 1980. URL: http://doi.acm.org/10.1145/355900. 355907, doi:10.1145/355900.355907. [7] Hasanuzzaman Bhuiyan, Jiangzhuo Chen, Maleq Khan, and Madhav V. Marathe. Fast parallel algorithms for edge-switching to achieve a target visit rate in heterogeneous graphs. In 43rd International Conference on Parallel Processing, ICPP 2014, Minneapolis, MN, USA, September 9-12, 2014, pages 60–69, 2014. URL: http://dx.doi.org/10.1109/ICPP. 2014.15, doi:10.1109/ICPP.2014.15. [8] Vincent Blondel, Jean-Loup Guillaume, Renaud Lambiotte, and Etienne Lefebvre. Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment, 2008(10), 2008. URL: http://dx.doi.org/10.1088/1742-5468/2008/10/ P10008. [9] Nazar Buzun, Anton Korshunov, Valeriy Avanesov, Ilya Filonenko, Ilya Kozlov, Denis Turdakov, and Hangkyu Kim. Egolp: Fast and distributed community detection in billionnode social networks. In 2014 IEEE International Conference on Data Mining Workshop, pages 533–540, Dec 2014. doi:10.1109/ICDMW.2014.158. [10] Corrie Jacobien Carstens, Annabell Berger, and Giovanni Strona. Curveball: a new generation of sampling algorithms for graphs with fixed degree sequence. ArXiv e-prints, September 2016. arXiv:1609.05137. [11] Kyrylo Chykhradze, Anton Korshunov, Nazar Buzun, Roman Pastukhov, Nikolay Kuzyurin, Denis Turdakov, and Hangkyu Kim. Distributed generation of billion-node social graphs with overlapping community structure. In Complex Networks V: Proceedings of the 5th Workshop on Complex Networks CompleNet, pages 199–208. Springer International, 2014. URL: http://dx.doi.org/10.1007/978-3-319-05401-8_19, doi: 10.1007/978-3-319-05401-8_19. 24 [12] Roman Dementiev, Lutz Kettner, and Peter Sanders. STXXL: standard template library for XXL data sets. Softw., Pract. Exper., 38(6):589–637, 2008. URL: http://dx.doi.org/ 10.1002/spe.844, doi:10.1002/spe.844. [13] Roger B. Eggleton and Derek Allan Holton. Simple and multigraphic realizations of degree sequences. In Proceedings of the Eighth Australian Conference on Combinatorial Mathematics, Lecture Notes in Mathematics, pages 155–172. Springer, 1980. URL: http://dx.doi.org/10.1007/BFb0091817. [14] Scott Emmons, Stephen G. Kobourov, Mike Gallant, and Katy Börner. Analysis of Network Clustering Algorithms and Cluster Quality Metrics at Scale. PLoS ONE, 11(7):1–18, July 2016. URL: http://dx.doi.org/10.1371/journal.pone.0159161. [15] Alcides Viamontes Esquivel and Martin Rosvall. Comparing network covers using mutual information, 2012. arXiv:1202.0425. URL: http://arxiv.org/abs/1202.0425. [16] Santo Fortunato. Community detection in graphs. Physics Reports, 486(35):75 – 174, 2010. URL: http://www.sciencedirect.com/science/article/pii/ S0370157309002841, doi:10.1016/j.physrep.2009.11.002. [17] Santo Fortunato and Darko Hric. Community detection in networks: A user guide. Physics Reports, 659:1–44, 2016. URL: https://dx.doi.org/10.1016/j.physrep.2016.09.002. [18] Christos Gkantsidis, Milena Mihail, and Ellen W. Zegura. The Markov Chain Simulation Method for Generating Connected Power Law Random Graphs. In Proceedings of the 5th Workshop on Algorithm Engineering and Experiments (ALENEX’03), pages 16–25. SIAM, 2003. [19] Catherine S. Greenhill and Matteo Sfragara. The switch markov chain for sampling irregular graphs and digraphs. CoRR, abs/1701.07101, 2017. URL: http://arxiv.org/abs/1701. 07101. [20] Seifollah L. Hakimi. On realizability of a set of integers as degrees of the vertices of a linear graph. i. Journal of the Society for Industrial and Applied Mathematics, 10(3):496–506, 1962. doi:10.1137/0110037. [21] Michael Hamann, Ulrich Meyer, Manuel Penschuck, and Dorothea Wagner. I/O-efficient Generation of Massive Graphs Following the LFR Benchmark. In Proceedings of the 19th Meeting on Algorithm Engineering and Experiments (ALENEX’17), pages 58–72. SIAM, 2017. URL: http://dx.doi.org/10.1137/1.9781611974768.5. [22] Steve Harenberg, Gonzalo Bello, L. Gjeltema, Stephen Ranshous, Jitendra Harlalka, Ramona Seay, Kanchana Padmanabhan, and Nagiza Samatova. Community detection in largescale networks: a survey and empirical evaluation. Wiley Interdisciplinary Reviews: Computational Statistics, 6(6):426–439, 2014. URL: http://dx.doi.org/10.1002/wics.1319, doi:10.1002/wics.1319. [23] Václav Havel. Poznámka o existenci konečných grafů. Časopis pro pěstovánı́ matematiky, 080(4):477–480, 1955. URL: http://eudml.org/doc/19050. [24] Lawrence Hubert and Phipps Arabie. Comparing partitions. Journal of Classification, 2(1):193–218, December 1985. URL: http://dx.doi.org/10.1007/BF01908075. [25] Marcus Kaiser. Mean clustering coefficients: the role of isolated nodes and leafs on clustering measures for small-world networks. New Journal of Physics, 10(8), 2008. URL: http: //dx.doi.org/10.1088/1367-2630/10/8/083042. 25 [26] Panqanamala Ramana Kumar, Martin J Wainwright, and Riccardo Zecchina. Mathematical Foundations of Complex Networked Information Systems: Politecnico Di Torino, Verrès, Italy 2009, volume 2141. Springer, 2015. [27] Andrea Lancichinetti and Santo Fortunato. Benchmarks for testing community detection algorithms on directed and weighted graphs with overlapping communities. Phys. Rev. E, 80:016118, Jul 2009. Source code available at https://sites.google.com/site/ santofortunato/inthepress2. URL: http://link.aps.org/doi/10.1103/PhysRevE. 80.016118, doi:10.1103/PhysRevE.80.016118. [28] Andrea Lancichinetti, Santo Fortunato, and Filippo Radicchi. Benchmark graphs for testing community detection algorithms. Phys. Rev. E, 78:046110, Oct 2008. URL: http://link. aps.org/doi/10.1103/PhysRevE.78.046110, doi:10.1103/PhysRevE.78.046110. [29] Andrea Lancichinetti, Filippo Radicchi, José J. Ramasco, and Santo Fortunato. Finding Statistically Significant Communities in Networks. PLoS ONE, 6(4):1–18, April 2011. URL: http://dx.doi.org/10.1371/journal.pone.0018961. [30] Jure Leskovec, Jon M. Kleinberg, and Christos Faloutsos. Graphs Over Time: Densification Laws, Shrinking Diameters and Possible Explanations. In Proceedings of the 11th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 177– 187. ACM Press, 2005. URL: http://portal.acm.org/citation.cfm?id=1081893. [31] Anil Maheshwari and Norbert Zeh. A survey of techniques for designing I/O-efficient algorithms. In Algorithms for Memory Hierarchies, pages 36–61, 2003. URL: http: //dx.doi.org/10.1007/3-540-36574-5_3, doi:10.1007/3-540-36574-5_3. [32] Ulrich Meyer and Manuel Penschuck. Generating massive scale-free networks under resource constraints. In Proceedings of the Eighteenth Workshop on Algorithm Engineering and Experiments (ALENEX 2016), pages 39–52, 2016. URL: http://epubs.siam.org/doi/ abs/10.1137/1.9781611974317.4, doi:10.1137/1.9781611974317.4. [33] Ulrich Meyer and Peter Sanders. Delta-stepping: A parallel single source shortest path algorithm. In Algorithms - ESA ’98, 6th Annual European Symposium, Venice, Italy, August 24-26, 1998, Proceedings, pages 393–404, 1998. URL: http://dx.doi.org/10.1007/ 3-540-68530-8_33, doi:10.1007/3-540-68530-8_33. [34] Ulrich Meyer, Peter Sanders, and Jop F. Sibeyn, editors. Algorithms for Memory Hierarchies, Advanced Lectures [Dagstuhl Research Seminar, March 10-14, 2002], volume 2625 of Lecture Notes in Computer Science. Springer, 2003. [35] Ron Milo, Nadav Kashtan, Shalev Itzkovitz, Mark E. J. Newman, and Uri. Alon. On the uniform generation of random graphs with prescribed degree sequences. eprint arXiv:condmat/0312028, December 2003. arXiv:cond-mat/0312028. [36] Rajeev Motwani and Prabhakar Raghavan. Randomized algorithms. Chapman & Hall/CRC, 2010. [37] Mark Newman. Networks: An Introduction. Oxford University Press, Inc., New York, NY, USA, 2010. [38] Rasmus Pagh. Basic external memory data structures. In Algorithms for Memory Hierarchies, pages 14–35. Springer, 2003. [39] Jaideep Ray, Ali Pinar, and C. Seshadhri. Are We There Yet? When to Stop a Markov Chain while Generating Random Graphs. In Algorithms and Models for the 26 Web-Graph, Proceedings of the 9th International Workshop, WAW 2012, Lecture Notes in Computer Science, pages 153–164. Springer, 2012. URL: https://doi.org/10.1007/ 978-3-642-30541-2_12. [40] Martin Rosvall, Daniel Axelsson, and Carl T. Bergstrom. The map equation. The European Physical Journal Special Topics, 178(1):13–23, 2009. URL: http://dx.doi.org/10.1140/ epjst/e2010-01179-1. [41] Peter Sanders. Random permutations on distributed, external and hierarchical memory. Inf. Process. Lett., 67(6):305–309, 1998. URL: http://dx.doi.org/10.1016/S0020-0190(98) 00127-6, doi:10.1016/S0020-0190(98)00127-6. [42] Peter Sanders. Fast priority queues for cached memory. Journal of Experimental Algorithmics (JEA), 5, 2000. [43] Wolfgang E. Schlauch, Emőke Ágnes Horvát, and Katharina A. Zweig. Different flavors of randomness: comparing random graph models with fixed degree sequences. Social Network Analysis and Mining, 5(1):1–14, 2015. URL: http://dx.doi.org/10.1007/ s13278-015-0267-z, doi:10.1007/s13278-015-0267-z. [44] Christian Staudt, Aleksejs Sazonovs, and Henning Meyerhenke. Networkit: A tool suite for large-scale complex network analysis. Network Science, 4(4):508530, 2016. doi:10.1017/ nws.2016.20. [45] The Lemur Project. ClueWeb12 Web Graph, Nov 2013. http://www.lemurproject.org/ clueweb12/webgraph.php. [46] Johan Ugander, Brian Karrer, Lars Backstrom, and Cameron Marlow. The anatomy of the facebook social graph. CoRR, abs/1111.4503, 2011. URL: http://arxiv.org/abs/1111. 4503. [47] Fabien Viger and Matthieu Latapy. Fast generation of random connected graphs with prescribed degrees. CoRR, abs/cs/0502085, 2005. Source code available at https: //www-complexnetworks.lip6.fr/~latapy/FV/generation.html. URL: http://arxiv. org/abs/cs/0502085. [48] Jeffrey Scott Vitter. An efficient algorithm for sequential random sampling. ACM Trans. Math. Softw., 13(1):58–67, 1987. URL: http://doi.acm.org/10.1145/23002.23003, doi: 10.1145/23002.23003. [49] Jianping Zeng and Hongfeng Yu. A study of graph partitioning schemes for parallel graph community detection. Parallel Computing, 2016. URL: http://dx.doi.org/10.1016/j. parco.2016.05.008. [50] James Y. Zhao. Expand and contract: Sampling graphs with given degrees and other combinatorial families. CoRR, abs/1308.6627, 2013. URL: http://arxiv.org/abs/1308. 6627. 27 A Summary of Definitions Symbol [k] [u, v] B dmin , dmax din v D DD n m µ M Pld ([a, b), γ) smin , smax scan(n) sort(n) Description [k] := {1, . . . , k} for k ∈ N+ (section 2) Undirected simple edge with implication u < v (section 2) Number of items in a block transferred between IM and EM (section 2.1) Min/max degree of nodes in LFR benchmark (section 3) din v = (1−µ) · dv , intra-community degree of node v (section 3) D = (d1 , . . . , dn ) with di ≤ di+1 ∀i. Degree sequence of a graph (section 4) DD = {di : 1 ≤ i ≤ n} where D = (d1 , . . . , dn ), degree support (section 4) Number of vertices in a graph (section 2) Number of edges in a graph (section 2) Mixing parameter in LFR benchmark, i.e. ratio of neighbors that shall be in other communities (section 3) Number of items fitting into internal memory (section 2.1) Powerlaw distribution with exponent −γ on the interval [a, b) (section 2) Min/max size of communities in LFR benchmark (section 3) scan(n) = Θ(n/B) I/Os, scan complexity (section 2.1) sort(n) = Θ((n/B) · logM/B (n/B)) I/Os, sort complexity (section 2.1) Table 2: Definitions used in this paper. 28 B Comparing LFR Implementations Mixing µ = 0.4, Cluster: Infomap 0.8 0.8 0.8 0.6 0.6 0.2 0 0.4 Orig NetworKit EM 103 0.2 104 0 105 AR 1 0.4 103 104 0.8 Orig NetworKit EM 0.6 0.4 0.2 0 105 NMI 0.8 0.6 104 Orig NetworKit EM 103 0.4 0 105 Mixing µ = 0.4, Cluster: Louvain 0.8 0.8 0.6 0.6 0.4 Orig NetworKit EM 0.2 104 0 105 AR 0.8 AR 1 103 103 0.2 104 0 105 Mixing µ = 0.4, Cluster: Louvain 0.8 0.8 0.6 Orig NetworKit EM 0.6 0.4 0.2 104 0 105 NMI 0.8 NMI 1 103 Orig NetworKit EM 103 Number n of nodes 0.4 0.2 104 0 105 Mixing: µ = 0.4 106 106 106 Orig NetworKit EM 105 Orig NetworKit EM 104 103 105 103 Mixing: µ = 0.2 Orig NetworKit EM 1 0.8 0.6 0.4 0.2 0 103 104 Number n of nodes Orig NetworKit EM 104 103 105 103 105 Orig NetworKit EM 0.8 0.6 0.4 0.2 0 103 104 Number n of nodes 104 105 Number n of nodes Mixing: µ = 0.4 1 105 105 Number n of nodes Avg. Local Clustering Coeff. Number n of nodes 104 105 Avg. Local Clustering Coeff. 104 Edges 107 Edges 108 107 105 104 Mixing: µ = 0.6 108 103 103 Number n of nodes 107 103 Orig NetworKit EM Number n of nodes Mixing: µ = 0.2 105 0.6 108 104 104 Mixing µ = 0.6, Cluster: Louvain 1 0 103 Number n of nodes 1 0.2 Orig NetworKit EM Number n of nodes Mixing µ = 0.2, Cluster: Louvain 105 0.6 0.4 Orig NetworKit EM Number n of nodes 0.4 104 Mixing µ = 0.6, Cluster: Louvain 1 0 103 Number n of nodes 1 0.2 Orig NetworKit EM Number n of nodes Mixing µ = 0.2, Cluster: Louvain 105 0.6 0.2 104 Number n of nodes 0.4 104 Mixing µ = 0.6, Cluster: Infomap 0.8 103 103 Number n of nodes 1 0.4 Orig NetworKit EM Mixing µ = 0.4, Cluster: Infomap NMI NMI 105 1 0 AR 0 1 0.2 NMI 0.2 Number n of nodes Mixing µ = 0.2, Cluster: Infomap Edges 0.6 0.4 Orig NetworKit EM Number n of nodes Avg. Local Clustering Coeff. Mixing µ = 0.6, Cluster: Infomap 1 AR AR Mixing µ = 0.2, Cluster: Infomap 1 Mixing: µ = 0.6 Orig NetworKit EM 1 0.8 0.6 0.4 0.2 0 103 104 105 Number n of nodes Comparison of the original LFR implementation, the NetworKit implementation and our EM solution for values of 103 ≤ n ≤ 106 , µ∈{0.2, 0.4, 0.6}, γ=2, β= − 1 dmin =10, dmax =n/20, smin =10, smax =n/20. Clustering is performed using Infomap and Louvain and compared to the ground-truth emitted by the generator using AdjustedRandMeasure (AR) and Normalized Mutual Information (NMI); S ≥ 8. Due to the computational costs, graphs with n ≥ 105 have a reduced multiplicity. In case of the original implementation it may be based on a single run which accounts for the few outliers. 29 Mixing: µ = 0.2, Cluster: OSLOM, Overlap: ν = 2 Mixing: µ = 0.4, Cluster: OSLOM, Overlap: ν = 2 Orig EM 1 0.4 0.2 0.6 0.4 0.2 104 0 105 Orig EM 0.8 0.6 0.4 0.2 103 104 105 Orig EM 0.6 0.4 0.2 0 103 0.6 0.4 0.2 104 0.6 0.4 0.2 103 0 0 104 105 0.4 0.2 104 105 0.4 0.2 104 Orig EM 0.8 0.6 0.4 0.2 0 103 105 0.4 103 0.6 0.4 0.2 104 Mixing: µ = 0.6 Orig EM 0.8 0.6 0.4 0.2 0 103 0.8 0.6 0.4 0.2 103 0.2 Orig EM 0.8 0.6 0.4 0 105 0.6 0.4 0.2 103 104 0 105 103 Number n of nodes Orig EM 0.8 0.6 0.4 0.2 105 Avg. Local Clustering Coeff. Number n of nodes Mixing: µ = 0.2 Orig EM 0.8 0.6 0.4 0.2 103 104 Number n of nodes 30 104 105 Number n of nodes Mixing: µ = 0.4 1 0 105 Mixing: µ = 0.6, Cluster: OSLOM, Overlap: ν = 4 1 0.2 104 104 Number n of nodes NMI NMI 0.4 105 Orig EM 1 0 105 0.8 0.6 104 Mixing: µ = 0.6, Degree Assortativity, Overlap: ν = 3 Orig EM 1 105 Number n of nodes Orig EM 103 104 1 Mixing: µ = 0.4, Cluster: OSLOM, Overlap: ν = 4 Orig EM 105 Orig EM Number n of nodes 0.8 NMI 105 0.8 Mixing: µ = 0.2, Cluster: OSLOM, Overlap: ν = 4 Avg. Local Clustering Coeff. 104 1 0 104 Number n of nodes Mixing: µ = 0.4 Number n of nodes Number n of nodes 103 0.6 0 105 Degree Assortativity 0.6 104 0.2 Mixing: µ = 0.4, Degree Assortativity, Overlap: ν = 3 Degree Assortativity Degree Assortativity Orig EM 103 0.4 Number n of nodes 0.8 0 104 1 Mixing: µ = 0.2, Degree Assortativity, Overlap: ν = 3 1 0.6 0.2 103 Number n of nodes 103 0.8 1 Avg. Local Clustering Coeff. 0.6 Avg. Local Clustering Coeff. Avg. Local Clustering Coeff. 0.8 0 Orig EM 1 Number n of nodes Orig EM 105 0.8 Number n of nodes Mixing: µ = 0.2 104 Mixing: µ = 0.6, Cluster: OSLOM, Overlap: ν = 3 Orig EM 0.4 0.2 1 103 Number n of nodes 0.6 0.2 103 0 0 105 NMI NMI NMI 0.4 1 0.2 Mixing: µ = 0.6, Degree Assortativity, Overlap: ν = 2 0.8 0.6 0 104 1 0.8 103 0.4 Mixing: µ = 0.4, Cluster: OSLOM, Overlap: ν = 3 Orig EM 1 0.6 Number n of nodes Mixing: µ = 0.2, Cluster: OSLOM, Overlap: ν = 3 103 Orig EM 0.8 Number n of nodes Orig EM Number n of nodes 0 105 0.8 0 105 105 Mixing: µ = 0.6 Mixing: µ = 0.4, Degree Assortativity, Overlap: ν = 2 Degree Assortativity Degree Assortativity 0.8 1 104 1 104 1 Number n of nodes Orig EM 103 103 Number n of nodes 0.8 Mixing: µ = 0.2, Degree Assortativity, Overlap: ν = 2 0 0 105 Mixing: µ = 0.4 1 Number n of nodes 1 104 Avg. Local Clustering Coeff. Mixing: µ = 0.2 0 0.4 Number n of nodes Avg. Local Clustering Coeff. Avg. Local Clustering Coeff. Number n of nodes 1 0.6 0.2 103 Degree Assortativity 103 105 Avg. Local Clustering Coeff. 0 0.8 NMI 0.6 Orig EM 1 0.8 NMI NMI 0.8 Mixing: µ = 0.6, Cluster: OSLOM, Overlap: ν = 2 Orig EM 1 Mixing: µ = 0.6 Orig EM 1 0.8 0.6 0.4 0.2 0 103 104 Number n of nodes 105 0.8 0.6 0.4 0.2 0 103 104 Number n of nodes Mixing: µ = 0.4, Degree Assortativity, Overlap: ν = 4 105 Orig EM 1 0.8 0.6 0.4 0.2 0 103 104 Number n of nodes Mixing: µ = 0.6, Degree Assortativity, Overlap: ν = 4 Degree Assortativity Orig EM Degree Assortativity Degree Assortativity Mixing: µ = 0.2, Degree Assortativity, Overlap: ν = 4 1 105 Orig EM 1 0.8 0.6 0.4 0.2 0 103 104 105 Number n of nodes Comparison of the original LFR implementation and our EM solution for values values of 103 ≤ n ≤ 106 , µ∈{0.2, 0.4, 0.6}, ν∈{2, 3, 4}, O = n, γ=2, β= − 1 dmin =10, dmax =n/20, smin =10ν, smax =ν · n/20. Clustering is performed using OSLOM and compared to the groundtruth emitted by the generator using a generalized Normalized Mutual Information (NMI); S ≥ 5. 31 Comparing EM-ES and EM-CM/ES Mean after 5m swaps EM-ES, S = 1170 EM-CMES, S = 600 ×107 2 3 Number of swaps per edge k/m 4 3.05 3.00 2.95 2.90 2.85 7.5 7.0 6.5 6.0 5.5 5 Nodes: 105 , Mixing: 1.0 ×109 Mean after 5m swaps EM-ES, S = 1138 EM-CMES, S = 600 ×108 0 1 Nodes: 105 , Mixing: 0.2 0.90 0.88 0.86 0.84 0.82 0.80 0.50 0.48 0.46 0.44 0.42 0.40 1 4 1 2 3 Number of swaps per edge k/m 4 2.4 2.2 2.0 1.8 1.6 1.4 1.2 5 1 2 3 Number of swaps per edge k/m 4 5 Nodes: 107 , Mixing: 1.0 ×1010 Mean after 5m swaps EM-ES, S = 20 EM-CMES, S = 20 ×109 0 0.50 0.48 0.46 0.44 0.42 4 6.38 6.36 6.34 6.32 6.30 6.28 6.26 5 Mean after 5m swaps EM-ES, S = 40 EM-CMES, S = 40 2 3 Number of swaps per edge k/m 5 Mean after 5m swaps EM-ES, S = 1138 EM-CMES, S = 720 0 Nodes: 107 , Mixing: 0.2 1 4 0.39 0.38 0.37 0.36 0.35 0.34 0.33 0.32 5 Mean after 5m swaps EM-ES, S = 40 EM-CMES, S = 40 1.00 0.98 0.96 0.94 0.92 0.90 0 0.61 0.60 0.59 0.58 0.57 0.56 0.55 0.54 Nodes: 107 , Mixing: 0.2 ×108 5.12 5.10 5.08 5.06 5.04 5.02 ×107 2.0 1.8 1.6 1.4 1.2 1.0 0 2 3 Number of swaps per edge k/m 2 3 Number of swaps per edge k/m Nodes: 105 , Mixing: 1.0 Mean after 5m swaps EM-ES, S = 1170 EM-CMES, S = 600 0 Number of triangles 1 Number of triangles Degree assortativity 0 Degree assortativity Number of triangles 1.3 1.2 1.1 1.0 0.9 0.8 Nodes: 105 , Mixing: 0.2 ×108 Degree assortativity 1.25 1.24 1.23 1.22 1.21 1.20 Degree assortativity Number of triangles C 1 2 3 Number of swaps per edge k/m 4 5 Nodes: 107 , Mixing: 1.0 1.00 0.98 0.96 0.94 0.92 0.90 Mean after 5m swaps EM-ES, S = 20 EM-CMES, S = 20 0.50 0.48 0.46 0.44 0.42 0 1 2 3 Number of swaps per edge k/m 4 5 Triangle count and degree assortativity of a graph ensemble obtained by applying random swaps/the Configuration Model to a common seed graph. Refer to section 10.6 for experimental details. 32	8cs.DS
arXiv:1510.08568v2 [cs.NE] 8 Apr 2016 Feature-Based Diversity Optimization for Problem Instance Classification Wanru Gao Optimisation and Logistics School of Computer Science The University of Adelaide Adelaide, Australia Samadhi Nallaperuma Department of Computer Science The University of Sheffield Sheffield,UK Frank Neumann Optimisation and Logistics School of Computer Science The University of Adelaide Adelaide, Australia April 11, 2016 Abstract Understanding the behaviour of heuristic search methods is a challenge. This even holds for simple local search methods such as 2-OPT for the Traveling Salesperson problem. In this paper, we present a general framework that is able to construct a diverse set of instances that are hard or easy for a given search heuristic. Such a diverse set is obtained by using an evolutionary algorithm for constructing hard or easy instances that are diverse with respect to different features of the underlying problem. Examining the constructed instance sets, we show that many combinations of two or three features give a good classification of the TSP instances in terms of whether they are hard to be solved by 2-OPT. 1 Introduction Heuristic search methods such as local search, simulated annealing, evolutionary algorithms and ant colony optimization have been shown to be very successful for various combinatorial optimization problems. Although they usually don’t come with performance guarantees on their runtime and/or approximation behaviour, they often perform very well in several situations. Understanding the conditions under which optimization algorithms perform well is essential for automatic algorithm selection, configuration and effective algorithm design. In both the artificial intelligence (AI) [23, 9, 22, 5, 7] and operational research communities [17, 21], this topic has become a major point of interest. 1 The feature-based analysis of heuristic search algorithms has become an important part in understanding such type of algorithms [11, 18]. This approach characterizes algorithms and their performance for a given problem based on features of problem instances. Thereby, it provides an important tool for bridging the gap between pure experimental investigations and mathematical methods for analysing the performance of search algorithms [15, 10, 6]. Current methods for the feature-based analysis are based on constructing hard and easy instances for an investigated search heuristic and a given optimization problem by evolving instances using an evolutionary algorithm [11, 14, 13]. This evolutionary algorithm constructs problem instances where the examined algorithm either shows a bad (good) approximation behaviour and/or requires a large (small) computational effort to come up with good or optimal solutions. Although the evolutionary algorithm for constructing such instances is usually run several times to obtain a large set of hard (easy) instances, the question arises whether the results in terms of features of the instances obtained give a good characterization of problem difficulty. In the paper, we propose a new approach of constructing hard and easy instances. Following some recent work on using evolutionary algorithms for generating diverse sets of instances that are all of high quality [19, 20], we introduce an evolutionary algorithm which maximizes diversity of the obtained instances in terms of a given feature. Our approach allows to generate a set of instances that is much more diverse with respect to the problem feature at hand. Carrying out this process for several features of the considered problem and algorithm gives a much better classification of instances according to their difficulty of being solved by the considered algorithm. To show the benefit of our approach compared to previous methods, we consider the classical 2-OPT algorithm for the TSP. Previous feature-based analyses have already considered hard and easy instances in terms of approximation ratio and analyzed the features of such hard (easy) instances obtained by an evolutionary algorithm. The experimental results of our new approach show that diversity optimization of the features results in an improved coverage of the feature space over classical instance generation methods. In particular, the results show that for some combinations of two features it is possible to classify hard and easy instances into two clusters with a wider coverage of the feature space compared to the classical methods. Moreover, the three-feature combinations further improve the classification of hard and easy instances for most of the feature combinations. Furthermore, a classification model is built using these diverse instances that can classify TSP instances based on hardness for 2-OPT. The remainder of this paper is organized as follows. Firstly, we introduce the Euclidean TSP and the background on feature based analysis. Afterwards, we state our diversity optimization approach for evolving instances according to feature values and report on the impact of diversity optimization in terms of the range of feature values. As features value can be very diverse both for easy and hard instances, we consider the combinations of several features for instance classification afterwards. We then build a classification model that can classify instances based on hardness and finally finish with some conclusions. 2 2 Background We consider the classical NP-hard Euclidean Traveling Salesperson problem (TSP) as the example problem for evolving hard and easy instances which have a diverse set of features. Our methodology can be applied to any optimization problem, but using the TSP in our study has the advantage that it has already been investigated extensively from different perspectives including the area of feature-based analysis. The input of the problem is given by a set V = {v1 , ..., vn } of n cities in the Euclidean plane and Euclidean distances d : V ×V → R≥0 between the cities. The goal is to find a Hamiltonian cycle whose sum of distances is minimal. A candidate solution for the TSP is often represented by a permutation π = (π1 , . . . , πn ) of the n cities and the goal is to find a permutation π ∗ which minimizes the tour length given by n−1 c(π) = d(πn , π1 ) + ∑ d(πi , πi+1 ). i=1 For our investigations cities are always in the normalized plane [0, 1]2 , i. e. each city has an x- and y-coordinate in the interval [0, 1]. In following, a TSP instance always consists of a set of n points in [0, 1]2 and the Euclidean distances between them. Local search heuristics have been shown to be very successful when dealing with the TSP and the most prominent local search operator is the 2-OPT operator [4]. The resulting local search algorithm starts with a random permutation of the cities and repeatedly checks whether removing two edges and reconnecting the two resulting paths by two other edges leads to a shorter tour. If no improvement can be found by carrying out any 2-OPT operation, the tour is called locally optimal and the algorithm terminates. The key factor in the area of feature-based analysis is to identify the problem features and their contribution to the problem hardness for a particular algorithm and problem combination. This can be achieved through investigating hard and easy instances of the problem. Using an evolutionary algorithm, it is possible to evolve sets of hard and easy instances by maximizing or minimizing the fitness (tour length in the case of the TSP) of each instance [21, 17, 11, 18]. However, none of these approaches have considered the diversity of the instances explicitly. Within this study we expect to improve the evolutionary algorithm based instance generation approach by introducing diversity optimization. The structural features are dependent on the underlying problem. In [11], there are 47 features in 8 groups used to provide an understanding of algorithm performance for the TSP. The different feature classes established are distance features, mode features, cluster features, centroid features, MST features, angle features and convex hull features. The feature values are regarded as indicators which allow to predict the performance of a given algorithm on a given instance. 3 Feature-Based Diversity Optimization In this section, we introduce our approach of evolving a diverse set of easy or hard instances which are diverse with respect to important problem features. As in previous 3 Algorithm 1: (µ + λ )-EAD 1 2 3 4 5 Initialize the population P with µ TSP instances of approximation ratio at least αh . Let C ⊆ P where \|C\| = λ . For each I ∈ C, produce an offspring I 0 of I by mutation. If αA (I 0 ) > αh , add I 0 to P. While \|P\| > µ, remove an individual I = arg minJ∈P d(J, P) uniformly at random. Repeat step 2 to 4 until termination criterion is reached. studies, we measure hardness of a given instance by the ratio of the solution quality obtained by the considered algorithm and the value of an optimal solution. The approximation ratio of an algorithm A for a given instance I is defined as αA (I) = A(I)/OPT (I) where A(I) is value of the solution produced by algorithm A for the given instance I, and OPT (I) is value of an optimal solution for instance I. Within this study, A(I) is the tour length obtained by 2-OPT for a given TSP instance I and OPT (I) is the optimal tour length which we obtain in our experiments by using the exact TSP solver Concorde [1]. We propose to use an evolutionary algorithm to construct sets of instances of the TSP that are quantified as either easy or hard in terms of approximation and are diverse with respect to underlying features of the produced problem instances. Our evolutionary algorithm (shown in Algorithm 1) evolves instances which are diverse with respect to given features and meet given approximation ratio thresholds. The algorithm is initialized with a population P consisting of µ TSP instances which have an approximation ratio at least αh in the case of generating a diverse set of hard instances. In the case of easy instances, we start with a population where all instances have an approximation ratio of at most αe and only instances of approximation ratio at most αe can be accepted for the next iteration. In each iteration, λ ≤ µ offspring are produced by selecting λ parents and applying mutation to the selected individuals. Offsprings that don’t meet the approximation threshold are rejected immediately. The new parent population is formed by reducing the set consisting of parents and offsprings satisfying the approximation threshold until a set of µ solutions is achieved. This is done by removing instances one by one based on their contribution to the diversity according to the considered feature. The core of our algorithm is the selection among individuals meeting the threshold values for the approximation quality according to feature values. Let I1 , . . . , Ik be the elements of P and f (Ii ) be their features values. Furthermore, assume that f (Ii ) ∈ [0, R], i.e. feature values are non-negative and upper bounded by R. We assume that f (I1 ) ≤ f (I2 ) ≤ . . . ≤ f (Ik ) holds. The diversity contribution of an instance I to a population of instances P is defined as d(I, P) = c(I, P) 4 where c(I, P) is a contribution based on other individuals in the population Let Ii be an individual for which f (Ii ) 6= f (I1 ) and f (Ii ) 6= f (Ik ). We set c(Ii , P) = ( f (Ii ) − f (Ii−1 )) · ( f (Ii+1 ) − f (Ii )) which assigns the diversity contribution of an individual based on the next smaller and next larger feature values. If f (Ii ) = f (I1 ) or f (Ii ) = f (Ik ), we set c(Ii , P) = R2 if there is no other individual I 6= Ii in P with f (I) = f (Ii ) and c(Ii , P) = 0 otherwise. This implies an individual Ii with feature value equal to any other instances in the population gains c(Ii , P) = 0. Furthermore, an individual with the unique smallest and largest feature value always stays in the population when working with µ ≥ 2. In [11], 47 features of TSP instances for characterizing easy and hard TSP instances have been studied. We consider 7 features coming from different feature classes which have shown to be well suited for classification and prediction. These features are: • angle mean :- mean value of the angles made by each point with its two nearest neighbor points • centroid mean distance to centroid :- mean value of the distances from the points to the centroid • chull area : - area covered by the convex hull • cluster 10pct mean distance to centroid :- mean value of the distances to cluster centroids at 10% levels of reachability • mst depth mean :- mean depth of the minimum spanning tree • nnds mean :- mean distance between nearest neighbours • mst dists mean :- mean distance of the minimum spanning tree We refer the reader to [11] for a detailed explanation for each feature. We carry out our diversity optimization approach for these features and use the evolutionary algorithm to evolve for each feature a diverse population of instances that meets the approximation criteria for hard/easy instances given by the approximation ratio thresholds. All programs in our experiments are written in R and run in R environment [16]. We use the functions in tspmeta package to compute the feature values [11]. The setting of the evolutionary algorithm for diversity optimization used in our experiments is as follows. We use µ = 30 and λ = 5 for the parent and offspring population size, respectively. The 2-OPT algorithm is executed on each instance I five times with different initial solutions and we set A(I) to the average tour length obtained. The examined instance sizes n are 25, 50 and 100, which are denoted by the number of cities in one instance. Based on previous investigations in [11] and initial experimental investigations, we set αe = 1 for instances of size 25 and 50, and αe = 1.03 for instances of size 100. Evolving hard instances, we use αh = 1.15, 1.18, 1.2 for instances of size n = 25, 50, 100, respectively. The mutation operator picks in each step one city for the 5 Figure 1: (left)The boxplots for centroid mean distance to centroid feature values of a population consisting of 100 different hard or easy TSP instances of different number of cities without or with diversity mechnism.(right)The boxplots for cluster 10% distance distance to centroid feature values of a population consisting of 100 different hard or easy TSP instances of different number of cities without or with diversity mechnism. Easy and hard instances from conventional approach and diversity optimization are indicated by e(a), h(a) and e(b), h(b) respectively. given parent uniformly at random and changes its x- and y-coordinator by choosing an offset according to the Normal-distribution with standard deviation σ . Coordinates that are out of the interval are reset to the value of the parent. Based on initial experiments we use two mutation operators with different values of σ . We use σ = 0.025 with probability 0.9 and σ = 0.05 with probability 0.1 in a mutation step. The evolutionary algorithm terminates after 10, 000 generations which allows to obtain a good diversity for the considered features. For each n = 25, 50, 100 and each of the 7 features, a set of easy and hard instances are generated, which results in 42 independent runs of the (µ+λ )-EAD . 4 Range of Feature Values We first evaluate our diversity optimization approach in terms of the diversity that is obtained with respect to a single feature. Focusing on a single feature in each run provides the insight of the possible range of a certain feature value for hard or easy instances. The previous study [11], suggests that there are some differences in the possible range of feature values for easy and hard instances. We study the effect of the diversity optimization on the range of features by comparing the instances generated by diversity optimization to the instances generated by the conventional approach in [11]. Evolving hard instances based on the conventional evolutionary algorithm, the obtained instances have mean approximation ratios of 1.12 for n = 25, 1.16 for n = 50, and 1.18 for n = 100. For easy instances, the mean approximation ratios are 1 for n = 25, 50 and 1.03 for n = 100. Figure 1 (left) presents the variation of the mean distance of the distances between points and the centroid feature (centroid mean distance to centroid) for hard and easy instances of the three considered sizes 25, 50 and 100. Each set consists of 100 instances generated by independent runs [11]. As shown in Figure 1 (left) the hard in6 stances have higher feature values than for easy instances for all instance sizes. For example, for instance size 100 and for the hard instances the median value (indicated by the red line) is 0.4157 while its only 0.0.4032 for the easy instances. The respective range of the feature value is 0.0577 for the hard instances and 0.0645 for the easy instances. For the instances generated by diversity optimization (easy and hard instances are indicated by e (b) and h (b) respectively), there is a difference in the median feature values for the hard and easy instances similar to the instances generated by the conventional approach. Additionally, the range of the feature values for both the hard and easy instances has significantly increased. For example, for the instance size 100, the median value for easy instances is 0.4028 and the range is 0.2382. For the hard instances of the same size, the median is 0.04157 while the range is 0.1917 (see Figure 1 (left)). Similarly, Figure 1 (right) presents the variation of cluster 10% distance to centroid (cluster 10pct distance to centroid) feature for the hard and easy instances generated by the conventional approach (indicated by (e (a) and h (a)) and for the hard and easy instances generated by diversity optimization (indicated by (e (b) and h (b))). The general observations from these box plots are quite similar to the observations from the mst dist mean shown in Figure 1 (left). For the easy instances of size 100, the range of the feature value is 0.0919 for conventional instances and 0.3471 for the instances generated by diversity optimization. Similarly, for the hard instances the range of the feature values has increased from 0.0577 to 0.1776 by the diversity optimization approach. As shown in Figure 1 (right), there is a significant increase in the range for other instance sizes as well. Improved ranges in feature values are observed for all considered features however, due to space limitations these are not included in the paper. The above results suggest that the diversity optimization approach has resulted in a significant increase in the coverage over the feature space. Having the threshold for approximation ratios (αe and αh ) our method guarantees the hardness of the instances. These approximation thresholds are more extreme than the mean approximation values obtained by the conventional method. Furthermore, starting with initial population of duplicated instances and a hard coded threshold, the modified (µ + λ )-EA is able to achieve hard instances with approximation ratio 1.347, 1.493 and 1.259, respectively for instance size 25, 50 and 100. The majority of the instances are clustered in a small region in the feature space while some other points are dispersed across the whole space. This is evident in the median values similar to the values for the instances with respect to conventional approach and with significantly larger range in feature value. The conventional approach has failed to explore certain regions in the feature space and missed some instances existing in those regions. Being able to discover all these instances spread in the whole feature space, our approach provides a strong basis for more effective feature based prediction. As a result of the increased ranges and the similar gap in median feature values for hard and easy instances compared to the conventional instances, there is a strong overlap in the ranges of the features for easy and hard instances generated by the diversity optimization. This is observed in the results for mst dist mean and cluster 10pct distance to centroid shown in Figure 1. Similar pattern holds for the other features as well. This prevents a good classification of problem instances based on single feature value. 7 Figure 2: 2D Plots of feature combinations which provide a separation between easy and hard instances. The blue dots and orange dots represent hard and easy instances respectively. 5 Classification Based on Multiple Features As a single feature is not capable in clearly classifying the hard/easy instances, combinations of two or three different features are examined in the following. Our analysis mainly focuses on combinations of the 7 previously introduced features. 5.1 Diversity Maximization over Single Feature Value Firstly, we represent the instances according to the combination of two different features in the 2-dimensional feature value space (see Figure 2 for an example). According to the observation and discussion in [11], the two features distance max and angle mean can be considered together to provide an accurate classification of the hard and easy instances. Whereas after increasing the diversity over the seven different feature values and a wider coverage of the 2D space is achieved, the separation of easy and hard instances is not so obvious. The clusters of dots representing hard and easy instances have some overlapping as shown in the left graphs of Figure 2. There are large overlapping areas lying between the two groups of instances. Another example of some separation given by two-feature combination is mst dists mean and chull area which measure the mean distance of the minimum spanning tree and the area of the convex hull. However, as the number of cities in an instance increases, the overlapping area becomes larger. It is hard to do classification based on this. 8 Figure 3: 2D Plots of feature combinations which does not provide a clear separation between easy and hard instances. The blue dots and orange dots represent hard and easy instances respectively. After examining the 21 different combinations of two features out of the seven features, we found out that some combinations of two features provide a fair separation between hard and easy instances after increasing the diversity over different feature values. As shown in Figure 2, taking both mst dists mean and chull area features into consideration, some separations can be spotted between hard and easy instances. However, most combinations are not able to give a clear classification between hard and easy instances, for example in Figure 3, neither the combination of features nnds mean and centroid mean distance to centroid nor features mst depth mean and chull area shows clear classification between instances of different hardness. Moreover, along with the instance size increment, the overlapping area of the dots standing for hard and easy instances grows. Since the majority of two-feature combinations are not capable of classifying easy and hard instances, the idea of combining three different feature is put forward. As in the analysis of two-feature combination, the values of the three selected features are plotted in 3D space. By considering a third feature in the combination, in the 35 different combinations, it is clear that there are some separations between the two groups of 210 instances from the 3D-plots. A good selection of features results in an accurate classification of the instances. The three-feature combinations with the features measuring statistics about minimum spanning tree always provide good separation between hard and easy instances as shown in Figure 4 and Figure 5. Although there is an overlapping in the 9 area between the two clusters of hard and easy instances, from the 3D-plots, we can spot some areas where there are only dots for instances of certain hardness. Taken another feature value into consideration, the two-feature combination that is not able to provide good separation can give some clear classification in hard and easy instances. An example illustrating this is included as Figure 6, where together with an additional feature mst dists mean, the two-feature combination of features mst depth mean and chull area shows a clear separation between easy and hard instances comparing to the results shown in the left graph in Figure 3. From the investigation of both the two-feature combination and three-feature combination, we found out that the range of feature values for larger TSP instances is smaller. Some of the good combinations for classifying the hardness of smaller instances may not work for larger instances, such as centroid features which performs well when combining with another feature in classifying the hardness of instances of 25 cities while does not show a clear separation with instance size 50 and 100 in our study. However, there exist some three-feature combinations that give good classification of easy and hard instances without regarding to the instance size, for example mst dists mean, chull area and nnds mean, and mst dists mean, chull area and mst depth mean. 5.2 Diversity Maximization over Multiple Feature Values In order to examine the relationship between feature combination and hardness of the instances, a weighted population diversity based on multiple features is introduced. The weighted population diversity for a certain set of features { f1 , f2 , ..., fk } is defined as the weighted sum of the normalised population diversity over these k features. The contribution of an instance I to the weighted population diversity is defined as k d 0 (I, P) = ∑ (wi · d fi (I, P)), i=1 where d fi (I, P) denotes the normalised contribution to the population diversity d(I, P) over certain feature i and wi represents the weight of feature i. The contribution of an individual to the population diversity on certain feature is normalised based on the maximum population diversity on the feature, in order to reduce the bias among different features. This weighted population diversity is used in Algorithm 1 to gain some insight of the relationship between features combination and instance quality. The same parent and offspring population sizes are used for these experiments, which are µ = 30 and λ = 5. The instance sizes examined are still 25, 50 and 100. The 2-OPT algorithm is executed five times to obtain the approximation quality. In the experiments, (µ+λ )-EAD execute for 10, 000 generation as previous. Since it is shown in Section 4 that a combination of three features is able to provide a good separation between hard and easy instances, some of the good three-feature combinations are chosen for exploration. The weight distributions for { f1 , f2 , f3 } considered in the experiments are {1, 1, 1}, {2, 1, 1}, {1, 2, 1}, {1, 1, 2}, {2, 2, 1}, {2, 1, 2}, {1, 2, 2}. The same hardness thresholds are used in these experiments as previous. After the seven independent runs 10 Figure 4: 3D Plots of combining experiment results from maximizing the diversity over features mst dists mean, nnds mean and chull area, which provides a good separation of easy and hard instances. Hard and easy instances are represented as blue dots and orange dots respectively. 11 Figure 5: 3D Plots of combining experiment results from maximizing the diversity over features mst dists mean, chull area and centroid mean distance to centroid, which provides a good separation of easy and hard instances. Legend is the same as that in Figure 4. 12 Figure 6: 3D Plots of combining experiment results from maximizing the diversity over features mst dists mean, chull area and mst depth mean, which provides a good separation of easy and hard instances. Legend is the same as that in Figure 4. 13 Figure 7: 3D Plots of combining experiment results from maximizing the diversity over features mst dists mean, nnds mean and chull area with considering of weighting, which provides a good separation of easy and hard instances. Legend is the same as that in Figure 4. 14 for easy and hard instances, the final solution sets are put together. Therefore the results set has 210 instances for each instance size and hardness type, which is the same as previous experiments. The results are plotted in 3D space and compared to the previous experiments on single feature discussed in Section 3 and 4. The weighted population diversity offers a way to examine the overlapping area of hard and easy instances. With the weighting technique, it takes consideration about the relationship between the different features examined. Since most of these features are not independent from each others and the weighted population diversity considers multiple features at the same time, it is predictable that with the weighted population diversity the extreme value for each single feature may not reach. An example is shown in Figure 7 focusing on maximizing the weighted population diversity over the combination of features mst dists mean, nnds mean and chull area, which is shown to be a good combination for separating the hard and easy instances. From the comparison between Figure 4 and Figure 7, we can see that although the results from maximizing weighted population diversity does not cover a wider search space, it provides a detailed insight into the intersection between the hard and easy instances. The 3D plots of different instance sizes show that the combination of these three certain features provide a clear separation between hard and easy instances. There are some overlapping areas in the search space, but it is clear that this combination of features provide some hints for predicting of hard or easy instances. 6 Instances Classification Using Support Vector Machine Support vector machines (SVMs) are well-known supervised learning models in machine learning which can be used for classification, regression and outliers detection [3, 8]. In order to quantify the separation between instances of different hardness based on the feature values, SVM models are constructed for each combination of features. 6.1 Linear SVM The linear classifier is the first model tried in classifying the dataset. In SVM the linear classifiers that can separate the data with maximum margin is termed as the optimal separating hyper-plane. From the plots in Figure 2, 3, 4 and 5, it is clear that none of the datasets are linearly separable. Taken the trade-off between maximizing the margin and minimizing the number of misclassified data points into consideration, the soft-margin SVM is used for classification. Let ACCn be the training accuracy of a feature combination in separating the hard and easy instances of size n. We define ACCn as the ratio of number of instances which are correctly classified by the model to the total number of instances in the dataset. All classification experiments are done in R with library{e1071} [12]. The training data of the SVM models are the population of 420 instances generated as in Section 3 and the training accuracy is regarded as a quantified measurement of the separation between hard and easy instances. The feature combinations used for classification are the 21 two-feature combinations and 35 three-feature combinations discussed in Section 5. 15 From experiment results, ACC25 for two-feature combinations lie in the range of 0.5095 to 0.7548 with an average accuracy of 0.6672, while the ACC25 for three-feature combination lie between 0.6286 to 0.7786 with average value 0.7079. In the case of instances with city number of 50, two-feature combination results in ACC50 lying in the range of 0.5286 to 0.7738 with an average of 0.6544 while ACC50 of three-feature combinations are from 0.5381 to 0.85 with average accuracy equal to 0.6969. For larger instance size, ACC100 are in the range between 0.5738 and 0.8119 with average 0.6986 for two-feature combination, whereas those for three-feature combination lie in the scope of 0.6238 to 0.8524 with average 0.7382. Although three-feature combinations show better accuracy in separation of hard and easy instances than those two-feature combinations, there is no significant difference in ACC for two-feature combinations and three-feature combinations. Moreover, the general low accuracy implies the high possibility that the linear models are not suitable for separating the hard and easy instances based on most of the feature combinations. We then move to applying kernel function for non-linear mapping of the feature combination. 6.2 Nonlinear Classification with RBF Kernel The linearly non-separable features can become linearly separable after mapped to a higher dimension feature space. The Radial Basis Function (RBF) kernel is one of the well-known kernel function used in SVM classification. There are two parameters need to be selected when applying RBF, which are C(cost) and γ. The parameter setting for RBF is crucial, since increasing C and γ leads to accurate separation of the training data but at the same time causes over-fitting. The SVMs here are generated for quantifying the separation rate between hard and easy instances rather than classifying other instances. After some initial trials, (C, γ) is set to (100, 2) in all the tests to avoid over-fitting. This parameter setting may not be the best parameters for the certain feature combination in SVM classifying, but it helps us to gain some understanding of the separation of hard and easy instances generated from previous experiments based on the same condition. Table 1 and 2 show the accuracy of different two features or three features combination in hard and easy instances separation. With RBF kernel, SVM with certain parameter setting can generate a model separating the dataset with average accuracy of 0.8170, 0.8244 and 0.8346 in 2D feature space for instance size 25, 50 and 100 respectively. Whereas with three features, SVM with the same parameter setting provides a separation with average accuracy of 0.9503, 0.9584 and 0.9422 for instance size 25, 50 and 100 respectively. From the results, it can be concluded that there are better separations between hard and easy instances in the 3D feature space. 16 Feature 1 angle mean angle mean angle mean angle mean angle mean angle mean centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid chull area chull area chull area chull area cluster 10pct mean distance to centroid cluster 10pct mean distance to centroid cluster 10pct mean distance to centroid mst depth mean mst depth mean nnds mean Feature 2 centroid mean distance to centroid chull area cluster 10pct mean distance to centroid mst depth mean nnds mean mst dists mean chull area cluster 10pct mean distance to centroid mst depth mean nnds mean mst dists mean cluster 10pct mean distance to centroid mst depth mean nnds mean mst dists mean mst depth mean nnds mean mst dists mean nnds mean mst dists mean mst dists mean ACC25 0.8476 0.7857 0.7810 0.7524 0.8167 0.8119 0.8619 0.8524 0.8381 0.8786 0.8905 0.8000 0.7429 0.8071 0.8619 0.7619 0.8190 0.8095 0.7786 0.8095 0.8500 ACC50 0.8071 0.7810 0.7786 0.7381 0.8833 0.8024 0.7667 0.8357 0.7643 0.9524 0.8571 0.7881 0.7429 0.8905 0.8643 0.7714 0.8833 0.8095 0.8595 0.8214 0.9143 ACC100 0.8071 0.7929 0.8000 0.8000 0.8452 0.8405 0.8381 0.7548 0.8095 0.8476 0.8762 0.8548 0.7571 0.8452 0.9024 0.7929 0.8643 0.8738 0.8405 0.8810 0.9024 Table 1: This table lists the accuracy of SVM with RBF kernel separating the hard and easy instances in 21 different two-feature space. Feature 1 angle mean angle mean angle mean angle mean angle mean angle mean angle mean angle mean angle mean angle mean angle mean angle mean angle mean angle mean angle mean centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid chull area chull area chull area chull area chull area chull area cluster 10pct mean distance to centroid cluster 10pct mean distance to centroid cluster 10pct mean distance to centroid mst depth mean Feature 2 centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid centroid mean distance to centroid chull area chull area chull area chull area cluster 10pct mean distance to centroid cluster 10pct mean distance to centroid cluster 10pct mean distance to centroid mst depth mean mst depth mean nnds mean chull area chull area chull area chull area cluster 10pct mean distance to centroid cluster 10pct mean distance to centroid cluster 10pct mean distance to centroid mst depth mean mst depth mean nnds mean cluster 10pct mean distance to centroid cluster 10pct mean distance to centroid cluster 10pct mean distance to centroid mst depth mean mst depth mean nnds mean mst depth mean mst depth mean nnds mean nnds mean Feature 3 chull area cluster 10pct mean distance to centroid mst depth mean nnds mean mst dists mean cluster 10pct mean distance to centroid mst depth mean nnds mean mst dists mean mst depth mean nnds mean mst dists mean nnds mean mst dists mean mst dists mean cluster 10pct mean distance to centroid mst depth mean nnds mean mst dists mean mst depth mean nnds mean mst dists mean nnds mean mst dists mean mst dists mean mst depth mean nnds mean mst dists mean nnds mean mst dists mean mst dists mean nnds mean mst dists mean mst dists mean mst dists mean ACC25 0.9500 0.9405 0.9548 0.9452 0.9571 0.9524 0.9357 0.9405 0.9667 0.9214 0.9476 0.9571 0.9310 0.9476 0.9429 0.9476 0.9595 0.9667 0.9714 0.9476 0.9643 0.9500 0.9500 0.9548 0.9667 0.9286 0.9524 0.9595 0.9381 0.9476 0.9714 0.9214 0.9500 0.9643 0.9429 ACC50 0.9190 0.9357 0.9548 0.9952 0.9500 0.9310 0.9238 0.9714 0.9619 0.9143 0.9690 0.9143 0.9762 0.9262 0.9762 0.9333 0.8762 0.9881 0.9714 0.9286 0.9905 0.9595 0.9881 0.9548 1.0000 0.9524 0.9667 0.9595 0.9857 0.9738 0.9857 0.9857 0.9476 0.9833 0.9929 ACC100 0.9452 0.8214 0.9214 0.9833 0.9524 0.8881 0.9500 0.9571 0.9143 0.9810 0.9333 0.9405 0.9238 0.9476 0.8833 0.9310 0.9762 0.9929 0.8381 0.8571 0.8810 0.9190 0.9595 0.9595 0.9952 0.9333 0.9667 0.9929 0.9476 0.9833 0.9667 0.9738 0.9643 0.9976 0.9929 Table 2: This table lists the accuracy of SVM with RBF kernel separating the hard and easy instances in 35 different three-feature space. 17 7 Conclusions With this paper, we have introduced a new methodology of evolving easy/hard instances which are diverse with respect to feature sets of the optimization problem at hand. Using our diversity optimization approach we have shown that the easy and hard instances obtained by our approach covers a much wider range in the feature space than previous methods. The diversity optimization approach provides instances which are diverse with respect to the investigated features. The proposed population diversity measurements provide good evaluation of the diverse over single or multiple feature values. Our experimental investigations for 2-OPT and TSP have shown that our large set of diverse instances can be classified quite well into easy and hard instances when considering a suitable combination of multiple features which provide some guidance for predication as the next step. In particular, the SVM classification model built with the diverse instances that can classify TSP instances based on problem hardness provides a strong basis for future performance prediction models that lead to automatic algorithm selection and configuration. Building such models would require further experimentation to determine the minimal set of strong features that can predict performance accurately. References [1] D. Applegate, W. Cook, S. Dash, and A. Rohe. Solution of a min-max vehicle routing problem. INFORMS Journal on Computing, 14(2):132–143, Apr. 2002. [2] B. Bonet and S. Koenig, editors. Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, January 25-30, 2015, Austin, Texas, USA. AAAI Press, 2015. [3] C. Cortes and V. Vapnik. Support-vector networks. Machine Learning, 20(3):273– 297, 1995. [4] G. A. Croes. A method for solving traveling-salesman problems. Operations Research, 6(6):791–812, 1958. [5] K. Eggensperger, F. Hutter, H. H. Hoos, and K. Leyton-Brown. Efficient benchmarking of hyperparameter optimizers via surrogates. In Bonet and Koenig [2], pages 1114–1120. [6] M. Englert, H. Röglin, and B. Vöcking. Worst case and probabilistic analysis of the 2-opt algorithm for the TSP. Algorithmica, 68(1):190–264, 2014. [7] M. Feurer, J. T. Springenberg, and F. Hutter. Initializing bayesian hyperparameter optimization via meta-learning. In Bonet and Koenig [2], pages 1128–1135. [8] S. R. Gunn. Support vector machines for classification and regression, 1998. [9] F. Hutter, L. Xu, H. H. Hoos, and K. Leyton-Brown. Algorithm runtime prediction: Methods & evaluation. Artif. Intell., 206:79–111, Jan. 2014. 18 [10] T. Kötzing, F. Neumann, H. Röglin, and C. Witt. Theoretical analysis of two ACO approaches for the traveling salesman problem. Swarm Intelligence, 6(1):1– 21, 2012. [11] O. Mersmann, B. Bischl, H. Trautmann, M. Wagner, J. Bossek, and F. Neumann. A novel feature-based approach to characterize algorithm performance for the traveling salesperson problem. Annals of Mathematics and Artificial Intelligence, 69(2):151–182, 2013. [12] D. Meyer, E. Dimitriadou, K. Hornik, A. Weingessel, and F. Leisch. e1071: Misc Functions of the Department of Statistics, Probability Theory Group (Formerly: E1071), TU Wien, 2015. R package version 1.6-7. [13] S. Nallaperuma, M. Wagner, and F. Neumann. Parameter prediction based on features of evolved instances for ant colony optimization and the traveling salesperson problem. In PPSN XIII - 13th International Conference, Ljubljana, Slovenia, September 13-17, 2014. Proceedings, pages 100–109, 2014. [14] S. Nallaperuma, M. Wagner, F. Neumann, B. Bischl, O. Mersmann, and H. Trautmann. A feature-based comparison of local search and the christofides algorithm for the travelling salesperson problem. In FOGA ’13, pages 147–160, 2013. [15] F. Neumann and C. Witt. Bioinspired Computation in Combinatorial Optimization:Algorithms and Their Computational Complexity. Springer-Verlag New York, Inc., New York, NY, USA, 1st edition, 2010. [16] R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, 2015. [17] K. Smith-Miles and L. Lopes. Measuring instance difficulty for combinatorial optimization problems. Computers & OR, 39(5):875–889, 2012. [18] K. Smith-Miles, J. van Hemert, and X. Y. Lim. Understanding TSP difficulty by learning from evolved instances. In 4th International Conference on Learning and Intelligent Optimization (LION), LION’10, pages 266–280. Springer, 2010. [19] T. Ulrich, J. Bader, and L. Thiele. Defining and optimizing indicator-based diversity measures in multiobjective search. In R. Schaefer, C. Cotta, J. Kolodziej, and G. Rudolph, editors, PPSN (1), volume 6238 of Lecture Notes in Computer Science, pages 707–717. Springer, 2010. [20] T. Ulrich, J. Bader, and E. Zitzler. Integrating decision space diversity into hypervolume-based multiobjective search. In M. Pelikan and J. Branke, editors, GECCO, pages 455–462. ACM, 2010. [21] J. I. van Hemert. Evolving combinatorial problem instances that are difficult to solve. Evolutionary Computation, 14(4):433–462, Dec. 2006. [22] R. Vilalta and Y. Drissi. A perspective view and survey of meta-learning. Artificial Intelligence Review, 18(2):77–95, Oct. 2002. 19 [23] L. Xu, F. Hutter, H. H. Hoos, and K. Leyton-Brown. Satzilla: portfolio-based algorithm selection for sat. Journal of Artificial Intelligence Research, 32(1):565– 606, June 2008. 20	9cs.NE
Variants of Plane Diameter Completion∗† Petr A. Golovach‡ Clment Requil§ Dimitrios M. Thilikos¶ arXiv:1509.00757v1 [cs.DS] 2 Sep 2015 Abstract The Plane Diameter Completion problem asks, given a plane graph G and a positive integer d, if it is a spanning subgraph of a plane graph H that has diameter at most d. We examine two variants of this problem where the input comes with another parameter k. In the first variant, called BPDC, k upper bounds the total number of edges to be added and in the second, called BFPDC, k upper bounds the number of additional edges per face. We prove that both problems are NP-complete, the first even for 3-connected graphs of face-degree at most 4 and the second even when k = 1 on 3-connected graphs of face-degree at most 5. In this paper we give parameterized algorithms for both problems that run in O(n3 ) + 22 1 O((kd)2 log d) · n steps. Introduction In 1987, Chung [3, Problem 5] introduced the following problem: find the optimum way to add q edges to a given graph G so that the resulting graph has minimum diameter. (Notice that in all problems defined in this paper we can directly assume that G is a simple graph as loops do not contribute to the diameter of a graph and the same holds if we take simple edges instead of multiple ones.) This problem was proved to be NP-hard if the aim is to obtain a graph of diameter at most 3 [19], and later the NP-hardness was shown even for the Diameter-2 Completion problem [14]. It is also know that Diameter-2 Completion is W[2]-hard when parameterized by q [9]. For planar graphs, Dejter and Fellows introduced in [5] the Planar Diameter Completion problem that asks whether it is possible to obtain a planar graph of ∗ The first author was supported by the European Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement n. 267959. The second author was supported by the FP7-PEOPLE-2013-CIG project CountGraph (ref. 630749), the collateral PROCOPE-DAAD project RanConGraph (ref. 57134837), and the Berlin Mathematical School. The research of the third author was co-financed by the European Union (European Social Fund ESF) and Greek national funds through the Operational Program “Education and Lifelong Learning” of the National Strategic Reference Framework (NSRF), Research Funding Program: ARISTEIA II. † Emails: Petr.Golovach@ii.uib.no, requile@math.fu-berlin.de, sedthilk@thilikos.info. ‡ Department of Informatics, University of Bergen, Bergen, Norway. § Freie Universität Berlin, Institut für Mathematik und Informatik, Berlin, Germany. ¶ AlGCo project team, CNRS, LIRMM, France, Department of Mathematics, University of Athens, Athens, Greece, and Computer Technology Institute & Press “Diophantus”, Patras, Greece. 1 diameter at most d from a given planar graph by edge additions. It is not known whether Planar Diameter Completion admits a polynomial time algorithm, but Dejter and Fellows showed that, when parameterized by d, Planar Diameter Completion is fixed parameter tractable [5]. The proof is based on the fact that the yes-instances of the problem are closed under taking minors. Because of the Robertson and Seymour theorem [18] and the algorithm in [16], this implies that, for each d, the set of graphs G for which (G, d) is a yes-instance can be characterized by a finite set of forbidden minors. This fact, along with the minor-checking algorithm in [17] implies that there exists an O(f (d) · n3 )-step algorithm (i.e. an FPT-algorithm) deciding whether a plane graph G has a plane completion of diameter at most d. Using the parameterized complexity, this means that Planar Diameter Completion is FPT, when parameterized by d. To make this result constructive, one requires the set of forbidden minors for each d, which is unknown. To find a constructive FPT-algorithm for this parameterized problem remains a major open problem in parameterized algorithm design. Our results. We denote by S0 the 3-dimensional sphere. By a plane graph G we mean a simple planar graph G with the vertex set V (G) and the edge set E(G) drawn in S0 such that no two edges of this embedding intersect. A plane graph H is a a plane completion (or, simply completion) of another plane graph G if H is a spanning subgraph of G. A q-edge completion of a plane graph G is a completion H of G where \|E(H)\| − \|E(G)\| ≤ q. A k-face completion of a plane graph G is a completion H of G where at most k edges are added in each face of G. In this paper we consider the variants of the Plane Diameter Completion problem: Plane Diameter Completion (PDC) Input: a plane graph G and d ∈ N≥1 . Output: is there a completion of G with diameter at most d? Notice that the important difference between PDC and the aforementioned problems is that we consider plane graphs, i.e., the aim is to reduce the diameter of a given embedding of a planar graph preserving the embedding. In particular we are interested in the following variants: Bounded Budget PDC (BPDC) Input: a plane graph G and q ∈ N, d ∈ N≥1 Question: is there a completion H of G of diameter at most d that is also a q-edge completion? Bounded Budget/Face PDC (BFPDC) Input: a plane graph G and k ∈ N, d ∈ N≥1 . Question: is there a completion H of G of diameter at most d that is also a k-face completion? We examine the complexity of the two above problems. Our hardness results are the following. 2 Theorem 1. Both BPDC and BFPDC are NP-complete. Moreover, BPDC is NP-complete even for 3-connected graphs of face-degree at most 4, and BFPDC is NP-complete even for k = 1 on 3-connected graphs of face-degree at most 5. The hardness results are proved in Section 6 using a series of reductions departing from the Planar 3-Satisfiability problem that was shown to be NP-hard by Lichtenstein in [15]. The results of Theorem 1 prompt us to examine the parameterized complexity1 of the above problems. For this, we consider the following general problem: Bounded Budget and Budget/Face BDC (BBFPDC) Input: a plane graph G, q ∈ N ∪ {∞}, k ∈ N, and d ∈ N≥1 . Question: is there a completion H of G of diameter at most d that is also a q-edge completion and a k-face completion? Notice that when q = ∞ BBFPDC yields BFPDC and when q = k BBFPDC yields BPDC. Our main result is that BBFPDC is fixed parameter tractable (belongs in the parameterized class FPT) when parameterized by k and d. O((kd) log d) Theorem 2. It is possible to construct an O(n3 )+22 for BBFPDC. ·(α(q))2 ·n-step algorithm (In the above statement and in the rest of this paper we use the function α : N ∪ {∞} → N such that if q = ∞, then α(q) = 1, otherwise α(q) = q.) The main ideas of the algorithm of Theorem 2 are the following. We first observe that yes-instances of PDC and all its variants have bounded branchwidth (for the definition of branchwidth, see Section 2). The typical approach in this case is to derive an FPT-algorithm by either expressing the problem in Monadic Second Order Logic – MSOL (using Courcelle’s theorem [4]) or to design a dynamic programming algorithm for this problem. However, for completion problems, this is not really plausible as this logic can quantify on existing edges or vertices of the graph and not on the “non-existing” completion edges. This also indicates that to design a dynamic programming algorithm for such problems is, in general, not an easy task. In this paper we show how to tackle this problem for BBFPDC (and its special cases BPDC and BFPDC). Our approach is to deal with the input G as a part of a more complicated graph with O(k 2 · n) additional edges, namely its cylindrical enhancement G0 (see Section 3 for the definition). Informally, sufficiently large cylindrical grids are placed inside the faces of G and then internally vertex disjoint paths in these grids can be used to emulate the edges of a solution of the original problem placed inside the corresponding faces. Thus, by the enhancement we reduce BBFPDC to a new problem on G0 certified by a suitable 3-partition of the additional edges. Roughly, this partition consists of the 1-weighted edges that should be added in the completion, the 0-weighted edges that should link these edges to the boundary of the face of G where they will be 1 For more on parameterized complexity, we refer the reader to [8]. 3 inserted, and the ∞-weighted edges that will be the (useless) rest of the additional edges. The new problem asks for such a partition that simulates a bounded diameter completion. The good news is that, as long as the number of edges per face to be added is bounded, which is the case for BBFPDC, the new graph G0 has still bounded branchwidth and it is possible, in the new instance, to quantify this 3-partition of the graph G0 . However, even under these circumstances, to express the new problem in Monadic Second Order Logic is not easy. For these reasons we decided to follow the more technical approach of designing a dynamic programming algorithm that leads to the (better) complexity bounds of Theorem 2. This algorithm is quite involved due to the technicalities of the translation of the BBFPDC to the new problem. It runs on a sphere-cut decomposition of the plane embedding of G0 and its tables encode how a partial solution is behaving inside a closed disk whose boundary meets only (a few of) the edges of G0 . We stress that this encoding takes into account the topological embedding and not just the combinatorial structure of G0 . Sphere-cut decompositions as well as some necessary combinatorial structures for this encoding are presented in Section 4. The dynamic programming algorithms is presented in Section 5 and is the most technical part of this paper. 2 Definitions and preliminaries Given a graph G, we denote by V (G) (respectively E(G)) the set of vertices (respectively edges) of G. A graph G0 is a subgraph of a graph G if V (G0 ) ⊆ V (G) and E(G0 ) ⊆ E(G), and we denote this by G0 ⊆ G. Also, in case V (G) = V (G0 ), we say that H is a spanning subgraph of G. If S is a set of vertices or a set of edges of a graph G, the graph G \ S is the graph obtained from G after the removal of the elements of S. If S is a set of edges, we define G[E] as the graph whose vertex set consists of the endpoints of the edges of E and whose edge set of E. Distance and diameter. Let G be a graph and let w : E(G) → N ∪ {∞} (w is a weighting of the edges of G). Given two vertices x, x0 ∈ V (G) we call (x, x0 )-path every path of G with x and x0 as endpoints. We also define w-distG (x, x0 ) = min{w(E(P )) \| P is an (x, x0 )-path in G} and w-diam(G) = max{w-distG (x, y) \| x, y ∈ V (G)} (if G is not connected then w-diam(G) is infinite). When the graph is unweighted then we use distG and diam instead of w-distG and w-diam. Plane graphs. To simplify notations on plane graphs, we consider a plane graph G as the union of the points of S0 in its embedding corresponding to its vertices and edges. That way, a subgraph H of G can be seen as a graph H where H ⊆ G. The faces of a plane graph G, are the connected components of the set S0 \ G. A vertex v (an edge e resp.) of a plane graph G is incident to a face f and, vice-versa, f is incident to v (resp. e) if v (resp., e) lies on the boundary of f . Two faces f1 , f2 are adjacent if they have a common incident edges. We denote by F (G) the set of all faces of G. The degree of a face f ∈ F (G) is the number of edges incident to f where bridges of G 4 count double in this number. The face-degree of G is the maximum degree of a face in F (G). Given a face f of G, we define BG (f ) as the graph whose set of points is the boundary of f and whose vertices are the vertices incident to f . A set ∆ ⊆ S0 is an open disc if it is homeomorphic to {(x, y) : x2 + y 2 < 1}. Also, ∆ is a closed disk of S0 if it is the closure of some open disk of S0 . Branch decomposition. Given a graph H with n vertices, a branch decomposition of H is a pair (T, µ), where T is a tree with all internal vertices of degree three and µ : L → E(H) is a bijection from the set of leaves of T to the edges of H. For every edge e of T , we define the middle set mid(e) ⊆ V (H) as follows: if T \ {e} has two connected components T1 and T2 , and for i ∈ {1, 2}, let Hie = H[{µ(f ) : f ∈ L ∩ V (Ti )], and set mid(e) = V (H1e ) ∩ V (H2e ). The width of (T, µ) is the maximum order of the middle sets over all edges of T , i.e. max{\|mid(e)\| : e ∈ T }. The branchwidth of H is the minimum width of a branch decomposition of H and is denoted by bw(H). A grid annulus Γk,h is the graph obtained by the cartesian product of a cycle of k vertices and a path of h vertices. We need the following result. Proposition 1 ([11]). Let G be a planar graph and k, h be integers with k ≥ 3 and h ≥ 1. Then G has either a minor isomorphic to Γk,h or a branch decomposition of width at most k + 2h − 2. An central feature of the PDC problem and its variants is that its yes-instances have bounded branchwidth. Lemma 1. There exists a constant c1 such that if (G, d) is a yes-instance of PDC, then bw(G) ≤ c1 · d. The same holds for the graphs in the yes-instances of BPDC, BFPDC, and BBFPDC. Proof. We examine only the case of PDC as a yes-instance of BPDC, BFPDC, and BBFPDC is also a yes-instance of PDC. Notice first that if G has a completion of diameter at most d and G0 is a minor2 of some G, then also G0 has a completion H of diameter at most d. Notice also that every completion of the grid annulus Γr+2,r+2 has diameter > r, therefore, if (G, d) is a yes-instance of PDC, then G cannot contain a Γr+2,r+2 as a minor. From Proposition 1, G has branchwidth bounded by a linear function of d and the lemma follows. 2A graph G0 is a minor of a graph G if it can obtained applying edge contractions to some subgraph of G. 5 3 3.1 The reduction cylindrical enhancements Grid-annulus. Let k and r be positive integers where k ∈ N≥3 , r ∈ N≥3 . We define the graph Γk,r as the (k × r)-grid annulus, which is the Cartesian product of a path of k vertices and a cycle of r vertices. Notice that Γk,r is uniquely embeddable (up to homeomorphism) in the plane and has exactly two non-square faces (i.e., faces incident to 4 edges) f1 and f2 that are incident only with vertices of degree 3. We call one of the faces f1 and f2 the interior of Γk,r and the other the exterior of Γk,r . We call the vertices incident to the interior (exterior) of Γk,r base (roof ) of Γk,r . Given an edge e in the base of Γk,r , we define its ceilings as the set of edges of Γk,r that contains e and whose dual edges in Γ∗k,r form a minimum length path between the duals of the interior and the exterior face of Γ∗k,r . Cylindrical enhancement of a plane graph. Let G be a plane graph. We next give the definition of the graph G(k) for k ∈ N≥3 . Let fi ∈ F (G) and let C1i , . . . , Cρi i be the connected components of BG (fi ). For each Cji , we denote by σji the number of its edges, agreeing that, in this number, bridge edges count twice and that if Cji consists of only one vertex, then σji = 1. We then add a copy Γij of Γk,k·σji in the embedding of G i i such that Cji is contained in the interior of Γij and all C1i , . . . , Cj−1 , . . . , Cj+1 . . . , Cρi i are contained in the exterior of Γij (In Figure 1 the edges of each Γij are colored red). We then add, for each v ∈ Cji , κ(v) · k edges (those around the disks C1 , . . . , C4 in Figure 1) from v to the base of Γij , where κ(v) is the number of connected components in Cji \ v (in the trivial case where Cji consists of only one vertex v, then κ(v) = 1). We add these edges in a way that the resulting embedding remains plane and no more than a set Vv,i,j of k consecutive vertices of the base of Cji are connected with the same vertex v of Cji ; observe that there is only one way to add edges so to fulfill these restrictions. Notice that the set Vv,i,j always induces a path Pv,i,j in the resulting graph except in the case where Cji consists of a single vertex v where Vv,i,j induces a cycle. In the later case we pick a maximal path in this cycle and we denote it by Pv,i,j . In the example of Figure 1 the Pv,i,j ’s are the bold paths of the innermost cycle of each Γij . We apply this enhancement for each connected component of the boundary of each (k) face of G and we denote the resulting graph by RG . (k) We call a face fi of RG non-trivial if BR(k) (fi ) has more than one connected G components C1i , . . . , Cρi i . Notice that if fi is non-trivial, each Cji is the roof of some previously added grid-annulus. For each such grid-annulus, let Jj be k consecutive vertices of its roof. We add inside fi a copy of Γk,k·ρi such that its base is a subset of fi and let {I1 , . . . , Iρi } be a partition of its roof in ρi parts, each consisting of k consecutive base vertices. In the example of Figure 1, the annulus Γk,k·ρi is the one with the edges in the middle of the figure and its base is its innermost cycle. For each j ∈ {1, . . . , ri } we add k edges (depicted as the “interconnecting” edges in Figure 1) each connecting a vertex of Jj with some vertex of Ij in a way that the resulting 6 C4 C3 C1 C2 Figure 1: An example of a cylindrical enhancement for k = 3 inside a non-trivial face of a graph with 4 connected components (i.e., the boundaries of the disks C1 , . . . , C4 ). embedding remains plane (again, there is a unique way for this to be done). We apply (k) this enhancement for each non-trivial face of RG and we denote the resulting graph by G(k) . Notice that G(k) is not uniquely defined as its definition depends on the choice of the sets Jj . From now on, we always consider an arbitrary choice for G(k) and we call G(k) the k-th cylindrical enhancement of G. By the construction of G(k) , it directly follows that \|V (G(k) )\| = O(k 2 · n). We say that an edge of G(k) is an expansion edge if it is an edge of Pv,i,j for some i, j, and v ∈ V (Ci,j ). Also we denote by Ḡ(k) the graph created by G(k) if we contract all its expansion edges and all their ceilings of the (k) grid-annuli that were added during the construction of RG . Primal-dual drawings. Let G be a connected plane graph. We denote by D(G) the graph obtained if we draw G together with its dual so that dual edges are intersecting to a single point and then introduce a vertex to each of these intersection points. We recursively define D(k) (G) such that D(0) (G) = G and D(k) (G) = Dk−1 (D(G)) for every k ≥ 1. The next proposition is a direct consequence of [13, Lemma 4]. Proposition 2. There exists some constant c such that for every connected plane graph G, it holds that bw(D(G)) ≤ 2 · bw(G). 7 Corollary 1. For every connected plane graph G and k ∈ N≥1 , it holds that bw(D(k) (G)) ≤ 2k · bw(G). Lemma 2. If G is a connected plane graph and k ∈ N≥3 , then Ḡ(k) is a minor of D(dlog(k+1)e+1) (G). Proof. Notice first that Ḡ(3) is a minor of D(3) (G). It is then enough to observe that for every i ≥ 3, if Ḡ(i) is a minor of D(i) (G), then Ḡ(2i+1) is a minor of D(i+1) (G). The following lemma indicates that cylindrical enhancements do not considerably increase the branchwidth of a graph. Lemma 3. There is a constant c2 such that if G is an n-vertex plane graph and k ∈ N≥3 , then G(k) is 3-connected, bw(G(k) ) ≤ c2 · k · bw(G). Proof. Let H be the graph created from G if we add a vertex vf to each non trivial face f and for each of the connected components of BG (f ), we arbitrarily pick a vertex and make it adjacent to vf by a path of 2k internal vertices. As the branchwidth of a non-acyclic graph is the maximum branchwidth of its connected components, it follows that bw(H) = bw(G). It is also easy to see that G(k) is a minor of H̄ (k) . From Lemma 2, H̄ (k) is a minor of D(r) (H), where r = dlog(k + 1)e + 1. By Corollary 1, it follows that bw(D(r) (H)) ≤ 2r · bw(H) = O(k · bw(G)). 3.2 Edge colorings of new edges. Let G and H be two plane graphs such that G is a subgraph of H and let q ∈ N ∪ {∞}, k ∈ N, and d ∈ N≥1 . Given a 3-partition p = {E 0 , E 1 , E ∞ } of E(H) \ E(G), we define the function wp : E(H) → N such that wp = {(e, 1) \| e ∈ E(G)} ∪ {(e, 0) \| E ∈ E 0 } ∪ {(e, 1) \| e ∈ E 1 } ∪ {(e, d + 1) \| E ∈ E ∞ }. We say that G has (q, k, d)-extension in H if there is a 3-partition p = {E 0 , E 1 , E ∞ } of E(H) \ E(G) such that the following conditions hold A. There is no path in H with endpoints in V (G) that consists of edges in E 0 , B. every face F of G contains at most k edges of E 1 , C. ∀x, y ∈ V (G), wp -distH (x, y) ≤ d, and D. \|E 1 \| ≤ q. Given a 3-partition p = {E 0 , E 1 , E ∞ } of E(H) \ E(G) we refer to its elements as the 0-edges, the 1-edges, and the ∞-edges respectively. We also call the edges of G old-edges. Our first step towards our algorithm is to reduce BBFPDC to a problem about (q, k, d)-extensions of G. 8 Given a plane graph G and an open set Λ of S0 , we define GhΛi as the graph whose edge set consists of the edges of G that are subsets of Λ and whose vertex set consists of their endpoints. Disjoint paths. Let G be a graph. We say that two paths in G are disjoint if none of the internal vertices of a path is a vertex of the other. Given a collection P of pairwise disjoint paths of G, we define L(P) = {{x, y} \| x and y are the endpoints of a path in P}. The proofs of the following proposition can be found in [2]. Proposition 3. Let G be a graph k ∈ N≥1 and let H be a k-face completion of G. For every face f ∈ F (G), there is a collection P of k disjoint paths in the graph G(max{3,k}) hf i such that E(Ghf i) = L(P). Lemma 4. Let G be a plane graph, with q ∈ N ∪ {∞}, k ∈ N≥1 and d ∈ N≥1 . Then (G, q, k, d) is a yes-instance of BBFPDC if and only if G has a (q, k, d)-extension in G(max{3,k}) . Proof. Assume first that (G, q, k, d) is a yes-instance of BBFPDC and let H be a completion H of G of diameter at most d that is also a q-edge completion and a k-face completion. This means that for every f ∈ F (G), the graph Hf = Hhf i contains at S most k edges and that the graph H new = f ∈F (H) Hf contains at most q edges. From Proposition 3, there is a collection Pf of yf = \|E(Hf )\| internally disjoint paths in P G(max{3,k}) . Let E 1 be a set of y = f ∈F (G) yf edges obtained if, for every f ∈ F (G), S S we pick one edge from each of the paths in Pf . Let E 0 = E( f ∈F (G) P ∈Pf ) \ E 1 and let E ∞ = E(H new )\(E 0 ∪E 1 ). We now observe that p = {E 0 , E 1 , E ∞ } is a 3-partition of E(H new ) = E(G(max{3,k}) ) \ E(G). By its construction, p satisfies conditions 1–4 of the definition of a (q, k, d)-extension of G in G(max{3,k}) as required. Let now p = {E 0 , E 1 , E ∞ } is a 3-partition of E(H new ) = E(G(max{3,k}) ) \ E(G) that is a (q, k, d)-extension of G in G(max{3,k}) . We construct the graph H by removing from G(max{3,k}) all edges in E ∞ and then, in the resulting graph, contract all edges in E 0 . It is easy to observe that H is a completion of G that is also an q-edge completion and a k-face completion 4 Structures for dynamic programming For our dynamic programming algorithm we need a variant of branchwidth for plane graphs whose middle sets have additional topological properties. Sphere-cut decomposition. Let H be a plane graph. An arc is a subset O of the plane homeomorphic to a circle and is called a noose of H if it meets H only in vertices. We also set VO = V (H) ∩ O. An arc of a noose O is a connected component of O \ VO while in the trivial case where VO = ∅, O does not have arcs. A sphere-cut decomposition or sc-decomposition of H is a triple (T, µ, π) where (T, µ) is a branch decomposition of H and π is a function mapping each e ∈ E(T ) to cyclic orderings 9 of vertices of H, such that for every e ∈ E(T ) there is a noose Oe of H where the following properties are satisfied. • Oe meets every face of H at most once, • H1e is contained in one of the closed disks bounded by Oe and H2e is contained in the other (H1e and H2e are as in the definition of branch decomposition). • π(e) is a cyclic ordering of VOe defined by a clockwise traversal of Oe in the embedding of H. We denote Xe = VOe and we always assume that its vertices are clockwise enumerated according to π(e). We denote by Ae the set containing the arcs of Oe . Also, if π(e) = [a1 , . . . , ak , a1 ], then we use the notation Ae = {a1,2 , a2,3 , . . . , ak−1,k , ak,1 } where the boundary of the arc ai,i+1 consists of the vertices ai and ai+1 . We also define He+ = (V (H), E(H ∪ Ae )), i.e., He+ is the embedding occurring if we add in H the arcs of Oe as edges. A face of He+ is called internal if it is not incident to an arc in Ae , i.e., it is also a face of H. A face of He+ is marginal if it is a properly included is some face of H. For our dynamic programming we require to have in hand an optimal sphere-cut decomposition. This is done combining the main result of [10] and [20, (5.1)] (see also [7]) and is summarized to the following. Proposition 4. There exists an algorithm that, with input a 3-connected plane graph G and w ∈ N, outputs a sphere-cut decomposition of G of width at most w or reports that bw(G) > w. Our next step is to define a series of combinatorial structures that are necessary for our dynamic programming. Given two sets A and B we denote by AB the set of all functions from B to A. (d, k, q)-configurations. Given a set X and a non-negative integer t, we say that the pair (X , χ) is a t-labeled partition of X if X is a collection of pairwise disjoint non-empty subsets of X and χ is a function mapping the integers in {1, . . . , \|X \|} to integers in {0, . . . , t}. In case X = ∅, a t-labeled partition corresponds to the pair {∅, ∅} where ∅ is the “empty” function, i.e. the function whose domain is empty. Let X and A be two finite sets. Given d, k ∈ N and q ∈ N ∪ {∞}, we define a (d, k, q)-configuration of (X, A) as a quintuple ((X , χ), (A, α), (F, E), δ, z) where 1. (X , χ) is a 1-labeled partition of X, 2. (A, α) is a k-labeled partition of A, 3. (F, E) is a graph (possibly with loops) where F ⊆ {0, . . . , d + 1}X , 2 4. δ ∈ {0, . . . , d + 1}X , and 10 5. if q ∈ N, then z ≤ q, otherwise z = ∞. Fusions and restrictions. Let (X1 , χ1 ) and (X2 , χ2 ) be two t-labeled partitions of the sets X1 and X2 respectively such that Xi = {X1i , . . . , Xρi 1 }, i ∈ {1, 2}. We define X1 ⊕ X2 as follows: if x, x0 ∈ X1 ∪ X2 we say that x ∼ x0 if there is a set in X1 ∪ X2 that contains both of them. Let ∼T be the transitive closure of ∼. Then X1 ⊕X2 contains the equivalence classes of ∼T . We now define χ1 ⊕ χ2 as follows: let X1 ⊕ X2 = {Y1 , . . . , Yρ }. Then for each i ∈ {1, . . . , ρ}, we define P P χ1 ⊕ χ2 (i) = min{t, X 10 ⊆Yi χ1 (i0 ) + X 20 ⊆Yi χ2 (i0 )}. i i The fusion of the t-labeled partitions (X1 , χ1 ) and (X2 , χ2 ) is the pair (X1 ⊕ X2 , χ1 ⊕ χ2 ) that is a (t + 1)-labeled partition and is denoted by (X1 , χ1 ) ⊕ (X2 , χ2 ). Given a t-labeled partition (X , χ) of a set X and given a subset X 0 of X we define the restriction of (X , χ) to X 0 as the t-labeled partition (X 0 , χ0 ) of X 0 where X 0 = {Xi ∩ X 0 \| Xi ∈ X } \ {∅} and χ0 = {(i, χ(i)) \| Xi ∩ X 0 6= ∅} and we denote it by (X , χ)\|X 0 . We also define the intersection of (X , χ) with X 0 as the t-labeled partition (X 0 , χ0 ) where X 0 = {Xi ∈ X \| Xi ∩ (X \ X 0 ) 6= ∅} and χ0 = {(i, χ(i)) \| Xi ∩ X 00 6= ∅} where X 00 = ∪Xi0 ∈X 0 Xi and we denote it by (X , χ) ∩ X 0 . Notice that (X , χ)\|X 0 and (X , χ) ∩ X 0 are not always the same. 5 Dynamic programming The following result is the main algorithmic contribution of this paper. Lemma 5. There exists an algorithm that, given (G, H, q, k, d, D, b) as input where G and H are plane graphs such that G is a subgraph of H, H is 3-connected, q ∈ N ∪ {∞}, k ∈ N, d ∈ N≥1 , b ∈ N, and D = (T, µ, π) is a sphere-cut decomposition of H with width at most b, decides whether G has (q, k, d)-extension in H in (α(q))2 · 2 O(b log d) 2O(b log d)+2 · n steps. Proof. We use the notation E old = E(G) and E new = E(H) \ E(G), V old = V (G) and V new = V (H) \ V (G). We choose an arbitrary edge e∗ ∈ E(T ), subdivide it by adding a new vertex vnew and update T by adding a new vertex r adjacent to vnew . We then root T at this vertex r and we extend µ by setting µ(r) = ∅. In T we call leaf-edges all its edges that are incident to its leaves except from the edge er = {r, vnew }. An edge of T that is not a leaf-edge is called internal. We denote by L(T ) the set of the leaf-edges of T and we denote by I(T ) the internal edges of T . We also call er root-edge. For each e ∈ E(T ), let Te be the tree of the forest T \ {e} that does not contain r as a leaf and let Ee be the edges that are images, via µ, of the leaves of T that are also leaves of Te . We denote He = H[Ee ] and Ve = V (He ) and observe that Her = H. For each edge e ∈ I(T ), we define its children as the two edges that both belong in the connected component of T \ e that does not contain the root r and that share a common endpoint with e. Also, for each edge e ∈ E(T ), we define ∆e as the closed disk bounded by 11 Oe such that G ∩ ∆e = He . Finally, for each edge e ∈ E(T ), we set Xe = mid(e), Venew = Ve ∩ V new , Veold = Ve ∩ V old , Eenew = Ee ∩ E new , and Eeold = Ee ∩ E old . Distance signatures and dependency graphs. Let p = {Ee0 , Ee1 , Ee∞ } be a 3partition of Eenew . For each vertex v ∈ Ve , we define the (Xe , p)-distance vector of v as the function φv : Xe → {0, . . . , d + 1} such that if x ∈ Xe then φv (x) = min{wp -distGe (v, x), d + 1}. We define the (e, p)-dependency graph Ge,p = (Fe,p , Ee,p ) (that may contain loops) where Fe,p = {φv \| v ∈ Ve } and such that two (not necessarily distinct) vertices φ and φ0 of Fe,p are connected by an edge in Ee,p if and only if there exist v, v 0 ∈ Ve such that φ = φv , φ0 = φv0 and wp -distHe (v, v 0 ) > d. Notice \|Xe \| that the set Φe = {Ge,p \| p is a 3-partition of Eenew } has at most 2(d+2) elements new Xe because {Fe,p \| p is a 3-partition of Ee } ⊆ {0, . . . , d + 1} and, to each Fe,p , assign a unique edge set Ee,p . Intuitively, each Fe,p corresponds to a partition of the elements of Ve such that vertices in the same part have the same (Xe , p)-distance signature. Moreover the existence of an edge in the (e, p)-dependency graph between two such parts implies that they contain vertices, one from each part, whose wp -distance in He is bigger than d. The tables. Our aim is to give a dynamic programming algorithm running on the sc-decomposition T . For this, we describe, for each e ∈ E(T ), a table T(e) containing information on partial solutions of the problem for the graph Ge in a way that the table of an edge e ∈ E(T ) can be computed using the tables of the two children of e, the size of each table does not depend on G and the final answer can be derived by the table of the root-edge er . We define the function T mapping each e ∈ E(T ) to a collection T(e) of (d, k, q)configurations of (Xe , Ae ). In particular, Q = ((X , χ), (A, α), (F, E), δ, z) ∈ T(e) iff there exists a 3-partition p = {Ee0 , Ee1 , Ee∞ } of Eenew such that the following hold: 1. C1 , . . . , Ch are the connected components of (V (He ), Ee0 ), then • X = {V (C1 ) ∩ Xe , . . . , V (Ch ) ∩ Xe } and • ∀i∈{1,...,h} χ(i) = 1 if Ci contains some vertex of Veold , otherwise χ(i) = 0. (The pair (X , χ) encodes the connected components of the 0-edges that contain vertices of Xe and for each of them registers the number (0 or 1) of the vertices in Veold in them. This information is important to control Condition A.) 2. A is a partition of Ae such that two arcs A, A0 ∈ Ae belong in the same set, say Ai of A if and only if they are incident to the same marginal face fi of He+ . Moreover, for each i ∈ {1, . . . , \|A\|}, α(i) is equal to the number of edges in Ee1 that are inside fi . (Here (A, α) encodes the “partial” faces of the embedding of Ge that are inside ∆e . To each of them we correspond the number of 1-edges that they contain in He . This is useful in order to guarantee that during the algorithm, faces that stop being marginal do not contain more than k 1-edges, as required by Condition B.) 12 3. (F, E) is the (e, p)-dependency graph, i.e., the graph Ge,p = (Fe,p , Ee,p ). (Recall that F is the collection of all the different distance vectors of the vertices of Ve . Notice also that there might be pairs of vertices x, x0 ∈ Ve whose wp -distance in Ge is bigger than d. In order for G to have a completion of diameter d, these two vertices should become connected, at some step of the algorithm, by paths passing outside ∆e . To check this possibility, it is enough to know the distance vectors of x and x0 and these are encoded in the set F. Moreover the fact that x and x0 are still “far away” inside Ge is certified by the existence of an edge (or a loop) between their distance vectors in F.) 4. For each pair x, x0 ∈ Xe , δ(x, x0 ) = min{wp -distHe (x, x0 ), d + 1}. (This information is complementary to the one stored in F and registers the distances of the vertices in Xe inside He . As we will see, F and δ will be used in order to compute the distance vectors as well as their dependencies during the steps of the algorithm. ) 5. There is no path in He with endpoints in Veold that consists of edges in Ee0 . (This ensures that Condition A is satisfied for the current graph Ge .) 1 6. Every internal face of G+ e contains at most k edges in Ee . (This ensures that Condition B holds for all the internal faces of Ge .) 7. ∀v, v 0 ∈ Ve , either wp -distHe (v, v 0 ) ≤ d or there are two vertices x, x0 ∈ Xe such that φv (x) + φv0 (x0 ) ≤ d. (Here we demand that if two vertices x1 , x2 of Ve are “far away” (have wp -distance > d) inside He then they have some chance to come “close” (obtain wp -distance ≤ d) in the final graph, so that Condition C is satisfied. This fact is already stored by an edge in E between the two distance vectors of x and x0 and the possibility that x1 and x2 may come close at some step of the algorithm, in what concerns the graph Ge , depends only on these distance vectors and not on the vertices x1 and x2 themselves.) 8. There are at most z edges of Ee1 inside the internal faces of G+ e (clearly, this last condition becomes void when q = ∞). (This information helps us control Condition D during the algorithm.) Notice that in case Xe = ∅ the only graph that can correspond to the 6th step is the graph ({∅}, ∅) which, from now on will be denoted by G∅ . Bounding the set of characteristics. Our next step is to bound T(e) for each e ∈ E(T ). Notice first that \|Xe \| = \|Ae \| ≤ b. This means that there are 2O(b log b) instantiations of (X , χ) and 2O(k+b log b) instantiations of (A, α). As we previously O(b log d) noticed, the different instantiations of (F, E) are \|Φe \| = 22 . Moreover, there 13 2 are 2O(b log d) instantiations of δ and α(q) instantiations of z. We conclude that there exists a function f such that for each e ∈ V (T ), \|T(e)\| ≤ f (k, q, b, d). Moreover, 2 O(b log d) f (k, q, b, d) = α(q) · 2O(b log d)+2 . The characteristic function on the root edge. Observe that Enew is (k, d, q, w)edge colorable in H if and only if T(er ) 6= ∅, i.e., ((∅, ∅), (∅, ∅), G∅ ,∅, z) ∈ T(er ) for some z ≤ q. Indeed, if this happens, conditions 1–4 become void while conditions 5, 6, 7, and 8 imply that H = He satisfies the conditions A, B, C, and D respectively in the definition of the (k, d, q, w)-edge colorability of E new . The computation of the tables. We will now show how to compute T(e) for each e ∈ E(T ). We now give the definition of T(e) in the case where e is a leaf of T is the following: Given a q ∈ N ∪ {∞}, we define A(q) = {∞} if q = ∞, otherwise A(q) = {z \| z ≤ q}. Suppose now that el is a leaf-edge of T where π(el ) = [a1 , a2 , a1 ] and Ael = {a1,2 , a2,1 }. 1. If {a1 , a2 } ∈ Eeold , then T(el ) { ({{a1 }, {a2 }}, {(1, 1), (2, 1)}), = ({{a1,2 }, {a2,1 }}, {(1, 0), (2, 0)}), {(a1 , 0), (a2 , w({a1 , a2 }))}, {(a1 , w({a1 , a2 })), (a2 , 0)} , ∅ , {((a1 , a2 ), w({a1 , a2 }))}, z \| z ∈ A(q)}, 2. if {a1 , a2 } ∈ Eenew and {a1 , a2 } ⊆ Veold , then T(el ) = Q1 ∪ Q∞ where Q1 { = ({{a1 }, {a2 }}, {(1, 1), (2, 1)}) ({{a1,2 , a2,1 }}, {(1, 1)}) ( {(a1 , 0), (a2 , 1)}, {(a1 , 1), (a2 , 0)} , ∅) {((a1 , a2 ), s)}, z \| z ∈ A(q) − {0}} Q∞ { = ({{a1 }, {a2 }}, {(1, 1), (2, 1)}) ({{a1,2 , a2,1 }}, {(1, 0)}) ( {(a1 , 0), (a2 , d + 1)}, {(a1 , d + 1), (a2 , 0)} , K) {((a1 , a2 ), d + 1)}, z \| z ∈ A(q)} (the set K above contains a single edge that is not a loop), and if {a1 , a2 } ∈ Eenew and {a1 , a2 } * Veold , then T(el ) = Q1 ∪ Q∞ ∪ Q0 where Q0 = { ({{a1 , a2 }}, {(1, 1 − h{a1 , a2 } ⊆ Venew i)}) ({{a1,2 , a2,1 }}, {(1, 0)}) ({{(a1 , 0), (a2 , 0)}}, ∅) {((a1 , a2 ), 0)}, z \| z ∈ A(q)} 14 Assume now that e is a non-leaf edge of T with children el and er , the collection T(e) is given by join(T(e1 ), T(e2 )) where join is a procedure that is depicted below. Notice that Ae is the symmetric difference of Ael and Aer and Xe consists of the endpoints of the arcs in Ae . We also set XeF = (Xel ∪ Xer ) \ Xe . Procedure join Input: two collections Cel and Cer of (d, k, q)-configurations of (Xel , Ael ) and (Xer , Aer ). Output: a collection Cr of (d, k, q)-configurations of (Xe , Ae ) (1) set Ce = ∅ (2) for every pair (Qel , Qer ) ∈ Cel × Cer , if merge(Qel , Qer ) 6= void, then let Ce ← Ce ∪ {merge(Qel , Qer )}. (3) return Ce It remains to describe the routine merge. For this, assume that it receives as inputs the (d, k, q)-configurations Ql = ((Xl , χl ), (Al , αl ), (Fl , El ), δl , zl ) and Qr = ((Xr , χr ), (Ar , αr ), (Fr , Er ), δr , zr ) of (Xel , Ael ) and (Xer , Aer ) respectively. Procedure merge(Qel , Qer ) returns a (d, k, q)-configuration ((X , χ), (A, α), (F, E), δ, z) of (Xe , Ae ) constructed as follows: 1. If zr + zr > q, then return void, otherwise z = zl + zr (This controls the number of 1-edges that are now contained in ∆e ) 2. Let (X 0 , χ0 ) = (Xl , χl ) ⊕ (Xr , χr ) and if χ0−1 (2) 6= ∅ then return void. (This compute the “fusion” of the connected components of (V (Hel , Ee0l )) and (V (Her , Ee0r )) with vertices in Vel and Ver and makes sure that none of the created components contains 2 or more 0-vertices.) 3. Let (X , χ) = (Xl0 , χ0l )\|Ve (This computes the fusion (Xl0 , χ0l ) is restricted on the boundary Oe of ∆e .) 4. Let (A0 , α0 ) = (Al , αl ) ⊕ (Ar , αr ) and if α0−1 (k + 1) 6= ∅ then return void. 5. Let (A, α) = (Al , αl ) ⊕ (Ar , αr )\|Ae . 6. Compute the function γ : (Fel ∪ Fer ∪ Xe ) × (Fel ∪ Fer ∪ Xe ) → {0, . . . , d + 1}, whose description is given latter. 7. Take the disjoint union of the graphs (Fl , El ) and (Fr , Er ) and remove from it every edge {φ1 , φ2 } for which γ(φ1 , φ2 ) ≤ d. Let G + = (F + , E + ) be the obtained graph. 8. If for some edge {φ1 , φ2 } ∈ E + it holds that for every x1 , x2 ∈ Ve , γ(φ1 , x1 ) + γ(φ2 , x2 ) > d, then return void. 9. Consider the function λ : Fl ∪ Fr → {1, . . . , d}Xe such that λ(φ) = {(x, γ(φ, x)) \| x ∈ Xe }. 15 10. For every φ0 ∈ λ(Fl ∪ Fr ), do the following for every set F = λ−1 (φ0 ): identify in G + all vertices in F and if at least one pair of them is adjacent in G + , then add an loop on the vertex created after this identification. Let G = (F, E) be the resulting graph (notice that F = λ(Fl ∪ Fr )). 11. δ = {((x, x0 ), γ(x, x0 )) \| x, x0 ∈ Ve }. The definition of function γ. We present here the definition of the function γ used in the above description of the tables of the dynamic programming procedure. Given a non-empty set X and q ∈ {0, 1} we define ordq (X) = {π \| ∃X 0 ⊆ X : X 0 6= ∅ ∧ \|X 0 \| mod 2 = q ∧ π is an ordering of X 0 } Given γl and γr , we define γ : (Fel ∪ Fer ∪ Xe ) × (Fel ∪ Fer ∪ Xe ) → {0, . . . , d + 1} by distinguishing the following cases: 1. If (x ∈ Xe \ Xer ∧ φ ∈ Fel ) or (x ∈ Xe \ Xel ∧ φ ∈ Fer ), then X δs(i) (pi , pi+1 ) + γ(φ, x) = min φ(x), min{φ(p1 ) + J1,ρ−1K δs(ρ) (pρ , x) \| [p1 , . . . , pρ ] ∈ ord0 (XeF )} , where s(i) = “l” if hx ∈ Xe \ Xel i = (i mod 2), otherwise s(i) = “r”. 2. If (x ∈ Xe \ Xel ∧ φ ∈ Fel ) or (x ∈ Xe \ Xer ∧ φ ∈ Fer ), then X γ(φ, x) = min φ(p1 ) + δt(i) (pi , pi+1 ) + δt(ρ) (pρ , x) J1,ρ−1K \| [p1 , . . . , pρ ] ∈ ord1 (XeF )} , where t(i) = “l” if hx ∈ Xe \ Xel i = 6 (i mod 2), otherwise t(i) = “r”. 3. If x is one of the (at most two) vertices in (Xer ∩ Xer ) \ XeF and φ ∈ Fel ∪ Fer , then γ(φ, x) = min φ(x), X min{φ(p1 ) + δu(i) (pi , pi+1 ) + δu(q) (pρ , x) J1,ρ−1K \| [p1 , . . . , pρ ] ∈ ordq (XeF )} \| q ∈ {0, 1} where u(i) = “r” if hφ ∈ Fel i = (i mod 2), otherwise u(i) = “l”. 4. If φ, φ0 ∈ Fl ∪ Fr , then γ(φ, φ0 ) = X min φ(p1 ) + δu(i) (pi , pi+1 ) + φ0 (pρ ) J1,ρ−1K \| [p1 , . . . , pρ ] ∈ ordq (XeF ) 16 In this equality, q = 1 if φ and φ0 belong in different sets in {Fl , Fr }, otherwise q = 0. The function u is the same as in the previous case. 5. If x1 , x2 ∈ Xe \ Xer or x1 , x2 ∈ Xe \ Xel , then δ(x1 , x2 ) = min δy(0,x1 ) (x1 , x2 ), min{δy(0,x1 ) (x1 , p1 ) + X δy(i,x1 ) (pi , pi+1 ) + i∈J1,ρ−1K δy(0,x2 ) (pρ , x2 ) \| [p1 , . . . , pρ ] ∈ ord0 (XeF )} In this equality y(i, x) = “l” if hx ∈ Xe \ Xer i = hi mod 2 = 0i otherwise y(i, x) = “r”. 6. If x1 , x2 belong in different sets is {Xe \ Xer , Xe \ Xel }, then X δ(x1 , x2 ) = min δy(0,x1 ) (x1 , p1 ) + δy(i,x1 ) (pi , pi+1 ) + J1,ρ−1K δy(0,x2 ) (pρ , x2 ) \| [p1 , . . . , pρ ] ∈ ord1 (XeF ) The function y is the same as in the previous case. 7. If exactly one, say x2 , of x1 , x2 belongs in Xer ∩ Xer ) \ XeF , then δ(x1 , x2 ) = min δy(0,x1 ) (x1 , x2 ), X min min{δy(0,x1 ) (x1 , p1 ) + δy(i,x1 ) (pi , pi+1 ) + J1,ρ−1K δy(0,x2 ) (pρ , x2 ) \| [p1 , . . . , pρ ] ∈ ord q (XeF )} \| q ∈ {0, 1} The function y is the same as in the two previous cases. In case x1 belongs in Xer ∩ Xer ) \ XeF , then just swap the positions of x1 and x2 in the above equation. 8. If both x1 , x2 belong in Xer ∩ Xer ) \ XeF , then δ(x1 , x2 ) = min δl (x1 , x2 ), δr (x1 , x2 ), min{min{δz(0,j) (x1 , p1 ) + X J1,ρ−1K δz(i,j) (pi , pi+1 ) + δz(q,j) (pρ , x2 ) \| [p1 , . . . , pρ ] ∈ ordq (XeF )} \| (q, j) ∈ {0, 1}2 } In the previous equality, z(i, j) = “l” if (i + j mod 2) = 0, otehrwise z(i, x) = “r”. Running time analysis. It now remains to prove that procedure join runs in 2 O(b log d) (α(q))2 · 2O(k )+2 steps. Recall that there exists a function f such that \|T(e)\| ≤ f (k, q, b, d). Therefore merge will be called in Step (2) at most (f (k, q, b, d))2 times. 17 The first computationally non-trivial step of merge is Step 5, where function γ is computed. Notice that γ has at most ((d + 1)\|Xel \| + (d + 1)\|Xer \| + \|Xe \|)2 = 2O(b·log d) entries and each of their values require running over all permutations of the subsets of XeF that are at most b! = 2O(b·log b) . These facts imply that the computation of γ takes 2O(b·log b) steps. As Steps 610 deal with graphs of 2O(b·log d) vertices, the running time of join is the claimed one. We are now in position to prove the main algorithmic result of this paper. Proof of Theorem 2. Given an input I = (G, q, k, d) of BBFPDC, we consider the graph H = G(max{3,k}) whose construction takes O(k 2 n) steps, because of Lemma 3. Then run the algorithm of Proposition 4 with (H, w) as input, where w = c1 · c2 · k · d. If the answer is that bw(H) > w, then, From Proposition 3, tw(G) > c1 · d, therefore, from Lemma 1, we can safely report that I is a no-instance. If the algorithm of Proposition 4 outputs a sphere-cut decomposition D = (T, µ) of width at most w = O(k · d) then we call the dynamic programming algorithm of Lemma 5, with input (G, H, q, k, d, D, b). This, from Lemma 4, provides an answer to BBFPDC for the 2 O((kd) log d) O((kd) log d) instance I in (α(q))2 · 2O((kd) log d)+2 · n = (α(q))2 · 22 · n steps and this completes the proof of the theorem. 6 NP-hardness proofs In this section we show that the Bounded Budget Plane Diameter Completion and Bounded Budget/Face Plane Diameter Completion problems are NPcomplete. Here we consider R2 -plane graphs, i.e., graphs embedded in the plane R2 . Each 2 R -plane graph has exactly one unbounded face, called the outer face, and all other faces are called inner faces. Take in mind that every S0 -plane graph has as many embeddings in R2 as the number of its faces (each correspond on which face of the embedding in S0 will be chosen to be the outer face in R2 ). All our problems can be equivalently restated on R2 -plane graphs. We choose such embeddings because they facilitate the presentation of the result of this section. We also need some additional terminology. A walk in a graph G of is a sequence P = v0 , e1 , v1 , e2 , . . . , es , vs of vertices and edges of G such that v0 , . . . , vs ∈ V (G), e1 , . . . , es ∈ E(G), the edges e1 , . . . , es are pairwise distinct, and for i ∈ {1, . . . , s}, ei = {vi−1 , vi }; v0 , vs are the end-vertices of the walk. A walk is closed if its end-vertices are the same. The length of a walk P is the number of edges in P . For a walk P with end-vertices u, v, we say that P is a (u, v)-walk. A walk is a path if v0 , . . . , vs and e1 , . . . , es are pairwise distinct with possible exception v0 = vs , and a cycle is a closed path. We write P = v0 . . . vs to denote a walk P = v0 , e1 , . . . , es , vs omitting edges. Recall that the 3-Satisfiability problem for a given Boolean formula φ = C1 ∧ . . . ∧ Cm with clauses C1 , . . . , Cm with 3 literals each over variables x1 , . . . , xn , asks 18 whether x1 , . . . , xn have an assignment that satisfies φ. We write that a literal xi ∈ Cj (xi ∈ Cj resp.) if this interval is in Cj . For an instance φ of 3-Satisfiability, we define the graphs Gφ and G0φ as follows. The vertex set of Gφ is {x1 , . . . , xn } ∪ {C1 , . . . , Cm }, and for i ∈ {1, . . . , n} and j ∈ {1, . . . , m}, {xi , Cj } ∈ E(Gφ ) if and only if Cj contains either xi or xi . Respectively, V (G0φ ) = {x1 , x1 , . . . , xn , xn }∪{C1 , . . . , Cm } and E(G0φ ) = {{xi , xi }\|1 ≤ i ≤ n}∪{{xi , Cj }\|xi ∈ Cj , 1 ≤ i ≤ n, 1 ≤ j ≤ m}∪{{xi , Cj }\|xi ∈ Cj , 1 ≤ i ≤ n, 1 ≤ j ≤ m}. Let φ over variables x1 , . . . , xn be an instance of 3-Satisfiability such that G0φ is planar, and let G0 be a plane embedding of G0φ . Let also Rφ = {{xi , xi }\|1 ≤ i ≤ n} ⊆ E(G0 ). We define the bipartite graph H(G0 ) as the graph with the vertex set Rφ ∪ F (G0 ) and the edge set {{e, f }\|e ∈ Rφ , f ∈ F (G0 ) such that e is incident to f }. We consider the following special variant of Satisfiability. Plane Satisfiability with Connectivity of Variables Input: A Boolean formula φ = C1 ∧ . . . ∧ Cm with clauses C1 , . . . , Cm with at most 3 literals each over variables x1 , . . . , xn such that G0φ is planar, and a plane embedding G0 of G0φ such that H(G0 ) is connected. Output: Is it possible to satisfy φ? We show that this problem is hard. Lemma 6. Plane Satisfiability with Connectivity of Variables is NPcomplete. Proof. It is straightforward to see that Plane Satisfiability with Connectivity of Variables is in NP. To show NP-hardness, we reduce Planar 3-Satisfiability, i.e. the 3-Satisfiability problem restricted to instances φ such that Gφ is planar. This problem was shown to be NP-complete by Lichtenstein in [15]. Let φ = C1 ∧ . . . ∧ Cm over variables x1 , . . . , xn be an instance of Planar 3Satisfiability. For the plane graph Gφ , we construct its plane embedding G. It is well known that it can be done in polynomial time, e.g., by the classical algorithm of Hopcroft and Tarjan [12] or by the algorithm of Boyer and Myrvold [1]. We consequently consider variables x1 , . . . , xn and modify φ and G. Suppose that a variable xi occurs in the clauses Cj1 , . . . , Cjp(i) . Without loss of generality we assume that the edges {xi , Cj1 }, . . . , {xi , Cjp(i) } are ordered clockwise in G as shown in Fig. 2 a). We perform the following modifications of φ and G. • Replace xi by 2p(i) new variables xi,1 , . . . , xi,2p(i) . • For k ∈ {1, . . . , p(i)}, replace xi in Cjk by xi,2k−1 . 2p(i) • Construct 2p(i) clauses Ci1 , . . . , Ci we assume that xi,0 = xi,2p(i) . where Cik = xi,k−1 ∨xi,k for k ∈ {1, . . . 2p(i)}; • Modify the current plane graph as it is shown in Fig. 2.3 3 Here and further we demonstrate constructions of plane embeddings in figures instead of long technical formal descriptions. 19 Denote the obtained Boolean formula and plane graph by φ̂ and Ĝ respectively. By the construction, Ĝ is a plane embedding of Gφ̂ . C j1 xi,1 Ci1 Cj1 Ci2 xi,2 xi,2p xi Ci2p xi,3 xi,2p−1 Cj2 Cjp C j2 Cjp a) b) Figure 2: Modification of φ and G: a) before the modification and b) after; p = p(i). We show that φ can be satisfied if and only if φ̂ has a satisfying assignment. Suppose that the variables have assigned values such that φ = true. For each i ∈ {1, . . . , n}, we assign the same value as xi for all the variables xi,1 , . . . , xi,2p(i) that replace xi in φ̂. It is straightforward to verify that φ̂ = true for this assignment. Assume now that φ̂ = true for some values of the variables. Observe that for each i ∈ {1, . . . , n}, the variables 2p(i) xi,1 , . . . , xi,2p(i) that replace xi should have the same value to satisfy Ci1 , . . . , Ci . It remains to observe that if each xi has the same value as xi,1 , . . . , xi,2p(i) , then φ = true by the construction of φ̂. C j1 xi,1 xi,1 Ci2 Ci1 xi,2 xi,2p xi,2 xi,2p Ci2p fi xi,2p−1 xi,2p−1 xi,3 xi,3 C j2 Cjp Figure 3: Construction of Ĝ0 ; it is assumed that Cj1 contains xi , Cj2 contains xi and Cjp(i) contains xi , and p = p(i). Observe that each variable xi,k in φ̂ occurs in at most 3 clauses, and it occurs at least once in positive and at least once with negation. It implies that a plane embedding Ĝ0 of G0φ̂ can be constructed from Ĝ by “splitting” the variable vertices as shown in Fig. 3. Clearly, Ĝ0 can be constructed in polynomial time. We claim that H(Ĝ0 ) is connected. To see it, observe that Ĝ0 is constructed from G by replacing each variable-vertex xi by the cycle Li = Ci1 xi,1 xi,1 Ci2 . . . Ci1 (see Fig. 3). 20 Respectively, this graph has n new faces that are inner faces of these cycles. All other faces correspond to the faces of G. Denote by fi the inner face of Li for i ∈ {1, . . . , n}. Notice that Rφ̂ contains edges from the cycles Li . It follows that each vertex of Rφ̂ is adjacent to some vertex fi in H(Ĝ0 ). Hence, to prove the connectivity of H(Ĝ0 ), it is sufficient to show that for any two vertices h1 , h2 ∈ F (Ĝ0 ), H(Ĝ0 ) has a (h1 , h2 )-walk. Consider the dual G∗ of Ĝ0 . Recall that V (G∗ ) = F (Ĝ0 ) and two vertices of G∗ are adjacent if and only if the corresponding faces of Ĝ0 are adjacent. It is straightforward to observe that the dual of any plane graph is always connected. Hence, to show that for any two vertices h1 , h2 ∈ F (Ĝ0 ) of H(Ĝ0 ), H(Ĝ0 ) has a (h1 , h2 )-walk, it is sufficient to prove that it holds for any two h1 , h2 that are adjacent vertices of G∗ , i.e., adjacent faces of Ĝ0 . Suppose that h1 = fi for some i ∈ {1, . . . , n}. Then h2 is a face corresponding to a face h02 of G such that the vertex xi lies on the boundary of h02 . Then by the construction of Ĝ0 , there is an edge e = {xi,j , xi,j } of Ĝ0 that lies on the boundaries of h1 and h2 . Because e is a vertex of H(Ĝ0 ) adjacent to h1 , h2 , there is a (h1 , h2 )-walk in H(Ĝ0 ). Assume now that h1 , h2 are faces of Ĝ0 distinct from fi for i ∈ {1, . . . , n}. Because h1 , h2 are adjacent in G∗ , the faces h1 , h2 correspond to faces h01 , h02 of G such that h01 , h02 has a common vertex xi on their boundaries. It implies that h1 , h2 are adjacent to fi in G∗ . We already proved that H(Ĝ0 ) has (fi , h1 ) and (fi , h2 )-walks. Therefore, H(Ĝ0 ) has an (h1 , h2 )-walk. It completes the proof of connectedness of H(Ĝ0 ) and the proof of the lemma. For the proof of our main result, we need some special gadgets. We introduce them and prove their properties that will be useful further. Let r ≥ 3 be a positive integer. We construct the graph Wr (v1 , . . . , vr ) as follows (see Fig. 4). • Construct vertices v1 , . . . , vr and a vertex u. • For i ∈ {1, . . . , r}, construct a (vi , u) path xi0 . . . xir of length r, vi = xi0 , u = xir . • For j ∈ {1, . . . , r − 1}, construct a cycle x1j . . . xrj x1j . i • For i ∈ {1, . . . , r} and j ∈ {1, . . . , r − 1}, construct an edge {xi−1 j−1 , xj }; we assume that x0j = xrj for j ∈ {0, . . . , r}. We say that the vertices of V (Wr (v1 , . . . , vr )) \ {v1 , . . . , vr } are the inner vertices of the gadget. Let G be a plane graph with a face f , and let v1 . . . vr v1 , r ≥ 3, be a facial walk for f . We say that G0 is obtained from G by attaching a web to f if G0 is constructed by adding a copy of Wr (v1 , . . . , vr ), where the vertices v1 , . . . , vr of the gadget are identified with the vertices with the same names in the facial walk, and embedding Wr (v1 , . . . , vr ) if f as is shown in Fig. 5. Notice that some vertices in the facial walk can occur several times. Lemma 7. Let G be a plane graph with a face f that has a facial walk of length r ≥ 3, and let G0 be a plane graph obtained from G by attaching a web to f . 21 v1 v2 v5 v3 v4 Figure 4: Construction of W5 (v1 , . . . , v5 ). Wr (v1 , . . . , vr ) f Figure 5: Attachment of a web. i) For any two vertices u, v ∈ V (G), distG0 (u, v) = distG (u, v). Moreover, any shortest (u, v)-path in G0 has no inner vertices of Wr (v1 , . . . , vr ) attached to f . ii) For any vertex v ∈ V (Wr (v1 , . . . , vr )), there is a vertex u ∈ V (G) such that distG0 (u, v) ≤ r. Proof. Let v1 . . . vr v0 be a facial walk for f . To prove i), it is sufficient to observe that for all vi , vj , the length of any (vi , vj )-path in Wr (v1 , . . . , vr ) is greater that the length of a shortest (vi , vj )-path in G that lies on the boundary of f . The definition of Wr (v1 , . . . , vr ) immediately implies ii). Let h be a positive integer. The graph Mh (u1 , u2 , u3 ) is defined as follows (see Fig. 6). • Construct vertices u1 , u2 , u3 and v1 , v2 , v3 . • For i ∈ {1, 2, 3}, construct a (ui , vi ) path xi0 . . . xir of length `, ui = xi0 , vi = xih . • For j ∈ {1, . . . , h}, construct a cycle x1j . . . xrj x1j . • For j ∈ {1, . . . , h}, construct edges {x1j−1 , x2j }, {x1j−1 , x3j } and {x2j−1 , x3j }. We say that the vertices of V (Mh (u1 , u2 , u3 )) \ {u1 , u2 , u3 } are the inner vertices of the gadget. We also say that u1 is the root and v1 is the pole of M` (u1 , u2 , u3 ). Let G be a plane graph, and let u1 ∈ V (G) be a vertex incident to a face f with a triangle facial walk u1 u2 u3 u1 . Let also ` be a positive integer. We say that G0 is 22 u2 u3 v1 u1 Figure 6: Construction of M3 (u1 , u2 , u3 ). obtained from G by attaching a mast of height h rooted in u1 to f if G0 is constructed by adding a copy of Mh (u1 , u2 , u3 ), where the vertices u1 , u2 , u3 of the gadget are identified with the vertices with the same names in the facial walk, and embedding Mh (u1 , u2 , u3 ) in f . We need the properties summarized in the following straightforward lemma. Lemma 8. Let ` be a positive integer. Let G be a plane graph, and let u1 be a vertex of G incident to a face f with a triangle facial walk u1 u2 u3 u1 . Let also G0 be a plane graph obtained from G by attaching a mast of height h rooted in u1 to f . i) For any two vertices u, v ∈ V (G), distG0 (u, v) = distG (u, v). Moreover, any shortest (u, v)-path in G0 has no inner vertices of Mh (u1 , u2 , u2 ) attached to f . ii) For any vertex v ∈ V (Mh (u1 , u2 , u2 )), distG0 (u1 , v) ≤ h. iii) If v is the pole of Mh (u1 , u2 , u2 ), then distG0 (u1 , v) = h and distG0 (u2 , v) > h, distG0 (u3 , v) > h. iv) For any inner vertices x, y of Mh (u1 , u2 , u2 ), distG0 (x, y) ≤ h. Now we are ready to prove the main result of the section. Proof of Theorem 1. It is straightforward to see that BPDC and BFPDC are in NP. To show NP-hardness, we reduce Plane Satisfiability with Connectivity of Variables that was shown to be NP-complete in Lemma 6. First, we consider BPDC. Let (φ, G0 ) be an instance of Plane Satisfiability with Connectivity of Variables, where φ = C1 ∧ . . . ∧ Cm is a Boolean formula with clauses C1 , . . . , Cm with at most 3 literals each over variables x1 , . . . , xn such that G0φ is planar, and G0 is a plane embedding of G0φ such that H(G0 ) is connected. Recall that H(G0 ) is the bipartite graph with the bipartition of the vertex set (Rφ , F (G0 )), where Rφ = {{xi , xi }\|1 ≤ i ≤ n} ⊆ E(G0 ), and F (G0 ) is the set of faces of G0 , and for e ∈ Rφ and f ∈ F (G0 ), {e, f } ∈ E(H(G0 )) if and only if the edge e is incident to the face f in G0 . Notice that degH(G0 ) (e) ≤ 2 for any e ∈ Rφ . We select an arbitrary vertex r ∈ F (G0 ) of H(G0 ). Using the connectedness of H(G0 ), we find in polynomial time a tree T of shortest (r, e)-paths for e ∈ Rφ by the 23 breadth-first search. We assume that T is rooted in r and it defines the parent-child relation on T . Let L ⊆ Rφ be the set of leaves of T , and let s = max{distT (r, e)\|e ∈ L}. vp vp yi yi fi yi fi fi zi xi xi xi yi fi zi xi wi vc b) a) Figure 7: Construction of gadgets for {xi , xi }. We construct the plane graph Ĝ as follows. i) Construct a copy of G0 . ii) For each vertex f ∈ V (T ) such that f ∈ F (G0 ), crate a vertex vf embedded in the face f . iii) For each e = {xi , xi } ∈ Rφ \ L, denote by p its parent and by c its child in T , construct vertices yi , y i , zi and edges {xi , yi }, {yi , vp }, {xi , vc }, {xi , zi }, {xi , y i }, {y i , vp }, {xi , vc }, {xi , zi }, {zi , vp } and embed them as is shown in Fig. 7 a). Denote by fi the inner face of the cycle xi yi vp zi xi and by f i the inner face of the cycle xi y i vp zi xi . iv) For each e = {xi , xi } ∈ L, denote by p its parent in T , construct vertices yi , y i , zi , wi and edges {xi , yi }, {yi , vp }, {xi , wi }, {xi , zi }, {xi , y i }, {y i , vp }, {xi , wi }, {xi , zi }, {zi , vp } and embed them as is shown in Fig 7 b). Denote by fi the inner face of the cycle xi yi vp zi xi and by f i the inner face of the cycle xi y i vp zi xi . v) For each i ∈ {1, . . . , n} and j ∈ {1, . . . , m}, if {xi , Cj } ∈ E(G0 ) ({xi , Cj } ∈ E(G0 ) resp.), replace this edge by a (xi , Cj )-path (by (xi , Cj )-path resp.) of length 2s − distT (r, {xi , xi }). We denote the constructed at this stage graph by Ĝ1 . Observe that Ĝ1 is connected. Hence, each face has a facial walk. Denote by ` the length of a longest facial walk in Ĝ1 . Now we proceed with construction of Ĝ. vi) For each face f ∈ F (Ĝ1 ) distinct from the faces fi , f i for i ∈ {1, . . . , n}, attach a web to f . Denote the constructed at this stage graph by Ĝ2 . Notice that Ĝ2 is 3-connected due to attached webs. vii) For j ∈ {1, . . . , m}, select a face f of the obtained graph such that Cj is incident to f and attach a mast of height `+2s rooted in Cj to f (notice that the boundary of f is a triangle because of attached webs). 24 viii) For each e = {xi , xi } ∈ L, attach a mast of height ` + 4s − 1 − distT (r, e) rooted in wi to the face with the facial walk wi xi xi wi . ix) For the vertex vr , select a face f with a triangle boundary such that vr is incident to f (such a face always exists due to attached webs) and attach a mast of height ` + 8s rooted in vr to f . Notice that the obtained graph Ĝ is 3-connected because Ĝ2 is 3-connected and attachments of masts cannot destroy 3-connectivity. Also only the faces fi , f i for i ∈ {1, . . . , n} have degree 4, and all other faces have degree 3. To complete the construction of an instance of BPDC, we set q = n and d = 2`+12s. We show that (φ, G0 ) is a yes-instance of Plane Satisfiability with Connectivity of Variables if and only if (Ĝ, q, d) is a yes-instance of BPDC. Suppose that (φ, G0 ) is a yes-instance of Plane Satisfiability with Connectivity of Variables. Assume that the variables x1 , . . . , xn have values such that φ = true. For i ∈ {1, . . . , n}, if xi = true, then we add an edge {xi , vp } for the parent p of {xi , xi } in T and embed this edge in fi . Respectively, we add an edge {xi , vp } and embed this edge in f i if xi = f alse. Denote the obtained graph by Ĝ0 . We show that diam(Ĝ0 ) ≤ d. By the construction of Ĝ1 , for any vertex v ∈ V (Ĝ1 ), distĜ1 (vr , v) ≤ 3s. By Lemma 7, any vertex v ∈ V (Ĝ2 ) is at distance at most ` from a vertex of Ĝ1 in Ĝ2 . Hence, for any vertex v ∈ V (Ĝ2 ), distĜ2 (vr , v) ≤ ` + 3s. Observe also that for any e = {xi , xi } ∈ L, distĜ0 (vr , wi ) = distT (r, e) + 1. To show that for any u, v ∈ V (Ĝ0 ), distĜ0 (u, v) ≤ d, we consider five cases. Case 1. u, v ∈ V (Ĝ2 ). Because distĜ2 (vr , u) ≤ ` + 3s and distĜ2 (vr , v) ≤ ` + 3s, distĜ0 (u, v) ≤ distĜ2 (u, v) ≤ 2` + 6s ≤ d. Case 2. u, v are vertices of the same mast attached to a face of Ĝ2 . By Lemma 8, distĜ0 (u, v) is at most the height of the mast, and we have that distĜ0 (u, v) ≤ `+8s ≤ d. Case 3. u ∈ V (Ĝ2 ) and v is a vertex of a mast attached to a face of Ĝ2 . By Lemma 8, distĜ0 (u, vr ) ≤ ` + 8s if the mast is rooted in vr . Suppose that this mast is rooted in some other vertex z, i.e., z = wi or z = Cj for some i ∈ {1, . . . , n}, j ∈ {1, . . . , m}. Then distĜ0 (u, vr ) ≤ ` + 4s − 1 + distĜ1 (z, r) ≤ ` + 8s. Because distĜ0 (vr , v) ≤ distĜ2 (vr , v) ≤ ` + 3s, distĜ0 (u, v) ≤ 2` + 11s ≤ d. Case 4. u, v are vertices of distinct masts attached to faces of Ĝ2 that are rooted in z, z 0 6= vr respectively. If z = wi for some i ∈ {1, . . . , n}, then distĜ0 (u, vr ) ≤ ` + 4s − 1 − distT (r, e) + distĜ0 (vr , wi ) ≤ (` + 4s − 1 − distT (r, e)) + (distT (r, e) + 1) ≤ ` + 4s where e = {xi , xi }. If z = Cj for some j ∈ {1, . . . , m}, then distĜ0 (u, vr ) ≤ ` + 2s + distĜ1 (Cj , vr ) ≤ ` + 5s. Clearly, the same bounds hold for distĜ0 (v, vr ). We have that distĜ0 (u, v) ≤ distĜ0 (u, vr ) + distĜ0 (vr , v) ≤ 2` + 10s ≤ d. It remains to consider the last case. Case 5. u, v are vertices of masts attached to faces of Ĝ2 such that u is in the mast rooted in vr and v is in a mast rooted in z 6= vr . Suppose that z = wi for some i ∈ {1, . . . , n}. Then e = {xi , xi } ∈ L. We have that distĜ0 (u, v) ≤ distĜ0 (u, vr ) + 25 distĜ0 (vr , wi ) + distĜ0 (wi , v) ≤ (` + 8s) + (distT (r, e) + 1) + (` + 4s − 1 − distT (r, e)) ≤ 2` + 12s ≤ d. Assume that z = Cj for j ∈ {1, . . . , m}. Then the clause Cj in φ contains a literal that has the value true. Let xi be such a literal (the case when Cj contains some xi = true is symmetric). Notice that if xi = true, then for the vertex xi ∈ V (Ĝ0 ), distĜ0 (xi , vr ) = distT (e, r) for e = {xi , xi } by the construction of Ĝ and the selection of the added edges. Then, distĜ0 (u, v) ≤ distĜ0 (u, vr )+distĜ0 (vr , xi )+distĜ0 (xi , Cj )+ distĜ0 (Cj , v) ≤ (` + 8s) + distT (r, e) + (2s − distT (r, e)) + (` + 2s) ≤ 2` + 12s ≤ d. Suppose now that (Ĝ, q, d) is a yes-instance of BPDC. Let A be a set of at most q edges such that the graph Ĝ0 obtained from Ĝ by the addition of A has diameter at most d. Because only the faces fi , f i for i ∈ {1, . . . , n} have degree 4 and all other faces have degree 3, each edge of A has its end-vertices in the boundary of some fi or f i and is embedded in this face. Using this observation, denote by Ĝ01 and Ĝ02 the graphs obtained from Ĝ1 and Ĝ2 respectively by the addition of A. Let vr0 be the pole of the mast rooted in vr . Because diam(Ĝ0 ) ≤ d, for any u ∈ V (Ĝ0 ), distĜ0 (vr0 , u) ≤ d and, in particular, it holds for poles of other masts. Consider masts rooted in wi for e = {xi , xi } ∈ L. For a mast rooted in wi , denote by wi0 its pole. By Lemma 8, distĜ0 (vr0 , wi0 ) = distĜ0 (vr0 , vr ) + distĜ0 (vr , wi ) + distĜ0 (wi , wi0 ) = (` + 8s) + distĜ0 (vr , wi ) + (` + 4s − 1 − distT (r, e)), and by Lemma 7, 2 distĜ0 (vr , wi ) = distĜ0 (vr , wi ). We conclude that distĜ0 (vr , wi ) ≤ distT (r, e) + 1. 2 1 1 Because distT (r, e) + 1 ≤ s + 1, a shortest (vr , wi )-path in Ĝ01 does not contain the vertices Cj for j ∈ {1, . . . , m}. We obtain that for every edge e0 = {xh , xh } that lies on the unique (r, e)-path in T , {xi , vp } ∈ A or {xi , vp } ∈ A where p is the parent of e0 in T . This holds for each leaf of T . Because Rφ ⊆ V (T ) and k = n, we have that for each h ∈ {1, . . . , n}, either {xi , vp } ∈ A or {xi , vp } ∈ A where p is the parent of {xh , xh } in T . For h ∈ {1, . . . , n}, we let the variable xh = true if {xi , vp } ∈ A and xh = f alse otherwise. We show that this assignment satisfies φ. Consider a clause Cj for j ∈ {1, . . . , m}. To simplify notations, assume that Cj contains literals xi1 , xi2 , xi3 (the cases when Cj contains two literals and/or some literals are negations of variables are considered in the same way). Let Cj0 be the pole of the mast rooted in the vertex Cj . We have that distĜ0 (vr0 , Cj0 ) ≤ d. By Lemma 8,distĜ0 (vr0 , Cj0 ) = distĜ0 (vr0 , vr ) + distĜ0 (vr , Cj ) + distĜ0 (Cj , Cj0 ) = (` + 8s) + distĜ0 (vr , Cj ) + (` + 2s), and by Lemma 7, distĜ0 (vr , Cj ) = distĜ0 (vr , Cj ). Therefore, 2 2 1 distĜ0 (vr , Cj ) ≤ 2s. Let eh = {xih , xih } for h ∈ {1, 2, 3}. By the construction 1 of Ĝ0 , distĜ0 (vr , Cj ) = min{distĜ0 (vr , xih ) + (2s − distT (r, eh ))\|1 ≤ h ≤ 3}. Let 1 1 distĜ0 (vr , Cj ) = distĜ0 (vr , xih ) + (2s − distT (r, eh )) for h ∈ {1, 2, 3}. It follows that 1 1 distĜ0 (vr , xih ) ≤ distT (r, eh ), and this immediately implies that {vp , xih } ∈ A where 1 p is the parent of eh in T . By the definition, xih = true and, therefore, Cj = true. This holds for each Cj for j ∈ {1, . . . , m}, and we conclude that φ = true. To complete the proof of the NP-hardness of BPDC, it remains to observe that Ĝ can be constructed in polynomial time. To show NP-hardness of BFPDC, we use similar arguments. 26 Let (φ, G0 ) be an instance of Plane Satisfiability with Connectivity of Variables, where φ = C1 ∧ . . . ∧ Cm is a Boolean formula with clauses C1 , . . . , Cm with at most 3 literals each over variables x1 , . . . , xn such that G0φ is planar, and G0 is a plane embedding of G0φ such that H(G0 ) is connected. As before, we pick an arbitrary vertex r ∈ F (G0 ) of H(G0 ) and find a tree T rooted in r of shortest (r, e)-paths for e ∈ Rφ with the set of leaves L ⊆ Rφ . Let s = max{distT (r, e)\|e ∈ L}. vp vp yi yi yi yi fi fi xi xi xi xi wi vc b) a) Figure 8: Construction of gadgets for {xi , xi }. We construct the plane graph G̃ similarly to the construction of Ĝ above. The only difference is that Steps iii) and iv) are replaced by the following steps iii∗ ) and iv∗ ). iii∗ ) For each e = {xi , xi } ∈ Rφ \ L, denote by p its parent and by c its child in T , construct vertices yi , y i and edges {xi , yi }, {yi , vp }, {xi , vc }, {xi , y i }, {y i , vp }, {xi , vc } and embed them as is shown in Fig. 8 a). Denote by fi the inner face of the cycle xi yi v p y i xi xi . iv∗ ) For each e = {xi , xi } ∈ L, denote by p its parent in T , construct vertices yi , y i , wi and edges {xi , yi }, {yi , vp }, {xi , wi }, {xi , y i }, {y i , vp }, {xi , wi } and embed them as is shown in Fig 8 b). Denote by fi the inner face of the cycle xi yi vp y i xi xi . Observe that G̃ can be obtained from Ĝ by the deletion of the vertices z1 , . . . , zn , and for any u, v ∈ V (G̃), distG̃ (u, v) = distĜ (u, v). Notice that the obtained graph G̃ is 3-connected, the faces f1 , . . . , fn have degree 5, and all other faces have degree 3. To complete the construction of an instance of BFPDC, we set k = 1 and d = 2` + 12s. We show that (φ, G0 ) is a yes-instance of Plane Satisfiability with Connectivity of Variables if and only if (G̃, k, d) is a yes-instance of BFPDC. Suppose that (φ, G0 ) is a yes-instance of Plane Satisfiability with Connectivity of Variables. Assume that the variables x1 , . . . , xn have values such that φ = true. For i ∈ {1, . . . , n}, if xi = true, then we add an edge {xi , vp } for the parent p of {xi , xi } in T and embed this edge in fi . Respectively, we add an edge {xi , vp } and embed this edge in fi if xi = f alse. Denote the obtained graph by G̃0 . By exactly the same arguments as for the proof of the inequality diam(Ĝ0 ) ≤ d, we have that diam(G̃0 ) ≤ d. Suppose now that (Ĝ, k, d) is a yes-instance of BFPDC. Let A be a set of edges such that the graph G̃0 obtained from G̃ by the addition of A has diameter at most 27 d. Because only the faces f1 , . . . , fn have degree 5 and all other faces have degree 3, each edge of A has its end-vertices in the boundary of some fi and is embedded in this face. Because k = 1, at most one edge of A is embedded in fi for i ∈ {1, . . . , n}. Let vr0 be the pole of the mast rooted in vr . Because diam(G̃0 ) ≤ d, for any u ∈ V (G̃0 ), distG̃0 (vr0 , u) ≤ d and, in particular, it holds for poles of other masts. Consider masts rooted in wi for e = {xi , xi } ∈ L. For a mast rooted in wi , denote by wi0 its pole. Because distG̃0 (vr0 , wi0 ) ≤ d, by the same arguments that were used above in the proof of the NP-hardness of BPDC, we obtain that it implies that for each h ∈ {1, . . . , n}, either {xi , vp } ∈ A or {xi , vp } ∈ A where p is the parent of {xh , xh } in T . For h ∈ {1, . . . , n}, we let the variable xh = true if {xi , vp } ∈ A and xh = f alse otherwise. To prove that this assignment satisfies φ, we again use the same arguments as above: it follows from the fact that for each clause Cj , distG̃0 (vr0 , Cj0 ) ≤ d where Cj0 is the pole of the mast rooted in the vertex Cj . To complete the proof of the NP-hardness of BPDC, it remains to observe that G̃ can be constructed in polynomial time. We proved that BPDC is NP-complete for 3-connected planar graphs. By the Whitney’s theorem (see, e.g., [6]), any two plane embeddings of a 3-connected plane graphs are equivalent. It gives the following corollary. Bounded Budget Planar Diameter Completion Input: A planar graph G, non-negative integers k and d. Output: Is it possible to obtain a planar graph G0 of diameter at most d from G by adding at most k edges? Corollary 2. Bounded Budget Planar Diameter Completion is NP-complete for 3-connected planar graphs. 7 Discussion We remark that our algorithm still works for the classic PDC problem when the face-degree of the input graph is bounded. For this we define the following problem: Bounded Face BDC (FPDC) Input: a plane graph G with face-degree at most k ∈ N≥3 , and d ∈ N Question: is it possible to add edges in G such that the resulting embedding remains plane and has diameter at most d? We directly have the following corollary of Theorem 2. O((kd) log d) Theorem 3. It is possible to construct an O(n3 ) + 22 FPDC. · n-step algorithm for To construct an FPT-algorithm for PDC when parameterized by d remains an insisting open problem. The reason why our approach does not apply (at least directly) for PDC is that, as long as a completion may add an arbitrary number of edges in each 28 face, we cannot guarantee that our dynamic programming algorithm will be applied on a graph of bounded branchwidth. We believe that our approach and, in particular, the machinery of our dynamic programming algorithm, might be useful for further investigations on this problem. All the problems in this paper are defined on plane graphs. However, one may also consider the “non-embedded” counterparts of the problems PDC and BPDC by asking that their input is a planar combinatorial graphs (without a particular embedding). Similarly, such a counterpart can also be defined for the case of BFPDC if we ask whether the completion has an embedding with at most k new edges per face. Again, all these parameterized problems are known to be (non-constructively) in FPT, because of the results in [18, 16]. However, our approach fails to design the corresponding algorithms as it strongly requires an embedding of the input graph. For this reason we believe that even the non-embedded versions of BPDC and BFPDC are as challenging as the general Planar Diameter Completion problem. Acknowledgement. We would like to thank the anonymous referees of an earlier version of this paper for their remarks and suggestions that improved the presentation of the paper. References [1] John M. Boyer and Wendy J. Myrvold. On the cutting edge: Simplified O(n) planarity by edge addition. J. Graph Algorithms Appl., 8(2):241–273, 2004. [2] Dimitris Chatzidimitriou, Archontia C. Giannopoulou, Spyros Maniatis, Clment Requil, Dimitrios M. Thilikos, and Dimitris Zoros. Fixed parameter algorithms for completion problems on planar graphs. Manuscript, 2015. [3] F. R. K. Chung. Diameters of graphs: Old problems and new results. Congressus Numerantium, 60, 1987. [4] Bruno Courcelle. The expression of graph properties and graph transformations in monadic second-order logic. Handbook of Graph Grammars, pages 313–400, 1997. [5] Italo J. Dejter and Michael R. Fellows. Improving the diameter of a planar graph. Manuscript, may 1993. [6] Reinhard Diestel. Graph Theory, 4th Edition, volume 173 of Graduate texts in mathematics. Springer, 2012. [7] Frederic Dorn, Eelko Penninkx, Hans L. Bodlaender, and Fedor V. Fomin. Efficient exact algorithms on planar graphs: Exploiting sphere cut decompositions. Algorithmica, 58(3):790–810, 2010. 29 [8] Rodney G. Downey and Michael R. Fellows. Fundamentals of Parameterized Complexity. Texts in Computer Science. Springer, 2013. [9] Yong Gao, Donovan R. Hare, and James Nastos. The parametric complexity of graph diameter augmentation. Disc. Appl. Math., 161(10-11):1626–1631, 2013. [10] Qian-Ping Gu and Hisao Tamaki. Optimal branch-decomposition of planar graphs in O(n3 ) time. ACM Transactions on Algorithms, 4(3), 2008. [11] Qian-Ping Gu and Hisao Tamaki. Improved bounds on the planar branchwidth with respect to the largest grid minor size. In Algorithms and Computation - 21st International Symposium, (ISAAC 2010), pages 85–96, 2010. [12] John E. Hopcroft and Robert Endre Tarjan. Efficient planarity testing. J. ACM, 21(4):549–568, 1974. [13] Athanassios Koutsonas and Dimitrios M. Thilikos. Planar feedback vertex set and face cover: Combinatorial bounds and subexponential algorithms. Algorithmica, 60(4):987–1003, 2011. [14] Chung-Lun Li, S. Thomas McCormick, and David Simchi-Levi. On the minimumcardinality-bounded-diameter and the bounded-cardinality-minimum-diameter edge addition problems. Oper. Res. Lett., 11(5):303–308, 1992. [15] David Lichtenstein. Planar formulae and their uses. SIAM Journal on Computing, 11(2):329–343, 1982. [16] Neil Robertson and Paul D. Seymour. Graph minors XIII. The disjoint paths problem. J. Comb. Theory, Ser. B, 63(1):65–110, 1995. [17] Neil Robertson and Paul D. Seymour. Graph Minors. XIII. The disjoint paths problem. J. Combin. Theory, Ser. B, 63(1):65–110, 1995. [18] Neil Robertson and Paul D. Seymour. Graph minors. XX. Wagner’s conjecture. J. Combin. Theory Ser. B, 92(2):325–357, 2004. [19] Anneke A. Schoone, Hans L. Bodlaender, and Jan van Leeuwen. Diameter increase caused by edge deletion. Journal of Graph Theory, 11(3):409–427, 1987. [20] Paul D. Seymour and Robin Thomas. Call routing and the ratcatcher. Combinatorica, 14(2):217–241, 1994. 30	8cs.DS
A Centralized Local Algorithm for the Sparse Spanning Graph Problem Christoph Lenzen ∗ Reut Levi † arXiv:1703.05418v2 [cs.DS] 18 Jul 2017 July 20, 2017 Abstract Constructing a sparse spanning subgraph is a fundamental primitive in graph theory. In this paper, we study this problem in the Centralized Local model, where the goal is to decide whether an edge is part of the spanning subgraph by examining only a small part of the input; yet, answers must be globally consistent and independent of prior queries. Unfortunately, maximally sparse spanning subgraphs, i.e., spanning trees, cannot be constructed efficiently in this model. Therefore, we settle for a spanning subgraph containing at most (1 + )n edges (where n is the number of vertices and is a given approximation/sparsity parameter). We achieve a query complexity of Õ(poly(∆/ε)n2/3 ),1 where ∆ is the maximum degree of the input graph. Our algorithm is the first to do so on arbitrary bounded degree graphs. Moreover, we achieve the additional property that our algorithm outputs a spanner, i.e., distances are approximately preserved. With high probability, for each deleted edge there is a path of O(log n · (∆ + log n)/) hops in the output that connects its endpoints. 1 Introduction When operating on very large graphs, it is often impractical or infeasible to (i) hold the entire graph in the local memory of a processing unit, (ii) run linear-time (or even slower) algorithms, or even (iii) have only a single processing unit perform computations sequentially. These constraints inspired the Centralized Local model [17], which essentially views the input as being stored in a (likely distributed) database that provides query access to external processing units. To minimize the coordination overhead of such a system, it is furthermore required that there is no shared memory or communication between the querying processes, except for shared randomness provided alongside the access to the input. The idea is now to run sublinear-time algorithms that extract useful global properties of the graph and/or to examine the input graph locally upon demand by applications. Studying graphs in this model leads to the need for query access to a variety of graph-theoretical structures like, e.g., independent or dominating sets. In such a case, it is crucial that locally evaluating whether a node participates in such a set is consistent with the same evaluation for other nodes. This is a non-trivial task, as local decisions can only be coordinated implicitly via the structure of the input (which is to be examined as little as possible) and the shared randomness. ∗ MPI for Informatics, Saarbrücken 66123, Germany, clenzen@mpi-inf.mpg.de MPI for Informatics, Saarbrücken 66123, Germany, rlevi@mpi-inf.mpg.de 1 Õ-notation hides polylogarithmic factors in n. † 1 Nonetheless, this budding field brought forth a number of elegant algorithms solving, e.g., maximal independent set, hypergraph coloring, k-CNF, approximate maximum matching and approximate minimum vertex cover for bipartite graphs [1, 4, 5, 11, 12, 13, 17]. In this work, we consider another very basic graph structure: sparse spanning subgraphs. Here, the task is to select a subset of the edges of the (connected) input graph so that the output is still connected, but has only few edges. By “few” we mean that, for some input parameter ε > 0, the number of selected edges is at most (1 + ε)n, where n denotes the number of nodes. One may see this as a relaxed version of the problem of outputting a spanning tree of the graph, which is a too rigid requirement when looking for fast algorithms: on a cycle, a single edge has to be deleted, but this necessitates to first verify that the input graph is not, in fact, a line. Definition 1 ([9]) An algorithm A is a Local Sparse Spanning Graph (LSSG) algorithm if, given n, ∆ ≥ 1, a parameter ≥ 0, and query access to the incidence list representation of a connected graph G = (V, E) over n vertices and of degree at most ∆, it provides oracle access to a subgraph G0 = (V, E 0 ) of G such that: 1. G0 is connected. 2. \|E 0 \| ≤ (1 + ) · n with high probability (w.h.p.),2 where E 0 is determined by G and the internal randomness of A. By “providing oracle access to G0 ” we mean that on input {u, v} ∈ E, A returns whether {u, v} ∈ E 0 , and for any sequence of edges, A answers consistently with respect to the same G0 . We are interested in LLSG algorithms that, for each given edge, perform as few queries as possible to G. Observe that Item 2 implies that the answers of an LLSG algorithm to queries cannot depend on previously asked queries. We note that relaxing from requiring a tree as output makes it possible to ask for additional guarantees that, in general, cannot be met by a spanning tree. Instead of merely preserving connectivity, it becomes possible to maintain distances up to small factors. Such subgraphs are known as (sparse, multiplicative) spanners [14, 15]. In fact, choosing ε ∈ o(1) then yields ultra-sparse spanners that are o(n) edges away from being trees. 1.1 Our Contribution We give the first non-trivial LSSG algorithm in the Centralized Local model that runs on arbitrary graphs. We achieve a query complexity of Õ(poly(∆/ε)n2/3 ) per edge, w.h.p. Moreover, we guarantee that for each edge that is not selected into the spanner, w.h.p. there is a path of O(log n · (∆ + log n)/) hops consisting of edges that are selected into the spanner; this is referred to as a stretch of O(log n · (∆ + log n)/). For simplicity, assume for the moment that ∆ and ε are constants. Our algorithm combines the following key ideas. • We classify edges as expanding if there are sufficiently many (roughly n1/3 ) nodes within O(log n) hops of their endpoints. For non-expanding edges, we can efficiently simulate a standard distributed spanner algorithm at small query complexity, as solutions of running time O(log n) are known (e.g. [3]). 2 That is, with probability at least 1 − 1/nc for an arbitrary constant c > 0 that is chosen upfront. 2 • On the node set induced by expanding edges, we can construct a partition into Voronoi cells with respect to roughly n2/3 randomly sampled centers. The Voronoi cells are spanned by trees of depth O(log n), as expanding nodes have their closest center within O(log n) hops w.h.p. Finding the closest center has query complexity Õ(n1/3 ). • We refine the partition into Voronoi cells further into clusters of Õ(n1/3 ) nodes. We simply let a node be a singleton cluster if its subtree in the spanning tree of its cell contains more than Õ(n1/3 ) nodes. This construction has query complexity Õ(n2/3 ) for constructing a complete cluster, yet ensures that there are Õ(n2/3 ) clusters in total due to the low depth of the trees; moreover, each cluster is completely contained in some Voronoi cell. • It remains to select few edges to interconnect the Voronoi cells. This is the main challenge, for which the above properties of the partition are crucial. To keep the number of selected edges small in expectation, we mark a subset of expected size Θ̃(n1/3 ) of the clusters by marking each Voronoi cell (and thereby its constituent clusters) with probability n−1/3 . We then try to ensure that (i) clusters select an edge to each adjacent marked Voronoi cell and (ii) for each marked Voronoi cell adjacent to an adjacent cluster, they select one edge connecting to some cluster adjacent to this cell. • The main issue with the previous step is that we cannot afford to construct each adjacent cluster, preventing us from guaranteeing (ii). We circumvent this obstacle by identifying for adjacent clusters in which cell they are and keeping an edge for the purpose of (ii) if it satisfies a certain minimality requirement with respect to the rank of the cell used for symmetry breaking purposes. This way, we avoid construction of adjacent clusters, instead needing to determine the rank of their Voronoi cells only. This way, we maintain query complexity Õ(n2/3 ). • However, this now entails an inductive argument for ensuring connectivity, which also affects stretch. By choosing Voronoi cell ranks uniformly at random, we ensure that the length of such an inductive chain is bounded by O(log n) w.h.p. Together with the depth of the Voronoi cell trees of O(log n) and the stretch of the spanner algorithm for non-expanding edges (also O(log n)), this yields the total bound of O(log2 n) on the stretch of our scheme. Finally, we note that we can place the above routine in a wrapper verifying that, w.h.p., the number of globally selected edges does not significantly exceed the expectation. If this is not the case, the wrapper starts the process all over. Since in each attempt the success probability is constant (and the verifier succeeds w.h.p.), we get within O(log n) attempts that the bound on the number of selected edges is satisfied w.h.p. and the routine terminates. Relation to Property Testing. As observed in [10], testing cycle-freeness with one sidederror in the bounded-degree model can be reduced to the LSSG problem. From this reduction it follows that we obtain a tester for cycle-freeness that works in Õ(n2/3 ) time. Czumaj et al. [2] studied the problem of Ck -minor freeness with one sided-error. For cycle-freeness (C3 -minor freeness) √ their complexity is Õ( n), therefore their complexity is better than ours. However, we would like to point out that an LSSG algorithm gives a stronger guarantee than a one-sided error tester for cycle-freeness. An LSSG algorithm can be used to find for all but at most (1 + )\|V \| edges e a cycle that e belongs to, i.e., a witness for the violation by e can be provided. In constrast, a one-sided error tester merely guarantees to find a single cycle in instances that are -far from being cycle-free. Perhaps more importantly, our approach proves useful when testing for other minors. Recently, Fichtenberger et al. [6] built on our work to construct one-sided error testers for each minor of the 3 (k × 2)-grid. 1.2 Related work The problem of finding a sparse spanning subgraph in the Centralized Local model was first studied √ in [9], where the authors show a lower bound of Ω( n) queries for constant and ∆ (see also survey by Rubinfeld [18]). They also present an upper bound with nearly tight query complexity for graphs that have very good expansion properties. However, for general (bounded degree) graphs their algorithm might query the entire graph for completing a single call to the oracle. They also provide an efficient algorithm for minor-free graphs that was later improved in [8]. The algorithm presented in [8] achieves a query complexity that is polynomial in ∆ and 1/ and is independent of n. The stretch factor of this algorithm is also independent of n and depends only on ∆, 1/, and the size of the excluded minor. A characterization of the query complexity of the problem was presented in [7]. Specifically, this work provides an upper bound (which builds on an algorithm in [9]) that has a query complexity that is independent of n (however, super-exponential in 1/) for families of graphs which are, roughly speaking, sufficiently non-expanding everywhere. On the other hand, they show that, for a family of graphs with expansion properties that are slightly better, any local algorithm must have a query complexity that depends on n. In the (distributed) Local model, Ram and Vicari [16] study the same problem and provide an algorithm that runs in min{D, O(log n)} rounds, where D denotes the diameter of the input graph. Their algorithm achieves the sparsity property by breaking all short cycles. 2 Preliminaries The graphs we consider are undirected and have a known degree bound ∆, and we assume we have query access to their incidence list representation. Namely, for any vertex v and index 1 ≤ i ≤ ∆, it is possible to obtain the ith neighbor of v by performing a query to the graph (if v has less than i neighbors, then a special symbol is returned). Without loss of generality, we assume that graphs are simple, i.e., contain neither loops nor parallel edges.3 The number of vertices in the graph is n and we assume that each vertex v has a unique id, which for simplicity we also denote by v. There is a total order on the ids, i.e., given any two distinct ids u and v, we can decide whether u < v or v < u. Let G = (V, E) be a graph, where V = [n]. We denote the distance between two vertices u and v in G by dG (u, v). For vertex v ∈ V and an integer r, let Γr (v, G) denote the set of vertices at distance at most r from v. When the graph G is clear from the context, we shall use the shorthands d(u, v) and Γr (v) for dG (u, v) and Γr (v, G), respectively. The total order on the vertices induces a total order r on the edges of the graph in the following straightforward manner: r({u, v}) < r({u0 , v 0 }) if and only if min{u, v} < min{u0 , v 0 } or min{u, v} = min{u0 , v 0 } and max{u, v} < max{u0 , v 0 }. The total order over the vertices also induces an order over those vertices visited by a Breadth First Search (BFS) starting from any given vertex v, and whenever we refer to a BFS, we mean that it is performed according to this order. 3 The answer on a self-loop can always be negative, and we can default to rejecting all but the first edge between two nodes. 4 Figure 1: Partition of a graph into cells and clusters, where k = 6. Black lines are the borders of Voronoi cells, whose centers have black fillings. Red edges belong to the BFS trees spanning the clusters, while dashed gray lines are non-tree edges. Red circles indicate singleton clusters (if the node has a red child) or the roots of subtrees that form a cluster (if the children are black). Whenever referring to one of the above orders, we may refer to the rank of an element in the respective order. This is simply the index of the respective element when listing all elements ascendingly with respect to the order. For a graph G = (V, E) and a pair of disjoint subsets of vertices A ⊂ V and B ⊂ V let def EG (A, B) = {(u, v) ∈ E \| u ∈ A ∧ v ∈ B}. When it is clear from the context, we omit the subscript. We say that a pair of subsets of vertices A and B are adjacent if EG (A, B) 6= ∅. 3 An Algorithm that Works under a Promise We begin by describing an LSSG algorithm which works under the following promise on the input graph G = (V, E). Sample ` uniformly at random from [2 log n/ log(1 + ), 2 log n/ log(1 + ) + ∆/}], def and let k = cn1/3 ln n · `∆/, where c is a constant that will be determined later. For every v ∈ V , def let iv = minr {\|Γr (v)\| ≥ k}. We are promised that maxv∈V {iv } ≤ `. In words, we assume that the `-hop neighborhood of every vertex in G contains at least k vertices. First, we fix a simple partition of V . 5 3.1 The Underlying Partition def Centers. Pick a set S ⊂ \|V \| of r = Θ(n2/3 / ln n) vertices at random. We shall refer to the vertices in S as centers. For each vertex v ∈ V , its center, denoted by c(v), is the center which is closest to v amongst all centers (break ties between centers according to the id of the center). Voronoi cells. The Voronoi cell of a vertex v, denoted by Vor(v), is the set of all vertices u for which c(u) = c(v). Additionally, we assign to each cell a random rank, so that there is a uniformly random total order on the cells; note carefully that the rank of a cell thus differs from the rank of its center (which is given by its identifier, which is not assigned randomly). We remark that we can determine the rank of the cell from the shared randomness and the cell’s identifier, for which we simply use the identifier of its center. Clusters. For each Voronoi cell, consider the BFS tree spanning it, which is rooted at the respective center. For every v ∈ V , let p(v) denote the parent of v in this BFS tree. If v is a center then p(v) = v. For every v ∈ V \ S, let T (v) denote the subtree of v in the above-mentioned BFS tree when we remove the edge {v, p(v)}; for v ∈ S, T (v) is simply the entire tree. Now consider a Voronoi cell. If the cell contains at most k vertices, then the cluster of all the vertices in the Voronoi cell is the cell itself. Otherwise, there are two cases. If T (v) contains at least k vertices, then the cluster of v is the singleton {v}. Otherwise, v has a unique ancestor u (including v) for which T (u) contains less than k vertices and T (p(u)) contains at least k vertices. The cluster of v is the set of vertices in T (u). For a cluster C, let c(C) denote the center of the vertices in C (all the vertices in the cluster have the same center). Let Vor(C) denote the Voronoi cell of the vertices in C. This describes a partition of V into Voronoi cells, and a refinement of this partition into clusters. See Figure 1 for an illustration. 3.2 The Edge Set Our spanner, H = (V, E 0 ), initially contains, for each Voronoi cell Vor the edges of the BFS tree that spans Vor, i.e., the BFS tree rooted at the center of Vor spanning the subgraph induced by Vor (see Section 2 for more details). Clearly, these edges also span the clusters. Next, we describe which edges we add to E 0 in order to connect the clusters. Marked Clusters and Clusters-of-Clusters def Each center is marked independently with probability p = 1/n1/3 . If a center is marked, then we say that its Voronoi cell is marked and all the clusters in this cell are marked as well. Cluster-of-clusters. For every marked cluster, C, define the cluster-of-clusters of C, denoted by C(C), to be the set of clusters which consists of C and all the clusters which are adjacent to C. A cluster B is participating in C(C) if B ∈ C(C) and the edge of minimum rank in E(B, Vor(C)) also belongs to E(B, C). Thus, if B is adjacent to Vor(C) and Vor(C) is marked, then there is a unique cluster D ⊆ Vor(C) such that B participates in C(D). See Figure 2 for a visualization. 6 E F C(C) D C Figure 2: Illustration of marked clusters and clusters of clusters. Thick red and black ovals are marked and unmarked cells, respectively. Thin circles are clusters, where cluster C comprises its entire cell. Thick edges are the ones of minimum rank between their incident clusters, while the dotted edges do not meet this criterion. The arrows of red edges indicate participation in the respective adjacent marked cluster; note that D does not participate in C(F ), as for each adjacent marked cell Vor it exclusively participates in the cluster-of-clusters connected to it by the edge of minimum rank in E(D, Vor). C(C) is marked in blue; all its constituent clusters also participate in C(C), as Vor(C) = C. The Edges between Clusters By saying that we connect two adjacent subsets of vertices A and B, we mean that we add the def minimum ranked edge in E(A, B) to E 0 . For a cluster A, define its adjacent centers Vor(∂A) = {Vor(v) \| u ∈ A ∧ {u, v} ∈ E} \ {Vor(A)}, i.e., the set of Voronoi cells that are adjacent to A. This definition explicitly excludes Vor(A), as there is no need to connect A to its own Voronoi cell. We next describe how we connect the clusters. The high-level idea is to make sure that every marked cluster and the clusters that participate in the respective cluster-of-clusters remain connected. For clusters which are not adjacent to any marked cluster we make sure to keep them connected to all adjacent Voronoi cells. Formally: 1. We connect every cluster to every adjacent marked cluster. 2. Each cluster A that is not participating in any cluster-of-clusters (i.e., no cell adjacent to A is marked) we connect to each adjacent cell. 3. Suppose cluster A is adjacent to cluster B, where B is adjacent to a marked cell Vor. Denote by C the (unique) cluster in cell Vor for which B participates in C(C). We connect A with B if the following conditions hold: • Vor(B) has minimum rank amongst Vor(∂A) ∩ Vor(∂C) 7 C A B D Figure 3: Illustration of the third edge selection rule. In this example, the thick black edge has minimum rank in E(A, Vor(B)) and B has minimum rank in ∂A ∩ ∂C. However, the rank of Vor(D) is smaller than that of Vor(B), and hence the dashed edge is not selected. Here, A will select the direct edge connecting it to D, due to the first rule. • the minimum ranked edge in E(A, Vor(B)) is also in E(A, B) Figure 3 showcases the third rule. 3.3 Sparsity Lemma 1 The number of clusters, denoted by s, is at most \|S\| + n`(∆ + 1)/k. Proof: We first observe that, due to the promise on G, it follows that for every v ∈ V , d(v, c(v)) ≤ ` w.h.p. Recall the terminology from Subsection 3.1. Consider v for which \|T (v)\| ≥ k and therefore its cluster is the singleton {v}. We say that a vertex u is special if \|T (u)\| ≥ k and for every child of u in T (u), t, it holds that \|T (t)\| < k. By an inductive argument, it follows that v is an ancestor of a special vertex. Since for every pair of special vertices u and w, T (u) and T (w) are vertex disjoint, we obtain that there are at most n/k special vertices. Since for every special vertex, there are at most ` ancestors, the total number of vertices v with \|T (v)\| ≥ k is bounded by n`/k. Observe that any cluster either (i) is a singleton {v} with \|T (v)\| ≥ k, (ii) contains a node v such that \|T (p(v))\| ≥ k, or (iii) is an entire Voronoi cell. We just bounded the number of clusters of type (i) by n`/k, and immediately get a bound on the number of type (ii) clusters of n`∆/k. The number of type (iii) clusters is bounded by the number of Voronoi cells \|S\|, showing the desired bound on s. Lemma 2 Exp(\|E 0 \|) ≤ (1 + O())\|V \|. 8 Proof: The number of edges we add due to the BFS trees of the Voronoi cells is at most \|V \| − 1. The number of edges which are taken due to Condition 1 is at most s times the number of marked clusters. In expectation, there are s · p marked clusters, yielding at most s2 p edges in expectation. Since s = O(n2/3 / ln n) and p = 1/n1/3 we obtain that s2 p = O(n/ ln n). Let A be a cluster. The number of edges which are adjacent to A and are taken due to Condition 3 is bounded by the total number of clusters-of-clusters. The number of clusters-of-clusters is exactly the number of marked clusters. Thus, the total number of edges which are taken due to Condition 3 is bounded by s2 p. Observe that the probability that cluster A is not adjacent to a marked cell is (1 − p)\|Vor(∂A)\| ≤ e−p\|Vor(∂A)\| . Hence, if \|Vor(∂A)\| ≥ 3p−1 ln n, A is adjacent to a marked cell w.h.p. Using a union bound over all clusters, it follows that w.h.p. each cluster A without an adjacent marked cell satisfies that \|Vor(∂A)\| ≤ 3p−1 ln n with probabilty at least 1 − 1/n2 ; the probability at most 1/n2 event that this bound is violated cannot contribute more than \|E\|/n2 < 1 to the expectation. Therefore, the total number of edges which are taken due to Condition 2 is bounded by (3s ln n)/p + 1 = O(n). To conclude, the total number of edges in E 0 is at most \|V \|(1 + O()) in expectation, as desired. 3.4 Connectivity and Stretch Lemma 3 H is connected. Proof: Recall that H contains a spanning tree on every Voronoi cell, hence it suffices to show that we can connect any pair of Voronoi cells by a path between some of their vertices. Moreover, the facts that G is connected and the Voronoi cells are a partition of V imply that it is sufficient to prove this for any pair of adjacent Voronoi cells. Accordingly, let Vor and Vor1 be two cells such that E(Vor, Vor1 ) 6= ∅. Consider clusters A ⊆ Vor and B ⊆ Vor1 such that the edge e of minimum rank in E(Vor, Vor1 ) is in E(A, B). If B is not adjacent to a marked cell, then Condition 2 implies that e is selected into H. Thus, we may assume that B is adjacent to a marked cell Vor0 . Accordingly, there exists a marked cluster C ⊆ Vor0 such that B is participating in C(C). If the rank of Vor1 is minimum in Vor(∂C) ∩ Vor(∂A), then e is selected into H by Condition 3 and we are done. Otherwise, observe that Vor1 is connected to Vor0 , as the edge of minimum rank in E(B, C) is selected into H by Condition 1. Therefore, it suffices to show that Vor gets connected to Vor0 . Let Vor2 be the cell of minimum rank among Vor(∂C) ∩ Vor(∂A). Let D ⊆ Vor2 be the cluster satisfying that the edge e0 of minimum rank in E(A, Vor2 ) is in E(A, D). Note that Vor2 is connected to Vor0 (which we saw to be connected to Vor1 ), as there is some cluster D0 ⊆ Vor(D) that is adjacent to C and selects the edge of minimum rank in E(D0 , C) by Condition 1. Overall, we see that it is sufficient to show that Vor gets connected to Vor2 , where Vor2 has smaller rank than Vor1 . We now repeat the above reasoning inductively. In step i, we either succeed in establishing connectivity between Vor and Vori , or we determine a cell Vori+1 that has smaller rank than Vori and is connected to Vori . As any sequence of Voronoi cells of descending ranks must be finite, the induction halts after finitely many steps. Because the induction invariant is that Vori+1 is connected to Vori , this establishes connectivity between Vor and Vor1 , completing the proof. 9 Lemma 4 Denote by GVor the graph obtained from G by contracting Voronoi cells and by HVor its subgraph obtained when doing the same in H. If the cells’ ranks are uniformly random, w.h.p. HVor is a spanner of GVor of stretch O(log n). Proof: Recall the proof of Lemma 3. We established connectivity by an inductive argument, where each step increased the number of traversed Voronoi cells by two. Hence, it suffices to show that the induction halts after O(log n) steps w.h.p. To see this, observe first that GVor is independent of the ranks assigned to Voronoi cells and pick any pair of adjacent cells Vor, Vor1 , i.e., neighbors in GVor . We perform the induction again, assigning ranks from high to low only as needed in each step, according to the following process. In each step, we query the rank of some cells, and given an answer of rank r, the ranks of all cells of rank at least r are revealed as well. In step i, we begin by querying the rank of Vori . Consider the cluster Di ⊆ Vori adjacent to A satisfying that the edge with minimum rank in E(Vori , A) is also in E(Di , A). We can assume without loss of generality that Di is adjacent to a marked cluster Fi and that it is participating in C(Fi ) (as otherwise Di connects to A directly and we can terminate the process). If the ranks of all the cells adjacent to both Fi and A were already revealed, then the process terminates. Otherwise, we query the rank of all such cells whose rank is still unrevealed. We set the cell of the queried cluster that has minimum rank to be Vori+1 and we continue to the next step. We claim that, in each step i, either the process terminates, or the rank of Vori+1 is at most half of the rank of Vori with probability at least 1/2. To verify this, observe that in the beginning of step i, any cell center whose rank was not revealed so far has rank which is uniformly distributed in [ri − 1], where ri is the rank of Vori .4 With probability at least 1/2, such a rank is at most ri /2. If Vori has no adjacent cells whose ranks have not been revealed yet, the process terminates. Hence, regardless of whether the process terminates or not, the claim holds. By Chernoff’s bound, we conclude that the process terminates within O(log n) steps w.h.p., as r1 is bounded by the number of Voronoi cells, which itself is trivially bounded by n. By the union bound over all pairs of cells Vor and Vor1 , we get the desired guarantee. Corollary 1 W.h.p., H is a spanner of G of stretch O(log n · (∆ + log n)/). Proof: Due to the promise on G, w.h.p. the spanning trees on Voronoi cells have depth ` ∈ O((∆ + log n)/). Hence, for any edge within a Voronoi cell, the claim holds w.h.p. Moreover, for an edge connecting different Voronoi cells, by Lemma 4, w.h.p. there is a path of length O(log n) in HVor connecting the respective cells. Navigating with at most 2` ∈ O((∆ + log n)/) hops in each traversed cell, we obtain a suitable path of length O(log n · (∆ + log n)/) in H. 4 The Algorithm for General Graphs We use a combination of the algorithm in Section 3 with the algorithm by Elkin and Neiman for ultra-sparse spanners [3]. We call a vertex v remote, with respect to a set of centers, if the `-hop neighborhood of v does not include a center. Fix S, let R denote the set of remote vertices with def respect to S, and abbreviate R̄ = V \ R. 4 In step 1, we first query Vor1 and then observe that this statement holds. 10 First Step. Run the algorithm from Section 3 on the subgraph induced by R̄, i.e., {u, v} ∈ E with u, v ∈ R̄ is added to E 0 if and only if the algorithm outputs the edge. Second Step. Run the algorithm of Elkin and Neiman [3] on the subgraph induced by R, i.e., {u, v} ∈ E with u, v ∈ R is added to E 0 if and only if the algorithm outputs the edge.5 Their algorithm proceeds as follows. Given an integer h, each vertex v draws rv according to the exponential distribution with parameter β = ln(n/δ)/h, where δ is a parameter that controls the success probability of the algorithm. Each vertex v receives ru from every vertex u within distance h, and stores mu (v) = ru −d(u, v) and a neighbor on a shortest path between v and u, denoted by nu (v). The edges that are added to the spanner are C(v) = {{v, nu (v)} \| mu (v) ≥ maxw∈V {mw (v) − 1}}, for every v ∈ R. With probability at least 1 − δ, it holds that rv < h for all v ∈ V (Claim 2.3 in [3]). We choose δ = 1/nc (i.e., the algorithm succeeds w.h.p.) and h = `, where w.l.o.g. we require c ≥ 1. The following lemma implies that the total number of edges that we add to H in the second step is at most \|R\| · (n2 )1/` ≤ \|R\|(1 + ) in expectation. Lemma 5 (Proof of Lemma 2.2 in [3]) For every v ∈ R, Exp[C(v)] ≤ (n/δ)1/h . Third Step. Add to E 0 all edges e ∈ E(R, R̄). The following lemma implies that the expected number of edges which are added in the third step is at most n. Lemma 6 Exp[\|E(R, R̄)\|] ≤ n. Proof: Observe that for an edge {u, v} ∈ E, there is at most one integer r such that Γr (u) ∩ S = ∅ and Γr (v) ∩ S 6= ∅ (or vice versa). If there is no such r or ` 6= r, then the edge is not in E(R, R̄). Over the random choice of `, the probability of the event that the edge is included is at most Pr[` = r] ≤ /∆. The desired bound follows by linearity of expectation. 4.1 Stretch Factor Consider any edge e = {u, v} ∈ E \ E 0 we removed. If u, v ∈ R̄, e was removed by the Algorithm from Section 3, which was applied to the subgraph induced by R. Applying Lemma 4 to the connected component of e, we get that w.h.p. there is a path of length O(log n · (∆ + log n)/) from u to v in H. If u, v ∈ R, by Lemma 5 and the choice of parameters, w.h.p. there is a path of length O(log n) from u to v in H. As e ∈ / E(R, R̄) by the third step, we arrive at the following corollary. Corollary 2 The above algorithm guarantees stretch O(log n · (∆ + log n)/) w.h.p. and satisfies that Exp[\|E 0 \|] ∈ (1 + O())n. 5 The Local Algorithm In this section we prove the our main theorem. 5 The algorithm is described for connected graphs; we simply apply it to each connected component of R. 11 Algorithm 1 LSSG for general graphs Input: {u, v} ∈ E Output: whether {u, v} is in E 0 or not. 1. If u, v ∈ R, compute the output of algorithm of Elkin and Neiman at u and v when running it on the connected component of u and v in the subgraph induced by R. Return true if {u, v} ∈ C(u) ∪ C(v) and false otherwise. 2. If {u, v} ∈ E(R, R̄), return true. 3. Otherwise, u, v ∈ R̄ and we proceed according to Section 3, where all nodes in R are ignored: (a) If Vor(u) = Vor(v), return true if {u, v} is in the BFS tree of Vor(u) and false otherwise. (b) Otherwise, let Q and W denote the clusters of u and v, respectively. Return true if at least one of the following conditions hold for A = Q and B = W , or symmetrically, for A = W and B = Q, and false otherwise. i. A is a marked cluster and {u, v} has minimum rank amongst the edges in E(A, B). ii. A is not participating in any cluster-of-clusters. Namely, all the clusters which are adjacent to A are not marked. In this case, we take {u, v} if it has minimum rank amongst the edges in E(A, Vor(B)). iii. There exists a marked cluster C such that A is participating in C(C), and the following holds: • Vor(A) has minimum rank amongst Vor(∂B) ∩ Vor(∂C) • {u, v} has minimum rank amongst the edges in E(B, Vor(A)). Theorem 3 Algorithm 1 is an LSSG algorithm. For any graph G over n vertices of maximum degree at most ∆ and > 0, its query complexity, space complexity (length of the random seed), and running time are Õ(n2/3 ) · poly(1/, ∆). Proof: The correctness of the algorithm follows from the previous sections. We shall prove that its complexity is as claimed. We analyze the complexity in terms of n. There are additional factors that depend polynomially in ∆ and 1/. The following claims hold w.h.p., simultaneously, for all vertices. Recall that a vertex u is remote if Γ` (u) does not contain a center. Due to the sampling probability for centers, a center is found after exploring Õ(n1/3 ) vertices. Therefore, we can decide for any vertex u whether it is in R with query and time complexity Õ(n1/3 ). Moreover, if u ∈ R̄, without additional cost the respective subroutine can return c(u), the center of u, and d(u, c(u)) (as we explore in a BFS fashion). For Step 1, we need to determine Γ` (x) ∩ R. Since \|Γ` (u)\| = Õ(n1/3 ) for every vertex u ∈ R, we obtain that the query and time complexity of this step is Õ(n2/3 ), in total. Accordingly, Step 2 has query and time complexity Õ(n1/3 ). If u, v ∈ R̄, the algorithm proceeds as in Section 3. We claim that reconstructing the clusters of u and v, determining the centers of all adjacent nodes, and deciding whether {u, v} is a BFS edge (i.e., Vor(u) = Vor(v) and u is the parent of v in the BFS of Vor(u) rooted at c(u) or vice versa) takes Õ(n2/3 ) queries and time. 12 We show first that we can reconstruct clusters efficiently. W.l.o.g., consider u. We determine c(u) and d(u, c(u)). For each neighbor w of u, we determine whether it is in R̄ (if not, the node is discarded), its center c(w), and the distance d(w, c(w)). As u ∈ R̄, and assuming that u = 6 c(u), it must have at least one neighbor w in distance d(u, c(u)) − 1 of c(u); any such w satisfies that c(w) = c(u), as otherwise d(u, c(w)) = d(w, c(w)) and c(w) < c(u), a contradiction to the tiebreaking rule for centers. Among these candidates w, we know that the one with minimum rank is the parent of u in the BFS tree of Vor(u) rooted at c(u), due to the tie-breaking rule for the BFS construction. Otherwise, we can use this subroutine to partially explore the BFS of Vor(u): given any node w ∈ R̄, we can determine its parent and children in Vor(w) at query and time complexity Õ(n1/3 ). We conclude that we can determine p(w) and whether T (w) < k or T (w) ≥ k at query complexity Õ(n2/3 ), by partially or completely exploring T (w); if T (w) < k, we determine T (w) completely. We collect this information for u and its ` − 1 ancestors, and determine the cluster Q of u. Finally, we repeat the procedure to reconstruct the cluster W of v, and determine for all nodes adjacent to either cluster whether they are in R̄ and, if so, their centers. The query and time complexity of this operation is Õ(n2/3 ) in total. For the cases A = Q, B = W and A = W , B = Q, respectively, with the above information we can determine • whether Vor(u) = Vor(v), • if Vor(u) = Vor(v), whether p(u) = v or p(v) = u, • if Vor(u) 6= Vor(v), – whether A is marked (i.e., c(A) has been marked) and whether {u, v} has minimum rank in E(A, B), – whether A is not adjacent to any marked cluster (i.e., none of the adjacent nodes’ centers has been marked) and whether {u, v} has minimum rank in E(A, Vor(B)), and – whether there is a marked cluster C adjacent to A so that A participates in C(C), Vor(A) has minimum rank in Vor(∂B) ∩ Vor(∂C), and {u, v} has minimum rank in E(B, Vor(A)). We note that since we have degree bounded by ∆, the number of vertices in A is Õ(n1/3 ), and the probability that a cell is marked is n−1/3 , the number of cluster-of-clusters that A participates in is Õ(∆) w.h.p. In other words, we can perform all necessary checks to decide whether {u, v} ∈ E 0 or not. The algorithm in Section 3 requires Õ(n2/3 ) random bits for the selection of centers and marked clusters. For the emulation of the algorithm of Elkin and Neiman it suffices that the random variables {ru } be Õ(n1/3 )-wise independent, because the outcome of the algorithm for a vertex v ∈ R depends only on the random variables of at most Õ(n1/3 ) vertices (the vertices in Γ` (v)). Thus, overall Õ(n2/3 ) random bits (up to poly(∆/) factors) are sufficient. References [1] N. Alon, R. Rubinfeld, S. Vardi, and N. Xie. Space-efficient local computation algorithms. In Proceedings of the Twenty-Third Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1132–1139, 2012. 13 [2] Artur Czumaj, Oded Goldreich, Dana Ron, C. Seshadhri, Asaf Shapira, and Christian Sohler. Finding cycles and trees in sublinear time. Random Structures and Algorithms, 45(2):139–184, 2014. [3] Michael Elkin and Ofer Neiman. Efficient algorithms for constructing very sparse spanners and emulators. In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2017, Barcelona, Spain, Hotel Porta Fira, January 16-19, pages 652–669, 2017. [4] G. Even, M. Medina, and D. Ron. Deterministic stateless centralized local algorithms for bounded degree graphs. In Algorithms - ESA 2014 - 22th Annual European Symposium, Wroclaw, Poland, September 8-10, 2014. Proceedings, pages 394–405, 2014. [5] U. Feige, Y Mansour, and R. E. Schapire. Learning and inference in the presence of corrupted inputs. In Proceedings of The 28th Conference on Learning Theory, COLT 2015, Paris, France, July 3-6, 2015, pages 637–657, 2015. [6] Hendrik Fichtenberger, Reut Levi, Yadu Vasudev, and Maximilian Wötzel. On testing minorfreeness in bounded degree graphs with one-sided error. Unpublished manuscript, 2017. [7] R. Levi, G. Moshkovitz, D. Ron, R. Rubinfeld, and A. Shapira. Constructing near spanning trees with few local inspections. Random Structures & Algorithms, pages n/a–n/a, 2016. [8] R. Levi and D. Ron. A quasi-polynomial time partition oracle for graphs with an excluded minor. ACM Trans. Algorithms, 11(3):24:1–24:13, 2015. [9] R. Levi, D. Ron, and R. Rubinfeld. Local algorithms for sparse spanning graphs. In Proceedings of the Eighteenth International Workshop on Randomization and Computation (RANDOM), pages 826–842, 2014. [10] R. Levi, D. Ron, and R. Rubinfeld. Local algorithms for sparse spanning graphs. CoRR, abs/1402.3609, 2014. [11] R. Levi, R. Rubinfeld, and A. Yodpinyanee. Local computation algorithms for graphs of non-constant degrees. Algorithmica, pages 1–24, 2016. [12] Y. Mansour, A. Rubinstein, S. Vardi, and N. Xie. Converting online algorithms to local computation algorithms. In Automata, Languages and Programming: Thirty-Ninth International Colloquium (ICALP), pages 653–664, 2012. [13] Y. Mansour and S. Vardi. A local computation approximation scheme to maximum matching. In Proceedings of the Sixteenth International Workshop on Approximation Algorithms for Combinatorial Optimization Problems (APPROX), pages 260–273, 2013. [14] D. Peleg and A. A. Schäffer. Graph spanners. Journal of Graph Theory, 13:99–116, 1989. [15] D. Peleg and J. D. Ullman. An optimal synchronizer for the hypercube. SIAM Journal on Computing, 18:229–243, 1989. [16] L. S. Ram and E. Vicari. Distributed small connected spanning subgraph: Breaking the diameter bound. Technical report, Zürich, 2011. 14 [17] R. Rubinfeld, G. Tamir, S. Vardi, and N. Xie. Fast local computation algorithms. In Proceedings of The Second Symposium on Innovations in Computer Science (ICS), pages 223–238, 2011. [18] Ronitt Rubinfeld. Can we locally compute sparse connected subgraphs? In Computer Science Theory and Applications - 12th International Computer Science Symposium in Russia, CSR 2017, Kazan, Russia, June 8-12, 2017, Proceedings, pages 38–47, 2017. 15	8cs.DS
An Adaptive Sampling Scheme to Efficiently Train Fully Convolutional Networks for Semantic Segmentation Lorenz Berger1,2 , Eoin Hyde1,2 , M. Jorge Cardoso1 , and Sébastien Ourselin1 arXiv:1709.02764v4 [cs.CV] 22 Dec 2017 1 TIG, CMIC, University College London, London, UK 2 Innersight Labs, London, UK Abstract Deep convolutional neural networks (CNNs) have shown excellent performance in object recognition tasks and dense classification problems such as semantic segmentation. However, training deep neural networks on large and sparse datasets is still challenging and can require large amounts of computation and memory. In this work, we address the task of performing semantic segmentation on large data sets, such as three-dimensional medical images. We propose an adaptive sampling scheme that uses a-posterior error maps, generated throughout training, to focus sampling on difficult regions, resulting in improved learning. Our contribution is threefold: 1) We give a detailed description of the proposed sampling algorithm to speed up and improve learning performance on large images. 2) We propose a deep dual path CNN that captures information at fine and coarse scales, resulting in a network with a large field of view and high resolution outputs. 3) We show that our method is able to attain new state-of-the-art results on the VISCERAL Anatomy benchmark. 1 Introduction This paper addresses the problem of efficiently training convolutional neural networks (CNNs) on large and imbalanced datasets. We propose a training strategy that adaptively samples the training data to effectively speed up training and avoid over-sampling data that contains little extra information. In this work, we investigate the problem of automatic segmentation from high resolution 3D CT scans. Several deep learning techniques [1–4] have recently been proposed for 3D segmentation of medical datasets. To overcome the problem of dealing with these large datasets, such as Computed Tomography (CT) volumes, commonly of dimension 512 × 512 × 700, previous approaches train a CNN on a cropped region of interest which reduces the size of individual training images by around 100 fold [2, 4]. By reducing the size of training images, they can now be fit into memory and a network can be trained effectively on the selected data. However, identifying regions of interest requires an additional pre-processing step which may not be easy in many applications. Also, training CNNs on cropped images limits the field of view of the CNN and subsequently 2 Lorenz Berger et al. can introduce unwanted image boundary induced effects during testing. Other applications, where training CNNs on very large images is a problem, includes the segmentation of histology datasets [5] or the segmentation of aerial images. For example in aerial image segmentation, training a CNN to segment ships [6] can be difficult because large portions of the image contain water which provide little information during training, resulting in slow learning. Some ideas to address this have already been proposed, for example in [3] a fixed, handcrafted, pre-computed weight map is used to help learn small separation borders between touching cells for biomedical image segmentation. In this work the proposed sampling scheme ends up dynamically learning such a weight mapping, making it generally applicable to many applications. Curriculum learning [7] and derivative methods like self-paced learning [8] build on the intuition that, rather than considering all samples simultaneously, the algorithm should be presented with the training data in a meaningful order that facilitates learning. These ideas have already successfully been applied to image classification [9, 10], by ordering images from easy to hard during training. Also for the problem of weakly supervised semantic segmentation [11] similar ideas are applied, where predictions from previous training iterations are used to iteratively learn segmentation maps from just a single class label per image. The focus of this paper is fully supervised semantic segmentation where a representative training set is available with dense manual label annotations and the challenge lies in efficiently learning from this large datasets. We give a detailed account of the implementation, which is a straightforward extension to any existing CNN segmentation system, and present state-of-the-art segmentation results on the VISCERAL anatomy benchmark. 2 2.1 Methods Neural Network Architecture For the dual path network architecture we build on several previous ideas [1, 12, 13]. Compared to the 3D network outlined in [1], we further develop the architecture by replacing the standard convolution layers with popular resnet blocks [14], and increase the maximum network depth from 11 layers in [1] to 19 layers. By having a deeper network and a down sampled pathway with input resolution 1/4 of the original resolution, we obtain a large receptive field of size 1243 whilst maintaining a deep high resolution pathway that does not compromise the resolution through pooling layers. The architecture results in a total of 649,251 parameters. A sketch of the architecture is given in Figure 1a. In Figure 1a, numbers inside round brackets give the input dimensions of each block. For the training stage these dimensions have been chosen carefully to balance memory usage with processing speed. During testing the dimensions may be chosen as large as can be fit into memory, to take advantage of the fully convolutional inference. Numbers in square brackets refer to the number of feature maps used at each layer. The proposed configuration allows for a large number of samples isample: Adaptive Sampling for Fully Convolutional Networks 3 (3D patches) per batch to ensure balanced class sampling and effective optimization, whilst maintaining a deep and wide enough network to capture the high variability and spatial semantics of the data. The blocks labeled ‘Conv’ are standard convolutional layers with kernel size 3 × 3 × 3. The blocks labeled ‘Res block’ are standard and bottleneck resnet blocks, respectively, as detailed in [14]. Each fully connected layer is preceded by a dropout layer with probability 0.5, and a softmax non-linearity is used as a final classification layer. The rationale for having a deep low resolution path is to further increase the receptive field and allow for complex higher level features to be learned i.e where an organ is positioned in relation to other structures. To minimize the memory footprint, the high definition path is chosen to be slightly shorter than the low resolution path. This seems reasonable as this path should learn texture information which is likely to require fewer layers and non-linearites. Further details on training hyper-parameters are given in section 3.1. (a) (b) Fig. 1: (a) The proposed dual path CNN architecture. (b) Coronal slice of a CT scan with overlaid segmentation output, described in section 3.3. The organs visible in this slice are: lungs (green), liver (red), spleen (light blue), psoas major muscle (dark blue), kidneys (brown) and bladder (yellow). 4 Lorenz Berger et al. 2.2 Adaptive sampling strategy The problem of class imbalance as described in [1] can be dealt with by choosing small patch sizes and evenly sampling from each class [1], and through weighted loss functions [3, 15, 16]. Both of these methods either load the whole image into GPU memory, which is not feasible for large images, or select a small subset of patches, which can lead to inefficient training on sparse datasets. To overcome both of these issues we propose the simple sampling Algorithm 1. Algorithm 1 isample: adaptive sampling algorithm Initialize error maps for every image in the training data: E i (x) = 1. while CNN training do while training for 1 epoch do while filling batch with patches do Pick an image I j from the training set I ∗ . Pick a class k from the corresponding label map Lj . Pick a patch in image I j , centered at location c, where Lj (c) = k. Accept patch into batch if E i (c) > U(0, 1) − . end while Back-propagate loss of batch and update the current CNN weights: w. end while Select a subset of images, I ∗ , and label maps, L∗ , from the training set: for [I k , Lk ] ∈ [I ∗ , L∗ ] do Update error maps: E k (x) = 1 − CNN(w, I k (x))Lk (x) end for end while In Algorithm 1, U(0, 1) is a random number drawn from the uniform distribution and E i refers to the error map of the ith training image. Error maps can easily be calculated, either after each epoch or concurrently to the training process, as E k (x) = 1 − CNN(w, I k (x))Lk (x) , (1) where CNN(w, I k (x))Lk (x) is a map of the CNN predictions over the full training image I k , evaluated using the most current weights, w, and outputting the probability of the true class label Lk (x), at position x. Examples of error maps produced throughout training are show in Figure 3. The additional parameter controls the strength of the isample scheme. Setting = 0, corresponds to choosing patches based entirely on the amount of error that they currently produce by the network. When = 1 the condition E i (c) > U(0, 1) − is always satisfied and we are left with a standard sampling scheme that accepts every chosen patch. For all results shown in this paper we have chosen = 0.01, since we are interested in using the isample scheme to full effect. Detailed investigations into how to best set this parameter for different datasets with varying amounts of sparsity is left for future work. The subset of images, I ∗ , and label maps, L∗ , of the training set may be chosen in line with how quickly to introduce the isample scheme during training and the amount of computational resources available. In isample: Adaptive Sampling for Fully Convolutional Networks 5 our experiments we had access to four GPUs, three were used to train the CNN continuously and one GPU was used in parallel to continuously perform full predictions of the validation dataset and the training dataset. From this, full dice scores of the validation dataset and full error maps of the training dataset could be calculated. A future extension of this work could log the loss of individual batches during training, and use this information to get an estimate of where the network is having difficulty. This modification would avoid the need for having to allocate additional resources for segmenting the training data, which instead could be used to speed up training. However for practical reasons, and the increasing availability of large multi-GPU cards, we have found this not to be an issue. Also, having access to dice scores calculated over the full images throughout training has been helpful in development since these true dice scores provide more meaningful information than dice scores calculated from individual batches which are biased by the sampling scheme and the size of patches. The error maps produced can also be useful for debugging purposes during development. 3 Results We trained, validated and tested the automatic segmentation method on contrast enhanced CT scans from the VISCERAL Anatomy 3 dataset, made up of 20 training scans, and 10 unseen testing scans (currently not available to download) [17]. The scans are form a heterogeneous dataset with various topological changes between patients, and manual segmentations are available for a number of different anatomical structures. We randomly split the training set into 16 scans for training (80%) and 4 scans for validation (20%), we also present results of our online submission on the unseen test dataset. For illustrative purposes, the first experiment, in section 3.2, focuses on segmenting only the kidneys from full body CT scans. In section 3.3 we present results on simultaneously segmenting multiple organs from the CT data. 3.1 CNN training setup During training we perform data augmentation by re-sampling the 3D patches to a [1mm, 1mm, 1.5mm] + U(−0.1, 0.1) resolution. We also rotate each patch by [U(−10, 10), U(−4, 4), U(−4, 4)] degrees. We set voxels with values greater than 1000 to 1000, and values less than −1000 to −1000, and divide all values by a constant factor of 218 (the standard deviation of the dataset). We use Glorot initializations [18] on all convolution layers. For batchnorm layers we use the initializations technique described in [19]. We impose L2 weight decay of size 0.0001, on all convolutional layers except on the last fully convolutional layer before the final softmax non-linearity. Using techniques described in [19] we make use of large batch sizes and large learning rates. We use SGD with Nestrov momentum set at 0.8 [20], the initial learning rate is 0.001, and each batch contains 12 patches, sampled from one randomly selected scans in the training set. We run each epoch for 100 batches. We also employ a learning rate 6 Lorenz Berger et al. warm up schedule as described in [19] for the first 5 epochs. We use a standard cross-entropy loss function. 3.2 Segmenting the kidneys from full body CT scans In this experiment we use labels for kidneys to train the CNN, resulting in a simple two class, foreground (kidneys) and background (everything else), segmentation problem. Figure 2 shows curves of the training loss (2a) and mean validation dice score (2b) for segmented kidneys throughout training, averaged over three separate runs. The blue curve represents training runs where patches are sampled randomly but evenly from background and kidney foreground, the red curve represents training runs where patches are sampled using the proposed sampling algorithm 1. Because isample adaptively selects more difficult patches as training progresses, the loss is higher, as seen in Figure 2a. In Figure 2b, the sampler achieves faster generalization, and our current results indicate that the final generalization of the CNN trained with the proposed sampling scheme is slightly improved for this sparse segmentation setup, where the kidneys only make up ∼ 0.3% of the voxels within the whole scan. Table 4a shows the average dice scores achieved by the Dual CNN, with and without isample, throughout training. When using the isample scheme the CNN is able to achieve a dice score of 0.855 after only 5k training iterations. This is close to the end of training performance, a dice score of 0.899 after 40k, of the Dual CNN without isample in use. Fast training of CNNs using isample can be useful for debugging and evaluating changes in neural network architectures, as more experiments can be run using the same computational resources, that quickly estimate the end performance of the network. 1.4 Dual CNN Dual CNN + isample 1.0 0.8 0.6 0.4 0.9 Mean Validation Dice Training Loss 1.2 0 100 200 Epoch (a) 300 400 0.8 0.7 0.6 0.5 0.4 0 100 Dual CNN Dual CNN + isample 200 300 400 Epoch (b) Fig. 2: (a) Training loss (a) and mean validation dice scores (b), averaged over 3 runs. In Figure 3 we show coronal slices of a training error volume E k (x), calculated by the algorithm. As seen in Figure 3c, initially there is significant produced isample: Adaptive Sampling for Fully Convolutional Networks 7 by the CNN prediction at epoch 16, for example misclassifying the aorta (part of the background class) because it has similar intensity values to the kidneys. After more training, at epoch 50, Figure 3d shows that the error is much lower. The CNN has now learned that the aorta is part of the background class. However more subtle regions such as the collecting system and large vessels within the kidney (see small hole in the true segmentation of the left kidney in Figure 3b) still produce high errors, and further focused training is required to optimize the weights until they are correctly classified. There also remains a high error around the border of the kidneys, which will result in the sampling process selecting more patches from the border region, and thus ends up learning to train the network with a similar loss to the hand-crafted border weighted loss function designed in [3]. (a) (b) (c) (d) Fig. 3: Coronal slices: (a) Raw CT scan from the training set (b) Kidney segmentation overlaid onto scan (c) Error map, E k (x), of foreground and background classification on a training scan after 16 epochs. (d) E k (x) after 50 epochs. For the error maps, white corresponds to voxels that are incorrectly classified and black to correctly classified voxels. # its 5k 10k 20k 40k Dual CNN 0.797 0.849 0.905 0.899 (a) + isample 0.855 0.897 0.920 0.927 Method Kidney Dice scores Dual CNN (validation data) Left + Right 0.899 (0.058) Dual CNN + isample (validation data) Left + Right 0.927 (0.037) Dual CNN + isample + CRF (test data) Left 0.954 Right 0.96 Wang et al [21] (test data) Left 0.945 (0.027) Right 0.959 (0.011) Vincent et al [22] (test data) Left 0.943 (0.015) Right 0.927 (0.040) Gass et al [23] (test data) Left 0.913 (0.029) Right 0.914 (0.027) (b) Fig. 4: (a) Mean dice scores and standard deviations at different number of iterations, throughout training. (b) Dice scores and standard deviations, where available, for different methods automatically segmenting kidneys on the VISCERAL CT enhanced dataset. Table 4b shows dice scores for segmenting both kidneys using different methods. The proposed method with isample performs significantly better than without. We also submitted our method, with the addition of a CRF [24] as a postprocessing step, to segment the test dataset, and achieved the top score for 8 Lorenz Berger et al. segmenting the left and right kidneys. Inference on a full size CT scans takes ∼ 65 seconds using four Tesla K50 GPU cards, each with 4GB of RAM. 3.3 Multi-organ segmentation We now extended the previously described algorithm to include a multi-class classification output and trained the model on the main organs available on the VISCERAL CT-enhanced dataset. We post-processes the output segmentation maps (maximum class probability at each voxel), by applying a filter that only retains the largest connected binary object within the segmentation, thus removing small objects. The segmentation output of one of the the validation scans is shown in Figure 1b. The results of our proposed method and other state-of-the-art methods, also summarized in [17], are given in Table 1. Method Dual CNN + isample (val) Dual CNN (val) Ga et al [23] Jimenez et al [25] Kéchichian et al [26] Vincent et al [22] Inter-annotator agreement Aorta 0.893 0.843 0.785 0.762 0.681 0.838 0.859 Lung 0.980 0.985 0.963 0.961 0.966 0.972 0.973 Kidney 0.938 0.934 0.914 0.899 0.912 0.935 0.917 PMajor 0.824 0.779 0.813 0.797 0.802 0.869 0.823 Liver 0.941 0.927 0.908 0.887 0.933 0.942 0.965 Abdom 0.769 0.755 0.463 0.538 0.673 Spleen 0.951 0.946 0.781 0.730 0.895 0.934 Sternum 0.900 0.904 0.635 0.721 0.713 0.810 Trachea 0.926 0.926 0.847 0.855 0.824 0.877 Bladder 0.912 0.918 0.683 0.679 0.823 0.857 Table 1: Dice scores for different automatic multi-organ segmentation methods and inter-annotator agreement results [17] on the VISCERAL dataset. We note that because the cloud-based evaluation service [17] containing the test data was closed at the time of running these experiments, we were not able to evaluate our method on the test data, thus making direct comparisons to previous methods difficult. As previously mentioned, we trained our method on 80% of the data (16 scans) and validated it on the remaining 20% (4 scans). From having evaluated the kidney only CNN on the test data, we found that the testing dataset gave better dice scores than the validation set. We are therefore confident that our results in Table 1 are representative. As soon as the test data set is made available we will update our results. Also for organs ‘Lung’, ‘Kidney’, ‘PMajor’ (psoas major) and ‘Abdom’ (rectus abdominis) we give mean dice scores of both the left and right organs. For this experiment we modified the training schedule such that the initial learning rate is 0.05, and each batch contains 24 patches, sampled from 2 randomly selected scans from the training set. We run each epoch for 200 batches, and halve the learning rate after every 25 epochs. The Dual CNN without using the isample scheme (average organ dice 0.8917) slightly underperformed compared to when using the isample (average organ dice 0.9034). However this difference is far less notable than during previous experiments, shown in Table 4b. We hypothesize this is because the background class in the multi-organ segmentation problem is split into background and other organs such as the Lung, Liver etc, thus making the dataset, isample: Adaptive Sampling for Fully Convolutional Networks 9 especially the background class, easier to sample from. The potential benefit of using the isample method is therefore problem dependent. 4 Conclusion We proposed and evaluated a sampling scheme to deal with very large images such as 3D CT scans. As shown in section 3 the sampler enables fast training, and our results indicate that the final generalization performance can be improved. This is inline with previous research that shows the positive effect of curriculum learning on optimization and end performance of machine learning systems [9, 8]. Our experimental results suggests our algorithm gives new state of the art performance for the aorta, lung, kidney, rectus abdominis, spleen, sternum, trachea and bladder, on the VISCERAL anatomy benchmark, and improves upon human inter-annotator agreement scores on the following organs: aorta, lung, kidney, psoas major, rectus abdominis, spleen, sternum, trachea and bladder. These encouraging results pave the way for using CNNs for robust automatic segmentation within clinical practice, such as surgical planning. References 1. K. Kamnitsas, C. Ledig, V. F. Newcombe, J. P. Simpson, A. D. Kane, D. K. Menon, D. Rueckert, and B. Glocker, “Efficient multi-scale 3d cnn with fully connected crf for accurate brain lesion segmentation,” Medical image analysis, vol. 36, pp. 61–78, 2017. 2. P. F. Christ, M. E. A. Elshaer, F. Ettlinger, S. Tatavarty, M. Bickel, P. Bilic, M. Rempfler, M. Armbruster, F. Hofmann, M. DAnastasi, et al., “Automatic liver and lesion segmentation in ct using cascaded fully convolutional neural networks and 3d conditional random fields,” in International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 415–423, Springer, 2016. 3. O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 234–241, Springer, 2015. 4. Q. Dou, H. Chen, Y. Jin, L. Yu, J. Qin, and P.-A. Heng, “3D deeply supervised network for automatic liver segmentation from CT volumes,” in International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 149– 157, Springer, 2016. 5. Y. Xu, Z. Jia, L.-B. Wang, Y. Ai, F. Zhang, M. Lai, I. Eric, and C. Chang, “Large scale tissue histopathology image classification, segmentation, and visualization via deep convolutional activation features,” BMC bioinformatics, vol. 18, no. 1, p. 281, 2017. 6. R. Zhang, J. Yao, K. Zhang, C. Feng, and J. Zhang, “S-cnn-based ship detection from high-resolution remote sensing images.,” International Archives of the Photogrammetry, Remote Sensing & Spatial Information Sciences, vol. 41, 2016. 7. Y. Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum learning,” in Proceedings of the 26th annual international conference on machine learning, pp. 41–48, ACM, 2009. 10 Lorenz Berger et al. 8. M. P. Kumar, B. Packer, and D. Koller, “Self-paced learning for latent variable models,” in Advances in Neural Information Processing Systems, pp. 1189–1197, 2010. 9. V. Avramova, “Curriculum learning with deep convolutional neural networks,” 2015. 10. X. Chen and A. Gupta, “Webly supervised learning of convolutional networks,” in Proceedings of the IEEE International Conference on Computer Vision, pp. 1431– 1439, 2015. 11. X. Qi, Z. Liu, J. Shi, H. Zhao, and J. Jia, “Augmented feedback in semantic segmentation under image level supervision,” in European Conference on Computer Vision, pp. 90–105, Springer, 2016. 12. H. Zhao, X. Qi, X. Shen, J. Shi, and J. Jia, “Icnet for real-time semantic segmentation on high-resolution images,” arXiv preprint arXiv:1704.08545, 2017. 13. G. Lin, A. Milan, C. Shen, and I. Reid, “Refinenet: Multi-path refinement networks with identity mappings for high-resolution semantic segmentation,” arXiv preprint arXiv:1611.06612, 2016. 14. K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016. 15. L. Fidon, W. Li, L. C. Garcia-Peraza-Herrera, J. Ekanayake, N. Kitchen, S. Ourselin, and T. Vercauteren, “Generalised wasserstein dice score for imbalanced multi-class segmentation using holistic convolutional networks,” arXiv preprint arXiv:1707.00478, 2017. 16. C. H. Sudre, W. Li, T. Vercauteren, S. Ourselin, and M. J. Cardoso, “Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations,” arXiv preprint arXiv:1707.03237, 2017. 17. O. Jimenez-del Toro, H. Müller, M. Krenn, K. Gruenberg, A. A. Taha, M. Winterstein, I. Eggel, A. Foncubierta-Rodrı́guez, O. Goksel, A. Jakab, et al., “Cloudbased evaluation of anatomical structure segmentation and landmark detection algorithms: Visceral anatomy benchmarks,” IEEE transactions on medical imaging, vol. 35, no. 11, pp. 2459–2475, 2016. 18. X. Glorot and Y. Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pp. 249–256, 2010. 19. P. Goyal, P. Dollár, R. Girshick, P. Noordhuis, L. Wesolowski, A. Kyrola, A. Tulloch, Y. Jia, and K. He, “Accurate, large minibatch sgd: Training imagenet in 1 hour,” arXiv preprint arXiv:1706.02677, 2017. 20. Y. Nesterov, Introductory lectures on convex optimization: A basic course, vol. 87. Springer Science & Business Media, 2013. 21. C. Wang and Ö. Smedby, “Multi-organ segmentation using shape model guided local phase analysis,” in International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 149–156, Springer, 2015. 22. G. Vincent, G. Guillard, and M. Bowes, “Fully automatic segmentation of the prostate using active appearance models,” MICCAI Grand Challenge: Prostate MR Image Segmentation, vol. 2012, 2012. 23. T. Gass, G. Szekely, and O. Goksel, “Multi-atlas segmentation and landmark localization in images with large field of view,” in International MICCAI Workshop on Medical Computer Vision, pp. 171–180, Springer, 2014. 24. P. Krähenbühl and V. Koltun, “Efficient inference in fully connected crfs with gaussian edge potentials,” in Advances in neural information processing systems, pp. 109–117, 2011. isample: Adaptive Sampling for Fully Convolutional Networks 11 25. O. A. J. del Toro and H. Müller, “Hierarchic multi–atlas based segmentation for anatomical structures: Evaluation in the visceral anatomy benchmarks,” in International MICCAI Workshop on Medical Computer Vision, pp. 189–200, Springer, 2014. 26. R. Kéchichian, S. Valette, M. Sdika, and M. Desvignes, “Automatic 3d multiorgan segmentation via clustering and graph cut using spatial relations and hierarchicallyregistered atlases,” in International MICCAI Workshop on Medical Computer Vision, pp. 201–209, Springer, 2014.	1cs.CV
arXiv:1411.4691v3 [math.ST] 8 Nov 2016 Group Regularized Estimation under Structural Hierarchy Yiyuan She, Zhifeng Wang and He Jiang Department of Statistics, Florida State University Abstract Variable selection for models including interactions between explanatory variables often needs to obey certain hierarchical constraints. The weak or strong structural hierarchy requires that the existence of an interaction term implies at least one or both associated main effects to be present in the model. Lately, this problem has attracted a lot of attention, but existing computational algorithms converge slow even with a moderate number of predictors. Moreover, in contrast to the rich literature on ordinary variable selection, there is a lack of statistical theory to show reasonably low error rates of hierarchical variable selection. This work investigates a new class of estimators that make use of multiple group penalties to capture structural parsimony. We give the minimax lower bounds for strong and weak hierarchical variable selection and show that the proposed estimators enjoy sharp rate oracle inequalities. A general-purpose algorithm is developed with guaranteed convergence and global optimality. Simulations and real data experiments demonstrate the efficiency and efficacy of the proposed approach. Keywords: hierarchical variable selection, scalable computation, oracle inequality, minimax optimality 1 Introduction In statistical applications, it is often noticed that an additive model including main effects only is inadequate. Including some higher-order terms, such as 1 interactions, in particular, are often of great help in prediction and modeling. Sometimes, interactions may be of independent interest; one example is the moderation analysis in behavioral sciences (Cohen et al., 2013). In this paper, we focus on the full quadratic model with all two-term interactions taken into account. Let X = [x1 , x2 , . . . , xp ] ∈ Rn×p be the (raw) predictor matrix and y ∈ Rn be the response vector. We assume the following nonlinear additive regression model X X y = b∗0 1 + b∗j xj + φ∗jk xj xk + ε, (1) 1≤j≤p 1≤j,k≤p where ε ∼ N (0, σ 2 I) and denotes the Hadamard product. The num p ber of predictors is then p + 2 , posing a challenge in variable selection even when p is moderate. Moreover, in this scenario, statisticians are often interested in obtaining a model satisfying certain logical relations, such as the structural hierarchy discussed in Nelder (1977), McCullagh and Nelder (1989), and Hamada and Wu (1992). Hierarchy is a natural requirement in gene regulatory network studies (Davidson and Erwin, 2006), banded covariance matrix estimation (Bien et al., 2016) and lagged variable selection in time series. Hierarchical variable selection leads to reduced number of variables in measurement, referred to as practical sparsity (Bien et al., 2013). For instance, a model consisting of x1 , x2 , and x1 x2 may be more parsimonious to practitioners than a model involving x1 , x2 , and x3 x4 . In our setting, there are two types of hierarchy (Chipman, 1996; Bien et al., 2013): strong hierarchy (SH) and weak hierarchy (WH). Let φjk be the coefficient of xj xk and xk xj . SH means that if an interaction term exists in the model, then both of its associated main effects must be present, i.e., φjk 6= 0 → bj 6= 0 and bk 6= 0, while WH requires that the inclusion of an interaction implies at least one of its associated main effects to be added into the model, i.e., φjk 6= 0 → bj 6= 0 or bk 6= 0. We will show in Section 3 that WH is relatively easy to realize compared with SH. SH is invariant to linear transformations of predictors (Peixoto, 1990) and is the primary concern in this work. It is a nontrivial task to maintain hierarchy in model selection using conventional approaches. LASSO (Tibshirani, 1996) may violate SH and WH as well. We refer to Nelder (1977), Peixoto (1987), Bickel et al. (2010), Wu et al. (2010), and Hao and Zhang (2014) for some well-developed multi-step 2 procedures which, however, might be ad hoc and greedy. This paper focuses on regularization-based approaches. The past works in this direction include SHIM (Choi et al., 2010), VANISH (Radchenko and James, 2010) and HL (Bien et al., 2013). SHIM reparametrizes φjk as ρjk bj bk and enforces sparsity in both b = [bj ] and ρ = [ρjk ]. The formulation is motivating, and we could also use φjk = ρjk (b2j + b2k ) for WH. However, the corresponding optimization problem is nonconvex and the computational algorithm of SHIM is quite slow in large-scale problems. VANISH is one of the main motivations of our work and will be discussed in detail in Section 3. HL is a recent breakthrough in hierarchical variable selection. One of its key ideas is to enforce a magnitude constraint on the coefficients, kφj k1 ≤ \|bj \|, to make hierarchy naturally hold. Here, φj is a vector of coefficients of the predictors xj xk , 1 ≤ k ≤ p. To handle the nonconvex constraint, Bien et al. (2013) rephrased it with the − pesudo-positive and pesudo-negative parts b+ j , bj of bj but dropped all zero+ − product constraints bj bj = 0, 1 ≤ j ≤ p. The quality of such a convex relaxation seems to have no theoretical justification in the literature. In our experience, HL has excellent performance when the main effects are strong and p is not very large. But it can miss some interaction effects and become computationally prohibitive on large datasets. For example, when p = 1000, HL can take days to obtain a 20-point solution path. In this work, we propose and study group regularized estimation under structural hierarchy (GRESH). In theory, we are able to establish nonasymptotic oracle inequalities to show the error rates of the proposed estimators are minimax optimal up to some logarithm factors. We come up with a new recipe to conquer the theoretical difficulties when analyzing overlapping regularization terms in pursuing structural parsimony. Moreover, we develop a computational algorithm which guarantees the convergences of iterates and function values; it is not only efficient but also simple to implement. The rest of the paper is organized as follows. Some notation and symbols are introduced in Section 2. Section 3 presents the general framework of GRESH. A fast computational algorithm with theoretical support is given in Section 4. Section 5 builds oracle inequalities for GRESH, and Section 6 shows the minimax optimal rates. In Section 7, simulation studies and real data analysis are conducted to show the prediction accuracy and computational efficiency of the proposed approach. All technical proofs are given in the Appendices. 3 2 Notation We introduce some convenient notation and symbols to be used through> out the paper. First, for any matrix A = [a1 , . . . , aP ∈ Rp×n , define its p] p (2, 1)-norm, (2, ∞)-norm and `1 -norm as kAk2,1 = i=1 kai k2 , kAk2,∞ = max1≤i≤p kai k2 and kAk1 = k vec (A)k1 , respectively, where vec is the standard vectorization operator. The spectral norm and Frobenius norm of A are denoted by kAk2 and kAkF , respectively. For any p-dimensional vector a that is divided into K groups with aj representing the j-th subvector, its P (2, 1)-norm is defined by kak2,1 = K j=1 kaj k2 . The following two operators diag and dg are introduced for notational simplicity. For a square matrix A := [aij ]n×n , diag(A) := [a11 , . . . , ann ]> , and for a vector a = [a1 , . . . , an ]> ∈ Rn , diag{a} is defined as an n × n diagonal matrix with diagonal entries given by a1 , . . . , an . Define dg(A) := diag{diag(A)} = diag{[a11 , . . . , ann ]> }. We use A[I, J ] to denote a submatrix of A with rows and columns indexed by I and J , respectively. For any arbitrary b ∈ Rp , Φ ∈ Rp×p , we define J 11 (b, Φ) : = {j ∈ [p] : bj 6= 0, φj 6= 0}, J 10 (b, Φ) : = {j ∈ [p] : bj 6= 0, φj = 0}, J 01 (b, Φ) : = {j ∈ [p] : bj = 0, φj 6= 0}, J 00 (b, Φ) : = {j ∈ [p] : bj = 0, φj = 0}, (2) Je (Φ) : = {j ∈ [p2 ] : (vec (Φ))j 6= 0}, JG (b, Φ) : = {j ∈ [p] : b2j + kφj k22 6= 0}, where [p] = {1, . . . , p}, and the coefficient vector φj denotes the j-th column of Φ. With \|·\| standing for set cardinality, we define J 11 (b, Φ) := \|J 11 (b, Φ)\|, J 10 (b, Φ) := \|J 10 (b, Φ)\|, J 01 (b, Φ) := \|J 01 (b, Φ)\|, J 00 (b, Φ) := \|J 00 (b, Φ)\|, Je (Φ) := \|Je (Φ)\| and JG (b, Φ) := \|JG (b, Φ)\|. Clearly, J 11 +J 10 +J 01 +J 00 = p, and J 11 + J 10 + J 01 = JG . In addition, under SH, JG (b, Φ) equals the number of nonzero elements of b. Given the true signal (b∗ , Φ∗ ), the following abbreviated symbols are used: J 11∗ = J 11 (b∗ , Φ∗ ), J 10∗ = J 10 (b∗ , Φ∗ ), J 01∗ = J 01 (b∗ , Φ∗ ), Je∗ = Je (Φ∗ ), and JG∗ = JG (b∗ , Φ∗ ). In the paper, we frequently use the concatenated coefficient matrix for convenience Ω = [b, Φ> ]> , (3) 4 > and its j-th column is denoted by Ωj = [bj , φ> j ] . Given Ω, Ωb and ΩΦ stand > for (Ω[1, :]) and Ω[2 : (p + 1), :], respectively. Let X = [x1 , . . . , xp ] ∈ Rn×p be the raw predictor matrix. Define X̄ = [x1 x1 , . . . , x1 X̆ = [x1 , x1 xp , . . . , xp x1 , . . . , x1 x1 , . . . , xp xp , . . . , xp , xp 2 xp ] ∈ Rn×p , x1 , . . . , xp xp ] ∈ R (4) n×(p2 +p) . (5) Then X̄ consists of all interactions, and X̆ includes all p2 +p predictors in the quadratic model. Given any subset J ⊂ [p], we abbreviate X[:, J ] as X J . It is also easy to see that diag(XΦX > ) = X̄ vec (Φ) and Xb+diag(XΦX > ) = X̆ vec (Ω). For any two real numbers a and b, a . b means that a ≤ b holds up to a multiplicative numerical constant. For two equally sized matrices A = (aij ) and B = (bij ), A ≥ B means aij ≥ bij , for ∀ i, j. 3 Group regularized estimation under structural hierarchy For simplicity, we assume for now that there exists no intercept term in the model. Then (1) can be written as y = Xb∗ + diag(XΦ∗ X > ) + ε, (6) where ε ∼ N (0, σ 2 I) with σ 2 > 0, y = [y1 , . . . , yn ]> ∈ Rn is the response vector, and X = [x1 , . . . , xp ] ∈ Rn×p is the design matrix consisting of main effects only. We describe a general framework for hierarchical variable selection, referred to as group regularized estimation under structural hierarchy or GRESH. GRESH has two different types, depending on which objects to regularize. Denoting the squared error loss by `, i.e., 1 (7) `(b, Φ) = ky − Xb − diag(XΦX > )k22 , 2 the first type is given by Type-A : min Ω=[b,Φ> ]> ∈R(p+1)×p `(b, Φ) + λ1 kΦk1 + λ2 p X j=1 s.t. Φ = Φ> (for SH only), 5 k[bj , z(φj )]kq (8) where λ1 , λ2 are regularization parameters, 1 < q ≤ +∞ and z(x) is a function satisfying the property that z(x) = 0 implies x = 0 for any vector x ∈ Rp . For instance, z can take the `r -norm function (r > 0) z(x) = kxkr , (9) z(x) = x> . (10) or simply the identity function The first `1 penalty imposes elementwise sparsity on Φ and the second group`1 penalty enforces column sparsity in Ω. We argue that with the two penalties and the constraint, (8) can be used for strong hierarchical variable selection. Indeed, the sparsity of b comes from the second group-penalty alone, i.e., bj = 0 implies k[bj , z(φj )]kq = 0 (with probability 1) or z(φj ) = 0. By the property of the z function, φj = 0, and thus φjk = 0. The symmetry condition indicates further that φkj = 0. Hence (φjk + φkj )/2, the coefficient for xj xk , is zero. Consequently, whenever bj = 0, xj xk will be removed from the model and SH is automatically obeyed. We can describe the reasoning as follows bj = 0 ⇒ k[bj , z(φj )]kq = 0 ⇒ z(φj ) = 0 ⇒ φj = 0 ⇒ φjk = 0 φjk + φkj = 0. ⇒ φkj = 0 ⇒ 2 (11) Without the symmetry constraint, we can only complete the argument in the first line of (11), and so SH dose not hold. But interestingly, WH is guaranteed, because from bj = bk = 0, we have (φjk + φkj )/2 = 0. Therefore, WH gives a relatively simpler problem. As pointed out by a reviewer, when the model contains an intercept, centering the response and the p raw predictors does not make it vanish due to the presence of nonlinear terms, and so 1b0 + Xb + diag(XΦX > ) should be used to approximate y. In the SH scenario, if at least one x-predictor is relevant, substituting X 0 = [1, X] for X in (7) suffices. We focus on convex forms of GRESH in this work. But surely the `1 penalty and the group-`1 penalty in (8) can be replaced by their nonconvex alternatives; see, e.g., She (2012). GRESH is related to some methods in the literature. HL makes a special case of (8) because one of its formulations corresponds to q = ∞ and r = 1, 6 with a single regularization parameter being used. Another instance is given by q = 2 and z(x) = x> : min Ω=[b,Φ> ]> `(b, Φ) + λ1 kΦk1 + λ2 kΩ> k2,1 s.t. Φ = Φ> . (12) Bien et al. (2013) incorrectly described VANISH (Radchenko and James, 2010) in this form, without the symmetry condition. We will focus on (12) in the theoretical and computational studies of Type-A GRESH. As a matter of fact, Radchenko and James (2010) defined VANISH in a different way, which motivates another type of GRESH X Type-B: min `(b, Φ) + λ1 kφjk xj xk k2 + Ω=[b,Φ> ]> 1≤j,k≤p p λ2 X k[bj xj , z(φj1 xj x1 , . . . , φjp xj xp )]kq , s.t. Φ = Φ> (SH only), j=1 (13) where λ1 , λ2 , q, and z are defined as in (8). Similarly, we can argue that (13) keeps hierarchy. When q = 2 and z takes the form of (10), the penalty part in (13) become X X X 1 (14) kφjk xj xk k22 ) 2 , λ1 kφjk xj xk k2 + λ2 (kbj xj k22 + j,k j k as considered by Radchenko and James (2010). VANISH constructs main effects and interactions from two small sets of orthonormal basis functions in a functional regression setting. We do not pose such a restriction on the design matrix, and p can be arbitrarily large. The key difference between the two types of GRESH is that the penalties are imposed on the coefficients in (8), but on the terms in (13). A common practice before calling a shrinkage method is normalizing/standardizing all predictors, so that it is more reasonable to use a common regularization parameter in penalizing different coefficients. In this way, (8) builds a model on the normalized predictors and their interactions, while Type-B amounts to forming the overall design X̆ first and then performing the standardization. They are not equivalent because in general, (xj /kxj k2 ) (xk /kxk k2 ) 6= (xj xk )/kxj xk k2 . Then, which type of GRESH is preferable? An answer will be given in Section 5. 7 GRESH offers some general schemes for hierarchical variable selection. But it is no ordinary lasso or group lasso, since Φ appears in both penalties as well as the symmetry constraint. The main goal of this paper is to tackle some computational and theoretical challenges arising from the overlapping regularization terms in high dimensions. In computation, we would like to develop fast and scalable algorithms (cf. Section 4); in theory, how to treat the penalties and the constraint jointly to derive a sharp error bound for GRESH is intriguing and challenging (cf. Sections 5, 6). 4 Computation It is perhaps natural to think of using the alternating direction method of multipliers (ADMM, cf. Boyd et al. (2011)) to deal with the computational challenge. ADMM recently gains its popularity among statisticians. In fact, Bien et al. (2016) designed an algorithm of HL based on ADMM, where one of the main ingredients is the augmented Lagrangian ρ + − − 2 > + `(b − b , Φ) + λ1 (b + b ) + λkΦk 1 + hΦ − Ψ, Li + kΦ − ΨkF , ± p p 2 b ∈R ,Φ∈R min − + − s.t. Ψ = Ψ> , kφj k1 ≤ b+ j + bj , bj ≥ 0, bj ≥ 0, 1 ≤ j ≤ p. (15) Here, L is a Lagrange multiplier matrix, and ρ > 0 is a given constant, sometimes referred as the penalty parameter. Although ADMM enjoys some nice convergence properties in theory, practically only when ρ is large enough can we obtain a solution with good statistical accuracy. But often the larger the value of ρ is, the slower the (primal) convergence is. For example, the R package HierNet (version 1.6) for computing HL recommends ρ = n, but for p = 1000, the algorithm may take several days to compute a single solution path. There are some empirical schemes on how to vary ρ during the iteration, but they are ad hoc and do not always behave well. In this section, we consider a slightly more general optimization problem which includes both types of GRESH as particular instances min Ω=[b,Φ> ]> 1 ky − Xb − diag(ZΦZ > )k22 + kλb 2 +kΛΦ Φk1 + kΛ> Ω > Ω k2,1 8 bk1 > s.t. Φ = Φ , (16) where X, Z ∈ Rn×p , and λb , ΛΦ and ΛΩ are non-negative regularization vector and matrices. Let Z̄ = [z 1 z 1 , . . . , z 1 z p , . . . , z p z 1 , . . . , z p z p ], and Z̆ = [x1 , z 1 z 1 , . . . , z 1 z p , . . . , xp , z p z 1 , . . . , z p z p ]. We assume p that ΛΩ = 1λ> Ω for some λΩ ∈ R in developing the algorithm. (Since our 2 algorithm applies to a general Z̄ ∈ Rn×p that is not necessarily symmetric, this is without loss of generality.) In (16), the `1 -type penalties are imposed on overlapping groups of variables. It is worth noting that the symmetry constraint considerably complicates the grouping structure. Without it, the variable groups can be shown to follow a tree structure, for which efficient algorithms can be developed on Jenatton et al. (2011) or Simon et al. (2013). Our algorithm follows a different track than ADMM. The details are presented in Algorithm 1. Step 1 updates Ξ and results from a linearizationbased surrogate function. Step 2 carries out a Dykstra-like splitting—see, e.g., Bauschke and Combettes (2008) and She et al. (2014), by use of two ~ S and ΘS . Concretely, for any real number a, ΘS (a; λ) proximity operators, Θ is given by sgn(a)(\|a\| − λ)+ with sgn(·) representing the sign function. For any vector a, ΘS (a; λ) is defined componentwise and the multivariate version ~ S (a; λ) is given by aΘS (kak2 ; λ)/kak2 if a 6= 0 and 0 otherwise. Θ The GRESH algorithm is easy to implement and involves no complicated matrix operations such as matrix inversion. Moreover, it does not contain sensitive algorithmic parameters like ρ in ADMM, and needs no line search. Theorem 1 provides a universal theoretical choice for τ to guarantee the global optimality of Ω̂. In particular, strict iterate convergence, in addition to function-value convergence, can be established, which is considerably stronger than an “every accumulation point” type conclusion in many numerical studies. For clarity, we assume that the inner iteration runs till convergence, but this is unnecessary; see Remark 2 below. Theorem 1. Suppose λb ≥ 0, ΛΦ ≥ 0, λΩ > 0. For any τ > kZ̆k2 and any starting point Ω(0) , the sequence of iterates {Ω(i) } converges to a globally optimal solution of (16). Remark 1. The conclusion in the theorem holds for “`1 + `2 ” type penalties as well (Zou and Hastie, 2005; Owen, 2007). For the associated proximity operators, see He et al. (2013). In hierarchical variable selection, adding an `2 -type shrinkage is particularly helpful to compensate for model collinearity. Remark 2. Neither the convergence of iterates nor the optimality guarantee requires the full convergence of the inner loop; see Appendix A.1 for more 9 Algorithm 1 The GRESH algorithm for solving the general problem (16) Inputs: Data:X, Z, y. Regularization parameters: λb , ΛΦ , λΩ . Initialization: i ← 0, τ large enough (say τ = kZ̆k2 ); 2 2 λb ← λb /τ 2 , ΛΦ ← (ΛΦ + Λ> Φ )/(2τ ), λΩ ← λΩ /τ . repeat 1. ΞΦ ← Φ(i) + Z > diag{y − Xb(i) − Z̄ vec (Φ(i) ))}Z/τ 2 , > Ξb ← b(i) + X > (y − Xb(i) − Z̄ vec (Φ(i) ))/τ 2 , Ξ ← [Ξb , Ξ> Φ] ; 2. P ← 0, Q ← 0; repeat ~ S (Ξ[:, k] + P [:, k]; λΩ [k]), ∀k : 1 ≤ k ≤ p; (i) Ω[:, k] ← Θ (ii) P ← P + Ξ − Ω; (iii) Ξ[1, :] ← ΘS (Ω[1, :] + Q[1, :]; λ> b ); (iv) Ω[2:end, :] ← (Ω[2:end, :] + Ω[2:end, :]> )/2; (v) Ξ[2:end, :] ← ΘS (Ω[2:end, :] + Q[2:end, :]; ΛΦ ); (vi) Q ← Q + Ω − Ξ; until convergence 3. Ω(i+1) ← Ξ; 4. i ← i + 1; until convergence Output Ω̂ detail. Various stopping criteria can be employed, e.g., Schmidt et al. (2011). In our experience, running (i)–(vi) for a few steps (say 10) usually suffices. Remark 3. Algorithm 1 and Theorem 1 can be extended beyond quadratic loss functions. When ` takes the binomial deviance in classification problems, the first step of Algorithm 1 becomes Ξb ← b(i) + X > (y − π(Xb(i) + Z̄ vec (Φ(i) )))/τ 2 , ΞΦ ← Φ(i) + Z > diag{y − π(Xb(i) + Z̄ vec (Φ(i) ))}Z/τ 2 , > Ξ ← [Ξb , Ξ> Φ ] , where π(t) = 1/(1 + exp(−t)) and extends componentwise to vectors. We can show that theoretically, choosing τ > kZ̆k2 /2 guarantees the convergence of the algorithm. Remark 4. We recommend applying Nesterov’s first acceleration in implementations (Nesterov, 2007). In more detail, it uses a momentum update of Ξ in Step 1: If i = 0, Ξb ← b(i) + X > (y − Xb(i) − Z̄ vec (Φ(i) ))/τ 2 , 10 ΞΦ ← Φ(i) + Z > diag{y − Xb(i) − Z̄ vec (Φ(i) ))}Z/τ 2 ; if i > 0, Ξb ← (1 − ωi )Ξb + ωi (b(i) + X > (y − Xb(i) − Z̄ vec (Φ(i) ))/τ 2 ), ΞΦ ← (1 − ωi )ΞΦ + ωi (Φ(i) +Z > diag{y −Xb(i) − Z̄ vec (Φ(i) )}Z/τ 2 ), where ωi = (2i+3)/(i+3). Empirically, the number of iterations can be reduced by about 40% in comparison to the non-relaxed form. 5 Non-asymptotic analysis In this section, given any ∆ ∈ R(p+1)×p and JG ⊂ [p], we use ∆JG to denote the submatrix ∆[:, JG ]. Given any Je ⊂ [p2 ], k vec (∆Φ )Je k1 and k vec (∆Φ )Je k2 are abbreviated as k(∆Φ )Je k1 and k(∆Φ )Je k2 , respectively, when there is no ambiguity. In this multi-regularization setting, some standard treatments of the stochastic term do not give sharp error rates. In particular, applying hε, > > X̄ vec (∆Φ )i ≤ kX̄ εk∞ k∆Φ k1 and hε, X̆ vec (∆)i ≤ kX̆ εk2,∞ k∆> k2,1 as commonly used in the literature (Bickel et al., 2009; Lounici et al., 2011; Negahban et al., 2012; van de Geer, 2014) would yield a prediction error bound of the order σ 2 (Je log p + JG p), which is, ironically, much worse than the error rate of LASSO or Group LASSO (G-LASSO). Our analysis relies on two interrelated inequalities derived from the statistical and computational properties of GRESH estimators. See Appendix A.2 for more technical detail. First, let’s consider the Type-A problem (12), with λ1 , λ2 redefined: 1 ky−Xb−X̄ vec (Φ)k22 +λ1 kX̆k2 kΦk1 +λ2 kX̆k2 kΩ> k2,1 s.t. Φ = Φ> . Ω=[b,Φ> ]> 2 (17) > > Let Ω̂ = [b̂, Φ̂ ] be any global minimizer of (17). We are interested in its prediction accuracy measured by M (b̂ − b∗ , Φ̂ − Φ∗ ), where min M (b, Φ) = kXb + X̄ vec (Φ)k22 . (18) The predictive learning perspective is always legitimate in evaluating the quality of the estimator regardless of the signal-to-noise ratio. To guarantee small predictor errors when using a convex method, the design matrix must satisfy certain incoherence conditions, one of the most popular being the restricted eigenvalue (RE) assumption (Bickel et al., 2009; Lounici et al., 2011). In the following, we give an extension of RE in the hierarchy setting, with the restricted cone defined with both `1 and group-`1 penalties. A less 11 intuitive but technically much less demanding condition is used in the proof. Assumption A(Je , JG , ϑ, κ). Given Je ⊂ [p2 ], JG ⊂ [p], κ ≥ 0 and a > (p+1)×p constant ϑ ≥ 0, for any ∆ = [∆b , ∆> satisfying ∆Φ = ∆> Φ] ∈ R Φ and > k(∆Φ )Jec k1 + k(∆JGc ) k2,1 ≤ (1 + ϑ)(k(∆Φ )Je k1 + k(∆JG )> k2,1 ), the following inequality holds κkX̆k22 (k(∆Φ )Je k22 + k∆JG k2F ) ≤ kX∆b + X̄ vec (∆Φ )k22 . (19) The rate choices of the regularization parameters play a major role in prediction. We choose λ1 and λ2 in (17) according to p p (20) λ1 = A1 σ log(ep), λ2 = A2 σ log(ep), where A1 , A2 are large constants. (20) is quite different from the typical choice in group-`1 penalization; see Remark 2 for more detail. The following theorem states a non-asymptotic oracle inequality as well as a model cardinality bound for GRESH estimators. For convenience, we use abbreviated symbols JˆG = JG (b̂, Φ̂) and Jˆe = Je (Φ̂) for the estimate, and JG = JG (b, Φ) and Je = Je (Φ) for the reference signal. > > Theorem 2. Assume ∼ N (0, σ 2 I). Let Ω̂ = b̂, Φ̂ be a global minimizer of (17). Then under (20), for any sufficiently large constants A1 , A2 , the following oracle inequality holds for any (b, Φ) ∈ Rp × Rp×p 1 E[M (b̂ − b∗ , Φ̂ − Φ∗ )] . M (b − b∗ , Φ − Φ∗ ) + (1 ∨ )σ 2 (Je + JG ) log p + σ 2 , κ (21) provided that (X, X̄, X̆) satisfies A(Je , JG , ϑ, κ) for some κ > 0 and some constant ϑ ≥ 0. Furthermore, under the same regularity condition, the overall sparsity of the obtained model is controlled by E[Jˆe ] + E[JˆG ] . {M (b∗ − b, Φ∗ − Φ) + σ 2 }/{σ 2 log(ep)} + Je + JG . (22) Remark 1. Letting b = b∗ and Φ = Φ∗ in (21), we obtain an error bound no larger than σ 2 (Je∗ + JG∗ ) log p (omitting constant factors). This indicates that GRESH not only guarantees SH, but can give an error rate as low as that of LASSO. The existence of the bias term M (b − b∗ , Φ − Φ∗ ) makes our results applicable to approximately sparse signals, which is of practical significance. The theorem does not require the spectral norms of the design 12 √ matrices X, X̄ and X̆ to be bounded above by O( n) as assumed in, for example, Zhang and Huang (2008) and Bickel et al. (2009). In addition, the true signal Ω∗ and the reference signal Ω in the theorem need not obey SH. Remark 2. It is widely acknowledged that the penalty parameter for a grouped `1 penalty should be adjusted √ by the group size (Yuan and Lin, 2006). In fact, λ2 would be of order σ p + log p from Lounici et al. (2011) and Wei and Huang (2010), in light of the fact that there are p groups of size (p + 1) in kΩk2,1 . Perhaps surprisingly, this parameter choice becomes suboptimal in hierarchical variable selection. In fact, due to the presence of multiple penalties, we show in the proof that (20) suffices to suppress the noise, which in turn leads to a reduced error rate. Such a novel finding is owing to the careful treatment of the stochastic term, which is generally applicable to overlap group lasso (Jenatton et al., 2011). The conclusion that λ1 and λ2 are of essentially the same rate also facilitates parameter tuning, since one just needs to search along a one-dimensional grid. Remark 3. Theorem 2 can be extended to sub-Gaussian vec (ε) with mean 0 and its ψ2 -norm bounded by σ, which covers more noise distributions. High probability form results of the prediction error can be obtained as well: (21) and (22), without the expectation and the additive σ 2 term, hold with 2 2 probability at least 1 − Cp−c min{A1 ,A2 } for some universal constants C and c, and so Jˆe + JˆG . Je∗ + JG∗ with high probability. Moreover, in Appendix B, we show how to adapt our proof to deliver a coordinatewise error bound which can be used for recovering the sparsity pattern of the true signal. Remark 4. For the WH version of (17) (without the symmetry condition), following the lines of the proof of Theorem 2, we can show its error rate is of the order σ 2 {Jew (Φ)+JGw (Ω)} log p, where JGw (Ω) = JG (Ω0 ), Jew (Φ) = Je (Φ0 ), with Ω0 = [b, Φ0> ]> and φ0kj = φkj + φjk for k ≥ j and 0 otherwise. The associated regularity condition uses Jew , JGw , in place of Je , JG , respectively, and does not require ∆Φ to be symmetric. Details are not reported in the paper. Similarly, we can derive an oracle inequality for GRESH estimators of Type-B. Let X s be the column-scaled X such that the `2 -norm of each of s s its columns equals 1. X̄ and X̆ are similarly defined. The corresponding coefficients, denoted by bs and Φs , satisfy Xb = X s bs , X̄ vec (Φ) = s s s X̄ vec (Φs ). Let (b̂ , Φ̂ ) be a global minimizer of the scaled Type-B s s problem: minΩs =[bs ,Φs> ]> 21 ky − X s bs − X̄ vec (Φs )k22 + λ1 kX̆ k2 kΦs k1 + 13 s λ2 kX̆ k2 k(Ωs )> k2,1 s.t. Φs = (Φs )> . As aforementioned, the problem can s not be reduced to (17) because X̄ [:, jk] does not equal X s [:, j] X s [:, k] in general. Assumption A0 (Je , JG , ϑ, κ0 ). Given Je ⊂ [p2 ], JG ⊂ [p], and positive > constants κ0 and ϑ, for any ∆ = [∆b , ∆> satisfying ∆Φ = ∆> Φ] Φ and > > c c k(∆Φ )Je k1 + k(∆JG ) k2,1 ≤ (1 + ϑ)(k(∆Φ )Je k1 + k(∆JG ) k2,1 ), the following inequality holds s s κ0 kX̆ k22 (k(∆Φ )Je k22 + k∆JG k2F ) ≤ kX s ∆b + X̄ vec (∆Φ )k22 . Theorem 2’ Under the same conditions as in Theorem 2 and with A0 (Je , JG , ϑ, δJ0 e ,JG ) in place of A(Je , JG , ϑ, κ), (21) and (22) hold. The error bounds of the two types of GRESH are of the same order, but their regularity conditions place different requirements on the design. We performed extensive simulation studies to compare A and A0 , and found that for the same ϑ, κ < κ0 usually holds, which suggests the penalization on the basis of terms seems more appropriate than that on the coefficients. Therefore, we recommend Type-B regularization for hierarchical variable selection. 6 Minimax lower bound and error rate comparison In this section, we show that in a minimax sense, the error rate we obtained in Theorem 2 is minimax optimal up to some logarithmic factors. Consider two signal classes having hierarchy and joint sparsity: SH(JG , Je ) = {Ω = [b, Φ> ]> : Ω obeys SH, Φ = Φ> , JG (Ω) ≤ JG , Je (Φ) ≤ Je }, (23) WH(JG , Je ) = {Ω = [b, Φ> ]> : Ω obeys WH, JGw (Ω) ≤ JG , Jew (Φ) ≤ Je }, (24) where 1 ≤ JG ≤ p, 1 ≤ Je ≤ pJG . Recall the definitions of JGw and Jew in Remark 4 following Theorem 2. Let `(·) be a nondecreasing loss function with `(0) = 0, ` 6≡ 0. Under some regularity assumptions, we study the 14 minimax lower bounds for strong and weak hierarchical variable selection. Assumption B S (JG , Je ). For any Ω = [b, Φ> ]> ∈ R(p+1)×p satisfying that Φ is symmetric, Je (Φ) ≤ Je and JG (Ω) ≤ JG , κkΩk2F ≤ kX̆ vec (Ω)k22 ≤ κkΩk2F holds, where κ/κ is a positive constant. Assumption B W (JG , Je ). For any Ω = [b, Φ> ]> ∈ R(p+1)×p satisfying that Jew (Φ) ≤ Je and JGw (Ω) ≤ JG , κkΩk2F ≤ kX̆ vec (Ω)k22 ≤ κkΩk2F holds, where κ/κ is a positive constant. Theorem 3. (i) Strong hierarchy. Assume y = Xb∗ + X̄ vec (Φ∗ ) + ε with ε ∼ N (0, σ 2 I), JG ≥ 2, p ≥ 2, Je ≥ 1, n ≥ 1, JG ≤ p/2, Je ≤ JG2 /2, and B S (2JG , 2Je ) is satisfied. Then there exist positive constants C, c (depending on `(·) only) such that sup inf ∗ Ω̂ Ω ∈SH(JG ,Je ) E[`(M (b̂ − b∗ , Φ̂ − Φ∗ )/(CPo (Je , JG )))] ≥ c > 0, (25) where Ω̂ denotes any estimator, and Po (Je , JG ) = σ 2 {Je log(eJG2 /Je ) + JG log(ep/JG )}. (26) (ii) Weak hierarchy. Let JG ≥ 1, Je ≥ 1, n ≥ 1, p ≥ 2, JG ≤ p/2, Je ≤ JG p/2. Under the same model assumption and B W (2JG , 2Je ), (25) holds if SH(JG , Je ) is replaced by WH(JG , Je ) and Po is replaced by Po0 (Je , JG ) = σ 2 {Je log(eJG p/Je ) + JG log(ep/JG )}. (27) We give some examples of ` to illustrate the conclusion. For SH, using the indicator function `(u) = 1u≥1 , we know that for any estimator (b̂, Φ̂), M (b̂ − b∗ , Φ̂ − Φ∗ ) & σ 2 Je log(eJG2 /Je ) + JG log(ep/JG ) occurs with positive probability, under some mild conditions. For `(u) = u, Theorem 3 shows that the risk E[M (b̂ − b∗ , Φ̂ − Φ∗ )] is bounded from below by Po (Je , JG ) up to some multiplicative constant. Because JG ≤ Je ≤ JG p and JG ≤ p, it is easy to see that the minimax rates are no larger than the error rate obtained in Theorem 2. A comparison of some popular methods follows, where we can see the benefits of hierarchical variable selection. In our context, LASSO solves 15 minb,Φ:Φ=Φ> `(b, Φ) + λ(kbk1 + kΦk1 ). From Bickel et al. (2009), the estimator has a prediction error of the order σ 2 (Je + J 10 + J 11 ) log p. G-LASSO, with the optimization problem defined by minb,Φ:Φ=Φ> `(b, Φ) + λkΩ> k2,1 , automatically maintains SH and has an error rate of σ 2 JG p (Lounici et al., 2011). In general, there is no clear winner between the two. Let’s turn to a particularly interesting case where Je∗ JG∗ p, i.e., the existence of a main effect in the model does not indicate that all its associated interactions must be relevant. In this scenario, LASSO always outperforms G-LASSO, although it does not possess the SH property. GRESH achieves the same low error rate and guarantees hierarchy, because under SH, J 11 + J 10 = JG . The error rate proved in (21) does not always beat that of G-LASSO, because only large values of A are considered in Theorem 2. Yet, even in the worst case when Je∗ JG∗ p, GRESH is only a logarithmic factor worse. In practical data analysis, there will be no performance loss, because when λ1 = 0, GRESH degenerates to G-LASSO. Table 1: Error rate comparison between LASSO, G-LASSO, and GRESH, where σ 2 and other constant factors are omitted. LASSO G-LASSO GRESH Minimax: 7 7.1 (Je + J 11 + J 10 ) log p JG p (JG + Je ) log p 2 /J ) JG log(ep/JG ) + Je log(eJG e Experiments Simulations In this part, we perform some simulation studies to compare the performance of HL and GRESH (of Type B, cf. (14)) in terms of prediction accuracy, selection consistency, and computational efficiency. We use a Toeplitz design to generate all main predictors, with the correlation between xi and xj given by 0.5\|i−j\| . The true coefficients b∗ and Φ∗ (symmetric) are generated according to the following three setups. Example 1. n = 40, p = 100 or 200 (and so p + p2 = 5050 or 20100). b∗ = [3, 1.5, 0, 0, 2, 2, 0, · · · , 0]> , Φ∗ = 0, σ 2 = 1. No interactions are relevant 16 to the response variable. SH is satisfied. Example 2. n = 150, p = 50 or 100 (and so p + p2 = 1275 or 5050). b∗ = [3, 3, 3, 3, 3, 3, 3, 3, 3, 0, · · · , 0]> , Φ∗ [1, 2] = Φ∗ [1, 3] = Φ∗ [4, 5] = Φ∗ [4, 6] = Φ∗ [7, 8] = Φ∗ [7, 9] = 3, σ 2 = 1. The model involves both main and interaction effects and obeys SH. Example 3. n = 100, p = 50 or 100 (and so p + p2 = 1275 or 5050). b∗ = [1, 1, 1, 1, 1, 1, 1, 0, · · · , 0]> , Φ∗ [i, j] = 5, 1 ≤ i, j ≤ 5, i 6= j, Φ∗ [4, 5] = Φ∗ [4, 6] = Φ∗ [4, 7] = 5, σ 2 = 1. The true model does not have very strong main effects but satisfies SH. All regularization parameters are tuned on a (separate) large validation dataset containing 10K observations. There is no need to perform a full twodimensional grid search to find the optimal parameters in GRESH. Rather, motivated by Theorem 2, we set λ2 = cλ1 , and chose c = 0.5 according to experience. Because of the convex nature of the problem, pathwise computation with warm starts is used. After variable selection, a ridge regression model is always refitted to be used for prediction. The official R package for HL is HierNet, implemented in C. We set strong=TRUE and post-calibrate HL by a restricted ridge refitting, which substantially enhances its accuracy. To make a fair comparison between HL and GRESH, we use the same error tolerance (1e-5) and the same number of grid values (20). All other algorithmic parameters in HierNet are set to their default values. Given each setup, we repeat the experiment for 50 times and evaluate the performance of each algorithm according to the measures defined below. The test error (Err) is the mean squared error between the true mean of y and its estimate; for robustness and stability, we report the median test error from all runs. The joint detection (JD) rate is the fraction of \|{(i, j) : Ω∗ij 6= 0}\| ⊆ \|{(i, j) : Ω̂ij 6= 0}\| among all experiments. The missing (M) rate and the false alarm (FA) rate are the mean of \|{(i, j) : Ω∗ij 6= 0, Ω̂ij = 0}\|/\|{(i, j) : Ω∗ij 6= 0}\| and the mean \|{(i, j) : Ω∗ij = 0, Ω̂ij 6= 0}\|/\|{(i, j) : Ω∗ij = 0}\|, respectively. The path computational cost is the average running time of an algorithm in seconds. All the experiments were run on a PC with 3.2GHz CPU, 32GB memory and 64-bit Windows 8.1. Table 2 and Table 3 summarize the statistical and computational results. From Table 2, GRESH and HL behaved equally well in Example 1, the model of which contains main effects only, but GRESH is faster. In Example 2 and Example 3, the two methods show more differences; see their test errors 17 Table 2: Statistical performance of HL and GRESH, measured in test error, joint detection rate, missing rate, and false alarm rate on simulation data. All numbers are multiplied by 100. Ex 1 (p = 100) Ex 2 (p = 50) Ex 3 (p = 50) Err JD M FA Err JD M FA Err JD M FA HL 13.7 100 0.00 0.00 14.2 95 0.24 0.31 68.0 90 0.65 2.71 GRESH 13.6 100 0.00 0.00 11.6 100 0.00 0.12 26.2 100 0.00 0.23 Ex 1 (p = 200) Ex 2 (p = 100) Ex 3 (p = 100) Err JD M FA Err JD M FA Err JD M FA HL 13.8 100 0.00 0.00 17.3 90 0.48 0.11 92.2 15 24.03 1.44 GRESH 13.5 100 0.00 0.00 14.3 100 0.00 0.05 26.9 100 0.00 0.08 Table 3: Path computation costs of GRESH and HL when p = 200 and 1000. The computational times are in seconds unless otherwise specified. p = 200 HL GRESH p = 1000 Ex1 Ex2 Ex3 Ex1 Ex2 Ex3 574 110 3057 158 2.9 hours 128 7.6 hours 1066 — 2.5 hours — 3194 and joint identification rates, for example. We also noticed that GRESH often gave a more parsimonious model. When the main effects are weak as in Example 3, HL may miss some genuine interaction effects. Overall, GRESH showed comparable or better test errors. In fact, this is observed even when SH is not satisfied (results not shown in the table). We suspect that the performance differences between HL and GRESH largely result from the fact that HL compares \|bj \| with kφj k1 , the `1 -norm of the overall φj , to realize SH, while (14) groups bj xj , φj1 xj x1 , . . ., and φjp xj xp , on the term basis, to select main effects. The computational times in Table 3 show the scalability of each algorithm as p varies. When p = 1000, there are 500500 variables in total, and so HL became computationally prohibitive, also evidenced by Lim and Hastie (2015). GRESH offered impressive computation gains in the experiment. 18 7.2 Comparison with ADMM This part shows that directly applying ADMM does not give a scalable algorithm for solving the optimization problem (16) which has a large number of groups with large group size. The detailed algorithm design is given in Appendix C. We set ρ = 1 in ADMM and compared it to Algorithm 1. The results are reported in Table 4 and Table 5. The statistical performances of the two algorithms are close. This is reasonable because they solve the same optimization problem. However, ADMM is much slower. In the experiments, ADMM became infeasible when p = 200 or larger. Table 4: Statistical performance of GRESH and ADMM, measured in test error, joint detection rate, missing rate, and false alarm rate. Ex 1 (p = 50) Ex 2 (p = 50) Ex 3 (p = 50) Err JD M FA Err JD M FA Err JD M FA GRESH 11.6 100 0.00 0.02 11.6 100 0.00 0.12 26.2 100 0.00 0.23 ADMM 11.7 100 0.00 0.02 11.9 100 0.00 0.12 26.3 100 0.00 0.22 Ex 1 (p = 100) Ex 2 (p = 100) Ex 3 (p = 100) Err JD M FA Err JD M FA Err JD M FA GRESH 13.6 100 0.00 0.00 14.3 100 0.00 0.05 26.9 100 0.00 0.08 ADMM 16.8 100 0.00 0.01 14.5 100 0.00 0.06 27.1 100 0.00 0.09 Table 5: Path computation costs of GRESH and ADMM. p = 50 p = 100 Ex1 Ex2 Ex3 GRESH ADMM 7.3 8 66 11 70 9 60 Ex1 Ex2 Ex3 27 36 30 1027 1571 1439 Real data example We performed hierarchical variable selection on the California housing data (Pace and Barry, 1997). The dataset consists of 9 summary characteristics for 20640 neighborhoods in California. The response variable is the median 19 house value in each neighborhood. Following Hastie et al. (2009), we obtained eight household-related predictor variables: median income, housing median age, average number of rooms and bedrooms per household, population, average occupancy (population/households), latitude, and longitude, denoted by MedInc, Age, AvgRms, AvgBdrms, Popu, AvgOccu, Lat, and Long, respectively. Similar to Ravikumar et al. (2007) and Radchenko and James (2010), 50 nuisance features generated as standard Gaussian random variables were added, to make the problem more challenging. The full quadratic model on this enlarged dataset contains 3422 unknowns. To prevent from getting over-optimistic error estimates, we used a hierarchical cross-validation procedure where an outer 10-fold cross-validation (CV) is for performance evaluation and the inner 10-fold selective CVs (She, 2012) are for parameter tuning. We managed to run both HL and GRESH for hierarchical variable selection, with the estimates post-calibrated by a local ridge fitting as described in Section 7.1. It took us approximately one and half days to complete the CV experiment for HL, and about 1.6 hours for GRESH. The median and mean test errors of the models obtained by HL are 530.8 and 553.5, respectively, and the average number of selected variables is 31.2. GRESH gave 516.9 and 521.1 for the median and mean test errors, , respectively, and selected 17.1 variables on average, about half of the model size of HL. To help the reader get an intuition of the selection frequencies of all predictors, we display heat maps in Figure 1. The two heat maps in the top panel include all variables, and the bottom panel only shows the heat maps restricted to the original covariates and their interactions. According to the figure, both methods successfully removed most of the artificially added noisy features. On average, only 9.3 nuisance covariates exist in the models obtained by HL, and 5.4 in the GRESH models. The heat maps of GRESH are however neater. The HL selection results are less parsimonious and are perhaps more difficult to interpret. The nonlinear terms in GRESH include the interaction between MedInc and Age, in addition to the quadratic effects of MedInc, Age, and AvgBdrms. Popu and all its associated interaction terms never got selected by GRESH. The insignificance of Popu can be confirmed by more elaborate analysis based on gradient boosting (Hastie et al., 2009). 20 Main Main 5 5 10 10 15 15 20 20 25 25 30 30 35 35 40 40 45 45 50 50 55 55 5 10 15 20 25 30 35 40 45 50 55 5 Main Main 1 1 2 2 3 3 4 4 5 5 6 6 7 7 8 10 15 20 25 30 35 40 45 50 55 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 Figure 1: Top panel: heat maps of HL (left) and GRESH (right) on California housing data. Bottom panel: heat maps of HL (left) and GRESH (right) restricted to the original 8 variables and their interactions. Appendices A A.1 Proofs of main theorems Proof of Theorem 1 Denote the object function of (16) by f (Ω; λb , ΛΦ , λΩ ) = `(Ω) + P1 (Ω; λb , ΛΦ ) + P2 (Ω; λΩ ), where P1 (Ω; λb , ΛΦ ) = kλb bk1 + kΛΦ Φk1 , and P2 (Ω; λΩ ) = k(λΩ 1> ) Ω> k2,1 . Define g(Ω, Ω0 ; λb , ΛΦ , ΛΩ ) = `(Ω0 ) + τ2 kΩ0 − Ωk2F + hX > (y − Xb0 − Z̄ vec (Φ0 )), b − b0 i + hZ > diag{y − Xb0 − 2 Z̄ vec (Φ0 )}Z, Φ−Φ0 i+P1 (Ω; λb , ΛΦ )+P2 (Ω; λΩ ). To simplify the notation, we sometimes write them as f (Ω) and g(Ω, Ω0 ) when there is no ambiguity. 21 Given Ω0 , by simple algebra, the minimization of g(Ω, Ω0 ; λb , ΛΦ , λΩ ) subject to Φ = Φ> reduces to min Ω∈R(p+1)×p 1 kΩ − Ξk2F + P1 (Ω; λb /τ 2 , ΛΦ /τ 2 ) + P2 (Ω; λΩ /τ 2 ) + ι{Φ=Φ> } , 2 (28) > where Ξ = [Ξb , Ξ> with Ξb = b0 + X > (y − Xb0 − Z̄ vec (Φ0 ))/τ 2 and Φ] ΞΦ = Φ0 + Z > diag{y − Xb0 − Z̄ vec (Φ0 )}Z/τ 2 , and ι{Φ=Φ> } = 0 if Φ = Φ> and +∞ otherwise, which is a lower semicontinuous convex function. Lemma 1. Let P ← 0, Q ← 0, Ξ ← Ξ0 . Consider the following iterative procedure: ~ S (Ξ[:, k] + P [:, k]; λΩ [k]), ∀k : 1 ≤ k ≤ p, Ω[:, k] ← Θ P ← P + Ξ − Ω, Ξ[1, :] ← ΘS (Ω[1, :] + Q[1, :]; λb ), Ω[2 : end, :] ← (Ω[2 : end, :] + Ω[2 : end, :]> )/2, (29) Ξ[2 : end, :] ← ΘS (Ω[2 : end, :] + Q[2 : end, :]; (ΛΦ + Λ> Φ )/2), Q ← Q + Ω − Ξ, till convergence. Then both the sequence of Ξ and the sequence of Ω converge to the globally optimal solution to minΩ∈R(p+1)×p 12 kΩ − Ξ0 k2F + kλb Ωb k1 + kΛΦ ΩΦ k1 + k(λΩ 1> ) Ω> k2,1 s.t. ΩΦ = Ω> Φ. Now we prove the convergence of Ω(i) in Algorithm 1. First, we notice the following fact. ~ S (·; λ) is non-expansive, that is, kΘ(a; ~ Lemma 2. Given any λ ≥ 0, Θ λ) − p ~ Θ(ã; λ)k2 ≤ ka − ãk2 for all a, ã ∈ R . Therefore, the mapping of the iteration in Algorithm 1 is non-expansive. We use the tool provided by Opial (1967) for nonexpansive operators to prove the strict convergence of Ω(i) . First, the fix point set of the mapping is non-empty due to the convexity and the KKT conditions. The mapping is also asymptotically regular in the sense that for any starting point Ω(0) , kΩ(i+1) − Ω(i) k → 0 as i → ∞. This property is implied by the following lemma with τ > kZ̆k2 . 22 Lemma 3. For the sequence {Ω(i) } generated by Algorithm 1, 1 f (Ω(i) ; λb , ΛΦ , λΩ ) − f (Ω(i+1) ; λb , ΛΦ , λΩ ) ≥ (τ 2 − kZ̆k22 )kΩ(i+1) − Ω(i) k2F , 2 (30) (i) for any i ≥ 0. Moreover, {Ω } is uniformly bounded. (30) actually does not require the optimality of Ω(i+1) in minimizing g(·, Ω(i) ) (cf. (28)). From the proof of Lemma 3, any Ω(i+1) satisfying g(Ω(i+1) , Ω(i) ) ≤ g(Ω(i) , Ω(i) )(= f (Ω(i) )) makes (30) hold. There is no need to run the inner loop till convergence. Moreover, to get the asymptotic regularity and the uniform boundedness of {Ω(i) }, it suffices to terminate the inner loop when g(Ω(i+1) , Ω(i) ) ≤ g(Ω(i) , Ω(i) ) + θkΩ(i+1) − Ω(i) k2F /2 or g(Ω(i+1) , Ω(i) ) − g(Ω(i) , Ω(i) ) ≤ θ, kΩ(i+1) − Ω(i) k2F /2 where θ is a pre-specified constant satisfying −τ 2 ≤ θ < τ 2 − kZ̆k22 . When Ω(i+1) exactly solves (28), θ = −τ 2 is satisfied. With all Opial’s conditions satisfied, the sequence {Ω(i) } has a unique limit point Ω∗ = [b∗ , Φ∗> ]> , and Ω∗ is a fixed point of Algorithm 1. Next we prove the optimality of any fixed point Ω∗ regarding problem (16). By Lemma 1, Ω∗ satisfies the Karush-Kuhn-Tucker (KKT) conditions of (28): ( 0 ∈ Ω∗ − Ξ + ∂P1 (Ω; λb /τ 2 , ΛΦ /τ 2 ) + ∂P2 (Ω; ΛΩ /τ 2 ) + ∂h(Φ∗ ), Φ∗ − Φ∗> = 0, where h(Φ; L) = hL, Φ − Φ> i. With the fixed point property established above, substituting Ξ with Ξb = b∗ + X > (y − Xb∗ − Z̄ vec (Φ∗ ))/τ 2 and ΞΦ = Φ∗ + Z > diag{y − Xb∗ − Z̄ vec (Φ∗ )}Z/τ 2 , we have  0 ∈ X > (y − Xb∗ − Z̄ vec (Φ∗ )) + ∂b P1 (Ω; λb , ΛΦ ) + ∂b P2 (Ω; ΛΩ ),    0 ∈ Z > diag{y − Xb∗ − Z̄ vec (Φ∗ )}Z + ∂ (P (Ω; λ , Λ ) + P (Ω; Λ )) Φ 1 b Φ 2 Ω ∗ 2  +∂h(Φ ; τ L),    0 = Φ∗ − Φ∗> , which are exactly the KKT conditions for the convex problem (16). Hence Ω∗ is a global minimizer of problem (16). 23 Similarly, the global optimality of limi→∞ Ω(i) = Ω∗ does not require solving (28) exactly. In the following, we use kAkΣ to denote {tr(A> ΣA)}1/2 . Let Ωo be an arbitrary optimal solution to (16). From the proof of Lemma 3, 2 the optimal Ω(i+1) satisfies g(Ω(i+1) , Ω(i) ) + τ2 kΩ(i+1) − Ωo k2F ≤ g(Ωo , Ω(i) ). We can relax it to 1 g(Ω(i+1) , Ω(i) ) + kΩ(i+1) − Ωo k2τ 2 I−θX > X ≤ g(Ωo , Ω(i) ) 2 for some θ : 0 ≤ θ ≤ 1. Then by the construction of g, we obtain 1 1 f (Ω(i+1) ) − f (Ωo ) ≤ kΩ(i) − Ωo k2τ 2 I−X > X − kΩ(i+1) − Ωo k2τ 2 I−θX > X 2 2 1 (i+1) − kΩ − Ω(i) k2τ 2 I−X > X 2 1 ≤ kΩ(i) − Ωo k2τ 2 I−X > X − kΩ(i+1) − Ωo kτ 2 I−θX > X 2 1 (i) 1 ≤ kΩ − Ωo k2τ 2 I−θX>X − kΩ(i+1) − Ωo kτ 2 I−θX>X . 2 2 It follows that f¯i+1 − f (Ωo ) ≤ 1 kΩ(0) − Ωo k2τ 2 I−θX>X , 2(i + 1) (31) Pi+1 (k) ∗ (i) 1 where f¯i+1 = i+1 k=1 f (Ω ). Letting i → ∞, we get f (Ω ) = f (limi→∞ Ω ) = limi→∞ f¯i = f (Ωo ), and so Ω∗ must also be a global minimizer of (16). Proof of Lemma 1 First, we consider min Ω=[b,Φ]> 1 kΩ − Ξ0 k2F + kλb 2 bk1 + kΛΦ Φk1 + ι{Φ=Φ> } . Let Φ0 = Ξ0Φ , b0 = Ξ0b . By some simple algebra, kΦ − Φ0 k2F = kΦ − (Φ0 + Φ0> )/2k2F + C(Φ0 ), where the last term does not depend on Φ. Based on Lemma 1 of She (2012), it is not difficult to show that the global minimizer is given by bo = ΘS (b0 ; λb ), Φo = ΘS ((Φ0 + Φ0> )/2; (ΛΦ + Λ> Φ )/2). 1 Similarly, the globally optimal solution to minΩ=[b,Φ]> 2 kΩ − Ξ0 k2F + ~ S (Ξ0 [:, k]; λΩ [k]), 1 ≤ k(λΩ 1> ) Ω> k2,1 is given by Ωo with Ωo [; , k] = Θ k ≤ p. Applying Theorem 3.2 and 3.3 in Bauschke and Combettes (2008) guarantees the strict convergence of the iterates and the global optimality of the limit point. 24 ~ is a Proof of Lemma 2 The conclusion follows from the fact that Θ proximity operator associated with a convex function. Proof of Lemma 3 From the optimality of Ω(i+1) and the convexity of the penalties in (28), it is easy to see that for any Ω, g(Ω(i+1) , Ω(i) ; λb , ΛΦ , ΛΩ ) + τ 2 (i+1) kΩ − Ωk2F ≤ g(Ω, Ω(i) ; λb , ΛΦ , λΩ ). 2 (32) It follows from the construction of g that g(Ω(i+1) , Ω(i) ; λb , ΛΦ , ΛΩ ) ≤ g(Ω(i) , Ω(i) ; λb , ΛΦ , λΩ ) = f (Ω(i) ; λb , ΛΦ , λΩ ). On the other hand, noticing that the gradient of l with respect to Φ is Z > diag{y − Xb0 − Z̄ vec (Φ0 ))}Z, Taylor expansion yields g(Ω, Ω0 ; λb , ΛΦ , ΛΩ ) ≥ f (Ω; λb , ΛΦ , λΩ ) + (τ 2 − kZ̆k22 ) 0 kΩ − Ωk2F 2 for any Ω, Ω0 . Furthermore, k(λΩ 1> ) Ω(i)> k2,1 ≤ f (Ω(i) ; λb , ΛΦ ) ≤ f (Ω(0) ; λb , ΛΦ ). Because λΩ > 0, the conclusion follows. A.2 Proof of Theorem 2 In this proof, we use C, c to denote universal constants. They are not necessarily the same at each occurrence. Throughout the proof, P (b, Φ; λ1 , λ2 ) is short for λ1 kΦk1 + λ2 kΩ> k2,1 , λ01 is short for λ1 kX̆k2 , and λ02 is short for λ2 kX̆k2 . We prove the theorem under a less restrictive condition. First, because Ω̂ is a global minimizer of (8) and P is convex, for any (b, Φ) ∈ Rp × Rp×p with Φ symmetric (not necessarily satisfying SH), we have 1 `(b̂, Φ̂) + P (b̂, Φ̂; λ01 , λ02 ) + M (b − b̂, Φ − Φ̂) ≤ `(b, Φ) + P (b, Φ; λ01 , λ02 ). 2 Based on the model assumption, it is easy to see that 1 1 M (b∗ − b̂, Φ∗ − Φ̂) + P (b̂, Φ̂; λ01 , λ02 ) + M (∆b , ∆Φ ) 2 2 1 ∗ ∗ 0 0 ≤ M (b − b, Φ − Φ) + P (b, Φ; λ1 , λ2 ) + hε, X∆b + diag(X∆Φ X > )i, 2 25 where ∆b = b̂ − b, ∆Φ = Φ̂ − Φ. ˘ = [∆> , (vec (∆Φ ))> ]> and J˜ = J ∪ Jˆ, where J := {j : Introduce ∆ b vec (Ω)j 6= 0}. and Jˆ := {j : vec (Ω̂)j 6= 0}. Denote by P J˜ the orthogonal projection matrix onto the column space of X̆ J˜ . Then hε, X∆b + diag(X∆Φ X > )i ˘ =hε, X̆ ∆i ˘ ˜i =hε, P J˜ X̆ J˜ ∆ J =hP J˜ ε, X∆b + diag(X∆Φ X > )i 1 a1 ≤ M (∆b , ∆Φ ) + kP J˜ εk22 , 2a1 2 for any a1 > 0. Representing kΩ̂k2,1 as k(Ω̂JG )> k2,1 + k(Ω̂JGc )> k2,1 and kΦ̂k1 as kΦ̂Je k1 + kΦ̂Jec k1 , respectively, and applying the triangle inequalities of k · k1 and k · k2,1 , we have 1 1 M (b∗ − b̂, Φ∗ − Φ̂) + λ01 k(∆Φ )Jec k1 + λ02 k(∆JGc )> k2,1 + M (∆b , ∆Φ ) 2 2 1 1 M (∆b , ∆Φ ) + λ01 k(∆Φ )Je k1 + λ02 k(∆JG )> k2,1 ≤ M (b∗ − b, Φ∗ − Φ) + 2 2a1 a1 + kP J˜ εk22 . 2 (33) How to treat the key term kP J˜ εk22 is nontrivial. We derive two inequalities from statistical analysis and computational analysis. First, we bound kP J˜ εk22 with some cardinality measures. The goal of this step is to show for a large constant L, supJ ⊂[p(p+1)] {kP J εk22 −LPo (JG (J ), Je (J ))} is negative with high probability, where Po (JG , Je ) = σ 2 {JG log(ep/JG ) + Je log(epJG /Je )}. (If the reference signal (b, Φ) satisfies SH, so does J˜, for which Po can be strengthened to σ 2 {JG log(ep/JG ) + Je log(eJG2 /Je )} following the lines of the proof of Lemma 4.) For notational convenience, we extend the J -measures defined in (2) to an index set: Given any J ⊂ [p(p + 1)], with Ω = [b, Φ> ]> a binary matrix satisfying (vec (Ω))J = 1 and (vec (Ω))J c = 0, define JG (J ) = JG (Ω), Je (J ) = Je (Φ). Given J ⊂ [p(p + 1)], JG (J ) ⊂ [p] and Je (J ) ⊂ [p2 ]. 26 Lemma 4. Given 1 ≤ JG ≤ p and 0 ≤ Je ≤ p2 , for any t ≥ 0, ! sup P \|JG \|=JG ,\|Je \|=Je kP J εk22 ≥ L0 Po (JG , Je ) + tσ 2 ≤ C exp(−ct), (34) where L0 , C, c > 0 are universal constants. Here, we omit the dependence of JG and Je on J for brevity. Define ! RJG ,Je = sup \|JG \|=JG ,\|Je \|=Je kP J εk22 − Lσ 2 {Je log(ep) + JG log(ep)} , + and R = sup1≤JG ≤p,0≤Je ≤p2 RJG ,Je (when JG (J ) = 0, P J ε = 0). Then kP J˜ εk22 ≤Lσ 2 {J˜e log(ep) + J˜G log(ep)} + R ≤Lσ 2 (Je + JG ) log(ep) + Lσ 2 {Jˆe log(ep) + JˆG log(ep)} + R. (35) Set L > L0 . Then, by Lemma 4 and Je ≥ JG , P(R ≥ tσ) 2 ≤ p p X X P(RJG ,Je ≥ tσ) JG =1 Je =0 2 ≤ p p X X C exp(−ct) exp {−c (L − L0 )(Je log(ep) + JG log(ep))} JG =1 Je =0 0 ≤C exp(−ct2 ), from which it follows that ER ≤ Cσ 2 . It is also easy to see that for sufficiently large L, R ≤ 0 occurs with probability at least 1 − Cp−cL . Next, we derive an inequality based on the computational optimality of (b̂, Φ̂). Due to the convexity of the problem, (b̂, Φ̂) is a stationary point of 1 ky − Xb − diag(XΦX > )k22 + λ01 kΦk1 2 X + λ02 kΩ> k2,1 + lj,k (φj,k − φk,j ), j<k 27 where lj,k are Lagrangian multipliers. By the KKT conditions, for any j ∈ JˆG , b̂j is nonzero (with probability 1) and so satisfies λ02 b̂j = −x> j (X̆ vec (Ω̂) − y). kΩ̂j k2 (36) 0 Similarly, for any j 0 ∈ Jˆe , letting jc = d jp e, jr = j 0 − (jc − 1)p, φ̂jr ,jc satisfies > φ̂ λ01 sgn(φ̂jr ,jc ) + λ02 kΩ̂jr ,jkc + ljr ,jc sgn(jr − jc ) = −X̄ j 0 (X̆ vec (Ω̂) − y), where jc 2 sgn(t) = 1, 0, −1 for t > 0, = 0, < 0, respectively. Adding the equations for φ̂jr ,jc and φ̂jc ,jr can cancel ljr ,jc . So when jr 6= jc , we have 2λ01 sgn(φ̂jr ,jc ) + λ02 φ̂jr ,jc kΩ̂jc k2 + λ02 φ̂jr ,jc kΩ̂jr k2 > = −2X̄ j 0 (X̆ vec (Ω̂) − y), (37) and when jc = jr , we have λ01 sgn(φ̂jr ,jc ) + λ02 φ̂jr ,jc > kΩ̂jc k2 = −X̄ j 0 (X̆ vec (Ω̂) − y). (38) Notice that sgn(φ̂jr ,jc )φ̂jr ,jc is always non-negative. Squaring both sides of (36), (37) and (38) and summing over all j ∈ JˆG and j 0 ∈ Jˆe , we obtain ˆ λ02 1 Je + λ02 2 X b̂2j + j∈JˆG P k:φ̂k,j 6=0 φ̂2k,j kΩ̂j k22 > ≤ 2kX̆ Jˆ (X̆ vec (Ω̂) − y)k22 , > 02 ˆ 2 ˆ ˆ ˜ and so λ02 1 Je + λ2 JG ≤ 2kX̆ Jˆ (X̆ vec (Ω̂) − y)k2 . It follows from J ⊂ J and the Cauchy-Schwarz inequality that λ21 Jˆe + λ22 JˆG ≤ 3M (b∗ − b̂, Φ∗ − Φ̂) + 6kP J˜ εk22 . (39) Now, combining the optimization inequality (39) and the statistical inp equality (35) can yield a bound of kP J˜ εk22 . In fact, with λo = σ log(ep), λ1 == A1 λo , λ2 = A2 λo , Ai ≥ A, A2 = A0 L for some A0 > 0, we get (1 − 3 6 )kP J˜ εk22 ≤ Lσ 2 (Je + JG ) log(ep) + M (b∗ − b̂, Φ∗ − Φ̂) + R. A0 A0 28 Plugging this into (33) gives 1 M (b∗ − b̂, Φ∗ − Φ̂) + Aλo kX̆k2 k(∆Φ )Jec k1 + Aλo kX̆k2 k(∆JGc )> k2,1 2 3a1 1 1 1 ∗ ∗ 1+ M (b − b, Φ − Φ) + − 1 M (∆b , ∆Φ ) ≤ 2 A0 − 6 2 a1 a1 L + Aλo kX̆k2 k(∆Φ )Je k1 + Aλo kX̆k2 k(∆JG )> k2,1 + σ 2 (Je + JG ) log(ep) 2(1 − 6/A0 ) a1 + R. 2(1 − 6/A0 ) Assume the following condition holds kX̆k2 (k(∆Φ )Je k1 + k(∆JG )> k2,1 ) ≤kX̆k2 (k(∆Φ )Jec k1 + k(∆JGc )> k2,1 ) + K p Je + JG kX∆b + X̄ vec (∆Φ )k2 , (40) for some K large enough. Then we have 1 M (b∗ − b̂, Φ∗ − Φ̂) 2 1 3a1 1 1 a1 R ∗ ∗ ≤ 1+ M (b − b, Φ − Φ) + − 1 M (∆b , ∆Φ ) + 2 A0 − 6 2 a1 2(1 − 6/A0 ) p a L 1 + Aλo K Je + JG kX∆b + X̄ vec (∆Φ )k2 + σ 2 (Je + JG ) log(ep) 2(1 − 6/A0 ) 1 3a1 1 1 1 ∗ ∗ ≤ 1+ M (b − b, Φ − Φ) + + − 1 M (∆b , ∆Φ ) 2 A0 − 6 2 a1 a2 a1 a2 A0 2 a1 + K Lσ 2 (Je + JG ) log(ep) + R. + 2(1 − 6/A0 ) 2 2(1 − 6/A0 ) Hence we get EM (b∗ − b̂, Φ∗ − Φ̂) . M (b∗ − b, Φ∗ − Φ) + (1 ∨ K 2 )σ 2 (Je + JG )σ 2 log(ep)+σ 2 if we choose the constants a1 , a2 , A0 satisfying, say, a1 = 2, a2 = 2, A0 ≥ 7. To complete the proof, we show that A(Je , JG , ϑ, κ) with κ > 0 and ϑ ≥ 0 implies (40). Consider two cases. The case (1 + ϑ)(k(∆Φ )Je k1 + k(∆JG )> k2,1 ) ≤ k(∆Φ )Jec k1 + k(∆JGc )> k2,1 is trivial. Suppose the reverse 29 inequality holds. Then kX̆k2 (k(∆Φ )Je k1 + k(∆JG )> k2,1 ) ≤kX̆k2 (Je + JG )1/2 (k(∆Φ )Je k22 + k(∆JG )> k2F )1/2 ≤(Je + JG )1/2 kX∆kF /κ1/2 , and so (40) holds with K = 1/κ1/2 . Finally, we can also plug (35) into (39), resulting in λ21 E[Jˆe ] + λ22 E[JˆG ] . M (b∗ − b, Φ∗ − Φ) + λ21 Je + λ22 JG + σ 2 , (41) and so (22) follows. The proof for Theorem 2’ follows the same lines; the details are omitted. Proof of Lemma 4 First, notice that for fixed J , kP J εk22 /σ 2 ∼ χ2 (D) with D ≤ J. The standard tail bound for X ∼ χ2 (D) gives P(X − D ≥ t) ≤ exp(−t2 /(4(D + t))), ∀t ≥ 0 (Laurent and Massart, 2000, Lemma 1). It is easy to see that sup \|JG \|=JG ,\|Je \|=Je kP J εk22 ≤ sup sup sup J 11 +J 10 +J 01 =JG Q(J 11 ,J 10 ,J 01 ) (e) \|Je \|=Je ,Je ⊂JG kP J εk22 , where Q(J 11 , J 10 , J 01 ) = {(J 11 , J 10 , J 01 ) ⊂ [p]3 : \|J 11 \| = J 11 , \|J 10 \| = (e) J 10 , \|J 01 \| = J 01 , J 11 ∩ J 10 = J 10 ∩ J 01 = J 11 ∩ J 01 = ∅} and JG denotes the index set {1 ≤ j ≤ p2 : d pj e ∈ JG }. Q(J 11 , J 10 , J 01 ) contains p 11 10 01 J , J , J , p − J 11 − J 10 − J 01 many elements. This multinomial coefficient is bounded by JpG , and by Stir ling’s approximation, log JpG ≤ JG log(ep/JG ). Similarly, sup\|Je \|=Je ,Je ⊂J (e) G (J 11 +J 01 )p JG +2 and supJ 11 +J 10 +J 01 =JG involve and 2 terms, respectively, and Je we have 11 J + J 01 )p JG + 2 log ≤ Je log(eJG p/Je ), log ≤ C log(eJG ). Je 2 Applying the union bound gives the desired result. 30 A.3 Proof of Theorem 3 The proof is based on the general reduction scheme in Chapter 2 of Tsybakov (2009). The key is to design proper least favorable signals in different situations. We consider two cases. (i) JG log(ep/JG ) ≤ Je log(eJG2 /Je ). Define a signal subclass B 1 (JG , Je ) = Ω = [b, Φ]> : bj = 1 if j ∈ [JG ] and 0 if JG < j ≤ p, φk,j = 0 or γR if (k, j) ∈ [JG ] × [JG ] and 0 otherwise, Φ = Φ> , Je (Φ) ≤ Je , σ where R = κ1/2 (log(eJG2 /Je ))1/2 and γ > 0 is a small constant to be chosen later. Because bj = 1 (1 ≤ j ≤ JG ), any Ω ∈ B 1 (JG , Je ) satisfies SH, and thus B 1 (JG , Je ) ⊂ SH(JG , Je ). By Stirling’s approximation, 2 J 1 log \|B (JG , Je )\| = log G ≥ Je log(JG2 /Je ) ≥ cJe log(eJG2 /Je ) Je for some universal constant c. Let ρ(Ω1 , Ω2 ) = kΩ1 − Ω2 k0 be the Hamming distance. By Lemma 8.3 in Rigollet and Tsybakov (2011), there exists a subset B 10 (JG , Je ) ⊂ B 1 (JG , Je ) such that log \|B 10 (JG , Je )\| ≥ c1 Je log(eJG2 /Je ), ρ(Ω1 , Ω2 ) ≥ c2 Je , ∀Ω1 , Ω2 ∈ B 10 , Ω1 6= Ω2 for some universal constants c1 , c2 > 0. Then kΩ1 −Ω2 k2F = γ 2 R2 ρ(Ω1 , Ω2 ) ≥ c2 γ 2 R2 Je . It follows from the restricted conditional number assumption that kX̆ vec (Ω1 ) − X̆ vec (Ω2 )k22 ≥ c2 κγ 2 R2 Je (42) for any Ω1 , Ω2 ∈ B 10 , Ω1 6= Ω2 , where κ/κ is a positive constant. For Gaussian models, the Kullback-Leibler divergence of N (X̆ vec (Ω2 ), σ 2 I) (denoted by PΩ2 ) from N (X̆ vec (Ω1 ), σ 2 I) (denoted by PΩ1 ) is K(PΩ1 , PΩ2 ) = 1 kX̆ vec (Ω1 ) − X̆ vec (Ω2 )k22 . Let P0 be N (0, σ 2 I). By the assumption 2σ 2 again, for any Ω ∈ B 1 (JG , Je ), 1 γ2 2 2 2 K(P0 , PΩ ) ≤ 2 κγ R ρ(0, Ω) ≤ 2 κR Je , 2σ σ 31 where we used ρ(Ω1 , Ω2 ) = Je (Φ1 − Φ2 ) ≤ 2Je . Therefore, 2 γ2 J 1 X 2 2 log G . K(P0 , PΩ ) ≤ γ Je log(eJG /Je ) ≤ 10 \|B \| c Je 10 (43) Ω∈B Combining (42) and (43) and choosing a sufficiently small value for γ, we can apply Theorem 2.7 of Tsybakov (2009) to get the desired lower bound. (ii) JG log(ep/JG ) > Je log(eJG2 /Je ). Define a signal subclass B 2 (JG ) = {Ω = [b, Φ> ]> : bj ∈ {0, 1} ∀j ∈ [p], Φ = 0, JG (Ω) ≤ JG }, σ where R = κ1/2 {log(ep/JG )}1/2 and γ > 0 is a small constant. The afterward treatment is similar to (i). The details are omitted. For WH, we redefine B 1 as B 1 (JG , Je ) = {Ω = [b, Φ> ]> : bj = 1 if j ∈ [JG ] and 0 otherwise, φk,j = 0 or γR if k ≥ j and 1 ≤ j ≤ JG , σ and 0 otherwise, Jew (Φ) ≤ Je }, where R = κ1/2 (log(eJG p/Je ))1/2 . Then B 1 (JG , Je ) ⊂ WH(JG , Je ). The rest follows the same lines as the proof for SH. B Coordinatewise error bound for support recovery In this part, we show that GRESH estimators can recover the sparsity pattern of the true signal, i.e., J ∗ = {j : vec (Ω∗ )j 6= 0}, with high probability. The result is of the type of Lounici (2008) and Ravikumar et al. (2010), but we do not have to assume the stringent irrepresentable conditions or mutual coherence conditions. For more discussions, see Zhang (2009). Assumption C(Je , JG , ζ, λ1 , λ2 ). Given Je ⊂ [p2 ], JG ⊂ [p], ζ ≥ 0, λ1 , λ2 > > 0, for any ∆ = [∆b , ∆> ∈ R(p+1)×p satisfying ∆Φ = ∆> Φ] Φ , the following inequality holds kX̆k2 (k(∆Φ )Je k1 + k(∆JG )> k2,1 ) + (Je + JG )k vec (∆)k2∞ /(2λ1 ∧ 2λ2 ) p ≤ kX̆k2 (k(∆Φ )Jec k1 + k(∆JGc )> k2,1 ) + ζ Je + JG kX∆b + X̄ vec (∆Φ )k2 . (44) p Theorem 4. Assume ∼ N (0, σ 2 I). Take λ1 = A1 σ log(ep) and λ2 = p A2 σ log(ep) in (17)and suppose (X, X̄, X̆) satisfies C(Je∗ , JG∗ , ζ, λ1 , λ2 ) for 32 > > be a some ζ ≥ 0, where Je∗ = Je (Φ∗ ), JG∗ = JG (b∗ , Φ∗ ). Let Ω̂ = b̂, Φ̂ p global minimizer of (17). Define Jˆ = {j : \| vec (Ω̂)j \| > 1 + ζ 2 (λ1 ∧ λ2 )}, and J ∗ = {j : vec (Ω∗ )j 6= 0}. Then for sufficiently large constants A1 and 2 2 A2 , with probability at least 1 − Cp−c(A1 ∧A2 ) , the estimate Ω̂ satisfies p k vec (Ω̂ − Ω∗ )k∞ ≤ 1 + ζ 2 (λ1 ∧ λ2 ). If, in addition, the minimum signal strength satisfies p min∗ \| vec (Ω∗ )j \| > 2 1 + ζ 2 (λ1 ∧ λ2 ), j∈J (45) 2 2 then Jˆ = J ∗ , with probability at least 1 − Cp−c(A1 ∧A2 ) . To recover the true support, signal strength conditions like (45) must be imposed. Interestingly, compared with Theorem 5.1 in Lounici et al. (2011), hierarchical variable selection can accommodate smaller signals than group variable selection. Proof. From the proof of Theorem 2, we get the following inequality with probability at least 1 − Cp−cL 1 M (b∗ − b̂, Φ∗ − Φ̂) + Aλo ρk(∆Φ )Jec k1 + Aλo ρk(∆JGc )> k2,1 2 1 3a1 1 1 ≤ (1 + )M (b∗ − b, Φ∗ − Φ) + ( − 1)M (∆b , ∆Φ ) 2 A0 − 6 2 a1 a1 L + Aλo ρk(∆Φ )Je k1 + Aλo ρk(∆JG )> k2,1 + σ 2 (Je + JG ) log(ep), 2(1 − 6/A0 ) p where ρ = kX̆k2 , L is a universal constant, λo = σ log(ep), A2 = A0 L, and a1 , A0 are positive constants to be determined. Also recall that λ1 = A1 λo , λ2 = A2 λo , A = A1 ∧ A2 . 33 By the regularity condition, we have 1 1 (Je + JG )k vec (∆)k2∞ + M (b∗ − b̂, Φ∗ − Φ̂) 2 2 1 3a1 1 1 ∗ ≤ (1 + )M (b − b, Φ∗ − Φ) + ( − 1)M (∆b , ∆Φ ) 2 A0 − 6 2 a1 p a1 L + Aλo ζ Je + JG kX∆b + X̄ vec (∆Φ )k2 + σ 2 (Je + JG ) log(ep) 2(1 − 6/A0 ) 1 3a1 1 1 1 )M (b∗ − b, Φ∗ − Φ) + ( + − 1)M (∆b , ∆Φ ) ≤ (1 + 2 A0 − 6 2 a1 a2 a2 A0 2 a1 + ζ Lσ 2 (Je + JG ) log(ep), + 2(1 − 6/A0 ) 2 for any a2 > 0. Taking b = b∗ , Φ = Φ∗ and choosing a1 = a2 = 1, A0 ≥ 7 give (Je∗ +JG∗ )k vec (Ω̂−Ω∗ )k2∞ ≤ A0 (1 + ζ 2 )Lσ 2 (Je∗ +JG∗ ) log(ep). The conclusion follows. C An ADMM algorithm We describe an ADMM algorithm for solving problem (16). Recall Z̄ = [z 1 z 1 , · · · , z 1 z p , · · · , z p z 1 , · · · , z p z p ], Z̆ = [x1 , z 1 z 1 , · · · , z 1 z p , · · · , xp , z p z 1 , · · · , z p z p ], and ΛΩ = 1λ> Ω . To apply ADMM, we rewrite (16) into 1 min ky − Xb − diag(ZΦZ > )k22 + kΛ> Ω + kλb Γb k1 + kΛΦ ΓΦ k1 Ω,Γ,Υ 2 Υ> k2,1 (46) s.t. ΓΦ = Γ> Φ , Γ = Ω, Υ = Ω, 2 2 2 > p p ×p where Υ ∈ Rn×(p +p) , Γ = [Γb , Γ> . The augΦ ] with Γb ∈ R , ΓΦ ∈ R mented Lagrangian of (46) can be formed by use of two Lagrangian multiplier matrices (L1 , L2 ) and a penalty parameter ρ: Lρ (Ω, Γ, Υ, L1 , L2 ) 1 = ky − Z̆vec(Ω)k22 + kλb Γb k1 + kΛΦ ΓΦ k1 + ι{ΓΦ =Γ>Φ } + kΛ> Υ> k2,1 Ω 2 ρ + ρ1> (L1 (Ω − Γ) + L2 (Ω − Υ))1 + (kΩ − Γk2F + kΩ − Υk2F ). 2 34 Based on the proof of Theorem 1, we can solve the sub-optimization problems of Γ and Υ by some proximity operators; the details are omitted. The full ADMM algorithm is given as follows. Algorithm 2 An ADMM algorithm Inputs: Data: X, Z, y. Regularization parameters: λb , ΛΦ , λΩ . repeat −1 > > 1. vec(Ω) ← Z̆ Z̆ + 2ρI Z̆ y + ρ vec(Γ + Υ − L1 − L2 ) ; ~ S (Ω[:, k] + L2 [:, k]; λΩ [k]/ρ), ∀k : 1 ≤ k ≤ p; 2. Υ[:, k] ← Θ 3. Ω[2 : end, :] ← (Ω[2 : end, :] + Ω[2 : end, :]> )/2; > 4. Γ ← ΘS (Ω + L1 ; [λb , Λ> Φ ] /ρ); 5. L2 ← L2 + Ω − Υ; 6. L1 ← L1 + Ω − Γ; until convergence Output: Ω References Bauschke, H. H. and P. L. Combettes (2008). “A Dykstra-like Algorithm for Two Monotone Operators”. Pacific Journal of Optimization 4 (3), 383– 391. Bickel, P. J., Y. Ritov, and A. B. Tsybakov (2009). “Simultaneous Analysis of Lasso and Dantzig Selector”. The Annals of Statistics, 1705–1732. Bickel, P. J., Y. Ritov, and A. B. Tsybakov (2010). “Hierarchical Selection of Variables in Sparse High-Dimensional Regression”. In Borrowing Strength: Theory Powering Applications–A Festschrift for Lawrence D. Brown, pp. 56–69. Institute of Mathematical Statistics. Bien, J., F. Bunea, and L. Xiao (2016). “Convex Banding of the Covariance Matrix”. Journal of the American Statistical Association 111 (514), 834– 845. Bien, J., J. Taylor, and R. Tibshirani (2013). “A Lasso for Hierarchical Interactions”. The Annals of Statistics 41 (3), 1111–1141. 35 Boyd, S., N. Parikh, E. Chu, B. Peleato, and J. Eckstein (2011). “Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers”. Foundations and Trends R in Machine Learning 3 (1), 1–122. Chipman, H. (1996). “Bayesian Variable Selection with Related Predictors”. Canadian Journal of Statistics 24 (1), 17–36. Choi, N. H., W. Li, and J. Zhu (2010). “Variable Selection With the Strong Heredity Constraint and its Oracle Property”. Journal of the American Statistical Association 105 (489), 354–364. Cohen, J., P. Cohen, S. G. West, and L. S. Aiken (2013). Applied Multiple Regression/Correlation Analysis for the Behavioral Sciences. Routledge. Davidson, E. H. and D. H. Erwin (2006). “Gene Regulatory Networks and the Evolution of Animal Body Plans”. Science 311 (5762), 796–800. Hamada, M. and C. J. Wu (1992). “Analysis of Designed Experiments With Complex Aliasing”. Journal of Quality Technology 24 (3), 130–137. Hao, N. and H. H. Zhang (2014). “Interaction Screening for Ultra-High Dimensional Data”. Journal of the American Statistical Association 109 (1), 1285–1301. Hastie, T., R. Tibshirani, and J. Friedman (2009). The Elements of Statistical Learning (2nd ed.). New York: Springer-Verlag. He, Y., Y. She, and D. Wu (2013). “Stationary-Sparse Causality Network Learning”. Journal of Machine Learning Research 14 (1), 3073–3104. Jenatton, R., J. Mairal, G. Obozinski, and F. Bach (2011). “Proximal Methods for Hierarchical Sparse Coding”. Journal of Machine Learning Research 12, 2297–2334. Laurent, B. and P. Massart (2000). “Adaptive Estimation of a Quadratic Functional by Model Selection”. The Annals of Statistics, 1302–1338. Lim, M. and T. Hastie (2015). “Learning Interactions via Hierarchical GroupLasso Regularization”. Journal of Computational and Graphical Statistics 24 (3), 627–654. 36 Lounici, K. (2008). “Sup-Norm Convergence Rate and Sign Concentration Property of Lasso and Dantzig Estimators”. Electronic Journal of Statistics 2, 90–102. Lounici, K., M. Pontil, S. Van De Geer, and A. B. Tsybakov (2011). “Oracle Inequalities and Optimal Inference Under Group Sparsity”. The Annals of Statistics 39 (4), 2164–2204. McCullagh, P. and J. A. Nelder (1989). Generalized Linear Models. London England Chapman and Hall. Negahban, S. N., P. Ravikumar, M. J. Wainwright, and B. Yu (2012). “A Unified Framework for High-Dimensional Analysis of M -Estimators With Decomposable Regularizers”. Statistical Science 27 (4), 538–557. Nelder, J. A. (1977). “A Reformulation of Linear Models”. Journal of the Royal Statistical Society. Series A (General), 48–77. Nesterov, Y. (2007). “Gradient Methods for Minimizing Composite Objective Function”. Technical report, Université catholique de Louvain, Center for Operations Research and Econometrics (CORE). Opial, Z. (1967). “Weak Convergence of the Sequence of Successive Approximations for Nonexpansive Mappings”. Bulletin of the American Mathematical Society 73 (4), 591–597. Owen, A. B. (2007). “A Robust Hybrid of Lasso and Ridge Regression”. Contemporary Mathematics 443, 59–72. Pace, R. K. and R. Barry (1997). “Sparse Spatial Autoregressions”. Statistics & Probability Letters 33 (3), 291–297. Peixoto, J. L. (1987). “Hierarchical Variable Selection in Polynomial Regression Models”. The American Statistician 41 (4), 311–313. Peixoto, J. L. (1990). “A Property of Well-Formulated Polynomial Regression Models”. The American Statistician 44 (1), 26–30. Radchenko, P. and G. M. James (2010). “Variable Selection Using Adaptive Nonlinear Interaction Structures in High Dimensions”. Journal of the American Statistical Association 105 (492), 1541–1553. 37 Ravikumar, P., M. J. Wainwright, and J. D. Lafferty (2010). “HighDimensional Ising Model Selection Using `1 -Regularized Logistic Regression”. The Annals of Statistics 38 (3), 1287–1319. Ravikumar, P. D., H. Liu, J. D. Lafferty, and L. A. Wasserman (2007). “SpAM: Sparse Additive Models”. In In Advances in Neural Information Processing Systems 20, pp. 1201–1208. MIT Press. Rigollet, P. and A. Tsybakov (2011). “Exponential Screening and Optimal Rates of Sparse Estimation”. The Annals of Statistics 39 (2), 731–771. Schmidt, M., N. L. Roux, and F. R. Bach (2011). “Convergence Rates of Inexact Proximal-Gradient Methods for Convex Optimization”. In J. ShaweTaylor, R. Zemel, P. Bartlett, F. Pereira, and K. Weinberger (Eds.), Advances in Neural Information Processing Systems 24, pp. 1458–1466. She, Y. (2012). “An Iterative Algorithm for Fitting Nonconvex Penalized Generalized Linear Models With Grouped Predictors”. Computational Statistics & Data Analysis 56 (10), 2976–2990. She, Y., Y. He, and D. Wu (2014). Learning topology and dynamics of large recurrent neural networks. IEEE Transactions on Signal Processing 62, 5881–5891. Simon, N., J. Friedman, T. Hastie, and R. Tibshirani (2013). “A SparseGroup Lasso”. Journal of Computational and Graphical Statistics 22 (2), 231–245. Tibshirani, R. (1996). “Regression Shrinkage and Selection via the Lasso”. Journal of the Royal Statistical Society. Series B (Methodological), 267– 288. Tsybakov, A. (2009). Introduction to Nonparametric Estimation. Springer Series in Statistics. New York: Springer. van de Geer, S. (2014). “Weakly Decomposable Regularization Penalties and Structured Sparsity”. Scandinavian Journal of Statistics 41 (1), 72–86. Wei, F. and J. Huang (2010). “Consistent Group Selection in HighDimensional Linear Regression”. Bernoulli 16 (4), 1369–1384. 38 Wu, J., B. Devlin, S. Ringquist, M. Trucco, and K. Roeder (2010). “Screen and Clean: A Tool for Identifying Interactions in Genome-Wide Association Studies”. Genetic epidemiology 34 (3), 275–285. Yuan, M. and Y. Lin (2006). “Model Selection and Estimation in Regression With Grouped Variables”. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 68 (1), 49–67. Zhang, C.-H. and J. Huang (2008). “The Sparsity and Bias of the Lasso Selection in High-Dimensional Linear Regression”. The Annals of Statistics, 1567–1594. Zhang, T. (2009). “Some Sharp Performance Bounds for Least Squares Regression With L1 Regularization”. The Annals of Statistics 37 (5A), 2109– 2144. Zou, H. and T. Hastie (2005). “Regularization and Variable Selection via the Elastic Net”. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 67 (2), 301–320. 39	10math.ST
Making up for the deficit in a marathon run Iztok Fister Jr.,1 Dušan Fister,2 Suash Deb,3, 4 Uroš Mlakar,5 Janez Brest,5 and Iztok Fister5 1 University of Maribor, Faculty of Electrical Engineering and Computer Science, Smetanova 17, 2000 Maribor, Slovenia; Email: iztok.fister1@um.si 2 University of Maribor, Faculty of Mechanical Engineering, Smetanova 17, 2000 Maribor, Slovenia 3 arXiv:1705.03302v1 [cs.NE] 9 May 2017 4 IT & Educational Consultant, Ranchi, Jharkhand, India Distinguished Professorial Associate, Decision Sciences and Modelling Program, Victoria University, Melbourne, Australia 5 University of Maribor, Faculty of Electrical Engineering and Computer Science, Smetanova 17, 2000 Maribor, Slovenia (Dated: 10 February 2017) To predict the final result of an athlete in a marathon run thoroughly is the eternal desire of each trainer. Usually, the achieved result is weaker than the predicted one due to the objective (e.g., environmental conditions) as well as subjective factors (e.g., athlete’s malaise). Therefore, making up for the deficit between predicted and achieved results is the main ingredient of the analysis performed by trainers after the competition. In the analysis, they search for parts of a marathon course where the athlete lost time. This paper proposes an automatic making up for the deficit by using a Differential Evolution algorithm. In this case study, the results that were obtained by a wearable sports-watch by an athlete in a real marathon are analyzed. The first experiments with Differential Evolution show the possibility of using this method in the future. 2 I. INTRODUCTION Running a marathon is a very challenging task for every athlete. Good and efficient preparation for a marathon is a very complex process that must be performed many time before the competition. The marathon is not just an ordinary running competition, but it is an epic distance having roots in ancient Greece. The marathon name was coined by the legend of soldier Philippides, who ran from the Battle of Marathon to Athens in order to announce the victory of the united Greeks against the Persian force. Nowadays, most of the big cities organize marathon competitions annually that attract a lot of people. Some athletes just want to finish a marathon, while others have higher goals. In line with this, the amateur athletes are satisfied with an achievement of around four hours, while more competitive amateur athletes try to beat the so-called magic line, the three hours. Although this magic result was set many years ago, running a marathon race sub three hours is still considered as a good result for a well-trained amateur athletes. However, those who would like to reach the result of sub three hours not only need the proper sports training, but also a lot of knowledge about a race course. Additionally, awareness of characteristics about an athlete’s body is a crucial step in the evolution of runners, i.e., when to go faster, when to eat, when to slow down, how to pace on hills are some of the most significant questions that runners ask themselves when planning their running pace. Making up for the deficit time in a marathon run is a very demanding task4 that comes into play when the final results as predicted by sports trainers are slower than the achieved one by only a few seconds. Thus, more experienced athletes also calculate the pace for each kilometer of a race many weeks before the race start in order to predict their final achievement. However, the achieved result can distinguish from the predicted, because of objective, as well as subjective factors. The objective factors represent environmental conditions such as weather, temperature, wind position and even humidity, while the subjective factors refer to the athlete’s psycho-physical conditions such as feelings on the race day. Nowadays, predicting the final result is a little bit easier using the modern mobile technology. In this group, we can count professional sports-watches that consist also of Global Positioning System (GPS) sensors3 , heart rate measuring sensors, cadence meters, etc. Additionally, there is a comprehensive software support and some intelligent systems have also been arisen recently. In this paper, we present a novel solution for making up for the deficit marathon run that is based on the data collected from sports-watches worn by the athlete during the run. The primary aim of 3 this study is to help athletes who did not achieve the predicted final results by a few seconds and show them on which parts of the predefined course they could improve the intermediate pace so that the predicted results could be reached. This solution is implemented using Differential Evolution (DE)6 , which calculates the improvement of the intermediate pace, based on the configuration of the running course. These improvements are considered as constraints in the DE and they base on the traditional rule of the marathon run, i.e., more seconds can be caught up when the running course is flat. The secondary aim of this approach is also to help athletes in their analysis of the already performed marathon race. In our case study, we focus on the situation where an athlete tries to achieve the magic marathon time of three hours. The paper is structured in the remainder as follows. In Section II, making up for the deficit in a marathon run is formulated as a Constraint Satisfaction Problem (CSP). Section III is devoted to Differential Evolution. Section IV illustrates the case study, in which the case of an athlete in the Three hearts marathon in 2012 is taken into consideration. Experiments are presented in the second part of Section IV. The paper is concluded with Section V, where the work done is reviewed and directions are outlined for further research. II. PROBLEM FORMULATION Making up for the deficit time can be formulated as a CSP2 , where the solution is specified as a vector p = (p1 , . . . , pn )T of n problem variables representing the intermediate pace deficit for each of the observed kilometer. Additionally, a variable bound constraint is attached to each problem (L) (U ) (L) variable that limits the proper values of each variable to the interval pi ∈ [pi , pi ], where pi (U ) and pi are lower and upper bounds, respectively. In the case of a marathon race, where the length of the whole course equals 42.195 km, the length of each vector is, consequently, n = 43 elements. Then, the problem is defined in generalized form as follows: n X max f (p) = pi ≤ Ta − Tp , (1) i=1 subject to (L) pi ≤ pi ≤ (U ) pi , i = 1, . . . , n; where Ta − Tp denotes the deficit time that must be made up and the relation Ta ≥ Tp is valid. Thus, the following inequality constraints must be satisfied: n X (L) pi ≥ Ta − Tp , (2) i=1 which ensures that the problem is solvable. Although the upper bounds can be set as negation (U ) of the corresponding lower bounds (i.e., pi (L) = −pi in general, here the upper bounds are fixed to 4 (U ) zero (i.e., pi = 0). This mean that only negative values of intermediate pace deficits are allowed in our study. (L) The lower bound values pi are calculated from the configuration of the course according to the following assumptions:    = 0, flat, 7 −2 sec, →   Alt i = < 0, downhill, → 7 −4 sec,     > 0, uphill, 7 0 sec, → (3) where Alt i is an average altitude obtained in the appropriate kilometer. The speculation behind the assumptions is to speed up, when the course is flat or downward and to retain the speed when running upward. The quantitative values for estimation of deficit time is also taken from the real marathon practice, where it holds that the deficit time can be a maximum of 4 seconds in one kilometer when the course is flat. III. DIFFERENTIAL EVOLUTION Differential Evolution (DE) is an Evolutionary Algorithm appropriate for continuous and combinatorial optimization that was introduced by Storn and Price in 19951,6 . DE is a population-based algorithm5 that consists of Np real-coded vectors representing the candidate solutions, as follows: (t) (t) (t) xi = (xi,1 , . . . , xi,n ), for i = 1, . . . , Np, (t) (L) (U ) (4) (L) where each element of the solution is in the interval xi,1 ∈ [xi , xi ], and xi (U ) and xi denote the lower and upper bounds of the i-th variable, respectively. The variation operator in DE supports a differential mutation and a differential crossover. In particular, the differential mutation selects two solutions randomly and adds a scaled difference between these to the third solution. This mutation can be expressed as follows: (t) (t) (t) (t) ui = xr1 + F · (xr2 − xr3 ), for i = 1, . . . , Np, (5) where F denotes the scaling factor as a positive real number that scales the rate of modification while r1, r2, r3 are randomly selected values in the interval [1 . . . Np]. Note that, typically, the interval F ∈ [0.1, 1.0] is used in the DE community. As a differential crossover, uniform crossover is employed by the DE, where the trial vector is built from parameter values copied from two different solutions. Mathematically, this crossover 5 can be expressed as follows: (t+1) wi,j =   u(t) i,j randj (0, 1) ≤ CR ∨ j = jrand ,  x(t) i,j otherwise, (6) where CR ∈ [0.0, 1.0] controls the fraction of parameters that are copied to the trial solution. Note, (t) the relation j = jrand ensures that the trial vector is different from the original solution xi . A differential selection is, in fact, a generalized one-to-one selection that can be expressed mathematically as follows: (t+1) xi =   w(t) i if f (wi ) ≤ f (xi ),  x(t) i otherwise . (t) (t) (7) In a technical sense, crossover and mutation can be performed in several ways in Differential Evolution. Therefore, a specific notation is used to describe the varieties of these methods (also strategies) generally. For example, ’DE/rand/ 1/bin’ denotes that the base vector is selected randomly, 1 vector difference is added to it, and the number of modified parameters in the mutant vector follows binomial distribution. A. Making up for the deficit time using DE Modifying the original DE to solve a problem of making up for the deficit time in a marathon run (let us call a new variant marathon DE or simply mDE) is relatively simple, because DE implements the variable bounds implicitly. Therefore, the main problem remains how to set the appropriate variable bounds for elements of solution vectors. The answer to this question is given when intermediate pace of an athlete is taken into consideration. The intermediate paces are, nowadays obtained using the mobile sports-watches worn by the marathon runner during the run. There are four phases of modifying the original DE for making up for the deficit time in a marathon run, as follows: • defining the objective • preparing the bounds for problem variables • mapping the solution from decision space to problem space • evaluating the solution in problem space 6 Indeed, defining the objective refers to the expression Ta − Tp in Eq. (2), where we define the total deficit of time that must be made up. Preparing the bounds for problem variables pi,j bases on tracked data from sports-watches referring to the information about altitude as overcome by the marathon runner in each kilometer (Eq. (3)). According to these data, the appropriate boundary (t) values are calculated according to Eq. (3). The mapping of solution from decision space xi to its (t) counterpart in the problem space pi is straightforward and obeys the following equation: & (t) pi,j = (U ) pj (U ) xj (L) − pj − (L) xj ' · (t) xi,j · 10 · 10−1 , (8) where multiplying by 10−1 ensures that the given result in seconds is extended with the most significant digit of milliseconds after decimal point. Finally, the fitness function presented in Eq. (1) is used in our study. (t) In order to simplify solving of the problem, the last element in vector pi that does not represent the whole kilometer is excluded from the optimization, i.e., it is taken as an uphill part, although it can be downhill in practice as is in our case. IV. CASE STUDY: THE THREE HEARTS MARATHON 2012 The Three Hearts marathon has been organized annually in the city of Radenci (Slovenia) since 1981. During the years, the organizers have changed the marathon course slightly. In 2012, the marathon course consisted of two 21.1 kilometer long laps as can be seen in Figure 1. The goal of FIG. 1. Three Hearts marathon course in 2012. an athlete in our study was to run the marathon sub 3:00 hours. However, to run the marathon in this time, the intermediate pace should be on average 4:15 minutes per kilometer (min/km) (Figure 2) or better (less than 4:15). The athlete’s plan was to run the first kilometers a little bit faster than the others. Although the course was flat on average, he did not focus enough on 7 FIG. 2. Average pace for each kilometer. some ascents with a slight uphill that were large enough to lose a power, especially in bad weather conditions. After 30 km, the time was still good enough for achieving the result under 3:00 hours, but the last kilometers were pretty hard to overcome because of a formerly broken nail, abrasions from the heat and many blisters. Therefore, the athlete was one minute slower at the finish than planned. FIG. 3. Ascents and descents on the course of Three Hearts marathon at 2012. The proposed approach for making up for the deficit time works as follows. The objective is expressed as: 3 : 01 : 09.4 − 3 : 00 : 00.0 = 1 : 09.4 min ≈ 70 sec. An altitude graph of the marathon course, from which problem variable bounds were prepared, is presented in Figure 3. 8 The graph in Figure 3 presents differences of altitudes in meters according to the run distances in kilometers that is divided into two parts denoting 21.1 km by a vertical line. Although both parts must be the same in theory, this assumption does not hold in practice, because GPS technology, used by sports-watches, is far from ideal for now. However, the tracked data are accurate enough for the purposes of our study. The results of an analysis of altitude data are as follows. The GPS receiver detected 4 flat parts, 22 uphill‘s and 17 downhills. The last 195 meters were downward. Essentially, the difference between the number of uphill‘s and downhills is a consequence of errors in the GPS receiver. On the other hand, there were no serious uphill‘s and downhills on the marathon course as can be seen from the maximum and minimum differences of the altitude (i.e., ±10 meters). In fact, the difference between average ascent and average descent values at the end of each kilometer higher than ±1 meter is declared as uphill or downhill in the study. (t) Bounds for problem variables pi were prepared based on the analysis of the altitude data according to Eq. (3). The results of this phase are illustrated in Table I under columns ”Lower P (L) bounds” and ”Upper bounds”. The sum of the lower bounds of all problem variables is nj=1 pi,j = 82, whereby a condition of the problem solvability in Eq. (2) was satisfied. The last two phases, i.e., mapping and evaluating the solution, were performed using the mDE. For our experiments, mDE was coded in Python programming language and the ’DE/rand/1/bin’ strategy was employed. The input of the algorithm were data obtained from the Three Hearts marathon. The dimension of the problem was n = 43, population size was set to Np = 100, scaling factor F = 0.5 and crossover rate CR = 0.9. Lower and upper bounds were set in the (t) interval xi,j ∈ [0, 1]. The algorithm terminated when either fitness function fulfilled the objective and, thus, all constraints were satisfied 100 times, or the number of fitness function evaluations exceeded 80,000. Only one run of the mDE was performed in this study. The obtained results of the mDE algorithm are presented in Table I that consists of six columns. The first column counts kilometers during the marathon run, while the second illustrates actual intermediate pace, as already presented in Figure 2. The third column depicts the predicted intermediate pace obtained after decrementing the actual intermediate pace by the intermediate pace deficit found in the first run that is presented in column four. The last values are obtained after conducting the mDE algorithm. As already mentioned, the last two columns present lower and upper bounds. 9 TABLE I. Results of making up for the deficit in the Three Hearts marathon Distance Actual pace (sec) Predicted pace (sec) Difference Lower bounds Upper bounds [km] [min/km] [min/km] [sec] [sec] [sec] 1 04:03.80 04:03.80 0 0 0 2 04:05.40 04:03.60 1.8 2 0 3 04:09.70 04:05.90 3.8 4 0 4 04:12.20 04:08.50 3.7 4 0 5 04:15.00 04:15.00 0 0 0 6 04:07.70 04:04.30 3.4 4 0 7 04:09.70 04:09.70 0 0 0 8 04:11.20 04:11.20 0 0 0 9 04:11.00 04:07.80 3.2 4 0 10 04:10.30 04:10.30 0 0 0 11 04:13.80 04:10.00 3.8 4 0 12 04:08.80 04:07.00 1.8 2 0 13 04:09.10 04:07.30 1.8 2 0 14 04:10.70 04:09.70 1 2 0 15 04:07.90 04:07.90 0 0 0 16 04:14.40 04:12.70 1.7 2 0 17 04:11.20 04:07.30 3.9 4 0 18 04:13.00 04:13.00 0 0 0 19 04:10.00 04:06.40 3.6 4 0 20 04:13.60 04:13.60 0 0 0 21 04:03.00 04:03.00 0 0 0 22 04:06.40 04:06.40 0 0 0 23 04:10.00 04:08.30 1.7 2 0 24 04:12.00 04:08.50 3.5 4 0 25 04:23.80 04:21.20 2.6 4 0 26 04:08.40 04:08.40 0 0 0 27 04:08.00 04:04.30 3.7 4 0 28 04:07.60 04:03.80 3.8 4 0 29 04:24.40 04:22.80 1.6 2 0 30 04:17.00 04:14.40 2.6 4 0 31 04:09.30 04:09.30 0 0 0 32 04:13.70 04:10.30 3.4 4 0 33 04:14.70 04:10.80 3.9 4 0 34 04:12.60 04:11.00 1.6 2 0 35 04:18.60 04:18.60 0 0 0 36 04:31.90 04:31.90 0 0 0 37 04:27.80 04:27.80 0 0 0 38 04:37.70 04:33.90 3.8 4 0 39 04:37.40 04:34.60 2.8 4 0 40 04:45.20 04:43.70 1.5 2 0 41 04:53.70 04:53.70 0 0 0 42 04:52.40 04:52.40 0 0 0 42.195 01:45.30 01:45.30 0 0 0 Total: 03:01:09.40 02:59:59.40 70 82 0 A. Discussion At a first glance, the conducted experiments are in contrast with the principles of stochastic population-based optimization, where we are interested in the mean values obtained after the number of runs. Strictly speaking, this is true, when the results are obtained from the aforementioned 10 point of view. However, the aim of our study was twofold: On the one hand, to show that the mDE algorithm can be employed for making up the deficit in a marathon race, while on the other hand, to prove that there are many solutions of this problem. Both assumptions were justified by finding numerous feasible solutions in just one run. Here, the question arises ”Could we say, which of the solutions found is really optimal?”. However, the answer to the question must be answered by professionals, i.e., sports trainers and coaches together with athletes. V. CONCLUSION This paper reports the first successful application of the mDE for making up for the deficit in a marathon run based on the history data obtained by a sports-tracker on the corresponding marathon course. Making up for the deficit in a marathon run is a complex task that is usually performed by professional trainers and even athletes themselves. We defined this problem as a constraint problem and solved it by the mDE. Preliminary experiments shown that proposed approach may be used in the real-world. However, there are still many open directions for further research in this problem, e.g., to determine the most appropriate solution for a specific athlete from a huge set of solutions. In line with this, context dependent information need to be accumulated, on which basis such a solution can be proposed by the mDE algorithm automatically. However, extending input data with a heart rate monitor might also increase the prediction possibilities of the proposed algorithm. 1 S. Das, S. S. Mullick, and P. Suganthan. Recent advances in differential evolution–an updated survey. Swarm and Evolutionary Computation, 27:1–30, 2016. 2 A. E. Eiben and J. E. Smith. Introduction to evolutionary computing, volume 53. Springer, 2003. 3 I. Fister, D. Fister, S. Fong, and I. Fister Jr. Widespread mobile devices in applications for real-time drafting detection in triathlons. Journal of emerging technologies in web intelligence, 5(3):310–321, 2013. 4 I. Fister Jr., K. Ljubič, P. N. Suganthan, M. Perc, and I. Fister. Computational intelligence in sports: Challenges and opportunities within a new research domain. Applied Mathematics and Computation, 262:178–186, 2015. 5 I. Fister Jr., X.-S. Yang, I. Fister, J. Brest, and D. Fister. A brief review of nature-inspired algorithms for optimization. Elektrotehniški vestnik, 80(3):116–122, 2013. 6 R. Storn and K. Price. Differential evolution–a simple and efficient heuristic for global optimization over continuous spaces. Journal of global optimization, 11(4):341–359, 1997.	9cs.NE
1 Hybrid Precoder and Combiner Design with Low Resolution Phase Shifters in mmWave MIMO Systems arXiv:1710.06192v1 [cs.IT] 17 Oct 2017 Zihuan Wang, Student Member, IEEE, Ming Li, Senior Member, IEEE, Qian Liu, Member, IEEE, and A. Lee Swindlehurst, Fellow, IEEE Abstract—Millimeter wave (mmWave) communications have been considered as a key technology for next generation cellular systems and Wi-Fi networks because of its advances in providing orders-of-magnitude wider bandwidth than current wireless networks. Economical and energy-efficient analog/digial hybrid precoding and combining transceivers have been often proposed for mmWave massive multiple-input multiple-output (MIMO) systems to overcome the severe propagation loss of mmWave channels. One major shortcoming of existing solutions lies in the assumption of infinite or high-resolution phase shifters (PSs) to realize the analog beamformers. However, low-resolution PSs are typically adopted in practice to reduce the hardware cost and power consumption. Motivated by this fact, in this paper, we investigate the practical design of hybrid precoders and combiners with low-resolution PSs in mmWave MIMO systems. In particular, we propose an iterative algorithm which successively designs the low-resolution analog precoder and combiner pair for each data stream, aiming at conditionally maximizing the spectral efficiency. Then, the digital precoder and combiner are computed based on the obtained effective baseband channel to further enhance the spectral efficiency. In an effort to achieve an even more hardware-efficient large antenna array, we also investigate the design of hybrid beamformers with one-bit resolution (binary) PSs, and present a novel binary analog precoder and combiner optimization algorithm with quadratic complexity in the number of antennas. The proposed low-resolution hybrid beamforming design is further extended to multiuser MIMO communication systems. Simulation results demonstrate the performance advantages of the proposed algorithms compared to existing lowresolution hybrid beamforming designs, particularly for the onebit resolution PS scenario. Index Terms—Millimeter wave (mmWave) communications, hybrid precoder, multiple-input multiple-output (MIMO), phase shifters, one-bit quantization. I. I NTRODUCTION The past decade has witnessed the exponential growth of data traffic along with the rapid proliferation of wireless deZihuan Wang and Ming Li are with the School of Information and Communication Engineering, Dalian University of Technology, Dalian, Liaoning 116024, China, (e-mail: wangzihuan@mail.dlut.edu.cn, mli@dlut.edu.cn). Qian Liu is with the School of Computer Science and Technology, Dalian University of Technology, Dalian, Liaoning 116024, China (e-mail: qianliu@dlut.edu.cn). A. L. Swindlehurst is with the Center for Pervasive Communications and Computing, University of California, Irvine, CA 92697 USA and also with the Institute for Advanced Study, Technical University of Munich,Munchen 80333, Germany (e-mail: swindle@uci.edu). This paper is supported by the National Natural Science Foundation of China (Grant No. 61671101 and 61601080) and the Fundamental Research Funds for the Central Universities (Grant No. DUT15RC(3)121 and DUT17JC10). vices. This flood of mobile traffic has significantly exacerbated spectrum congestion in current frequency bands, and therefore stimulated intensive interest in exploiting new spectrum bands for wireless communications. Millimeter wave (mmWave) wireless communications, operating in the frequency bands from 30-300 GHz, have been demonstrated as a promising candidate to fundamentally solve the spectrum congestion problem [1]-[3]. However, challenges always come along with opportunities. MmWave communications still need to overcome several technical difficulties before real-world deployment. As a negative result of the ten-fold increase of the carrier frequency, the propagation loss in mmWave bands is much higher than that of conventional frequency bands (e.g. 2.4 GHz) due to atmospheric absorption, rain attenuation, and low penetration [4]. From a positive perspective, the smaller wavelength of mmWave signals allows a large antenna array to be packed in a small physical dimension [5]. With the aid of pre/post-coding techniques in massive multiple-input multiple-output (MIMO) systems, the large antenna array can provide sufficient beamforming gain to overcome the severe propagation loss of mmWave channels. It also enables simultaneous transmission of multiple data streams resulting in significant improvements to spectral efficiency. For MIMO systems operating in conventional cellular frequency bands, the full-digital precoder and combiner are completely realized in the digital domain by adjusting both the magnitude and phase of the baseband signals. However, these conventional full-digital schemes require a large number of expensive and energy-intensive radio frequency (RF) chains, analog-to-digital converters (ADCs), and digital-toanalog converters (DACs). Since mmWave communication systems operate at much higher carrier frequencies and wider bandwidths, the enormous cost and power consumption of the required RF chains and ADCs/DACs make the adoption of full-digital precoding and combining schemes impractical for mmWave systems. Recently, economical and energy-efficient analog/digital hybrid precoders and combiners have been advocated as a promising approach to tackle this issue. The hybrid precoding approaches adopt a large number of phase shifters (PSs) to implement high-dimensional analog precoders to compensate for the severe path-loss at mmWave bands, and a small number of RF chains and DACs to realize lowdimensional digital precoders to provide the necessary flexibility to perform advanced multiplexing/multiuser techniques. The investigation of hybrid precoder and combiner design has attracted extensive attention in recent years because of its potential energy efficiency for mmWave MIMO communications. The major challenges in designing hybrid precoders are the practical constraints associated with the analog components, such as the requirement that the analog precoding be implemented with constant modulus PSs. Thus, hybrid precoder design typically requires the solution of various matrix factorization problems with constant modulus constraints. In particular, a popular solution to maximize the spectral efficiency of point-to-point transmission is to minimize the Euclidean distance between the hybrid precoder and the fulldigital precoder [6]-[10]. Hybrid precoder design for partiallyconnected architectures are also studied in [11]-[13]. Due to the special characteristics of mmWave channels, codebookbased hybrid precoder designs are commonly proposed [14][18], in which the columns of the analog precoder are selected from certain candidate vectors, such as array response vectors of the channel and discrete Fourier transform (DFT) beamformers. Extensions of the hybrid beamformer design to multiuser mmWave MIMO systems have also been investigated in [19]-[27]. The aforementioned existing hybrid precoder and combiner designs generally assume that infinite or high-resolution PSs are used for implementing the analog beamformers in order to achieve satisfactory performance close to the full-digital scheme. However, implementing infinite/high-resolution PSs at mmWave frequencies would significantly increase the energy consumption and complexity of the required hardware circuits [28], [29]. Obviously, it is impractical to employ infinite/high-resolution PSs for mmWave systems and realworld analog beamformers will be implemented with lowresolution PSs. Consequently, an important research direction is the exploration of signal processing techniques for hybrid analog/digital architectures that can mitigate the loss of beamforming accuracy due to the low-resolution PSs. A straightforward approach to obtain the finite-resolution beamformer is to design the infinite-resolution analog beamformer first, and then directly quantize each phase term to a finite set [30]. However, this solution becomes inefficient when the PSs have very low resolution. An alternative solution for hybrid beamforming with finite-resolution PSs is codebook-based design [14]-[18]. However, for low-resolution PSs, the size of the codebook is very small and the resulting performance is not satisfactory. In [31], [32], Sohrabi and Yu proposed to iteratively design the low-resolution hybrid precoder to maximize the spectral efficiency. However, the performance of this algorithm often suffers when one-bit quantized PSs are applied. In this paper, we first consider the problem of designing hybrid precoders and combiners with low-resolution PSs for a point-to-point mmWave MIMO system. The objective of the proposed algorithm is to minimize the performance loss caused by the low-resolution PSs while maintaining a low computational complexity. To achieve this goal, we propose to successively design the low-resolution analog precoder and combiner pair for each data stream, aiming at conditionally maximizing the spectral efficiency. An iterative phase match- ing algorithm is introduced to implement the low-resolution analog precoder and combiner pair. Then, the digital precoder and combiner are computed based on the obtained effective baseband channel to further enhance the spectral efficiency. Note that the power consumption and cost of the PS are proportional to its resolution. For example, a 4-bit (i.e. 22.5◦) resolution PS at mmWave frequencies requires 45-106 mW, while a 3-bit (i.e. 45◦ ) resolution PS needs only 15 mW [29]. In an effort to achieve maximum hardware efficiency, we also investigate the design of hybrid beamformers with one-bit resolution (binary) PSs. Inspired by the findings in [33], we present a binary analog precoder and combiner optimization algorithm under a rank-1 approximation of the interference-included equivalent channel. This algorithm has quadratic complexity in the number of antennas and can achieve almost the same performance as the optimal exhaustive search method. Finally, our investigation of low-resolution hybrid precoders and combiners is extended to multiuser mmWave MIMO systems. Numerical results in the simulation section demonstrate that the proposed algorithms can offer a performance improvement compared with existing lowresolution hybrid beamforming schemes, especially for the one-bit resolution PS scenario. Notation: The following notation is used throughout this paper. Boldface lower-case and upper-case letters indicate column vectors and matrices, respectively. (·)T and (·)H denote the transpose and transpose-conjugate operations, respectively. E{·} represents statistical expectation. Re{·} extracts the real part of a complex number; sign(·) denotes the sign operator; angle{·} represents the phase of a complex number. IL indicates an L × L identity matrix. C denotes the set of complex numbers. \|A\| denotes the determinant of matrix A. \|A\| denotes the cardinality of set A. \|a\| and kak are the magnitude and norm of a scalar a and vector a, respectively. kAkF denotes the Frobenius norm of matrix A. Finally, we adopt a Matlab-like matrix indexing notation: A(:, i) denotes the i-th column of matrix A; A(i, j) denotes the element of the i-th row and the j-th column of matrix A; a(i) denotes the i-th element of vector a. II. S YSTEM M ODEL AND P ROBLEM F ORMULATION A. Point-to-Point mmWave MIMO System Model We first consider a point-to-point mmWave MIMO system using a hybrid precoder and combiner with low-resolution PSs, as illustrated in Fig. 1. The transmitter employs Nt antennas and NtRF RF chains to simultaneously transmit Ns data streams to the receiver which is equipped with Nr antennas and NrRF RF chains. To ensure the efficiency of the mmWave communication with a limited number of RF chains, the number of data streams and the number of RF chains are constrained as Ns = NtRF = NrRF . The transmitted symbols are first processed by a baseband RF digital precoder FBB ∈ CNt ×Ns , then up-converted to the RF RF domain via Nt RF chains before being precoded with an analog precoder FRF of dimension Nt × NtRF . While the baseband digital precoder FBB enables both amplitude and phase modifications, the elements of the analog precoder Fig. 1. The point-to-point mmWave MIMO system using hybrid precoder and combiner. FRF , which are implemented by the PSs, have a constant amplitude √1N and quantized phases: FRF (i, j) = √1N ejϑi,j , t t in which the phase ϑi,j is quantized as ϑi,j ∈ B , { 2πb 2B \| b = 1, 2, . . . , 2B }, and B is the number of bits to control the phase. We denote the constraint set of the analog precoder as follows: 2πb FRF (i, j) ∈ F , { √1N ej 2B \| b = 1, 2, . . . , 2B }. Obviously, t a larger number of bits B leads to finer resolution for the PSs and potentially better performance, but also results in higher hardware complexity and power consumption. The discrete-time transmitted signal can be written in the following form √ (1) x = P FRF FBB s where s is the Ns × 1 symbol vector, E{ssH } = N1s INs , P represents transmit power and this power constraint is enforced by normalizing FBB such that kFRF FBB k2F = Ns . We consider a narrow-band slow-fading propagation channel, which yields the following received signal y = = Hx + n √ P HFRF FBB s + n R = log2 INs + P −1 H H R WBB WRF HFRF FBB × Ns n ! H H FH BB FRF H WRF WBB where WRF is the Nr × NrRF analog combiner whose elements have the same constraint as FRF , i.e. WRF (i, j) = √1 ejϕi,j , ϕi,j ∈ B and thus WRF (i, j) ∈ W , N , (4) H H WRF WRF WBB is the noise covariwhere Rn , σn2 WBB ance matrix after combining. We aim to jointly design the digital beamformers FBB , WBB as well as the low-resolution analog beamformers FRF , WRF to maximize the spectral efficiency: o n ⋆ ⋆ = arg max R , WBB F⋆RF , F⋆BB , WRF s. t. (2) where y is the Nr × 1 received signal vector, H is the Nr × Nt channel matrix, and n ∼ CN (0, σ 2 INr ) is the complex Gaussian noise vector corrupting the received signal. The receiver employs an analog combiner implemented by the PSs and a digital combiner using NrRF RF chains to process the received signal. The signal after the spatial processing has the form √ H H H H b WRF HFRF FBB s + WBB WRF n (3) s = P WBB r hardware constraint, we aim to jointly design the hybrid precoder and combiner for a mmWave MIMO system. When Gaussian symbols are transmitted over the mmWave MIMO channel, the achievable spectral efficiency is given by FRF (i, j) ∈ F , ∀i, j, WRF (i, j) ∈ W, ∀i, j, kFRF FBB k2F (5) = Ns , kWRF WBB k2F = Ns . Obviously, the optimization problem (5) is a non-convex NPhard problem. In the next section, we attempt to decompose the original problem into a series of sub-problems and seek a sub-optimal solution with low-complexity and satisfactory performance. III. L OW-R ESOLUTION H YBRID P RECODER C OMBINER D ESIGN AND 2πb { √1N ej 2B \| b = 1, 2, . . . , 2B }, WBB is the NrRF ×Ns digital r baseband combiner and the combiner matrices are normalized such that kWRF WBB k2F = Ns . B. Problem Formulation We consider the practical and hardware-efficient scenario in which the PSs have very low-resolution (e.g. B = 1, 2) to reduce the power consumption and complexity. Under this To simplify the joint hybrid precoder and combiner design, the objective problem is decomposed into two separate optimizations. We first focus on the joint design of the analog precoder FRF and combiner WRF . Then, having the effective baseband channel associated with the obtained optimal analog precoder and combiner, the digital precoder FBB and combiner WBB are computed to further maximize the spectral efficiency. A. Low-Resolution Analog Precoder and Combiner Design We observe that under the assumption of high signal-tonoise-ratio (SNR), the achievable spectral efficiency in (4) can be approximated as R ≈ log2 P −1 H H R WBB WRF HFRF FBB × Ns n ! H H FH BB FRF H WRF WBB . (6) While the per-antenna SNR in mmWave systems is typically low, the post-combining SNR should be high enough to justify this approximation. In addition, it has been verified in [32] that for large-scale MIMO systems, the optimal analog beamformers are approximately orthogonal, i.e. FH RF FRF ∝ INtRF . RF 2 = ≈ ζ I This enables us to assume FBB FH N s when Nt BB 2 Ns , where ζ is a normalization factor. Similarly, we have H H H WRF WRF WBB ≈ INs . WBB WBB ≈ ξ 2 INs and WBB 2 2 2 Let γ , ζ ξ , then (6) can be further simplified as ! P γ2 H R ≈ log2 (7) WH HFRF FH RF H WRF Ns σ 2 RF ! P γ2 (a) H (8) + 2 × log2 WRF HFRF = Ns log2 Ns σ 2 where (a) follows since \|XY\| = \|X\|\|Y\| when X and Y are both square matrices. Therefore, the analog precoder and combiner design with low-resolution PSs can be approximately reformulated as: ⋆ ⋆ H {FRF , WRF } = arg max log2 WRF HFRF s. t. FRF (i, j) ∈ F , ∀i, j, WRF (i, j) ∈ W, ∀i, j. (9) Unfortunately, the optimization problem (9) is still NP-hard RF RF and has exponential complexity O(\|F \|Nt Nt \|W\|Nr Nr ). Therefore, we propose to further decompose this difficult optimization problem into a series of sub-problems, in which each transmit/receive RF chain pair is considered one by one and the analog precoder and combiner for each pair are successively designed. In particular, we define the singular value decomposition (SVD) of H as H = UΣVH (10) where U is an Nr ×Nr unitary matrix, V is an Nt ×Nt unitary matrix, and Σ is a rectangular diagonal matrix of singular values. Due to the sparse nature of the mmWave channel, the matrix H is typically low rank. In particular, the effective rank of the channel serves as an upper bound for the number of data streams Ns that the channel can support. Thus, we assume that the channel H can be well approximated by retaining only the bΣ bV b H , where U b , U(:, 1 : Ns strongest components H ≈ U b b Ns ), Σ , Σ(1 : Ns , 1 : Ns ), and V , V(:, 1 : Ns ). Then, the objective in (9) can be converted to H b b bH H ≈ log2 WRF UΣV FRF .(11) log2 WRF HFRF Next, we write the analog precoding and combining matrices as FRF , [fRF,1 . . . fRF,Ns ] and WRF , [wRF,1 . . . wRF,Ns ], respectively, where fRF,l and wRF,l , l = 1, . . . , Ns , are the analog precoder and combiner pair for the l-th data stream. Furthermore, we denote FRF,\l as the precoding matrix excluding the l-th precoder vector fRF,l and WRF,\l as the combining matrix excluding the lth combiner vector wRF,l . Then, the formulation (11) can be further transformed to (12)-(16), which are presented at the top of following page, where α is a very small scalar to assure invertibility. Thus, the objective in (9) can be reformulated as: log2 H WRF HFRF ≈ log2 +log2 H HFRF,\l WRF,\l H wRF,l Ql fRF,l ! ! (17) where we define the interference-included channel matrix Ql as H b −1 Σ bV b H. b bH b Ql , U(αI Ns + ΣV FRF,\l WRF,\l U) (18) According to (17), if FRF,\l and WRF,\l are known, the problem (9) can be reformulated as finding a corresponding precoder fRF,l and combiner wRF,l pair to conditionally maximize the achievable spectral efficiency: ⋆ ⋆ H fRF ,l , wRF ,l = arg max wRF,l Ql fRF,l s. t. fRF ,l (i) ∈ F , i = 1, . . . , Nt , (19) wRF ,l (j) ∈ W, j = 1, . . . , Nr . This motivates us to propose an iterative algorithm, which starts with appropriate initial RF precoding and combining matrices then successively designs fRF ,l and wRF ,l according to (19) with an updated Ql as in (18) until the algorithm converges. The complexity of obtaining an optimal solution to (19) for each iteration is now reduced to O(\|F \|Nt \|W\|Nr ), which is still too high. To practically solve the problem (19), in what follows we present an iterative phase matching algorithm, which searches the conditionally optimal phase of each element of the analog precoder fRF ,l and combiner wRF ,l . Specifically, we first design the analog precoder fRF ,l assuming the analog combiner wRF,l is fixed. Let ϑl,i be the phase of the i-th element of the analog precoder fRF ,l and let ϕl,j be the phase of the j-th element of the analog combiner wRF,l . If we temporarily remove the discrete phase constraint, the optimal continuous phase ϑ̃l,i of the i-th element of the analog precoder fRF,l is given by the following proposition, whose proof is provided in Appendix A. Proposition 1: Given the phases ϕl,j of the analog combiner wRF,l and the phases ϑl,u , u 6= i, of the analog precoder fRF,l , the optimal continuous phase ϑ̃l,i of the i-th element of analog log2 H b b bH WRF UΣV FRF = log 2 ≈ log 2 = log 2 = log 2 precoder fRF,l is ϑ̃l,i bV b H FRF,\l Σ H b bV b H FRF WRF Σ U ! H b fRF,l WRF,\l wRF,l U = log2 = log2 H bV b H FRF,\lWRF,\l b Σ U H bV b H FRF,\l WRF,\l b Σ U H bΣ bV b H FRF,\l WRF,\l U (12) H H bV b H FRF,\l WRF,\l b +Σ bV b H fRF,l wRF,l b Σ U U h −1 i H H bV b H FRF,\l WRF,\l b bV b H fRF,l wRF,l b INs + αINs + Σ U Σ U ! ! + log 2 + log 2 ! h −1 i H H bV b H FRF,\lWRF,\l b bV b H fRF,l wRF,l b INs + αINs + Σ U Σ U h −1 i H H b αINs + Σ bV b H fRF,l bV b H FRF,\l WRF,\l b 1 + wRF,l U Σ U  Nt  X ejϕl,j ejϑl,u Ql (j, u) = angle   j=1 u6=i   Nr  X ejϕl,j Ql (j, i) . (20) −angle    Nr X j=1 Then, after finding the optimal continuous phase ϑ̃l,i by (20), we reconsider the discrete phase constraint and find the optimal low-resolution phase ϑl,i by quantization: ϑl,i = arg min ϑ̃l,i − ϑ̂l,i . (21) ϑ̂l,i ∈B Similarly, if the analog precoder fRF,l is determined, the optimal continuous phase ϕ̃l,j of the j-th element of wRF,l is   Nt Nt  X X ejϑl,i ejϕl,u Ql (u, i) ϕ̃l,j = angle   i=1 u6=j ) (N t X jϕl,i e Ql (j, i) , (22) −angle ! ! (13) (14) ! (15) (16) Algorithm 1 Iterative Phase Matching Algorithm for LowResolution Analog Precoder and Combiner Design Input: F , W, H. ⋆ Output: F⋆RF and WRF . ⋆ ⋆ 1: Initialize FRF = 0, WRF = 0. 2: for l = 1 : Ns do ⋆ 3: Obtain FRF,\l from F⋆RF and WRF,\l from WRF . −1 b b H H H b b b b 4: Update Ql =U(αINs+ΣV FRF,\l WRF,\l U) ΣV . 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: while no convergence of ϑl,i and ϕl,j do for i = 1 : Nt do Obtain quantized phase ϑl,i by (20) and (21). end for for j = 1 : Nr do Obtain quantized phase ϕl,j by (22) and (23). end for end while ⋆ ⋆ Construct fRF,l by ϑl,i and wRF,l by ϕl,j . end for ⋆ ⋆ ⋆ and WRF by wRF,l . Construct F⋆RF by fRF,l ⋆ ⋆ Goto Step 2 until convergence of FRF and WRF . i=1 and the optimal low-resolution phase ϕl,j is obtained by ϕl,j = arg min ϕ̃l,j − ϕ̂l,j . ϕ̂l,j ∈B (23) Motivated by (20)-(23), the iterative procedure to design the precoder fRF ,l and combiner wRF ,l as in (19) is straightforward. With appropriate initial ϑl,i , ϕl,j , we design the precoder fRF ,l by finding the conditionally optimal phases ϑl,i as in (20) and (21). Then, with the obtained ϑl,i , i = 1, . . . , Nt , we design the combiner wRF ,l by finding the conditionally optimal phases ϕl,j as in (22) and (23). We alternate the designs of fRF ,l and wRF ,l iteratively until the obtained phase of each element of fRF ,l and wRF ,l does not change and the convergence is achieved. Note that since in each precoder and combiner design step, the objective function of (19) is monotonically non-decreasing, and thus our proposed algorithm is guaranteed to converge to at least a locally optimal solution. We summarize the proposed joint low-resolution analog precoder and combiner design in Algorithm 1. B. Digital Precoder and Combiner Design After all analog precoder-combiner pairs have been detere as mined, we can obtain the effective baseband channel H e , (W⋆ )H HF⋆ , H RF RF (24) e =U eΣ eV eH H (25) e F⋆BB = V, ⋆ e W = U. (26) ⋆ ⋆ ⋆ where F⋆RF , [fRF,1 , . . . , fRF,N ] and WRF , s ⋆ ⋆ [wRF,1 , . . . , wRF,Ns ]. For the baseband precoder and combiner design, we define the SVD of the effective e as baseband channel H e and V e are Ns × Ns unitary matrices, Σ e is an where U Ns × Ns diagonal matrix of singular values. Then, to further enhance the spectral efficiency, an SVD-based baseband digital precoder and combiner are employed: BB (27) Finally, the baseband precoder and combiner are normalized F⋆BB = ⋆ WBB = √ N s F⋆ BB ⋆ kF⋆ RF FBB kF √ ⋆ Ns WBB ⋆ W⋆ kWRF BB kF , (28) . Now, the joint optimization problem (33) can be decoupled into individually designing the analog precoder fRF ,l and combiner wRF ,l : (29) ⋆ fRF ,l = arg max fRF ,l ∈ √1 Nt IV. O NE -B IT R ESOLUTION A NALOG P RECODER C OMBINER D ESIGN AND ⋆ wRF ,l = arg {±1}Nt max 1 wRF ,l ∈ √N {±1}Nr H fRF,l gl,1 , H wRF,l pl,1 . (34) (35) r In the previous section, we proposed a novel hybrid beamformer design for maximizing the spectral efficiency of a mmWave MIMO system, in which the analog precoder and combiner are implemented with low-resolution PSs. In order to achieve maximum hardware efficiency, in this section we focus on the design of analog precoders and combiners using “one-bit” resolution (binary) PSs, which can maximally reduce the power consumption and simplify the hardware complexity. Although the iterative phase matching algorithm proposed in the previous section can also be applied, a simpler approach is possible in the one-bit case. Therefore, in this section, we present an efficient one-bit resolution analog beamformer design, which can achieve good performance with much lower complexity. We follow the procedure of the hybrid beamforming design proposed in the previous section, but only modify the optimization problem (19), which attempts to determine the l-th analog precoder and combiner pair. Particularly, we reformulate this analog beamformer design problem (19) with the constraint of one-bit resolution PSs as ⋆ H ⋆ wRF,l Ql fRF,l . max fRF ,l , wRF ,l = arg fRF ,l ∈ √1 {±1}Nt Nt 1 wRF ,l ∈ √N {±1}Nr r (30) The optimization problem (30) can be solved through exhaustive search with exponential complexity O(2Nt Nr ), which would not be possible with large antenna arrays. Therefore, in the following we attempt to develop an efficient one-bit resolution beamformer design with polynomial complexity in the number of antennas. We first define the SVD of Ql as Ql = Ns X H λl,i pl,i gl,i , (31) i=1 where pl,i and gl,i are the i-th left and right singular vectors of Ql , respectively, and λl,i is the i-th largest singular value, λl,1 ≥ λl,2 ≥ . . . ≥ λl,Ns . Then, the objective in (30) can be rewritten as H Ql fRF,l \| = \|wRF,l Ns X H H λl,i wRF,l pl,i gl,i fRF,l . (32) i=1 If we utilize a rank-1 approximation by keeping only the H strongest term, i.e. Ql ≈ λl,1 pl,1 gl,1 , the optimization function in (30) can be approximated by ⋆ ⋆ H H fRF ,l , wRF max wRF,l pl,1 gl,1 fRF,l . ,l = arg fRF ,l ∈ √1 Nt 1 wRF ,l ∈ √N r {±1}Nt {±1}Nr (33) These two optimization problems (34) and (35) require only the singular vectors pl,1 and gl,1 associated with the largest singular value, which can be quickly obtained by the power iteration algorithm [34] instead of the complete SVD calculation. However, solving (34) and (35) by exhaustive search still has exponential complexity in the number of antennas. In order to further reduce the complexity without a significant loss of performance, we propose to construct a smaller dimension candidate beamformer set, from which the optimal beamformer can be found with linear complexity. In the following, we present this algorithm for the precoder design (34) as an example, while the combiner design (35) follows the same procedure. We introduce an auxiliary variable φ ∈ [−π, π) and we reformulate the optimization problem (34) as: o n o n H ⋆ (36) Re fRF,l gl,1 e−jφ = arg max φ⋆ , fRF,l φ∈[−π,π) fRF ,l ∈ √1 {±1}Nt Nt = arg max Nt P φ∈[−π,π) i=1 fRF ,l ∈ √1 {±1}Nt Nt fRF,l (i)\|gl,1 (i)\| cos(φ − ψi ) (37) where ψi denotes the phase of gl,1 (i). Obviously, given any φ ∈ [−π, π), the corresponding binary precoder that maximizes (37) is 1 fRF,l (i) = √ sign (cos (φ − ψi )) , i = 1, . . . , Nt . Nt (38) With the conditionally optimal fRF,l for any given φ shown in (38), we will now show that we can always construct a set of Nt candidate binary precoders Fl , {fl,1 , . . . , fl,Nt } and ⋆ guarantee fRF ,l ∈ Fl . Then, the maximization in (34) can be carried out over a set of only Nt candidates without loss of performance. We first define the angles ψbi , i = 1, . . . , Nt , as  π 3π   ψi − π, if ψi ∈ , , 2 2 ψbi , (39) h  π π ψ , , if ψi ∈ − , i 2 2 so that ψbi ∈ [− π2 , π2 ). Then, we map the angles ψbi to ψei , i = 1, . . . , Nt , which are rearranged in ascending order, i.e. ψe1 ≤ ψe2 ≤ . . . ≤ ψeNt . Because of the periodicity of the cosine function, the maximization problem (37) with respect to φ can be carried out over any interval of length π. If we construct Nt non-overlapping sub-intervals [ψe1 − π2 , ψe2 − π2 ), [ψe2 − π2 , ψe3 − π e π π ⋆ e 2 ), . . . , [ψNt − 2 , ψ1 + 2 ), then the optimal φ must be located in one of Nt sub-intervals since the full interval is [ψe1 − π e 2 , ψ1 + π2 ) of length π. Therefore, the optimization problem (37) can be solved by examining each sub-interval separately. Assuming the optimal φ⋆ is in the k-th sub-interval, the corresponding optimal binary can be obtained by (38) precoder as e fl,k (i) = √1N sign cos φ⋆ − ψei , i = 1, . . . , Nt , and t has the form 1 e . . 1} −1 . . . − 1]T . (40) fl,k = √ [1\| .{z \| {z } Nt k Nt −k After that, given the inverse sorting that maps ψei to ψbi , we rearrange the corresponding elements of e fl,k and obtain b fl,k . Then, based on the relationship between ψi and ψbi defined in (39), we can achieve the conditionally optimal precoder fl,k by  π 3π  b  − fl,k (i), if ψi ∈ , i = 1, . . . , Nt , , 2 2 fl,k (i) , h π π   b fl,k (i), if ψi ∈ − , , i = 1, . . . , Nt , 2 2 (41) for the case that φ⋆ is in the k-th sub-interval. Since the optimal φ⋆ must be located in one of Nt subintervals, we can obtain Nt conditionally optimal precoders by examining all Nt sub-intervals and construct a candidate precoder set Fl as Fl , {fl,1 , . . . , fl,Nt }, (42) ⋆ which must contain the optimal precoder fRF,l . Therefore, without loss of performance, the problem in (34) can be transformed to an equivalent maximization task over only the set Fl H ⋆ fRF,l gl,1 , (43) fRF ,l = arg max fRF ,l ∈Fl which has linear complexity O(Nt ). Similarly, we can also construct a candidate analog combiner set Wl and obtain ⋆ wRF ,l by the same procedure. The rank-1 solution returned by (43) is based on the rank-1 approximation of the interference-included equivalent channel Ql . The approximation of Ql may cause a performance degradation when we revisit the original problem (30). Therefore, in order to enhance the performance, we propose to jointly select the precoder and combiner over candidate sets Fl and Wl as n o ⋆ ⋆ H fRF , w (44) wRF,l Ql fRF,l ,l RF ,l = arg max fRF ,l ∈Fl wRF ,l ∈Wl which may return the rank-1 or a better solution with quadratic complexity O(Nt Nr ). This low-complexity analog beamformer design with one-bit resolution PSs is summarized in Algorithm 2. V. H YBRID P RECODER AND C OMBINER D ESIGN M ULTIUSER MM WAVE MIMO S YSTEMS FOR In this section, we consider a mmWave multiuser MIMO uplink system and extend the low-resolution hybrid precoder and combiner designs proposed in the previous sections to the multiuser system. Algorithm 2 One-Bit Resolution Analog Beamformer Design Input: Ql . ⋆ ⋆ Output: fRF ,l and wRF ,l . 1: Calculate pl,1 and gl,1 by an SVD of Ql . bi , i = 1, . . . , Nt , by (39). 2: Define the angles ψ b e 3: Map ψi to ψi , i = 1, . . . , Nt , in an ascending order. 4: for k = 1 : Nt do 5: Obtain e fl,k by (40). 6: Obtain b fl,k from e fl,k based on inverse mapping from ψei b to ψi , i = 1, . . . , Nt . 7: Obtain fl,k from b fl,k by (41). 8: end for 9: Construct Fl = {fl,1 , . . . , fl,Nt }. 10: Construct Wl by a similar procedure as Steps 2-9. ⋆ ⋆ 11: Find the optimal fRF ,l and wRF ,l by (44). A. System Model and Problem Formulation We consider a multiuser mmWave MIMO uplink system as presented in Fig. 2, where a base-station (BS) is equipped with Nr antennas and NRF RF chains and simultaneously serves K mobile users. Due to power consumption and hardware limitations, each mobile user has Nt antennas and a single RF chain to transmit only one data stream to the BS. We further assume the number of RF chains at the BS is equal to the number of users, i.e. NRF = K. Let fRF,k be the analog precoder of the k-th user, where each element of fRF,k has a constant magnitude √1N and lowt resolution discrete phases, i.e. fRF,k (i) ∈ F , ∀i = 1, . . . , Nt . The transmitted signal of the k-th user after precoding can be formulated as p xk = Pk fRF,k sk (45) 2 where sk is the symbol of the k-th user, E{\|sk \| } = 1, and Pk is the k-th user’s transmit power. Let Hk ∈ CNr ×Nt , k = 1, . . . , K, denote the uplink channel from the k-th user to the BS. The received signal at the BS can be written as r= K p X Pk Hk fRF,k sk + n (46) k=1 where n ∼ CN (0, σ 2 INr ) is complex Gaussian noise. The BS first applies an Nr × K analog combining matrix WRF , [wRF ,1 . . . wRF ,K ] to process the received signal, in which the analog combiner wRF,k corresponding to the k-th user is also implemented by low-resolution PSs, i.e. wRF ,k (j) ∈ W, j = 1, . . . , Nr . Then, a baseband digital combiner wBB,k ∈ CK×1 is employed to retrieve the information of the k-th user. Let wk , WRF wBB,k denote the hybrid combiner corresponding to the k-th user. After the combining process at the BS, the estimated symbol of the k-th user can be expressed as ŝk = K p X p Pk wkH Hk fRF,k sk + wkH Pi Hi fRF,i si + wkH n. i=1 i6=k (47) Fig. 2. The multiuser mmWave MIMO system using hybrid precoder and combiner. Given the received signal at the BS in (47), the signal-tointerference-plus-noise ratio (SINR) of the k-th user can be expressed as √ \| Pk wkH Hk fRF,k \|2 γk = K √ P (48) \| Pi wkH Hi fRF,i \|2 +σ 2 k wk k2 i=1,i6=k and the achievable sum-rate of the multiuser uplink system is Ru = K X log(1 + γk ). (49) k=1 We aim to jointly design the analog precoders and combiners implemented by low-resolution PSs as well as the digital combiners to maximize the sum-rate of the uplink multiuser system: n ⋆ ⋆ ⋆ wRF ,k , wBB,k , fRF,k s. t. K k=1 o = arg max fRF,k (i) ∈ F , ∀k, i, K X log (1 + γk ) k=1 wRF ,k (j) ∈ W, ∀k, j. (50) B. Low-Resolution Hybrid Precoder and Combiner Design Obviously, the optimization problem (50) cannot be directly solved. Thus, we adopt an approach similar to [26] and propose to successively design the low-resolution analog beamformer pair for each user, aiming at enhancing the channel gain as well as suppressing the inter-user interference. Then, the baseband combiner at the BS is calculated to further mitigate the interference and maximize the sum-rate. In particular, for the first user, the analog precoder and combiner pair is designed to maximize the corresponding channel gain, which can be formulated as follows: ⋆ H ⋆ wRF ,1 , fRF ,1 = arg max wRF ,1 H1 fRF ,1 s. t. fRF ,1 (i) ∈ F , i = 1, . . . , Nt , wRF ,1 (j) ∈ W, j = 1, . . . , Nr . (51) This analog precoder and combiner design problem can be efficiently solved by the algorithm presented in Sec. III-A when low-resolution PSs are utilized, or the algorithm proposed in Sec. IV if only one-bit resolution PSs are available. Then, the analog precoders fRF,k and combiners wRF ,k , k = 2, 3, . . . , K, for the remaining K − 1 users are successively designed by an iterative procedure. In each iteration, we attempt to find the analog beamformer pair that suppresses the interference from the users whose analog beamformers have already been determined. To achieve this goal, the channel of the user whose combiner is to be calculated is projected onto the space orthogonal to the collection of previously designed analog combiners. This approach leads to orthogonal analog combiners that suppress the inter-user interference. Specifically, to design the k-th user’s analog beamformer pair, we first extract the orthonormal components di of the pre⋆ viously determined analog combiners wRF ,i , i = 1, . . . , k − 1 by the Gram-Schmidt procedure: ⋆ qi = wRF ,i − i−1 X ⋆ dH j wRF ,i dj , (52) j=1 di = qi /kqi k. (53) ⋆ ⋆ Note that d1 = wRF ,1 and wRF ,1 is the analog combiner calculated for the first user. Then, the combiner components are removed from the k-th user’s channel to obtain the b k as modified channel H bk = H INr − k−1 X i=1 di dH i ! Hk . (54) b k , the analog beamFinally, based on the modified channel H former pair for the k-th user is found by solving the following optimization using the algorithms proposed in the previous s. t. fRF,k (i) ∈ F , i = 1, . . . , Nt , 120 (55) wRF ,k (j) ∈ W, j = 1, . . . , Nr . After finding the analog beamformers for all users, the effective baseband channel for each user can be obtained √ H ⋆ ⋆ as hek , Pk (WRF ) Hk fRF ,k . Then, a minimum mean square error (MMSE) baseband digital combiner for the k-th user is employed to further suppress the interference: h i−1 ⋆ ⋆ ⋆ wBB,k = He (He )H + σ 2 (WRF )H WRF hek , (56) 100 where He , [he1 , . . . , heK ]. The proposed low-resolution hybrid precoder and combiner design for multiuser mmWave systems is summarized in Algorithm 3. wRF ,1 (i)∈W fRF,1 (j)∈F 3: 4: ⋆ d1 = wRF ,1 . for k = 2: K do k−1 P H b di di H k . Hk = INr − i=1 5: ⋆ ⋆ Obtain wRF ,k and fRF ,k for user-k by solving ⋆ ⋆ H b max \|wRF wRF ,k , fRF ,k = arg ,k Hk fRF,k \|. wRF ,k (i)∈W fRF,k (j)∈F 6: 7: 8: 9: ⋆ qk = wRF ,k − k−1 P i=1 ⋆ dH i wRF ,k di ; dk = qk /kqk k. end for ⋆ Obtain digital combiners wBB,k , k = 1, . . . , K, by h i−1 H ⋆ ⋆ ⋆ ) WRF wBB,k = He (He )H + σ 2 (WRF hek . VI. S IMULATION R ESULTS In this section, we provide simulation results for the proposed joint hybrid precoder and combiner designs with lowresolution PSs for point-to-point mmWave systems as well as multiuser mmWave systems. MmWave channels are expected to be sparse and have a limited number of propagation paths. In the simulations, we adopt a geometric channel model with L paths [32]. In particular, the discrete-time narrow-band mmWave channel H is formulated as r L Nt Nr X H= αi ar (θir )at (θit )H (57) L i=1 90 80 70 60 102 50 100 40 98 30 96 34 20 0 5 10 15 20 34.5 25 35 30 35.5 35 40 SNR (dB) Fig. 3. Spectral efficiency versus SNR (Nt = 64, Nr = 64, NtRF = NrRF = 6, Ns = 6). 65 60 Spectral Efficiency (b/s/Hz) Algorithm 3 Low-Resolution Hybrid Precoder and Combiner Design for Multiuser mmWave Systems Input: F , W, Hk , k = 1, . . . , K. ⋆ ⋆ ⋆ Output: fRF,k , wRF ,k , wBB,k , k = 1, . . . , K. ⋆ ⋆ 1: Obtain wRF ,1 and fRF ,1 for user-1 by solving ⋆ ⋆ H = arg max \|wRF wRF ,1 , fRF,1 ,1 H1 fRF,1 \|. 2: Full−digital PE−AltMin [6] (B=∞) CDM [30] (B=2) CDM [30] (B=1) HBF [32] (B=2) HBF [32] (B=1) Proposed (B=2) Proposed (B=1) 110 Spectral Efficiency (b/s/Hz) sections: ⋆ H ⋆ b wRF ,k , fRF ,k = arg max wRF ,k Hk fRF,k 55 50 45 Full−digital PE−AltMin [6] (B=∞) CDM [30] (B=2) CDM [30] (B=1) HBF [32] (B=2) HBF [32] (B=1) Proposed (B=2) Proposed (B=1) 40 35 30 25 20 40 60 80 100 120 140 Number of Antenna Fig. 4. Spectral efficiency versus number of antenna (NtRF = NrRF = 4, Ns = 4, SNR = 20dB). where αi ∼ CN (0, L1 ) are the independent and identically distributed complex gains of the i-th propagation path (ray) θit and θir ∈ [− π2 , π2 ] are the angles of departure (AoD) and the angles of arrival (AoA), respectively. Finally, the array response vectors at (θt ) and ar (θr ) depend on the antenna array geometry. We assume that the commonly used uniform linear arrays (ULAs) are employed, and the transmit antenna array response vector at (θt ) and the receive antenna array response vector ar (θr ) can be written as t t 2π 2π 1 at (θt ) = √ [1, ej λ d sin(θ ) , . . . , ej(Nt −1) λ d sin(θ ) ]T , Nt (58) r r 2π 1 j 2π d sin(θ ) j(N −1) d sin(θ ) T r r λ ,...,e ] , ar (θ ) = √ [1, e λ Nr (59) respectively, where λ is the signal wavelength, and d is the distance between antenna elements. In the following simulations, we consider an environment with L = 6 scatterers between the 60 20 58 18 Exhaustive search Proposed (B=1) Spectral Efficiency (b/s/Hz) Spectral Efficiency (b/s/Hz) 56 54 52 50 48 Full−digital CDM [30] (B=2) CDM [30] (B=1) HBF [32] (B=2) HBF [32] (B=1) Proposed (B=2) Proposed (B=1) 46 44 42 40 38 1 2 3 4 5 6 7 8 9 10 52 Spectral Efficiency (b/s/Hz) 50 51.4 51.2 51 44 4.95 40 38 5 5.05 Full−digital PE−AltMin [6] (B=∞) CDM [30] HBF [32] Proposed 42 1 2 3 4 12 10 8 6 2 −10 0 10 20 30 40 SNR (dB) Fig. 5. Spectral efficiency versus number of iteration (Nt = 64, Nr = 64, NtRF = NrRF = 4, Ns = 4, SNR = 20dB). 46 14 4 Number of Iterations 48 16 5 6 Resolution (bits) Fig. 6. Spectral efficiency versus resolution of PSs (Nt = 64, Nr = 64, NtRF = NrRF = 4, Ns = 4, SNR = 20dB). transmitter and the receiver. The antenna spacing is d = λ2 . A. Simulation Results of a Point-to-Point mmWave System We first consider a point-to-point mmWave communication system, in which the transmitter and receiver are both equipped with 64-antenna ULAs. The number of RF chains at the transmitter and receiver are NtRF = NrRF = 6, so the number of data streams is also assumed to be Ns = 6. Fig. 3 shows the average spectral efficiency versus SNR over 106 channel realizations. We evaluate the spectral efficiency of the algorithm proposed in Sec. III for the case of 2-bit (B = 2) resolution PSs and the algorithm proposed in Sec. IV for the case of 1-bit (B = 1) resolution PSs. For comparison purposes, we also plot the spectral efficiency of two state-of-the-art lowresolution hybrid beamformer designs: the coordinate descent method (CDM) algorithm in [30] and the hybrid beamforming (HBF) algorithm in [32]. To the best of our knowledge, the Fig. 7. Spectral efficiency versus SNR (Nt = 8, Nr = 8, Ns = 1, B = 1). algorithm in [32] achieves the best performance with lowresolution PSs in the existing literature. The performance of a fully digital approach using SVD-based beamforming and the hybrid beamforming scheme with infinite-resolution (B = ∞) PSs using the phase extraction (PE-AltMin) algorithm in [6] are also included as performance benchmarks. Fig. 3 illustrates that the proposed algorithm outperforms the competitors, particularly for the case of 1-bit resolution PSs. Moreover, it can be observed that the proposed algorithm with B = 2 achieves performance close to optimal full-digital beamforming and hybrid beamforming with infinite-resolution PSs. For additional simulation validation, Fig. 4 illustrates the spectral efficiency versus the number of antennas and similar conclusions can be drawn. In order to illustrate the convergence of the proposed algorithm, we show the spectral efficiency versus the number of iterations in Fig. 5, which also includes other algorithms for comparison. It is observed that our proposed algorithms converge faster than the other two iterative schemes, which is a highly favorable property. In Fig. 6, we show the spectral efficiency as a function of B to illustrate the impact of the resolution of PSs on the spectral efficiency. As expected, increasing the PS resolution will improve the system performance, but using only B = 3 bits is sufficient to closely approach the performance of the ideal unquantized case. Beyond B = 3, the additional cost and complexity associated with using higher-resolution PSs is not warranted given the very marginal increase in spectral efficiency. Moreover, our proposed algorithms outperform the other two low-resolution beamforming methods for all PS resolutions. To examine the impact of the approximations used in deriving the proposed one-bit resolution hybrid beamformer scheme, in Fig. 7 we compare it with the optimal exhaustive search approach. The number of antennas at both transmitter and receiver is chosen to be 8 and the number of data streams is Ns = 1. A relatively simple case is examined here due to the exponential complexity of the exhaustive search method. We see from Fig. 7 that the spectral efficiency achieved by (I-HB) in [26]. All three algorithms are codebook-based approaches and the size of the beamsteering codebook is set at 32 (i.e. B = 5 quantization bits) for fairness of the comparison. It can be observed from Fig. 8 that our proposed lowresolution hybrid beamforming design outperforms the other three algorithms using only 2-bit resolution PSs. Moreover, the performance with 1-bit resolution PSs is also comparable. Fig. 9 further shows the sum-rate versus the number of users K. From Fig. 9, we see that our proposed algorithm with 2-bit resolution PSs always outperforms the other codebook-based algorithms. Furthermore, even with 1-bit resolution PSs, the proposed algorithm can still achieve competitive performance compared with the TS-HB and MMSE-HB approaches when K > 5. 60 Sum−rate (b/s/Hz) 50 40 Full−digital TS−HB [24] MMSE−HB [25] I−HB [26] Proposed (B=2) Proposed (B=1) 30 20 10 0 −20 −15 −10 −5 0 5 10 15 20 SNR (dB) VII. C ONCLUSIONS Fig. 8. Spectral efficiency versus SNR (Nt = 64, Nr = 16, K = 4). NtRF = 4, 100 90 Sum−rate (b/s/Hz) 80 70 Full−digital TS−HB [24] MMSE−HB [25] I−HB [26] Proposed (B=2) Proposed (B=1) 60 50 40 This paper considered the problem of hybrid precoder and combiner design for mmWave MIMO systems with lowresolution quantized PSs. We proposed an efficient iterative algorithm which successively designs the low-resolution analog precoder and combiner pair for each data stream. Then, the digital precoder and combiner were computed based on the obtained effective baseband channel to further enhance the spectral efficiency. The design of low-resolution hybrid beamformers for multiuser MIMO communication systems was also investigated. Simulation results verified the effectiveness of the proposed algorithms, particularly for scenarios in which onebit resolution phase shifters are used. 30 A PPENDIX A P ROOF OF P ROPOSITION 1 20 10 1 2 3 4 5 6 7 8 Number of users The optimization problem (19) can be equivalently formulated as Fig. 9. Spectral efficiency versus K (Nt = 64, Nr = 16, NtRF = K, SNR = 20dB). the proposed algorithm is the same as that of the optimal exhaustive search method, suggesting that the proposed hybrid beamforming algorithm with one-bit resolution PSs can provide optimal or near-optimal performance. B. Simulation Results of a Multiuser mmWave System Next, we evaluate the performance of the proposed lowresolution beamformer algorithm in a multiuser uplink system. We assume there are K = 4 users, each of which is equipped with Nt = 16 antennas and only one RF chain to transmit a single data stream. The BS has Nr = 64 antennas and NrRF = 4 RF chains. Fig. 8 illustrates the sum-rate versus SNR for various hybrid beamformer designs. In particular, we include three state-of-the-art multiuser hybrid beamforming approaches for comparison: i) two-stage hybrid beamforming (TS-HB) in [24], ii) MMSE-based hybrid beamforming (MMSE-HB) in [25], and iii) iterative hybrid beamforming Nt Nr X l l X 1 ejϑi Ql (j, i) . ejϕj max √ Nt Nr j=1 i=1 (60) √ By discarding the constant coefficient 1/ Nt Nr , (60) can be further transformed as   Nt Nr   X X l l l ejϑu Ql (j, u) . (61) ejϕj ejϑi Ql (j, i) + max   u6=i j=1 l Since the term ejϑi does not involve the summation index j, it can be put outside the first summation, resulting in max e jϑli Nr X j=1 e jϕlj Ql (j, i) + Nr X j=1 e jϕlj Nt X l ejϑu Ql (j, u) . u6=i (62) Obviously, the optimal value of ϑli makes the phases of the first and second term equal to obtain the largest amplitude, and (20) is proved. R EFERENCES [1] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broadband systems,” IEEE Commun. Mag., vol. 49, no. 6, pp. 101-107, June 2011. [2] T. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N. Wong, J. K. Schulz, M. Samimi, and F. Gutierrez “Millimeter wave mobile communications for 5G cellular: It will work!” IEEE Access, vol. 1, pp. 335-349, 2013. [3] A. Lee Swindlehurst, E. Ayanoglu, P. Heydari, and F. Capolino, “Millimeter-wave massive MIMO: The next wireless revolution?” IEEE Commun. Mag., vol. 52, no. 9, pp. 56-62, Sept. 2014. [4] T. S. Rappapport, G. R. MacCartney, M. K. Samimi, and S. Sun, “Wideband millimeter-wave propagation measurements and channel models for future wireless communication system design”, IEEE Trans. Commun., vol. 63, no. 9, pp. 3029-3056, Sept. 2015. [5] R. W. Heath Jr., N. González-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed, “An overview of signal processing techniques for millimeter wave MIMO systems,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 436-453, April 2016. [6] X. Yu, J.-C. Shen, J. Zhang, and K. B. Letaief, “Alternating minimization algorithms for hybrid precoding in millimeter wave MIMO systems,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 485-500, April 2016. [7] C. Rusu, R. Méndez-Rial, N. González-Prelcic, and R. W. Heath Jr., “Low complexity hybrid precoding strategies for millimeter wave communication systems,” IEEE Trans. Wireless Commun., vol. 15, no. 12, pp. 8380-8393, Sept. 2016. [8] R. López-Valcarce, N. González-Prelcic, C. Rusu, and R. W. Heath Jr., “Hybrid precoders and combiners for mmWave MIMO systems with per-antenna power constraints.” in Proc. IEEE Global Communication Conference (GLOBECOM), Washington, DC, Dec. 2016, pp. 1-6. [9] C.-E. Chen, “An iterative hybrid transceiver design algorithm for millimeter wave MIMO systems,” IEEE Wireless Commun. Lett., vol. 4, no. 3, pp. 285-288, June 2015. [10] W. Ni, X. Dong, and W.-S. Lu, “Near-optimal hybrid processing for massive MIMO systems via matrix decomposition,” IEEE Trans. Signal Process., vol. 65, no. 15, pp. 3922-3933, Aug. 2017. [11] X. Gao, L. Dai, S. Han, C.-L. I, and R. W. Heath Jr., “Energy-efficient hybrid analog and digital precoding for mmWave MIMO systems with large antenna arrays,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 998-1009, April 2016. [12] L. Dai, X. Gao, J. Quan, S. Han, and C.-L. I, “Near-optimal hybrid analog and digital precoding for downlink mmWave massive MIMO systems”, in Proc. IEEE Int. Conf. Commun. (ICC), London, UK, June 2015, pp. 1334-1339. [13] S. He, C. Qi, Y. Wu, and Y. Huang, “Energy-efficient transceiver design for hybrid sub-array architecture MIMO systems,” IEEE Access, vol. 4, pp. 9895-9905, 2016. [14] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath Jr., “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE Trans. Wireless Commun., vol. 13, no. 3, pp. 1499-1513, Mar. 2014. [15] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath Jr., “Channel estimation and hybrid precoding for millimeter wave cellular systems,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831-846, Oct. 2014. [16] C. Rusu, R. Méndez-Rial, N. González-Prelcic, and R. W. Heath Jr., “Low complexity hybrid sparse precoding and combining in millimeter wave MIMO systems,” IEEE Int. Conf. on Commun. (ICC), London, UK, June 2015, pp. 1340-1345. [17] J.-C. Chen, “Efficient codebook-based beamforming algorithm for millimeter-wave massive MIMO systems,” IEEE Trans. Veh. Technol., to Appear. [18] X. Gao, L. Dai, C. Yuen, and Z. Wang, “Turbo-like beamforming based on Tabu search algorithm for millimeter-wave massive MIMO systems,” IEEE Trans. Veh. Technol., vol. 65, no. 7, pp. 5731-5737, July 2016. [19] S. Han, C.-L. I, Z. Xu, and C. Rowell, “Large-scale antenna systems with hybrid analog and digital beamforming for millimeter wave 5G,” IEEE Commun. Mag., vol. 53, no. 1, pp. 186-194, Jan. 2015. [20] A. Li and C. Masouros, “Hybrid analog-digital millimeter-wave MUMIMO transmission with virtual path selection,” IEEE Commun. Lett., vol. 21, no. 2, pp. 438-441, Feb. 2017. [21] L. Liang, W. Xu, and X. Dong, “Low-complexity hybrid precoding in massive multiuser MIMO systems,” IEEE Wireless Commun. Lett., vol. 3, no. 6, pp. 653-656, Dec. 2014. [22] A. Li and C. Masouros, “Hybrid precoding and combining design for millimeter-wave multi-user MIMO based on SVD,” IEEE Int. Conf. on Commun. (ICC), Paris, France, May 2017, pp. 1-6. [23] M. Kim and Y. H. Lee, “MSE-based hybrid RF/baseband processing for millimeter-wave communication systems in MIMO interference channels,” IEEE Trans. Veh. Technol., vol. 64, no. 6, pp. 2714-2720, June 2015. [24] A. Alkhateeb, G. Leus, and R. W. Heath Jr., “Limited feedback hybrid precoding for multi-user millimeter wave systems,” IEEE Trans. Wireless Commun., vol. 14, no. 11, pp. 6481-6494, Nov. 2015. [25] D. H. N. Nguyen, L. B. Le, and T. Le-Ngoc, “Hybrid MMSE precoding for mmWave multiuser MIMO systems,” IEEE Int. Conf. on Commun. (ICC), Kuala Lumpur, Malaysia, May 2016. [26] Z. Wang, M. Li, X. Tian, and Q. Liu, “Iterative hybrid precoder and combiner design for mmWave multiuser MIMO systems,” IEEE Commun. Lett., vol. 21, no. 7, pp. 1581-1584, July 2017. [27] T. E. Bogale, L. B. Le, A. Haghighat, and L. Vandendorpe, “On the number of RF chains and phase shifters, and scheduling design with hybrid analog-digital beamforming,” IEEE Trans. Wireless Commun., vol. 15, no. 5, pp. 3311-3326, May 2016. [28] A. S. Y. Poon and M. Taghivand, “Supporting and enabling circuits for antenna arrays in wireless communications,” Proc. IEEE, vol. 100, no. 7, pp. 2207-2218, July 2012. [29] R. Méndez-Rial, C. Rusu, N. González-Prelcic, and A. Alkhateeb, “Hybrid MIMO architectures for millimeter wave communications: Phase shifters or switches?”, IEEE Access, vol. 4, pp. 247-267, March 2016. [30] J.-C. Chen, “Hyrbrid beamforming with discrete phase shifters for millimeter-wave massive MIMO systems,” IEEE Trans. Veh. Technol., to Appear. [31] F. Sohrabi and W. Yu, “Hybrid beamforming with finite-resolution phase shifters for large-scale MIMO systems,” in Proc. IEEE Workshop Signal Process. Adv. Wireless Commun. (SPAWC), Stockholm, Sweden, June 2015, pp. 136-140. [32] F. Sohrabi and W. Yu, “Hybrid digital and analog beamforming design for large-scale antenna arrays,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 501-513, April 2016. [33] G. N. Karystinos and D. A. Pados, “Rank-2-optimal adaptive design of binary spreading codes,” IEEE Trans. Inf. Theory, vol. 53, no. 9, pp. 3075-3080, Sept. 2007. [34] G. H. Golub and C. F. Van Loan, Matrix Computations. Baltimore, USA: JHU Press, 2012.	7cs.IT
Phaseless compressive sensing using partial support information arXiv:1705.04048v2 [cs.IT] 13 Dec 2017 Zhiyong Zhou∗, Jun Yu Department of Mathematics and Mathematical Statistics, Umeå University, Umeå, 901 87, Sweden December 14, 2017 Abstract: We study the recovery conditions of weighted ℓ1 minimization for real-valued signal reconstruction from phaseless compressive sensing measurements when partial support information is available. A strong restricted isometry property condition is provided to ensure the stable recovery. Moreover, we present the weighted null space property as the sufficient and necessary condition for the success of k-sparse phaseless recovery via weighted ℓ1 minimization. Numerical experiments are conducted to illustrate our results. Keywords: Phaseless compressive sensing; Partial support information; Strong restricted isometry property; Weighted null space property. 1 Introduction Compressive sensing aims to recover an unknown signal from the underdetermined linear measurements (see [8, 9] for a comprehensive view). It is known as phase retrieval or phaseless compressive sensing when there is no phase information. The phaseless compressive sensing problem has recently attracted considerable research interests and many algorithms have been proposed to solve this problem. Existing literature include [2–4,7,12,14,16], to name a few. Specifically, the goal of phaseless compressive sensing is to recover x ∈ RN up to a unimodular scaling constant from noisy magnitude measurements y = \|Ax\|+e ∈ Rm with the measurement matrix A = (a1 , · · · , am )T ∈ Rm×N , \|Ax\| = (\|ha1 , xi\|, · · · , \|ham , xi\|)T and the noise term e ∈ Rm . When x is sparse or compressible, the stable recovery can be guaranteed by solving the following ℓ1 minimization problem min kzk1 subject to k\|Az\| − yk2 ≤ ε, z∈RN (1) provided that the measurement matrix A satisfies the strong restricted isometry property (SRIP) [11, 17]. In the noiseless case, the first sufficient and necessary condition was presented in [18] by proposing a new version of null space property for the phase retrieval problem. In this paper, we generalize the existing theoretical framework for phaseless compressive sensing to incorporate partial support information, where we consider the case that an estimate of the support of the signal is available. We follow the similar notations and arguments in [10, 20]. For an arbitrary signal x ∈ RN , let xk be its best k-term approximation, so that xk minimizes kx − f k1 over all k-sparse vectors f . Let T0 be the support of xk , where T0 ⊂ {1, · · · , N } and \|T0 \| ≤ k. Let T̃ , the support estimate, be a subset of {1, 2 · · · , N } with cardinality \|T̃ \| = ρk, where ρ ≥ 0 and \|T̃ ∩ T0 \| = αρk with 0 ≤ α ≤ 1. Here the parameter ρ determines the ratio of the size of the estimated support to the size of the actual support of xk ∗ Corresponding author, zhiyong.zhou@umu.se. 1 (or the support of x if x is k-sparse), while the parameter α determines the ratio of the number of indices in the support of xk that are accurately estimated in T̃ to the size of T̃ , i.e., α = prior support information T̃ , we adopt the weighted ℓ1 minimization N X ( ω ∈ [0, 1] min wi \|zi \|, subject to k\|Az\| − yk2 ≤ ε, where wi = N z∈R 1 i=1 \|T̃ ∩T0 \| . \|T̃ \| To incorporate i ∈ T̃ , i ∈ T̃ c . (2) We present the SRIP condition and weighted null space property condition to guarantee the success of the recovery via the weighted ℓ1 minimization problem above. The paper is organized as follows. In Section 2, we introduce the definition of SRIP and present the stable recovery condition with this tool. In Section 3, the sufficient and necessary weighted null space property condition for the real sparse noise free phase retrieval is given. In Section 4, some numerical experiments are presented to illustrate our theoretical results. Finally, Section 5 is devoted to the conclusion. P Throughout the paper, for any vector x ∈ RN , we denote the ℓp norm by kxkp = ( pi=1 \|xi \|p )1/p for P p > 0 and the weighted ℓ1 norm as kxk1,w = N i=1 wi \|xi \|. For any matrix X, kXk1 denotes the entry-wise ℓ1 norm. For any set T , we denote its cardinality as \|T \|. The vector x ∈ RN is called k-sparse if at most k of its entries are nonzero, i.e., if kxk0 = \|supp(x)\| ≤ k, where supp(x) denotes the index set of the nonzero entries. We denote the index set [N ] := {1, 2, · · · , N }. For a matrix A = (a1 , · · · , am )T ∈ Rm×N and an index set I ⊂ [m], we denote AI the sub-matrix of A where only rows with indices in I are kept, i.e., AI = (aj , j ∈ I)T . 2 SRIP To recover sparse signals via ℓ1 minimization in the classical compressive sensing setting, [5] introduced the notion of restricted isometry property (RIP) and established a sufficient condition. We say a matrix A satisfies the RIP of order k if there exists a constant δk ∈ [0, 1) such that for all k-sparse vectors x we have (1 − δk )kxk22 ≤ kAxk22 ≤ (1 + δk )kxk22 . (3) q Cai and Zhang [1] proved that the RIP of order tk with δtk < t−1 t where t > 1 can guarantee the exact recovery in the noiseless case and stable recovery in the noisy case via ℓ1 minimization. This condition is sharp when t ≥ 43 , see [1] for details. Very recently, Chen and Li [6] generalized this sharp RIP condition to the weighted ℓ1 minimization problem when partial support information was incorporated. We first present the following useful lemma, which is an extension of the result in [6]. N m Lemma 1 Let qx ∈ R , y = Ax + e ∈ R with kek2 ≤ ζ, and η√≥ 0. Suppose that A satisfies RIP of order t−d 1 + ρ − 2αρ and tk with δtk < t−d+γ 2 for some t > d, where γ = ω + (1 − ω) d= ( 1, 1 − αρ + a, ω=1 0≤ω<1 (4) with a = max{α, 1 − α}ρ. Then for any x̂ ∈ {z ∈ RN : kzk1,w ≤ kxk1,w + η, kAz − yk2 ≤ ε}, we have kx̂ − xk2 ≤ C1 (ζ + ε) + C2 2(ωkxT0c k1 + (1 − ω)kxT̃ c ∩T c k1 ) η 0 √ + C2 √ , k k 2 (5) where p 2(t − d)(t − d + γ 2 )(1 + δtk ) q , t−d (t − d + γ 2 )( t−d+γ − δ ) 2 tk r q √ t−d 2δtk γ + (t − d + γ 2 )( t−d+γ 2 − δtk )δtk 1 q +√ . C2 = t−d d (t − d + γ 2 )( t−d+γ 2 − δtk ) C1 = Remark 1 Note that if xℓ2 is the solution of the weighted ℓ1 minimization problem: min kzk1,w , subject to kAz − yk2 ≤ ε, z∈RN then xℓ2 ∈ {z ∈ RN : kzk1,w ≤ kxk1,w + η, kAz − yk2 ≤ ε} with η = 0. Therefore, this lemma is an extension of Theorem 3.1 in [6] by letting ζ = ε and η = 0. The proof follows from almost the same procedure for the proof of Theorem 3.1 in Section 4 of [6] via replacing the P = 2(ωkxT c k1 +(1−ω)kxT̃ c ∩T c k1 ) 0 0 √ kγ 2(ωkxT c k1 +(1−ω)kxT̃ c ∩T c k1 )+η 0 0 √ with P ′ = , and letting ζ = ε. In order not to repeat, we leave out all the kγ details. In addition, this result also generalizes the Lemma 2.1 in [11], which is the special case with the noise term e = 0, ζ = 0 and ω = 1. This lemma will play a crucial role in establishing the stable phaseless recovery result via weighted ℓ1 minimization later on. To address the phaseless compressive sensing problem (2), a stronger version of RIP is needed. Its definition is provided as follows. Definition 1 (SRIP [11, 17]) We say a matrix A = (a1 , · · · , am )T ∈ Rm×N has the Strong Restricted Isometry Property (SRIP) of order k with bounds θ− , θ+ ∈ (0, 2) if θ− kxk22 ≤ min kAI xk22 ≤ I⊆[m],\|I\|≥m/2 max kAI xk22 ≤ θ+ kxk22 I⊆[m],\|I\|≥m/2 (6) holds for all k-sparse vectors x ∈ RN , where [m] = {1, · · · , m}. We say A has the Strong Lower Restricted Isometry Property of order k with bound θ− if the lower bound in (6) holds. Similarly, we say A has the Strong Upper Restricted Isometry Property of order k with bound θ+ if the upper bound in (6) holds. Next, we present the conditions for the stable recovery via weighted ℓ1 minimization by using SRIP. Theorem 1 Let x ∈ RN , y = \|Ax\| + e ∈ Rm with kek2 ≤ ζ. Adopt the notations in Lemma 1 and assume that A ∈ Rm×N satisfies the SRIP of order tk with bounds θ− , θ+ ∈ (0, 2) such that γ 2 (1 − θ+ )2 γ 2 (1 − θ− )2 . (7) t ≥ max d + 2 , d + 2θ − θ 2 2θ− − θ− + + Then any solution x♯ of (2) satisfies min{kx♯ − xk2 , kx♯ + xk2 } ≤ C1 (ζ + ε) + C2 where C1 and C2 are constants defined in Lemma 1. 3 2(ωkxT0c k1 + (1 − ω)kxT̃ c ∩T c k1 ) 0 √ . k (8) Remark 2 As it has been proved in [17] that Gaussian matrices with m = O(tk log(N/k)) satisfy SRIP of order tk with high probability, thus the stable recovery result (8) can be achieved by using Gaussian measurement matrix with appropriate number of measurements m. Remark 3 Note that when the weight ω = 1, we have γ = d = 1. Then, by assuming ζ = ε = 0 and x is exactly k-sparse, our theorem reduces to Theorem 2.2 in [17]. That is, if A satisfies the SRIP of order tk with bounds θ− , θ+ and t ≥ max{ 2θ 1−θ2 , 2θ 1−θ2 }, then for any k-sparse signal x ∈ RN we have − − + + arg minz∈RN {kzk1 : \|Az\| = \|Ax\|} = {±x}. Similarly, if we let the noise term e = 0, ζ = 0 and ω = 1, this theorem goes to Theorem 3.1 in [11]. Remark 4 If α = 21 , we have γ = d = 1. The sufficient condition (7) of Theorem 1 is identical to that √ 2(1+δtk ) of Theorem 2.2 in [17] and that of Theorem 3.1 in [11]. And the constants C1 = c1 = √ , C2 = 1− t/(t−1)δtk q √ √ 2δtk + ( t(t−1)−δtk t)δtk √ c2 = (see Theorem 3.1 in [11]). In addition, if 0 ≤ ω < 1 and α > 21 , then d = 1 t(t−1)−δtk t and γ < 1. The sufficient condition (7) in Theorem 1 is weaker than that of Theorem 2.2 in [17] and that of Theorem 3.1 in [11]. In this case, the constants C1 < c1 , C2 < c2 . o n 2 2 γ 2 (1−θ+ )2 −) . We illustrate how the constants tω , C1 and C2 change , d + Set tω = max d + γ 2θ(1−θ −θ 2 2θ −θ 2 − − + + with ω for different values of α in Figure 1. In all the plots, we set ρ = 1. In the plot of tω , we set θ− = γ2 1 2 and θ+ = 23 , then tω = d + 3 . In the plots of C1 and C2 , we fix t = 4 and δtk = 0.3. Note that if ω = 1 or α = 0.5, then tω ≡ 1 + 13 = 34 , C1 ≡ c1 and C1 ≡ c2 . And it shows that tω decreases as α increases, which means that the sufficient condition (7) becomes weaker as α increases. For each α > 0.5, the sufficient condition becomes stronger (tω increases) as ω increases. For instance, if 90% of the support estimate is accurate (α = 0.9) and ω = 0.6, we have tω = 1.2022, while tω = 1.3333 for standard ℓ1 minimization (ω = 1). The opposite conclusion holds for the case α < 0.5. In addition, as α increases, the constant C1 decreases with t = 4 and δtk = 0.3. Meanwhile, the constant C2 with α 6= 0.5 is smaller than that with α = 0.5. Proof of Theorem 1. For any solution x♯ of (2), we have kx♯ k1,w ≤ kxk1,w and k\|Ax♯ \| − \|Ax\| − ek2 ≤ ε. If we divide the index set {1, 2, · · · , m} into two subsets T = {j : sign(haj , x♯ i) = sign(haj , xi)} and T c = {j : sign(haj , x♯ i) = −sign(haj , xi)}, then it implies that kAT x♯ − AT x − ek2 + kAT c x♯ + AT c x − ek2 ≤ ε. (9) Here either \|T \| ≥ m/2 or \|T c \| ≥ m/2. If \|T \| ≥ m/2, we use the fact that kAT x♯ − AT x − ek2 ≤ ε. Then, we obtain x♯ ∈ {z ∈ RN : kzk1,w ≤ kxk1,w , kAT z − AT x − ek2 ≤ ε}. 4 (10) 2.4 2.75 1.56 α=0.1 α=0.3 α=0.5 α=0.7 α=0.9 2.7 2.2 α=0.1 α=0.3 α=0.5 α=0.7 α=0.9 2 1.54 2.65 1.52 2.6 1.8 1.5 2.5 1.6 α=0.1 α=0.3 α=0.5 α=0.7 α=0.9 C2 tω C1 2.55 1.48 2.45 1.4 1.46 2.4 1.2 1.44 2.35 1 2.3 0 0.2 0.4 0.6 0.8 1 1.42 0 0.2 Weights (ω) 0.4 0.6 0.8 1 0 0.2 Weights (ω) 0.4 0.6 0.8 1 Weights (ω) Figure 1: Comparison of the constants tω , C1 and C2 for various of α. In all the plots, we set ρ = 1. In the plot of tω , we set θ− = 21 and θ+ = 32 . In the plots of C1 and C2 , we fix t = 4 and δtk = 0.3. Since A satisfies SRIP of order tk with bounds θ− , θ+ and γ 2 (1 − θ− )2 γ 2 (1 − θ+ )2 t ≥ max d + > d, 2 , d + 2θ − θ 2 2θ− − θ− + + therefore, the definition of SRIP implies that AT satisfies the RIP of order tk with s t−d . δtk ≤ max{1 − θ− , θ+ − 1} ≤ t − d + γ2 (11) Thus, by using Lemma 1 with η = 0, we have kx♯ − xk2 ≤ C1 (ζ + ε) + C2 2(ωkxT0c k1 + (1 − ω)kxT̃ c ∩T c k1 ) 0 √ . k Similarly, if \|T c \| ≥ m/2, we obtain the other corresponding result kx♯ + xk2 ≤ C1 (ζ + ε) + C2 2(ωkxT0c k1 + (1 − ω)kxT̃ c ∩T c k1 ) 0 √ . k The proof of Theorem 1 is now completed. 3 Weighted Null Space Property In this section, we consider the noiseless weighted ℓ1 minimization problem, i.e., ( ω ∈ [0, 1], i ∈ T̃ . min kzk1,w , subject to \|Az\| = \|Ax\|, where wi = N z∈R 1, i ∈ T̃ c 5 (12) We denote the kernel space of A by N (A) := {h ∈ RN : Ah = 0} and denote the k-sparse vector space N ΣN k := {x ∈ R : kxk0 ≤ k}. Definition 2 The matrix A satisfies the w-weighted null space property of order k if for any nonzero h ∈ N (A) and any T ⊂ [N ] with \|T \| ≤ k it holds that khT k1,w < khT c k1,w , (13) where T c is the complementary index set of T and hT is the restriction of h to T . Remark 5 Obviously, when the weight ω = 1, the weighted null space property reduces to the classical null space property. And according to the specific setting of wi , the expression (13) is equivalent to ωkhT ∩T̃ k1 + khT ∩T̃ c k1 < ωkhT c ∩T̃ k1 + khT c ∩T̃ c k1 ⇔ ωkhT k1 + (1 − ω)khG k1 < khT c k1 , where G = (T ∩ T̃ c ) ∪ (T c ∩ T̃ ) (see [13] for more arguments). It is known that a signal x ∈ ΣN k can be recovered via the weighted ℓ1 minimization problem if and only if the measurement matrix A has the weighted null space property of order k. We state it as follows (see [19]): Lemma 2 Given A ∈ Rm×N , for every k-sparse vector x ∈ RN it holds that arg min {kzk1,w : Az = Ax} = x z∈RN if and only if A satisfies the w-weighted null space property of order k. Next, we extend Lemma 2 to the following theorem on phaseless compressive sensing for the real-valued signal reconstruction. Theorem 2 The following statements are equivalent: (a) For any k-sparse x ∈ RN , we have arg min{kzk1,w : \|Az\| = \|Ax\|} = {±x}. (14) ku + vk1,w < ku − vk1,w (15) z∈RN (b) For every S ⊆ [m], it holds for all nonzero u ∈ N (AS ) and v ∈ N (AS c ) satisfying ku + vk0 ≤ k. Remark 6 If ω = 1, then Theorem 2 reduces to Theorem 3.2 in [18]. Since wi = ω when i ∈ T̃ , and wi = 1 otherwise, the expression (15) is equivalent to ωku + vk1 + (1 − ω)k(u + v)T̃ c k1 < ωku − vk1 + (1 − ω)k(u − v)T̃ c k1 . 6 Proof of Theorem 2. The proof follows from the proof of Theorem 3.2 in [18] with minor modifications. First we show (a) ⇒ (b). Assume (b) is false, that is, there exist nonzero u ∈ N (AS ) and v ∈ N (AS c ) such that ku + vk1,w ≥ ku − vk1,w N and u + v ∈ ΣN k . Now set x = u + v ∈ Σk , obviously for i = 1, · · · , m, we have \|hai , xi\| = \|hai , u + vi\| = \|hai , u − vi\|, since either hai , ui = 0 or hai , vi = 0. In other words \|Ax\| = \|A(u− v)\|. Note that u− v 6= −x, for otherwise we would have u = 0, which is a contradiction. Then, it follows from (a) that we obtain kxk1,w = ku + vk1,w < ku − vk1,w , This is a contradiction. Thus, (b) holds. T Next we prove (b) ⇒ (a). Let b = (b1 , · · · , bm )T = \|Ax\| where x ∈ ΣN k . For a fixed σ = (σ1 , · · · , σm ) ∈ {−1, 1}m , we set bσ = (σ1 b1 , · · · , σm bm )T . We now consider the following weighted ℓ1 minimization problem: min kzk1,w subject to Az = bσ . (16) z∈RN Its solution is denoted as xσ . Then, we claim that for any σ ∈ {1, −1}m , if xσ exists (it may not exist), we have kxσ k1,w ≥ kxk1,w and the equality holds if and only if xσ = ±x. ⋆ To prove the claim, we assume σ ⋆ ∈ {1, −1}m such that bσ = Ax. First note that the statement (b) implies the classical weighted null space property of order k. To see this, for any nonzero h ∈ N (A) and T ⊆ [N ] with \|T \| ≤ k, we set u = h, v = hT − hT c and S = [m]. Then, we have u ∈ N (AS ) and v ∈ N (AS c ). Therefore, the statement (b) now implies 2khT k1,w = ku + vk1,w < ku − vk1,w = 2khT c k1,w . ⋆ As a consequence, we have xσ = x by Lemma 2. And, σ ∈ {−1, 1}m 6= ±σ ⋆ , if xσ doesn’t exist then we have S⋆ = {i : σi = σi⋆ }. Then ( hai , xi hai , xσ i = −hai , xi ⋆ similarly we have x−σ = −x. Next, for any nothing to prove. Assume it does exist, set i ∈ S⋆ , i ∈ S⋆c . Set u = x − xσ and v = x + xσ . Obviously, u ∈ N (AS⋆ ) and v ∈ N (AS⋆c ). Furthermore, u + v = 2x ∈ ΣN k . Then, by the statement (b), we have 2kxk1,w = ku + vk1,w < ku − vk1,w = 2kxσ k1,w . This proves (a) and the proof is completed. 7 4 Simulations In this section, we present some simple numerical experiments to illustrate the benefits of using weighted ℓ1 minimization to recover sparse and compressible signals when partial prior support information is available in the phaseless compressive sensing case. In order to facilitate the computation, we follow a nonstandard noise model: b = \|Ax\|2 + e = {aTi xxT ai }1≤i≤m + e, where e ∈ Rm is a noise term with kek2 ≤ ε. Then the weighted ℓ1 minimization goes to ( N X ω ∈ [0, 1] i ∈ T̃ , 2 wi \|zi \|, subject to k\|Az\| − bk2 ≤ ε, where wi = min z∈RN 1 i ∈ T̃ c . i=1 (17) (18) Here we adopt the compressive phase retrieval via lifting (CPRL) algorithm developed in [15] to solve this phaseless recovery problem. By using a lifting technique, this problem can be rewritten as a semidefinite program (SDP). More specifically, given the ground truth signal x ∈ RN , let X = xxT ∈ RN ×N be an induced rank-1 semidefinite matrix. We further denote Φi = ai aTi , a linear operator B of Z = zz T ∈ RN ×N as B : Z 7→ {Tr(Φi Z)}1≤i≤m ∈ Rm and the weight matrix W = diag{wi , 1 ≤ i ≤ N } ∈ RN ×N . Then the phaseless vector recovery problem (18) can be cast as the following rank-1 matrix recovery problem: min Z∈RN×N kW ZW T k1 , subject to kB(Z) − bk2 ≤ ε, rank(W ZW T ) = 1, Z 0. This is of course still a non-convex problem due to the rank constraint. The lifting approach addresses this issue by replacing rank(W ZW T ) with Tr(W ZW T ). This leads to an SDP: min Z∈RN×N Tr(W ZW T ) + λkW ZW T k1 , subject to kB(Z) − bk2 ≤ ε, Z 0, (19) where λ > 0 is a design parameter. Then the estimate of x can be finally be found by computing the rank-1 decomposition of the recovered matrix via singular value decomposition. The recovery performance is assessed by the average reconstruction signal to noise ratio (SNR) over 10 experiments. The SNR is measured in dB and it is given by kxk2 ♯ , (20) SNR(x, x ) = 20 log10 min{kx♯ − xk2 , kx♯ + xk2 } where x is the true signal and x♯ is the recovered signal. For all the experiments, we fix the parameter i.i.d λ = 1. In the experiments where the measurements are noisy, we set the noise {ei , 1 ≤ i ≤ m} ∼ N (0, σ 2 ) with σ = 0.1 and ε = kek2 . 8 4.1 Sparse Case We first consider the case that x is exactly sparse with an ambient dimension N = 32 and fixed sparsity k = 4. The sparse signals are generated by choosing k nonzero positions uniformly at random, and then choosing the nonzero values from the standard normal distribution for these k nonzero positions. The recovery is done via (19) using a support estimate of size \|T̃ \| = 4 (i.e., ρ = 1). Figure 2 shows the recovery performances for different α and ω with an increasing number of measurements m, both in the noise free and noisy cases. It can be observed that when α = 0.75 > 0.5, the best recovery is achieved for very small ω whereas a ω = 1 results in the lowest SNR for both cases. On the other hand, when α = 0.25 < 0.5, the performance of the recovery algorithms is better for large ω than that for small ω. The case ω = 0 results in the lowest SNR. When α = 0.5, the performance gaps for different ω are not particularly large and it seems that a medium ω (ω = 0.5) achieves the best recovery. In the noise free case, a perfect recovery can be achieved as long as the number of measurements m is large enough. As is also expected that in all settings, comparing to the noise free case, we have a lower SNR in the noisy case. These findings are largely consistent with the theoretical results provided in Section 2. α=0.75 α=0.5 200 α=0.25 200 120 100 150 100 0 10 100 ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 50 14 18 22 26 80 30 40 ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 50 0 10 60 SNR SNR SNR 150 14 18 22 26 number of measurements m number of measurements m α=0.75 α=0.5 ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 20 0 30 -20 10 14 18 22 26 30 number of measurements m (a) α=0.25 25 30 25 15 20 10 20 15 SNR SNR SNR 15 5 10 10 ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 5 0 10 14 18 22 26 number of measurements m ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 5 30 0 10 14 18 22 26 number of measurements m ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 0 30 -5 10 14 18 22 26 30 number of measurements m (b) Figure 2: Performance of weighted ℓ1 recovery in terms of SNR averaged over 10 experiments for sparse signals x with N = 32, k = 4, while varying the number of measurements m. From left to right, α = 0.75, α = 0.5 and α = 0.25. (a) Noise Free. (b) σ = 0.1. 4.2 Compressible Case Here we generate a signal x whose coefficients decay like j −θ where j ∈ {1, · · · , N } and θ = 4.5. This kind of signal itself is not sparse, but can be well approximated by an exactly sparse signal. For this experiment, we set k = 4, i.e., we use the best 4-term approximation. We fix ρ = 1 as in the sparse case. The phaseless recovery results are presented in Figure 3. It shows that on average a mediate value of ω (ω = 0.5) results in the best recovery. In general, when α > 0.5, smaller ω favours better reconstruction results. The opposite conclusion holds for the case that α < 0.5. Therefore, as is expected that the behaviors that occur in the exactly sparse case also occur in the compressible case. 9 α=0.75 α=0.5 α=0.25 80 120 100 100 80 60 80 60 60 SNR SNR SNR 40 40 20 40 ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 20 0 10 14 18 22 26 30 -20 10 20 ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 0 14 18 22 26 number of measurements m number of measurements m α=0.75 α=0.5 ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 0 30 -20 10 14 18 22 26 30 number of measurements m (a) 25 30 20 25 15 20 ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 15 10 14 18 22 26 number of measurements m 30 30 ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 20 10 15 10 5 5 0 0 -5 10 14 ω=0 ω=0.2 ω=0.5 ω=0.8 ω=1 25 SNR 35 5 10 α=0.25 30 SNR SNR 40 18 22 26 number of measurements m 30 -5 10 14 18 22 26 30 number of measurements m (b) Figure 3: Performance of weighted ℓ1 recovery in terms of SNR averaged over 10 experiments for compressible signals x with N = 32, θ = 4.5, while varying the number of measurements m. From left to right, α = 0.75, α = 0.5 and α = 0.25. (a) Noise Free. (b) σ = 0.1. 5 Conclusion In this paper, we established the sufficient SRIP condition and the sufficient and necessary weighted null space property condition for phaseless compressive sensing using partial support information via weighted ℓ1 minimization, and we conducted some numerical experiments to illustrate the theoretical results. Some further problems are left for future work. As we only consider the real-valued signal reconstruction case, it will be challenging to generalize the present results to the complex-valued signal case. Besides it will be very interesting to construct the measurement matrix A ∈ Rm×N satisfying the weighted null space property given in (15) directly. Acknowledgements This work is supported by the Swedish Research Council grant (Reg.No. 340-2013-5342). References [1] Cai, T. T. and Zhang, A. (2014). Sparse representation of a polytope and recovery of sparse signals and low-rank matrices. IEEE Transactions on Information Theory 60(1) 122–132. [2] Candes, E. J., Eldar, Y. C., Strohmer, T. and Voroninski, V. (2015). Phase retrieval via matrix completion. SIAM review 57(2) 225–251. [3] Candes, E. J., Li, X. and Soltanolkotabi, M. (2015). Phase retrieval via Wirtinger flow: Theory and algorithms. IEEE Transactions on Information Theory 61(4) 1985–2007. 10 [4] Candes, E. J., Strohmer, T. and Voroninski, V. (2013). Phaselift: Exact and stable signal recovery from magnitude measurements via convex programming. Communications on Pure and Applied Mathematics 66(8) 1241–1274. [5] Candes, E. J. and Tao, T. (2005). Decoding by linear programming. IEEE Transactions on Information Theory 51(12) 4203–4215. [6] Chen, W. and Li, Y. (2016). Recovery of signals under the high order RIP condition via prior support information. arXiv preprint arXiv:1603.03464. [7] Chen, Y. and Candes, E. (2015). Solving random quadratic systems of equations is nearly as easy as solving linear systems. In Advances in Neural Information Processing Systems (pp. 739–747). [8] Eldar, Y. C. and Kutyniok, G. (2012). Compressed Sensing: Theory and Applications. Cambridge University Press. [9] Foucart, S. and Rauhut, H. (2013). A Mathematical Introduction to Compressive Sensing. New York, NY, USA: Springer-Verlag. [10] Friedlander, M. P., Mansour, H., Saab, R. and Yilmaz, O. (2012). Recovering compressively sampled signals using partial support information. IEEE Transactions on Information Theory 58(2) 1122–1134. [11] Gao, B., Wang, Y. and Xu, Z. (2016). Stable signal recovery from phaseless measurements. Journal of Fourier Analysis and Applications 22(4) 787–808. [12] Gao, B. and Xu, Z. (2017). Phaseless recovery using the GaussNewton method. IEEE Transactions on Signal Processing 65(22) 5885–5896. [13] Mansour, H. and Saab, R. (2017). Recovery analysis for weighted ℓ1 -minimization using the null space property. Applied and Computational Harmonic Analysis 43(1) 23–38. [14] Netrapalli, P., Jain, P. and Sanghavi, S. (2015). Phase retrieval using alternating minimization. IEEE Transactions on Signal Processing 18(63) 4814–4826. [15] Ohlsson, H., Yang, A. Y., Dong, R. and Sastry, S. S. (2011). Compressive phase retrieval from squared output measurements via semidefinite programming. arXiv preprint arXiv:1111.6323. [16] Shechtman, Y., Beck, A. and Eldar, Y. C. (2014). GESPAR: Efficient phase retrieval of sparse signals. IEEE Transactions on Signal Processing 62(4) 928–938. [17] Voroninski, V. and Xu, Z. (2016). A strong restricted isometry property, with an application to phaseless compressed sensing. Applied and Computational Harmonic Analysis 40(2) 386–395. [18] Wang, Y. and Xu, Z. (2014). Phase retrieval for sparse signals. Applied and Computational Harmonic Analysis 37(3) 531–544. [19] Zhou, S., Xiu, N., Wang, Y. and Kong, L. (2013). Exact recovery for sparse signal via weighted ℓ1 minimization. arXiv preprint arXiv:1312.2358. [20] Zhou, Z. and Yu, J. (2017). Recovery analysis for weighted mixed ℓ2 /ℓp minimization with 0 < p ≤ 1. arXiv preprint arXiv:1709.00257. 11	7cs.IT
Inferential results for a new measure of inequality Youri Davydova , Francesca Greselinb a Laboratoire Paul Painlevé, Université des Sciences et Technologies de Lille (Lille 1), arXiv:1706.05576v1 [math.ST] 17 Jun 2017 Lille, France and Saint Petersburg State University, Saint Petersburg, Russia b Dipartimento di Statistica e Metodi Quantitativi, Università di Milano Bicocca, Milan, Italy Abstract. In this paper we derive inferential results for a new index of inequality, specifically defined for capturing significant changes observed both in the left and in the right tail of the income distributions. The latter shifts are an apparent fact for many countries like US, Germany, UK, and France in the last decades, and are a concern for many policy makers. We propose two empirical estimators for the index, and show that they are asymptotically equivalent. Afterwards, we adopt one estimator and prove its consistency and asymptotic normality. Finally we introduce an empirical estimator for its variance and provide conditions to show its convergence to the finite theoretical value. An analysis of real data on net income from the Bank of Italy Survey of Income and Wealth is also presented, on the base of the obtained inferential results. Keywords and phrases: Income inequality, Lorenz curve, Gini Index, consistency, asymptotic normality, economic inequality, confidence interval, nonparametric estimator. 1 Introduction In view of measuring economic inequality in a society, suppose that we are interested, for instance, in incomes. Let X be an ’income’ random variable with non negatively supported cdf F (x). 1 Next, define Q(p) = F −1 (p) = inf{x : F (x) ≥ p, p ∈ [0, 1]} as the p-th quantile of X, and suppose that X possesses a finite mean Z Z ∞ xF (dx) = µF = 0 1 F −1 (p)dp. 0 The Lorenz curve, introduced by Lorenz (1905), is an irreplaceable tool in this domain. It is defined by 1 lF (p) = µF Z p F −1 (t)dt. (1.1) 0 The curve lF (p) expresses the share of income possessed by the p% poorer part of population. It has been expressed firstly by Pietra (1915, with English translation now available as Pietra, 2014), and mathematically formulated as in (1.1) by Gastwirth (1971). In the following we will also employ mF (p) = 1 − lF (1 − p), which provides the share of income owned by the richer p% of the population. Obviously, mF (p) is the curve obtained by applying a central symmetry to lF (p), with respect to the center of the unit square, as R1 shown in Figure 1.1 and allows us also to rephrase the Gini into GF = 0 (mF (p)−lF (p))dp. 1.0 We recall that the Gini can be rephrased as the weighted average of all comparisons made 0.0 0.2 0.4 0.6 0.8 l F (t ) m F (t ) 0.0 0.2 0.4 0.6 0.8 1.0 Figure 1.1: Curves lF (p) and mF (p) among the mean income of the poorest and the overall mean (Greselin et al. 2012, Greselin 2014). When dealing with skewed distributions, as it is the case for many economic size distributions, the median should be preferred to the mean, in such a way that Gastwirth (2014) proposed to modify the Gini accordingly. Very recently, motivated by the observed shifts toward the extreme values in income distributions, a new focus is introduced in Gastwirth (2016), almost contemporarily to 2 Davydov and Greselin (2016). Policy makers are nowadays interested in understanding what happens in the more critical portions of the population, as significant changes have been observed both in the left and in the right tail of the income distributions in countries like US, Germany, UK, France in the last decades. Notice that the classical Lorenz curve provides useful pointwise information with reference to poorest people, while on the other hand, as L(p) approaches 1 as p approaches 1, it does not display the variation within the upper end (f.i., top 5% or 1%) of the distribution clearly. The novel approach is to consider equally sized and opposite groups of population, and compare their mean income. Aiming at contrasting the economic position of the group of the poorer p% to the one of the p% of the richest, the following inequality curve has been introduced DF (p) = mF (p) − lF (p) . mF (p) (1.2) In the case of perfect equality, each fraction p of population has same mean income, hence DF (p) = 0 for all p ∈ [0, 1]. While the income distribution moves toward more variability, the mean income income µF lF (p) p µF mF (p) p of the p% of richest people will be moving far from the mean of the p% of poorest part of the population, and DF (p) raises toward 1. Hence, we can represent the pointwise measure of inequality in the population by plotting DF (p). Naturally, we can summarize all the information given by the curve DF (p) in a single measure of inequality DF , by taking the expected value Z 1 DF (p)dp. DF = (1.3) 0 Notice that DF is the area between the observed inequality curve DF (p) and the curve of perfect equality, which is the horizontal line passing through the origin of the axes. The structure of the paper is as follows. Section 2 introduces two estimators for the new inequality measure, and provide reasons for selecting them in view of their main purpose. The third section, which is the core of the paper, states the main inferential results, in more detail we will show the consistency of the estimators, state their asymptotical distribution, and the asymptotic negligibility of their difference. We also introduce an empirical estimator for the variance, and establish its convergence to the finite variance of the estimator. Some lemmas useful for the inferential theory have been presented in Section 4, along with their proof. Section 5 shows how the inferential results can be employed to develop an analysis on real income data. Final considerations are given in Section 6. 3 2 Estimators Economic data on the entire (or complete) population is rarely available, so most studies are based on data obtained from well-designed sample surveys. Hence we usually have to estimate summary measures from samples. We introduce here two empirical estimators, b n and D e n for estimating DF . The first one is derived, in a very natural way, say D by replacing the population cdf F (x) and mean value µF in (1.3) by their empirical P P counterparts Fn (x) = n1 ni=1 1[0,x] Xi , and X n = n1 ni=1 Xi , and then considering the empirical Lorenz curve, say Z p 1 ln (p) = Fn−1 (s)ds, Xn 0 and its dual mn (p) = 1 Xn Z 1 Fn−1 (s)ds, 1−p as follows 1 mn (p) − ln (p) dp mn (p) 0 Z 1 := 1 − Gn (p)dp Z bn = D (2.1) 0 where we set Gn (p) = ln (p)/mn (p). Then, the second estimator is defined in terms of the order statistics X1:n ≤ · · · ≤ Xn:n of the i.i.d. sample X1 , X2 , ..., Xn drawn from X, therefore we define Pi n X X 1 en = 1 − Pn k=1 k:n D n i=1 k=n−i+1 Xk:n n 1X := 1 − Gn (i/n) n i=1 (2.2) where Gn (i/n) expresses the ratio between the mean income of the poorest i and of the richest i elements in the sample. b n and D e n are asympWe will show later, in Theorem 3.5, that the two estimators D b n is suitable for developing inferential results, totically equivalent. While the estimator D e n is much simpler when it comes to implement code for the analysis of real data. D 3 Inferential results In this Section we will present our main results, starting from the consistency of the b n , next we state its asymptotic normal distribution, and then we deal with estimator D its variance estimation. Finally, we will show the asymptotic equivalence of the two e n and D b n. estimators D 4 Unless explicitly stated otherwise, we assume throughout that the cdf F (x) of X is a continuous function. This is a natural choice when modeling income or wealth distributions, and for many other economic size distributions. 3.1 bn Consistency of D b n is a consistent estimator for DF . Theorem 3.1. D Proof. From the normalized definition of the empirical Lorenz curve and its dual, say ln (p) and mn (p), it is useful here to introduce their absolute versions, given by Ln (p) = R p −1 R1 b n as Fn (s)ds and Mn (p) = 1−p Fn−1 (s)ds. We may rephrase D 0 Z bn = D 0 1 Mn (p) − Ln (p) dp Mn (p) (3.1) For all p ∈ [0, 1], we have that Ln (p) converges, with probability 1, uniformly to Rp L(p) = 0 F −1 (s)ds (see Goldie, 1977). With the same approach, we have that Mn (p) R1 converges, with probability 1, uniformly to M (p) = 1−p F −1 (s)ds. As L(p) ≤ M (p), due to the Lebesgue dominated convergence theorem we get the thesis. 3.2 bn Asympthotic normality of the estimator D Theorem 3.2. If the moment E\|X\|2+δ is finite for some δ > 0, then we have the asymptotic representation n √ 1 X b n Dn − D = √ h(Xi ) + oP (1), n i=1 (3.2) where oP (1) denotes a random variable that converges to 0 in probability when n → ∞, and Z h(Xi ) = +∞ 1[0,x] (Xi ) − F (x) ω F (x) dx 0 with the weight function ω(t) = ω1 (t) + ω2 (t), where Z t Z 1 1 L(1 − s) ω1 (t) = ds and ω2 (t) = 2 ds. t M (s) 0 [M (1 − s)] (3.3) b n is asymptotically Corollary 3.3. Under the conditions of Theorem 3.2, we have that D normally distributed, that is √ b n − D =⇒ N (0, σ 2 ), n D F n→∞ 5 where σF2 ∞ Z Z = 0 y F (x) ω F (x) dx 1 − F (y) ω F (y) dy. 0 The proof follows immediately from (3.2) by applying the Central Limit Theorem of P. Lévy. Proof of Theorem 3.2 b n and D, we get From the definition of D Z 1 √ √ Ln (t) L(t) b n Dn − D = − n − dt Mn (t) M (t) 0 Z 1 Z 1 √ √ Ln (t) − L(t) L(t) [Mn (t) − M (t)] =− n dt + n dt + Rn (t) (3.4) M (t) [M (t)]2 0 0 (1) (2) where the remainder term is given by Rn (t) = Rn + Rn and Z 1 √ 1 1 (1) Ln (t) − L(t) − dt, Rn = n M (t) Mn (t) 0 Z 1 √ L(t) 1 1 (2) Rn = n − Mn (t) − M (t) dt. Mn (t) M (t) 0 M (t) (1) (3.5) (3.6) (2) We will later show (Lemma 4.1 and 4.2, respectively) that Rn and Rn are of order oP (1). The proof follows the approach of Greselin, Pasquazzi and Zitikis (2010), to state the asymptotic normality for the Zenga inequality index (Zenga, 2007). Hence we now proceed our analysis of the first two terms in (3.4), by using the Vervaat process Z p Z F −1 (p) −1 −1 Vn (p) = Fn (t) − F (t) dt + Fn (x) − F (x) dx 0 (3.7) 0 and its dual, Vn∗ (p) Z = 1 Fn−1 (t) −F −1 (t) dt + Z ∞ Fn (x) − F (x) dx (3.8) F −1 (p) p for which we know that Vn∗ (p) = −Vn (p). For mathematical and historical details on the Vervaat process, see Zitikis (1998), Davydov and Zitikis (2004), and Greselin et al. (2009). √ Now, denoting the uniform on [0, 1] empirical process by en (p) = n(Fn (F −1 (p)) − p) and using one property of the Vervaat process, namely √ nVn (p) ≤ \|en (p)\|\|Fn−1 (p) − F −1 (p)\|, (3.9) we find a bound for the first term in (3.4) as follows "Z −1 # Z 1 Z 1 F (t) √ √ Ln (t) − L(t) 1 − n dt = n Fn (x) − F (x) dx dt + O(Rn(3) ) M (t) 0 0 M (t) 0 6 where Rn(3) √ =− n 1 Z 0 1 Z 1 Vn (t) dt M (t) 1 \|en (t)\| Fn−1 (t) − F −1 (t) dt. M (t) ≤ 0 (3) We will later show (Lemma 4.3) that Rn = oP (1). Now we deal with the second term in (3.4), and we obtain, using similar arguments as before √ 1 Z n 0 L(t) [Mn (t) − M (t)] dt [M (t)]2 Z +∞ Z 1 √ L(t) =− n Fn (x) − F (x) dx dt + O(Rn4 ), 2 −1 [M (t)] 0 F (1−t) where Rn(4) = √ Z 1 n Vn∗ (1 − t) 0 Z L(t) dt [M (t)]2 1 \|en (t)\| Fn−1 (t) − F −1 (t) ≤ 0 L(t) dt. [M (t)]2 (4) In Lemma 4.4 we show that Rn = oP (1), therefore we have "Z −1 # √ Z 1 1 F (t) √ bn − D = n n D Fn (x) − F (x) dx dt 0 M (t) 0 Z +∞ Z 1 √ L(t) − n Fn (x) − F (x) dx dt + oP (1). 2 0 [M (t)] F −1 (1−t) (3.10) We notice that the first term in the right hand side of equation (3.10) could be rewritten as √ n Z 0 1 1 M (t) "Z # Fn (x) − F (x) dx dt F −1 (t) 0 √ Z 1 Z +∞ Fn (x) − F (x) 1[0,F −1 (t)] (x)dx 1 dt M (t) 0 0 Z +∞ Z 1 √ 1 = n Fn (x) − F (x) 1[F (x),1] (t) dt dx M (t) 0 0 Z 1 Z +∞ √ 1 = n Fn (x) − F (x) dt dx 0 F (x) M (t) n 1 X =√ h1 (Xi ) n i=1 = n 7 where +∞ Z h1 (Xi ) = 1[0,x] (Xi ) − F (x) ω1 F (x) dx, 0 and 1 Z ω1 (t) = t 1 ds. M (s) For the second term in the right hand side of equation (3.10), using the change of variable t = 1 − s, we obtain: √ − n Z 0 1 L(1 − s) [M (1 − s)]2 +∞ Z F −1 (s) where Z n 1 X Fn (x) − F (x) dx ds = − √ h2 (Xi ), n i=1 +∞ h2 (Xi ) = 1[0,x] (Xi ) − F (x) ω2 F (x) dx, 0 and t L(1 − s) ω2 (t) = 2 ds = 0 [M (1 − s)] This completes the proof of Theorem 3.2. Z 3.3 Z 1 1−t L(s) ds. [M (s)]2 Convergence of the empirical variance We deal here with the theoretical variance V ar h(X1 ) , that is Z ∞ Z y 2 F (x) ω F (x) dx 1 − F (y) ω F (y) dy σF = 0 and its empirical counterpart Z ∞ Z y 2 Fn (x) ω Fn (y) dx 1 − Fn (y) ω Fn (y) dy. σn = 0 (3.11) 0 (3.12) 0 Let x0 ≥ 0 be the minimum value in the support of F (x), i.e. the value such that F (y) = 0 for y < x0 and F (y) > 0 if y > x0 . Analogously, let T0 be maximum value in the support of F (x), i.e. such that F (x) < 1 ∀x < T0 and F (T0 ) = 1. Notice that we may have T0 = +∞. Then we have • Fn (x) = 0 ∀x < x0 because X ≥ x0 a.s., • σF2 = R∞ • σn2 = R∞ x0 x0 [. . .] . . . dy, [. . .] . . . dy. Therefore, without loss of generality, we can take x0 = 0. 8 Theorem 3.4. Assume that E\|X\|2+δ < +∞ for some δ > 0. Then, we have a.s. σn2 −→ σF2 . (3.13) n→∞ Proof. The proof is composed by three steps. Step 1 : For all , T such that 0 < < T < T0 , we will show that, with probability 1, for almost all y ∈ [, T ], and with ω(t) = ω1 (t) + ω2 (t) given by (3.3) we have Z y Z y Fn (x) ω Fn (y) dx −−−→ F (x) ω(F (x)) dx. n→∞ 0 (3.14) 0 We begin by the study of the first part of (3.14) related to ω1 (t), i.e. Z yZ 1 Z y Z yZ 1 1 Fn (x) ω1 (Fn (x)) dx = 1[Fn (x),1] (s)Fn (s) ψn (x, s) ds dx. ds dx := Mn (s) 0 0 0 0 0 We know that • as Fn (x) → F (x) a.s. (uniformly), we have the convergence 1[Fn (x),1] (s) → 1[F (x),1] (s) with probability 1 for almost all s ∈ [0, 1] and x ∈ R+ , • ∀s ∈ [0, 1] and with probability 1, we have that Z 1 Z −1 Mn (s) = Fn (y)dy −−−→ n→∞ 1−s • Mn (s) ≥ sX n 1 F −1 (y)dy = M (s), 1−s ∀s ∈ [0, 1]. As X n → µF a.s., with probability 1 there exists a constant C > 0 such that ψn (x, s) ≤ 1[Fn (x),1] (s)Fn (x) 1 sX n ≤ C. Hence, Lebesgue theorem gives (3.14), with ω replaced by ω1 . Now we consider the second part of (3.14), where ω2 (t) takes the place of ω: Z y Z yZ 1 Z yZ 1 Ln (1 − s) Fn (x) ω2 (Fn (x)) dx = 1[0,Fn (x)] (s)Fn (x) ds dx := ψen (x, s) ds dx [Mn (s)]2 0 0 0 0 0 and observe that • 1[0,Fn (x)] (s) → 1[0,F( x)] (s) with probability 1 for almost all s ∈ [0, 1] and x ∈ R+ , 9 • ∀s ∈ [0, 1], with probability 1, we have that Ln (1 − s) → L(1 − s), and Mn (1 − s) → M (1 − s), • Ln (1 − s) ≤ M (1 − s) ∀s ∈ [0, 1]. Therefore ψen (x, s) ≤ 1[0,Fn (x)] (s)Fn (x) ≤ 1[0,Fn (x)] (s)Fn (x) ≤ Fn (x) 1 − Fn (x) X n 1 Mn (1 − s) 1 (1 − s)X n 1 ≤ 1 − Fn (T ) X n Once more using Lebesgue theorem we get ∀y ≤ T Z y Z a.s. Fn (x) ω2 (Fn (x)) dx −→ 0 → 1 < +∞. 1 − F (T ) µF y F (x) ω2 (F (x)) dx 0 which completes the proof of (3.14). Step 2 : For all , T such that 0 < < T < T0 and given Z T 2 σn [, T ] = Ψn (y)dy where y Z Fn (x)ω Fn (x) dx Ψn (y) = 1 − Fn (y) ω Fn (y) , 0 we will show that, with probability 1, σn2 [, T ] −−−→ σ 2 [, T ]. (3.15) n→∞ Due to the previous step, for every y we know that Ψn (y) converges, with probability 1, to Z y Ψ(y) = F (x) ω F (x) dx 1 − F (y) ω F (y) . 0 We have shown that, with probability 1, lim sup sup Fn (x) ω Fn (x) ≤ 1 + n ≤x≤T and using ω1 Fn (y) = Z 1 Fn (y) 1 1 − F (T ) 1 1 ds ≤ , Mn (s) Fn (y)X n 10 1 , µF it follows that we have a.s. sup ω1 Fn (y) lim sup n ≤ ≤y≤T 1 . F ()µF Hence lim sup n sup Ψn (y) ≤ y 1 + ≤x≤T 1 1 − F (T ) 1 1 + F () 1 − F (T ) 1 . µ2F (3.16) Observing now that the function at the right hand side in (3.16) is integrable on [, T ], we can apply Lebesgue’s dominated convergence theorem and prove (3.15). Step 3 : To complete the proof of Theorem 3.4 we need to obtain a bound for the integrals σn2 [0, ] and σn2 [T, ∞]. We use the following more delicate estimation of ω(t), for all γ > 0: there exists a positive constant Cγ such that Cγ 1 . γ µF t (1 − t)γ ω(t) ≤ (3.17) Indeed, as L(t) ≤ tµF ≤ M (t), for t ∈ [0, 1], we have Z 1 1 1 1 ω1 (t) ≤ ds = \| ln t\|. µF t s µF and 1 ω2 (t) ≤ µF Z 0 t 1 1 \| ln(1 − t)\|. ds = 1−s µF Hence, for every γ > 0, there exists a constant 0 < Cγ < +∞, depending only on γ, such that for all t ∈ (0, 1) ω1 (t) ≤ Cγ , µF tγ ω2 (t) ≤ which jointly give the estimation (3.17). 11 Cγ , µF (1 − t)γ For γ ≤ 1/2, we have Z Z y 2 Fn (x) ω Fn (x) dx 1 − Fn (y) ω Fn (y) dy σn [0, ] = 0 0 Z "Z y # Cγ Cγ dx 1 − Fn (y) γ dy γ γ 1 − Fn (x) X n Fn (x) 1 − Fn (x) X n 0 0 Z Z y 1−2γ Cγ2 1−γ ≤ Fn (x) dx 1 − Fn (y) Fn (y)−γ dy 2 0 Xn 0 Z 1−2γ Cγ2 dy y Fn1−2γ (y) 1 − Fn (y) ≤ 2 Xn 0 Z Cγ2 ≤ y dy 2 Xn 0 ≤ = Fn (x) 2 Cγ2 2 Fnγ (x) . (3.18) 2 Xn We have similarly σn2 [T, ∞] ≤ ≤ Cγ2 Z 2 Xn Cγ2 2 Xn Z ∞ 1−2γ yFn1−2γ (y) 1 − Fn (y) dy T ∞ 1−2γ y 1 − Fn (y) dy. (3.19) T e and a new random variable Y , e F, e P) Let us introduce a new probability space (Ω, e Y = Xi = 1 . Then taking the values Xi , for i = 1, . . . , n, such that P n e {Y > y} 1 − Fn (y) = P e \|2+p E\|Y y 2+p ) ( n 1 1 X 2+p = 2+p X . y n i=1 i ≤ If p ≤ δ then, due to the strong law of large numbers, n 1 X 2+p X n i=1 i a.s. −→ E\|X\|2+p < +∞, hence, with probability 1, we have lim sup n σn2 [T, ∞] Cγ2 ≤ E(X 2+δ ) µ2F 12 Z T ∞ y (1−(1−2γ)(2+δ)) dy. (3.20) R∞ Observing now that, for γ < δ/[2(2 + δ)], the integral T y (1−(1−2γ)(2+δ)) dy converges, and defining β such that 1 − (1 − 2γ)(2 + δ) = −(1 + β), then (3.20) takes the form lim sup σn2 [T, ∞] n Cγ2 1 ≤ E(X 2+δ ) . 2 µF β Tβ (3.21) Evidently, replacing Fn (x) by F (x) in (3.18) and (3.21), we obtain their theoretical counterparts σF2 [0, ] 2 Cγ2 , ≤ 2 µ2F (3.22) and σF2 [T, ∞] ≤ Cγ2 1 E(X 2+δ ) . 2 µF β Tβ (3.23) Now, collecting the bounds (3.18) and (3.22), the convergence stated in (3.15), and finally bounds (3.21) and (3.23) from the three steps lim sup σn2 − σF2 n ≤ lim sup σn2 [0, ] − σF2 [0, ] + lim sup σn2 [, T ] − σF2 [, T ] + lim sup σn2 [T, ∞] − σF2 [T, ∞] ≤ n C2 2 γ 2 µF n + 2 β Cγ2 µ2F E(X 2+δ ) n 1 . Tβ (3.24) Taking → 0 and T → ∞ in (3.24), we arrive at (3.13). b n, Having established the consistency and asymptotic normality for the estimator D e n defined in (2.2). we would like to prove similar properties for the second estimator D bn − D e n and prove its asymptotic To do this, we will focus on their difference ∆n := D negligibility. Theorem 3.5. If the moment E\|X\|2+δ is finite for some δ > 0, then we have √ p n\|∆n \| −→ 0. (3.25) Before proving Theorem 3.5, it is worth to state two useful Corollaries. Corollary 3.6. If the moment E\|X\|2+δ is finite for some δ > 0, then we have √ e n − DF ) =⇒ N (0, σF2 ) n(D where σF2 = V ar h(X1 ) is the theoretical variance. 13 (3.26) Corollary 3.7. Under the same assumptions, we have also e n − DF ) √ (D n =⇒ N (0, 1) σn (3.27) where σn2 is the empirical counterpart for σF2 given by (3.12). e n by D b n. The same is true if we replace D Proof. (of Theorem 3.5). Let := n = ∼ n−α−1/2 , where 0 < α < mn n δ . 2(2+δ) We have Z 1 mn n 1 X 1X Gn (t)dt − Gn (t)dt − Gn (i/n) + Gn (i/n) \|∆n \| ≤ n i=1 n i=m +1 0 n Z i/n n X ≤ 2 + [Gn (t) − Gn (i/n)] dt Z i=mn +1 (i−1)/n [,1] ≤ 2 + ωGn (1/n) (3.28) [,1] where ωGn is the modulus of continuity of Gn on the interval [, 1] given by [,1] ωGn (h) = \|Gn (t) − Gn (s)\|. sup ≤t, s≤1, \|t−s\|≤h Let t ∈ [(i − 1)/n, i/n], then Ln (t) Ln (i/n) − Mn (t) Mn (i/n) \|Gn (t) − Gn (i/n)\| = 1 Ln (i/n) \|Ln (t) − Ln (i/n)\| + \|Mn (i/n) − Mn (t)\| Mn (t) Mn (t)Mn (i/n) ≤ 2 ≤ t Xn [,1] ωLn (1/n), (3.29) where we used Mn (s) = 1 − Ln (1 − s) and the inequalities Ln (s) ≤ Mn (s), s X n ≤ Mn (s) that hold true ∀s ∈ [0, 1]. From the bounds in (3.28) and (3.29) we get \|∆n \| ≤ 2 + 2 X n [,1] ωLn (1/n). (3.30) As for t ∈ [(i − 1)/n, i/n] Z \|Ln (t) − Ln (i/n)\| = i/n Fn−1 (s)ds = Xi:n (i/n − t) ≤ t 1 Xi:n , n we get [,1] ωLn (1/n) ≤ 1 n max Xk:n ≤ mn ≤k≤n 1 Mn n 14 where Mn = max Xk:n . k≤n Therefore, due to (3.30), we obtain √ n\|∆n \| ≤ 2 2Mn . + α n X n n1/2−α As 2/nα → 0 and X n → µF , it is sufficient to state the convergence in probability of Mn /n1/2−α to 0. We have, for t > 0 : Mn P ≥t = P Mn ≥ tn1/2−α ≤ n P X ≥ tn1/2−α 1/2−α n E\|X\|2+δ ≤ n 1/2−α 2+δ −−−→ 0 n→∞ (tn ) as we have chosen α < 4 δ . 2(2+δ) Proofs Lemma 4.1. Under the conditions of Theorem 3.2, we have that Rn(1) = oP (1). (1) Proof. We estimate Rn by splitting the integral in two parts, by choosing ρ ∈ (0, 1) Z 1 √ 1 1 (1) Ln (t) − L(t) − dt Rn = n M (t) Mn (t) 0 Z ρ Z 1 √ √ = n . . . dt + n . . . dt 0 ρ := Rn(1,1) + Rn(1,2) . (4.1) (1,1) We now look for getting a bound for Rn and initially deal with its first part, given by Z ρ 1 √ n Ln (t) − L(t) dt. (4.2) M (t) 0 Provided that t < ρ, we have that M (t) ≥ t F −1 (1 − ρ), and Z ρ √ 1 n Ln (t) − L(t) dt 0 t Z ρ Z F −1 (t) √ 1 Vn (t) − Fn (x) − F (x) dx dt = n 0 t 0 Z ρ Z ρ Z F −1 (t) √ √ 1 1 = n Vn (t)dt − n Fn (x) − F (x) dx dt. 0 t 0 t 0 (4.3) Now we consider the lefthand term in (4.3), and set Kn = sup t∈(0,1) \|en (t)\| , 1/2− t (1 − t)1/2− 15 (4.4) where en (t) = √ n Fn (F −1 (t)) − t . We know that (see (9.2) in Greselin et al. 2010) Kn = OP (1). (4.5) Therefore, employing the inequality in (3.9) related to the Vervaat process, and choosing such that 0 < < 21 , we obtain Z ρ Z ρ √ 1 1 n Vn (t)dt ≤ \|en (t)\|\|Fn−1 (t) − F −1 (t)\| dt t t 0 0 Z ρ 1 1/2− t (1 − t)1/2− Fn−1 (t) − F −1 (t) dt ≤ Kn t Z0 ρ 1 ≤ Kn Fn−1 (t) − F −1 (t) dt. 1/2+ 0 t a.s. As \|Fn−1 (t)−F −1 (t)\| ≤ Fn−1 (ρ)+F −1 (ρ) −→ 2 F −1 (ρ), by Lebesgue dominated convergence theorem the integral Z ρ 1 a.s. Fn−1 (t) − F −1 (t) dt −→ 0 as n → ∞. t1/2+ 0 Hence √ Z ρ n 0 1 Vn (t)dt = oP (1). t The righthand term in (4.3) may be estimated as follows Z 0 ρ 1 t F −1 (t) Z √ n Fn (x) − F (x) dx dt 0 Z Z F −1 (t) 1/2− 1 1/2− F (x) 1 − F (x) dx dt t 0 Z +∞ 1/2− 1 1/2− t 1 − F (x) dx dt t 0 ρ ≤ Kn 0 Z ρ ≤ Kn 0 Z ≤ C1 Kn ρ t−1/2− dt = C1 Kn φ(ρ) (4.6) 0 Z for all 0 < < 1/2, where we set φ(ρ) = ρ t−1/2− dt, and C1 = 0 R +∞ 0 1/2− 1−F (x) dx < +∞. The latter quantity is finite, due to the existence of the 2 + δ moment of X. (1,1) To complete the analysis of Rn we have to deal now with its second part, given by Z ρ 1 √ Ln (t) − L(t) n dt. (4.7) Mn (t) 0 a.s. As Mn (t) ≥ tFn−1 (1 − ρ) for t ∈ [0, 1] and Fn−1 (1 − ρ) −→ F −1 (1 − ρ), the bound for (4.7) can be found by following the same steps as for (4.2). 16 (1,2) We continue our proof now by finding a bound for the second term Rn ρ ≤ t ≤ 1, then Z in (4.1). As 1 F −1 (s)ds > 0 M (t) ≥ 1−ρ and 1 Z Fn−1 (s)ds Mn (t) ≥ Z ≥ 1−ρ Therefore, setting Hn := F −1 (s)ds + oP (1). 1−ρ 1 = OP (1), we have Mn (s) M (s) sup s∈[1−ρ,1] √ 1 1 Z Ln (t) − L(t) Mn (t) − M (t) dt n ρ √ 1 Z ≤ Hn n Ln (t) − L(t) Mn (t) − M (t) dt. (4.8) ρ We observe that (4.8) is oP (1) if the following two equalities hold true: Z 1 √ n X n − µF Ln (t) − L(t) dt = oP (1), (4.9) ρ and √ Z 1 Ln (t) − L(t) n Ln (1 − t) − L(1 − t) dt = oP (1), (4.10) ρ by recalling that Mn (t) − M (t) = (X n − µF ) − [Ln (1 − t) − L(1 − t)] due to Z 1−t −1 Fn (s) − F −1 (s) ds + X n − µF = Z 0 1 Fn−1 (s) − F −1 (s) ds. 1−t To get (4.9), remark that √ Z n X n − µF = OP (1), and 1 Ln (t) − L(t) dt ρ Z 1 Z ≤ Z0 1 ≤ t Fn−1 (s) − F −1 (s) ds dt 0 Fn−1 (s) − F −1 (s) ds = oP (1). 0 Finally, to get (4.10), we begin with the inequality Z 1 √ n Ln (t) − L(t) Ln (1 − t) − L(1 − t) dt ρ ≤ √ Z n 1 2 Ln (t) − L(t) dt, ρ 17 (4.11) and use the following bound for the latter integrand 2 2 √ √ √ n Ln (t) − L(t) ≤ 2 n Vn (t) + 2 n Z F −1 (t) 2 Fn (x) − F (x) dx . 0 Recalling that √ Z n 1 2 Vn (t) dt = oP (1), ρ and exploiting (4.4) we get √ Z F −1 (t) 2 Fn (x) − F (x) dx n 0 1 =√ n Z F −1 (t) √ 2 n Fn (x) − F (x) dx 0 Z F −1 (t) 2 1/2− 1 1/2− ≤ Kn √ F (x) 1 − F (x) dx n 0 Z 1/2− 2 1 +∞ 1 − F (x) dx = oP (1). ≤ Kn √ n 0 Integrating in dt on [ρ, 1] we hence obtain the desired bound. From the previous estimates it follows that ∀ρ : 0 < ρ < 1 we have Rn(1) = Un (ρ) + Tn (ρ), where Un (ρ) = oP (1), Tn (ρ) ≤ C1 Kn φ(ρ), ρ→0 with Kn = OP (1) and φ(ρ) −→ 0. Fixing > 0, let C > 0 be such that P {\|Kn \| > C} < ∀n, and let ρ > 0 be such that for ρ < ρ we have φ(ρ) < /2C1 C. Then, having P Rn(1) > ≤ P {Un (ρ) > /2} + P {Tn (ρ) > /2} , we get, for ρ < ρ , lim sup P {Tn (ρ) > /2} ≤ P \|Kn \| > n 2C1 φ(ρ) ≤ P {\|Kn \| > C} < , (1) which finally gives Rn = oP (1). Lemma 4.2. Under the conditions of Theorem 3.2, we have that Rn(2) = oP (1). 18 (2) Proof. We start from the definition of Rn in (3.6) here recalled for convenience Z 1 √ L(t) 1 1 (2) Rn = n − Mn (t) − M (t) dt. Mn (t) M (t) 0 M (t) Observing that L(t) ≤ M (t) for t ∈ [0, 1] and using (4.11) to rewrite Mn (t) − M (t) , the proof can be established following the proof of Lemma (4.1) with minor modifications. Lemma 4.3. Under the conditions of Theorem 3.2, we have that Rn(3) = oP (1). (3) Proof. We estimate Rn by splitting it in two integrals as follows Z 1 1 (3) \|en (t)\| Fn−1 (t) − F −1 (t) dt Rn = M (t) 0 Z 1 Z 1/2 . . . dt . . . dt + = 0 Rn(3,1) := (3,1) Let us consider the first term Rn 1/2 + Rn(3,2) . (4.12) and observe that 1 1 ≤ −1 1 M (t) F (2) t where we assume that F −1 ( 12 ) > 0, otherwise we may replace F −1 ( 12 ) by F −1 (a) > 0, with a ∈ (0, 1) appropriately chosen. Hence, by choosing ≤ √ en (t) = n Fn F −1 (t) − t , we arrive at Rn(3,1) 1 ≤ F −1 ( 12 ) Z Kn −1 F ( 12 ) Z ≤ 1/2 0 1 t 1/2 and recalling that √ n \|Fn (F −1 (t)) − t\| \|Fn−1 (t) − F −1 (t)\| 1/2− t (1 − t)1/2− dt t1/2− (1 − t)1/2− 1 Fn−1 (t) − F −1 (t) = oP (1). t1/2+ 0 1 , 2 a.s. as Kn = OP (1) and Fn−1 (t) −→ F −1 (t) for t ∈ [0, 1]. (3,2) Now we deal with Rn , i.e. the second term in (4.12). Observing that M (t) = R1 R 1 −1 −1 F (s)ds ≥ F (s)ds = c > 0, we obtain 1−t 1/2 Rn(3,2) 1 ≤ c Z 1 \|en (t)\| Fn−1 (t) − F −1 (t) dt 1/2 Kn ≤ c Z 1 Fn−1 (t) − F −1 (t) dt = oP (1). 1/2 19 Lemma 4.4. Under the conditions of Theorem 3.2, we have that Rn(4) = oP (1). (4) Proof. We will deal with Rn , as for the previous Lemma, by splitting it as follows Z 1 L(t) (4) \|en (1 − t)\| Fn−1 (t) − F −1 (t) Rn = dt [M (t)]2 0 Z 1 Z 1/2 . . . dt . . . dt + = 0 Rn(4,1) := 1/2 + Rn(4,2) (4.13) (4,1) . Observing that M (t) ≥ tF −1 (1/2) for t < 1/2, we have Z Kn 21/2− 1 ≤ /2t−1/2− Fn−1 (t) − F −1 (t) dt = oP (1). F −1 (1/2) 0 and we initially consider Rn Rn(4,1) (4,2) Finally, the result on Rn comes from observing that for t ∈ (1/2, 1) there exists a constant C such that M (t) ≥ C, and that supt∈(0,1) en (1 − t) ≤ Kn , we have Rn(4,2) Kn ≤ C Z 1 Fn−1 (t) − F −1 (t) dt = oP (1), 1/2 due to the assumption on the second (hence the first) moment finite on X. 5 The new inequality measure on real data The purpose of this section is to show, through a real data application, the theoretical results obtained in the previous sections. We employ the Bank of Italy Survey on Household Income and Wealth (hereafter named by its acronym, SHIW) dataset, published in 2016. This survey contains information on household post-tax income and wealth in the year 2014, covering 8,156 households, and 19,366 individuals. The sample is representative of the Italian population, which is composed of about 24,7 million households and 60,8 million individuals. The SHIW provides information on each individual’s Personal Income Tax net income, but does not contain the corresponding gross income. We employ an updated version of the microsimulation model described in Morini and Pellegrino (2016) to estimate the latter for each taxpayer. A comparison of the results from the microsimulation model with the official statistics published by the Italian Ministry of Finance (2016) shows that the distribution of gross income and of net tax, according to bands of gross income and type of employment, are close to each other. The empirical 20 analysis we develop here is based on the observed net income from the SHIV, while tax data and gross income arise from the microsimulation model. e n , we To appreciate the asymptotic results of Section 3 on the empirical estimator D calculate four types of confidence intervals: the normal, the basic, the percentile and the BCa confidence intervals. After drawing the bootstrap samples, the empirical estimator is evaluated at each sample, and Figure 5.1 show the histograms of the obtained values when considering Gross Income in panel (a), Net Income in panel (b) and Taxes in panel (c). While inequality estimators have a skewed distribution in case of low sample size, here the accuracy of the normal approximation is apparent, due to the large sample size. As a further check of the quality of the first order approximation, Figure 5.2 shows the 0.655 0.660 0.665 0.670 0.675 0.680 0.685 0.690 1400 1000 400 0 0 0 200 200 400 500 600 600 800 800 1000 1000 1200 1200 1400 1500 Q-Q plots obtained for the three cases. 0.615 0.620 (a) 0.625 0.630 0.635 0.640 0.645 0.795 0.800 (b) 0.805 0.810 0.815 0.820 0.825 (c) e n on Gross Income in panel (a), on Net Income in panel (b) and on Taxes Figure 5.1: Histograms for D −2 0 2 4 Theoretical Quantiles (a) 0.825 0.820 0.815 0.810 Sample Quantiles 0.800 0.795 0.620 0.615 0.660 0.655 −4 0.805 0.635 0.630 Sample Quantiles 0.625 0.675 0.670 0.665 Sample Quantiles 0.680 0.640 0.685 0.645 in panel (c) −4 −2 0 Theoretical Quantiles (b) 2 4 −4 −2 0 2 4 Theoretical Quantiles (c) e n vs Gaussian quantiles on Gross Income in panel (a), on Net Income in Figure 5.2: Q-Q plots for D panel (b) and on Taxes in panel (c) 21 0.85 0.80 0.75 0.60 0.65 0.70 Normal CI Basic CI Percentile CI BCa CI Gross Income Net Income Taxes e n on Gross Income, on Net Income and on Taxes Figure 5.3: Confidence Intervals for D Finally, from Figure 5.3 we observe that the four methods for constructing Confidence Intervals have a substantial agreement. They all agree in assuring that there is a statistically significant increase in inequality when passing from Net Income to Gross Income (the redistributive effect of taxation) and from Gross Income to Taxes. We recall that adjusted percentile methods (also named BCa) for calculating confidence limits are inherently more accurate than the basic bootstrap and percentile methods (Davison and Hinkley, 1997). 6 Concluding remarks Moved from the considerations that nowadays, in many developed countries, the more critical (i.e the extremes) portions of the population are facing great reshaping of their economic situation, a new index for measuring inequality have been proposed in Davydov and Greselin (2016). In the cited paper, a discussion of the properties of the index has been given, to motivate its introduction in the literature and to show its descriptive features. Inferential results for the index were still missing, and this paper is a first contribution to fill the gap. After proposing two empirical estimators, we have shown their asymptotic 22 equivalence. Then, consistency and asymptotic normality for the first estimator have been derived. We also proved the convergence of the empirical estimator for the variance to its finite theoretical value. Finally, we used the new statistical inferential results to analyze data on Net Income from the Bank of Italy Survey on Household Income and Wealth, and to compare them with Gross Income and Taxes. References Atkinson A. B. 1970. On the Measurement of Inequality, Journal of Economic Theory, 2, 244–263. Cobham A., Sumner A. 2013. Putting the Gini back in the bottle? The Palma as a Policy-Relevant Measure of Inequality, Technical report, Kings College London. Davydov Yu., Zitikis R. 2004. Convex rearrangements of random elements in Asymptotic Methods in Stochastics, vol. 44 of Fields Institute Communications, pp. 141–171, American Mathematical Society, Providence, RI, USA. Davydov Yu., Greselin F. 2016. Comparisons between poorest and richest to measure inequality, (under revision) Davison A. C., Hinkley D. V. 1997. Bootstrap methods and their application (Vol. 1). Cambridge University Press. Gastwirth J.L. 2016. Measures of Economic Inequality Focusing on the Status of the Lower and Middle Income Groups, Statistics and Public Policy, 3:1, 1-9. Gastwirth J.L. 2014. Median-based measures of inequality: reassessing the increase in income inequality in the U.S. and Sweden, Journal of the International Association for Official Statistics, 30, 311–320. Gini C. 1914. Sulla misura della concentrazione e della variabilità dei caratteri, Atti del Reale Istituto Veneto di Scienze, Lettere ed Arti, 73, 1203–1248. (English translation in Gini, C. 2005. On the measurement of concentration and variability of characters. Metron, 63, 3–38.) Greselin F. 2014. More equal and poorer, or richer but more unequal?, Economic Quality Control, 29, 2, 99–117. Greselin F., Pasquazzi L., Zitikis R. 2010. Zenga’s new index of economic inequality, its estimation, and an analysis of incomes in Italy, Journal of Probability and Statistics, 26 pp., DOI:10.1155/2010/718905. 23 Greselin F., Pasquazzi L., Zitikis R. 2012. Contrasting the Gini and Zenga indices of economic inequality, Journal of Applied Statistics, 40, 2, 282–297. Greselin F., Puri M. L., Zitikis R. 2009. L-functions, processes, and statistics in measuring economic inequality and actuarial risks, Statistics and Its Interface, 2, 2, pp. 227–245. Goldie C. M. 1977. Convergence theorems for empirical Lorenz curves and their inverses Advances in Applied Probability, 9, 765–791. Lorenz M.C. 1905. Methods of measuring the concentration of wealth, Journal of the American Statistical Association, 9, 209–219. Morini M., Pellegrino S. 2016. Personal income tax reforms: A genetic algorithm ap- proach, European Journal of Operational Research, first online, https://doi-org.proxy.unimib.it/10.1016/ Pietra G. 1915. Delle relazioni tra gli indici di variabilità. Note I. Atti del Reale Istituto Veneto di Scienze, Lettere ed Arti, 74, 775–792 Pietra G. 2014. On the relationships between variability indices. Note I. Metron, 72, 5–16 Zenga M. 2007. Inequality curve and inequality index based on the ratios between lower and upper arithmetic means, Statistica & Applicazioni, 5, 3–27. Zitikis R. 1998. The Vervaat process, in Szyszkowicz, B. (ed.), Asymptotic Methods in Probability and Statistics, Elsevier Science, Amsterdam, pp. 667-694. 24	10math.ST
1 Scalable Co-Optimization of Morphology and Control in Embodied Machines Nick Cheney1,2∗ , Josh Bongard3 , Vytas SunSpiral4 , and Hod Lipson5 Department of Computational Biology and Biological Statistics, Cornell University, Ithaca, NY 2 Department of Computer Science, University of Wyoming, Laramie, WY 3 Department of Computer Science, University of Vermont, Burlington, VT Intelligent Robotics Group, Intelligent Systems Division, NASA Ames/SGT Inc., Mountain View, CA 5 Department of Mechanical Engineering, Columbia University, New York, NY ∗ nac93@cornell.edu 1 arXiv:1706.06133v2 [cs.AI] 12 Dec 2017 4 Abstract—Evolution sculpts both the body plans and nervous systems of agents together over time. In contrast, in AI and robotics, a robot’s body plan is usually designed by hand, and control policies are then optimized for that fixed design. The task of simultaneously co-optimizing the morphology and controller of an embodied robot has remained a challenge. In psychology, the theory of embodied cognition posits that behavior arises from a close coupling between body plan and sensorimotor control, which suggests why co-optimizing these two subsystems is so difficult: most evolutionary changes to morphology tend to adversely impact sensorimotor control, leading to an overall decrease in behavioral performance. Here, we further examine this hypothesis and demonstrate a technique for “morphological innovation protection”, which temporarily reduces selection pressure on recently morphologically-changed individuals, thus enabling evolution some time to “readapt” to the new morphology with subsequent control policy mutations. We show the potential for this method to avoid local optima and converge to similar highly fit morphologies across widely varying initial conditions, while sustaining fitness improvements further into optimization. While this technique is admittedly only the first of many steps that must be taken to achieve scalable optimization of embodied machines, we hope that theoretical insight into the cause of evolutionary stagnation in current methods will help to enable the automation of robot design and behavioral training – while simultaneously providing a testbed to investigate the theory of embodied cognition. I. I NTRODUCTION Designing agile, autonomous machines has been a longstanding challenge in the field of robotics [31]. Animals, including humans, have served as examples of inspiration for many researchers, who meticulously and painstakingly attempt to reverse engineer the biological organisms that navigate even the most dynamic, rugged, and unpredictable environments with relative ease [11], [32], [38]. However, another competing approach is the use of evolutionary algorithms to search for robotic designs and behaviors without presupposing what those designs and behaviors may be. These methods often take inspiration from the evolutionary method itself, rather than the exact specifications of any given organism produced by it. The use of an evolutionary algorithm for automated design comes with many benefits: It removes the costly endeavour of determining which traits of a given organism are specific to its biological niche, and which are useful design features that can provide the same beneficial functions if instantiated in a machine. It can yield machines that do not resemble any animals currently found on earth [23], as it allows for machines that are specialized for behaviors and environments that differ from those of the model organism. Additionally, the optimization process can serve as a controlled and repeatable test-bed for the study of evolutionary, developmental, or behavioral theory [25], [26], [4]. However, the generalization of design automation to include both the optimization of robot neural controllers and body plans has proven to be problematic. While recent successes have demonstrated the potential of effective optimization for the control policies of agents with fixed morphologies [17], [13], [27] or – to a lesser extent – the optimization of morphologies (body plans) for agents with minimal and fixed control policies [12], [10], [1], the co-optimization of the two has seen very limited success [16], [8]. This inability to perform robot brain-body co-optimization at scale has been experienced and noted informally by many researchers involved with robot optimization, yet published rarely. Thus the lack of publication is presumably because the field lacks incentives for the publication of negative results [14], rather than a lack of negative results in unpublished works. Joachimczak et al. provide an anecdotal example of premature convergence in the co-optimization of robot brain and body plan (Fig. 19 of [22]). Cheney et al. [8] further analyze the phenomenon of premature convergence in embodied machines and suggest that traditional evolutionary algorithms are hindered in this setting primarily in their ability to perform continued optimization on the morphology of the robot. They hypothesized that the premature convergence may be due to an effect of embodied cognition, in which an individual’s body plan and brain have an incentive to specialize their behaviors to complement one another. This specialization makes improvements to either subsystem difficult without complementary changes in the other (a highly unlikely event given current algorithms) and thus results in an embodied agent which is fragile with respect to design perturbations. II. BACKGROUND Attempts to solve this problem of fully-automated robot evolution are frequently traced back to the work of Sims [34]. 2 This work introduced the use of evolutionary algorithms to produce goal-directed behaviors and morphologies simultaneously. Despite the advance this work represented, the evolved robots tended to be composed of a small number of components (figures show a mean of 6.042 segments per robot, with each typically controlled by a few neurons). It has been suggested that current computational power should have vastly increased the scale and complexity of robots evolved using Sims’ and similar methods. Yet surveys of evolved robots (e.g. [16], [5]) fail to exhibit any significant increase in size or complexity. A wide range of hypotheses for the lack of scalability have been proposed. Some focused on a lack of efficient evolutionary search algorithms [20], [24], [29] or genetic encodings [21], [6], [10], while others pointed to a lack of incentives for complexity in the simple tasks and environments [1], [7]. Yet attempts to evolve robots using methods designed to overcome these challenges have yet to obviously surpass Sims’ work in terms of complexity and scale. This work investigates a different hypotheses, first suggested in [8], that considers the way in which an agent’s brain and body plan interact during the optimization process. III. M ETHODS A. The Interdependency of Body and Brain We investigate the notion that the specialization of brain and body plan to one another during evolution creates a fragile co-dependent system that is not easily amenable to change. This specialization creates local optima in the search space and premature convergence to suboptimal designs. In this paper, we explore a direct solution to the problem of fragile coupled systems: explicitly readapting one subsystem (e.g. the body plan or the brain) after each evolutionary perturbation to the other. The proposed method differs from a traditional evolutionary algorithm, which evaluates the fitness of a newly proposed variation immediately (i.e. with no readaptation), and uses only this valuation of fitness to determine the long term potential of that variation. As a thought experiment, consider a hypothetical robot with a partially optimized body plan (for example a quadrupedal form) and a partially optimized controller (for example the legs swing forward and back through the sagittal plane). Suppose that this controller has co-adapted to the morphology during evolution such that for each step forward, the robot contracts a muscle near its hip with enough force to swing one of its legs forward with just enough force to land in front of the body and successfully take a step. Consider a variation of this morphology proposed by an evolutionary algorithm in which the new robot possesses longer legs but the controller has not changed. If this machine were being evolved for rapid locomotion, having longer legs and being able to take longer strides would be beneficial and this variation should be more successful than the original design. However, during evaluation of this new robot, the original controller applies the same amount of force to the now-longer leg, failing to move it, and thus frustrating the robot’s ability to walk in a coordinated fashion. Current evolutionary methods would treat this robot as the recipient of a detrimental mutation and remove it from the population. If all such variations display an immediate detrimental fitness impact and are rejected (regardless of their long term potential), then the evolutionary algorithm has converged to a local optima in the search space. However, the newly proposed morphology would have resulted in a robot which outperformed its predecessor, if coupled with a controller suited to that morphology. We can determine that the newly proposed morphology is superior by suppressing mutations to that body plan and allowing readaptation of its controller to properly coordinate behavior. Herein lies the foundation for our proposed algorithm: readapt each controller until the new proposed morphology is more fit than its predecessor. B. Controller Readaptation The most obvious method for modeling controller readaptation would be to protect any lineage that has recently experienced a mutation to the body plan by allowing it to undergo several generations of evolutionary change restricted to the control subsystem. If any member of the lineage achieves higher fitness than the pre-mutation ancestor during that time period, the descendant is retained. Otherwise, the new morphological variant dies out. However, it is unclear how to set the time period for this protection a priori. Surely the amount of time a controller takes to readapt to a new morphology depends on many specific features of the complexity, genetic encoding, desired behavior, and current ability level of the robot (which changes over optimization time). Determining the correct value of this parameter would require a full parameter sweep over various values of readaptation time for each new combination of brain, body, and environment. If our goal is simply to optimize a robot, then searching for this value in each unique optimization scenario is intractable. C. Proposed Method: Morphological Innovation Protection In response to the unintuitive nature of the optimal value for readaptation length, our proposed approach is free of this parameter. Descendants of robots that experience morphological mutations are retained in the population and the number of generations that have elapsed since that mutation occurred are tracked (referred to as the “morphological age” of the robot). If two individuals are found in the population such that the latter robot exhibits better performance on the desired task and has experienced fewer generations since a morphological mutation than the former robot, this latter robot is said to “dominate” the former robot in this mutli-objective optimization (on “age” and “fitness”) – and the former robot is removed from the population. In effect, the latter robot has exhibited an ability to not only recover from its ancestor’s morphological mutation but improve upon it (and others in the population). The concept of tracking age using it as an optimization objective are borrowed from [20], [33]. The major difference here being that the age refers to the length of time that a subsystem of the agent (e.g. the morphology) has remained unvaried, rather 3 than (the original definition of) the total time since a random individual was introduced to the population. This procedure has the effect of “protecting” new morphologies with poorly adapted controllers, and thus we will henceforth refer to this procedure as “morphological innovation protection.” This protection is a form a diversity maintenance, though reduced selection pressure for newly mutated morphologies. Various other methods exist for encouraging diversity (e.g. fitness sharing, crowding, random restart parallel hillclimbers [18], novelty [24], or speciation [36]), however age was chosen for its simplicity of implementation, its parallels to multi-timescale learning in biology, and because it helps to avoid the cost of extended control re-optimization for non-promising morphologies – since the the age-pareto optimization allows fitness comparisons between all new “child” morphologies that have had equal readaptation time, even if they are not yet fully readapted (rather than making comparisons only after a predefined amount of readaptation). D. Evolutionary Algorithm All optimization is performed by a population-based evolutionary algorithm. All trials follow a (µ, λ)-Evolutionary Strategy [3] with µ = 25 parents and λ = 25 mutants for a population size of 50. Trials last for 5000 generations. Crossover was not considered in this work. Mutation had a 50% chance of creating a variation to either the morphology or the controller of a given robot, but not both. Other ratios of morphology:controller mutations were considered (1:99, 20:80, 50:50, 80:20, and 99:1), but none showed a significant effect on resolving the premature convergence and resulting fitness in preliminary trials without innovation protection. E. Genetic Encoding for Soft Robot Morphologies Consistent with prior work studying the co-optimization of robot morphologies [10] and controllers [8], we choose soft robots as our model system due to the open-ended complexity of deformable voxel-based morphologies and distributed controllers. The soft robot morphologies are encoded with a Compositional Pattern Producing Network (CPPN) [35]. The CPPN encoding produces the cell fate of each voxel in the robot through a type of neural network that takes each cell’s geometric location (x, y, z Cartesian coordinates and r radial polar coordinate) and outputs a variety of “morphogens” (in this work, there is one to determine whether a cell is present in that location and one to determine whether a present voxel should be a muscle or a passive tissue cell). Since nearby voxels tend to have similar coordinate inputs, they also tend to produce similar outputs from the network – creating continuous muscle or tissue patches. CPPNs also produce complex geometric patterns, as the activation functions at each node can take on a variety of functions (here: sigmoid, sine, absolute value, negative absolute value, square, negative square, square root, and negative square root). These functions tend to produce regular patterns and features across the coordinate inputs (for example: an absolute value node with an x input would produce left-right symmetry, or a sine node with a y input would produce front-to-back repetition). This network is optimized to produce high performing morphologies by iterating through various proposed perturbations to it. These include the addition or removal of a node, or edge to the network, as well as the mutation of the weight of any edge or the activation function at each node. F. Soft Robot Resolution As the CPPN genetic encoding is a continuous function (mapping the location of a cell to its cell fate) it may be discretized into a phenotype at any resolution (i.e. creating any number of voxels in the morphology, and a unique controller for each voxel), and in practice this resolution is only limited by computational resources (as more elements are more computationally expensive to simulate). In the default lower-resolution treatment, this discretization occurs over a 5 × 5 × 5 space. The higher-resolution robots use phenotypes created at a 10 × 10 × 10 resolution. Note that the distance values are noted in absolute number of voxels, so higher fitness values tend to be produced by phenotypes of higher resolution. G. Controllers and their Genetic Encoding A unique controller is optimized for each muscle cell in the robot’s morphology. For each voxel’s controller, two parameters are optimized (the outputs of a separate CPPN with the same inputs as before). One is the phase offset between each individual cell’s muscle oscillations and that of the global sinusoidal oscillator (which acts as a central pattern generator), and the second is the frequency of this global clock (since CPPNs don’t currently enable global parameters, this is done by averaging the local values at each cell to produce a single global value). All controllers output a value between −1 and 1 at each time step, which corresponds to a linear change in each dimension of a muscle cell (±14% of its original length, or ±48% of its original volume), defining a robot’s behavior. Passive tissue cells remain at their original size (though they also deform based on their intrinsic compliance). While this encoding is simple and straightforward, it has the ability to produce complex behaviors, such as multiple patches of muscle groups that are in sync, counter-sync, or any real valued phase offset from each other. It also has the ability to produce gradually varying sweeps of phase offset, resulting in propagating waves of excitation across large muscle groups. Furthermore, the optimization of the global frequency is able to produce oscillation speeds which are fine tuned to the properties of individual morphologies (such as optimizing to maximize the resonance of soft tissues in appendages). The encoding of the morphology and the controller of the robot into two separate CPPN networks emphasizes the false dichotomy of robot brains and body plans. However, this explicit separation allows us to make changes specific and isolated to either the brain or the body. This is necessary for the proposed algorithm, as controller readaptation requires iterating through controller changes without affecting the morphology. 4 H. Physics Simulation for Evaluation Once the morphology and controller for a given robot are specified, the fitness (locomotion distance) of that robot is determined by constructing and simulating that robot in the VoxCad soft-body physics simulator [19]. Simulations last for 20 actuation cycles (which may be a variable amount of time, depending on the length of the globally optimized frequency – though this method of normalizing for the number of “steps” taken leads to a more fair comparison than normalizing by the amount of time per simulation). I. Morphological Innovation Protection In our newly proposed method we set the “morphological age” to zero for each new “child” that was the result of a morphological mutation to a current individual in the population. This means that for an individual to have a large value in their age objective, that individual must have been the result of a large number of successive controller mutations without any change to that robot’s body plan. This setup thus allows a simple comparison method for individuals who have had similar amounts of controller (re)adaptation to their current morphologies – as a dominated individual would have to have been out-competed in fitness ability by a morphology that is paired with a controller that is less-well adapted to it. This diversity maintenance mechanism encourages the exploration of new peaks in the rugged landscape of brain-body plans – with the implicit assumption that unique morphologies correspond to peaks in this landscape. The method of age resets corresponding to morphological mutations to existing individuals differs from the prior technique (e.g. [20], [33]) of inserting completely random individuals, as it allows the improvement in the fitness levels of age-zero individuals over time. In the case of traditional age-fitness optimization, age-zero individuals are drawn from the same distribution of fitness values for random genotypes regardless of when they are created. However, in the case of morphological innovation protection, age-zero individuals are not random and inherit many of the properties of their parents – meaning that they show higher fitness values over time. To demonstrate this empirically, we performed linear regression on the age-zero individuals from Fig. 2 (containing morphological innovation protection), which showed a significant (p < 0.001) increase in fitness over time (from 18.715 at generation 0 to 22.916 at generation 5000). This confirms a major difference between our proposed technique and the standard approach of age-based diversity maintenance through the introduction of random individuals. J. Statistical Analysis All treatments were performed for 30 independent trials, with random seeds consistent between treatments. All plots show mean values averaged across the most fit individual of 30 trials for each condition with shaded areas representing 95% bootstrapped confidence intervals of this average, and all p-values are generated by a Wilcoxon Rank-Sum Test [37]. IV. R ESULTS A. The Effect of Morphological Innovation Protection on Fitness For the task of locomotion ability (a standard task in evolutionary robotics [5]), we first optimize the robots using the traditional method of “greedy” fitness evaluations for our selection criteria (where immediate locomotion ability determines survival in the population of candidate morphologies – i.e. with no “innovation protection”). In this setup, the traditional method produces robots with an average fitness of 21.717 voxels (with 95% bootstrapped confidence (CI) interval of 19.457 to 22.426 voxels). Additionally, we optimized robots in the same task and environment setup, but this time using “morphological innovation protection” for our selection method – in which individuals can only be out-competed by those with equal or lesser amounts of controller (re)adaptation to their current morphologies. This treatment produces significantly more effective robots (p = 6.067 ∗ 10−6 ), with a mean distance travelled of 31.953 voxels (and 95% CI of 28.157 to 36.511 voxels). The increase in fitness shows that morphological innovation protection is a more effective way of optimizing robots, yet it does not conclusively demonstrate that the intuition of [8] is correct, as that work demonstrated the asymmetric difficulty in optimizing the morphology of a robot (as compared to optimizing its controller) and drew the hypotheses that this because the morphology encapsulated the controller (acting as a translator between the “language” of the “cognitive” functions and the outside environment). While the above experiment does help to support the intuition that the controller must readapt to a new morphological encoding, it also introduces confounding effects, such as the added population diversity afforded by “protection” and the added dimensionality of the search space from its this protection age – moving search from a singleobjective to multi-objective optimization problem. To tease apart the influence of these two confounds, we present a treatment where the controllers of the robot undergo an equivalent protection to which the morphologies did in the above experiment. In this treatment, individuals can only be out-competed by others whose morphology has had equal or lesser amounts of readaptation to their newly mutated controllers – deemed “controller innovation protection”. This condition provides the potential advantages of multi-dimensional search and the added diversity from temporary reductions in selection pressure. Yet it does not rely on the idea of a broken “morphological language.” In this treatment, robots locomote 22.049 voxels on average (95% CI of 20.726 to 22.641 voxels), which fails to show a significant improvement over the singleobjective case of no protection (p = 0.240), and performs significantly worse (p = 1.211 ∗ 10−4 ) than the protection of morphological innovations. The fitness trajectories over evolutionary time (Fig. 1) demonstrate a typical logarithmic fitness improvement over the first 1000 generations or so, but then show a stagnation for the traditional optimization procedure without innovation protection. The mean fitness value of the treatment without protection shows no significant improvement (p = 0.085) 5 turnover of morphologies, with just 10.115 unique robot body plans doing so. Fig. 1. The fitness impact (distance travelled, in voxels) over optimization time (in generations) for various types of brain/body plan protection mechanisms. Values plotted represent the mean value of 30 independent trials, with 95% bootstrapped confidence intervals denoted by colorized regions. from generation 1000 to 5000 (with average fitness values of 20.988 and 21.717 voxels, respectively). Contrary to this, the treatment with morphological innovation protection shows significant improvements over this time (p = 0.013) from 25.925 at generation 1000 to 31.953 voxels at generation 5000. The rapid improvement in the controller innovation protection and no protection cases during the first 1000 generations does not contradict the hypotheses of a fragile “morphological language,” as the coupling between the morphology and controller takes time to become established – and would not introduce fragility into the system before then. B. The Effect of Morphological Innovation Protection on Population Stagnation Perhaps more telling than the average locomotion ability at the end of optimization time is the examination of the optimization process within each individual run. Figs. 2 and 3 represent typical runs, and help to give an intuition of the optimization process. In these figures, each randomly colored line represents a unique morphology, plotted by its locomotion ability over optimization time. Note the continued improvement in performance of the most fit individual over optimization time in the case of morphological innovation protection (Fig. 2) – which is consistent with the trend seen on average in Fig. 1. This is not seen in the case without innovation protection (Fig. 3), where the best individual was found before generation 2000 – and by generation 1000, fitness has reached 99.6% of its final value. Consistent with the above observation, we see that the most fit individual in Fig 2 changes rapidly, continually turning over in the trial with morphological innovation protection. As each color in the figure represents a unique morphology, we notice that a wide variety of different morphologies hold the title of “best-so-far”. On average, in runs with morphological innovation protection, 24.179 unique morphologies are the best-so-far at some point in optimization, where the runs without protection show significantly (p = 1.555 ∗ 10−6 ) less The question of how temporary reduced selection pressure (via the morphological-age dimension) of morphological innovation protection may help to improve overall fitness and continued optimization may be best demonstrated in the pop out box for Fig. 2. Here we see the current best morphology, in teal. This morphology was unable to improve on itself for some time, as we see its fitness value (y-xais) flatlining. This “parent” morphology has a “child,” a new proposed variation of its morphology, highlighted in red. As the original fitness value of this morphology falls below its parent, this individual empirically shows worse performance than its parent – and thus would not be considered as a viable solution in a traditional evolutionary method. However, since this new morphology does not have a controller that is well adapted to it (as the controller is specialized for the previous morphology, in teal), morphological innovation protection does not expect this new robot to immediately outperform its parent, and keeps this individual in consideration as one which could hold long term potential but does not show immediate promise. Indeed, we see that after a number of controller optimization iterations later (occurring in equal amounts to both the parent and child during this intermediate period), the child morphology (in red) overtakes the parent morphology (in teal) – achieving higher fitness and demonstrating that it did indeed hold a better long term potential than its parent, despite the immediate drop in fitness. As the fitness of the parent is outperformed by the child (which has had less time to fine-tune its controller to its morphology), we assume that the parent is unlikely to be the most promising robot body plan in the long run, and thus remove it from the population. We see this trend of “overtakes” – where children start out with worse fitness than their parents, but eventually outperform them – continuing throughout this run (as the blue child then overtakes the red parent, and green overtakes blue in the pop out of Fig. 2). We see morphological overtakes significantly (p ≤ 6.939 ∗ 10−10 ) more often in runs with morphological innovation protection (an average of 76.714 overtakes in the first 5000 generations) than without any innovation protection (where there are only 1.432 overtakes) or in the trials with controller innovation protection (where there are just 1.333 morphological overtakes on average). Interestingly, the number of morphological overtakes in the morphological innovation protection treatment is not significantly different (p = 0.533) than the number of controller overtakes in the controller innovation protection treatment (where a “child” is a robot with a new controller variation, that readapts its morphology to catch back up to its “parent” controller – which happens 74.542 times on average). Combined with the finding in Fig. 1 that morphological innovation protection outperforms the other two treatments, this suggests a greater potential for the relative importance of morphological overtakes over controller overtakes – and again reinforces the asymmetry between morphologies and controllers from an optimization perspective. 6 Fig. 2. A single optimization trial featuring morphological innovation protection. Each unique morphology is represented by a random color. The pop out in this figure highlights an example of an “overtake” where the new child morphology (in red) initially performs worse than its parent morphology (in teal), only to outperform that previous morphology after successive control optimization to both the parent and child. Fig. 3. The fitness over time of a single optimization trial with the same initial conditions as Fig. 2, but without any innovation protection. Again, each unique color represents a unique morphology. The prevalence of colored dots filling the space underneath the best performing individual, which represent new morphologies which initially performed worse than the current best individual and were thus rejected and thrown out of the population (as opposed to those individuals which were protected and eventually led to “overtakes” in Fig. 2). pronounced in Fig. 6. C. The Effects of Morphological Innovation Protection on the Progression of Morphologies over Evolutionary Time D. The Potential for Convergence Across Initial Conditions Visual inspection of actual morphologies over evolutionary time further supports the proposed method’s improved optimization efficiency and ability to escape local optima. Fig. 4 shows the current best individual at various points over evolutionary time for each of the top 10 runs in the treatment with morphological innovation protection. Note how the fitness values of the robots increase over time (from left to right, and indicated by the color of each robot). Also note how the final morphology of some robots (e.g. runs 24, 27, 18, 11, and 16) result in identical morphologies, despite starting from a range of starting morphologies and not finding this convergent morphology until hundreds or thousands of generations into the optimization process. In contrast to the sustained turnover of morphologies shown above, Fig. 5 shows the snapshots of the 10 best runs in the treatment without innovation protection. Notice how the colors of the robots tend to show little change over the evolutionary process, mirroring the stagnation shown in Fig. 1. While convergence of the final morphologies is present here as well, the gross morphologies found here (variants of the a full cube with no appendages) are found early on in the search – and often provided in the set of random initial morphologies. In this treatment, gross morphological changes tend to be absent after generation 50 (just 1% into the full 5000 generations). The differences between Figs. 4 and 5 suggests that the traditional method without morphological innovation protection tends to converge prematurely to morphologies early in the evolutionary search, while morphological innovation protection may allow search to escape these local optima and converge to “more global” optima. The convergence of morphologies across varied initial conditions is even further To explore the question of scale, we apply morphological innovation protection to the evolution of robots with higher resolution morphologies – up to 103 = 1000 voxels rather than the lower resolution morphologies (53 = 125 voxels) employed in the previous experiments. The increased number of voxel cells that make up each robot allows for greater expressiveness and finer details in its morphology. However this also presents a challenge for the above algorithm. As the number of cells increases, the effect of changing a single voxel (the minimum morphological variation) is reduced. In the extreme, the concept of readapting controllers since the last “morphological change” is less straightforward – as increasingly small changes can modify minor details of the morphology without affecting its overall function. To help address the problem of non-functional morphology changes, we introduce a parameter to represent the minimum percentage of voxels that must be varied in order to qualify as a “gross morphological change”. It is important to note that this parameter is specific to the voxel-based soft robot implementation employed in this work – and thus the optimal setting of this parameter is not of great importance for its generalization outside of this soft robot encoding. But the general concept of a threshold for the minimum morphological change is a universal concept that could be applied to any robot instantiation, as necessary. In the case of robots with higher-resolution morphologies, we find that resetting the “morphological age” of an individual only after a mutation that changes more than 20% of their voxels produces optimal results. The value was found via a parameter sweep (of minimum age-resets of 0%, 5%, 10%, 7 Fig. 4. The progression of morphologies over evolutionary time with morphological innovation protection. Rows represent the top 10 (out of 30) performing runs at generation 5000, while columns represent snapshots of the morphology at various points during the optimization process. Note how some of the runs converge upon the same morphology (a front and back “legged” robot), despite starting from varying initial conditions. The color of the morphologies represent their fitness values. 20%, 30%, 40%, and 50% of voxels changed). We also investigated a minimum change threshold for controller innovation protection and found that its benefit falls as the threshold increases (showing optimal performance with no threshold), so we ignore the threshold for minimum controller change here. In this setting, basic morphological innovation protection (with no threshold) travels significantly farther (60.728 voxels) than both the case with no protection (32.575 voxels, p = 1.208 ∗ 10−4 ) and the case with controller innovation protection (35.375 voxels, p = 0.018). While trials employing the threshold on morphological changes did achieve higher average fitness than the case of morphological innovation protection without a threshold (where robots traveled 43.340 voxels on average), this difference was not significant at the p = 0.05 level (p = 0.063). As a side note, in other more difficult trials (where the robots need to optimize a closed-loop neural network controller – not depicted visually due to space constraints), trials with the morphological change threshold significantly outperform all other treatments (as they travel 31.313 voxels – while trials with no protection, controller protection, and morphological innovation protection with no threshold travel 13.941, 15.356, and 17.224 voxels, respectively; all p ≤ 7.287 ∗ 10−8 ). Fig. 5. The progression of morphologies over evolutionary time for the top 10 performing runs with no innovation protection. The color legend and column snapshots times are identical to Fig. 4. Visualization of these soft robot morphologies over time in the trials with a 20% morphological change threshold is depicted in Fig. 6 and compared to the case with no protection in Fig. 7. The ability of evolution to converge to the same high-performing morphology across many independent trials, despite starting from different initial conditions suggests that (in this particular soft robot implementation) the inclusion of thresholded morphological innovation protection is able to escape the local optima around these starting conditions and find “more global” optima in this search space. In the case without any protection, search stagnates quickly and again appears unable to escape the local optima near its initial conditions (Fig. 7). Interestingly, the low resolution soft robot implementation (in Figs. 4 and 5) does not benefit from the inclusion of a threshold. As presumably the more discrete construction of these robots mean that all mutations are large enough to create a meaningful “morphological change”. V. D ISCUSSION The above results demonstrate a new method for entire robot evolution (i.e. both brain and body-plan) that is more scalable in terms of continued optimization for longer periods of time and better resulting fitness than the traditional evolutionary methods. This method for “morphological innovation protection” helps prevent premature convergence to the many local 8 Fig. 6. The progression of morphologies over evolutionary time in the setting with high-resolution morphologies and evolution with morphological innovation protection with a 20% threshold. The rows represent the top 10 (of 30) independent trials, while the columns represent the progression over evolutionary time. Color represent the fitness values of the robot (their locomotion speed), with warmer values depicting more fit individuals. Note the convergence of all 10 of these runs to the same final morphology. optima which appear to be present in the rugged search space of robot morphologies and controllers [8]. The hypothesis from [8] that the fragile co-optimization of brain and body plan caused by specialization of one subcomponent to the other is consistent with the findings above. This work also reveals that there is an asymmetry between the brain and body plan: protecting innovations to the morphology leads to more effective optimization that protecting innovations to the controller. The benefits of the temporarily reduced selection pressure provided by morphological innovation protection suggests that the long-term potential and immediate fitness impact of a morphological mutation are not always correlated. Thus, we require a form of diversity maintenance to help evolution to rate proposed solutions based on their long-term potential rather than on immediate fitness impact. As was shown here, using morphological innovation protection for this purpose can help to reduce premature convergence in the search space and stagnation at suboptimal values. We believe this to be the first example of a design automation algorithm for robotics that considers the interdependence of neural controllers and body plans arising from the psychological theory of embodied cognition [30] and uses this intuition to propose a method to escape local optima in the Fig. 7. The progression of high-resolution morphologies over evolutionary time for the top 10 performing runs with no innovation protection. The color legend and column snapshots times are identical to Fig. 6. fitness landscape. Despite the significant improvement to our ability to simultaneously optimize the brain and body plan of embodied robots, there is much work still to be done. Firstly, the proposed method was only applied to one class of robot. This class may actually represent the simplest form of brain-body co-optimization because the distributed sensing, actuation, and information processing that the cellular soft robot paradigm was designed to possess helps to blur the line between physical interactions of the morphology with the environment and information processing of a controller [10], [9]. In the case of centralized controllers and robots composed of rigid components, topological (rather than parametric) changes to the cognitive architecture would be required for control readaptation if morphological mutations add or remove physical components (e.g. limbs). Future work should explore the effect of morphological innovation protection in such a paradigm, where there is the potential for morphological changes to more drastically change the function of the robot – and thus for readaptation to those morphologies to play an even more critical role in optimization. In these experiments, the genotype encoding of the soft robots was modularized such that one part of the genome dictated the shape and material properties of the robot and a separate part encoded the actuations in the form of volumetric deformations of the voxels during behavior. In these robots (and those with even more complex phenotypes), one can 9 envision different splits as to what constitutes morphology and control – and what qualifies as a “morphological change” in this proposed algorithm. It would be of interest to investigate the effect of various splitting points of this dichotomy on the efficacy of morphological innovation protection. Furthermore, the very idea of a distinct split between body and brain is a false dichotomy, as information processing and physical processes happen throughout the agent [2]. Rather than various mutation operators which affect only the “brain” or the “body-plan” of a robot, we desire agents that grown and develop as the result of rich crosstalk and feedback loops between the processes responsible for creating both of these components (creating complex adaptive systems like we see in biological organisms). How this proposed method relates to, and informs, biological analogs not entirely clear, but some rough analogies can be drawn. For example, the observation that brains learn and adapt at a much faster rate than bodies grow fits into this paradigm, as does the even slower change of gross morphological features over evolutionary time. In this way, the readaptation of controller strategies for varying body plans is built into the neuroplasticity of the brain. One could also imagine periods of evolutionary time when an entire species is under relatively little morphological selection pressure before an environmental shock suddenly reapplies that pressure. Also possible are periods of a single individual’s lifetime when selection pressure varies: For example, human infants may not be selected highly for their locomotion speed, as their parents tend to physically carry and protect them early in life. We do not believe that the methods introduced here are restricted to this particular domain. The above algorithm is simple to implement (requiring only: an age counter, a check for variations in brain and/or body for each mutation, and – optionally – a criterion for the minimal gross morphological change), and thus we believe it will be widely applicable. Future work will test this supposition. Due to the recent interest in co-optimization of neural network topology and weights [15], [28], we should also note that this work – an agent’s controller embodied within its morphology – is closely related to that of neural network’s weights embodied within its topology. Future work will show whether the method proposed here will show similar optimization gains in the design of neural network topologies as well. VI. C ONCLUSION We demonstrate an example of a robot design automation algorithm that considers the interdependence of neural controllers and body plans (due to the theory of embodied cognition) on the optimization process. We use this intuition to temporarily reduce selection pressure on newly mutated robot morphologies, thus allowing the agents to readapt their controllers and better escape local optima in the fitness landscape. We have shown that this technique – deemed “morphological innovation protection” – produces evolutionary optimization which delays premature convergence and stagnation, and results in more efficient evolved robots. We showcase the ability of this technique to escape local optima in the search space by demonstrating the convergence to a similar morphology across many independent trials from randomly initial conditions. While we hope that this technique will be surpassed in the future by a developmental process with feedback loops between the body and brain, we propose the above algorithm as a short term improvement over the current techniques for the co-optimization of morphology and control in virtual creatures. VII. ACKNOWLEDGMENT We thank NASA Space Technology Research Fellowship #NNX13AL37H and Army Research Office Contract #W911NF1610304 for support, Steve Strogatz for advice, Kathryn Miller for copy editing, and the UVM Morphology, Evolution & Cognition Lab for feedback. R EFERENCES [1] J. E. Auerbach and J. C. Bongard. Environmental influence on the evolution of morphological complexity in machines. PLoS Comput Biol, 10(1):e1003399, 2014. [2] F. Baluska and M. Levin. On having no head: Cognition throughout biological systems. Frontiers in Psychology, 7:902, 2016. [3] H.-G. Beyer and H.-P. Schwefel. Evolution strategies–a comprehensive introduction. Natural computing, 1(1):3–52, 2002. [4] J. Bongard. Morphological change in machines accelerates the evolution of robust behavior. Proc. of the National Academy of Sciences, 108(4):1234–1239, 2011. [5] J. C. Bongard. Evolutionary robotics. Comm. of the ACM, 56(8):74–83, 2013. [6] J. C. Bongard and R. Pfeifer. Evolving complete agents using artificial ontogeny. In Morpho-functional Machines: The new species, pages 237– 258. Springer Japan, 2003. [7] N. Cheney, J. Bongard, and H. Lipson. Evolving soft robots in tight spaces. In Proceedings of the 2015 Annual Conference on Genetic and Evolutionary Computation, pages 935–942. ACM, 2015. [8] N. Cheney, J. Bongard, V. SunSpiral, and H. Lipson. On the difficulty of co-optimizing morphology and control in evolved virtual creatures. In ALIFE XV, The Fifteenth International Conference on the Synthesis and Simulation of Living Systems, volume 15, pages 226–233. MIT Press. [9] N. Cheney, J. Clune, and H. Lipson. Evolved electrophysiological soft robots. In ALIFE 14: The Fourteenth Conference on the Synthesis and Simulation of Living Systems, volume 14, pages 222–229, 2014. [10] N. Cheney, R. MacCurdy, J. Clune, and H. Lipson. Unshackling evolution: evolving soft robots with multiple materials and a powerful generative encoding. In Proc. of the 15th Genetic and Evol. Comp. Conf., pages 167–174. ACM, 2013. [11] J. Chestnutt, M. Lau, G. Cheung, J. Kuffner, J. Hodgins, and T. Kanade. Footstep planning for the honda asimo humanoid. In Robotics and Automation, 2005. ICRA 2005. Proceedings of the 2005 IEEE International Conference on, pages 629–634. IEEE, 2005. [12] S. Collins, A. Ruina, R. Tedrake, and M. Wisse. Efficient bipedal robots based on passive-dynamic walkers. Science, 307(5712):1082– 1085, 2005. [13] A. Cully, J. Clune, D. Tarapore, and J.-B. Mouret. Robots that can adapt like animals. Nature, 521(7553):503–507, 2015. [14] D. Fanelli. Negative results are disappearing from most disciplines and countries. Scientometrics, 90(3):891–904, 2011. [15] C. Fernando, D. Banarse, C. Blundell, Y. Zwols, D. Ha, A. A. Rusu, A. Pritzel, and D. Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint arXiv:1701.08734, 2017. [16] T. Geijtenbeek and N. Pronost. Interactive character animation using simulated physics: A state-of-the-art review. In Computer Graphics Forum, volume 31, pages 2492–2515. Wiley Online Library, 2012. [17] T. Geijtenbeek, M. van de Panne, and A. F. van der Stappen. Flexible muscle-based locomotion for bipedal creatures. ACM Transactions on Graphics (TOG), 32(6):206, 2013. [18] V. S. Gordon and D. Whitley. Serial and parallel genetic algorithms as function optimizers. In ICGA, pages 177–183, 1993. [19] J. Hiller and H. Lipson. Dynamic simulation of soft multimaterial 3dprinted objects. Soft Robotics, 1(1):88–101, 2014. 10 [20] G. S. Hornby. Alps: the age-layered population structure for reducing the problem of premature convergence. In Proceedings of the 8th annual conference on Genetic and evolutionary computation, pages 815–822. ACM, 2006. [21] G. S. Hornby and J. B. Pollack. Body-brain co-evolution using lsystems as a generative encoding. In Proceedings of the 3rd Annual Conference on Genetic and Evolutionary Computation, pages 868–875. Morgan Kaufmann Publishers Inc., 2001. [22] M. Joachimczak, R. Suzuki, and T. Arita. Artificial metamorphosis: evolutionary design of transforming, soft-bodied robots. Artificial life, 2016. [23] C. G. Langton et al. Artificial life. Addison-Wesley Publishing Company Redwood City, CA, 1989. [24] J. Lehman and K. O. Stanley. Evolving a diversity of virtual creatures through novelty search and local comp. In Proc. of 13th Genetic & Evol. Comp. Conf, pages 211–8. ACM, 2011. [25] R. E. Lenski, C. Ofria, T. C. Collier, and C. Adami. Genome complexity, robustness and genetic interactions in digital organisms. Nature, 400(6745):661–664, 1999. [26] R. E. Lenski, C. Ofria, R. T. Pennock, and C. Adami. The evolutionary origin of complex features. Nature, 423(6936):139–144, 2003. [27] S. Levine, C. Finn, T. Darrell, and P. Abbeel. End-to-end training of deep visuomotor policies. Journal of Machine Learning Research, 17(39):1– 40, 2016. [28] R. Miikkulainen, J. Liang, E. Meyerson, A. Rawal, D. Fink, O. Francon, B. Raju, A. Navruzyan, N. Duffy, and B. Hodjat. Evolving deep neural networks. arXiv preprint arXiv:1703.00548, 2017. [29] J.-B. Mouret and J. Clune. Illuminating search spaces by mapping elites. arXiv preprint arXiv:1504.04909, 2015. [30] R. Pfeifer and J. C. Bongard. How the body shapes the way we think: a new view of intelligence. MIT press, 2006. [31] R. Pfeifer, M. Lungarella, and F. Iida. Self-organization, embodiment, and biologically inspired robotics. science, 318(5853):1088–1093, 2007. [32] M. Raibert, K. Blankespoor, G. Nelson, R. Playter, and T. B. Team. Bigdog, the rough-terrain quadruped robot. In Proceedings of the 17th world congress, volume 17, pages 10822–10825. Proceedings Seoul, Korea, 2008. [33] M. Schmidt and H. Lipson. Age-fitness pareto optimization. In Genetic Programming Theory and Practice VIII, pages 129–146. Springer, 2011. [34] K. Sims. Evolving virtual creatures. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 15– 22. ACM, 1994. [35] K. O. Stanley. Compositional pattern producing networks: A novel abstraction of development. Genetic Programming and Evolvable Machines, 8(2):131–162, 2007. [36] K. O. Stanley and R. Miikkulainen. Evolving neural networks through augmenting topologies. Evolutionary computation, 10(2):99–127, 2002. [37] F. Wilcoxon and R. A. Wilcox. Some rapid approximate statistical procedures. Lederle Laboratories, 1964. [38] R. J. Wood. The first takeoff of a biologically inspired at-scale robotic insect. IEEE transactions on robotics, 24(2):341–347, 2008.	2cs.AI
Criteria for universality of quantum gates Adam Sawicki1 and Katarzyna Karnas1∗ 1 Center for Theoretical Physics PAS, Al. Lotników 32/46, 02-668, Warsaw, Poland arXiv:1610.00547v6 [quant-ph] 17 May 2017 We consider the problem of deciding if a set of quantum one-qudit gates S = {U1 , . . . , Un } is universal. We provide the compact form criteria leading to a simple algorithm that allows deciding universality of any given set of gates in a ﬁnite number of steps. Moreover, for a non-universal S our criteria indicate what type of gates can be added to S to turn it into a universal set. Universal quantum gates play an important role in quantum computing and quantum optics [15, 24, 30]. The ability to effectively manufacture gates operating on many modes, using for example optical networks that couple modes of light [9, 29], is a natural motivation to consider the universality problems not only for qubits but also for higher dimensional systems, i.e. qudits (see also [27, 28] for fermionic linear optics and quantum metrology). For quantum computing with qudits, a universal set of gates consists of all one-qudit gates together with an additional two-qudit gate that does not map separable states onto separable states [10] (see [11, 35–37] for recent results in the context of universal Hamiltonians). The set of all one-qudit gates can be, however, generated using a finite number of gates [23]. We say that one-qudit gates S = {U1 , . . . , Un } ⊂ SU (d) are universal if any gate from SU (d) can be built, with an arbitrary precision, using gates from S. It is known that almost all sets of qudit gates are universal, i.e. non-universal sets S of the given cardinality are of measure zero and can be characterised by vanishing of a finite number of polynomials in the gates entries and their conjugates [17, 23]. Surprisingly, however, these polynomials are not known and it is hard to find operationally simple criteria that decide one-qudit gates universality. Some special cases of optical 3-mode gates have been recently studied in [5, 32] and the approach providing an algorithm for deciding universality of a given set of quantum gates that can be implemented on a quantum automata has been proposed [13] (see also [1, 2, 14] for algorithms deciding if a finitely generated group is infinite). The main obstruction in the problems considered in [5, 32] is the lack of classification of finite subgroups of SU (d) for d > 4. Nevertheless, as we show in this paper one can still provide some reasonable conditions for universality of one-qudit gates without this knowledge. The efficiency of universal sets is typically measured by the number of gates that are needed to approximate other gates with a given precision ǫ. The Solovay-Kitaev theorem states that all universal sets are roughly the same efficient. More precisely, the number of gates needed to approximate any gate U ∈ SU (d) is bounded by O(logc (1/ǫ)) [26], where c may depend only on d and c ≥ 1. Recently there has been a bit of flurry in the ∗ E-mail: a.sawicki@cft.edu.pl, karnas@cft.edu.pl area of single qubit gates [3, 21, 22, 34] showing that using some number theoretic results and conjectures one can construct universal sets with c = 1. The approach presented in these contributions has been unified in [31] where the author pointed out the connection of these new results with the seminal work about distributing points on the sphere S 2 [25] that uses results concerning optimal spectral gap for the averaging operator. Moreover, the authors of [20] showed that the existence of the spectral gap implies c = 1 for all symmetric universal sets of single qudit gates, where by symmetric we mean the set S = {U1 , . . . , Un } with n = 2k and Uk+i = Ui−1 for i ∈ {1, . . . , k}. Although there are still some problems to solve in this area it seems that any further progress would require development of more advanced methods in pure mathematics rather than in quantum information. These developments should include verification of the spectral gap conjecture. Currently it is known to be true under the additional assumption that gates have algebraic entries [6, 7]. In this paper we present an approach that allows to decide universality of S by checking the spectra of the gates and solving some linear equations whose coefficients are polynomial in the entries of the gates and their complex conjugates. Moreover, for non-universal S, our method indicates what type of gates can be added to make S universal. The paper is organised as follows. We start from presenting basic facts concerning the adjoint representation of SU (d). The adjoint representations assigns to every matrix U ∈ SU (d) a matrix AdU ∈ SO(d2 − 1). We give the explicit formula for AdU . The necessary condition for universality (Lemmas 1 and 2) is then formulated using matrices AdU and AdU −1 , where U ∈ S and boils down to checking the dimension of the kernel of the matrix MS given by (10). Next, we assume that the necessary condition for universality is satisfied and provide sufficient conditions for < S > to be infinite and thus dense in SU (d). More precisely, if < S > contains at least one element whose Hilbert-Schmidt distance from d Z(SU √ (d)) = {αI : α = 1} is both nonzero and less than 1/ 2 then < S > is infinite. Combining this with basic results in number theory we arrive at our main results. In Theorem 1 we state that S is universal if S contains at least one matrix whose spectrum does not belong to some finite list of exceptional spectra. We also provide the algorithm which allows deciding universality of any given set of gates S (also when S contains matrices with exceptional spectra) in a finite number of steps. We discuss 2 the correctness of the algorithm and provide instructive examples for S ⊂ SU (2). I. THE NECESSARY CONDITION FOR UNIVERSALITY Let us begin with introducing the basic notation used in this paper and explaining the adjoint representation. The set of gates S = {U1 , . . . , Un } ⊂ SU (d) is called universal if the set generated by elements of S < S >:= {Ui1 · . . . · Uim : Uij ∈ S, m ∈ N} (1) is dense in SU (d), i.e. the closure < S > = SU (d). In fact < S > is always a Lie group [33]. If this this group is SU (d) we will say that S generates SU (d). Let us denote by su(d) the Lie algebra of SU (d). Recall that X ∈ su(d) iff X is an antihermitian traceless matrix. Moreover, the Lie algebra su(d) is a real vector space equipped with a nondegenerate positive inner product defined by (X\|Y ) = − 21 trXY . For U ∈ SU (d) and X ∈ su(d) we define AdU X := U XU −1 . One easily checks that AdU is a linear operator acting on su(d). It is also invertible as (AdU )−1 = AdU −1 and preserves the inner product as 1 (AdU X\|AdU Y ) = − trU XU −1 U Y U −1 = 2 1 = − trXY = (X\|Y ) 2 (2) Therefore AdU is an orthogonal transformation acting on d2 − 1 dimensional vector space su(d). Upon a choice of d2 −1 an orthonormal basis {Xi }i=1 in su(d), i.e. basis that statisfies (Xi \|Xj ) = δij the transformation AdU can be expressed in this basis as a matrix belonging to SO(d2 − 1), i.e AdtU AdU = I and det AdU = 1. The entries of this matrix, (AdU )ij , are real and defined by the identity: AdU Xj = U −1 Xj U = d X (AdU )ij Xi , (3) i=1 thus they are give by 1 (AdU )ij = − tr Xi U Xj U −1 . 2 (4) Ad : SU (d) → SO(d2 − 1), (5) Note that we also have AdU1 U2 = AdU1 AdU2 and this way we obtain the homomorphism that is known as the adjoint representation. For a set of d × d real matrices M , let us denote the set of all d × d matrices commuting with matrices from M by C(M ) = {L : [L, m] = 0, ∀m ∈ M }. (6) The adjoint representation of SU (d) is an absolutely irreducible real representation and therefore by the extended version of Schur’s lemma [8, 19], the only (d2 −1)×(d2 −1) matrix that commutes with all matrices AdSU(d) = {AdU : U ∈ SU (d)} is proportional to the identity matrix, I. In other words C(AdSU(d) ) = {λI : λ ∈ R}. Example 1. The adjoint representation for d = 2, i.e. Ad : SU (2) → SO(3) has a particularly nice form. Any matrix from SU (2) can be written in a form U (φ, ~k) = I cos φ + sin φ(kx X + ky Y + kz Z), (7) where Y = iσ1 , X = iσ2 , Z = iσ3 and σi are Pauli matrices, ~k = (kx , ky , kz )T satisfies kx2 + ky2 + kz2 = 1. Similarly, any matrix from SO(3) has a form O(φ, ~k) = I + sin φ(−kx X12 + ky X13 − kz X23 )+ φ +2 sin2 (−kx X12 + ky X13 − kz X23 )2 , 2 (8) where Xij = Eij − Eji , and Eij is a matrix whose only non vanishing entry is (i, j). One easily verifies that the adjoint representation is given by AdU(φ,~k) = O(2φ, ~k). (9) For U ∈ SU (d), where d > 2 calculation of matrices AdU can be done using formula (4) upon the choice of orthonormal basis in su(d). For d = 3 this basis is given by, for example, the Gell-Mann matrices multiplied by imaginary unit i. For higher d one can construct an orthonormal basis of su(d) in an analogous way as for d = 3. General considerations that can be found in [33] show that the group < S > can be either: 1. < S > = SU (d), or 2. < S > is infinite and connected, or 3. < S > is infinite and consists of k < ∞ connected components, where each component has the same dimension (as a manifold), or 4. < S > is finite. Note that in cases 1, 2, and 3 the group < S > has infinite number of elements. Thus we first provide criteria that distinguish between case 1 and cases 2 and 3. To this end we will use the adjoint representation. For S = {U1 , . . . , Un } ⊂ SU (d) let AdS = {AdU : U ∈ S}. Note that if [L, AdU1 ] = 0 and [L, AdU2 ] = 0 then [L, AdU1 U2 ] = [L, AdU1 ]AdU2 + AdU1 [L, AdU2 ] = 0. Thus if S generates SU (d) and L is a matrix that commutes with AdS then L commutes with AdSU(d) . Therefore for universal S we have C(AdS ) = C(AdSU(d) ) = {λI : λ ∈ R}. It turns out (see [33]) that the converse is true under one additional assumption, namely that < S > is infinite. 3 Lemma 1. For a set of special unitary matrices S = {U1 , . . . , Un } assume that < S > is infinite and C(AdS ) = {λI : λ ∈ R}. Then < S > = SU (d). The proof of this lemma is based on the structure theory for semisimple Lie groups and can be found in [33]. Here we only make some additional remarks regarding calculation of C(AdS ). Let vec(L) be the vectorisation of matrix L, i.e. the vector obtained by stacking the columns of the matrix L on top of one another. One easily calculates that [L, AdU ] = 0 ⇔ (I ⊗ AdU − AdU † ⊗ I) vec(L) = 0, where U † is the complex conjugate and transpose of U , i.e. U † = Ū t . Let   I ⊗ AdU1 − AdU † ⊗ I 1   .. MS =  (10)  . I ⊗ AdUn − AdUn† ⊗ I Lemma 2. C(AdS ) = {λI : λ ∈ R} if and only if the kernel of MS is one-dimensional. We emphasise the role of the adjoint representation which is crucial in Lemma 1. In particular there are infinite subgroups < S > such that C(S) = C(SU (d)) but < S > 6= SU (d). In Example 1 we provide such a subgroup for d = 2. We next characterise space C(AdS ) for SU (2). Let us recall that the composition of two unitary matrices U (γ, ~k12 ) = U (φ1 , ~k1 )U (φ2 , ~k2 ) is a unitary matrix with γ and ~k12 determined by: ~k12 cos γ = cos φ1 cos φ2 − sin φ1 sin φ2~k1 · ~k2 , 1 ~ = (k1 sin φ1 cos φ2 + ~k2 sin φ2 cos φ1 + sin γ +~k1 × ~k2 sin φ1 sin φ2 ). (11) (12) Moreover, two unitary matrices U1 (φ1 , ~k1 ), U2 (φ2 , ~k2 ) commute iff ~k1 k ~k2 or φ = kπ. Similarly two orthogonal matrices O1 (φ1 , ~k1 ), O2 (φ2 , ~k2 )) commute if ~k1 k ~k2 or one of φi ’s is an even multiple of π, or φ1 = ±π = φ2 and ~k1 ⊥ ~k2 . Making use of these facts in [33] we show: Fact 1. For noncommuting U1 (φ1 , ~k1 ), U2 (φ2 , ~k2 ), the space C(AdU1 (φ1 ,~k1 ) , AdU2 (φ2 ,~k2 ) ) is larger than {λI : λ ∈ R} if and only if: (1) φ1 = kπ 2 = φ2 , (2) one of φi ’s is kπ ~ ~ equal to 2 and k1 ⊥ k2 , where k is an odd integer. II. WHEN IS < S > INFINITE? We next describe the conditions under which < S > is infinite. For U1 , U2 ∈ SU (d) the group commutator is defined as [U1 , U2 ]• = U1 U2 U1−1 U2−1 . Note that [U1 , U2 ] = 0 is equivalent to [U1 , U2 ]• = I. The distance between elements of SU (d) can be measured using the √ Hilbert-Schmidt norm \|\|U \|\| = trU U † . For two elements U1 , U2 we have the following relation between their distances from the identity and the distance of their group commutator from the identity [12]: √ \|\|[U1 , U2 ]• − I\|\| ≤ 2\|\|U1 − I\|\| · \|\|U2 − I\|\|, (13) [U1 , [U1 , U2 ]• ]• = I and \|\|U2 − I\|\| < 2 ⇒ [U1 , U2 ]• = I. Let 1 Bα = {U ∈ SU (d) : kU − αIk ≤ √ } ⊂ SU (d), 2 (14) be a ball of radius √12 that is centred at elements αI. As det αI = 1 we need to assume αd = 1. Let [ B= Bα . (15) αd =1 It turns out that noncommuting elements belonging to B generate infinite subgroups of SU (d): Lemma 3. Assume that [U1 , U2 ]• ∈ / Z(SU (d)) and U1 , U2 ∈ B. Then < U1 , U2 > is infinite. One of the steps in the proof of lemma 3 uses relations (13) to show that the sequence g0 = U1 , g1 = [U1 , U2 ]• , gk = [gk−1 , U2 ]• converges to I and gn 6= I for any integer n [17] (see [33] for the full discussion). We next describe when U ∈ Bαm , where αm = eiθm and θm = 2m d π. To this end note that kU − αm Ik2 = 2trI − αm trU † − α∗m trU. (16) As the trace of U is determined by its spectrum, the desired condition can be expressed in terms of the eigenvalues of U that are given by {eiφ1 , ..., eiφd }, φi ∈ [0, 2π[ P and di=1 φi = 0 mod 2π. Direct calculations lead to: U ∈ Bαm ⇔ d X i=1 sin2 φi − θm 1 < . 2 8 (17) Let us next assume that U ∈ SU (d) does not belong to B. Then one can show that there always exists an integer n such that U n belongs to some Bα [33], αI ∈ Z(SU (d)). For a given U , let nU be the smallest integer satisfying this condition. In [33] we prove the modified version of the Dirichlet’s approximation theorem and use it to find an upper bound for NSU(d) := maxU nU . This way, for every U ∈ SU (d) there is 1 ≤ n ≤ NSU(d) such that U n ∈ Bα for some αI ∈ Z(SU (d)). Thus by taking powers 1 ≤ n ≤ NSU(d) we can move every element of SU (d) into B. Assume next that the necessary condition for universality is satisfied, i.e. C(AdS ) = {λI : λ ∈ R}. From Lemma 1 one can easily deduce that under the assumption that < S > is infinite the intersection < S > ∩B is dense in B. As we have shown in [33] the necessary condition for universality places significant constrains on the structure of < S > also in the case when < S > is a 4 finite group, namely we have that < S > ∩B is a subset of Z(SU (d)). Thus < S > is finite if and only if there are no elements in < S > that belong to B other than those in Z(SU (d)). The above discussion is summarised by: Lemma 4. Let S = {U1 , . . . , Un } and assume that C(AdS ) = {λI : λ ∈ R}. Then < S > = SU (d) if and only if there is at least one matrix U ∈< S > that belongs to B \ Z(SU (d)). We know that every element of S can be put to B by taking powers (that are bounded by NSU(d) ). Hence when < S > is finite introducing U ∈< S > to B must be equivalent to introducing it to Z(G). This condition can be phrased in terms of specra of the matrices from S. Definition 1. Assume U ∈ / B. The spectrum of U is called exceptional if it consists of nth roots of α ∈ C where αd = 1 and 1 ≤ n ≤ NSU(d) . The set of exceptional spectra is a finite set. To illustrate the above ideas we find NSU(d) and the list of exceptional spectra for d = 2. Note that for any U ∈ SU (2) the spectrum is given by {eiφ , e−iφ } and therefore is determined by one angle φ. The angle corresponding to an exceptional spectrum will be called an exceptional angle. Moreover, the centre of SU (2) consists of two matrices Z(SU (2)) = {I, −I}. We start with recalling the Dirichlet approximation theorem [16]. Fact 2 (Dirichlet). For a given real number b and a positive integer N there exist integers 1 ≤ n ≤ N and p such, that nb differs from p by at most N1+1 , i.e. \|nb − p\| ≤ 1 . N +1 (18) Let [0, 2π) ∋ φ = 2bπ be the spectral angle of U . By Fact 2 for a given N there are integers p and 1 ≤ n ≤ N such that \|nb − p\| ≤ N1+1 . Multiplying this inequality by π2 we obtain \|n φ2 − p π2 \| ≤ 2(Nπ+1) . Note that (17) 1 simplifies to \| sin ψ2 \| < 14 or \| sin ψ−π 2 \| < 4 . Thus for a φ given φ we search for n satisfying \|n 2 − p π2 \| < arcsin 14 . π < arcsin 41 and Thus 2(n+1) n≤ π 2 − arcsin arcsin 14 1 4 = 6. (19) The above upper bound for NSU(d) is attained for φ = arcsin 14 (see figure 17). Hence NSU(2) = 6. Exceptional spectra for SU (2), are determined by roots of 1 or −1 of order 1 ≤ n ≤ 6, or equivalently, by primitive roots of unity of order 1 ≤ n ≤ 6 and 8, 10, 12. More precisely, they are given by {eiθ , e−iθ }, where k2 π k3 π k4 π k5 π k6 π , , , , }, θ ∈ {kπ, 2 3 4 5 6 (20) 1 0.5 0 FIG. 1. Condition (17) for SU (2). The gray segments are determined by \| sin φ2 \| < 14 and the white segments by \| sin φ−π \| < 14 . 2 and gcd(ki , i) = 1. The number of exceptional angles can be calculated using P6 P6 the Euler totient function, and is equal i=1 φ(i) + i=4 φ(2i) = 24. For higher dimensional groups, as we discuss in [33], the number NSU(d) grows exponentially with d. Our main result is: Theorem 1. Assume that C(AdU1 , . . . , AdUn ) = {λI : λ ∈ R} and at least one matrix Ui has a nonexceptional spectrum. Then < U1 , . . . , Un > = SU (d). III. THE ALGORITHM FOR DECIDING UNIVERSALITY OF S = {U1 , . . . , Un } ⊂ SU (d) The case when all matrices {U1 , . . . , Un } have exceptional spectra requires an algorithm which we next present. Our algorithm allows deciding universality of any given set of SU (d) gates in a finite number of steps. The algorithm Step 1: Check if C(AdS ) = {λI : λ ∈ R}. This can be done by checking the dimension of the kernel of the matrix MS (10) constructed from the entries of matrices {AdU1 , . . . , AdUn } and thus is a linear algebra problem. If the answer is NO stop as the set S is not universal. If YES, set l = 1 and go to step 2. Step 2: Check if there is a matrix U ∈ S for which U nU belongs to B but not to Z(SU (d)), where 1 ≤ nU ≤ NSU(d) . This can be done using formula (16). If the answer is YES S is universal. If the answer is NO, set l = l + 1. Step 3: Define the new set S by adding to S words of length l, i.e products of elements from S of length l. If the new S is equal to the old one, the group < S > is finite. Otherwise go to step 2. The major advantage of our approach is the fact that we can make decisions in steps 2 and 3 in finite ‘time’. It is also clear that for randomly chosen matrices S = 5 {U1 , . . . , Un } ⊂ SU (d) our algorithm terminates with probability 1 in Step 2 for l = 1. This is a direct consequence of the fact that exceptional spectra form a finite set. Let us next discuss the correctness of our algorithm. Assume that S passes positively the necessary condition for universality, i.e. the Step 1. If the group < S > is finite the algorithm terminates in Step 3 for some finite l. On the other hand, as a direct consequence of Lemma 4, for an infinite < S > the algorithm must terminate in Step 2 for some finite l. One can also argue that if all finite-length words have exceptional spectra then they cannot form a dense subset. Thus if < S > is dense then it must contain words of a finite length that have non-exceptional spectra. These words terminate the algorithm in Step 2. Moreover, we have the following: Fact 3. Assume < S > is dense in SU (d). The length of a word that gives termination of the universality algorithm is at most the length l such that words of length k ≤ l form an ǫ-net that covers SU (d), where ǫ = 2√12+δ and δ > 0 is arbitrary small. Proof. Assume that words of the length k ≤ l built from elements S form an ǫ-net for SU (d), where ǫ = 2√12+δ and δ > 0 is arbitrary small. Let U be an element of 1 SU (d) whose distance from the identity is exactly 2√ 2 (see Figure 2). Then by the definition of ǫ-net there must be at least one word w ∈< S > of length k ≤ l contained in the ball C of radius ǫ = 2√12+δ centred at U . But this ball is contained in B1 \ I. Hence w gives termination of the universality algorithm in Step 2. The result follows. both φ1 and φ2 are exceptional angles and the product U (φ12 , ~k12 ) = U (φ1 , ~k1 )U (φ2 , ~k2 ) has exceptional φ12 as otherwise the algorithm terminates in Step 2 with either l = 1 or l = 2. For each such pair (there are finitely many of them) we find the Step and l that gives termination of our algorithm. The detailed discussion of the results and their connection to finite subgroups of SU (2) can be found in [33]. In short words, it turns out that our algorithm terminates at Step 2 if and only if 1 ≤ l ≤ 4 and at Step 3 when 5 ≤ l ≤ 13. Fact 4. For S = {U (φ1 , ~k1 ), U (φ2 , ~k2 )} ⊂ SU (2) the algorithm for checking universality terminates for l ≤ 13. Moreover, the set S is universal if and only if the algorithm terminates for l ≤ 4. The main conclusion from Fact 4 is that one can decide universality of any two-element subset of SU (2) by looking at words of the length at most 4. IV. EXAMPLES FOR SU (2) In the remaining part of this paper we demonstrate our approach calculating a few examples. They are chosen particularly to elucidate the importance of the conditions given by Theorem 1. a. Example 1 Let S = {U (φ, ~k1 ), U (π/2, ~k2 )}, where ~k1 ⊥ ~k2 and φ1 is an irrational multiple π. For example, when ~k1 = (0, 0, 1) and ~k2 = (1, 0, 0), we have −iφ e 0 0 1 ~ . U (φ, ~k1 ) = , U (π/2, k ) = 2 −1 0 0 eiφ (21) U (φ, ~k1 ) is of an infinite order and since U (φ, ~k1 ) and U (π/2, k~2 ) do not commute we have that < S > is infinite and not abelian. By Fact 1, however, C(AdU(φ,~k1 ) , AdU(π/2,~k2 ) ) 6= {λI : λ ∈ C}, and hence < S > 6= SU (2). For example O(π, ~k1 ) ∈ SO(3) commutes with both AdU(φ1 ,~k1 ) = O(2φ1 , ~k1 ) and Ad = O(π, ~k2 ). Interestingly, however ~ U(π/2,k2 ) C(U1 , U2 ) = {λI : λ ∈ C}. FIG. 2. The proof of Fact 3. The formulation of Fact 3 is related to the results contained in [18]. In order to demonstrate how efficient is our algorithm we determine the maximal l which gives its termination in Step 2 and Step 3 respectively for SU (2). For simplicity we consider S of the form S = {U (φ1 , ~k1 ), U (φ2 , ~k2 )} ⊂ SU (2). To this end it is enough to consider the case when To understand the structure of the group < S > note that U (π/2, k~2 )U (φ1 , ~k1 )U −1 (π/2, k~2 ) = U −1 (φ1 , ~k1 ). Hence U (π/2, k~2 ) is a normaliser of < U (φ1 , ~k1 ) >. Thus the group < S > consists of two connected components. The first one is given by one-parameter group U (t, ~k1 ), where t ∈ R and the other one by elements of the form U (π/2, k~2 )U (t, ~k1 ). The adjoint representation is able to identify infinite disconnected subgroups whereas the defining representation is not. Moreover, we know exactly how to fix non-universality of the set S. For example, we can add one matrix U (γ, ~kγ ) such that γ 6= kπ and ~kγ is neither parallel nor orthogonal to ~k1 and ~k2 . 6 non-exceptional spectrum and whose ~k is neither parallel nor orthogonal to ~kH and ~kTπ/2 . Case 3: For φ 6= kπ 2 , again by Fact 1, C(AdH , AdT (φ) ) = {λI : λ ∈ R} and we just need to check if < H, Tφ > is infinite. We distinguish three possibilities: 1. We first assume that φ is not exceptional. Then by Theorem 1 hH, Tφ i = SU (2). Our algorithm for deciding universality terminates at step 2 with l = 1. 2. We next consider the exceptional angles. For φ∈{ FIG. 3. Two component group generated by S {U (φ, ~k1 ), U (π/2, ~k2 )}. = b. Example 2 Let H be the Hadamard gate and Tφ a phase gate with an arbitrary phase φ: i H=√ 2 1 1 1 −1 , Tφ = e−iφ 0 0 eiφ . (22) Using our notation H = U (π/2, ~kH ), where ~kH = √1 (0, 1, 1) and Tφ = U (φ, ~ kTφ ), where ~kTφ = (0, 0, 1), 2 ~k1 · ~k2 = √1 . 2 Our goal is to check for which φ, < H, Tφ > = SU (2). Case 1: If φ = kπ then Tφ = ±I and the generated group is the finite cyclic group of the order 4 when φ = 0 or the order 8 when φ = π. Case 2: When φ = kπ 2 and k is odd, by Fact 1 we have that C(AdH , AdTφ ) is larger than {λI : λ ∈ R} and hence < H, T kπ > 6= SU (2). In fact it is the 2 finite dicyclic group of order 16 whose generators are HT and T . Fixing universality in this case requires, for example, adding a matrix that has a [1] L. Babai, Proceedings of the Third Annual ACM SIAM Symposium on Discrete Algorithms (Orlando, FL, 1992). ACM, New York, pp. 33-40, (1992). [2] L. Babai, R. Beals, D. N. Rockmore, Proc. of International Symposium on Symbolic and Algebraic Computation. ISSAC 93. ACM Press, pp. 117-126, (1993). [3] A. Bocharov et al., Phys. Rev. A 88, 012313 (2013). [4] A. Böttcher, D.Wenzel, Linear Algebra Appl. 429 (2008). k3 π k5 π k6 π , , }, gcd(ki , i) = 1, 3 5 6 we look at the product U (γ, ~kHT ) = HTφ = U (π/2, ~kH )U (φ, ~kT ). Using formula (11) we calculate cos γ, compare it with cos ψ for all exceptional angles ψ and find out they never agree. Hence γ is not exceptional. Thus by Theorem 1 we get < HTφ > = SU (2). Our algorithm for deciding universality terminates in Step 2 with l = 2. 3. We are left with φ = k44π where gcd(k4 , 4) = 1. There are exactly four such angles. Calculations of U (γ, ~kHTφ ) = HTφ shows that γ is exceptional, i.e γ = k33π , where gcd(k3 , 3) = 1. Moreover, taking further products results in a finite subgroup consisting of 48 elements (all have exceptional spectra) known as the binary octahedral group. Our algorithm for deciding universality terminates in Step 3 with l = 8. Fixing non-universality can be accomplished by, for example, adding one gate U (ψ, ~kψ ) with a non-exceptional ψ and an arbitrary ~kψ . As we can see from the above example our algorithm requires at most words of length l = 8 to terminate for any H and Tφ . V. ACKNOWLEDGMENT We would like to thank the anonymous referees for suggestions that led to improvements of the paper. This work was supported by National Science Centre, Poland under the grant SONATA BIS: 2015/18/E/ST1/00200. [5] A. Bouland, S. Aaronson, Phys. Rev. A 89, 062316 (2014). [6] J. Bourgain and A. Gamburd , J. Eur. Math. Soc. 014.5, 1455-1511, (2012). [7] J. Bourgain and A. Gamburd, Invent. Math., 171, Issue 1, 83-12, (2008). [8] T. Bröcker and T. Dieck, Representations of Compact Lie Groups, Springer-Verlag, New York, MR 86i:22023,(1985). 7 [9] Y. Bromberg et al., Phys. Rev. Lett. 102, 253904 (2009). [10] R. Bryliński, G. Chen, Mathematics of Quantum Computation (Boca Raton, FL: Chapman and Hall/CRC Press, 2002) [11] A. M. Childs et al. Quantum Info. Comput. 11, (2011). [12] W. Curtis, I. Reiner, Representation theory of finite groups and associative algebras (Interscience Publishers, John Wiley and sons, 1962) [13] H. Derksen et al., J. Symb. Comput. 39, 357-371, (2005). [14] A. S. Detinko, D. L. Flannery, Journal of Symbolic Computation 44 1037-1043, (2009). [15] D. Deutsch, Proc. Roy. Soc. Lond. A 425 (1989). [16] G. H. Hardy, E.M. Wright, An introduction to the Theory of Numbers, Oxford at the Clarendon Press, (1960). [17] M. Field, Proc. Amer. Math. Soc. 127 (2011). [18] M. H. Freedman et al., Math. Res. Lett. 10, 11, (2003). [19] B. C. Hall, Lie Groups, Lie Algebras and Representations: An Elementary Introduction, (Springer, GTM, 2nd ed., 2015). [20] A. W. Harrow et al., J. Math. Phys. 43, 4445-4451, (2002). [21] V. Kliuchnikov et al., arXiv:1510.03888, (2015). [22] V. Kliuchnikov et al., IEEE Transactions on Computers, vol.65, no. 1, pp. 161-172, (2016). [23] M. Kuranishi, Nagoya Math. J. 2, (1951). [24] S. Lloyd Phys. Rev. Lett. 75, 2 (1995). [25] A. Lubotzky, et al., Comm. Pur. Appl. Math. 40, 4 401420, (1987). [26] M. Nielsen, I. Chuang, Quantum Computation and Quantum Information (Cambridge University Press, 2000). [27] M. Oszmaniec et al., Phys. Rev. A, vol. 90, p. 020302, (2014). [28] M. Oszmaniec et al., Phys. Rev. X 6, 041044 (2016). [29] A. Politi et al., Science 320, (2008). [30] M. Reck et al., Phys. Rev. Lett. 73, 1 (1994). [31] P. Sarnak, Letter to Scott Aaronson and Andy Pollington on the Solovay-Kitaev theorem, (2015). [32] A. Sawicki, Quantum Info. Comput. 16, 3&4 (2016). [33] A. Sawicki, K. Karnas, Universality of single qudit gates, arXiv:1609.05780, (2016). [34] P. Selinger, Quant. Inf. Comp. 15, 159-180 (2015). [35] R. Zeier, T. Schulte-Herbrüggen, J. Math. Phys. 52, 113510, (2011). [36] R. Zeier, Z. Zimborás, J. Math. Phys. 56, 081702, (2015). [37] Z. Zimborás et al., Phys. Rev. A 92, 042309 (2015).	4math.GR
Iterators, Recursors and Interaction Nets Ian Mackie1 , Jorge Sousa Pinto2 , and Miguel Vilaça2 mackie@lix.polytechnique.fr,{jsp,jmvilaca}@di.uminho.pt 1 arXiv:0910.3321v1 [cs.PL] 17 Oct 2009 2 LIX, CNRS UMR 7161, École Polytechnique, 91128 Palaiseau Cedex, France Departamento de Informática / CCTC, Universidade do Minho, Braga, Portugal Abstract. We propose a method for encoding iterators (and recursion operators in general) using interaction nets (INs). There are two main applications for this: the method can be used to obtain a visual notation for functional programs; and it can be used to extend the existing translations of the λ-calculus into INs to languages with recursive types. 1 Introduction The use of visual notations for functional programs has long been an active research topic, whose main goal is to have a notation that can be used (i) to define functional programs visually, and (ii) to animate their execution. In this paper we propose a graphical system for functional programming, based on token-passing INs [9]. The system offers an adequate solution for classic problems of visual notations, including the treatment of higher-order functions, pattern-matching, and recursion (based on the use of recursion operators). The system implements a call-by-name semantics, with a straightforward correspondence between functional programs and graphical objects. Most approaches to visual programming simply propose a notation for programs. Program evaluation is animated by representing visually the intermediate programs that result from executing reduction steps on the initial program, using the operational semantics of the underlying functional language. Our approach is different in that we use a graph-rewriting formalism with its own semantics. The advantages of using INs for visual programming are: – – – – Both programs and data are represented in the same graphical formalism. Programs can be animated without leaving the interaction formalism. Pattern-matching for external constructors is in-built. Recursive definitions are expressed very naturally as interaction rules involving agents that are reintroduced on the right-hand side. But the interaction net formalism does not offer a satisfactory semantic interpretation for the behaviour of functional symbols. Moreover, many interaction net systems can be defined that do not have a functional reading. What is missing is a clear correspondence between functional definitions and interaction systems.In this paper we establish a correspondence between agents with “obviously functional” interaction rules and functions defined with recursion operators. 2 The Token-passing Encoding of the λ-calculus The token-passing encodings [9] use an interaction system where two different symbols exist for application: one is the syntactic symbol @ introduced by the translation; the corresponding agents have their principal ports facing the root b exists that will of the term and will be depicted by triangles. A second symbol @ be used for computation; to simplify the figures, the corresponding agents will be depicted by circles equally labelled with @. Their principal ports face the net that represents the applied function, to make possible interaction with λ agents. The translation Ttp (·) encodes terms in the system (Σtp , Rtp ) where Σtp = b λ, c, ε, δ}. It generates nets containing no active pairs.The special sym{⇓, @, @, bol ⇓ is used as an evaluation token: an agent ⇓ traverses the net, transforming b The evaluation rules involving ⇓ can be tailored to a occurrences of @ into @. specific evaluation strategy. For call-by-name, Rtp consists of the following rules which comprises evaluation rules involving ⇓ and a computation rule involving @ and λ. Management (copying and erasing) rules are omitted here. To start the reduction a ⇓ symbol must be connected to the root port of the term. Let ⇓ N denote such net; then the following correctness result holds: t ⇓ z iff ⇓ Ttp (t) −→∗ Ttp (z), where the evaluation relation · ⇓ · is defined by the standard evaluation rules for call-by-name. The language used in this paper is the simply-typed λ-calculus extended with natural numbers, booleans, lists, and iterators for these types. BNL is defined by the following syntax for terms (x, y range over a set of variables): t, u, v ::= x \| λx.t \| t u \| tt \| ff \| iterbool(t, u, v) \| 0 \| suc(t) \| iternat(λx.t, u, v) \| nil \| cons(t, u) \| iterlist(λxy.t, u, v) 3 A Token-passing Encoding of BNL We extend to BNL the token-passing call-by-name translation of the λ-calculus into the interaction system (Σtp , Rtp ). The novelty of this encoding is not the token-passing aspect, but rather the approach to recursion. We first consider data structures. In a token-passing implementation, there will be an interaction rule between the token agent and each constructor symbol that will stop evaluation. For BNL we define the system (ΣBNL , RBNL ) where ΣBNL consists of the symbols tt, ff, 0 and nil with arity 0; suc with arity 1; and cons with arity 2. Each recursive program will be encoded in an interaction system specifically generated for it. This is a major novelty of our approach. The interaction system for the λ-calculus will not be extended by introducing a fixed set of symbols; instead a new symbol will be introduced for each occurrence of a recursion operator, with an interaction rule for each different constructor, so a dedicated interaction system (Σt0 , Rt0 ) is generated for each term t. This system is constructed by a recursive function (Σt0 , Rt0 ) = S (t), defined as: . . . . . S (x) = S (tt) = S (ff) = S (0) = S (nil) = (∅, ∅) . . S (λx.t) = S (suc(t)) = S (t) . . S (t u) = S (cons(t, u)) = S (t) ∪ S (u) . d Bool S (iterbool(V, F, b)) = ({ItBool V,F , ItV,F } ∪ Σ, RItBool ∪ R), V,F where (Σ, R) = S (b) ∪ S (V ) ∪ S (F ). . Nat S (iternat(λx.S, Z, n)) = ({ItNat , Itd } ∪ Σ, R S,Z S,Z ItNat S,Z where (Σ, R) = S (n) ∪ S (S) ∪ S (Z) . List S (iterlist(λxy.C, N, l)) = ({ItList , Itd } ∪ Σ, R C,N C,N ∪ R) ItList C,N ∪ R) where (Σ, R) = S (l) ∪ S (C) ∪ S (N ) RItList consists of the following interaction rules (the others are similar). C,N ... c Iterator symbols are introduced in pairs (It... ... , It... ) where the first symbol is used for syntactic agents and the second for computation agents (similarly to b @, @). A BNL program t will be translated into a net defined in the system (Σt , Rt ) = (Σtp ∪ ΣBNL ∪ Σt0 , Rtp ∪ RBNL ∪ Rt0 ) where (Σtp , Rtp ) was defined in Section 2. Given a BNL program t, where t is iterbool(V, F, b), iternat(λx.S, Z, n) or iterlist(λxy.C, N, l), then the net T (t) is given as follows. In token-passing implementations, all terms are translated as syntax trees. Syntactic iterator agents i are turned into their computation counterparts bi by token agents. A first key aspect of our approach is that the interaction rules of the (computation) iterator agents internalise the iterator’s parameters. For instance the net T (iterlist(λxy.C, N, cons(h, t))) reduces to ⇓ T (C[h/x, iterlist(λxy.C, N, t)/y]). A second key aspect is that each such new symbol will have auxiliary ports in a one-to-one correspondence with the free variables in the iterator term. We end the section with a correctness result. The proofs can be found in [1]. Proposition 1. (Correctness) If t is a closed BNL term and z a canonical form, then: t ⇓ z ⇐⇒ ⇓ T (t) −→∗ T (z). 4 Conclusions and Future Work We have presented an approach to encoding in INs functional programs defined with recursion operators, and given the full details of the application of this approach to the token-passing implementation of a call-by-name language, which results in a very convenient visual notation for this language. The approach can be easily extended to richer sets of recursive types and other recursion operators and also to new strategies. The novel characteristics of the encoding are (i) the interaction system is generated dynamically from the program, and (ii) the internalisation of some of the parameters of the recursion operator in the interaction rules. With respect to previous work on encoding recursion in interaction nets, fixpoint operators have been studied elsewhere for interaction net implementations [4, 6], and we have shown elsewhere how a binding recursion operator (as in PCF) can be implemented in the token-passing setting [2]. A prototype system for visual functional programming has been developed, integrated in the tool INblobs [3, 10] for interaction net programming. The tool consists of an evaluator for interaction nets together with a visual editor and a compiler module that translates programs into nets. The latter module allows users to type in a functional program, visualize it, and then follow its evaluation visually step by step. The current compiler module automatically generates callby-name or call-by-value systems. Additionally, a visual editing mode is available that allows users to construct nets corresponding to functional programs. In the current implementation there is no way to convert visual programs back to textual ones. The token-passing translation is not however very representative of most work in this area, which has concentrated on designing efficient translations; [5, 7, 8] are some samples. Let T (·) be one such translation. Typically T (t u) is constructed from T (t) and T (u) by introducing an application symbol @ with its principal port connected to the root port of T (t). Our treatment of iterators can be adapted to this setting by removing the evaluator tokens and introducing the iterator agents with the principal port immediately facing the argument. When the iterated function is a closed term, a correctness result can be easily established. Let λx.S be a closed term, then T (iternat(λx.S, Z, 0)) −→ T (Z) and T (iternat(λx.S, Z, suc(n))) −→ T (S[iternat(λx.S, Z, n)/x]) References 1. J. B. Almeida, I. Mackie, J. S. Pinto, and M. Vilaça. Encoding iterators in interaction nets. Available from http://www.di.uminho.pt/~jmvilaca. 2. J. B. Almeida, J. S. Pinto, and M. Vilaça. Token-passing Nets for Functional Languages. In J. Giesl, editor, Proceedings of the 7th International Workshop on Reduction Strategies in Rewriting and Programming (WRS’07), volume 204 of Electronic Notes in Theoretical Computer Science, pages 181–198, 2007. 3. J. B. Almeida, J. S. Pinto, and M. Vilaça. A Tool for Programming with Interaction Nets. In Proceedings of the The Eighth International Workshop on Rule-Based Programming (RULE’07), 2007. To appear in Elsevier ENTCS. 4. A. Asperti and S. Guerrini. The Optimal Implementation of Functional Programming Languages, volume 45 of Cambridge Tracts in Theoretical Computer Science. Cambridge University Press, 1998. 5. G. Gonthier, M. Abadi, and J.-J. Lévy. The geometry of optimal lambda reduction. In Proceedings of the 19th ACM Symposium on Principles of Programming Languages (POPL’92), pages 15–26. ACM Press, Jan. 1992. 6. I. Mackie. The geometry of interaction machine. In Proceedings of the 22nd ACM Symposium on Principles of Programming Languages (POPL’95), pages 198–208. ACM Press, January 1995. 7. I. Mackie. YALE: Yet another lambda evaluator based on interaction nets. In Proceedings of the 3rd International Conference on Functional Programming (ICFP’98), pages 117–128. ACM Press, 1998. 8. I. Mackie. Efficient λ-evaluation with interaction nets. In V. van Oostrom, editor, Proceedings of the 15th International Conference on Rewriting Techniques and Applications (RTA’04), volume 3091 of Lecture Notes in Computer Science, pages 155–169. Springer-Verlag, June 2004. 9. F.-R. Sinot. Call-by-name and call-by-value as token-passing interaction nets. In P. Urzyczyn, editor, TLCA, volume 3461 of Lecture Notes in Computer Science, pages 386–400. Springer, 2005. 10. M. Vilaça. Inblobs webpage. http://haskell.di.uminho.pt/jmvilaca/INblobs/.	6cs.PL
Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks Brenden Lake 1 2 Marco Baroni 2 arXiv:1711.00350v2 [cs.CL] 11 Feb 2018 Abstract Humans can understand and produce new utterances effortlessly, thanks to their compositional skills. Once a person learns the meaning of a new verb “dax,” he or she can immediately understand the meaning of “dax twice” or “sing and dax.” In this paper, we introduce the SCAN domain, consisting of a set of simple compositional navigation commands paired with the corresponding action sequences. We then test the zero-shot generalization capabilities of a variety of recurrent neural networks (RNNs) trained on SCAN with sequence-to-sequence methods. We find that RNNs can make successful zero-shot generalizations when the differences between training and test commands are small, so that they can apply “mix-and-match” strategies to solve the task. However, when generalization requires systematic compositional skills (as in the “dax” example above), RNNs fail spectacularly. We conclude with a proof-of-concept experiment in neural machine translation, suggesting that lack of systematicity might be partially responsible for neural networks’ notorious training data thirst. 1. Introduction Human language and thought are characterized by systematic compositionality, the algebraic capacity to understand and produce a potentially infinite number of novel combinations from known components (Chomsky, 1957; Montague, 1970). For example, if a person knows the meaning and usage of words such as “twice,” “and,” and “again,” once she learns a new verb such as “to dax” she can immediately understand or produce instructions such as “dax twice and then dax again.” This type of compositionality is central to the human ability to make strong generalizations from very 1 Dept. of Psychology and Center for Data Science, New York University 2 Facebook Artificial Intelligence Research. Correspondence to: Brenden Lake <brenden@nyu.edu>, Marco Baroni <mbaroni@fb.com>. limited data (Lake et al., 2017). In a set of influential and controversial papers, Jerry Fodor and other researchers have argued that neural networks are not plausible models of the mind because they are associative devices that cannot capture systematic compositionality (Fodor & Pylyshyn, 1988; Marcus, 1998; Fodor & Lepore, 2002; Marcus, 2003; Calvo & Symons, 2014, a.o.). In the last few years, neural network research has made astounding progress in practical domains where success crucially depends on generalization. Perhaps most strikingly, end-to-end recurrent neural networks currently dominate the state-of-the-art in machine translation (Bojar et al., 2016; Wu et al., 2016). Since the overwhelming majority of sentences or even word sequences in a language only occur once, even in a large corpus (Baroni, 2009), this points to strong generalization abilities. Still, it is commonly observed that neural networks are extremely sample inefficient, requiring very large training sets, which suggests they may lack the same algebraic compositionality that humans exploit, and they might only be sensitive to broad patterns over lots of accumulated statistics (Lake et al., 2017). In this paper, we introduce a grounded navigation environment where the learner must translate commands given in a limited form of natural language into a sequence of actions. This problem is naturally framed as a sequence-to-sequence task, and, due to its simplicity, it is ideal to study systematic generalization to novel examples in a controlled setup. We thus use it to test a wide range of modern recurrent network architectures in terms of their compositional abilities. Our results suggest that standard architectures such as LSTMs with attention (Bahdanau et al., 2015) generalize very well when novel examples feature a mixture of constructions that have been observed in training. However, the models are catastrophically affected by systematic differences between training and test sentences, of the sort that would be trivial for an agent equipped with an “algebraic mind” (Marcus, 2003). Generalization without Systematicity 2. The SCAN tasks We call our data set SCAN because it is a Simplified version of the CommAI Navigation tasks (Mikolov et al., 2016).1 For a learner, the goal is to translate commands presented in simplified natural language into a sequence of actions. Since each command is unambiguously associated to a single action sequence, SCAN (unlike the original CommAI tasks) can be straightforwardly treated as a supervised sequenceto-sequence semantic parsing task (Dong & Lapata, 2016; Jia & Liang, 2016; Herzig & Berant, 2017), where the input vocabulary is given by the set of words used in the commands, and the output by the set of actions available to the learner. Several examples from SCAN are presented in Fig. 1. Formally, SCAN consists of all the commands generated by a phrase-structure grammar (presented in the Appendix) and the corresponding sequence of actions, produced according to a semantic interpretation function (see Appendix). Intuitively, the SCAN grammar licenses commands denoting primitive actions such as JUMP (denoted by “jump”; Fig. 1), WALK (denoted by “walk”) and LTURN (denoted by “turn left”). We will refer to these as primitive commands.2 It also accepts a set of modifiers and conjunctions that compositionally build expressions referring to action sequences. The “left” and “right” modifiers take commands denoting undirected primitive actions as input and return commands denoting their directed counterparts (“jump left”; Fig. 1). The “opposite” modifier produces an action sequence that turns the agent backward in the specified direction before executing a target action (“jump opposite left”), while “around” makes the agent execute the action at each step while turning around in the specified direction (“jump around right”; Fig. 1). The “twice/thrice” modifiers trigger repetition of the command they take scope over, and “and/after” combine two action sequences. Although the SCAN examples in Fig. 1 focus on the “jump”/JUMP primitive, each instance of JUMP can be replaced with either WALK, RUN, or LOOK to generate yet more commands. Many more combinations are possible as licensed by the grammar. The SCAN grammar, lacking recursion, generates a finite but large set of unambiguous commands (20,910, to be precise). Commands can be decoded compositionally by applying the corresponding interpretation function. This means that, if it discovers the right interpretation function, a learner can understand commands it has not seen during training. For example, the learner might have only observed the primitive “jump” command during training, but if it has 1 SCAN available at: https://ANONYMIZED Introducing the primitive turning actions LTURN and RTURN considerably simplifies the interpretation function, compared to capturing orientation by specifying arguments to the movement actions (e.g., JUMP[L], JUMP[R]). 2 learned the meaning of “after”, “twice” and “around left” from other verbs, it should be able to decode, zero-shot, the complex command: “jump around left after jump twice”. 3. Models and setup We approach SCAN through the successful sequence-tosequence (seq2seq) framework, in which two recurrent networks work together to learn a mapping between input sequences and output sequences (e.g., Sutskever et al., 2014).3 Fig. 2 illustrates the application of the seq2seq approach to a SCAN example. First, a recurrent network encoder receives the input sequence word-by-word, forming a lowdimensional representation of the entire command. Second, the low-dimensional representation is passed to a recurrent network decoder, which then generates the output sequence action-by-action. The decoder’s output is compared with the ground truth, and the backpropagation algorithm is used to update the parameters of both the encoder and decoder. Note that although the encoder and decoder share the same network structure (e.g., number of layers and hidden units), they do not otherwise share weights/parameters with each other. More details regarding the encoder-decoder RNN are provided in the Appendix. Using the seq2seq framework, we tested a range of standard recurrent neural network models from the literature: simple recurrent networks (SRNs; Elman, 1990), long short-term memory networks (LSTMs; Hochreiter & Schmidhuber, 1997), and gated recurrent units (GRUs; Chung et al., 2014). Recurrent networks with attention have become increasingly popular in the last few years, and thus we also tested each network with and without an attentional mechanism, using the model from Bahdanau et al. (2015) (see Appendix for more details). Finally, to make the evaluations as systematic as possible, a large-scale hyperparameter search was conducted that varied the number of layers (1 or 2), the number of hidden units per layer (25, 50, 100, 200, or 400), and the amount of dropout (0, 0.1, 0.5; applied to recurrent layers and word embeddings). Varying these hyperparameters leads to 180 different network architectures, all of which were run on each experiment and replicated 5 times each with different random initializations.4 In reporting the results, we focus on the overall-best architecture as determined by the extensive hyperparameter search. The winning architecture was a 2-layer LSTM with 200 hidden units per layer, no attention, and dropout applied at the 0.5 level. Although the detailed analyses to follow focus on this particular model, the top-performing ar3 Very recently, convolutional seq2seq networks have reached comparable or superior performance in machine translation (Gehring et al., 2017). We will investigate them in future work. 4 A small number of runs (23/3600) did not complete, and thus not every network had 5 runs. Generalization without Systematicity jump jump left jump around right turn left twice jump thrice jump opposite left and walk thrice jump opposite left after walk around left ⇒ ⇒ ⇒ ⇒ ⇒ ⇒ ⇒ JUMP LTURN JUMP RTURN JUMP RTURN JUMP RTURN JUMP RTURN JUMP LTURN LTURN JUMP JUMP JUMP LTURN LTURN JUMP WALK WALK WALK LTURN WALK LTURN WALK LTURN WALK LTURN WALK LTURN LTURN JUMP Figure 1. Examples of SCAN commands (left) and the corresponding action sequences (right). JUMP jump twice and walk <EOS> <SOS> JUMP JUMP WALK JUMP <EOS> WALK Figure 2. The seq2seq framework is applied to SCAN. The symbols <EOS> and <SOS> denote end-of-sentence and start-ofsentence, respectively. The encoder (left) ends with the first <EOS> symbol, and the decoder (right) begins with <SOS>. 4. Experiments In each of the following experiments, the recurrent networks are trained on a large set of commands from the SCAN tasks to establish background knowledge as outlined above. After training, the networks are then evaluated on new commands designed to test generalization beyond the background set in systematic, compositional ways. In evaluating these new commands, the networks must make zero-shot generalizations and produce the appropriate action sequence based solely on extrapolation from the background training. Experiment 1: Generalizing to a random subset of commands chitecture for each experiment individually is also reported and analyzed. Networks were trained with the following specifications. Training consisted of 100,000 trials, each presenting an input/output sequence and then updating the networks weights.5 The ADAM optimization algorithm was used with default parameters, including a learning rate of 0.001 (Kingma & Welling, 2014). Gradients with a norm larger than 5.0 were clipped. Finally, the decoder requires the previous step’s output as the next step’s input, which was computed in two different ways. During training, for half the time, the network’s self-produced outputs were passed back to the next step, and for the other half of the time, the groundtruth outputs were passed back to the next step (teacher forcing; Williams & Zipser, 1989). The networks were implemented in PyTorch and based on a standard seq2seq implementation.6 Training accuracy was above 99.5% for the overall-best network in each of the key experiments, and it was at least 95% or above for the top-performers in each experiment specifically. 5 Note that, in all experiments, the number of distinct training commands is well below 100k: we randomly sampled them with replacement to reach the target size 6 The code we used is publicly available at the link: http://pytorch.org/tutorials/intermediate/ seq2seq_translation_tutorial.html In this experiment, the SCAN tasks were randomly split into a training set (80%) and a test set (20%). The training set provides broad coverage of the task space, and the test set examines how networks can decompose and recombine commands from the training set. For instance, the network is asked to perform the new command, “jump opposite right after walk around right thrice,” as a zero-shot generalization in the test set. Although the conjunction as a whole is novel, the parts are not: The training set features many examples of the parts in other contexts, e.g., “jump opposite right after turn opposite right” and “jump right twice after walk around right thrice” (both bold sub-strings appear 83 times in the training set). To succeed, the network needs to generalize by recombining pieces of existing commands to interpret new ones. Overall, the networks were highly successful at generalization. The top-performing network for this experiment achieved 99.8% correct on the test set (accuracy values here and below are averaged over the five training runs). The topperforming architecture was a LSTM with no attention, 2 layers of 200 hidden units, and no dropout. The best-overall network achieved 99.7% correct. Interestingly, not every architecture was successful: Classic SRNs performed very poorly, and the best SRN achieved less than 1.5% correct at test time (performance on the training set was equally low). However, attention-augmented SRNs learned the commands much better, achieving 59.7% correct on average for the test set (with a range between 18.4% and 94.0% across SRN Generalization without Systematicity value of 4 times. Experiment 2: Generalizing to commands demanding longer action sequences Figure 3. Zero-shot generalization after training on a random subset of the SCAN tasks. The overall-best network was trained on varying proportions of the set of distinct tasks (x-axis) and generalization was measured on new tasks (y-axis). Each bar shows the mean over 5 training runs with corresponding ±1 SEM. architectures). For LSTMs and GRUs, attention was not essential, since many of the highest performing architectures did not use it. As indicated above, the main split was quite generous, providing 80% of the commands at training time for a total of over 16,700 distinct examples (with strong combinatorial coverage). We next re-trained the best-overall network with varying numbers of distinct examples (the actual number of training presentations was kept constant at 100K). The results are shown in Fig. 3. With 1% of the commands shown during training (about 210 examples), the network performs poorly at about 5% correct. With 2% coverage, performance improves to about 54% correct on the test set. By 4% coverage, performance is about 93% correct. Our results show that not only can networks generalize to random subsets of the tasks, they can do so from relatively sparse coverage of the compositional command space. This is well in line with the success of seq2seq architectures in machine translation, where most test sentences are likely never encountered in training. Still, even with this sparser coverage, differences between training and test instances are not dramatic. Let us for example consider the set of all commands without a conjunction (e.g., “walk around thrice”, “run”, “jump opposite left twice”). All the commands of this sort that occur in the test set of the 2% training coverage split (either as components of a conjunction or by themselves) also occur in the corresponding training set, with an average of 8 occurrences. Even for the 1% split, there is only one conjunction-less test command that does not also occur in the training split, and the frequency of occurrence of such commands in the training set is at a non-negligible average We study next a more systematic form of generalization, where models must bootstrap to commands requiring longer action sequences than those seen in training.7 Now the training set contains all 16,990 commands requiring sequences of up to 22 actions, whereas the test set includes all remaining commands (3,920, requiring action sequences of lengths from 24 to 48). Under this split, for example, at test time the network must execute the command “jump around left twice and walk opposite right thrice”, requiring a sequence of 25 actions. Although all the elements used in the command have been observed during training, the network has never been asked to produce a sequence of this length, nor it has ever seen an “around * twice” command conjoined with an “opposite * thrice” command (although it did observe both components conjoined with others). Thus, it must productively generalize familiar verbs, modifiers and conjunctions to generate longer action sequences. This is a fair task for a system that is correctly translating the input commands. If you know how to “walk around,” how to “jump,” and the function of the “and” conjunction, you will be immediately able to “walk around and jump,” even if you have never performed an action sequence of that length. This test turns out to be very challenging for all models. The best result (20.8% on average, again over 5 runs) is achieved by a GRU with attention, one 50-dimensional hidden layer, and dropout 0.5 (interestingly, a model with considerably less capacity than the best for the random-split setup). The overall-best model achieves 13.8% accuracy. Fig. 4 (top) shows partial success is almost entirely explained by generalization to the shortest action sequence lengths in the test set. Although we might not expect even humans to be able to generalize to very long action sequences, the sharp drop between extrapolating to 25 and 26 actions is striking. The bottom panel of Fig. 4 shows accuracy in the test set organized by command length (in word tokens). The model only gets right some of the longest commands (8 or 9 tokens). In the training set, the longest action sequences (≥20) are invariably associated to commands containing 8 or 9 tokens. Thus, the model is correctly generalizing only in those cases that are most similar to training instances. 7 We focus on action sequence length rather than command length since the former exhibits more variance (1-48 vs. 1-9). The longest commands (9 words) are given by the conjunction of two directed primitives both modified twice, e.g.: “jump around left twice and run opposite right thrice.” On the other hand, a relatively short command such as “jump around left thrice” demands 24 actions. Accuracy on new commands (%) Generalization without Systematicity a strong effect of action sequence length, with average accuracy ranging from 95.76% for commands requiring 24 actions to 22.8% for commands requiring 48 actions). 100 80 Experiment 3: Generalizing composition across primitive commands 60 40 20 0 24 25 26 27 28 30 32 33 36 40 48 Accuracy on new commands (%) Ground-truth action sequence length 100 80 60 40 20 0 4 6 7 8 Command length 9 Figure 4. Zero-shot generalization to commands with action sequence lengths not seen in training. Top: accuracy distribution by action sequence length. Bottom: accuracy distribution by command length (only lengths attested in the test set shown, in both cases). Bars show means over 5 runs of overall-best model with ±1 SEM. Finally, we studied whether the difficulty with long sequences can be mitigated if the proper length was provided by an oracle at evaluation time.8 If this difficulty is a relatively straightforward issue of the decoder terminating too early, then this should provide an (unrealistic) fix. If this difficulty is symptomatic of deeper problems with generalization, then this change will have only a small effect. With the oracle, the overall-best network performance improved from 13.8% to 23.6% correct, which was notable but insufficient to master the long sequences. The top-performing model showed a more substantial improvement (20.8% to 60.2%). Although improved, the networks were far from perfect and still exhibited key difficulties with long sequences of output actions (again, even for the top model, there was 8 Any attempt from the decoder to terminate the action sequence with an <EOS> was ignored (and the second strongest action was chosen) until a sequence with proper length was produced. Our next test is closest to the “dax” thought experiment presented in the introduction. In the training phase, the model is exposed to the primitive command only denoting a certain basic action (e.g., “jump”). The model is also exposed to all primitive and composed commands for all other actions (e.g., “run”, “run twice”, “walk”, “walk opposite left and run twice”, etc.). At test time, the model has to execute all composed commands for the action that it only saw in the primitive context (e.g., “jump twice”, “jump opposite left and run twice”, etc.). According to the classic thought experiments of Fodor and colleagues, this should be easy: if you know the meaning of “run”, “jump” and “run twice”, you should also understand what “jump twice” means. We run two variants of the experiment generalizing from “turn left” and “jump”, respectively. Since “turn right” is distributionally identical to “turn left” (in the sense that it occurs in exactly the same composed commands) and “walk”, “run” and “look” are distributionally identical to “jump”, it is redundant to test all commands. Moreover, to ensure the networks were highly familiar with the target primitive command (“jump” or “turn left”), it was overrepresented in training such that roughly 10% of all training presentations were of the command.9 We obtain strikingly different results for “turn left” and “jump”. For “turn left”, many models generalize very well to composed commands. The best performance is achieved by a GRU network with attention, one layer with 100 hidden units, and dropout of 0.1 (90.3% accuracy). The overallbest model achieved 90.0% accuracy. On the other hand, for “jump,” models are almost completely incapable to generalize to composed commands. The best performance was 1.2% accuracy (LSTM, attention, one layer, 100 hidden units, dropout 0.1). The overall-best model reached 0.08% accuracy. In the case of “turn left”, although models are only exposed to the primitive command during training, they will see the action it denotes (LTURN) many times, as it is used to accomplish many directed actions. For example, a training example is: “walk left and jump left”, with ground-truth interpretation: LTURN WALK LTURN JUMP. Apparently, seeing action sequences containing LTURN suffices for the model to understand composed commands with “turn left”, probably because the model receives direct evidence 9 Without over-sampling, performance was consistently worse than what we report. Generalization without Systematicity Table 1. Nearest training commands for a sample of commands, with the respective cosines. Here, “jump” was trained in isolation while “run” was trained compositionally. Italics mark low similarities (cosine <0.2). run jump look walk .73 .65 run walk .15 .13 walk after run .55 turn right .12 run thrice after run run twice after run .50 look right twice after walk twice turn right after turn right .09 .49 .09 run twice look twice run twice and look opposite right thrice run twice and run right twice run twice and look opposite right twice walk twice and run twice about how LTURN is used in context. On the other hand, the action denoted by “jump” (JUMP) only occurs with this primitive command in training, and the model does not generalize from this minimal context to new composed ones. We now take a closer look at the results, focusing on the median-performance run of the overall-best model (as the most representative run of this model). We observe that even in the successful “turn left” case model errors are surprising. One would expect such errors to be randomly distributed, or perhaps to pertain to the longest commands or action sequences. Instead, all 45 errors made by the model are conjunctions where one of the components is simple “turn left” (22 cases) or “turn left thrice” (23 cases). This is particularly striking because the network produced the correct mapping for “turn left” during training, as well as for “turn left thrice” at test time, and it gets many more conjunctions right (ironically, including “turn left thrice and turn left”, “turn left thrice after turn left” etc.). We conclude that, even when the network has apparently learned systematic composition almost perfectly, it got at it in a very non-human-like way. It’s hard to conceive of someone who understood the meaning of “turn left”, and “jump right and turn left twice” (which the network gets right), but not that of “jump right and turn left” (one of the examples the network missed). In the “jump” experiment, the network could only correctly decode two composite cases, both starting with the execution of primitive “jump”, conjoined with a different action: “jump and run opposite right”, “jump and walk around left thrice”. It is instructive to look at the representations that the network induced for various commands in the latter experiment. Table 1 reports the 5 nearest neighbours for a sample of commands. Command similarity is measured by the cosine between the final encoder hidden state vectors, and computed with respect to all commands present in the training set. “Run” is provided as an example primitive command for which the model has been exposed to the full composed paradigm in training. As one would expect, “run” is close .72 .65 .64 jump twice walk and walk run and walk .19 .16 .12 .63 walk opposite right and walk look right and walk .63 walk right and walk .11 .12 to the other primitive commands (“look”, “walk”), as well as to short conjoined commands that contain primitive “run” as one of the conjuncts (we observe a similar pattern for the non-degenerate “jump” representation induced in Experiment 1). Instead, since “jump” had a different training distribution than the other primitive commands, the model does not capture its similarity to them, as shown by the very low cosines of its nearest commands. Since it fails to establish a link to other basic commands, the model does not generalize modifier application from them to “jump”. Although “run twice” is similar to (conjunctions of) other primitive tasks composed with “twice”, “jump twice” is isolated in representational space, and its (far) nearest neighbours look arbitrary. We tested here systematicity in its purest form: the model was only exposed to “jump” in isolation, and asked to bootstrap to its compositional paradigm based on the behaviour of other primitive commands such as “walk”, “look” and “run”. Although we suspect humans would not have problems with this setup, it arguably is too opaque for a computational model, which could lack evidence for “jumping” being the same sort of action as “walking”. Suppose we give the network some evidence that “jumping” composes like “walking” by showing a few composed “jump” command during training. Is the network then able to generalize to the full composed paradigm? This question is answered in Figure 5. Again, the new primitive command (and its compositions) were over-sampled during training to make up 10% of all presentations. Here, even when shown 4 different composed commands with “jump” at training time, the network does not generalize to other composed commands (4.1% correct). Weak generalization starts appearing when the network is presented 8 composed tasks in training (15.3%), and significant generalization (still far from perfect) shows up when the training set contains 16 and especially 32 distinct composed commands (70.2% and 89.9%, respectively). We conclude that the network is not failing to generalize simply because, in the original setup, it had no evidence that “jump” should Generalization without Systematicity Figure 5. Zero-shot generalization after adding the primitive “jump” and some compositional jump commands. The overall-best network was trained on different numbers of composed “jump” commands (x-axis), and generalization was measured on new composed “jump” commands (y-axis). Each bar shows the mean over 5 runs with varying training commands along with the corresponding ±1 SEM. behave like the other commands. On the other hand, the runs with more composed examples confirm that, as we found in Experiment 1, the network does display powerful generalization abilities. Simply, they do not conform to the “all-or-nothing” rule-based behaviour we would expect from a systematically compositional device–and, as a consequence, they require more positive examples to emerge. Experiment 4: Compositionality in machine translation Our final experiment is a proof-of-concept that our findings are more broadly applicable; that is, the limitations of recurrent networks with regards to systematic compositionality extend beyond SCAN to other sequence-to-sequence problems such as machine translation. First, we trained our standard seq2seq code on short (≤ 9 words) English-French sentence pairs that begin with English phrases such as “I am,” “he is,” “they are,” and their contractions (randomly split with 10,000 for training and 1180 for testing).6 An informal hyperparameter search led us to pick a LSTM with attention, 2 layers of 400 hidden units, and 0.05 dropout. With these hyperparameters and the same training procedure used for the SCAN tasks (Section 3), the network reached a respectable 28.6 BLEU test score after 100,000 steps. Second, to examine compositionality with the introduction of a new word, we trained a fresh network after adding 1,000 repetitions of the sentence “I am daxy” (fr. “je suis daxiste”) to the training data (the BLEU score on the original test set dropped less than 1 point).10 We tested this network by embedding “daxy” into the following constructions: “you are daxy” (“tu es daxiste”), “he is daxy” (“il est daxiste”), “I am not daxy” (“je ne suis pas daxiste”), “you are not daxy” (“tu n’es pas daxiste”), “he is not daxy” (“il n’est pas daxiste”), “I am very daxy” (“je suis très daxiste”), “you are very daxy” (“tu es très daxiste”), “he is very daxy” (“il est très daxiste”). During training, the model saw these constructions occurring with 22 distinct predicates on average (limiting the counts to perfect matches, excluding, e.g., “you are not very X”). Still, the model could only get one of the 8 translations right (that of “he is daxy”). For comparison, for the adjective “tired”, which occurred in 80 different constructions in the training corpus, our network had 8/8 accuracy when testing on the same constructions as for “daxy” (only one of which also occurred with “tired” in the training set). Although this is a small-scale machine translation problem, our preliminary result suggests that models will similarly struggle with systematic compositionality in larger data sets, when adding a new word to their vocabulary, in ways that people clearly do not. 5. Discussion In the thirty years since the inception of the systematicity debate, many have tested the ability of neural networks to solve tasks requiring compositional generalization, with mixed results (e.g., Christiansen & Chater, 1994; Marcus, 1998; Phillips, 1998; Chang, 2002; van der Velde et al., 2004; Botvinick & Plaut, 2006; Wong & Wang, 2007; Bowers et al., 2009; Botvinick & Plaut, 2009; Brakel & Frank, 2009; Frank et al., 2009; Frank, 2014; Bowman et al., 2016). However, to the best of our knowledge, ours is the first study testing systematicity in modern seq2seq models, and our results confirm the mixed picture. On the one hand, Experiment 1 and the “turn left” results in Experiment 3 show how standard recurrent models can reach very high zeroshot accuracy from relatively few training examples. We would like to stress that this is an important positive result, showing in controlled experiments that seq2seq models can make powerful zero-shot generalizations. Indeed, an interesting direction for future work is to understand what are, precisely, the generalization mechanisms that subtend the networks’ success in these experiments. After all, human language does have plenty of generalization patterns that are not easily accounted for by algebraic compositionality (see, e.g., Goldberg, 2005). On the other hand, the same networks fail spectacularly when the link between training and testing data is dependent on the ability to extract systematic rules. This can 10 Results do not change if, instead of repeating “I am daxy” 1,000 times, we insert it 100 times; with just 1 or 10 occurrences of this sentence in the training data, we get 0/8 translations right. Generalization without Systematicity be seen as a trivial confirmation of the basic principle of statistical machine learning that your training and test data should come from the same distribution. But our results also point to an important difference in how humans and current seq2seq models generalize, since there is no doubt that human learners can generalize to unseen data when such data are governed by rules that they have learned before. Crucially, the training data of experiments 2 and 3 provide enough evidence to learn composition rules affording the correct generalizations. In Experiment 2, the training data contain examples of all modifiers and connectives that are needed at test time for producing longer action sequences. In Experiment 3, the usage of modifiers and connectives is illustrated at training time by their application to many combinations of different primitive commands, and, at test time, the network should apply them to a new command it encountered in isolation during training. We thus believe that the fundamental component that current models are missing is the ability to extract systematic rules from the training data. A model that can abstract away from surface statistical patterns and operate in “rule space” should extract rules such as: translate(x and y) = translate(x) translate(y); translate(x twice) = translate(x) translate(x). Then, if the meaning of a new command (translate(“jump”)) is learned at training time, and acts as a variable that rules can be applied to, no further learning is needed at test time. When represented in this more abstract way, the training and test distributions are quite similar, even if they differ in terms of shallower statistics such as word frequency. How can we encourage seq2seq models to extract rules from data rather than exploiting shallower pattern recognition mechanisms? We think there are several, non-mutually exclusive avenues to be explored. First, in a “learning-to-learn” approach (Thrun & Pratt, 1997; Risi et al., 2009; Finn et al., 2017, a.o.), a network can be exposed to a number of different learning environments regulated by similar rules. An objective function requiring successful generalization to new environments might encourage learners to discover the shared general rules. Another promising approach is to add more structure to the neural networks. Taking inspiration from recent neural program induction and modular network models (e.g., Reed & de Freitas, 2016; Hu et al., 2017; Johnson et al., 2017), we could endow RNNs with a set of manually-encoded or (ideally) learned functions for interpreting individual modifiers, connectives, and primitives. The job of the RNN would be to learn how to apply and compose these functions as appropriate for interpreting a command. Similarly, differentiable stacks, tapes, or random-access memory (e.g., Joulin & Mikolov, 2015; Graves et al., 2016) could equip seq2seq models with quasi-discrete memory structures, enabling separate storage of variables, which in turn might encourage abstract rule learning. Other solutions, such as ad-hoc copying mechanisms or special ways to initialize the embeddings of novel words, might help to solve the SCAN tasks specifically. But they are unlikely to help with more general seq2seq problems. It remains to be seen, of course, if any of our proposed approaches offer a truly general solution. Nonetheless, we see all of the suggestions as directions worth pursuing, perhaps simultaneously and in complementary ways, with the goal of achieving human-like systematicity on SCAN and beyond. Given the astounding successes of seq2seq models in challenging tasks such as machine translation, one might argue that failure to generalize by systematic composition indicates that neural networks are poor models of some aspects of human cognition, but it is of little practical import. However, systematicity is an extremely efficient way to generalize. Once a person learns the new English adjective “daxy”, he or she can immediately produce and understand an infinity of sentences containing it. The SCAN experiments and a proof-of-concept machine translation experiment (Experiment 4) suggest that this ability is still beyond the grasp of state-of-the-art neural networks, likely contributing to their striking need for very large training sets. These results give us hope that neural networks capable of systematic compositionality could greatly benefit machine translation, language modeling, and other applications. ACKNOWLEDGMENTS We thank Germán Kruszewski, Adam Liska, Tomas Mikolov, Kristina Gulordava, Gemma Boleda, Michael Auli, Matt Botvinick, Sam Bowman, Jeff Dean, Jonas Gehring, David Grangier, Angeliki Lazaridou, Gary Marcus, the CommAI team and the audiences at the Facebook Dialogue Summit, 2017 Paris Syntax and Semantics Colloquium and CLIC-it 2017 for feedback and advice. The SCAN tasks are based on the navigation tasks available at: https:// github.com/facebookresearch/CommAI-env References Bahdanau, Dzmitry, Cho, Kyunghyun, and Bengio, Yoshua. Neural machine translation by jointly learning to align and translate. In Proceedings of ICLR Conference Track, San Diego, CA, 2015. Published online: http://www. iclr.cc/doku.php?id=iclr2015:main. Baroni, Marco. Distributions in text. In Lüdeling, Anke and Kytö, Merja (eds.), Corpus Linguistics: An International Handbook, volume 2, pp. 803–821. Mouton de Gruyter, Berlin, Germany, 2009. Bojar, Ondřej, Chatterjee, Rajen, Federmann, Christian, Generalization without Systematicity Graham, Yvette, Haddow, Barry, Huck, Matthias, Jimeno Yepes, Antonio, Koehn, Philipp, Logacheva, Varvara, Monz, Christof, Negri, Matteo, Neveol, Aurelie, Neves, Mariana, Popel, Martin, Post, Matt, Rubino, Raphael, Scarton, Carolina, Specia, Lucia, Turchi, Marco, Verspoor, Karin, and Zampieri, Marcos. Findings of the 2016 Conference on Machine Translation. In Proceedings of the First Conference on Machine Translation, pp. 131–198, Berlin, Germany, 2016. Botvinick, Matthew and Plaut, David. Short-term memory for serial order: A recurrent neural network model. Psychological Review, 113(2):201–233, 2006. Dong, Li and Lapata, Mirella. Language to logical form with neural attention. In Proceedings of ACL, pp. 33–43, Berlin, Germany, 2016. Elman, Jeffrey. Finding structure in time. Cognitive Science, 14:179–211, 1990. Finn, Chelsea, Abbeel, Pieter, and Levine, Sergey. Modelagnostic meta-learning for fast adaptation of deep networks. In Proceedings of ICML, pp. 1126–1135, Sydney, Australia, 2017. Fodor, Jerry and Lepore, Ernest. The Compositionality Papers. Oxford University Press, Oxford, UK, 2002. Botvinick, Matthew and Plaut, David. Empirical and computational support for context-dependent representations of serial order: Reply to Bowers, Damian, and Davis (2009). Psychological Review, 116(4):998–1002, 2009. Fodor, Jerry and Pylyshyn, Zenon. Connectionism and cognitive architecture: A critical analysis. Cognition, 28: 3–71, 1988. Bowers, Jeffrey, Damian, Markus, and David, Colin. A fundamental limitation of the conjunctive codes learned in PDP models of cognition: Comment on Botvinick and Plaut (2006). Psychological Review, 116(4):986–997, 2009. Frank, Stefan. Getting real about systematicity. In Calvo, Paco and Symons, John (eds.), The architecture of cognition: Rethinking Fodor and Pylyshyn’s systematicity challenge, pp. 147–164. MIT Press, Cambridge, MA, 2014. Bowman, Samuel R, Manning, Christopher D, and Potts, Christopher. Tree-Structured Composition in Neural Networks without Tree-Structured Architectures. arXiv preprint, 2016. Frank, Stefan, Haselager, Willem, and van Rooij, Iris. Connectionist semantic systematicity. Cognition, 110(3):358– 379, 2009. Brakel, Philémon and Frank, Stefan. Strong systematicity in sentence processing by simple recurrent networks. In Proceedings of CogSci, pp. 1599–1604, Amsterdam, the Netherlands, 2009. Calvo, Paco and Symons, John (eds.). The architecture of cognition: Rethinking Fodor and Pylyshyn’s systematicity challenge. MIT Press, Cambridge, MA, 2014. Chang, Frankilin. Symbolically speaking: a connectionist model of sentence production. Cognitive Science, 26: 609–651, 2002. Chomsky, Noam. Syntactic Structures. Mouton, Berlin, Germany, 1957. Christiansen, Morten and Chater, Nick. Generalization and connectionist language learning. Mind & Language, 9(3): 273–287, 1994. Chung, Junyoung, Gulcehre, Caglar, Cho, Kyunghyun, and Bengio, Yoshua. Empirical evaluation of gated recurrent neural networks on sequence modeling. In Proceedings of the NIPS Deep Learning and Representation Learning Workshop, Montreal, Canada, 2014. Published online: http://www.dlworkshop.org/ accepted-papers. Gehring, Jonas, Auli, Michael, Grangier, David, Yarats, Denis, and Dauphin, Yann. Convolutional sequence to sequence learning. https://arxiv.org/abs/1705. 03122, 2017. Goldberg, Adele. Constructions at work: The nature of generalization in language. Oxford University Press, Oxford, UK, 2005. Graves, Alex, Wayne, Greg, Reynolds, Malcolm, Harley, Tim, Danihelka, Ivo, Grabska-Barwinska, Agnieszka, Colmenarejo, Sergio Gomez, Grefenstette, Edward, Ramalho, Tiago, Agapiou, John, Badia, Adrià Puigdomènech, Hermann, Karl Moritz, Zwols, Yori, Ostrovski, Georg, Cain, Adam, King, Helen, Summerfield, Christopher, Blunsom, Phil, Kavukcuoglu, Koray, and Hassabis, Demis. Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626): 471–476, 2016. Herzig, Jonathan and Berant, Jonathan. Neural semantic parsing over multiple knowledge-bases. In Proceedings of ACL (Volume 2: Short Papers), pp. 623–628, Vancouver, Canada, 2017. Hochreiter, S and Schmidhuber, J. Long short-term memory. Neural computation, 9:1735–1780, 1997. Generalization without Systematicity Hu, Ronghang, Andreas, Jacob, Rohrbach, Marcus, Darrell, Trevor, and Saenko, Kate. Learning to reason: End-toend module networks for visual question answering. In Proceedings of ICCV, pp. 804–813, Venice, Italy, 2017. Jia, Robin and Liang, Percy. Data recombination for neural semantic parsing. In Proceedings of ACL, pp. 12–22, Berlin, Germany, 2016. Johnson, Justin, Hariharan, Bharath, van der Maaten, Laurens, Hoffman, Judy, Fei-fei, Li, Zitnick, C Lawrence, and Girshick, Ross. Inferring and Executing Programs for Visual Reasoning. In International Conference on Computer Vision, 2017. Joulin, Armand and Mikolov, Tomas. Inferring algorithmic patterns with stack-augmented recurrent nets. In Proceedings of NIPS, Montreal, Canada, 2015. Kingma, Diederik P and Welling, Max. Efficient GradientBased Inference through Transformations between Bayes Nets and Neural Nets. In International Conference on Machine Learning (ICML 2014), 2014. Lake, Brenden, Ullman, Tomer, Tenenbaum, Joshua, and Gershman, Samuel. Building machines that learn and think like people. Behavorial and Brain Sciences, 2017. In press. Marcus, Gary F. Rethinking Eliminative Connectionism. Cognitive Psychology, 282(37):243–282, 1998. Marcus, Gary F. The Algebraic Mind: Integrating Connectionism and Cognitive Science. MIT Press, Cambridge, MA, 2003. Mikolov, Tomas, Joulin, Armand, and Baroni, Marco. A Roadmap towards Machine Intelligence. arXiv preprint, 2016. URL http://arxiv.org/abs/ 1511.08130. Montague, Richard. Universal Grammar. Theoria, 36:373– 398, 1970. Phillips, Steven. Are feedforward and recurrent networks systematic? analysis and implications for a connectionist cognitive architecture. Connection Science, 10:137–160, 1998. Reed, Scott and de Freitas, Nando. Neural programmerinterpreters. In Proceedings of ICLR, San Juan, Puerto Rico, 2016. Published online: http://www.iclr. cc/doku.php?id=iclr2016:main. Risi, Sebastian, Vanderbleek, Sandy, Hughes, Charles, and Stanley, Kenneth. How novelty search escapes the deceptive trap of learning to learn. In Proceedings of GECCO, pp. 153–160, Montreal, Canada, 2009. Sutskever, Ilya, Vinyals, Oriol, and Le, Quoc. Sequence to sequence learning with neural networks. In Proceedings of NIPS, pp. 3104–3112, Montreal, Canada, 2014. Thrun, Sebastian and Pratt, Lorien (eds.). Learning to Learn. Kluwer, Dordrecht, 1997. van der Velde, Frank, , van der Voort van der Kleij, Gwendid, and de Kamps, Marc. Lack of combinatorial productivity in language processing with simple recurrent networks. Connection Science, 16(1):21–46, 2004. Williams, Ronald J and Zipser, David. A Learning Algorithm for Continually Running Fully Recurrent Neural Networks. Neural Computation, 1:270–280, 1989. Wong, Francis and Wang, William. Generalisation towards combinatorial productivity in language acquisition by simple recurrent networks. In Proceedings of KIMAS, pp. 139–144, Waltham, MA, 2007. Wu, Yonghui, Schuster, Mike, Chen, Zhifeng, Le, Quoc, Norouzi, Mohammad, Macherey, Wolfgang, Krikun, Maxim, Cao, Yuan, Gao, Qin, Macherey, Klaus, Klingner, Jeff, Shah, Apurva, Johnson, Melvin, Liu, Xiaobing, Kaiser, Lukasz, Gouws, Stephan, Kato, Yoshikiyo, Kudo, Taku, Kazawa, Hideto, Stevens, Keith, Kurian, George, Patil, Nishant, Wang, Wei, Young, Cliff, Smith, Jason, Riesa, Jason, Rudnick, Alex, Vinyals, Oriol, Corrado, Greg, Hughes, Macduff, and Dean, Jeffrey. Google’s neural machine translation system: Bridging the gap between human and machine translation. http:// arxiv.org/abs/1609.08144, 2016.	2cs.AI
arXiv:1604.07227v1 [math.GR] 25 Apr 2016 Alternate modules are subsymplectic Clément Guérin∗ April 26, 2016 Abstract In this paper, an alternate module (A, φ) is a finite abelian group A with a Z-bilinear application φ : A × A → Q/Z which is alternate (i.e. zero on the diagonal). We shall prove that any alternate module is subsymplectic, i.e. if (A, φ) has a Lagrangian of cardinal n then there exists an abelian group B of order n such that (A, φ) is a submodule of the standard symplectic module B × B ∗ . Keywords : alternate modules, symplectic modules, Lagrangians, finite abelian groups. Mathematics Subject Classification : 20K01. Contents 1 Definitions and statement of the result 2 2 Lagrangians in alternate modules 5 3 Any alternate module is subsymplectic 10 3.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.2 Proof of the fundamental lemma . . . . . . . . . . . . . . . . . 15 3.3 Proof of the theorem . . . . . . . . . . . . . . . . . . . . . . . 20 4 Remarks 20 IRMA Institut de Recherche Mathématiques avancée, University of Strasbourg, 7 rue René Descartes, F-67084 Strasbourg , France, e-mail : guerin@math.unistra.fr ∗ 1 1 Definitions and statement of the result Before stating the result, we give some standard definitions along with some properties regarding the alternate modules. The concept is defined by Wall in [Wal64] and by Tignol, Amitsur in [AT86]. Among the results, we will establish here, some might be directly read in those references and also in [TW15]. Let A be an abelian group. If we are given an application φ : A × A → Q/Z, then we say that (A, φ) is a bilinear module if φ is bilinear when A is seen as a Z-module. By definition, the underlying group of the bilinear module (A, φ) is A and the associated bilinear form to (A, φ) is φ. In the sequel, all bilinear modules are assumed to be finite. Let (A, φ) a bilinear module, we say that (A, φ) is an alternate module if φ(a, a) = 0 for all a ∈ A. If (A, φ) is an alternate module, then for all a, b ∈ A : 0 = φ(a + b, a + b) = φ(a, a) + φ(a, b) + φ(b, a) + φ(b, b) = φ(a, b) + φ(b, a) since (A, φ) is alternate by ”bilinearity” since (A, φ) is alternate Whence for all a, b ∈ A, φ(a, b) = −φ(b, a). We will sum up this by saying that if (A, φ) is an alternate module then, in particular φ is anti-symmetric. Let (A, φ) be an alternate module and a, b ∈ A, we say that a is orthogonal to b if φ(a, b) = 0. Since A is alternate, any element of A is orthogonal to itself and the relation of orthogonality is symmetric by anti-symmetry of the bilinear form φ. Let (A, φ) be an alternate module and S a subset of A. We define S ⊥ the orthogonal of S as the subset of elements in A which are orthogonal to S. One can directly check that S ⊥ is a subgroup of A. Let (A, φ) be an alternate module and B a subgroup of A then the bilinear module induced by A on B : (B, φB×B ) is also an alternate module called the induced submodule on B by φ. 2 Let (A1 , φ1 ) and (A2 , φ2 ) be two alternate modules, we define the orthogonal sum of those modules by : ⊥ (A1 , φ1 ) ⊕ (A2 , φ2 ) := (A, φ) With A := A1 ⊕ A2 and φ((a1 , a2 ), (b1 , b2 )) := φ1 (a1 , b1 ) + φ2 (a2 , b2 ). It is a straightforward verification that (A, φ) is an alternate module, verifying ⊥ A1 ≤ A⊥ 2 , A2 ≤ A1 and the induced submodule by φ on Ai is (Ai , φi ). Conversly, if (A, φ) is an alternate module and B, C two subgroups of A such that A = B ⊕ C with B ≤ C ⊥ , then (A, φ) can be written as the orthogonal sum of the two induced modules on B and C by φ. Let A be a finite module, we define A∗ , the dual of A as the group A∗ := Mor(A, Q/Z). A non-trivial result (direct consequence of the classification of finite abelian groups) states that A is isomorphic to A∗ as a group. Remark that the isomorphism between A and A∗ is not canonical. Let (A, φ) be an alternate module, we define the dual application associated to (A, φ) as : φ∗ : A −→ A∗ a 7−→ b 7→ φ(a, b) The kernel Kφ of (A, φ) is, by definition, the kernel Ker(φ∗ ) of the dual application φ∗ . In other words, Kφ = A⊥ i.e. the set of elements in A which are orthogonal to any element of A. Let (A, φ) be an alternate module, we say that (A, φ) is a symplectic module if its kernel Kφ is trivial (i.e. the application φ is non-degenerate). To any alternate module (A, φ), we can associate a symplectic module (B, φB ) setting B := A/Kφ and : φB (a1 mod Kφ , a2 mod Kφ ) := φ(a1 , a2 ) The fact that (B, φB ) is well defined and is a symplectic module is clear. This symplectic module will always be denoted (A/Kφ , φ) and called the symplectic module associated to (A, φ). An important example of symplectic module is the following : 3 Example 1. Let B be an abelian group. We define A := B × B ∗ and : φ: A × A −→ Q/Z ((a1 , ψ1 ), (a2 , ψ2 )) 7−→ ψ2 (a1 ) − ψ1 (a2 ) Then (A, φ) is a symplectic module. In the sequel, such symplectic module will always be denoted B × B ∗ (the underlying bilinear form being the form φ as above). Proof. Clearly (A, φ) is an alternate module. Let (a, ψ) be in Kφ then for all b∈A: φ((a, ψ), (b, 0)) = 0 ⇒ ψ(b) = 0 This implies that ψ is the trivial morphism, i.e. ψ = 0. Furthermore for all ψ ′ ∈ Mor(A, Q/Z) : φ((a, 0), (0, ψ ′)) = 0 ⇒ ψ ′ (a) = 0 Since A is isomorphic to Mor(A, Q/Z) (although the isomorphism is not canonical) this implies that a = 0. Therefore Kφ is trivial and (A, φ) is a symplectic module. Let (A, φ) be an alternate module and S a subset of A. We say that S is isotropic if S ⊆ S ⊥ . Furthermore, if S = S ⊥ then S is a subgroup of A and is called a Lagrangian of (A, φ). In general (see proposition 3) the cardinal of a Lagrangian of (A, φ) only depends on \|A\| and \|Kφ \|. We shall p denote nA,φ := \|A\|\|Kφ \| the cardinal of any Lagrangian of (A, φ). In the example 1, the subgroups B and B ∗ of A are both Lagrangians of the module (A, φ). Let (A, φ) and (A′ , φ′ ) be alternate modules, we say that (A, φ) and (A′ , φ′ ) are isometric if there exists a group isomorphism g : A → A′ verifying φ = g ∗φ′ . The isometric relation is clearly an equivalence relation. Now, we come to the most important definition of this paper. Let (A, φ) be an alternate module. We say that (A, φ) is subsymplectic if there exists an abelian group B of order nA,φ such that (A, φ) is included in the standard symplectic module B × B ∗ . The result we are going to prove here is : Theorem. Any alternate module is subsymplectic. 4 Whereas the classification of symplectic modules is easily delt with, it seems hopeless to do the same thing for alternate modules in general. The statement that alternate modules are subsymplectic appears as the closest possible result to a classification of alternate modules. In another paper, we will use this result to classify conjugacy classes of centralizers of irreducible subgroups in P SL(n, C). In the second section, we shall characterize and study the Lagrangians in alternate modules. In the third section we will prove the theorem. In the last section, we make some remarks about this proof and alternate modules in general. 2 Lagrangians in alternate modules Some results in this section have been proven in other papers, we shall only give the references. We begin with an elementary propoosition : Proposition 1. Let (A, φ) be a symplectic module and B a subgroup of A then : 1. \|B\|\|B ⊥ \| = \|A\| and (B ⊥ )⊥ = B. 2. If the induced submodule on B by (A, φ) is a symplectic module then the induced submodule on B ⊥ by (A, φ) is also a symplectic module, Furthermore ⊥ A = B ⊕ B⊥ Proof. See proposition 2.2 of [AT86] or lemma 1 in [Wal64]. Whereas, alternate modules are very hard to classify, the following property implies that any symplectic module is isometric to one constructed as in the example 1. In particular, the set of symplectic modules is quite rigid. Corollary 1. Let (A, φ) be a symplectic module. There exists B ≤ A such that (A, φ) is isomorphic to B × B ∗ . 5 Proof. See lemma 7 in [Wal64] or theorem 4.1 in [AT86]. This construction can be seen as the analogous of a symplectic base for (A, φ) (i.e. a ”base” of B as a finite abelian group and its ”dual base” in B ∗ ). A direct consequence of this corollary is that any two symplectic modules (A, φ) and (A′ , φ′) are isometric if and only if A and A′ are isomorphic as groups. It is, somehow, surprising that such symplectic base exists for those finite modules (in a similar fashion than for bilinear forms on K-vector spaces). Whereas K-vector spaces endowed with an alternate form (not necessarily symplectic) are always the sum of the kernel of the form and a non-degenerate complementary, this is not the case for alternate modules in general. Indeed, one can consider the following counter-example : Counter-example 1. Let A := Z/2 × Z/4 × Z/8. We denote e1 := (1, 0, 0), e2 := (0, 1, 0) and e3 := (0, 0, 1). Then, on the base e1 , e2 , e3 , we define the alternate bilinear form φ on A with its associated matrix :   0 1/2 1/2 Mφ := 1/2 0 −1/4 1/2 1/4 0 Then the kernel Kφ of (A, φ) is not a direct factor of A, in particular, A is not the sum of a non-degenerate module and its kernel. Proof. With a straightforward computation, one can check that Kφ = he1 + 2e2 +2e3 i is isomorphic to Z/4. We remark that 2(e1 +2e2 +2e3 ) = 4e3 ∈ Kφ . Furthermore if (x, y, z) ∈ A then 4(x, y, z) = (0, 0, 4z) so any element a ∈ A of order 8 will verify that 4a = 4e3 ∈ Kφ . In particular, Kφ is not a direct factor of A. This little example shows that the kernel of (A, φ) is not necessarily a direct factor of A. Furthermore, one can check that A cannot be written as the orthogonal sum of strictly smaller submodules. In some sense, it is irreducible, this suggests that if there is a classification of alternate modules (in some sense, we would like to have one in the sequel), it should be complicated. Another corollary of the proposition 1 : Corollary 2. Let p (A, φ) be a symplectic module, then any Lagrangian L of A is of cardinal \|A\|. 6 Proof. By definition, L = L⊥ , applying the first point of proposition 1, we get \|L\|2 = \|A\|. In their paper, [AT86], J.-P. Tignol and S.A. Amitsur are interested in Lagrangian of symplectic modules, they apply this to division algebra (cf. [AT85]). We need to understand, Lagrangians of alternate modules. We recall some results about Lagrangians in this case : Proposition 2. Let (A, φ) be an alternate module. The Lagrangians of A are exactly the maximal elements for the inclusion within the set of isotropic subsets of A. In other words a subset L in A is a Lagrangian of A if and only if for any K ′ isotropic subset of A containing L, K ′ = L. Proof. Let L be an isotropic subset of A. Assume that L is a Lagrangian, then if K ′ is an isotropic subset of A containing L, then L⊥ ⊇ K ′ since K ′ is orthogonal to K ′ whence to L. Since L is a Lagrangian, L = L⊥ whence K ′ ⊆ L, since K ′ ⊇ L by assumption L = K ′ and L is thus maximal among the isotropic subsets of A. Assume that L is not a Lagrangian. Since L is isotropic, L ≤ L⊥ and there exists a ∈ L⊥ which does not belong to L. We define K ′ := hL, ai, clearly L ≤ K ′ and K ′ 6= L, furthermore L being isotropic is orthogonal to a by assumption, whence K ′ is an isotropic subspace of A. Whence L is not maximal among the isotropic subsets of A. Let (A, φ) be an alternate module then the set of isotropic subspaces of A is not empty (it contains the trivial group) and finite (since A is finite). Therefore, there exists a maximal element for the inclusion. Whence any alternate module admits a Lagrangian. Furthermore, if K is any isotropic subset, there will always be a maximal isotropic subset containing it, therefore any isotropic subset is contained in a Lagrangian. In particular, for any a ∈ A, there exists a Lagrangian containing a (applying what we have just done to K = hai). The next proposition is a generalization of corollary 2 : Proposition 3. Let (A, φ) p be an alternate module. For any Lagrangian L of A, the cardinal of L is \|A\|\|Kφ \|. 7 Proof. We define (A/Kφ , φ) to be the associated symplectic module to (A, φ) and π be the projection of A onto A/Kφ . Clearly the application π leads to a bijective correspondance between isotropic subgroups of A containing Kφ and isotropic subgroups of A/Kφ . Since maximal elements among isotropic subsets of A are isotropic subgroups of A containing Kφ , it follows that π induces a bijective correspondance between Lagrangians in A and Lagrangians in A/Kφ (by sending L to π(L)). In particular, any Lagrangian L of A is of the form π −1 (L) where L is a Lagrangian in A/Kφ . \|L\| = \|π −1 (L)\| = \|Kφ \|\|L\| s = \|Kφ \| = q \|A\| by corollary 2 \|Kφ \| \|A\|\|Kφ \| We see that the cardinal of Lagrangians is constant in alternate modules. However, the isomorphism class of Lagrangians may vary (see also [AT86]). Example 2. Let (A, φ) be the alternate module defined in counter-example 1. Then A = Z/2 × Z/4 × Z/8, and if we denote e1 := (1, 0, 0), e2 = (0, 1, 0) and e3 = (0, 0, 1) in A then the subgroups L1 := he3 , e1 + 2e2 i and L2 := he1 , 2e2 , 2e3 i are Lagrangians in A and : L1 is isomorphic to Z/2 × Z/8 L2 is isomorphic to Z/2 × Z/2 × Z/4 In particular, L1 is not isomorphic to L2 . Proof. From the counter-example 1, \|Kφ \| = 4, hence the cardinal of a La√ grangian in (A, φ) is 2 · 4 · 8 · 4 = 16 using proposition 3. We directly check that both L1 and L2 are isotropic. 8 Furthermore e1 + 2e2 is of order 2 and not in he3 i, therefore L1 is isomorphic to he1 + 2e2 i × he3 i. Whence L1 is isomorphic to Z/2 × Z/8. In particular, L1 is isotropic and has the cardinal of a Lagrangian, by maximality of Lagrangians (proposition 2) it is Lagrangian. Finally, L2 is clearly isomorphic to Z/2 × Z/2 × Z/4. Likewise, the L2 is isotropic of cardinal 16, by maximality of Lagrangians (proposition 2) it is Lagrangian. In general, it is much more convenient to work with abelian p-groups than finite abelian groups. We remark that : Proposition 4. Let (A, φ) be an alternate module such that (A, φ) is isometric to the orthogonal sum of (A1 , φ1 ) and (A2 , φ2) : ⊥ (A, φ) = (A1 , φ1 ) ⊕ (A2 , φ2 ) If both (A1 , φ1 ) and (A2 , φ2 ) are subsymplectic then (A, φ) is subsymplectic. Proof. We begin by showing that Kφ = Kφ1 ⊕ Kφ2 . Clearly, Kφ contains Kφ1 and Kφ2 and both are in direct sum. Finalely, if k = (l1 , l2 ) ∈ Kφ ≤ A1 ⊕ A2 then for all a1 ∈ A1 , we have : 0 = φ(k, (a1 , 0)) = φ1 (l1 , a1 ) + φ2 (l2 , 0) = φ1 (l1 , a1 ) This is true for any a1 ∈ A1 so l1 ∈ Kφ1 . Likewise, l2 ∈ Kφ2 , whence Kφ = Kφ 1 ⊕ K φ 2 . It follows that \|Kφ \| = \|Kφ1 \|\|Kφ2 \|, since \|A\| = \|A1 \|\|A2 \|, we have : nA,φ = nA1 ,φ1 nA2 ,φ2 (1) By hypothesis, for i = 1, 2, there exists Bi , an abelian gorup of order nAi ,φi such that (Ai , φi ) ≤ Bi × Bi∗ . Whence : ⊥ (A, φ) = (A1 , φ1 ) ⊕ (A2 , φ2 ) ⊥ ≤ (B1 × B1∗ ) ⊕ (B2 × B2∗ ) ≤ B × B ∗ where B := B1 × B2 9 We have included (A, φ) (as an alternate module) in B × B ∗ , since \|B\| = \|B1 \|\|B2 \| = nA1 ,φ1 nA2 ,φ2 = nA,φ by equation 1, it follows, by definition, that (A, φ) is subsymplectic. As a result : Corollary 3. Let (A, φ) be an alternate module then, we denote, for p any prime dividing \|A\|, Sp the unique p-Sylow of A. If for all p dividing \|A\|, the induced submodule on Sp by φ is subsymplectic then (A, φ) is subsymplectic. Proof. It is a well known result that if p1 , . . . , pr are the r distinct primes dividing \|A\| then, as a group, A is isomorphic to the product of its p-Sylows A = Sp 1 ⊕ · · · ⊕ Sp r Furthermore, if 1 ≤ i 6= j ≤ r and ai ∈ Spi , aj ∈ Spj then φ(ai , aj ) is of order dividing some power of pi and some power of pj , it follows that φ(ai , aj ) is of order 1, i.e. is trivial. In particular, for i 6= j, the induced modules Spi and Spj are orthogonal to each other. It follows that : ⊥ ⊥ (A, φ) = Sp1 ⊕ · · · ⊕ Spr The corollary is a direct consequence of this decomposition and proposition 4. With those classical properties, we will show the theorem. 3 Any alternate module is subsymplectic Basically, the idea is to make an induction on the cardinal of the kernel. The following lemma is the major step of the proof : Fundamental lemma. Let p be a prime number and (A, φ) an alternate module which is not symplectic with A a p-group. Then there exists an alternate module (Â, φ̂) such that (A, φ) is a submodule of (Â, φ̂), \|A\| < \|Â\| and nA,φ = nÂ,φ̂ . 10 In the sequel, we will say that (Â, φ̂) is an extension of modules of (A, φ) via the inclusion ιA if ιA : A → Â is an inclusion of alternate modules, i.e. ιA∗ φ̂ = φ and ιA is an inclusion of modules. The extension will be with constant Lagrangians if nÂ,φ̂ = nA,φ . The first subsection gathers some preliminary results. The second subection is the proof of the fundamental lemma. For subsections 3.1 and 3.2, p is a fixed prime number and any alternate module (A, φ) is assumed to have its underlying group A to be a p-group. In the third subsection, we prove the theorem. 3.1 Preliminaries The next lemma is interesting in itself, it generalizes the second point of proposition 1. Lemma 1. Let (A, φ) be an alternate module. If there exists a submodule B of (A, φ) such that B is a symplectic module then : ⊥ (A, φ) is isometric to B ⊕ B ⊥ Proof. Let (A, φ) be an alternate module and B be a submodule of A which is symplectic. Then B ⊥ is also a submodule of (A, φ). Since B is symplectic, B ⊥ ∩ B is trivial. It follows that B ⊕ B ⊥ is a subgroup of A. By definition, B is orthogonal to B ⊥ it follows that : ⊥ B ⊕ B ⊥ is a submodule of (A, φ) Let π be the canonical projection of A on A/Kφ . Since B is symplectic, B ∩ Kφ is trivial and B is isomorphic to B := π(B). Using the first point of proposition 1 : ⊥ \|A/Kφ \| \|B\| \|A\| = \|Kφ \|\|B\| B = 11 ⊥ ⊥ Clearly, π −1 (B ) = B ⊥ , it follows that \|B ⊥ \| = \|Kφ \|\|B \|. Finally, we have that : ⊥ \|B ⊕ B ⊥ \| = \|B\|\|B ⊥ \| ⊥ = \|B\|\|Kφ \|\|B \| \|A\| = \|B\|\|Kφ \| \|Kφ \|\|B\| = \|A\| ⊥ Since B ⊕ B ⊥ is a submodule of (A, φ) and both modules have the same cardinal, we have that : ⊥ (A, φ) is isometric to B ⊕ B ⊥ The next lemma gives a sufficient condition for an extension of modules to be with constant Lagrangians : Lemma 2. Let p be a prime number, (Â, φ̂) be an alternate module and A be a subgroup of Â of index p. We denote (A, φ) the induced submodule on A by φ̂. If Kφ is not included in Kφ̂ then nA,φ = nÂ,φ̂ . Proof. Let ê be an element of Â which is not in A. Since [Â : A] is of cardinal p, any element in â ∈ Â can uniquely be written as : Define : â = λê + a where 0 ≤ λ ≤ p − 1 and a ∈ A f: Then for all k ∈ Kφ : Kφ −→ Q/Z k 7−→ φ̂(ê, k) p · f (k) = p · φ̂(ê, k) = φ̂(pê, k) = φ(pê, k) since pê and k are in A = 0 since k ∈ Kφ and pê ∈ A 12 As a result Im(f ) is a finite subgroup of Q/Z of exponent dividing p. Whence Im(f ) is either trivial or the unique cyclic subgroup of order p in Q/Z. If Im(f ) is trivial, then Kφ is orthogonal to ê, since Kφ is orthogonal to A and Â is generated by ê and A, it follows that Kφ is orthogonal to Â, whence Kφ ≤ Kφ̂ . By assumption, this is a contradiction. It follows that Im(f ) is cyclic of cardinal p. Let x = λê + a ∈ Kφ̂ where 0 ≤ λ ≤ p − 1 and a ∈ A. Let k0 ∈ Kφ such that f (k0 ) is of order p. Since x ∈ Kφ̂ , we have φ̂(x, k0 ) = 0. On the other hand : φ̂(x, k0 ) = λφ̂(ê, k0 ) + φ̂(a, k0 ) = λf (k0 ) + φ(a, k0 ) = λf (k0 ) + 0 since k0 ∈ Kφ It follows that λf (k0 ) = 0. Since f (k0 ) is of order p, we have that p divides λ and since 0 ≤ λ ≤ p − 1, we end up with λ = 0 which implies that x ∈ A. As a result, we have shown that Kφ̂ ≤ Kφ . Finally : nÂ,φ̂ = q \|Â\|\|Kφ̂ \| p p\|A\|\|Kφ \| =q [Kφ : Kφ̂ ] s p nA,φ = [Kφ : Kφ̂ ] Let L be a Lagrangian of A, then L is isotropic in Â and by proposition 2, there exists a Lagrangian L̂ of Â containing L. Using Lagrange’s theorem, we have that \|L\| divides \|L̂\| and using proposition 3, nA,φ divides nÂ,φ̂ . It q follows that [Kφp:K ] is a positive integer which cannot but be equal to one φ̂ (because p is a prime number). In particular, we have that nÂ,φ̂ = nA,φ . 13 The last lemma in those preleminaries allows us to construct extension of modules in a fairly simple way. Lemma 3. Let p be a prime number, (A, φ) an alternate module where A is a p group. Assume that there exist r + 1 non-trivial elements e, e1 , . . . , er in A such that : A = hei × he1 i × · · · × her i Assume furthermore that for any 1 ≤ i ≤ r the order of φ(e, ei ) is strictly lesser than the order of ei then there exists an alternate module (Â, φ̂) such that Â = hêi × he1 i × · · · × her i where ê is of order p times the order of e and the inclusion : ιA : A −→ Â e 7−→ pê ei 7−→ ei for 1 ≤ i ≤ r is an inclusion of submodules. Proof. Let Â be hêi × he1 i × · · · × her i where ê is of order p times the order of e. Clearly, the mapping : ιA : A −→ Â e 7−→ pê ei 7−→ ei for 1 ≤ i ≤ r defines an injective morphism of groups. In order to define the bilinear form φ̂, it suffices to define it on a generating set of Â × Â. Define : φ̂(ei , ej ) := φ(ei , ej ) for all 1 ≤ i, j ≤ r We also define φ̂(ê, ê) := 0. Finally, remark that the group Q/Z is pdivisible (i.e. any element admits a p-th root). Then for any 1 ≤ i ≤ r, define λi to be one element verifying pλi = φ(e, ei ). Remark that the order of λi thus defined divides p times the order of φ(e, ei ). By assumption, this divides the order of ei and also divides the order of ê which is p times the order of e. It follows that λi divides both the order of ê and ei . Whence the following equations : 14 φ̂(ei , ej ) := φ(ei , ej ) for all 1 ≤ i, j ≤ r φ̂(ê, ê) := 0 φ̂(ê, ei ) := λi φ̂(ei , ê) := −λi allow us to define a group morphism on Â which is clearly an alternate module. Let ψ be the induced module on A via ιA by φ̂ : ψ(ei , ej ) = φ̂(ei , ej ) = φ(ei , ej )for all 1 ≤ i, j ≤ r ψ(e, ej ) = φ̂(pê, ej ) = pλj = φ(e, ej ) for all 1 ≤ j ≤ r Whence ψ = φ on A × A (since they are alternate forms). In particular we have shown that (A, φ) is a submodule of (Â, φ̂) via ιA . We have the tools to prove the fundamental lemma. 3.2 Proof of the fundamental lemma We recall first the statement of the fundamental lemma : Fundamental lemma. Let p be a prime number and (A, φ) an alternate module which is not symplectic with A a p-group. Then there exists an alternate module (Â, φ̂) such that (A, φ) is a submodule of (Â, φ̂), \|A\| < \|Â\| and nA,φ = nÂ,φ̂ . Proof. We prove the lemma doing a strong induction on the cardinal of the module. Let pk be such that the lemma is true for any alternate module M with \|M\| < pk . Let (A, φ) be an alternate module which is not symplectic and A be of cardinal pk . There are two different cases to consider : Case 1 : Kφ is not included in p · A 15 By the classification of abelian p-groups, decompose the group A as a product of cyclic subgroups : A = he0 i × he1 i × · · · × her i (2) Denote di the order of ei , then we may assume that di−1 divides di for 1 ≤ i ≤ r. Let k0 ∈ Kφ which is not in p · A. By the decomposition 2, we may write k0 = r X i=0 αi ei where α0 , . . . , αr ∈ N Since k0 is not in p · A there exists at least one coefficient among the α0 , . . . , αr which is not divisible by p. Denote i0 to be one index such that αi0 is not divisible by p. Define the group Â := Z/p × A. Let γ be a generator of Z/p. Then we define an alternate form φ̂ on Â by : φ̂(ei , ej ) := φ(ei , ej ) for 0 ≤ i, j ≤ r φ̂(γ, ei ) := 0 for 0 ≤ i ≤ r and i 6= i0 φ̂(ei , γ) := 0 for 0 ≤ i ≤ r and i 6= i0 φ̂(γ, γ) := 0 1 φ̂(γ, ei0 ) := p φ̂(ei0 , γ) := − 1 p Clearly, φ̂ is an alternate form on Â, furthermore, the inclusion of A in Â given by a is mapped to (0, a) clearly leads to an inclusion of modules. Finally : 16 φ̂(γ, k0 ) = r X αi φ̂(γ, ei ) i=0 = αi0 φ̂(γ, ei0 ) since γ is orthogonal to ei where i 6= i0 αi = 0 6= 0 in Q/Z since p does not divide αi0 p Hence we found an element in Kφ which is not in the kernel of Kφ̂ . Applying lemma 2, we have nA,φ = nÂ,φ̂ , hence (Â, φ̂) contains (A, φ), has the same cardinality of Lagrangians and is strictly greater than A. In this case, we have constructed the extension with constant Lagrangians (without using the induction hypothesis). Case 2 : Kφ is included in p · A Write : A = hei × he1 i × · · · × her i (3) Denote d the order of e, di the order of ei , then we may assume that d divides d1 and di divides di+1 for 1 ≤ i ≤ r − 1. Assume that there exists 1 ≤ i ≤ r such that the order of φ(e, ei ) is equal to di (the order of ei ), since the order of φ(e, ei ) divides d, the order of e (by bilinearity), it follows that di divides d. Since we also have d divides di by definition, we have two elements e and ei in A whose order is equal to d and such that φ(e, ei ) is of order d. As a result, the submodule B generated by e and ei is isometric to Z/d × (Z/d)∗ . In particular B is a submodule of (A, φ) which is symplectic, by lemma 1, it follows that : ⊥ (A, φ) = B ⊕ B ⊥ Let (B ⊥ , ψ) be the induced submodule on B ⊥ by φ. By induction hypothesis, there exists a module (C, ψ̂) such that nB⊥ ,ψ = nC,ψ̂ with \|B ⊥ \| < \|C\| and (B ⊥ , ψ) is a submodule of (C, ψ̂). Denote : ⊥ (Â, φ̂) := B ⊕ (C, ψ̂) 17 ⊥ Clearly A = B ⊕ B ⊥ is a submodule of (Â, φ̂) and \|A\| < \|Â\|. Remark that Kφ = Kψ and Kφ̂ = Kψ̂ since B is symplectic. Whence : q \|A\|\|Kφ \| p q = \|B\| \|B ⊥ \|\|Kψ \| p = \|B\|nB⊥ ,ψ p = \|B\|nC,ψ̂ by construction of (C, ψ̂) p q = \|B\| \|C\|\|Kψ̂ \| q = \|Â\|\|Kφ̂ \| nA,φ = = nÂ,φ̂ In this case we can also construct an extension (Â, φ̂) of (A, φ) with constant Lagrangians. Assume now that for 1 ≤ i ≤ r, the order of φ(e, ei ) strictly divides the order of di . Then we are in the condition of applications of lemma 3. Define Â = hêi × he1 i × · · · × her i where ê is of order p times the order of e and the inclusion : ιA : A −→ Â e 7−→ pê ei 7−→ ei for 1 ≤ i ≤ r then we have defined on Â a bilinear form φ̂ such that ιA∗ φ̂ = φ. We recall that we are in the second case where Kφ ≤ p · A. Denote π the canonical projection of A onto A/Kφ . Since Kφ ≤ p · A, it follows that π −1 (p · (A/Kφ )) = p · A. In particular (since e is clearly not an element of p · A) it follows that π(e) is not in p · (A/Kφ ). We know that A/Kφ is isomorphic to B × B ∗ by corollary 1. Denote g1 , . . . , gs a ”base” of B and g1∗ , . . . , gs∗ the corresponding ”dual base” of B ∗ and write : 18 π(e) = r X γi gi + γi∗ gi∗ i=1 Since π(e) is not in p · (A/Kφ ), it follows that at least one of the γi or γi∗ is not divisible by p. Up to exchanging gi and gi∗ , we may assume that γi∗0 is not divisible by p. Let pk (with k ≥ 1) be the order of gi0 and a0 ∈ A such that π(a0 ) = gi0 . Since gi0 is of order pk , pk · a0 ∈ Kφ . Furthermore : φ(e, pk−1 · a0 ) = φ(π(e), pk−1 · π(a0 )) by definition of (A/Kφ , φ) = pk−1 φ(π(e), gi0 ) r X k−1 γi φ(gi , gi0 ) + γi∗ φ(gi∗, gi0 ) =p = i=1 k−1 ∗ p γi0 φ(gi∗0 , gi0 ) Remark that, by definition φ(gi∗0 , gi0 ) is of order pk , since γi∗0 is not divisible by p : φ(e, pk−1 · a0 ) = pk−1 γi∗0 φ(gi∗0 , gi0 ) is of order p, whence not trivial. Remark now that, in (Â, φ̂) we have : φ̂(ê, pk · a0 ) = φ̂(p · ê, pk−1 · a0 ) by Z-bilinearity = φ(e, pk−1 · a0 ) since p · ê = ιA (e) and ιA∗ φ̂ = φ 6= 0 in Q/Z In particular, pk · a0 is an element of Kφ which is not in Kφ̂ . Applying lemma 2, we have that nA,φ = nÂ,φ̂ . Now that the fundamental lemma is proven, we shall see that the theorem follows by an easy induction on the cardinal of the kernel of the alternate module. 19 3.3 Proof of the theorem Recalling that the result we want to prove is the following theorem : Theorem. Any alternate module is subsymplectic. Proof. By corollary 3, it suffices to prove it for alternate modules (A, φ) where A is a p-group and p is a prime number. Let us show the following by induction on k : Let (A, φ) be an alternate module which is a p-group with \|Kφ \| = pk then (A, φ) is subsymplectic. If k = 0 then (A, φ) ispsymplectic. Therefore, by corollary 1, it is isometric to B × B ∗ where \|B\| = \|A\| = nA,φ . It follows that (A, φ) is subsymplectic by definition. If k > 0 then (A, φ) is an alternate module which is not symplectic (its kernel is not trivial) whence, by the fundamental lemma, there exists an alternate module (Â, φ̂) such that (A, φ) is a submodule of (Â, φ̂), \|A\| < \|Â\| and nA,φ = nÂ,φ̂ . We have n2A,φ = n2Â,φ̂ so \|A\|\|Kφ \| = \|Â\|\|Kφ̂ \|, since \|A\| < \|Â\| we have that \|Kφ̂ \| < \|Kφ \| = pk . It follows that we can apply the induction hypothesis to (Â, φ̂) and there exists an abelian group B of order n := nÂ,φ̂ such that : (Â, φ̂) ≤ B × B ∗ Since (A, φ) is a submodule of (Â, φ̂), we finally have included the module (A, φ) in B × B ∗ where \|B\| = n = nÂ,φ̂ = nA,φ . It follows that (A, φ) is subsymplectic by definition. 4 Remarks The proof is constructive. Indeed, it explicitely gives the construction (by induction) of a symplectic module B × B ∗ containing (A, φ) with \|B\| = nA,φ . Basically, in order to define an algorithm computing B, one needs an 20 algorithm that, given (A, φ) computes its kernel and a ”diagonalization” of its associated symplectic module (A/Kφ , φ). Remark that, given (A, φ), an alternate module, there might exist B1 6= B2 , both of cardinal nA,φ such that (A, φ) is included in Bi × Bi∗ for i = 1, 2 : Counter-example 2. Let A be the abelian group Z/2×Z/4 endowed with the trivial bilinear form. It follows that nA,φ = \|A\|. Let B1 := A and B2 := Z/8. Then we have two inclusions : ιA,1 : A −→ B1 × B1∗ (u, v) 7−→ ((u, v), (0, 0)) and ιA,2 : A −→ B2 × B2∗ (u, v) 7−→ (4u, 2v) Those two inclusions are inclusions of submodules (when Bi × Bi∗ is endowed with its natural structure of symplectic form, defined in example 1). Furthermore, it is clear that they are extensions with constant Lagrangians. However B1 6= B2 . As we have stated in the introduction of this paper, the theorem will be used in another paper in order to give a classification of centralizer of irreducible subgroups of P SL(n, C). An interesting thing to have would be a classification of alternate modules. Let (A, φ) be an alternate module, we say that (A, φ) is indecomposable if, whenever (A, φ) is the orthogonal sum of (A1 , φ1 ) and (A2 , φ2 ) then (A1 , φ1 ) or (A2 , φ2) is ({0}, 0). Since any alternate module is clearly the sum of indecomposable ones by induction, if we want to classify alternate modules, we only need to characterize indecomposable ones. This leads to the following question : Question. Let (A, φ) be an indecomposable alternate module. Does it follow that rk(A) ≤ 3? Does it follow that rk(A/Kφ ) ≤ 2? We computed many examples and this conjecture seems to be verified. However we still do not have a proof for it. 21 Acknowledgements All computations performed on examples leading to this result have been performed using the GAP system [GAP]. I would like to thank my thesis advisor Olivier Guichard, whose suggestions have been very helpful. I would also like to thank Gaël Collinet who encouraged me to fully investigate this topic and Jean-Pierre Tignol for pointing out some helpful references. Finally, I would like to thank my fellow Ph.D student, Mohamad Maassarani for carefully listening to my proofs and his helpful comments. References [AT85] S.A. Amitsur, J.-P. Tignol. Kummer subfields of Malcev-Neumann division algebras. Israel Journal of mathematics, Vol. 50, Nos 1-2, 1985. [AT86] S.A. Amitsur, J.-P. Tignol. Symplectic modules. Israel Journal of mathematics, Vol. 54, No 3, 1986. [GAP] The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.8.3 ; 2016, (http://www.gap-system.org). [TW15] J.-P. Tignol, A. Wadsworth. Value Functions on Simple Algebras, and Associated Graded Rings. Springer Monographs in Mathematics, 2015. [Wal64] C.T.C. Wall. Quadratic forms on finite groups, and related topics. Topology, Vol. 2:281-298, 1964. 22	4math.GR
arXiv:1206.5112v1 [cs.PL] 22 Jun 2012 Lucretia — a type system for objects in languages with reflection∗ Viviana Bono1 , Marcin Benke2 , and Aleksy Schubert2 1 Dipartimento di Informatica dell’Università di Torino† 2 University of Warsaw‡ May 1, 2012 Abstract Object-oriented scripting languages such as JavaScript or Python gain in popularity due to their flexibility. Still, the growing code bases written in the languages call for methods that make possible to automatically control the properties of the programs that ensure their stability in the running time. We propose a type system, called Lucretia, that makes possible to control the object structure of languages with reflection. Subject reduction and soundness of the type system with respect to the semantics of the language is proved. 1 Introduction Scripting object-oriented languages such as JavaScript, Python, Perl or Ruby became very popular [Sof12] due to their succinctness of program expressions and flexibility of idioms [WNV09]. These languages optimise the programmer time, rather than the machine time and are very effective when small programs are constructed [Pre00, WEF+ 09]. The advantages of the languages that help when a short programs are developed can be detrimental in case big applications are created. Short code that has clear advantages when small programs are constructed [Ous98] can provide less information for a person that looks for clues on what particualr piece of code is doing (and this is one of the most frequent activities ∗ See http://www.youtube.com/watch?v=IuezNswtRfo This work was partly supported by the PRIN 12008 DISCO grant. ‡ This work was partly supported by the Polish government grant no N N206 355836. † 1 1 INTRODUCTION 2 during software maintenance tasks, see e.g. [Sas96, KMCA06]). Moreover, the strong invariants a programmer can rely on in case of statically typed languages are no longer valid as type of a particular value can change with each instruction and mostly in an uncontrolled way with each function call in the program. Still, systems that handle complex and critical tasks such as Swedish premium pension system [Ste01] are deployed and continuously maintained. Therefore, it is desirable to study methods which help programmers in understanding their code and keeping it consistent. Therefore we propose a type system that handles dynamic features of scripting languages and can help in understanding of the existing code. The type system we propose is presented in the style where many type annotations must be present to guarantee correct typechecking. In particular, we require that function declarations as well as labelled instructions are decorated with their types. What is more, the type system is so strong that the reconstruction of the types is (most probably) undecidable. However, the system is designed in such a way that it will primarily be used on top of a type inference algorithm and admits a wide range of type inference heuristics that are not complete, but provide correct type annotations for wide range of programs. In addition, we would like to encourage programmers to add type annotations to their programs since the annotations can serve as an important documentation of the code invariants assumed by developers in the program construction phase. The inherent feature of the scripting languages is that a running time type of a particular variable may change in the course of program execution. This problem can be solved to some extent through introduction of single assignment form for local variables. Still, this cannot be applied to object fields. The natural semantics of programs is such that the fields change and the efforts to circumvent this principle will most probably result in a complicated solution. Therefore, the statement that statically describes evolution of the running time type of a variable cannot be just a type name, but must reflect the journey of the running time type throughout the control flow graf of the program. It would be very inconvenient to repeat the structure of the whole control flow graf at each variable in the program. Therefore, it makes more sense to describe the type of each visible variable at each statically available program point. This is the approach we follow in this paper and present a type system which is inspired by the works on type-and-effect systems [MM09, GL86]. However, we present our typing judgements in a slightly different manner, i.e. one where the effect is described by a pair of constraints that express the update performed by a particular instruction. In a sense the pair together with the typed expression can be viewed as a triple in a variant of Hoare logic. The constraints we use to express the type information for an object are approximations of the actual types by the matching constraints similar to the ones in [BB99]. 2 MODEL LANGUAGE AND ITS SEMANTICS 3 This paper is structured as follows. The language in which we model the scripting languages is presented in Section 2. The type system is presented in Section 3. Its properties are demonstrated in Section 4. At the end we present related work in Section 5 and conclude in Section 6. 2 Model language and its semantics The syntax of expressions we work with is presented in Fig 1. The figure presents not only the raw syntax of programs, but also the syntax of evaluation contexts that represent ifnormation on the particular point in the program the semantical reduction rules elaborate. Moreover, we present here the full syntax of the expressions that may show up in the evolution of the expression being reduced. This is reflected by the fact that we permit locations to occur in the expressions while the intent is they are not directly visible in programs in the source form. Definition 2.1 (Types) The type information associated with an expression of the language we consider here is combined of two items. One is a representation of the actual type and the second is the constraint expression which approximates the shape of the type. These components are generated with help of the following grammar: tb t tr Ψ ::= ::= ::= ::= c \| X \| tb,1 ∨ tb,2 \| [t1 , . . . , tn ; Ψ1 ] =⇒ [tn+1 ; Ψ2 ] tb \| t1 ∨ t2 \| ⊥ {l : t} \| {} X <# tr , Ψ \| ∅ where c is a type constant from ConstT , X ∈ VT is a type variable. The set Types is the set of elements generated from t, the set Typesb is the set of elements generated from tb , the set Typesr is the set of elements generated from tr , and the set Constr is the set of elements generated from Ψ. The set of objects O is the set of partial functions from identifiers to values. The set of stores H is the set of finite partial functions from the set of locations to the set of objects. The set of evaluation contexts EC is the set of sequences of basic evaluation contexts as presented in Figure 1. The set of expressions EX is the set of expressions as given in Figure 1. The semantics is defined by the relation ; which relates triples (σ, E, e) ∈ H × EC × EX . Let s be a (partial) function we write s[x/v] to denote function such that s[x/v](x) = v and s[x/v](y) = s(y) for y 6= x and y ∈ dom(s). When an evaluation context is a sequence E1 , E2 , . . . , En , we write it as E1 hE2 h. . . En h•i . . .ii. 2 MODEL LANGUAGE AND ITS SEMANTICS identifiers labels locations constants function value values objects fields list function expression expressions x n l c vf v o Lf ef e ::= . . . (identifiers). . . ::= . . . (identifiers). . . ::= . . . ::= func(x1 , · · · , xn ).t {e} ::= x \| c \| vf \| l ::= {} \| {Lf } ::= n : v \| n : v, L ::= x \| vf ::= v \| let x = e1 in e2 \| ef (e1 , · · · , en ) \| opn (e1 , · · · , en ) \| if (e1 ) then e2 else e3 \| ifhasattr (x, a) then e1 else e2 \| break n e \| n.t {e} \| new \| x1 .x2 = v \| x1 .x2 basic evaluation con- E ::= • \| let x = E in e texts \| vf (v1 , · · · , vi−1 , E, ei+1 , · · · , en ) \| opn (v1 , · · · , vi−1 , E, ei+1 , · · · , en ) \| if (E) then e2 else e3 \| break n E \| n.t {E} stores σ ::= · \| (l, o)σ Figure 1: Syntax of λM 4 2 MODEL LANGUAGE AND ITS SEMANTICS ~ ~ (Let-Enter) σ, Ehlet x = e1 in e2 i ; σ, Ehlet x = he1 i in e2 i ~ ~ (Let) σ, Ehlet x = hvi in ei ; σ, Ehe[x/v]i ~ (Op-Enter) σ, Ehop n (e1 , · · · , ek )i ; ~ σ, Ehop n (he1 i, · · · , ek )i ~ (Op-Prop) σ, Ehop (v , n 1 · · · , vi−1 , hvi i, ei+1 , · · · , ek )i ; ~ σ, Ehop n (v1 , · · · , vi−1 , vi , hei+1 i, · · · , vk )i ~ (Op-Back) σ, Ehopn (v1 , · · · , hvk i)i ; ~ σ, Ehop n (v1 , · · · , vk )i ~ ~ (Op-Eval) σ, Ehop (v , n 1 · · · , vk )i ; σ, Ehδn (opn , v1 , · · · , vk )i (β-Enter) (β-Prop) (βv) (If-Enter) (If-True) (If-False) ~ σ, Ehfunc(x 1 , · · · , xn ).t {e}(e1 , · · · , ek )i ; ~ σ, Ehfunc(x1 , · · · , xn ).t {e}(he1 i, · · · , ek )i ~ σ, Ehfunc(x 1 , · · · , xn ).t {e}(v1 , · · · , vi−1 , hvi i, ei+1 , · · · , en )i ; ~ σ, Ehfunc(x 1 , · · · , xn ).t {e}(v1 , · · · , vi−1 , vi , hei+1 i, · · · , en )i ~ σ, Ehfunc(x1 , · · · , xn ).t {e}(v1 , · · · , hvk i)i ; ~ σ, Ehe[x 1 /v1 , · · · , xn /vn ]i ~ (e1 ) then e2 else e3 i ; σ, Ehif ~ (he1 i) then e2 else e3 i σ, Ehif ~ (htruei) then e2 else e3 i ; σ, Ehe ~ 2i σ, Ehif ~ ~ 3i σ, Ehif (hfalsei) then e2 else e3 i ; σ, Ehe ~ ~ 1i (Ifhtr-True) σ, Ehifhasattr (l, a) then e1 else e2 i ; σ, Ehe when a ∈ dom(σ(l)) ~ ~ 2i (Ifhtr-False) σ, Ehifhasattr (l, a) then e1 else e2 i ; σ, Ehe when a 6∈ dom(σ(l)) ~ ~ (Brk-Enter) σ, Ehn.t {e}i ; σ, Ehn.t {hei}i ~ 1 hn.t {E ~ 2 hbreak n vi}i ; σ, E ~ 1 hvi, (Brk-P) σ, E ~2 when n does not occur as label in E ~ ~ (Lbl-Pop) σ, Ehn.t {hvi}i ; σ, Ehvi (New) (SetRef) (Deref) ~ ~ σ, Ehnewi ; (l, ∅)σ, Ehli l fresh ~ ~ σ, Ehl.f = vi ; σ[l/σ(l)[f /v]], Ehvi ~ ~ σ, Ehl.f i ; σ, Ehσ(l)(f )i when f ∈ dom(σ(l)) Figure 2: Semantic rules of λM 5 3 THE TYPE SYSTEM 6 When the relation holds between triples (σ, E, e), (σ 0 , E 0 , e0 ) we write the relation as σ, Ehei ; σ 0 , E 0 he0 i. The precise semantical rules are given in Figure 2 We do not model inheritance and multiple inheritance of the scripting languages. We assume that this can be viewed as a notational shortcut for direct presentation of objects and therefore we cannot model all features of the class models (e.g. the method updates in classes in Python). Constructs of the language The language we propose has object-oriented features such as object creation (new), field reference (the dot notation x1 .x2 ), and field modification (x1 .x2 = v). In addition we have object introspetion operation available through ifhasattr (·, ·) then · else · construct. The flow of information is controlled by let x = e1 in e2 expression that in addition to creation of a local variable x makes possible to execute e1 and e2 in sequence. The traditional split of control flow depending on a computed condition is realised by if (·) then · else ·. The exceptional flow can be realised by our labelled instructions combined with break statements. At last loops can be organised by recursively defined functions. Notably we use types in labelled instructions and in function declarations. These types can in fact be omitted since they do not play part in the respective reduction rules. The possible errors that may result by execution of an operation on a value that the operation is not prepared to take are present in our system through appropriate definition of δn that can check the types of the values supplied in the arguments. 3 The type system Typing judgements are of the form Ψ1 ; Γ; Σ ` e : tb ; Ψ2 where Γ is an environment (i.e. mapping from variables to tb ), Σ is a mapping from locations to type variables VT , Ψi for i = 1, 2 is a set of constraints of the form X <# t. The intended meaning of such a judgment is: evaluating e in the environment Γ with types of locations that match the mapping Σ and with the store satisfying Ψ1 leads to a value of type tb and store satisfying Ψ2 . Record update The information about the record being updated is represented by a constraint on a variable X. The rule for record update comes in two variants, depending on whether the constraint mentions the field being updated. Ψ1 ; Γ; Σ ` x : X; Ψ1 Ψ1 ; Γ; Σ ` e : t2 ; Ψ2 , X <# {a : t02 , l : s} Ψ1 ; Γ; Σ ` x.a = e : t2 ; Ψ2 , X <# {a : t2 , l : s} (update-old) 3 THE TYPE SYSTEM 7 If (we know that) the record contains the field a that is to be updated, we forget the old value, hence its type (t02 ) is ignored; the store after the update reflects the new value of a. Ψ1 ; Γ; Σ ` x : X; Ψ1 Ψ1 ; Γ; Σ ` e : t2 ; Ψ2 , X <# {l : s} a 6∈ l Ψ1 ; Γ; Σ ` x.a = e : t2 ; Ψ2 , X <# {a : t2 , l : s} (update-fresh) otherwise the constraint for X is amended with information about the new value of a. Record access We can access a field a provided the constraint on the record guarantees that it has the field a and it is of a definite type (it cannot be ⊥). ` u Ψ 3 X <# {a : u, l : t} (access) Ψ; Γ, x : X; Σ ` x.a : u; Ψ A type is definite (` u) if it is a type constant, type variable, function type or a disjunction of definite types. This is defined inductively as follows: `c `X ` t1 ` t2 ` t1 ∨ t2 ` [t1 , . . . , tn ; Ψ1 ] =⇒ [u; Ψ2 ] Conditional instruction The basic way to split the processing depending on some value is to use a conditional instruction. This instruction is typed in our system in the following way. Ψ; Γ; Σ ` e1 : bool ; Ψ1 Ψ1 ; Γ; Σ ` e2 : u; Ψ2 Ψ1 ; Γ; Σ ` e3 : u; Ψ2 (if) Ψ; Γ; Σ ` if (e1 ) then e2 else e3 : u; Ψ2 The typing of the branches need to be typable into the same type. However, the constraints in each of them can be in principle different. Therefore we need a rule to weaken the constraint so that they result in the same set. For this we need an operation that merges type constraint from two branches. This is defined in the following way. • Ψ, X <# {l : t} ] Ψ0 , X <# {l0 : t0 } = Ψ ] Ψ0 , X <# {l : t} ] {l0 : t0 } • Ψ]∅=Ψ 3 THE TYPE SYSTEM 8 • ∅]Ψ=Ψ • {l0 : t0 , l : t} ] {l0 : t00 , l0 : t0 } = {l0 : t0 ∨ t00 , l00 : t00 } where {l00 : t00 } = {l : t} ] {l0 : t0 } • {} ] {l0 : t0 } = {l0 : t0 ∨ ⊥} • {l0 : t0 } ] {} = {l0 : t0 ∨ ⊥} This combination of two sets of constraints need to be incorporated into type derivation and this can be done using the following weakening rule. Ψ; Γ; Σ ` e : t; Ψ1 (weaken-]) Ψ; Γ; Σ ` e : t; Ψ1 ] Ψ2 Object structure introspection Dynamic languages split the processing not only depending on some condition defined in terms of the actual values, but also depending on types of expressions. Therefore, we introduce the construct that performs appropriate check and provide a typing rule that handles it. Ψ[X ← {a}]; Γ, x : X; Σ ` e1 : t; Ψ2 Ψ; Γ, x : X; Σ ` e2 : t; Ψ02 (ifhasattr) Ψ; Γ, x : X; Σ ` ifhasattr (x, a) then e1 else e2 : t; Ψ2 ] Ψ02 The typing rule must update the typing information available in the branches of the instruction. Type information in one branch takes as granted that the attribute a is present. We need not pass the information that the attribute a is missing in the other branch since the case that the actual value does not have the field a must be expressed in the type description by an alternative with ⊥. • (Ψ, X <# t)[X ← {a}] = Ψ, X <# (t[X ← {a}]), • {a : t1 ∨ · · · ∨ tn , l : u}[X ← {a}] = {a : ti1 ∨ · · · ∨ tik , l : u} where {ti1 , . . . , tik } = {ti \| ti 6= ⊥ and ti ∈ {t1 , . . . , tn }}. We assume that base types are implemented so that they have a distinguished field which makes possible to check for base type with ifhasattr (?, ?) then ? else ?. Variable and location access The information about a type of a variable is recorded in the type environment so we use it when a variable is referred in an expression. Ψ; Γ, x : t; Σ ` x : t; Ψ (v-access) 3 THE TYPE SYSTEM 9 Similarly, the information about locations is stored in the location environment Σ and we exploit it in an analogous way. Ψ; Γ; Σ, l : t ` l : t; Ψ (l-access) Function definition and call Whenever we want to type a function definition we must rely on the type annotation that is associated with the function. We have to add the type decoration to functions since they can be used in a recursive way. The inference of a type in such circumstances must rely on some kind of a fixpoint computation which is a difficult task. We assume here that the type is given in the typechecking procedure either by hand with help of programmer or by some kind of automatic type inference algorithm. The type we obtain for a function definition is just the type that is explicitly given in the function. However, we have to check in addition that the body of the function indeed obeys the declared type. Therefore, we assume that the formal parameters have the types expressed in the precondition of the function type and obey the constraints noted in the constraint set from the same precondition. We expect then that the resulting type is equal to the one in the postcondition of the type and the constraints match the constraints in the postcondition. It is worth noting that the constraints in the precondition and postcondition should take into account not only the formal parameters, but also the global variables that are visible in the function body. Notably, we type the function body so that the typing does not interfere with the typing of the function definition. This reflects the dynamic character of the language. In particular the running time type of a global variable may differ at the function definition point from the expected one so that the typing is impossible at that particular moment. However, the situation may be different at the execution site. Therefore, we defer the check of the compatibility of global variables to the function call site. t ≡ [t1 , . . . , tn ; Ψ1 ] =⇒ [u; Ψ2 ] Ψ1 ; Γ, x1 : t1 , . . . , xn : tn ; Σ ` e : u; Ψ2 Ψ; Γ; Σ ` func(x1 , · · · , xn ).t {e} : t; Ψ (fdecl) In case a real application of a function expression is made to arguments we have to check that the application indeed does not lead to type error in this case. Therefore, the typing rule for application must check that the type information for global variables at the call site (i.e. after all actual parameters of the call are elaborated) is in accordance with the type information available in the functional type. This is checked with the relation c . 3 THE TYPE SYSTEM 10 Definition 3.1 (c ) The rules that operate on constraint sets are as follows. Ψ c ∅ Ψ c Ψ0 t c t0 Ψ, X <# t c Ψ0 , X <# t0 They reduce the problem to the one on record types where the relation is defined as follows. t c t {l : t} c {} {a : u, l : t} c {l : t} u c u0 {a : u, l : t} c {a : u0 , l : t} u c u u1 c u1 ∨ u2 u2 c u1 ∨ u2 u1 c u01 u2 c u02 u1 ∨ u2 c u02 ∨ u01 Proposition 3.2 The relation c is reflexive and transitive. Proposition 3.3 For any constraint sets Ψ1 , Ψ2 we have Ψ1 c Ψ1 ] Ψ2 . Execution of a function body causes changes of values held in global variables. Such changes may give rise to changes in the types. This must by reflected and the type information must be updated accordingly. Therefore, we need an operation of such an update. Definition 3.4 (type information update ←-) For two sets of constraints Ψ1 , Ψ2 we define inductively the set of constraints Ψ1 ←- Ψ2 , in words constraints Ψ1 updated with Ψ2 , as follows. • Ψ ←- ∅ = Ψ, • Ψ, X <# t ←- Ψ0 , X <# t0 = (Ψ ←- Ψ0 ), X <# t0 , • Ψ ←- Ψ0 , X <# t0 = (Ψ ←- Ψ0 ), X <# t0 when X 6∈ dom(Ψ). It is important to observe that the update preserves at least the type information that is present in the constraint with which we make the update. Lemma 3.5 For any constraint sets Ψ1 , Ψ2 , Ψ2 c Ψ1 ←- Ψ2 3 THE TYPE SYSTEM 11 We can now present the rule for function application. This rule elaborates first all the expressions in the actual parameters and updates the constraints so that their side effects are taken into account and then weakens the result to match the constraints for the input part of the function type. Then the resulting set of constraint is just the set of constraints after the last argument updated with the constraint information contained in the result type of the function. Ψi ; Γ; Σ ` Ni : ti ; Ψi+1 for i = 1, . . . , n Ψn+1 c Ψ01 0 Ψ1 ; Γ; Σ ` ef : [t1 , . . . , tn ; Ψ1 ] =⇒ [u; Ψ02 ]; Ψ1 Ψ1 ; Γ; Σ ` ef (N1 , . . . , Nn ) : u; Ψn+1 ←- Ψ02 (fapp) Control flow with break instructions The language we propose includes nonlocal control flow instructions that may be used to simulate exceptions or jumps out of loops (recursive calls in our case). The labelled instruction resembles a call to an anonymous function with no parameters. We only have to remember type information that we want to achieve as the result of the execution. Note that the prospective break instructions can reside inside of a recursive function call. Therefore, we cannot assume that the type we give here explicitely is a result of a finitary collecting of a non-local typing information from all the break statements inside e. Ψ; Γ, n : [; ∅] =⇒ [t; Ψ2 ]; Σ ` e : t; Ψ02 Ψ02 c Ψ2 (label) Ψ; Γ; Σ ` n.[; ∅] =⇒ [t; Ψ2 ]{e} : t; Ψ2 In case we encounter break instruction we have to check that the expression the value of which it given as the result indeed obeys the expected result type of the labelled instruction surrounding the break. Ψ; Γ, n : [; ∅] =⇒ [t; Ψ2 ]; Σ ` e : t; Ψ02 Ψ02 c Ψ2 (break) Ψ; Γ, n : [; ∅] =⇒ [t; Ψ2 ]; Σ ` break n e : t0 ; Ψ0 Let-expression The let expression is our language statement that does the instruction sequencing operation. Therefore, we have to compute the constraints for the first step, then for the second one and at last combine them together as in the logical cut rule that forgets the middle formula. Ψ1 ; Γ; Σ ` e1 : t1 ; Ψ2 Ψ2 ; Γ, x : t1 ; Σ ` e0 : t; Ψ3 (let) Ψ1 ; Γ; Σ ` let x = e1 in e0 : t; Ψ3 4 PROPERTIES OF THE TYPE SYSTEM 12 Object creation The object creation rule should just introduce information that a new object is created. This is expressed by the fact the set of constraints is extended with information on an object that has no known to the type system fields. X fresh (new) Ψ; Γ; Σ ` new : X; X <# {}, Ψ Disjunction We also add a rule to weaken the assumptions about an expression. This is primarily necessary to obtain the subject reduction property. In case an if instruction is reduced to one of its branches the type system infers only information from one of the branches and the information from the other one must be added artificially. Ψ1 ; Γ; Σ ` e : ti ; Ψ2 i = 1 or 2 (weaken-∨) Ψ1 ; Γ; Σ ` e : t1 ∨ t2 ; Ψ2 4 Properties of the type system Soundness and subject reduction need to take the store into account. Hence we need to express the fact that a store instance satisfies certain constraints: σ; Σ \|= Ψ where σ is a store, Σ a mapping between locations and types, and Ψ a set of constraints of the form X <# t. Definition 4.1 (\|=) If c is a constant of type tc : σ; Σ \|= c : tc For function values: Ψ; ∅; σ ` f unc(~x).t{e} : t; Ψ σ; Σ \|= Ψ (\|= fun) σ; Σ \|= f unc(~x).t{e} : t σ; Σ \|= ∅ If l is a location: σ; Σ \|= Ψ σ, Σ \|= v : t (\|= obj) σ[l/{a : v, y : s}]; Σ(l : X) \|= X <# {a : t}, Ψ 4 PROPERTIES OF THE TYPE SYSTEM 13 in particular, for every object o: σ(l/o); Σ(l : X) \|= X <# {} For any Ψ, Ψ0 σ; Σ \|= Ψ (\|= ]) σ; Σ \|= Ψ ] Ψ0 For any t1 , t2 , σ; Σ \|= v : ti (\|= ∨) σ; Σ \|= v : t1 ∨ t2 Lemma 4.2 (soundness of c ) For any σ, Σ and constraint sets Ψ1 , Ψ2 , if σ, Σ \|= Ψ1 and Ψ1 c Ψ2 then σ, Σ \|= Ψ2 Proof: It is sufficient to prove that if σ, Σ \|= X <# o1 and o1 c o2 then σ, Σ \|= X <# o2 . We proceed first by induction on the derivation of \|=. The only nontrivial rules are (\|= obj) — again, the only (marginally) nontrivial case is when o1 c o2 was derived via the disjunction rule: {l : t} c {l : t0 } {a : u, l : t} c {a : u ∨ u0 , l : t0 } In this case we can derive σ, Σ \|= X <# o2 by inserting (\|= ∨) before the “a : u” premise of the final (\|= obj) rule in the original derivation. (\|= ]) — we have σ, Σ \|= X <# o1 , o2 = o1 ] o0 , o2 c o3 and need to prove σ, Σ \|= X <# o3 , which can be done by induction on the definition of o2 c o3 . Lemma 4.3 If σ; Σ(l : X) \|= X <# {a : t; f : s}, Ψ0 and ` t (i.e. t contains no ⊥), then (i) σ(l)(a) is defined (ii) σ; Σ \|= σ(l)(a) : t (iii) σ; Σ \|= Ψ0 4 PROPERTIES OF THE TYPE SYSTEM 14 Proof: By induction on the derivation of σ; Σ(l : X) \|= X <# {a : t; f : s}. If the last rule is (\|= obj) then the thesis follows immediately. On the other hand, if the last rule is (\|= ]), we have X <# {a : t, f : u}, Ψ0 = Ψ1 ] Ψ2 σ; Σ \|= Ψi (\|= ]) σ; Σ \|= Ψ1 ] Ψ2 By the definition of ], one of the following holds: (a) Ψ1 = X <# {a : t, f : u}, Ψ01 , X 6∈ dom(Ψ2 ) (b) Ψ2 = X <# {a : t, f : u}, Ψ02 , X 6∈ dom(Ψ1 ) (c) Ψ1 = X <# {l : s}, Ψ01 , Ψ2 = X <# {l0 : s0 }, Ψ02 If (a) or (b) holds, the thesis follows directly from the induction hypothesis. In the (c) case, by the assumption that ` t, we have a ∈ l, a ∈ l0 as well as t = t1 ∨ t2 with ` t1 and ` t2 . Moreover, either σ; Σ \|= X <# {a : t1 , f : u}, Ψ01 or σ; Σ \|= X <# {a : t2 , f 0 : u0 }, Ψ01 Then by the induction hypothesis σ(l)(a) is defined, and σ; Σ \|= Ψ0i σ; Σ \|= σ(l)(a) : ti for appropriate i. Thus by the rule (\|= ∨) we have σ; Σ \|= σ(l)(a) : t1 ∨ t2 which completes the proof. 2 Lemma 4.4 (value completeness) For any σ, Σ, Γ, Ψ and value v if σ, Σ \|= Ψ and σ, Σ \|= v : t then Ψ; Γ; Σ ` v : t; Ψ 4 PROPERTIES OF THE TYPE SYSTEM 15 Proof: By induction on the derivation of σ, Σ \|= v : t. For the constant and function rules, the thesis follows directly. On the other hand if the last rule used was σ; Σ \|= v : ti (\|= ∨) σ; Σ \|= v : t1 ∨ t2 then by the induction hypothesis Ψ; Γ; Σ ` v : ti ; Ψ, hence Ψ; Γ; Σ ` v : t1 ∨ t2 ; Ψ. Lemma 4.5 (Progress) If Ψ1 ; Γ; Σ ` e : t; Ψ2 and σ; Σ \|= e : t; Ψ1 then either e is a value, or σ, e ; σ 0 , e0 Proof: The only expressions which may get stuck are record access (in case of lack of a field) and function application (in case of arity mismatch). Reducibility of record access follows from the Lemma 4.3. Lemma 4.6 (value stability) If v is a value and Ψ1 ; Σ ` v : t; Ψ2 then Ψ1 c Ψ2 . Moreover, for any Ψ we have Ψ; Σ ` v : t; Ψ Proof: By induction on the derivation. The thesis follows directly for axioms and the function definition. For the disjunction rule use the induction hypothesis. For the ] rule the thesis follows from Proposition 3.3. Since v is a value, no other rules can be used in the derivation. Lemma 4.7 (weakening) If Ψ1 ; Γ; Σ ` e : t; Ψ2 and Γ ⊆ Γ0 , Σ ⊆ Σ0 then Ψ1 ; Γ; Σ0 ` e : t; Ψ2 Proof: By routine induction on derivation. Lemma 4.8 (subject reduction) Reduction preserves typings, i.e. if Ψ1 ; Γ; Σ ` e : t; Ψ2 σ; Σ \|= Ψ1 4 PROPERTIES OF THE TYPE SYSTEM 16 ~ ~ 0i σ, Ehei ; σ 0 , Ehe then there exist Σ0 , Ψ0 such that Σ ⊆ Σ0 and Ψ0 ; Γ; Σ ` e0 : t; Ψ2 σ 0 ; Σ0 \|= Ψ0 Moreover if σ = σ 0 then the thesis holds for Σ0 = Σ and Ψ0 = Ψ. Proof: By induction on the derivation. We now consider several cases depending on the last rule used Case e = new: Ψ; Γ; Σ ` new : X; X <# {}, Ψ If σ; Σ \|= Ψ1 then σ(l/{}); Σ(l : X) \|= X <# {}, Ψ Case e = l.a: ` u Ψ 3 X <# {a : u, f : t} Ψ; Γ; l : X, Σ ` l.a : u; Ψ ~ hl.xi ; σ, σ(l)(a). Since σ; l : X, Σ \|= Ψ then by Lemma 4.3, σ(l) We have σ, E has field a of type u: σ; Σ \|= σ(l)(a) : u Hence by Lemma 4.4 Ψ; Γ; l : X, Σ ` σ(l)(a) : u; Ψ Case e = if (e1 ) then e2 else e3 : Ψ; Γ; Σ ` e1 : bool ; Ψ1 Ψ1 ; Γ; Σ ` e2 : u; Ψ2 Ψ1 ; Γ; Σ ` e3 : u; Ψ2 Ψ; Γ; Σ ` if (e1 ) then e2 else e3 : u; Ψ2 ~ hei reduces to σ, E ~ he0 i where e0 is either e2 or e3 . In If e1 is a value then σ, E both cases we have Ψ1 ; Γ; Σ ` e0 : t; Ψ2 σ 0 ; Σ \|= Ψ2 ~ he1 i ; σ 0 , E ~ he0 i. By induction hypothesis, there are Σ0 , Ψ0 Otherwise, σ, E 1 such that Ψ0 ; Γ; Σ ` e01 : t; Ψ2 4 PROPERTIES OF THE TYPE SYSTEM 17 σ 0 ; Σ0 \|= Ψ0 By the typing rule for if and Lemma 4.7, we have Ψ0 ; Γ; Σ0 ` if (e1 ) then e2 else e3 : t; Ψ2 Case e = func(x1 , · · · , xn ).t {eb }(v1 , · · · , vk ) : Ψ01 ; Γ, ~x : ~t; Σ ` eb : u; Ψ02 Ψ1 ; Γ; Σ ` func(~x).t {eb } : [t1 , . . . , tn ; Ψ01 ] =⇒ [u; Ψ02 ]; Ψ1 Ψi ; Γ; Σ ` Ni : ti ; Ψi+1 for i = 1, . . . , n Ψn+1 c Ψ01 Ψ1 ; Γ; Σ ` func(~x).t {eb }(~v ) : u; Ψn+1 ←- Ψ02 For simplicity consider case when n = 1: Ψ01 ; Γ, x : t1 ; Σ ` eb : u; Ψ02 Ψ1 ; Γ; Σ ` func(x).t {eb } : [t1 ; Ψ01 ] =⇒ [u; Ψ02 ]; Ψ1 Ψ1 ; Γ; Σ ` v1 : t1 ; Ψ2 Ψ2 c Ψ01 Ψ1 ; Γ; Σ ` func(x).t {eb }(v) : u; Ψ2 ←- Ψ02 We have ~ hfunc(x).t {eb }(v)i ; σ, E ~ h(eb [x/v])i σ, E Consider an occurrence of a variable axiom for x in the type derivation for eb : Ψx ; Γ0 ` x : t1 ; Ψx By the stability lemma (4.6), Ψx ; Γ0 ` v : t1 ; Ψx . Hence by replacing all such occurrences by appropriate type derivations for v, we get Ψ01 ; Γ, x : t1 ; Σ ` eb [x/v] : u; Ψ02 By the Lemma 4.6 Ψ1 c Ψ2 , and Ψ2 c Ψ01 is the side condition of the application. Hence by transitivity of c we have Ψ1 c Ψ01 . By Lemma 3.5, we have Ψ02 c Ψ2 ←- Ψ02 , hence Ψ1 ; Γ, x : t1 ; Σ ` eb [x/v] : u; Ψ2 ←- Ψ02 . Which was to be proved. 5 RELATED WORK Structural rules (], ∨) 18 If the last rule used is Ψ1 ; Γ; Σ ` e : ti ; Ψ2 Ψ1 ; Γ; Σ ` e : t1 ∨ t2 ; Ψ2 then by the induction hypothesis the thesis holds for ti , hence by reapplying the disjunction rule also for t1 ∨ t2 . If the last rule used is Ψ; Γ; Σ ` e : t; Ψ2 Ψ; Γ; Σ ` e : t; Ψ2 ] Ψ3 Then by the induction hypothesis Ψ0 ; Γ; Σ ` e0 : t; Ψ2 σ 0 ; Σ0 \|= Ψ0 And we can get the thesis by applying the ] rule. ~ Theorem 4.9 (soundness) If Ψ1 ; Γ; Σ ` \|Ehei\| : t; Ψ2 and σ; Σ \|= Ψ1 ~ then either σ, Ehei has an infinite reduction path, or it reduces to a value v, 0 0 i.e. there exist σ , Σ and value v such that ~ σ, Ehei ;∗ σ 0 , v and σ 0 ; Σ0 \|= v : t, Ψ2 . Proof: By the progress lemma (4.5), the reduction can stop only at a value. Required properties of the final value and state follow from lemmas 4.8 (by induction on the lreduction length) and 4.6. 5 Related work The starting point of our research is the paper by Guha et al [GSK11] where a type system for scripting languages is presented. The type system there does not address the types of objects and only infers the typing information concerning base types (including special type of references to take into account object in a very simple 6 CONCLUSIONS AND FUTURE WORK 19 way). We provide here a type system which in addition to the one by Guha et al can infer meaningful typing information for objects. In addition, the paper by Guha et al relies on runtime tags which are present in the semantics of dynamicly typed languages such as JavaScript. The checks over runtime type tags can be viewed as asserts that check if the particular value is of an expected type. This primitive has, however, one disadvantage. Namely, it does not reflect the split in the control flow that makes possible to use particular type. Still, the scripting languages have operators that check for the actual running time type of a value (e.g. typeof function and in operator in JavaScript, or hasattr, getattr, and getattr_static in Python). We consider it more natural to rely on these operators instead of the tagchecks. Our typing framework works in a fashion such that, in principle, all the typing information should be present. This scenario is not realistic when the expectations of the scripting languages programmers are taken into account. Each real effort to bring static information into the program development here must consider what was explicated in [WEF+ 09] and realised in a scpecific way in Thorn scripting language [BFN+ 09] and led to interesting type architecture [WNL+ 10a, WNL+ 10b]. 6 Conclusions and future work The type system presented in this paper gives an expressible framework for typing objects in dynamic languages such as Python or JavaScript. It follows the general view of type-and-effect systems where running time program property must be expressed as a dynamical change. We express this change by means of Hoarelike triples that describe the structure of relevant objects before an expression is executed and after the execution. The type system is presented in a fashion similar to Church-style. The necessary typing information must be provided by a programmer. We work now on helpful heuristics that will infer most of the type information to make the annotational effort of a programmer minimal and on the way a system with types can be integrated to real program development in scripting languages. References [BB99] Viviana Bono and Michele Bugliesi. Matching for the lambda calculus of objects. Theoretical Computer Science, 212(1–2):101–140, 1999. REFERENCES 20 [BFN+ 09] Bard Bloom, John Field, Nathaniel Nystrom, Johan Östlund, Gregor Richards, Rok Strniša, Jan Vitek, and Tobias Wrigstad. Thorn: robust, concurrent, extensible scripting on the jvm. In Proceedings of the 24th ACM SIGPLAN Conference on Object Oriented Programming Systems Languages and Applications, pages 117–136, New York, NY, USA, 2009. ACM. [GL86] David K. Gifford and John M. Lucassen. Integrating functional and imperative programming. In Proceedings of the 1986 ACM Conference on LISP and Functional Programming, LFP ’86, pages 28–38, New York, NY, USA, 1986. ACM. [GSK11] Arjun Guha, Claudiu Saftoiu, and Shriram Krishnamurthi. Typing local control and state using flow analysis. In Proceedings of the 20th European conference on Programming languages and systems: part of the joint European conferences on theory and practice of software, ESOP’11/ETAPS’11, pages 256–275, Berlin, Heidelberg, 2011. Springer-Verlag. [KMCA06] Andrew J. Ko, Brad A. Myers, Michael J. Coblenz, and Htet Htet Aung. An exploratory study of how developers seek, relate, and collect relevant information during software maintenance tasks. IEEE Transactions on Software Engineering, 32(12):971–987, December 2006. [MM09] Daniel Marino and Todd Millstein. A generic type-and-effect system. In Proceedings of the 4th International Workshop on Types in Language Design and Implementation, TLDI ’09, pages 39–50, New York, NY, USA, 2009. ACM. [Ous98] John K. Ousterhout. Scripting: Higher-level programming for the 21st century. Computer, 31(3):23–30, March 1998. [Pre00] Lutz Prechelt. An empirical comparison of seven programming languages. Computer, 33(10):23–29, October 2000. [Sas96] William Sasso. Cognitive processes in program comprehension: An empirical analysis in the context of software reengineering. Journal of Systems and Software, 34(3):177–189, September 1996. [Sof12] Tiobe Software. Tiobe programming community index for april 2012. Web page http://www.tiobe.com/index.php/ content/paperinfo/tpci/index.html, April 2012. REFERENCES 21 [Ste01] Ed Stephenson. A fallback application built in six months earns the prime role. a fallback application built in six months earns the prime role. http://oreilly.com/pub/a/oreilly/ perl/news/swedishpension_0601.html, June 2001. [WEF+ 09] Tobias Wrigstad, Patrick Eugster, John Field, Nate Nystrom, and Jan Vitek. Software hardening: a research agenda. In Proceedings for the 1st Workshop on Script to Program Evolution, STOP ’09, pages 58–70, New York, NY, USA, 2009. ACM. [WNL+ 10a] Tobias Wrigstad, Francesco Zappa Nardelli, Sylvain Lebresne, Johan Östlund, and Jan Vitek. Integrating typed and untyped code in a scripting language. SIGPLAN Notices, 45(1):377–388, January 2010. [WNL+ 10b] Tobias Wrigstad, Francesco Zappa Nardelli, Sylvain Lebresne, Johan Östlund, and Jan Vitek. Integrating typed and untyped code in a scripting language. In Proceedings of the 37th Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, pages 377–388, New York, NY, USA, 2010. ACM. [WNV09] Tobias Wrigstad, Nate Nystrom, and Jan Vitek, editors. STOP ’09: Proceedings for the 1st Workshop on Script to Program Evolution, New York, NY, USA, 2009. ACM.	6cs.PL
1 arXiv:1711.09305v2 [cs.PL] 28 Nov 2017 A Language for Probabilistically Oblivious Computation DAVID DARAIS, University of Maryland CHANG LIU, University of California, Berkeley IAN SWEET, University of Maryland MICHAEL HICKS, University of Maryland An oblivious computation is one that is free of direct and indirect information leaks, e.g., due to observable diﬀerences in timing and memory access patterns. This paper presents L obliv , a core language whose type system enforces obliviousness. Prior work on type-enforced oblivious computation has focused on deterministic programs. L obliv is new in its consideration of programs that implement probabilistic algorithms, such as those involved in cryptography. L obliv employs a substructural type system and a novel notion of probability region to ensure that information is not leaked via the distribution of visible events. The use of regions was motivated by a source of unsoundness that we discovered in the type system of ObliVM, a language for implementing state of the art oblivious algorithms and data structures. We prove that L obliv ’s type system enforces obliviousness and show that it is nevertheless powerful enough to check state-of-the-art, eﬃcient oblivious data structures, such as stacks and queues, and even tree-based oblivious RAMs. 1 INTRODUCTION Cloud computing allows clients to conveniently outsource computation, but they must trust that cloud providers do not exploit or mishandle sensitive information. To remove the provider from the trusted computing base, work in both industry and research has strived to realize a “secure” abstract machine comprising an execution engine and protected memory: The adversary cannot see sensitive data as it is being operated on, nor can it observe such data at rest in memory. Such an abstract machine can be realized by encrypting the data in memory and then performing computations using cryptographic mechanisms (e.g., secure multi-party computation [Yao 1986]) or secure processors [Hoekstra 2015; Suh et al. 2003; Thekkath et al. 2000]. Unfortunately, a secure abstract machine does not defend against an adversary that can observe memory access patterns [Islam et al. 2012; Maas et al. 2013; Zhuang et al. 2004] and instruction timing [Brumley and Boneh 2003; Kocher 1996], among other “side” channels of information. For cloud computing, such an adversary is the cloud provider itself, which has physical access to its machines, and so can observe traﬃc on the memory bus. A countermeasure against an unscrupulous provider is to augment the secure processor to store code and data in oblivious RAM (ORAM) [Maas et al. 2013; Suh et al. 2003]. First proposed by Goldreich and Ostrovsky [Goldreich 1987; Goldreich and Ostrovsky 1996], ORAM obfuscates the mapping between addresses and data, in eﬀect “encrypting” the addresses along with the data. Replacing RAM with ORAM solves (much of) the security problem but incurs a substantial slowdown in practical situations [Liu et al. 2015, 2013; Maas et al. 2013] as reads/writes add overhead that is polylogarithmic in the size of the memory. Recent work has explored methods for reducing the cost of programming with ORAM. Liu et al. [2015, 2013, 2014] developed a family of type systems to check when partial use of ORAM (alongside normal, encrypted RAM) results in no loss of security; i.e., only when the addresses of secret data could (indirectly) reveal sensitive information must the data be stored in ORAM. This optimization can provide order-of-magnitude (and asymptotic) performance improvements. Wang Working draft as of July 7, 2017. 1:2 David Darais, Chang Liu, Ian Sweet, and Michael Hicks et al. [2014] explored how to build oblivious data structures (ODSs), such as queues or stacks, that are more eﬃcient than their standard counterparts implemented on top of ORAM. Their technique involves specializing ideas from ORAM algorithms to particular data structures, resulting in asymptotic performance gains in common cases. In followup work, Liu et al. [2015] devised a programming language called ObliVM1 for implementing such oblivious data structures. While Liu et al.’s earlier work treats ORAM as a black box, in ObliVM one can program ORAM algorithms as well as ODSs. A key feature of ObliVM is careful treatment of random numbers, which are at the heart of state-of-the-art ORAM and ODS algorithms. While the goal of ObliVM is that well-typed programs are secure, no formal argument to this eﬀect is made. In this paper, we present L obliv , a core language for oblivious computation, inspired by ObliVM. L obliv extends a core imperative language equipped with an information ﬂow type system [Sabelfeld and Myers 2006] with primitives for generating and using uniformly distributed random numbers. We prove that L obliv ’s type system guarantees probabilistic memory trace obliviousness (MTO), i.e., that the possible distribution of adversary-visible execution traces is independent of the values of secret variables. This property generalizes the deterministic MTO property enforced by Liu et al. [2015, 2013], which did not consider the use of randomness. In carrying out this work, we discovered that the ObliVM type system is unsound, so an important contribution of L obliv is to address the problem without overly restricting or complicating the language. L obliv ’s type system aims to ensure that no probabilistic correlation forms between secrets and publicly revealed random choices. With oblivious computations it is frequently the case that a random choice is made (e.g., where to store a particular block in an ORAM), and eventually that choice is made visible to the adversary (e.g., when that or another block is later looked up). This visibility is OK as long as the revealed value does not communicate information about a secret, which L obliv ensures by guaranteeing the number is always (still) uniformly distributed. L obliv ’s type system, presented in Section 4, ensures uniform revelations in part by treating randomly generated numbers as aﬃne, meaning they cannot be freely copied. This prohibition prevents revealing the same number twice (where the second revelation will not be uniform). Unfortunately, aﬃnity is too strong for practical algorithms, which do make copies of random numbers. L obliv ’s type system thus allows random numbers to be coerced to merely-secret numbers but only so long as the secret version is never revealed. Moreover, L obliv requires that previously-random numbers not be allowed to inﬂuence the choice of to-be-revealed random numbers, since such inﬂuence might violate the latter’s uniformity. For example, we should not be able to coerce a random number to a secret, look at the secret, and then decide to reveal that random number or some other one. The type system prevents this problem by using a new mechanism we call probability regions to track the transitive (in)dependence of numbers in one region from those in another. (That probability regions are missing in ObliVM is the source of its unsoundness.) Section 5 outlines proof that L obliv enjoys MTO, i.e., if a L obliv program is well typed, then its observable behavior is oblivious. Further details of the proof are in the Appendix. While we have not retroﬁtted L obliv ’s type system into ObliVM, we have implemented a type checker for L obliv . Section 6 presents a version of Tree ORAM [Shi et al. 2011], a state-of-the-art ORAM implementation, that our language type checks. We also show, in Section 3, that oblivious stacks, a kind of oblivious data structure [Wang et al. 2014], can be type checked when implemented on top of a non-recursive ORAM. As far as we are aware, our implementations constitute the ﬁrst automated proofs that these data structures are indeed oblivious. We believe L obliv is a promising step forward, generalizing the work that came before it (Section 7 discusses related work). 1 http://www.oblivm.com A Language for Probabilistically Oblivious Computation 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1:3 let add (mem : ( int S ∗ int S) array ) (a : int S) (v : int S) = let len = length (m) in let rec iterate ( curr_idx : int P) ( pair : ( int S ∗ int S) ) : unit = if curr_idx = len then () else (∗ WRONG code first ∗) (∗ let (_, curr_addr ) = mem[curr_idx] in ∗) (∗ if curr_addr = −1 then mem[curr_idx] ← pair ∗) (∗ else iterate ( curr_idx + 1) pair ∗) (∗ CORRECT code below ∗) let (_, curr_addr) as curr_pair = mem[curr_idx] in let (_, addr) = pair in let ( t_pair , f_pair ) = mux (curr_addr = −1 && not(addr = −1)) pair curr_pair in mem[curr_idx] ← t_pair ; iterate ( curr_idx + 1) f_pair in iterate 0 (v , a) Fig. 1. Adding to an “association array” obliviously (aka add for trivial ORAM) 2 BACKGROUND This section presents the deﬁnition of the threat model and background on type-enforced deterministic oblivious execution. The next section motivates and sketches our novel type system for enforcing probabilistic oblivious execution. 2.1 Threat Model We assume a powerful adversary that can make ﬁne-grained observations about a program’s execution. In particular, we use a generalization of the program counter (PC) security model [Molnar et al. 2006]: The adversary knows the program being executed, can observe the PC during execution as well as the contents and patterns of memory accesses. Some secret memory contents may be encrypted (while public memory is not) but all memory addresses are still visible. As a relevant instantiation, consider an untrusted cloud provider using a secure processor, like SGX [Hoekstra 2015]. Reads/writes to/from memory can be directly observed, but secret memory is encrypted (using a key kept by the processor). The pattern of accesses, timing information, and other system features (e.g., instruction cache misses) provide information about the PC. Another instantiation is secure multi-party computation (MPC) using secret shares [Goldreich et al. 1987]. Here, two parties simultaneously execute the same program (and thus know the program and program counter), but certain values, once entered by one party or the other, are kept hidden from both using secret sharing. Our techniques can also handle weaker adversaries, such as those that can observe memory traﬃc but not the PC, or can make timing measurements but cannot observe the PC or memory. 2.2 Oblivious execution Our goal is to ensure memory trace obliviousness (MTO), which is a kind of noninterference property [Goguen and Meseguer 1982; Sabelfeld and Myers 2006]. This property states that despite being able to observe each address (of instructions and data) as it is fetched, and each public value, the adversary will not be able to infer anything about manipulated secret values. To see how such an inference could occur, consider the program in Figure 1 (in OCaml-style syntax). This program takes an array of pairs of secret integers mem, a secret address a, and a secret value v and adds the 1:4 David Darais, Chang Liu, Ian Sweet, and Michael Hicks pair (v , a) to a free slot in the array, as deﬁned by its value having address −1. The commented out code on lines 6–8 illustrate a non-oblivious implementation. The code works by (on line 6) extracting the current array element, (on line 7) checking if its address is −1 and updating the element if so, and otherwise (on line 8) continuing to consider the next array element. Notice that curr_index is designated as non-secret ( int P means “public int”) because the index to an array is eﬀectively revealed when the adversary can observe memory accesses. If the program executes mem[idx] then the adversary can observe some address a being read. Knowing the starting address s of array mem, the adversary can compute idx = a − s. By watching the address trace, an adversary can learn something about the array’s contents. In particular, a sequence of N reads followed by a write implies that the ﬁrst N − 1 slots were not free while slot N was; no writes means no free slots. The adversary can also watch the program counter to notice when the true vs. the false branch of the conditional is taken, revealing the same information. The code on lines 10–16 ﬁxes these problems. The mux on line 12 is the key part. It takes three arguments. The ﬁrst is a boolean condition, which here is whether the current slot is free and also whether pair ’s address is not −1. If so then the mux evaluates to a pair comprised of the mux’s second and third arguments, in that order; otherwise the pair’s contents are reversed. Line 13 updates mem’s current slot with the ﬁrst element of the pair, and recurses to consider the next element, passing the second element of the pair. This code is oblivious: it reads and writes every slot in the array, and always executes exactly the same statements, no matter the contents of the array, index, or value arguments. As such, observing the code’s execution reveals nothing about the secrets it manipulates. This code is actually the add operation of a “trivial” oblivious RAM (ORAM) implementation [Goldreich 1987; Goldreich and Ostrovsky 1996]. ORAM provides an API that is like an array, but makes sure the address trace does not reveal the relationship between the index and the value. That is, for read oram i and write oram i v operations, the index i is eﬀectively kept secret. We will later consider sophisticated ORAM algorithms, which use trivial ORAM as a building block. 2.3 Obliviousness by typing Liu et al. [2015, 2013] developed a type system that ensures programs are MTO. Types are extended to indicate where values are allocated; as per our above example data can be public or secret, but can also reside in ORAM. Liu et al.’s small-step operational semantics σ ; e −→t σ ′; e ′ reduces an expression e in memory σ to heap σ ′ and expression e ′ while emitting trace event t. Trace events include fetched instruction addresses, public values, and the addresses of public and secret values, that are read and written, but not the addresses or values of data stored in ORAM, which is modeled as a black box: Basically it is treated as if it were an array whose accesses are opaque to the adversary (similar to the Dolev-Yao modeling of encrypted messages [Dolev and Yao 1981]). Under this model, the MTO property means that running e in two low-equivalent memories σ1 and σ2 —meaning they agree on public values—will produce the exact same memory trace, along with low-equivalent output heaps and results. More formally, if σ1 ∼ σ2 then σ1 ; e −→t1 σ1′; e 1 and σ2 ; e −→t2 σ2′ ; e 2 imply t 1 = t 2 , σ1′ ∼ σ2′ , and e 1 ∼ e 2 , where operator ∼ denotes low-equivalence. 3 TYPE-BASED ENFORCEMENT OF PROBABILISTIC OBLIVIOUSNESS Liu et al’s type system enforces obliviousness of deterministic programs that use ORAM. However, work on oblivious data structures [Wang et al. 2014] has shown that probabilistic uses of ORAM can be far more eﬃcient. In such uses, visible memory and traces need not be identical when computing with diﬀerent secrets, but rather they only must be identically distributed; i.e., all possible memory traces are equally likely, no matter the secrets used. Indeed, this is the same property that ORAM A Language for Probabilistically Oblivious Computation 1 2 3 4 5 6 7 8 1:5 let stackop (( oram,rid_r ) :' a Stack) (ispush : Bool S) (d :' a) : ' a = let rid = ! rid_r in let old_d = read oram rid in let (d ', _) = mux ispush d old_d in let ( id , _) = mux ispush (rid +1) rid in write oram id d '; let ( rid ', _) = mux ispush (rid+1) ( rid −1) in rid_r := rid '; d' Fig. 2. A deterministic oblivious stack built using a full (recursive) ORAM. algorithms, if we open up the black box, are actually providing. This section motivates the beneﬁts of probabilistic programming with ORAMs, and sketches our type system that enforces MTO for such programs. Section 6 shows this type system is actually powerful enough to implement ORAM algorithms directly, so they need not be treated as black boxes. 3.1 Motivation: naive oblivious stack is ineﬀicient The goal of an oblivious stack is to hide both its data and which operations (pushes or pops) are taking place—only the total number of operations should be revealed. An incorrect approach would be to implement the stack as usual but using a secret (encrypted) array. While doing this would hide the contents of the array, indexes into that array would be visible to the adversary (as discussed in Section 2.2), which would reveal the operation—increasing indexes indicate pushes, and decreasing ones indicate pops. To hide the index we can store the stack in an ORAM instead of an array, and merge the code of the push and pop operations, as shown in the stackop function in Figure 2, so that observation of the PC does not reveal which operation is taking place. In the code, a stack consists of an ORAM of secret numbers and a reference storing the index of the root. Function stackop takes a stack, a ﬂag indicating whether the operation is a push or pop, and the value to push, and returns a value. The code reads the value at the root index (line 3). The next line copies that value to d' if the operation is pop, or else puts d there if it is a push. Line 5 determines the index of the write it will perform on line 6: this index ( id ) is one more than the root index if it’s a push and it’s the current root index if not. As such, the write on line 6 puts the given value in the next slot in case of a push, or writes back the value at the current root, if it’s a pop. Finally, line 7 adjusts the root index, and line 8 returns the result, which is either the popped value or pushed value (if it was a push). While this code works perfectly well (and would typecheck in Liu et al’s system) a probabilistic version of the stack, using a non-recursive ORAM, would be more space-eﬃcient. In particular, it will require only O(L) extra space where L is the current size of the stack, whereas this version requires O(N ) extra space, where N is the size of the ORAM. We will see why, next. 3.2 Recursive and non-recursive ORAM To understand the source of ineﬃciency, we now digress to consider features of state-of-the-art ORAM algorithms. In particular, consider a tree-based ORAM [Shi et al. 2011; Stefanov et al. 2013; Wang et al. 2015]. It has two parts: a tree-like structure for storing the actual data blocks, and a position map that maps logical data block indexes to position tags that indicate the block’s position in the tree. In the simplest instantiation, an ORAM of size N has a position map of size N which is hidden from the adversary; e.g., it could be stored in on-chip memory in a processor-based deployment of ORAM. The position tags mask the relationship between a logical index and the 1:6 David Darais, Chang Liu, Ian Sweet, and Michael Hicks location of its corresponding block in the tree. As blocks are read and written, they are shuﬄed around in the data structure, and their new locations are recorded in the position map. Note that the position map need not be hidden on chip; rather, much of it can be stored recursively in ORAM itself, reducing the space overhead on the client. As such, the tree part that contains the data blocks is sometimes called a non-recursive ORAM (or NORAM for short). The ORAM read and write operations are implemented as a sequence of two NORAM operations, called readAndRemove (or rr , for short) and add. rr : ' a NOram → (idx:int S) → (tag : int P) → ' a add: ' a NOram → (idx:int S) → (tag : int S) → ( block :' a) → unit Here, the type ' a NORam designates the tree portion of the ORAM data structure, which stores data blocks of type ' a. Operation rr ’s int S argument is the desired block’s logical index, which is hidden from the adversary; the int P argument is its position tag extracted from the map; and the ' a is the returned data block. The add operation is similar except that its position tag can be secret, and it adds a block to the tree, rather than removing it. A full ORAM read at index i requires looking up the tag in the position map, performing an rr to read and remove the value at i, and adding back that value (updating the position map with a fresh tag). A write to i requires a rr to remove the old value at i, and an add to write the new value (again updating the position map). Depending on the details of the NORAM algorithm, rr and add may perform many reads and writes to/from the tree. For simplicity, we can think of a call to rr noram id pos as producing a single abstract event RR(noram,pos). The event indicates the operation, the NORAM it is performed on (there may be more than one), and the position tag, which is visible because it correlates with the actual memory addresses accessed to retrieve the requested block. The add operation’s behavior does not depend on the position tag, so it simply generates event ADD(noram). The fact that rr ’s position tag is visible to the adversary means that traces involving rr operations might diﬀer not because of a dependence on particular secret values, but due to innocuous random variation. Hence we require a more general deﬁnition of obliviousness, discussed below. The running time of both rr and add are O(log N ). The position map requires O(N ) storage and O(1) access time when stored non-recursively on the client, hidden from the adversary. When stored recursively in ORAM, the position map imposes O(1) storage on the client, and O(log N access time; this is the standard deployment for (full, recursive) ORAM, and is necessary when the position map accesses are visible to the adversary (as in MPC-based deployments). It turns out that L obliv is powerful enough to implement tree-based ORAMs from scratch; we present more details about this in Section 6. 3.3 An optimized oblivious stack Returning to the question of how to more eﬃciently implement an oblivious stack, consider that we always access a stack via its head, using the root index. Thus, in the code in Figure 2, the full oram internally only ever uses one slot in its position map. Thus we can do better by using an NORAM directly, having the stack manage the position tag of the root. In short, we implement an oblivious stack as a triple comprising a NORAM, the index of the root element, and its position tag. The latter two act as a kind of pointer into the NORAM. Each block stored in the NORAM contains the data and the position tag of the next block in the stack. A Language for Probabilistically Oblivious Computation 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 1:7 type ' a OStack = (' a ∗ rnd R) noram ∗ int S ref ∗ rnd R ref let stackop (( noram,rid_r , pos_r) :' a OStack) (ispush : Bool S) (d :' a) : ' a = let ( rid , pos) = ! rid_r ,! pos_r in let ( rid ', pos ', d ') = if ispush then let (d ', _) = rr noram (−1) (reveal (rnd () ) ) in let b = (d,pos) in let pos ' = rnd () in let _ = add noram (rid+1) (use pos ') b in ( rid +1,pos ', d ') else let (d ', pos ') = rr noram rid (reveal pos) in let b = (d,rnd () ) in let _ = add noram (−1) 0 b in ( rid −1,pos ', d ') in rid_r := rid '; rpos_r := pos '; d' Fig. 3. A probabilistic oblivious stack built using a non-recursive ORAM. (Does not use mux, for simplicity.) Code implementing the stack following this design is given in Figure 3. Note the code branches on the ispush variable to make it easier to read; the actual implementation must use muxs to conditionally execute each statement in both branches to ensure obliviousness.2 Line 3 extracts the current root index and position tag. Lines 6–10 handle a push operation. Line 6 ﬁrst does a “dummy read” from the NORAM; just as we saw with the oblivious add in Figure 1, using index −1 results in a dummy block being returned (the position tag argument is unimportant in this case). Line 7 constructs a new block b to push: it consists of the given data d paired with the current root’s position tag pos, thus creating a “pointer” to that block. We then generate a fresh position tag pos' for this (the new root’s) block, add the block to the noram. Random numbers generated by rnd () have type rnd R; the coercion reveal ascribes a random number the type int P (per line 6), while use gives it type int S (line 9); we explain these constructs in the next subsection. The new root index (the old one plus one), the root’s tag, and the dummy block passed in are returned on line 10. Lines 12–15 handle a pop. Here, the ﬁrst rr does real work, extracting the block that corresponds to the root index and position tag. We then generate a dummy block to “add” to the ORAM. The updated root index (the old one minus one), its position tag (returned by the rr ) and the fetched block are returned. This version of an oblivious stack performs better than the version from Figure 2. The space overhead, due to the added pointers at the root and within the ORAM, is O(L) where L is the size of the current stack, not the size N of the ORAM. The running time is still O(loд N ). The random numbers generated on lines 6 and 8 are eventually revealed by rr operations on lines 4 and/or 12. As such, there can be many possible RR(...) events for the same sequence of secret inputs. While this code is safe, there is the possibility that a diﬀerence in the visible trace reveals secret information. We need to ensure that traces are identically distributed. More formally, if σ1 ∼ σ2 then for all traces t the probability that evaluating e under σ1 produces trace t is the 2 Notice that the structure of both branches is roughly parallel, which makes converting to the use of muxes straightforward. 1:8 David Darais, Chang Liu, Ian Sweet, and Michael Hicks same as the probability that evaluating e under σ2 also produces t. This property, which we call probabilistic trace obliviousness, is ensured by the underlying NORAM implementation, and we need to ensure that the programs that use it preserve it. It turns out this property holds for our stack, and is conﬁrmed by our type system, as described next. 3.4 L obliv : Obliviousness by typing for probabilistic programs The main contribution of this paper is L obliv , a core language for ensuring obliviousness of programs that use randomness. L obliv ’s type system establishes that programs like the one in Figure 3 are oblivious by employing two mechanisms: aﬃne treatment of random values, and probability regions to track dependences between random values that could leak information when a value is revealed. Aﬃnity. Random numbers, produced by executing rnd () , are given type rnd R, where R is a probability region, explained shortly. Values of this type are, like secret values, invisible to the adversary. They can be converted to int S numbers with the use coercion and can be converted to int P numbers with the reveal coercion. For example, in Figure 3, line 8 creates a random value pos ' , which it converts to a secret number and passes to the add operation on line 9, and sets to be the position tag of the root on line 10. On line 12, the root position tag is converted to a public number which is passed as an argument to rr (eﬀectively revealing it to the adversary). L obliv ’s type system enforces that a random number can be made public at most once. It does this by treating values of type Rnd R aﬃnely: they cannot be duplicated, e.g., through assignments, and they are consumed by operations other than use. For example, variable pos ' is not consumed on line 9 when passed to use, but is consumed when assigned to rpos_r on line 17; subsequent reference to n_rpos on line 18 would be illegal. Similarly, on line 3 variable pos is read from the root of the tree; it has type Rnd R. It is then consumed on line 6 when given to reveal. Why is aﬃnity important? Consider this example, in which variable s has type int S. 1 2 3 4 let ( r1 , r2 ) = (rnd () , rnd () ) in let (z , _) = mux (s = 0) r1 r2 in output (reveal z) ; (∗ OK ∗) output (reveal r1 ) ; (∗ Bad ∗) Lines 1–3 in this code are safe: we generate two random numbers that are invisible to the adversary, and then store one of them in z depending on whether the secret s is zero or not. Revealing z at line 3 is safe: regardless of whether z contains the contents of r1 or r2, the fact that both are uniformly distributed means that whatever is revealed, nothing can be learned about s. However, revealing r1 on line 4, after having revealed z, is not safe. This is because seeing two zeroes in a row is more likely when z is r1, which happens when s is zero. So this violates obliviousness. A similar problem would arise in our oblivious stack if we changed line 8 in Figure 3 to be let pos ' = pos in ... This would set the position tag for the new root to the old root’s tag, pos, which has just been revealed. As a result, we will leak information about the secret ispush variable at the next operation, i.e., whether a push was previously performed, vs. a pop. With aﬃne revelation, the type system catches this mistake; assigning pos to b on line 7 consumes it, so reading it again on line 8 would be disallowed. Probability regions. Unfortunately, aﬃnity on its own is insuﬃcient. This is because, for reasons of needed expressiveness, use can convert random numbers to non-aﬃne, secret ones, and then computing on these can aﬀect the distribution of revealed values. To see how, consider this program: A Language for Probabilistically Oblivious Computation 1 2 3 4 5 1:9 let (x , y) = (rnd () , rnd () ) in let s_x = use(x) in let ( r1 , r2 ) = mux (s_x = 0) x y in let (z , _) = mux (s = 0) r1 r2 in output (reveal z) ; (∗ Bad ∗) Line 1 generates two random numbers, and line 2 creates a secret copy of x, which does not consume it. Line 3 sets r1 to x if it is 0, and to y otherwise; the reverse is true for r2. The problem now is that r1 and r2 are not uniformly distributed: r1 is more likely to contain zero than any other number. As a result, outputting z on line 5 is dangerous: seeing a zero here is more likely if z was r1 and thus that says that s is more likely to be zero, than not. Notice that we have not violated aﬃnity here: no random number has been revealed publicly more than once. Our type system addresses this problem with a novel construct we call probability regions. Both normal and random number types are ascribed a region R. The region represents the set of random numbers generated at program points whose rnd instructions are annotated with R. (We have elided the region name so far, but normally should write rnd R () ). Regions allow the type system to reason that random numbers are probabilistically independent. In particular, regions follow a lattice ordering, and we can say that two regions R1 and R2 are independent if their meet is ⊥. Otherwise, regions may depend on one another, meaning that revelation of a random number in one region may compromise the uniformity of the distribution of a number in the other, which could result in a leak. We can see this in our example: the region of the number s_x = 0 is the same region as x, since s_x was derived from x. As such, the two are not independent. But this lack of independence is problematic when conditioning on s_x = 0 to return x; it means that the output is no longer uniform. On the other hand, if the guard of a mux has a region that is independent of the regions of its branches, then the uniformity of the output is not threatened. Normal numbers that never were (or were inﬂuenced by) random numbers have region ⊥ because they are independent of any random number; as a notational convenience, when we write only the secrecy label (e.g., int P or int S) this means the region is ⊥. Looking at our ﬁrst example, line 2 is perfectly safe assuming that s has type int S because then its region ⊥ is independent of the region of r1 and r2. Allowing such safe code is critical for implementing eﬃcient oblivious algorithms. 4 FORMAL LANGUAGE This section presents L obliv : Its syntax, sampling semantics, and type system. The next section presents our main metatheoretic result, that typing ensures memory trace obliviousness. 4.1 Syntax Figure 4 shows the syntax for L obliv . The syntax for terms is stratiﬁed into a kind of a-normal form to simplify the semantics. We have expressions e, which comprise let binding and conditionals; atomic expressions a, which comprise various computational forms discussed shortly; and pico expressions p, which comprise variables x or literals tagged with a security label ℓ. Literals are either bits b (which are either O or I) or natural numbers n, and security labels ℓ are either public P or secret S. Semantically, literals labeled with S are invisible to the adversary while those labeled with P are not. (Think of S-labeled literals as encrypted.) Returning to atomic expressions, p ◦ takes the complement of a bit; p ⊙ p represents an arbitrary arithmetic operation; asnat(p) converts from bits to nats, and asbit(p) does the reverse. Expression flipρ () randomly generates a bit of type flipρ ; we call values of this type “coin ﬂips.” The annotation ρ is the coin ﬂip’s probability region; we defer discussion of these to Section 4.3. Coin 1:10 David Darais, Chang Liu, Ian Sweet, and Michael Hicks ∈ label F P \| S ∈ region F (parameter) ∈ BFO\|I ∈ N F {0, 1, . . .} ∈ aop F (parameter) ∈ var ≔ {x, y, z, . . .} ∈ lit ≔ b \| n ∈ pico F x \| ι ℓ ∈ atom F p \| p◦ \| p ⊙p \| asnat(p) \| asbit(p) \| flipρ () \| use(x) \| reveal(x) \| mux(p, p, p) \| hp, pi \| arrayN (p, . . . N , p) \| p[p] \| p[p] ← p e ∈ expr F a \| let x ≔ a in e \| let x, x ≔ a in e \| if p then e else e ℓ ρ b n ⊙ x ι p a security label probability region bits natural numbers arithmetic ops lexical variables literals “pico” expressions “atomic” expressions bit complement arithmetic ops bit-to-nat conversion nat-to-bit conversion coin ﬂip convert ﬂip to bit reveal ﬂip to public atomic conditional tuple creation array creation array read atomic array read+write “compound” expressions local variable binding local tuple binding conditional Fig. 4. Syntax for L obliv ﬂips are semantically secret, and have limited use; we can coerce one to public bit via reveal(p) and to a secret bit via use(p). Expression mux(p, p2, p3 ) unconditionally evaluates p2 and p3 and returns them as a pair in the given order if p evaluates to I, or in the opposite order if it evaluates to O. This operation is critical for obliviousness because its operation is atomic. By contrast, normal conditionals if p then e 1 else e 2 evaluate either e 1 or e 2 depending on p, never both, so the instruction trace communicates which branch is taken. The components of tuples constructed as hp1 , p2 i can be accessed via let x 1 , x 2 ≔ a in e. Finally, the last three atomic expression forms are for (constant size-N ) array creation, read, and update, respectively. We designed L obliv to capture the key features used when programming oblivious data structures (notably, numbers, tuples, arrays, and random numbers), but it omits obviously useful features to avoid unimportant complexity. Many missing features can be added without diﬃculty (and we have them in our implementation). For example, we can easily encode random N-bit numbers as a tuple of bits, with ready conversions to/from nats, and we can allocate variable-sized arrays. We can support lambda terms with bounded recursion using standard ideas; we imagine L obliv programs as post inlining. Allowing non-termination (as opposed to bounded recursion) is A Language for Probabilistically Oblivious Computation 1:11 more technically delicate, but also doable. We further discuss implementation issues (including the support for richer types) in Section 6. 4.2 Sampling Semantics Figure 5 shows a small-step operational semantics for L obliv programs. The main judgment has q ς ′ which states that a conﬁguration ς steps to conﬁguration ς ′ with probability q form ς (a rational number). A conﬁguration contains an expression e, environment γ , and a store σ ; a special HALT form indicates a terminated computation, with the expression replaced with a value v. A value is a labeled literal ι ℓ , an address, or a pair of values hv, vi. The environment maps variables to values, while the store maps (array) addresses α to sequences of values. The bottom of the ﬁgure gives rules for the main judgment, and are basically standard; notice that let-bound variables are put in the environment γ , rather substituted into the target term. These rules inherit their probability annotation from subsidiary judgment γ ; σ ⊢ a ։q v ; σ ′ , which states that an atomic expression a under γ and σ steps to a value v and output store σ ′ with probability q. In turn, these rules invoke PJpK to evaluate pico expressions—variables x are looked up in the environment, and literals ι ℓ are returned. All of the atomic expression rules have probability q = 1 except for the rules for flip, which return OS or IS with probability 1/2. The rules preceding these are for computing on bits and numbers, and are straightforward; we assume a given semantics function for operations ⊙. The rules for mux evaluate both branches and their values as a pair; the order of these values depends on the bit to which the guard evaluates. Array allocation stores a sequence of the given values at fresh address α in the store; array read extracts a value at the given index n P (notice it must be public), and array write returns the existing value before updating the appropriate position in the sequence present at α in the store. Having array update return the old value is critical for coexisting with the type system’s aﬃne treatment of random numbers, as described below. Attempting to access an array out of bounds is undeﬁned; note that this does not represent an information leak because array indexes are public. Traces and adversary observations. An execution trace t is a sequence of conﬁgurations ς 1 . . . ς N q whereby each ς i and ς i +1 in the sequence is justiﬁed by proof of judgment ς i i +1 , and ς N is a halted conﬁguration. We deﬁne what parts of the trace the adversary can see according to the function obs. This function takes a sequence of conﬁgurations and produces a corresponding sequence of “bulleted” conﬁgurations. We omit this function for space reasons (it is given in the appendix, in Figure 21) but its operation is straightforward: a bullet conﬁguration is the same as its original except that all secret numbers (both random and non-random) have been replaced with •. As such, we capture the idea that the adversary can see the full instruction trace (the structure of expressions is visible) the address trace (addresses α are not aﬀected), and non-secret values (public literals are unaﬀected). Only speciﬁcally secret numbers are hidden. Our type system, given next, is able to ensure even when the adversary can see such a trace, he can learn nothing about the program’s secret values. 4.3 Typing Figure 7 shows the type system for L obliv , which deﬁnes a judgment Γ ⊢ e : τ ; Γ ′, which states that under type environment Γ expression e has type τ , and yields residual type environment Γ ′. Types ρ τ and environments Γ are deﬁned in Figure 6. Bits and numbers respectively have type bitℓ and ρ natℓ , where ℓ and ρ are the label and probability region, respectively. As mentioned earlier, flipρ characterizes a coin ﬂip in region ρ; it implicitly has label S. Types array(τ ) and τ ×τ characterize arrays and pairs, respectively. Environments map variables to either types τ or inaccessibility tags •; the latter is used to support aﬃnity, discussed shortly. 1:12 David Darais, Chang Liu, Ian Sweet, and Michael Hicks γ ∈ env ≔ var ⇀ V σ ∈ store ≔ addr ⇀ V ∗ ς ∈ conﬁg F he, γ , σ i \| HALT(v, γ , σ ) α ∈ addr ≔ N v ∈ V F ι ℓ \| α \| hv, vi PJ·K : pico → env ⇀ val PJxK(γ ) ≔ γ (x) PJxK(ι ℓ ) ≔ ι ℓ γ ; σ ⊢ a ։q v ; σ CompO CompI Aop AsNO AsNI AsB FlipO FlipI Use Rev MuxI MuxO Tup Arr Read Write ⊢ po ։1 Iℓ ; σ when Oℓ = PJpK(γ ) o ⊢p ։1 Oℓ ; σ when Iℓ = PJpK(γ ) ⊢ p1 ⊙ p2 ։1 J⊙K(PJp1 K(γ ), PJp2 K(γ )) ; σ ⊢ asnat(p) ։1 0ℓ ; σ when Oℓ = PJpK(γ ) ⊢ asnat(p) ։1 1ℓ ; σ when Iℓ = PJpK(γ ) ⊢ asbit(p) ։1 (n % 2)ℓ ; σ when n ℓ = PJpK(γ ) ρ ⊢ flip ։1/2 OS ; σ ρ ⊢ flip ։1/2 IS ; σ ⊢ use(x) ։1 γ (x) ; σ ⊢ reveal(x) ։1 γ (x) ; σ ⊢ mux(p1, p2 , p3 ) ։1 hPJp2 K(γ ), PJp3 K(γ )i ; σ when Iℓ = PJp1 K(γ ) ⊢ mux(p1, p2 , p3 ) ։1 hPJp3 K(γ ), PJp2 K(γ )i ; σ when Oℓ = PJp1 K(γ ) ⊢ hp1 , p2 i ։1 hPJp1 K(γ ), PJp2 K(γ )i ; σ N ⊢ array (p1, . . . , p N ) ։1 α ; σ [α 7→ PJp1 K(γ ) . . . PJp N K(γ )] when α = fresh(σ ) γ ; σ ⊢ p1 [p2 ] ։1 σ (α)[n] ; σ when α = PJp1 K(γ ), n P = PJp2 K(γ ) γ ; σ ⊢ p1 [p2 ] ← p3 ։1 σ (α)[n] ; σ [α 7→ σ (α)[n 7→ PJp3 K(γ )]] when α = PJp1 K(γ ), n P = PJp2 K(γ ) γ γ γ γ γ γ γ γ γ γ γ γ γ γ ;σ ;σ ;σ ;σ ;σ ;σ ;σ ;σ ;σ ;σ ;σ ;σ ;σ ;σ ς q Atom γ ;σ ⊢a ։ v ;σ ha, γ , σ i LetT q ′ q ′ HALT(v, γ , σ ) Let γ ;σ ⊢a ։ v ;σ hlet x ≔ a in e, γ , σ i q IfI q he, γ [x 7→ v], σ ′i he, γ [x 1 7→ v 1 , x 2 7→ v 2 ], σ ′i γ ; σ ⊢ a ։q IP ; σ ′ hif p then e 1 else e 2 , γ , σ i IfO q he 1 , γ , σ ′i q he 2 , γ , σ ′i γ ; σ ⊢ a ։q OP ; σ ′ hif p then e 1 else e 2 , γ , σ i Fig. 5. Sampling Semantics for L obliv ς ′ γ ; σ ⊢ a ։q hv 1 , v 2 i ; σ ′ hlet x 1 , x 2 ≔ a in e, γ , σ i q A Language for Probabilistically Oblivious Computation ρ ρ τ ∈ type F bitℓ \| natℓ \| flipρ \| array(τ ) \| τ × τ • Γ ∈ tenv ≔ var ⇀ type K : type → kind ρ ρ K(bitℓ ) ≔ U K(natℓ ) ≔ U • 1:13 • τ ∈ type F τ \| • κ ∈ kind F U \| A • τ ⊏• U⊏A K(τ1 × τ2 ) ≔ K(τ1 ) ⊔ K(τ2 ) K(flipρ ) ≔ A K(array(τ )) ≔ U Fig. 6. Type and Kind Language for L obliv At a high level, the type system is enforcing three properties: (1) No type errors occur at runtime (e.g., operating on an array as if it were a number). (2) No secrets can be inferred from adversary-visible information. (3) No probabilistic correlation forms between secrets and publicly revealed random numbers. Properties (1) and (2) are enforced using standard techniques. For example, no type-correct program can operate on an array as if it were a number because rule Aop only types expression p1 ⊙ p2 when p1 and p2 are numbers (not arrays). Moreover, this operation correctly captures the inﬂuence of secret information by annotating the result type’s security label with ℓ1 ⊔ ℓ2 , the join of the labels of p1 and p2 . (Recall that S ⊔ ℓ = S for all ℓ.) Likewise, the rules ensure that array indexes and conditionals do not leak information improperly. Property (3) is the main novelty of this work, and its need arises from L obliv ’s support for random numbers that are used to enforce secrecy but can be later be seen by the adversary. A key invariant is that coin ﬂips, which have type flipρ , always have the following properties: (1) Their distribution is independent of other ﬂip values. (2) Their distribution is stable, meaning the probability of each possible bit value O or I is 1/2. The type system only allows creating, manipulating, and eliminating values at ﬂip type in ways that preserve properties (1) and (2). In particular, property (1) is maintained by treating ﬂip values as aﬃne, which prevents their duplication, and property (2) is maintained by tracking probability regions of ﬂip values as they are combined and used in computations. Aﬃnity. To enforce non-duplicability, when an aﬃne variable is used by the program, its type is removed from the output environment. Figure 6 deﬁnes kinding metafunction K that assigns a type either the kind U (freely duplicatable) or A (non-duplicatable). Arrays, nats, and bits are always universal, and ﬂips are always aﬃne; a pair is considered aﬃne if either of its components is. Rule VarU in Figure 7 types universally-kinded variables; the output environment Γ is the same as the input environment. Rule VarA types an aﬃne variable by marking it • in the output environment. This rule is suﬃcient to rule out the ﬁrst problematic example in Section 3.4. There are a few other rules that are aﬀected by aﬃnity. Rules Use and Reveal permit converting flipρ types to bits via the use and reveal coercions, respectively. The ﬁrst converts a flipρ to ρ ρ a bitS and does not make x inaccessible, while the second converts to a bitP and does make it inaccessible. In essence, the type system is enforcing that any random number is made adversaryvisible at most once; secret copies (that are never themselves revealed) are allowed. Arrays are permitted to contain aﬃne or universal values. Only in the latter case is it safe to read out an array element directly, since in the former it would become immediately inaccessible. Hence rule Read requires that array element type τ have kind U. Rule Write applies to all arrays: 1:14 David Darais, Chang Liu, Ian Sweet, and Michael Hicks Γ ⊢e :τ ; Γ VarU K(Γ(x)) = U VarA Γ ⊢ x : Γ(x) ; Γ K(Γ(x)) = A LitB Γ ⊢ x : Γ(x) ; Γ[x 7→ •] ρ LitN AsN Γ ⊢ n ℓ : nat⊥ ℓ ; Γ Rev ρ Γ(x) = flipρ Γ ⊢ p1 : natℓ11 ; Γ ′ ρ ⊔ρ ρ Γ ′ ⊢ p2 : bitℓ22 ; Γ ′′ Γ ′′ ⊢ p3 : bitℓ33 ; Γ ′′′ ρ ⊔ρ ⊔ρ U ρ Γ ′ ⊢ p2 : natℓ22 ; Γ ′′ Γ ⊢ p1 ⊙ p2 : natℓ11⊔ℓ22 ; Γ ′′ ρ Γ ⊢ p1 : bitℓ11 ; Γ ′ ρ Γ ⊢ use(x) : bitS ; Γ ρ ρ ρ ⊔ρ ⊔ρ Γ ⊢ mux(p1 , p2 , p3 ) : bitℓ11⊔ℓ22⊔ℓ33 × bitℓ11⊔ℓ22⊔ℓ33 ; Γ ′′′ ρ Γ ⊢ p1 : bitℓ 1 ; Γ ′ MuxFlip Γ ′ ⊢ p2 : flipρ 2 ; Γ ′′ ρ1 ⊥ ⊥ ρ2 ρ1 ⊥ ⊥ ρ3 Γ ′′ ⊢ p3 : flipρ 3 ; Γ ′′′ A Γ ⊢ mux(p1, p2 , p3 ) : flipρ 2 ⊔ρ 3 × flipρ 2 ⊔ρ 3 ; Γ ′′′ Γ ⊢ p1 : τ1 ; Γ ′ Γ ′ ⊢ p2 : τ2 ; Γ ′′ Γ ⊢ hp1 , p2 i : τ1 × τ2 ; Γ Γ ⊢ p1 : array(τ ) ; Γ ′ ′′ Γ ′ ⊢ p2 : nat⊥ P ; Γ ′′ Arr K(τ ) = U Write Γ ⊢ p1 [p2 ] : τ ; Γ ′′ Let LetT ρ Γ ⊢ asbit(p) : bitℓ ; Γ ′ Use Γ ⊢ p ◦ : flipρ ; Γ ′ Aop Γ ⊢ reveal(x) : bit⊥ P ; Γ[x 7→ •] Read Γ ⊢ p : natℓ ; Γ ′ Γ ⊢ p : flipρ ; Γ ′ Comp Γ(x) = flipρ Tup AsB Γ ⊢ asnat(p) : natℓ ; Γ ′ Γ ⊢ flipρ () : flipρ ; Γ MuxBit ρ Γ ⊢ p : bitℓ ; Γ ′ ρ,⊥ Flip Γ ⊢ b ℓ : bit⊥ ℓ ; Γ Γ ⊢ a : τ1 ; Γ ′ Γ0 ⊢ Γn−1 ⊢ pn : τ ; Γn N array (p1, .., p N ) : array(τ ) Γ ⊢ p1 : array(τ ) ; Γ ′ ′′ Γ ′ ⊢ p2 : nat⊥ P ; Γ Γ ′′ ⊢ p3 : τ ; Γ ′′′ Γ ⊢ p1 [p2 ] ← p3 : τ ; Γ ′′′ Γ ′ ⊎ [x 7→ τ1 ] ⊢ e : τ2 ; Γ ′′ ⊎ [x 7→ ] Γ ⊢ let x ≔ a in e : τ2 ; Γ ′′ Γ ⊢ a : τ1 × τ2 ; Γ ′ Γ ′ ⊎ [x 1 7→ τ1 , x 2 7→ τ2 ] ⊢ e : τ3 ; Γ ′′ ⊎ [x 1 7→ , x 2 7→ ] Γ ⊢ let x 1 , x 2 ≔ a in e : τ3 ; Γ ′′ If ; ΓN ′ Γ ⊢ p : bit⊥ P ; Γ Γ ′ ⊢ e 1 : τ1 ; Γ1′′ Γ ′ ⊢ e 2 : τ2 ; Γ2′′ Γ ⊢ if p then e 1 else e 2 : τ1 ⊔ τ2 ; Γ1′′ ⊔ Γ2′′ Fig. 7. Typing Judgment for L obliv A Language for Probabilistically Oblivious Computation 1:15 the returned value is the old one (just as with a read), and the written one serves as its immediate replacement, guaranteeing the store remains well-typed. Finally, note that diﬀerent variables could be made inaccessible in diﬀerent branches of a conditional, so If types each branch in the same initial context, but then joins their the output contents; if a variable is made inaccessible by one branch, it will be inaccessible in the joined environment. Probability regions. Now we consider the role of probability regions ρ in the type system. Probability regions are a partially ordered set. The purpose of regions is to track ﬁne-grained, dynamic probabilistic dependencies between values. Regions must be a distributive lattice with a least element ⊥. A distributive lattice must support both join ⊔ and meet ⊓ operations, and meets must distribute through joins, that is: ρ 1 ⊓ (ρ 2 ⊔ ρ 3 ) = (ρ 1 ⊓ ρ 2 ) ⊔ (ρ 1 ⊓ ρ 3 ). We notate the meet of two ⊥ ρ 2 , which we call region independence. Distributivity is regions being bottom ρ 1 ⊓ ρ 2 = ⊥ as ρ 1 ⊥ necessary for a conjunction of region independence to coincide with region independence for the join, that is: ρ 1 ⊥ ⊥ ρ2 ∧ ρ1 ⊥ ⊥ ρ 3 ⇐⇒ ρ 1 ⊥ ⊥ (ρ 2 ⊔ ρ 3 ). Rule Flip types a coin ﬂip. The region annotation ρ is a static name for coin ﬂips generated at this (so-annotated) program point. The region must not be ⊥ because ⊥ represents independence from all random values, and a coin ﬂip is dependent on itself. When created, a coin ﬂip is independent ⊥ ρ ′, and that the distribution from any other value with an independent region, so ρ ′ where ρ ⊥ is stable, meaning each possible bit value O and I has equal probability. Given this interpretation of the ﬂip type, we are able to typecheck the bit-complement of a value at ﬂip type as also having ﬂip type, given as rule Comp. Rules for literals LitB and LitN track the security label ℓ of the literal, and are given probability region ⊥ to signify they are probabilistically independent of all other random values. Rules Use and Reveal, mentioned above, crucially preserve the probability region ρ from the input flipρ on the output type. As such, the type system can track their potential dependence on other values. Rules AsN and AsB convert between bit and nat types, also preserving the region. Per rule Aop, the result type of an arithmetic operation is given the join of operand labels and the join of operand regions. In particular, joining the regions indicates that whatever correlates with p1 or p2 may also correlate with the result of the operation, which is a safe over-approximation of correlation for the result of any binary arithmetic operation which operates directly on values (i.e., not distributions of values). A critical feature of the type system is the handling of mux expressions, which are treated by rules MuxBit and MuxFlip for bit and ﬂip types, respectively. The mux rule for bits treats mux as a ternary bit operation, and takes the join of security labels and probability regions of each argument. The mux rule for ﬂip types is designed to avoid passing on any probabilistic dependence from the guard p1 to the results—notice that the labels on the output pair of ﬂip types are ρ 2 ⊔ ρ 3 and not ρ 1 ⊔ ρ 2 ⊔ ρ 3 . This is acceptable when the probability region ρ 1 of the guard p1 is independent of the regions of the arguments, i.e., ρ 1 ⊥ ⊥ ρ 2 and ρ 1 ⊥ ⊥ ρ 3 . As such, any revelation of values in ρ 1 will not aﬀect the stability or independence of ρ 2 or ρ 3 . For example, the following program is well-typed: 1 2 3 4 5 6 let x = rndρ0 in let y = rndρ in let z = rndρ in let s_x = use(x) in let ( r1 , r2 ) = mux (s_x = 0) y z in output (reveal x) ; 1:16 David Darais, Chang Liu, Ian Sweet, and Michael Hicks According to the MuxFlip rule, both r1 and r2 have type flipρ —there is no inﬂuence from x’s region ρ 0 . This is as it should be: because y and z are probabilistically independent from x, the revelation of x on line 6 does not aﬀect their stability. On other hand, rule MuxFlip prevents the second problematic example from Section 3.4 (page 9). In that example the mux on line 3 is disallowed because the region of the guard s_x = 0 is not independent of the region of its argument x—they are the same. The remaining type rules handle tuples, arrays, let binding, and normal conditionals; we remark on a few important features of them. First, notice that Read and Write both require that the array index p2 has label P; this is important because we assume the adversary is able to observe array accesses. (It also avoids leaks due to programs accessing the array out of bounds.) Rule If similarly requires the guard p to have label P since the execution trace reveals which branch is taken. Second, both Let and LetT remove their bound variables from the output environment; we write Γ ′′ ⊎ [x 7→ ] to split a context into the part that binds x and Γ ′′ binds the rest; it is the latter part that is returned, dropping the x binding. 5 PROBABILISTIC MEMORY TRACE OBLIVIOUSNESS This section describes our proof of memory trace obliviousness for L obliv . We begin by presenting a model for distributions, probabilities, conditional probabilities, and probabilistic dependence. Next we give a denotational semantics based on this model, and which we prove coincides with the sampling semantics in Section 4.2. Equipped with a model for distributions and a denotational semantics we then state precisely the deﬁnition of memory trace obliviousness. To describe our proof approach, we proceed in more informal terms and describe the key semantic invariants which hold of the denotational semantics. These key invariants are suﬃcient to prove a strengthened memory trace obliviousness property, and this strengthened property implies the simpler property initially stated. Further details of the proof, including precise mathematical deﬁnitions and properties, are given in the Appendix. 5.1 Distributions We use a discrete model for distributions which are deﬁned over a ﬁxed universe with ﬁnite entropy. In our model, a distribution of elements x ∈ X in universe R is notated x̃ R . Here, R is the number of available coin ﬂips. So distributions are represented as a mapping from a bitvector b® of length R to an element of X , that is: x̃ R ∈ BR → X . This domain supports modeling the result of a terminating computation which ﬂips an arbitrary, ﬁnite number of coins. Our model does not support nonterminating computations, especially those which may ﬂip an inﬁnite number of coins. As an example, suppose we want ỹ R to represent the distribution of the count of coin ﬂips that have come up O. Thus ỹ R is a function that takes a length-R vector of coin ﬂips and returns a natural number that sums the occurrences of O in the vector. We deﬁne a probability measure for distributions p x̃ R = x by counting the occurrences of x in x̃ R and dividing by 2R . So for our example, the probability p ỹ R = 2 when R = 2 would be 41 . The model BR → X for distributions is particularly useful for expressing a measure for conditional probability. Conditional probabilities for distributions p x̃ R = x ỹ R = y are deﬁned by counting simultaneous occurrences of x in x̃ R and y in ỹ R and dividing by occurrences of y in ỹ R . Conditional probabilities are therefore a derived notion, as opposed to an axiomatic notion, as is typically the case when modeling distributions as x̃ ∈ X → [0..1]. When the universe R is not informative, we will omit it and just write x̃. We notate joint probabilities as p [x̃ = x ∧ ỹ = y]. We notate point distributions ⌊x⌋ where p [⌊x⌋ = x] = 1. A Language for Probabilistically Oblivious Computation 1:17 Key to our approach is careful bookkeeping of a set of coin ﬂips θ ∈ ℘(N ≤R ) upon which all information observable to the adversary depends. We notate probabilistic dependency x̃ ⋉ θ and prove that if two random variables x̃ and ỹ depend on disjoint sets of coin ﬂips θ 1 ∩ θ 2 = then they are probabilistically independent, written x̃ ⊥ ⊥ ỹ, which means their joint probability factors: p [x̃ = x ∧ ỹ = y] = p [x̃ = x] p [ỹ = y]. We will often notate a probability conditioned on some θ as p [x̃ = x \| θ ] = q, which is the probability that x̃ = x no matter the values of each of the coin ﬂips in θ . We notate probabilistic dependency conditioned on a set of coin ﬂips θ modulo θ ′ as x̃ ⋉ θ ′ θ , and probabilistic independence modulo θ as x̃ ⊥ ⊥ θ ỹ. We call two distributions x̃ 1 and x̃ 2 equivalent, notated x̃ 1 ≈ x̃ 2 , when their probability measures coincide for all elements, that is ∀x . p [x̃ 1 = x] = [x̃ 2 = x]. We call x̃ 1 and x̃ 2 equivalent modulo θ , notated x̃ 1 ≈ θ x̃ 2 , when their probability measures coincide for all elements conditioned on θ , that is ∀x . p [x̃ 1 = x \| θ ] = p [x̃ 2 = x \| θ ]. Another key invariant in our proof will regard a distribution’s stability, notated stable [x̃]. Stability is the property that each outcome of a distribution is equally likely, that is ∀x 1 x 2 . p [x̃ = x 1 ] = p [x̃ = x 2 ]. We call a distribution x̃ stable modulo θ , notated stable [x̃ \| θ ], when it is stable conditioned on θ , that is ∀x 1 x 2 . p [x̃ = x 1 \| θ ] = p [x̃ = x 2 \| θ ]. 5.2 Denotational Semantics To facilitate the proof of memory trace obliviousness for L obliv , we deﬁne a denotational semantics over distributions, and which we prove corresponds with the sampling semantics deﬁned previously in Section 4. We deﬁne the denotational semantics in two parts, an evaluation semantics for expressions, notated EJ·K, and a trace semantics for expressions, notated T J·K. The evaluation semantics EJ·K denotes an expression e to a partial function from distributions of environments γ̃ and stores σ̃ to distributions of values ṽ, environments γ̃ ′ and stores σ̃ ′. Recall that a conﬁguration ς is an expression, a (non-distributional) environment, and a store, while an execution trace t is a sequence of conﬁgurations. The trace semantics T J·K then denotes an expression to a partial function from distributions of environments γ̃ and stores σ̃ to distributions of traces t˜. e × eg g ⇀V g EJ·K : exp → eg nv × store nv × store ∗ g ⇀ conﬁg T J·K : exp → eg nv × store The semantics is partial due to the possibility of runtime type errors. However, type safety (proved later) guarantees that the denotational semantics is total for well-typed expressions. Theorem 5.1 (Sampling and Denotation Correspondence). The sampling semantics ς q ς and the denotational semantics T J·K correspond, that is for ς 1 ≔ he 1 , γ 1 , σ1 i and ς N ≔ HALT(v, γ N , σ N ): If: Then: ς 1 q1 · · · qN ς N and t˜ ≔ T Je 1 K(⌊γ 1 ⌋, ⌊σ1 ⌋) Î qn p t˜ = ς 1 . . . ς N = n ∈{1..n } 5.3 Memory Trace Obliviousness The design of L obliv is such that well typed programs have the property of Memory Trace Obliviousness (MTO). We state and prove MTO with respect to an adversary that is able to observe the memory and instruction access pattern of the execution. The MTO theorem establishes that the adversary learns nothing from these observations. 1:18 David Darais, Chang Liu, Ian Sweet, and Michael Hicks Theorem 5.2 (Memory Trace Obliviousness (MTO)). Given a well typed L obliv expression e and two value environments γ 1 and γ 2 which are both well-formed and agree on public values, then the publicly observable traces of executing e using γ 1 or γ 2 yields equivalent distributions, that is: If: Then: Γ ⊢ e : τ ; Γ ′ and Γ ⊢ γ 1 and Γ ⊢ γ 2 f JeK(⌊γ 1 ⌋, )) ≈ obs(T f JeK(⌊γ 2 ⌋, )) obs(T and obs(γ 1 ) = obs(γ 2 ) We deﬁne the public observations of a value, environment, store or conﬁguration using obs (see f Section 4.2), and notate its lifting to distributions as obs. 5.4 Proof outline This section outlines the key ideas and invariants required for the proof of MTO. We provide further mathematical detail of deﬁnitions and key properties in the Appendix. The ﬁrst property we prove is type safety, which shows that the partial denotation functions are actually total under the assumption that the term is well typed. Theorem 5.3 (Type Safety). Given a well-typed L obliv expression e, and distributions of value environments γ̃ and value stores σ̃ which are well-formed, then the evaluation semantics EJeK is always deﬁned, and the result is well-formed, that is: If: Then there exists: Such that: Γ ⊢ e : τ ; Γ ′ and Γ ; Σ ⊢ γ̃ Σ ′, ṽ, γ̃ ′ and σ̃ ′ EJeK(γ̃ , σ̃ ) = hṽ, γ̃ ′, σ̃ ′i and and Σ ′ ⊢ ṽ Σ ⊢ σ̃ and Γ ′ ; Σ ′ ⊢ γ̃ ′ and Σ ′ ⊢ σ̃ ′ We then prove type safety for the trace semantics as a consequence of type safety for the evaluation semantics. The next step in the proof is to: (1) Give a semantics to regions which appear in types; and (2) Relate this semantics for regions to the evaluation semantics We achieve (1) by deﬁning a region semantics, notated RJ·K, which when applied to a conﬁguration ς, produces a map from regions ρ to the concrete set of coin ﬂips θ that occurred at a flipρ () (i.e., with that region annotation) during the evaluation of ς. We call this map a region model, notated ξ . A key invariant of the region model is that disjoint regions map to disjoint sets, that is: ∀ρ 1 ρ 2 . ρ 1 ⊓ ρ 2 = ⊥ ⇒ ξ (ρ 1 ) ∩ ξ (ρ 2 ) = . We achieve (2) in two parts. First, we ﬁrst deﬁne a dependency semantics, notated DJ·K, which maps conﬁgurations to (in the case of base types) a set of coin ﬂips θ . This set over-approximates the ground-truth dependencies of the resulting value ṽ after conditioning on public revelations, which is given as a set of coin ﬂip dependencies Φ, that is ṽ ⋉ Φ θ . The desired relationship between regions and values is that whenever an expression annotated with region ρ evaluates to a value ṽ, then the dependency set θ predicted by the dependency semantics DJ·K is included in the set of coin ﬂips associated with the region model, that is ṽ ⋉ Φ θ ⊆ ξ (ρ). We make each of these steps precise as theorems used in the (ﬁnal) proof. Theorem 5.4 (Dependency Soundness). Given a well-typed L obliv expression e and distributions of value environments γ̃ and value stores σ̃ which are well-formed w.r.t. models for dependencies in the environment and store respectively, conditioned on the set of publicly revealed bits Φ, then the result of the dependency semantics DJ·K is well-formed w.r.t. the results of the evaluation semantics EJ·K conditioned on some equal or larger set of publicly revealed bits Φ′ . A Language for Probabilistically Oblivious Computation 1:19 Theorem 5.5 (Region Soundness). Given a well-typed L obliv expression e and distributions of value environments γ̃ and value stores σ̃ which are well-formed w.r.t. a region model ξ conditioned on the set of publicly revealed bits Φ, then the result of the region semantics RJ·K is well-formed w.r.t. the results of the dependency semantics DJ·K conditioned on some equal or larger set of publicly revealed bits Φ′ . The primary purpose of the region semantics, the dependency semantics, and their soundness ⊥ ρ 2 , then their comproperties is to establish that when terms are typed at disjoint regions ρ 1 ⊥ ⊥ ṽ 2 . We now turn to the two key invariants puted values will be probabilistically independent ṽ 1 ⊥ that justify memory trace obliviousness in the presence of reveal(p) expressions: partitioning and stability. We say that evaluation contexts hγ̃ , σ̃ i are well-partitioned if and only if any two bit distribution values b̃1 and b̃2 at type flip occurring at diﬀerent positions in the evaluation context are guaranteed to be independent of each other, conditioned on publicly revealed information Φ, that ⊥ Φ b̃2 . We prove a theorem that the evaluation semantics preserves well partitioning, conis b̃1 ⊥ ditioned on public information. Theorem 5.6 (Well-partitioning). Given a well-typed L obliv expression e and distributions of value environments γ̃ and value stores σ̃ where every contained flip value is well-partitioned w.r.t. other flip values conditioned on the set of publicly revealed bits Φ, then the result of the evaluation semantics EJ·K is well-partitioned for some equal or larger set of publicly revealed bits Φ′ . The key part of the proof of well-partitioning is for reveal(p) expressions. Before the revelation, it is assumed all ﬂip values are well partitioned. Immediately after the reveal expressions, ﬂip values will remain well-partitioned, but conditioned on a larger set of publicly revealed bits Φ′ , to which the publicly revealed ﬂip is added. In subsequent execution, the publicly revealed bit may end up correlating with ﬂip values (e.g. as a result of a public if expression), however because partitioning is quotiented by publicly revealed bits, this correlation ends up being benign. The mechanism in the type system that enforces this semantic property of well-partitioning is the aﬃne treatment of values at type flip; because flip values cannot be duplicated, there will always only be one copy of each coin ﬂip in the context, and therefore each ﬂip value will be independent of other ﬂip values. We deﬁne stability also as a property of evaluation contexts hγ̃ , σ̃ i, such that any two bit distribution values b̃ at type flip in the evaluation context (either γ̃ or σ̃ ) are hguaranteed to be stable distrii butions conditioned on publicly revealed information, that is: stable b̃ Φ . We prove a theorem that the evaluation semantics EJ·K preserves stability, again conditioned on public information. Theorem 5.7 (Stability). Given a well-typed L obliv expression e and distributions of value environments γ̃ and value stores σ̃ where every contained flip value is stable conditioned on the set of publicly revealed bits Φ, then the result of the evaluation semantics EJ·K is stable conditioned for some equal or larger set of publicly revealed bits Φ′ . The proof of stability uses the proof of partitioning as a lemma. The key part of the proof is for reveal(p) expressions. Before the revelation, it is assumed all ﬂip values are stable. Immediately after the revelation, it must be shown that all other coin ﬂips remain stable. This is a consequence of well-partitioning, which guarantees that the revealed coin ﬂip will not correlate with any other coin ﬂip, and therefore adding the revealed bit to the set of public information does not further perturb the remaining stable coin ﬂips in the context. The mechanism in the type system which enforces the semantic property of stability is the typing rules for bit complement and mux. Bit complement is safe w.r.t. stability, which we prove, 1:20 David Darais, Chang Liu, Ian Sweet, and Michael Hicks and the mux rule is stable under the conditions of region independence, which in turn imply that the underlying distributions of values are probabilistically independent. Other than public revelation, these are the only two language forms which manipulate coin ﬂips, and their typing rules guarantee that resulting coin ﬂips will remain stable. In order to prove the desired memory trace obliviousness property, we state a stronger property which implies the desired property, but which supports a direct proof by induction on the term. This stronger trace obliviousness property includes all of the assumptions for each of the above lemmas, so each lemma can be applied in each case of the proof by induction, for which we show the high-level details. Theorem 5.8 (Strong Memory Trace Obliviousness (SMTO)). Given a well-typed L obliv expression e and satisfying properties from the above lemmas applied to each of two sets of distribution value environments and stores hγ̃ 1 , σ̃1 i and hγ̃ 2 , σ̃2 i, and where each of γ̃ 1 and γ̃ 2 and σ̃1 and σ̃2 are distributed equally for publicly observable values conditioned on publicly revealed bits Φ (i.e., f (γ̃ 1 ) ≈ Φ obs f (γ̃ 2 ) and obs f (σ̃1 ) ≈ Φ obs f (σ̃2 )), then the observable traces for e applied to either set obs of contexts hγ̃ 1 , σ̃1 i or hγ̃ 2 , σ̃2 i will yield identical distributions conditioned on publicly revealed bits Φ f JeK(γ̃ 1 , σ̃1 )) ≈ Φ obs(T f JeK(γ̃ 2 , σ̃2 ))). (i.e., obs(T Proof. We prove this theorem by induction on the syntax for terms, and by applying each other theorem developed thus far. For example, for let expressions, the proof goal is of the form: f Jlet x ≔ a in eK(γ̃ 1 , σ̃1 )) ≈ obs(T Φ f Jlet x ≔ a in eK(γ̃ 2 , σ̃2 )) obs(T From the deﬁnition of the trace semantics T J·K, this goal becomes: f γ̃ ′′ ) ⊗ obs( f σ̃ ′) + f JeK(γ̃ ′′ , σ̃ ′)) ≈ ⌊e⌋ ⊗ obs( + obs(T 1 1 e 1 1 Φ f γ̃ ′′ ) ⊗ obs( f σ̃ ′) + f JeK(γ̃ ′′ , σ̃ ′)) ⌊e⌋ ⊗ obs( + obs(T 2 2 e 2 2 e where hṽ, γ̃i′, σ̃i′i ≔ EJaK(γ̃i , σ̃i ), γ̃i′′ ≔ γ̃i′ [x 7→ ṽ], and ⊗ and + + are product and concatenation operations respectively lifted to distributions. This goal is broken down using the product chain rule for distributions, which states, using our notation and in the case of three elements: p[x̃ ⊗ ỹ ⊗ z̃ = hx, y, zi] = p[x̃ = x]p[ỹ = y \| x̃ = x]p[z̃ = z \| x̃ = x ∧ ỹ = y] In the case that two elements of a tuple are independent, then it factors without conditioning, as per the deﬁnition of independence. Using the product chain rule, in order to show the two distributions in the goal are equivalent, it suﬃces to show each component of the three-tuple is independent of other elements of the threef γ̃ ′′ ) ⊥ f σ̃ ′)), that each element of the three-tuple is equivalent to its pairwise tuple (e.g. obs( ⊥ obs( 1 1 ′′ f f γ̃ ′′)), and that the recursive occurrence of the trace is equivalent element (e.g. obs(γ̃ 1 ) ≈ Φ obs( 2 pairwise conditioned on the three-tuple. That is, after establishing independence of elements in the three-tuple, it suﬃces to show: h i h i f γ̃ ′′ ) ≈ Φ p obs( f γ̃ ′′ ) (1) p obs( 1 2 h i h i f σ̃ ′) ≈ Φ p obs( f γ̃ ′′ ) (2) p obs( 1 2 h i h i ′′ ′ f f f σ̃ ′) ≈ Φ p obs(T f JeK(γ̃ ′′ , σ̃ ′ )) obs( f γ̃ ′′ ) ∧ obs( f σ̃ ′) (3) p obs(T JeK(γ̃ 1 , σ̃1 )) obs(γ̃ 1′′ ) ∧ obs( 1 2 2 2 2 Goals (1) and (2) are consequences of partitioning and stability theorems. Goal (3) is further reﬁned based on the observation that the new contexts hγ̃ 1′′ , σ̃1′i and hγ̃ 2′′, σ̃2′ i depend uniquely on an equal or larger set of publicly revealed bits Φ′ under which the contexts remain well-partitioned and stable. It is therefore suﬃcient to condition on this (potentially) larger set of publicly revealed bits A Language for Probabilistically Oblivious Computation Φ′ , and demonstrate equivalence of distributions module Φ′ , that is: h i h f JeK(γ̃ ′′ , σ̃ ′)) Φ′ ≈ Φ′ p obs(T f JeK(γ̃ ′′ , σ̃ ′ )) p obs(T 1 1 2 2 1:21 Φ′ i At this point, the inductive hypothesis applies and the proof for let is complete. The proofs for let on tuples and public if statements follow the same structure, and are analogous. After proving SMTO, we recover a proof of MTO as an instantiation of SMTO with the empty σ = ⌊⌋ and the empty set of publicly revealed bits Φ = {}. store b 6 CASE STUDY We have implemented a type checker for an extension to L obliv that adds functions, (bounded) recursion, and various arithmetic operations. We have used it to check two classical probabilistically oblivious algorithms: non-recursive ORAM (NORAM) and tree-based ORAMs. These were ﬁrst mentioned in Section 3.2, where we used NORAM as a building block to implement an oblivious stack. In this section, we discuss the implementation of both NORAM and tree ORAM, where the latter makes use of the former. In what follows, we write nat S to be a secret number in region ⊥; nat P for a public number in ⊥; nat R to be a secret in some region R. We also write type rnd R to be a random natural number in region R; in our implementation we encode these using tuples of flip Rs. 6.1 Non-recursive ORAM A non-recursive ORAM (NORAM) is the foundation all tree-based ORAMs. Such NORAMs provide the state-of-the-art in performance, and enable eﬃcient oblivious data structures. A NORAM organizes data blocks as a complete tree of depth log N where N is the capacity of the ORAM. Each node on the tree is a bucket, which is essentially a trivial ORAM of size b, which we call the bucket size. A trivial ORAM is basically an array in which reads and writes require iterating through each element in the array; we showed the trivial ORAM add operation in Section 2.2. Each NORAM data block bucket associated with a randomly generated position tag, which is a log N -bit random number, corresponding to one leaf of the tree. A key invariant of a tree-based NORAM is that each data block always resides in a bucket somewhere on the path from the root to the leaf corresponding to its position tag. In particular, we have the following signature: type noram R R' = (bucket R R ') array type bucket = ( block R R ') array type block R R' = bit R ∗ nat R ∗ nat R ∗ (rnd R' ∗ rnd R ') where R ⊥ ⊥ R' A NORAM is an array of 2N − 1 buckets which represent a complete tree in the style of a heap data structure: for the node at index i ∈ {0, ..., 2N − 2}, its parents, left child, and right child correspond to the nodes at index (i −1)/2, 2i +1, and 2i +2, respectively. Each bucket is an array of blocks, each of which is a 4-tuple, where the last element is the actual data. Ideally, an NORAM should be able to store data of an arbitrary type α, but since our language does not yet support polymorphism we ﬁx the data to type rnd R' ∗ rnd R' where R and R' are independent. We choose this type to illustrate the aﬃne values can be stored in the NORAM, and to set up our implementation of Tree ORAM next. (In our actual interpreter, we duplicate the code for each type we wish to store.) The other three components of the block are all secret and in region R; they are (1) a bit indicating whether the block is dummy or not; (2) the index of the block; and (3) the position tag for the block. NORAM supports two primitives, rr and add, which we introduced (in greater generality) in Section 3.2. Their signatures for our implementation are: 1:22 David Darais, Chang Liu, Ian Sweet, and Michael Hicks rr : noram R R' → ( idx : nat R) → (tag :nat P) → (data :rnd R' ∗ rnd R ') add: noram R R' → ( idx : nat R) → (tag : nat R) → (data :rnd R' ∗ rnd R ') → unit Due to the NORAM invariant, when the position tag corresponding to a block is known, rr can be implemented as a linear scan through the path corresponding to the tag to ﬁnd the block with the given index. In doing so, log N buckets are accessed, each with b blocks. Therefore, the rr routine can be implemented in O(b log N ) time. In the state-of-the-art ORAM constructions, such as Circuit ORAM [Wang et al. 2015], b can be parameterized as a constant (e.g., 4), which renders the overall time complexity of rr to be O(log N ), which is much faster than linear scanning over a trivial ORAM. Here is the code: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 let rr (noram:noram R R') ( idx :nat R) (tag : nat P) : rnd R '∗ rnd R' = let noram_cap = (length(noram) + 1) / 2 in let bucket_size = length(noram[0]) in let rec read_bucket bucket ( idx : nat R) ( i : nat P) (read : rnd R '∗ rnd R ') = if i =bucket_size then read else (∗ read out the current block ∗) let isdummy, blk_idx, pos_tag, data = (bucket[i ] ← ( true , 0, 0, mk_dummy())) in (∗ check if the current block is non−dummy and its index matches the queried one ∗) let toswap:bit R = isdummy && blk_idx = idx in let data, read = mux toswap read data in let isdummy, _ = mux toswap false isdummy in (∗ when toswap is false , this is equivalent to writting the data back ; otherwise , read stores the found block ' s data field , and is passed into the next iteration ∗) let _ = (bucket[i ] ← (isdummy, blk_idx, pos_tag, data) ) in read_bucket bucket idx ( i +1) read in let rec read_level ( level : nat P) (read : nat R) = let base :Nat P = (pow 2 level ) − 1 in (∗ compute the first index into the bucket array at depth level ∗) if base >= length (noram) then read else let bucket_index:Nat P = base + (tag & base) in (∗ the bucket on the path to access ∗) let bucket = noram[bucket_index] in let read = readBucket bucket idx 0 read in read_level ( level + 1) read in let read = mk_dummy () in read_level 0 read The implementation involves two nested loops: read_level (line 18-26) implements the outer loop, which iterates through all buckets on the path from the root to the leaf corresponding to tag; and read_bucket (line 4-17) iterates through all blocks in a bucket (this is essentially a Trivial ORAM read). Here are some things to notice. First, the NORAM itself is an array of arrays, and since array addresses are non-aﬃne, we can read out noram[0] directly (line 3). On the other hand, buckets contain aﬃne blocks (the data part), so we must replace a read block with a dummy one, as shown on lines 8 and 15. On line 10, toswap is computed to determine whether to swap out the block, i.e., whether its index matches the queried one. Line 10 muxes on toswap and picks either the block’s data, or the argument read, which are each pairs of type rnd R' ∗ rnd R' . While toswap is in region R, the pairs are in independent region R' , so rule MuxFlip ensures the returned value is still in R' . If we were to store non-random data in the NORAM, we would require it to be in region R for this code to type check. A Language for Probabilistically Oblivious Computation 1:23 The add routine of NORAM is also simple: it adds the data into the bucket corresponding to the root node of the tree, which is the bucket at index 0 in the array. This is similar to the read_bucket routine above, so we omit its implementation. To avoid overﬂowing the root’s bucket due to repeated adds, a tree-based ORAM employs an additional eviction routine to evict blocks in the buckets of NORAM nodes closer to the leaf buckets. This routine should also maintain the key invariant: each data block should always reside on the path corresponding to its position tag. Diﬀerent tree-based ORAM implementations diﬀer only in their choices of b and the eviction strategies. One simple eviction strategy picks random nodes at each level of the tree, reads a single non-empty block from its bucket, and then writes that block one level further down either to the left or right; a dummy block is written in the opposite direction to make the operation oblivious. 6.2 Tree-based (recursive) ORAM To use tree-based NORAM we need a way to get the position tag for reads and writes. A tree-based recursive ORAM employs an additional position map to solve this issue. Such a map PM maps any index i ∈ 0..N , where N is the capacity of ORAM, to randomly generated position tag, which will be given to the rr and add routines of NORAM. The Tree ORAM tree_rr routine will ﬁrst access PM to get the position tag, and invoke NORAM’s rr routine to retrieve the data block. As mentioned in Section 3.2, one way to implement the position map is to just use a regular array stored in hidden memory, e.g., on-chip in a secure processor deployment of ORAM. However, this is not possible for MPC-based deployments, the adversary can observe the access pattern on the map itself. To avoid this problem, we can use another smaller Tree-based ORAM to store PM. In particular, the ORAM PM has N /c blocks, where each block contain c elements, where c ≥ 2 is a parameter of the ORAM. Therefore, PM also contains another position map PM ′ , which can be stored as an ORAM of capacity N /c 2 . Such a construction can continue recursively, until the position map at the bottom is small enough to be constructed as a Trivial ORAM. Therefore, there are logc N NORAMs constructed to store those position maps, and thus the overall runtime of a recursive ORAM is logc N times of the runtime of a NORAM. In the following, we implement the ORAM with c = 2.3 A full ORAM thus has the type oram, given below. 1 2 3 type oram R R' = (( noram R R') array ) ∗ trivial_oram R R' type trivial_oram R R' = ( to_block R R ') array type to_block R R' = bit R ∗ nat R ∗ (rnd R' ∗ rnd R ') where R ⊥ ⊥ R' In short, an ORAM is a sequence of NORAMs, where the ﬁrst in the sequence contains the actual data, and the remainder serve as the position map, which eventually terminate with a trivial ORAM. A trivial ORAM is simply an array of blocks which contain the isdummy ﬂag, the index, and the data (recall that trivial ORAMs themselves do not use position tags in their implementations). The data is a pair (i.e., c = 2) of position tags. The main idea is to think of the position map as essentially an array of size N but “stored” as a 2-D array: array[N/2][2]. In doing so, to access PM[idx], we essentially access PM[idx/2][idx mod 2]. We implement the tree_rr as a call to the function tree_rr_rec , which takes an additional public level argument, to indicate at which point in the list of NORAMs to start its work; it’s initially called with level 0. 1 2 3 let rec tree_rr_rec (oram:oram R R') ( idx : nat R) ( level : nat P) : (rnd R' ∗ rnd R ') = let (norams,tr_oram) = oram in let levels :nat P = length(norams) in 3 We present a readable, almost-correct version of the code ﬁrst, and give the full picture in Section 6.3. 1:24 4 5 6 7 8 9 10 11 David Darais, Chang Liu, Ian Sweet, and Michael Hicks if ( level >= levels ) then ... (∗ do trivial ORAM case, returning block (rnd R' ∗ rnd R ') ∗) else let ( r0 , r1) : rnd R' ∗ rnd R' = tree_rr_rec oram (idx /2) ( level +1) in let r0 ', tag = mux ((idx mod 2) = 0) (rnd R' () ) r0 in let r1 ', tag = mux ((idx mod 2 ) = 1) tag r1 in let _ = tree_add_rec oram (idx /2) ( level +1) ( r0 ', r1 ') in rr norams[level] idx (reveal tag) let tree_rr (oram:oram R R') ( idx : nat R) : (rnd R' ∗ rnd R ') = tree_rr_rec oram idx 0 Line 4 checks whether we have hit the base case of the recursion, in which case we lookup idx in the tr_oram, returning back a rnd R' ∗ rnd R' pair (code not shown). Otherwise, we enter the recursive case. Here, line 6 essentially reads out PM[idx/2], and lines 7 and 8 obliviously read out PM[idx/2][idx mod 2] into tag, replacing it with a freshly generated tag, to satisfy the aﬃnity requirement. Finally, line 8 writes the updated block PM[idx/2] back (i.e., ( r0 ', r1 ') ), using an analogous tree_add_rec routine, for which a level can be speciﬁed. Finally, line 9 reveals the retrieved position tag for index idx , so that it can be passed into the rr routine of NORAM. Level 0 corresponds to the actual data of the ORAM, which is returned to the client. The tree_add routine is similar so we do not show it all. As with tree_rr it recursively adds the corresponding bits of the position tag into the array of NORAMs. At each level of the recursion there is a snippet like the following: 1 2 3 4 5 6 7 let new_tag = rnd R' () in let sec_tag = use new_tag in (∗ does NOT consume new_tag ∗) let ( r0 , r1 ) : rnd R' ∗ rnd R' = tree_rr_rec oram (idx /2) ( level +1) in let r0 ', tag = mux ((idx mod 2) = 0) new_tag r0 in (∗ replaces with new tag ∗) let r1 ', tag = mux ((idx mod 2 ) = 1) tag r1 in let _ = tree_add_rec oram (idx /2) ( level +1) ( r0 ', r1 ') in add norams[level] idx sec_tag data (∗ adds to Tree ORAM ∗) Lines 1 and 2 generate a new tag, and make a secret copy of it. The new tag is then stored in the recursive ORAM—lines 3–5 are similar to tree_add_rec but replace the found tag with new_tag, not some garbage value, at the appropriate level of the position map (line 6). Finally, sec_tag is used to store the data in the appropriate level of the NORAM. 6.3 Limitations The astute reader may have noticed that the code snippet for add will not type check. In particular, the sec_tag argument has type nat R' but add requires it to have type R. This is because the position tags for the NORAM at level are stored as the data of the NORAM at level +1, and these are in diﬀerent regions. We cannot put them in the same region because we require a single NORAM’s metadata and data to reside in diﬀerent regions. We can solve this problem by having each level use the opposite pair of regions as the one above it. This solves the type error and does not compromise the NORAM independence requirement. Unfortunately, it means our oram type deﬁnition is made a bit more awkward, essentially it will be 1 type oram R R' = (( noram R R' ∗ noram R' R) array) ∗ trivial_oram R R' Basically, tree_rr_rec will have to operate two levels at a time in order to satisfy the type checker, but the basic logic will be the same. In addition, tree_rr_rec and tree_rr require the type of idx to be nat S, i.e., probability region ⊥ rather than R. Because ⊥ ⊑ R and ⊥ ⊑ R' in the region lattice, idx can be passed to rr at both of the unrolled levels, one of which requires it to be R and the other, R' . A more unfortunate limitation is that we cannot implement an NORAM on which we can also implement an oblivious stack. To see why, consider the rr operation implemented above, but with A Language for Probabilistically Oblivious Computation 1:25 the data having type nat R ∗ rnd R' , as required for an oblivious stack (the ﬁrst component is the stack element’s value, and the second is the position tag of the stack’s next element). The issue is that the stack element’s position tag is from the same region as the position tags in the underlying NORAM, which violates the independence requirement of the mux on line 11 of rr . However, this violation is a false positive: While the tags happen to be in the same region, they are actually independent—no tag’s generation is conditioned on any other’s value, and the same tag is never stored in the same block. Indeed, using the more general NORAM signature presented in Section 3.2, L obliv ’s type system conﬁrms that the stack in Figure 3 is secure. Checking the implementation of the stack and the NORAM end-to-end likely requires a more ﬁne-grained, generative treatment of regions and the use of existential types. Such features may simplify the solution to recursion problem as well. We are pursuing them in ongoing work. 7 RELATED WORK Lampson ﬁrst pointed out various covert, or “side,” channels of information leakage during a program’s execution [Lampson 1973]. Defending against side-channel leakage is challenging. Previous works have attempted to thwart such leakage from various angles: • Processor architectures that mitigate leakage through timing [Kocher et al. 2004; Liu et al. 2012], power consumption [Kocher et al. 2004], or memory-traces [Fletcher et al. 2014; Liu et al. 2015; Maas et al. 2013; Ren et al. 2013]. • Program analysis techniques that formally ensure that a program has bounded or no leakage through instruction traces [Molnar et al. 2006], timing channels [Agat 2000; Molnar et al. 2006; Russo et al. 2006; Zhang et al. 2012, 2015], or memory traces [Liu et al. 2015, 2013, 2014]. • Algorithmic techniques that transform programs and algorithms to their side-channel-mitigating or side-channel-free counterparts while introducing only mild costs – e.g., works on mitigating timing channel leakage [Askarov et al. 2010; Barthe et al. 2010; Zhang et al. 2011], and on preventing memory-trace leakage [Blanton et al. 2013; Eppstein et al. 2010; Goldreich 1987; Goldreich and Ostrovsky 1996; Goodrich et al. 2012; Shi et al. 2011; Stefanov et al. 2013; Wang et al. 2015, 2014; Zahur and Evans 2013]. Often, the most eﬀective and eﬃcient is through a comprehensive co-design approach combining these areas of advances – in fact, several aforementioned works indeed combine (a subset of) algorithms, architecture, and programming language techniques [Fletcher et al. 2014; Liu et al. 2015; Ren et al. 2013; Zhang et al. 2012, 2015]. Our core language not only generalizes a line of prior works on timing channel security [Agat 2000], program counter security [Molnar et al. 2006], and memory-trace obliviousness [Liu et al. 2015, 2013, 2014], but also provides the ﬁrst distinct core language that captures the essence of memory-trace obliviousness without treating key mechanisms—notably, the use of randomness— as a black box. Thus we can express state-of-the-art algorithmic results and formally reason about the security of their implementations, thus building a bridge between algorithmic and programming language techniques. ObliVM [Liu et al. 2015] is a language for programming oblivious algorithms intended to be run as secure multiparty computations [Yao 1986]. Its type system also employs aﬃne types to ensure random numbers are used at most once. However, there it provides no mechanism to disallow constructing a non-uniformly distributed random number. When such random numbers are generated, they can be distinguished by an attacker from uniformly distributed random numbers when being revealed. Therefore, the type system in ObliVM does not guarantee obliviousness. L obliv ’s use of probability regions enforces that all random numbers are uniformly random, and thus eliminates 1:26 David Darais, Chang Liu, Ian Sweet, and Michael Hicks this channel of information leakage. Moreover, we prove that this mechanism (and the others in L obliv ) are suﬃcient to prove PMTO. Our work belongs to a large category of work that aims to statically enforce noninterference, e.g., by typing [Sabelfeld and Myers 2006; Volpano et al. 1996]. Our probabilistic memory trace obliviousness property bears some resemblance to probabilistic notions of noninterference. Much prior work [Ngo et al. 2014; Russo and Sabelfeld 2006; Sabelfeld and Sands 2000; Smith 2003] is concerned with how random choices made by a thread scheduler could cause the distribution of visible events to diﬀer due to the values of secrets. Here, the source of nondeterminism is the (external) scheduler, rather than the program itself, as in our case. Smith and Alpízar [2006, 2007] consider how the inﬂuence of random numbers may aﬀect the likelihood of certain outcomes, mostly being concerned with termination channels. Their programming model is not as rich as ours, as a secret random number is never permitted to be made public; such an ability is the main source of complexity in L obliv , and is crucial for supporting oblivious algorithms. Some prior work aims to quantify the information released by a (possibly randomized) program (e.g., Köpf and Rybalchenko [2013]; Mu and Clark [2009]) according to entropy-based measures. Work on verifying the correctness of diﬀerentially private algorithms [Barthe et al. 2013; Zhang and Kifer 2017] essentially aims to bound possible leakage; by contrast, we enforce that no information leaks due to a program’s execution. 8 CONCLUSIONS This paper has presented L obliv , a core language suitable for expressing computations whose execution should be oblivious to a powerful adversary who can observe an execution’s trace of instructions and memory accesses, but not see private values. Unlike prior formalisms, L obliv can be used to express probabilistic algorithms whose security depends crucially on the use of randomness. To do so, L obliv tracks the use of randomly generated numbers via a substructural (aﬃne) type system, and employs a novel concept called probability regions. The latter are used to track a random number’s probabilistic (in)dependence on other random numbers. We have proved that together these mechanisms ensure that a random number’s revelation in the visible trace does not perturb the distribution of possible events so as to make secrets more likely. We have demonstrated that L obliv ’s type system is powerful enough to accept sophisticated algorithms, including forms of oblivious RAMs and oblivious data structures. Such data structures were out of the reach of prior type systems. As such, our proof-of-concept implementations represent the ﬁrst automated proofs that these algorithms are secure. We are currently pursuing two threads of ongoing work. First, we are working on a more powerful notion of probability region that is ﬁner grained so as to accept oblivious data structure algorithms, end to end. Second, we are working to expand the general expressiveness of the language. For example, to support containers requires some form of parametric polymorphism, and to support some deployments of ORAM requires bit-width polymorphism, implementable with dependent types. Ultimately we hope to integrate our ideas into ObliVM [Liu et al. 2015], a stateof-the-art compiler for oblivious algorithms, and thereby ensure that well-typed programs are indeed secure. ACKNOWLEDGMENTS We thank Aseem Rastogi and Kesha Heitala for comments on earlier drafts of this paper, and Elaine Shi for helpful suggestions and comments on the work while it was underway. This material is based upon work supported by the National Science Foundation under Grant Nos. CNS-1563722, A Language for Probabilistically Oblivious Computation 1:27 CNS-1314857, and CNS-1111599, and by DARPA under contracts FA8750-15-2-0104 and FA875016-C-0022. Any opinions, ﬁndings, and conclusions or recommendations expressed in this material are those of the author and do not necessarily reﬂect the views of the National Science Foundation. REFERENCES Johan Agat. 2000. Transforming out Timing Leaks. In POPL. Aslan Askarov, Danfeng Zhang, and Andrew C. Myers. 2010. Predictive black-box mitigation of timing channels. In CCS. Gilles Barthe, Boris Köpf, Federico Olmedo, and Santiago Zanella Béguelin. 2013. Probabilistic Relational Reasoning for Diﬀerential Privacy. ACM Trans. Program. Lang. Syst. 35, 3 (2013), 9:1–9:49. Gilles Barthe, Tamara Rezk, Alejandro Russo, and Andrei Sabelfeld. 2010. Security of multithreaded programs by compilation. ACM Transactions on Information and System Security (TISSEC) 13, 3 (2010), 21. Marina Blanton, Aaron Steele, and Mehrdad Alisagari. 2013. Data-oblivious Graph Algorithms for Secure Computation and Outsourcing. In ASIA CCS. David Brumley and Dan Boneh. 2003. Remote Timing Attacks Are Practical. In USENIX Security. D. Dolev and A. C. Yao. 1981. On the Security of Public Key Protocols. In Proceedings of the 22nd Annual Symposium on Foundations of Computer Science (SFCS). David Eppstein, Michael T. Goodrich, and Roberto Tamassia. 2010. Privacy-preserving data-oblivious geometric algorithms for geographic data. In GIS. Christopher W. Fletcher, Ling Ren, Xiangyao Yu, Marten van Dijk, Omer Khan, and Srinivas Devadas. 2014. Suppressing the Oblivious RAM timing channel while making information leakage and program eﬃciency trade-oﬀs. In HPCA. J.A. Goguen and J. Meseguer. 1982. Security policy and security models. In IEEE S & P. O. Goldreich. 1987. Towards a theory of software protection and simulation by oblivious RAMs. In STOC. O. Goldreich, S. Micali, and A. Wigderson. 1987. How to play ANY mental game. In STOC. Oded Goldreich and Rafail Ostrovsky. 1996. Software protection and simulation on oblivious RAMs. J. ACM (1996). Michael T. Goodrich, Olga Ohrimenko, and Roberto Tamassia. 2012. Data-Oblivious Graph Drawing Model and Algorithms. CoRR abs/1209.0756 (2012). Matt Hoekstra. 2015. Intel SGX for Dummies (Intel SGX Design Objectives). https://software.intel.com/en-us/blogs/2013/09/26/protecting-application-secrets-with-intel-sgx. (2015). Mohammad Islam, Mehmet Kuzu, and Murat Kantarcioglu. 2012. Access Pattern disclosure on Searchable Encryption: Ramiﬁcation, Attack and Mitigation. In Network and Distributed System Security Symposium (NDSS). Paul Kocher, Ruby Lee, Gary McGraw, and Anand Raghunathan. 2004. Security As a New Dimension in Embedded System Design. In Proceedings of the 41st Annual Design Automation Conference (DAC ’04). 753–760. Moderator-Ravi, Srivaths. Paul C. Kocher. 1996. Timing Attacks on Implementations of Diﬃe-Hellman, RSA, DSS, and Other Systems. In CRYPTO. Boris Köpf and Andrey Rybalchenko. 2013. Automation of quantitative information-ﬂow analysis. In Formal Methods for Dynamical Systems. Butler W. Lampson. 1973. A Note on the Conﬁnement Problem. Commun. ACM (1973). Chang Liu, Austin Harris, Martin Maas, Michael Hicks, Mohit Tiwari, and Elaine Shi. 2015. GhostRider: A HardwareSoftware System for Memory Trace Oblivious Computation. In ASPLOS. Chang Liu, Michael Hicks, and Elaine Shi. 2013. Memory Trace Oblivious Program Execution. In CSF. Chang Liu, Yan Huang, Elaine Shi, Jonathan Katz, and Michael Hicks. 2014. Automating Eﬃcient RAM-Model Secure Computation. In IEEE S & P. Chang Liu, Xiao Shaun Wang, Kartik Nayak, Yan Huang, and Elaine Shi. 2015. ObliVM: A Programming Framework for Secure Computation. In IEEE S & P. Isaac Liu, Jan Reineke, David Broman, Michael Zimmer, and Edward A. Lee. 2012. A PRET microarchitecture implementation with repeatable timing and competitive performance. In ICCD. Martin Maas, Eric Love, Emil Stefanov, Mohit Tiwari, Elaine Shi, Kriste Asanovic, John Kubiatowicz, and Dawn Song. 2013. Phantom: Practical Oblivious Computation in a Secure Processor. In CCS. David Molnar, Matt Piotrowski, David Schultz, and David Wagner. 2006. The Program Counter Security Model: Automatic Detection and Removal of Control-ﬂow Side Channel Attacks. In ICISC. Chunyan Mu and David Clark. 2009. An abstraction quantifying information ﬂow over probabilistic semantics. In Workshop on Quantitative Aspects of Programming Languages (QAPL). Tri Minh Ngo, Mariëlle Stoelinga, and Marieke Huisman. 2014. Eﬀective veriﬁcation of conﬁdentiality for multi-threaded programs. Journal of computer security 22, 2 (2014). Ling Ren, Xiangyao Yu, Christopher W. Fletcher, Marten van Dijk, and Srinivas Devadas. 2013. Design space exploration and optimization of path oblivious RAM in secure processors. In ISCA. 1:28 David Darais, Chang Liu, Ian Sweet, and Michael Hicks Alejandro Russo, John Hughes, David A. Naumann, and Andrei Sabelfeld. 2006. Closing Internal Timing Channels by Transformation. In Annual Asian Computing Science Conference (ASIAN). Alejandro Russo and Andrei Sabelfeld. 2006. Securing interaction between threads and the scheduler. In CSF-W. A. Sabelfeld and A. C. Myers. 2006. Language-based Information-ﬂow Security. IEEE J.Sel. A. Commun. 21, 1 (Sept. 2006). Andrei Sabelfeld and David Sands. 2000. Probabilistic noninterference for multi-threaded programs. In CSF-W. Elaine Shi, T.-H. Hubert Chan, Emil Stefanov, and Mingfei Li. 2011. Oblivious RAM with O ((log N )3 ) Worst-Case Cost. In ASIACRYPT. Geoﬀrey Smith. 2003. Probabilistic noninterference through weak probabilistic bisimulation. In CSF-W. Geoﬀrey Smith and Rafael Alpízar. 2006. Secure Information Flow with Random Assignment and Encryption. In Workshop on Formal Methods in Security (FMSE). Geoﬀrey Smith and Rafael Alpízar. 2007. Fast Probabilistic Simulation, Nontermination, and Secure Information Flow. In PLAS. Emil Stefanov, Marten van Dijk, Elaine Shi, Christopher Fletcher, Ling Ren, Xiangyao Yu, and Srinivas Devadas. 2013. Path ORAM – an Extremely Simple Oblivious RAM Protocol. In CCS. G. Edward Suh, Dwaine Clarke, Blaise Gassend, Marten van Dijk, and Srinivas Devadas. 2003. AEGIS: architecture for tamper-evident and tamper-resistant processing. In ICS. David Lie Chandramohan Thekkath, Mark Mitchell, Patrick Lincoln, Dan Boneh, John Mitchell, and Mark Horowitz. 2000. Architectural support for copy and tamper resistant software. SIGOPS Oper. Syst. Rev. 34, 5 (Nov. 2000). Dennis Volpano, Cynthia Irvine, and Geoﬀrey Smith. 1996. A Sound Type System for Secure Flow Analysis. J. Comput. Secur. 4, 2-3 (Jan. 1996). Xiao Wang, Hubert Chan, and Elaine Shi. 2015. Circuit ORAM: On tightness of the Goldreich-Ostrovsky lower bound. In CCS. Xiao Shaun Wang, Kartik Nayak, Chang Liu, T-H. Hubert Chan, Elaine Shi, Emil Stefanov, and Yan Huang. 2014. Oblivious Data Structures. In CCS. Andrew Chi-Chih Yao. 1986. How to generate and exchange secrets. In FOCS. Samee Zahur and David Evans. 2013. Circuit Structures for Improving Eﬃciency of Security and Privacy Tools. In S & P. Danfeng Zhang, Aslan Askarov, and Andrew C. Myers. 2011. Predictive Mitigation of Timing Channels in Interactive Systems. In CCS. Danfeng Zhang, Aslan Askarov, and Andrew C. Myers. 2012. Language-based Control and Mitigation of Timing Channels. In PLDI. Danfeng Zhang and Daniel Kifer. 2017. LightDP: Towards Automating Diﬀerential Privacy Proofs. In POPL. Danfeng Zhang, Yao Wang, G. Edward Suh, and Andrew C. Myers. 2015. A Hardware Design Language for Timing-Sensitive Information-Flow Security. In ASPLOS. Xiaotong Zhuang, Tao Zhang, and Santosh Pande. 2004. HIDE: an infrastructure for eﬃciently protecting information leakage on the address bus. SIGARCH Comput. Archit. News 32, 5 (Oct. 2004). A Language for Probabilistically Oblivious Computation A APPENDIX 1:29 1:30 David Darais, Chang Liu, Ian Sweet, and Michael Hicks x x®N x®N x̃ R i θ ∈X ∈ XN ∈ X∗ eR ≔ BR → X ∈X ∈ idxN ≔ {1, .., N } ∈ dep ≔ ℘(N) Notation for elements x of a parameterized set X Sequence of elements X of length N Sequence of elements X of some length N Distributions of elements X over R coin ﬂips Index into sequences of length N Probabilistic Dependency Set n o b® x̃ b® = x c [x̃ = x] p [x̃ = x] ≔ 2R eR X c [x̃ = x] ≔ c[· = ·] : ∀R X . ×X → N eR × X → Q p[· = ·] : ∀R X . X c[· = · ∧ · = ·] : ∀R X 1 X 2 . p[· = · ∧ · = ·] : ∀R X 1 X 2 . p[· = · \| · = ·] : ∀R X 1 X 2 . eR × X 1 × X eR × X 2 → N X 1 2 R e × X1 × X eR × X 2 → Q X 1 2 eR × X 1 × X eR × X 2 → Q X 1 2 o x̃ 1 b® = x 1 ∧ x̃ 2 b® = x 2 c [x̃ 1 = x 1 ∧ x̃ 2 = x 2 ] p [x̃ 1 = x 1 ∧ x̃ 2 = x 2 ] ≔ 2R c [x̃ 1 = x 1 ∧ x̃ 2 = x 2 ] p [x̃ 1 = x 1 \| x̃ 2 = x 2 ] ≔ c [x̃ 2 = x 2 ] c [x̃ 1 = x 1 ∧ x̃ 2 = x 2 ] ≔ n b® eR × N eR → N eR ·e ⊙ · : ∀R. N g : ∀R X . e eR × X eR → X eR cond BR × X R R R e e eR × X eR e mux g : ∀R X . B × X × X → X ® J⊙K ñ 1 b® , ñ 2 b® ñ 1 e ⊙ ·ñ 2 ≔ λb.    x̃ 2 b® if x̃ 1 b® = O ® g (x̃ 1 , x̃ 2 , x̃ 3 ) ≔ λb. cond  x̃ 3 b® if x̃ 1 b® = I  D E g (x̃ 1 , x̃ 2 , x̃ 3 ) , cond g (x̃ 1 , x̃ 3 , x̃ 2 ) mux g (x̃ 1 , x̃ 2 , x̃ 3 ) ≔ cond eR ⌊·⌋ : ∀R X . X → X R eR sel[·] : ∀R X . idx → X eR → X eR+R ′ ↑˜ : ∀R R ′ X . X x̃ 1 ≈ x̃ 2 ≔ x̃ 1 ≈ θ x̃ 2 ≔ stable [x̃ θ ] ≔ x̃ 1 ⊥ ⊥ x̃ 2 ≔ x̃ ∝ θ ≔ ⌊x⌋ ≔ sel[i] ≔ ↑˜ x̃ R ≔ λ . x λb®R . bi ′ λb®R+R . x̃ R (b1 . . . bR ) ∀x . p [x̃ 1 = x] = p [x̃ 2 = x] ∀x . p [x̃ 1 = x \| θ ] = p [x̃ 2 = x \| θ ] ∀x 1 x 2 . p [x̃ = x 1 \| θ ] = p [x̃ = x 2 \| θ ] ∀x 1 x 2 . p [x̃ 1 = x 1 ∧ x̃ 2 = x 2 ] = p [x̃ 1 = x 1 ] p [x̃ 2 = x 2 ] ∀i < θ . x̃ ⊥ ⊥ sel[i] Fig. 8. Distributions A Language for Probabilistically Oblivious Computation α ∈ addr ≔ {0, 1, . . .} bR F ι̃ R α̃ R vbR ∈ V ℓ b bR → V b ↑ : ∀R R ′ . V bR ⊎ {•} γbR ∈ env c R ≔ var ⇀ V bR ∗ σ R ∈ env b c R ≔ addr ⇀ V vbR , vbR b ↑ (˜ι ) ≔ ↑˜ (ι̃) b ↑ (α̃) ≔ ↑D˜ (α̃) E b ↑ (hb v 1 , vb2 i) ≔ b ↑b v1, b ↑b v2 R+R ′ ′ b ↑ : ∀R R ′ . env c R → env c R+R env b d R → store d R+R ↑ : ∀R R ′ . store store 1:31 ′ env b ↑b γ (x) ≔ b ↑b γ (x) ! store b ↑b σ (α)n ≔ b ↑b σ (α)n d : ∀R. V bR × V bR × V bR ⇀ V bR cond bR × V bR × V bR ⇀ V bR × V bR mux d : ∀R. V d b̃ ℓ , ι˜1 , ι˜2 cond d b̃ ℓ , α̃ 1 , α̃ 2 cond d b̃ ℓ , hb cond v 11 , vb21 i , hb v 12 , vb22 i mux d b̃ ℓ , vb1, vb2 R g bR ⇀ V bR join : ∀R. V g b̃, ι˜1 , ι̃ 2 ≔ cond g b̃, α̃ 1 , α̃ 2 ≔ cond E D d b̃ ℓ , vb21, vb22 d b̃ ℓ , vb11 , vb12 , cond ≔ cond E D d b̃ ℓ , vb2, vb1 d b̃ ℓ , vb1, vb2 , cond ≔ cond ® ι˜˜ b® b® join ι˜˜ ≔ λb. ® α̃˜ b® b® join α̃˜ ≔ λb. E D join hb v 1 , vb2 i ≔ join vb̃1 , join vb̃2 gR ×N bR e R × store dR ⇀ V read : ∀R. addr R g ×N bR × store bR × store eR × V dR ⇀ V dR read+write : ∀R. addr ® b read (α̃, ñ ℓ , b σ ) ≔ join λb. σ α̃ b® [ñ b® ] read+write (α̃, ñ ℓ , vb, b σ ) ≔ hread (α̃, ñ, b σ),b σ ′i   ® = α and ñ b® = n α̃ b v b if   ª © ® where b σ ′(α)[n] ≔ join λb. ®  ® ® σ (α)[n] if α̃ b , α or ñ b , n b « ¬  Fig. 9. Domains 1:32 David Darais, Chang Liu, Ian Sweet, and Michael Hicks bR × env V cR ′ ′ bR ′ × env d R ⇀ ∃R ′ . V dR EJ·K : atom → ∀R. env c R × store c R × store exp ′ ′ bR ′ × env d R ⇀ ∃R ′ . V dR EJ·K : expr → ∀R. env c R × store c R × store EJ·K : pico → ∀R. env cR pico ⇀ atom γ ) ≔ hb γ (x), b γi EJxK (b pico U EJxK (b γ ) ≔ hb γ (x), b γ [x 7→ •]i pico EJpK (b γ,b σ ) ≔ hb v, b γ ′, b σ i where hb v, b γ ′i ≔ EJpK (b γ) D E D E pico atom ◦ ′ g b̃, ⌊0⌋, ⌊1⌋ , b EJp K (b γ,b σ ) ≔ cond γ ,b σ γ) b̃, b γ ′ ≔ EJpK (b where atom EJp1 ⊙ p2 K (b γ,b σ ) ≔ ñ 1 e ⊙ ñ 2 , b γ ′′, b σ atom EJι ℓ K (b γ ) ≔ h⌊ι ℓ ⌋, b γi pico L pico hñ 1 , b γ ′i ≔ EJp1 K (b γ ) hñ 2 , b γ ′′ i ≔ EJp2 K (b γ ′) D E D E pico atom b̃ , b EJasnat(p)K (b γ) b̃, γb′ ≔ EJpK (b γ,b σ ) ≔ asnat where γ ′, b σ D E pico atom g (ñ) , b γ,b σ ) ≔ asbit where hñ, b γ ′ i ≔ EJpK (b γ) γ ′, b σ EJasbit(p)K (b * + env store atom ρ EJflip K b γ R, b σ R ≔ sel[R + 1], b ↑b γ R, b ↑b σR pico pico where ℓ γ,b σ ) ≔ hb γ (x), b γ, b σi EJuse(x)K (b atom EJreveal(x)K (b γ,b σ ) ≔ hb γ (x), b γ [x 7→ •], b σi atom γ,b σ ) ≔ hmux d (b v 1 , vb2, vb3 ) , b γ ′′′ , b σi EJmux(p1, p2 , p3 )K (b atom hb v 1, b γ ′ i ≔ EJp1 K (b γ) pico where hb v2, b γ ′′ i ≔ EJp2 K (b γ ′) pico EJhp1 , p2 iK (b γ,b σ ) ≔ hhb v 1 , vb2 i , b γ ′′, b σi atom where hb v1, b γ ′i ≔ EJp1 K (b γ) pico hñ 1 , b γ ′i ≔ EJp1 K (b γ) pico where hb v 2, b γ ′′′ i ≔ EJp2 K (b γ ′′ ) pico hb v2, b γ ′′ i ≔ EJp2 K (b γ ′) pico hñ 2 , b γ ′′ i ≔ EJp2 K (b γ ′) pico Fig. 10. Denotational Evaluation Semantics (1) A Language for Probabilistically Oblivious Computation 1:33 EJarrayN (p1 , . . . , p N )K (b γ0, b σ ) ≔ h⌊α⌋, b γn , b σ [α 7→ [n 7→ vbn ]]i atom hb vn , b γn , i ≔ EJpn K (b γn−1 ) pico where α ≔ fresh (b σ) EJp1 [p2 ]K (b γ,b σ ) ≔ hread (b v 1 , vb2, b σ) ,b γ ′′, b σi atom hb v1, b γ ′i ≔ EJp1 K (b γ) pico where hb v2, b γ ′′ i ≔ EJp2 K (b γ ′) pico EJp1 [p2 ] ← p3 K (b γ,b σ ) ≔ hb v, b γ ′′′ , b σ ′i atom hb v 1, b γ ′ i ≔ EJp1 K (b γ ) hb v2, b γ ′′ i ≔ EJp2 K (b γ ′ ) hb v 3, b γ ′′′ i ≔ EJp2 K (b γ ′′ ) ′ hb v, b σ i ≔ read+write (b v 1 , vb2, vb3, b σ) pico where pico pico EJaK (b γ,b σ ) ≔ EJaK (b γ,b σ) exp atom EJlet x ≔ a in eK (b γ,b σ ) ≔ EJeK (b γ ′ [x 7→ vb], b σ ′) exp exp hb v, b γ ′, b σ ′i ≔ EJaK (b γ, b σ) atom where γ,b σ ) ≔ EJeK (b γ ′ [x 1 7→ vb1 , x 2 7→ vb2 ] , b σ ′) EJlet x 1 , x 2 ≔ a in eK (b exp exp hhb v 1, vb2 i , b γ ′, b σ ′ i ≔ EJaK (b γ,b σ) atom where EJif p then e 1 else e 2 K (b γ,b σ ) ≔ hb v, b γ ′′, b σ ′i E D pico R vbR1 ⊔R2 ≔ γ) vbR , γb′ ≔ EJpK (b where E D exp R1 ⊔R2 (x) ≔ γ ′, b σ ) γc′′ vb1R1 , b γ 1R1 , b σ1R1 ≔ EJe 1 K (b E D exp R ⊔R ′ 1 2 (α) ≔ γ ′, b σ ) σb vb2R2 , b γ 2R2 , b σ2R2 ≔ EJe 2 K (b exp d b cond ↑v, b ↑v 1 , b ↑v 2 d b cond ↑v, b ↑γ 1 (x), b ↑γ 2 (x) d b cond ↑v, b ↑σ1 (α), b ↑σ2 (α) Fig. 11. Denotational Evaluation Semantics (2) 1:34 David Darais, Chang Liu, Ian Sweet, and Michael Hicks l ∈ loc F . . . source code locations . . . v ∈ V F ι ℓ \| α \| hv, vi γ ∈ env ≔ var ⇀ V σ ∈ store ≔ addr ⇀ V ∗ t ∈ trace ≔ (loc × V × env × store)∗ R eR × X eR → X · ⊗ · : ∀R X 1 X 2 . X 1 × X2 1 2 f∗R × X f∗ R → X f∗ e ·+ + · : ∀R X . X E D ® x̃ 1 b® , x̃ 2 b® x̃ 1 ⊗ x̃ 2 ≔ λb. e ® e e x®2 b® x®1 + +e x®2 ≔ λb. x®1 b® ++ e bR → V eR δ : ∀R. V V δ (˜ι ℓ ) ≔ ι˜ℓ δ (α̃ ) ≔ α̃ δ (hb v 1 , vb2 i) ≔ δ (b v 1 ) ⊗ δ (b v2) h i x→ 7 δ (b γ (x)) b® h h ii store ® α→ δ (b σ ) ≔ λb. 7 n 7→ δ (b σ (α)[n]) b® env δ : ∀R. env c R → eg nvR ® δ (b γ ) ≔ λb. env d R → store gR δ : ∀R. store store R d R ⇀ ∃R ′ . t T J·K : expr → ∀R. env c R × store race ′ T Jal K (b γ,b σ ) ≔ ⌊l⌋ ⊗ δ (b v ) ⊗ δ (b γ ′) ⊗ δ (b σ ′) env store where hb v, b γ ′,b σ ′ i ≔ EJaK (b γ,b σ) atom env store e γ,b σ ) ≔ ⌊l⌋ ⊗ δ (b v ) ⊗ δ (b γ ′) ⊗ δ (b σ ′) + +t T Jlet x ≔ al in eK (b hb v, b γ ′, b σ ′i ≔ EJaK (b γ, b σ ) t ≔ T JeK (b γ ′ [x 7→ vb] , b σ ′) env store e T Jlet x 1 , x 2 ≔ al in eK (b γ,b σ ) ≔ ⌊l⌋ ⊗ δ (b v ) ⊗ δ (b γ ′) ⊗ δ (b σ ′) + +t atom where where hhb v 1 , vb2 i , b γ ′, b σ ′i ≔ EJaK (b γ,b σ ) t ≔ T JeK (b γ ′ [x 1 7→ vb1, x 2 7→ vb2 ] , b σ ′) ′ ′ T Jif p then e 1 else e 2 K (b γ,b σ ) ≔ cond (b v , T Je 1 K (b γ ,b σ ) , T Je 2 K (b γ ,b σ )) atom hb v, b γ ′i ≔ EJpK (b γ,b σ) pico where Fig. 12. Denotational Trace Semantics A Language for Probabilistically Oblivious Computation w ∈ W F θ \| hw, wi pico ω ∈ penv ≔ var ⇀ W n all (x̃) ≔ x eR → ℘(X ) all : ∀R X . X DJ·K : pico → 1:35 ψ ∈ pstore ≔ addr ⇀ W o ® x̃ b® = x ∃b. penv ⇀ W d R × penv × pstore × dep ⇀ W × pstore × dep DJ·K : atom → ∀R. env c R × store atom d R × penv × pstore × dep ⇀ W × pstore × dep DJ·K : exp → ∀R. env c R × store exp pico pico DJxK(ω) ≔ ω(x) DJlK(ω) ≔ DJpK (b γ, b σ , ω,ψ , θ ) ≔ DJpK(ω),ψ , θ pico atom ◦ DJp K (b γ, b σ , ω,ψ , θ ) ≔ DJpK(ω),ψ , θ pico pico atom DJp1 ⊙ p2 K (b γ, b σ , ω,ψ , θ ) ≔ DJp1 K(ω) ⊔ DJp2 K(ω),ψ , θ pico atom γ, b σ , ω,ψ , θ ) ≔ DJpK(ω),ψ , θ DJasnat(p)K (b pico atom γ, b σ , ω,ψ , θ ) ≔ DJpK(ω),ψ , θ DJasbit(p)K (b atom ρ σ R , ω,ψ , θ ≔ h{R + 1},ψ , θ i DJflipℓ K γbR , b atom pico DJuse(x)K (b γ, b σ , ω,ψ , θ ) ≔ hω(x),ψ , θ i atom γ, b σ , ω,ψ , θ ) ≔ hω(x),ψ , θ ⊔ ω(x)i DJreveal(x)K (b atom DJmux(p1 , p2, p3 )K (b γ, b σ , ω,ψ , θ ) ≔ hhw, wi,ψ , θ i atom U pico pico pico where w ≔ DJp1 K(ω) ⊔ DJp2 K(ω) ⊔ DJp3 K(ω) DJmux(p1 , p2, p3 )K (b γ, b σ , ω,ψ , θ ) ≔ hhw, wi,ψ , θ i atom A pico pico where w ≔ DJp2 K(ω) ⊔ DJp3 K(ω) pico pico atom DJhp1 , p2 iK (b γ, b σ , ω,ψ , θ ) ≔ DJp1 K(ω), DJp2 K(ω) ,ψ , θ Fig. 13. Dependency Semantics (1) 1:36 David Darais, Chang Liu, Ian Sweet, and Michael Hicks pico pico atom DJarrayN (p1 , . . . , p N )K (b γ,b σ , ω,ψ , θ ) ≔ ,ψ α → 7 DJp1 K(ω) ⊔ · · · ⊔ DJp N K(ω) , θ α ≔ fresh(b σ) Ã DJp1 [p2 ]K (b γ,b σ , ω,ψ , θ ) ≔ ψ (α),ψ , θ where atom α hb v, b γ i ≔ EJp γ ) α ∈ all (b v) 1 K (b pico Ã DJp1 [p2 ] ← p3 K (b γ,b σ , ω,ψ , θ ) ≔ ψ (α),ψ ⊔ α 7→ DJp3 K(ω) , θ where atom where ′ hb v, b γ ′i pico α pico ≔ EJp1 K (b γ) α ∈ all (b v) DJaK (b γ,b σ , ω,ψ , θ ) ≔ DJaK (b γ,b σ ,ψ , ω, θ ) exp atom γ,b σ , ω,ψ , θ ) ≔ DJeK (b γ ′′, b σ ′, ω ′′ ,ψ ′, θ ′) DJlet x ≔ a in eK (b exp exp hb v, b γ ′, b σ ′i ≔ EJaK (b γ,b σ) γb′′ ≔ γb′ [x 7→ vb] ′′ ′ hw,ψ ′, θ ′i ≔ DJaK (b γ,b σ , ω,ψ , θ ) ω ≔ ω [x 7→ w] atom where atom γ,b σ , ω,ψ , θ ) ≔ DJeK (b γ ′′, b σ ′, ω ′′ ,ψ ′, θ ′) DJlet x 1 , x 2 ≔ a in eK (b exp exp hhvb1 , vb2 i , b γ ′, b σ ′ i ≔ EJaK (b γ, b σ) γb′′ ≔ γb′ [x 1 7→ vb1 , x 2 7→ vb2 ] ′′ ′ hhw 1 , w 2 i, ω ′,ψ ′, θ ′ i ≔ DJaK (b γ, b σ , ω, θ ) ω ≔ ω [x 1 7→ w 1 , x 2 7→ w 2 ] atom where atom γ,b σ , ω,ψ , θ ) ≔ DJpK(ω) ⊔ w 1 ⊔ w 2 ,ψ 1 ⊔ ψ 2 , θ 1 ⊔ θ 2 DJif p then e 1 else e 2 K (b exp pico hw 1 ,ψ 1 , θ 1 i ≔ DJe 1 K (b γ,b σ , ω,ψ , θ ) exp where hw 2 ,ψ 2 , θ 2 i ≔ DJe 2 K (b γ,b σ , ω,ψ , θ ) exp Fig. 14. Dependency Semantics (2) A Language for Probabilistically Oblivious Computation 1:37 ξ ∈ renv ≔ region → dep d × renv ⇀ renv RJ·K : atom → ∀R. env c R × store atom R d × renv ⇀ renv RJ·K : exp → ∀R. env c R × store exp R atom ρ σ R , ω,ψ , ξ ≔ ξ ⊔ [ρ ′ 7→ {R + 1} \| ρ ⊑ ρ ′ ] RJflipℓ K γbR , b RJ K (b γ,b σ , ω,ψ , ξ ) ≔ ξ atom RJaK (b γ,b σ , ξ ) ≔ RJaK (b σ, ξ) γ , b exp exp atom ′′ ′ γ,b σ , ξ ) ≔ RJeK b γ ,b σ , RJaK (b γ,b σ, ξ ) RJlet x ≔ a in eK (b exp atom hb v, b γ ′, b σ ′i ≔ EJaK (b γ,b σ ) γb′′ ≔ b γ ′ [x 7→ vb] exp exp atom RJlet x 1 , x 2 ≔ a in eK (b γ,b σ , ξ ) ≔ RJeK b γ ′′, b σ ′, RJaK (b γ,b σ, ξ ) atom where hhvb1, vb2 i , b γ ′, b σ ′i ≔ EJaK (b γ,b σ ) γb′′ ≔ b γ ′ [x 1 7→ vb1 , x 2 7→ vb2 ] atom where RJif p then e 1 else e 2 K (b γ,b σ , ξ ) ≔ RJe 1 K (b γ,b σ , ξ ) ⊔ RJe 2 K (b γ,b σ, ξ ) exp exp Fig. 15. Region Semantics exp 1:38 David Darais, Chang Liu, Ian Sweet, and Michael Hicks Bit Arr ∀α ∈ all α̃ R . Σ(α) = τ ι˜Rℓ ⋉ Φ Φ α̃ R ∝ θ θ Arr R α̃ ⋉ Flip Σ ⊢ ñ R : nat Tup Σ ⊢ α̃ R : array(τ ) ι̃ R ∝ Dist Nat Σ ⊢ b̃ R : bit Φ Φ ξ ⊢ β@hℓ, ρi ⋉ θ Arr Tup vb1R ⋉ Φ w1 R R vb1 , vb2 Tup ξ ⊢ array(τ ) ⋉ θ Σ ⊢ vb2R : τ2 Σ ⊢ vb1R , vb2R : τ1 × τ2 θ ⊆ ξ (ρ) Base Σ ⊢ b̃ R : flip Σ ⊢ vb1R : τ1 θ θ Σ ⊢ vbR : τ ⋉ Φ vb2R ⋉ vbR ⋉ vbR ⋉ hw 1 , w 2 i ξ ⊢τ ⋉w ξ ⊢ τ1 ⋉ w 1 ξ ⊢ τ ⋉w Φw Σ ; ξ ⊢ vb : τ ⋉ R Φ w Fig. 16. Value Well-formedness w Φ w2 ξ ⊢ τ2 ⋉ w 2 ξ ⊢ hτ1 , τ2 i ⋉ hw 1 , w 2 i Σ ;ξ ⊢v :τ ⋉ Σ ⊢ vbR : τ Φ Φ w A Language for Probabilistically Oblivious Computation 1:39 Γ ; Σ⊢b γ ≔ ∀x . Σ ⊢ γb(x) : Γ(x) γ ⋉ ω ≔ ∀x . γb(x) ⋉ b Φ ω(x) Γ ; Σ ⊢ γb γ⋉ b Φ ω ξ ⊢ Γ ⋉ω ξ ⊢ Γ ⋉ ω ≔ ∀x . ξ ⊢ Γ(x) ⋉ ω(x) Γ ; Σ ⊢ γb γb ⋉ Φ ξ ⊢ Γ ⋉ω ω Γ ; Σ;ξ ⊢b γ⋉ ω Φ Σ⊢b σ ≔ ∀α . Σ ⊢ b σ (α) : Σ(α) b σ ⋉ ψ ≔ ∀α . b σ (α) ⋉ Φ ψ (α) ξ ⊢ Σ ⋉ ψ ≔ ∀α . ξ ⊢ Σ(α  ⋉ ψ (α) Σ⊢b σ σ⋉ b Φψ Σ;ξ ⊢b σ⋉ ξ ⊢ Σ ⋉ψ Φψ Fig. 17. Env and Store Well-formedness Γ ; Σ;ξ ⊢b γ⋉ Φ ω Σ⊢b σ σ⋉ b Φψ ξ ⊢ Σ ⋉ψ Σ;ξ ⊢b σ⋉ Φψ 1:40 David Darais, Chang Liu, Ian Sweet, and Michael Hicks z ϕ π β ∈ loc F ∈ cloc F ∈ rloc F ∈ btype F R cc xt R rc xt tcxt trxt \| hz, •i \| h•, zi env (x, z) \| store (α, z) ret (z) \| ϕ bit \| nat \| flip eR bR × loc ⇀ lit ·@· : ∀R. V ≔ ≔ ≔ ≔ ι˜ℓ @ ≔ ι˜ vb, cxt hb γ,b σ i @store (α, z) ≔ b σ (α)@z cxt hb v, b γ, b σ i @ret (z) ≔ vb@z rxt R e ·@· : ∀R. rc xt × rloc ⇀ lit γ (x)@z hb v, b γ,b σ i @env (x, z) ≔ b R rxt hb v, b γ,b σ i @store (α, z) ≔ b σ (α)@z rxt ρ type bitℓ @ ≔ bit type ·@· : type × loc ⇀ btype @hz, •i ≔ vb@z , vb @h•, zi ≔ vb@z hb γ,b σ i @env (x, z) ≔ γb(x)@z R eR ·@· : ∀R. cc xt × cloc ⇀ lit cxt rxt dR env c R × store bR × env dR V c R × store tenv × tstore type × tenv × tstore ρ type natℓ @ type flipρ @ τ, ≔ nat type type type type @hz, •i ≔ τ @z , τ @h•, zi ≔ τ @z ≔ flip type tcxt hΓ, Σi @env (x, z) ≔ Γ(x)@z tcxt ·@· : tcxt × cloc ⇀ btype type tcxt hΓ, Σi @store (α, z) ≔ Σ(α)@z trxt trxt type hτ , Γ, Σi @ret (z) ≔ τ @z ·@· : ∀R. trxt × rloc ⇀ btype trxt type hτ , Γ, Σi @env (x, z) ≔ Γ(x)@z trxt type hτ , Γ, Σi @store (α, z) ≔ Σ(α)@z Fig. 18. Locations A Language for Probabilistically Oblivious Computation 1:41 ι̃ R :: β # ι̃ R :: β Flip b̃1 ⊥ ⊥ b̃2 Φ b̃1 :: flip # b̃2 :: flip Γ ; Σ ⊢ P(b γ,b σ) Other β 1 , flip ∨ β 2 , flip Φ ι˜1 :: β 1 # ι˜2 :: β 2 Φ Γ ; Σ ⊢ P(b γ,b σ) Φ≔ ∀ϕ 1 , ϕ 2 . hb γ,b σ i @ϕ 1 :: hΓ, Σi @ϕ 1 # hb γ,b σ i @ϕ 2 :: hΓ, Σi @ϕ 2 τ ; Γ ; Σ ⊢ P(b v, b γ,b σ) cxt tcxt cxt tcxt Φ≔ ∀π1 , π2 . hb v, b γ, b σ i @π1 :: hτ , Γ, Σi @π1 # hv, b γ, b σ i @π2 :: hτ , Γ, Σi @π2 trxt rxt Φ Φ τ ; Γ ; Σ ⊢ P(b v, b γ,b σ) rxt Φ trxt Φ Φ Fig. 19. Partitioning S (˜ι :: β) h i stable b̃ Φ Flip Φ S b̃ :: flip Γ ; Σ ⊢ S(b γ,b σ) Γ ; Σ ⊢ S(b v, b γ, b σ) β , flip Other S (˜ι :: β) cxt tcxt Φ ≔ ∀ϕ. S hb γ,b σ i @ϕ :: hΓ, Σi @ϕ Φ Γ ; Σ ⊢ S(b γ,b σ) Φ Φ Γ ; Σ ⊢ S(b v, b γ,b σ) cxt tcxt Φ ≔ ∀π . S hb v, b γ, b σ i @π :: hτ , Γ, Σi @π Fig. 20. Stability Φ Φ Φ 1:42 David Darais, Chang Liu, Ian Sweet, and Michael Hicks l • ι • v • γ ∈ ∈ ∈ ∈ loc F • lit F • VF • env ≔ • • • . . . source code locations . . . bP \| nP \| • • • • ι \| α \| hv , v i • var ⇀ V ∗ σ ∈ store ≔ addr ⇀ V • • • • • • • ς ∈ conﬁg ≔ he, γ , σ i \| HALT (v , γ , σ ) • ∗ • • t ∈ trace ≔ conﬁg lit • •R b• R V F ι̃ lit lit • obs(ι P ) ≔ ι obs : lit × label → lit obs : V → V D R RE • • b α̃ R v v ,b • R •R b• b c γ ∈ env ≔ var ⇀ V • R •R b• d b ≔ addr ⇀ V ∗ σ ∈ store •R b v ∈ obs(ι S ) ≔ • lit obs(ι ℓ ) ≔ obs(ι ℓ ) env obs(hv 1 , v 2 i) ≔ hobs(v 1 ), obs(v 2 )i obs(α) ≔ α env • obs : env → env store obs(γ )(ν ) ≔ obs(γ (ν )) store • obs : store → store obs(σ )(α)[n] ≔ obs(σ (α)[n]) D env E env obs(he, γ , σ i) ≔ e, obs(γ ), obs(σ ) conﬁg V env env obs(HALT (v, γ , σ )) ≔ HALT obs(v), obs(γ ), obs(σ ) conﬁg conﬁg • obs : conﬁg → conﬁg trace • obs : conﬁg∗ → conﬁg ∗ b• R c : ∀R. V bR → V obs c (α̃ ) ≔ α̃ obs • R c c : ∀R. env obs c R → env env • R d c : ∀R. store d R → store obs store conﬁg trace conﬁg obs(ς 1 . . . ς N ) ≔ obs(ς 1 ) . . . obs(ς N )    ι if ι˜ℓ b® = ι P ® c (ι̃ ℓ ) ≔ λb. obs   • if ι˜ℓ b® = ι S  E D E D • • • • c b c b c b obs v2 v 1, b v 2 ≔ obs v 1 , obs env • • c b c b obs γ (ν ) ≔ obs γ (ν ) store • • c b c b obs σ (α)n ≔ obs σ (α)n Fig. 21. Adversary Observations A Language for Probabilistically Oblivious Computation 1:43 B THEOREMS Theorem B.1 (Type Safety). If Γ ; Σ ⊢ a : τ ; Γ′ then Γ ; Σ⊢b γ Σ⊢b σ ∃Σ ′ vb b γ′ b σ ′. EJaK (b γ, b σ ) = hb v, b γ ′, b σ ′i ′ ′ ′ Σ ⊢ vb : τ Γ ; Σ ⊢ γb Σ ′ ⊢ b σ′ ′ Deﬁnition B.2 (Well Typed Evaluation). Γ ; Σ ⊢ a : τ ; Γ ′ ; Σ ′ ;; hb γ,b σ i ⇓ hb v, b γ ′, b σ ′i Γ ; Σ ⊢ a : τ ; Γ′ Γ ; Σ⊢b γ hb v, b γ ,σ b i ≔ EJaK (b γ,b σ) Γ ′ ; Σ ′ ⊢ vb : τ ′ ′ Σ⊢b σ Γ ′ ; Σ ′ ⊢ γb′ Γ ; Σ ⊢ a : τ ; Γ ′ ; Σ ′ ;; hb γ,b σ i ⇓ hb v, b γ ′, b σ ′i Σ′ ⊢ b σ′ Theorem B.3 (Dependency Soundness). If Γ ; Σ ⊢ a : τ ; Γ ′ ; Σ ′ ;; hb γ,b σ i ⇓ hb v, b γ ′, b σ ′i γb ⋉ ω c (b obs γ) ⋉ Φ vb ⋉ w env c (b b ⋉ψ σ obs σ) ⋉ Φ DJaK (b γ, b σ , ω,ψ , Φ) = hw,ψ ′ , Φ′i store then b′ ⋉ ψ ′ σ c (b obs v ) ⋉ Φ′ c (b obs σ ′ ) ⋉ Φ′ store Theorem B.4 (Region Soundness). If Γ ; Σ ⊢ a : τ ; Γ ′ ; Σ ′ ;; hb γ,b σ i ⇓ hb v, b γ ′, b σ ′i ξ \|= Γ ⋉ ω γ ⋉ Φω b ξ \|= Σ ⋉ ψ σ ⋉ Φψ b DJaK (b γ,b σ , ω,ψ , Φ) = hw,ψ ′ , Φ′ i RJaK (ξ ) = ξ ′ then ξ ′ \|= τ ⋉ w \|= Σ ′ ⋉ ψ ′ ξ′ vb ⋉ Φ′ w b σ ′ ⋉ Φ′ ψ ′ Theorem B.5 (Well-partitioning). If Γ ; Σ ⊢ a : τ ; Γ ′ ; Σ ′ ;; hb γ,b σ i ⇓ hb v, b γ ′, b σ ′i Γ ; Σ ⊢ P (b γ, b σ) Φ DJaK (b γ,b σ , ω,ψ , Φ) = hw,ψ ′ , Φ′ i then τ ; Γ ′ ; Σ ′ ⊢ P (b v, b γ ′, b σ ′) Φ′ Γ ; Σ ⊢ a : τ ; Γ ′ ; Σ ′ ;; hb γ, b σ i ⇓ hb v, b γ ′, b σ ′i Γ ; Σ ⊢ S (b γ, b σ) Φ DJaK (b γ,b σ , ω,ψ , Φ) = hw,ψ ′ , Φ′ i then τ ; Γ ′ ; Σ ′ ⊢ S (b v, b γ ′, b σ ′) Φ′ Theorem B.6 (Stability). If 1:44 David Darais, Chang Liu, Ian Sweet, and Michael Hicks Theorem B.7 (Strong Memory Trace Obliviousness (SMTO)). Γ ⊢ e : τ ; Γ′ If Γ ; Σ ⊢ γbi Σ⊢b σi σi ⋉ b c (b obs γ1) ≈ env γbi ⋉ c (b obs γi ) ⋉ Φ env Φω Φψ c (b obs γ2) P (b γi , b σi ) env Φ c (b obs σi ) ⋉ Φ c (b obs σ1 ) ≈ S (b γi , b σi ) ξ \|= Σ ⋉ ψ then f (T JeK (b obs γ2, b σ2 )) trace c (b obs σ2 ) store store Φ f (T JeK (b obs γ1, b σ1 )) ≈ Φ trace ξ \|= Γ ⋉ ω store Corollary B.8 (Memory Trace Obliviousness (MTO)). If Γ ⊢ e : τ ; Γ′ Γ ; [] ⊢ ⌊γi ⌋ env env obs (γ 1 ) = obs (γ 2 ) f (T JeK (⌊γ 1 ⌋ , [])) ≈ obs f (T JeK (⌊γ 2 ⌋ , [])) then obs trace trace C LEMMAS REGARDING DISTRIBUTION MODEL Lemma C.1 (Cond Probability Measure Semantics). p[cond(b̃1, b̃2 , b̃3 ) = b] = p[b̃1 = O ∧ b̃2 = b] + p[b̃1 = I ∧ b̃3 = b] Proof. p[cond(b̃1 , b̃2, b̃3 ) = b] = p[cond(b̃1 , b̃2, b̃3 ) = b ∧ b̃1 = O] + p[cond(b̃1, b̃2 , b̃3 ) = b ∧ b̃1 = I] = p[cond(b̃1 , b̃2, b̃3 ) = b \| b̃1 = O]p[b̃1 = O] + p[cond(b̃1 , b̃2, b̃3 ) = b \| b̃1 = I]p[b̃1 = I] = p[b̃2 = b \| b̃1 = O]p[b̃1 = O] + p[b̃3 = b \| b̃1 = I]p[b̃1 = I] = p[b̃2 = b ∧ b̃1 = O] + p[b̃3 = b ∧ b̃1 = I] = p[b̃1 = O ∧ b̃2 = b] + p[b̃1 = I ∧ b̃3 = b] A Language for Probabilistically Oblivious Computation 1:45 Lemma C.2 (Cond Stability). Assume b̃1 ⊥ ⊥ b̃2 , b̃1 ⊥ ⊥ b̃3 and b̃2 ≈ b̃3 . Then: cond(b̃1, b̃2 , b̃3 ) ≈ b̃2 ≈ b̃3 Proof. Given b, the goal is to show p[cond(b̃1 , b̃2, b̃3 ) = b] = p[b̃2 = b] = p[b̃3 = b]. p[cond(b̃1, b̃2 , b̃3 ) = b] = * Cond Probability Measure Semantics (Lemma (C.1)) + p[b̃1 = O ∧ b̃2 = b] + p[b̃1 = I ∧ b̃3 = b] = * Conditional independence b̃1 ⊥ ⊥ b̃2 and b̃1 ⊥ ⊥ b̃3 + p[b̃1 = O]p[b̃2 = b] + p[b̃1 = I]p[b̃3 = b] = * Distribution equivalence b̃2 ≈ b̃2 + p[b̃1 = O]p[b̃2 = b] + p[b̃1 = I]p[b̃2 = b] = * Common factor + p[b̃2 = b](p[b̃1 = O] + p[b̃1 = I]) = * Proper Probability Measure (Lemma (C.5)) + p[b̃2 = b] = p[b̃3 = b] Lemma C.3 (Probabilistic Independence). Assume b̃1 ⋉ θ 1 , b̃2 ⋉ θ 2 and θ 1 ⊥ ⊥ θ 2 . Then: b̃1 ⊥ ⊥ b̃2 Lemma C.4 (Cond Independence). Assume b̃1 ⋉θ 1 , b̃2 ⋉θ 2 , b̃3 ⋉θ 3 , θ 1 ⊥ ⊥ θ2, θ1 ⊥ ⊥ θ 3 and b̃2 ≈ b̃3 . Then: cond(b̃1, b̃2, b̃3 ) ⋉ θ 2 ∪ θ 3 Our model for probability measures allows for proving the following lemmas. 1:46 David Darais, Chang Liu, Ian Sweet, and Michael Hicks Lemma C.5 (Proper Probability Measure). p[b̃1 = O] + p[b̃2 = I] = 1 Proof. p[b̃1 = O] + p[b̃2 = I] c[b̃1 = O] c[b̃1 = I] = + 2R 2R ′ {b \| b̃1 (b ′) = O} + {b ′ \| b̃1 (b ′) = I} = 2R ′ ′ {b \| b̃1 (b ) = O ∨ b̃1 (b ′) = I} = 2R {b ′ \| b ′ ∈ BR } = 2R 2R = R =1 2 Lemma C.6 (Joint Probability Symmetry). p[b̃1 = b1 ∧ b̃2 = b2 ] = p[b̃2 = b2 ∧ b̃1 = b1 ] Proof is immediate. Lemma C.7 (Chain Rule). p[b̃1 = b1 ∧ b̃2 = b2 ] = p[b̃1 = b1 \| b̃2 = b2 ]p[b̃2 = b2 ] Proof. p[b̃1 = b1 ∧ b̃2 = b2 ] = c[b̃1 = b1 ∧ b̃2 = b2 ] 2R = c[b̃1 = b1 ∧ b̃2 = b2 ] c[b̃2 = b2 ] 2R c[b̃2 = b2 ] = p[b̃1 = b1 \| b̃2 = b2 ]p[b̃2 = b2 ] Corollary C.8 (Alt Chain Rule). p[b̃1 = b1 ∧ b̃2 = b2 ] = p[b̃2 = b2 \| b̃1 = b1 ]p[b̃1 = b1 ] Consequence of Lemma (C.6). A Language for Probabilistically Oblivious Computation 1:47 Lemma C.9 (Bayes’s Rule). p[b̃1 = b1 \| b̃2 = b2 ] = p[b̃2 = b2 \| b̃1 = b1 ]p[b̃1 = b1 ] p[b̃2 = b2 ] Proof. p[b̃1 = b1 \| b̃2 = b2 ] = = = c[b̃1 = b1 ∧ b̃2 = b2 ] c[b̃2 = b2 ] c[b̃2 = b2 ∧ b̃1 = b1 ] c[b̃1 = b1 ] 2R c[b̃1 = b1 ] c[b̃2 = b2 ] 2R p[b̃2 = b2 \| b̃1 = b1 ]p[b̃1 = b1 ] p[b̃2 = b2 ] Corollary C.10 (Alt Bayes’s Rule). p[b̃1 = b1 \| b̃2 = b2 ] = Consequence of Lemma (C.7). p[b̃1 = b1 ∧ b̃2 = b2 ] p[b̃2 = b2 ] 1:48 David Darais, Chang Liu, Ian Sweet, and Michael Hicks Lemma C.11 (Total Probability). p[b̃1 = b1 ] = Õ p[b̃1 = b1 ∧ b̃2 = b2 ] b 2 ∈B Proof. p[b̃1 = b1 ] = c[b̃1 = b1 ] 2R {b ′ \| b̃1 (b ′) = b1 } = 2R {b ′ \| b̃1 (b ′) = b1 ∧ (b̃2 = O ∨ b̃2 = I)} = 2R {b ′ \| b̃1 (b ′) = b1 ∧ b̃2 = O ∨ b̃1 (b ′) = b1 ∧ b̃2 = I} = 2R {b ′ \| b̃1 (b ′) = b1 ∧ b̃2 (b ′) = O} + {b ′ \| b̃1 (b ′) = b1 ∧ b̃2 (b ′) = O} = = 2R c[b̃1 = b1 ∧ b̃2 = O] c[b̃1 = b1 ∧ b̃2 = I] + 2R 2R = p[b̃1 = b1 ∧ b̃2 = O] + p[b̃1 = b1 ∧ b̃2 = I] Í b 2 ∈B p[b̃1 = b1 ∧ b̃2 = b2 ]	6cs.PL
arXiv:0809.3187v1 [cs.CE] 18 Sep 2008 A Control Variate Approach for Improving Efficiency of Ensemble Monte Carlo ⋆ Tarik Borogovac a,b,∗ Francis J. Alexander b Pirooz Vakili c a Electrical b CCS-3, c Division and Computer Engineering Department, Boston University, 8 Saint Mary’s St. Boston, MA 02215, USA Los Alamos National Laboratory, MS-B256, Los Alamos, NM 87545, USA of Systems Engineering & Mechanical Engineering Department, Boston University, 15 Saint Mary’s St. Brookline, MA 02446, USA Abstract In this paper we present a new approach to control variates for improving computational efficiency of Ensemble Monte Carlo. We present the approach using simulation of paths of a time-dependent nonlinear stochastic equation. The core idea is to extract information at one or more nominal model parameters and use this information to gain estimation efficiency at neighboring parameters. This idea is the basis of a general strategy, called DataBase Monte Carlo (DBMC), for improving efficiency of Monte Carlo. In this paper we describe how this strategy can be implemented using the variance reduction technique of Control Variates (CV). We show that, once an initial setup cost for extracting information is incurred, this approach can lead to significant gains in computational efficiency. The initial setup cost is justified in projects that require a large number of estimations or in those that are to be performed under real-time constraints. Key words: Monte Carlo, Variance Reduction, Control Variates PACS: S05.10.Ln, 02.70.Uu, 02.70.Tt ⋆ Portions of this work, LA-UR-08-05399, were carried out at Los Alamos National Laboratory under the auspices of the US National Nuclear Security Administration of the US Department of Energy. Tarik Borogovac and Pirooz Vakili were supported in part by the National Science Foundation grants CMMI-0620965 and DGE-0221680. ∗ Corresponding author Email address: tarikb@bu.edu (Tarik Borogovac). Preprint submitted to Elsevier 13 January 2018 1 Introduction The purpose of this paper is to present a novel approach for efficient estimation via the Monte Carlo (MC) method. The approach is very broadly applicable but here, to present the main ideas, we narrow the focus to Ensemble Monte Carlo where estimation is based on stochastically independent trajectories of a system. To illustrate, we use simulation of time-dependent nonlinear processes for which Monte Carlo is a particularly general and powerful numerical method compared to available alternatives. Time-dependent nonlinear processes are very general models used, among others, in statistical mechanics [1], data assimilation in climate, weather and ocean modeling [2], financial modeling [3], and quantitative biology [4]. Hence developing efficient MC methods may significantly impact a wide range of applications. A known weakness of MC is its slow rate of convergence. Assume Y is a random quantity defined on paths of a process and let σY denote its standard deviation. The √ convergence rate of MC for estimating the expected value of Y is ≈ σY / n where n is the number of independent paths of the process. In general the canonical n−1/2 rate of convergence cannot be improved upon, hence, since the inception of the MC method, a number of variance reduction (VR) techniques have been devised to reduce σY (see, [5] for an early account and [3] and [6] for more recent discussions). Most VR techniques lead to estimators of the form w1 Y1 + · · · + wn Yn , i.e., a weighted average of the samples. These techniques prescribe (i) a recipe for selecting samples Y1 , · · · , Yn and (ii) a set of weights w1 , · · · , wn . To arrive at these prescriptions, one must rely on the existence of specific problem features and the ability of the user of the method to discover and effectively exploit such features. This lack of generality has significantly limited the applicability of VR techniques. The point of departure of a new strategy, called DataBase Monte Carlo (DBMC), is to address this shortcoming and to devise generic VR techniques that can be generically applied [7]. All VR techniques bring additional information to bear on the estimation problem, however, as mentioned above, this information is problem specific and relies on exploiting special features of the problem at hand. By contrast, as will be clarified in this paper, DBMC adds a generic computational exploration phase to the estimation problem that relies on gathering information at one (or more) nominal model parameter(s) to achieve estimation efficiency at neighboring parameters. The advantage of this approach is its generality and wide applicability: it is quite easy to im2 plement and it can wrap existing ensemble MC codes. On the other hand, the computational exploration phase of the DBMC approach may require extensive simulations and can be computationally costly. Therefore, the initial setup cost needs justification. The setup cost may be justified in projects that involve estimations at many model parameters and/or in projects where there is a real-time computational constraint. In the first type of project, the setup cost may lead to efficiency gain for each subsequent estimation, and for a large enough number of subsequent estimations it can be easily justified. In projects with a real-time constraint the setup cost is an off-line “passive” cost that can lead to estimates of significantly higher quality (lower statistical error); the higher quality in many such projects more than justifies the setup cost. In this paper we limit ourselves to presenting the implementation of the VR technique of Control Variates (CV) in the DBMC setting (see [7] for discussion of other VR techniques). The CV technique, which compared to the VR technique of Importance Sampling is less utilized in computational physics, requires identifying a number of random variables called control variates, say X1 , · · · , Xk , that are correlated with Y and have known means. The correlation with Y implies that Xi ’s carry information about Y . The CV technique is a way of utilizing the information included in the controls (their known means) to help with the estimation of the mean of variable Y . In the DBMC setting we assume that Y = Y (θ) depends on a model parameter θ and use Xi = Y (θi ) where θi ’s are in a neighborhood of θ (i = 1, · · · , k). In a departure from the classical CV technique, we use “high quality” estimates of E[Xi ] rather than precise values of E[Xi ] to arrive at the controlled estimator of E[Y ]. As we argue in this paper (and elsewhere [8]) this departure allows for substantially broader choices of control variates and makes the CV technique significantly more flexible and effective. The DBMC method shares a similar intent as the well-known histogram reweighing method [9] from the Markov chain Monte Carlo literature (e.g. [1]), but with a very different setting and implementation, and with broader applicability. For example, it does not rely on having a Boltzmann distribution or exp(−H/kT ) structure. Given its generality, it has potential applications, among others, in ensemble weather prediction, hydrological source location, climate and ocean, optimal control, and stochastic simulations of biological systems. The remainder of the paper is organized as follows. In section 2 we discuss preliminaries, including the details of the example numerical study – the timedependent Ginzburg-Landau (TDGL) equation – as well as the method of control variates. Estimation of mean outcomes of the TDGL equation over a range of temperatures is of interest, especially considering the large difference in behavior below and above the coexistence curve. In section 3 we describe the DBMC methodology and motivation in a general context. Section 4 discusses 3 the implementation and results of DBMC as applied to estimation of quantities generated by the TDGL equation, and the results of that numerical study. We conclude in section 5. 2 Preliminaries We present aspects of our approach and numerical results in the context of the time-dependent Ginzburg Landau (TDGL) model. It is worth noting that this model is chosen for illustrative purposes only and we do not make use of any of its specific features. 2.1 Time-Dependent Ginzburg Landau We use a canonical equation of phase-ordering kinetics [10,11] the stochastic TDGL equation in two spatial dimensions. This is written as ∂φ(x, t) = D∆φ(x, t) − V ′ (φ(x, t)) + η(x, t) ∂t (1) where φ(x, t) represents a local order parameter, e.g. a magnetization at point x = (x1 , x2 )⊤ and time t (⊤ denotes transpose). The noise has mean zero and covariance hη(x, t)η(x′, t′ )i = 2δ(x − x′ )δ(t − t′ ). We choose a double-well potential V (φ) = − 2θ φ2 + χ4 φ4 . As in [11] χ is a constant, and θ is a function of temperature such that a high θ corresponds to a low temperature. We use a discrete form of (1) using a forward Euler-Maruyama stochastic integrator and a 5-point stencil for the Laplacian (denoted ∆L ) for simulation: φ(x, t + δt) = φ(x, t) + Dδt∆L φ(x, t) − δt[−θφ(x, t) + χφ3 (x, t)] q + 2(δt/δx)N(x, t) with time step δt and lattice spacing δx, and where N(x, t) are independent and identically distributed standard normal random variables for each spacetime point (x, t). What follows applies to other discretization schemes as well. 2.2 Estimation problem To cover a broad range of estimation problems, we consider the estimation of quantities related to a specific space-time point, quantities that are global (en4 tire lattice at a particular time) and quantities that depend on the entire time evolution of the system. Specifically, we consider the following representative quantities: (P1) Point magnetization: φ(x, t), P (P2) Total magnetization at a specific time t: x φ(x, t), and P P (P3) Total space-time magnetization: t x φ(x, t). The problem of estimating the expected value of any one of the above quantities can be represented by: J(θ) = E[Y (ω; θ))] where ω is a vector of random numbers representing all the noise/uncertainty in a single complete path φ of the dynamics; θ is the temperature related parameter; Y (ω; θ) is the random sample of a quantity of interest (e.g., the magnetization from a single sample path), and E[·] denotes expectation. Note that knowing the noise ω and parameter θ completely determines the path φ and the sample quantity of interest Y (ω; θ). 2.3 The control variate technique Here we give a brief review of the classical control variate (CV) technique for variance reduction (see [3] [12]). Let Y = Y (ω; θ), J = E[Y ]. Assume X1 , · · · , Xk are random variables (called control variates) that are correlated with Y and assume their means E[Xi ] are known. Let X = (X1 , · · · , Xk )⊤ , E[X] = (E[X1 ], · · · , E[Xk ])⊤ , and β = (β1 , · · · , βk )⊤ . Then Z, defined below, is a controlled estimator of E[Y ] Z=Y + k X i=1 βi (Xi − E[Xi ]) = Y + β ⊤ (X − E[X]) The estimator Z uses information included in samples of the controls (the degree of their deviation from their known means) to “correct/adjust” the estimator Y and bring it closer to its unknown mean. This is the key idea of CV. (Alternatively, Z can be viewed as the fitted value of Y when Y is linearly regressed on variables X1 , · · · , Xk . In other words, Z includes the part of the variation in Y that cannot be “explained” by Xi ’s.) Z is an unbiased estimator of E[Y ] for all vectors β. The coefficient vector 5 that minimizes the variance of Z is: β o = Σ−1 X ΣXY where ΣX is the k × k covariance matrix of X and ΣXY is the k × 1 vector of covariances of Y and Xi ’s. When β o is used, the variance of Z is given by −1 2 (1 − R2 )σY2 where R2 = Σ⊤ XY ΣX ΣXY /σY . Therefore, V ar(Y ) = (1 − R2 )−1 V ar(Z) (2) and hence (1 − R2 )−1 is precisely the theoretical degree of variance reduction if the controlled estimator Z = Y + β o⊤ (X − E[X]) is used to estimate J as opposed to the crude MC estimator Y , and it is called the Variance Reduction Ratio (VRR) statistic for control variates. Note that there is no upper limit to the degree of achievable variance reduction since R2 can potentially be very close to 1 when the controls are highly correlated with the estimation variable Y . In other words, the CV technique can potentially be very effective leading to orders of magnitude of variance reduction. In practice and in general, ΣX and ΣXY (i.e. β o ) are not known exactly and need to be estimated from samples of Xi ’s and Y . Typically, β o is estimated from the same samples used to construct the controlled estimator Z. While this practice adds some bias for small sample sizes, and thus makes the effective decrease in estimator mean squared error not precisely equal to the variance reduction ratio (1 − R2 )−1 , this bias converges to zero faster than the standard error of Z. Thus, expending computational resources into generating separate pilot samples for estimating β o is not considered to be justifiable. For an insightful and detailed discussion of the CV technique, see [3]. 2.4 Challenges in using the CV technique The critical task for using the CV technique is in finding effective controls. Once the controls are selected, the rest of the procedure is fairly routine. An effective control, say X, needs to satisfy two requirements (to simplify the discussion we consider a scalar control): (R1) X needs to be correlated with Y , and (R2) E[X] needs to be available to the user, i.e., known. The main barrier to finding effective controls is the second requirement, namely the requirement of a known mean E[X]. A modification of the CV technique called Biased Control Variate (BCV) reduces the burden of requirement (R2) 6 by allowing for a good approximation of E[X] when E[X] cannot be evaluated analytically [13]. While BCV lowers the requirement barrier and expands the range of available choices for effective controls, it nonetheless limits its potential scope by implicitly assuming an analytic path to arriving at the approximate value. As we describe in the next section, in the DBMC approach we turn the second requirement into a computational task; in other words, we use statistical estimation to obtain a good estimate of E[X]. Therefore, barrier (R2) is completely removed and the range of choices of controls is dramatically expanded. The relevant question now becomes whether the computational investment in estimating E[X] pays enough dividends to make the investment worthwhile. 3 DBMC & Control Variate The starting point of the DBMC approach is the observation that in many parametric estimation settings, including in the example considered in this paper, quantities Y (θ) and Y (θ′ ) are highly correlated when the same random input ω is used to generate them and when θ and θ′ are close 1 . This suggests using control variates Xi = Y (θi ), i = 1, · · · , k, when estimating Y (θ) where θi ’s are “close” to θ. While we have identified potentially effective controls, we do not have sufficient information about them, i.e., J(θi ) = E[Xi ] is not known and needs to be evaluated. This brings us to the second feature of the DBMC method that corresponds to its initial computational information gathering/setup stage. This stage corresponds to statistical estimation of J(θi ). Details are given below. 3.1 DBMC + CV algorithm The DBMC approach consists of a setup stage and an estimation stage. 3.1.1 Setup stage The DBMC setup phase involves generating a “large” number of input random vectors ω and obtaining “high quality” estimates of J(θi ). Let DB = 1 A similar observation is the basis for the histogram re-weighting methods: “from a simulation at a single state point (characterized in an Ising model by choice of temperature T and magnetic field H) one does not gain information on properties at that point only, but also in the neighboring region,” ([1], page 116) 7 {ω1 , ω2 , · · · , ωN } (N “large”) denote a large set of random inputs. This set represents the database. Given the database, the averages of the controls are precisely calculated. A schematic of this stage is given in Figure 1. (1) For j = 1, · · · , N (a) Generate ωj according to the distribution of the inputs; (b) For i = 1, · · · , k (i) Simulate the path φ(ωj ; θi ) (ii) Evaluate the value of the control Xi (ωj ) = Y (ωj ; θi ) (2) For i = 1, · · · , k (a) Find JDB (θi ), the average of the ith control on the datebase, as JDB (θi ) = N 1 X Y (ωj ; θi ) N j=1 Fig. 1. DBMC setup stage 3.1.2 Estimation stage To estimate J(θ), at a θ close to θi ’s (i = 1, · · · , k) select a “small” sample (say of size n ≪ N) uniformly from the database. For each sample ωj resimulate the equation using ωj and θ to obtain Y (ωj ; θ). For these samples the values of the controls Xi (ωj ) are available in the database. Using these evaluate a controlled estimate of J(θ). A schematic version of these steps is given in Figure 2. (1) For (a) (b) (c) j = 1, · · · , n Select ωj uniformly from the database; Simulate the path φ(ωj ; θ); Evaluate the estimation variable Y (ωj ; θ). (2) Find the controlled estimator of J(θ): Jbcv (θ) = n k X 1X [Y (ωj ; θ) + βio (Xi (ωj ) − JDB (θi ))] n j=1 i=1 Fig. 2. DBMC estimation stage 8 (3) 3.2 Implementation choices There are two general schemes for implementation of our CV approach: (I1) corresponding to what is described above, requires storing simulation inputs {ωj } and outputs {Xi (ωj )} in a database for later resampling; (I2) does not utilize resampling, so there is no storage of data beyond recording the calculated control means. Both implementations are feasible, the first is preferable in most cases; the second may be preferred in some cases. We elaborate below. Implementation (I1). • The database of random inputs, i.e., ωj ’s, are either directly stored or enough information about them (e.g. input seeds of a pseudo-random number generator) is stored to be able to regenerate ωj ’s precisely. • The k paths corresponding to θi , i = 1, · · · , k, i.e., φ(ωj ; θi ) are generally simulated “in parallel” as elements of a random vector, ωi , are progressively generated. • For each random input, say ωj , the value of the controls, Xi (ωj ), j = 1, . . . , N, i = 1, . . . , k are stored. Implementation (I2). • Once the setup stage is completed, the only values stored are the “high quality” estimates of the means of J(θi )’s, i.e., the k values JDB (θi ), i = 1, · · · , k. • At the estimation phase, n random input vectors ωj , j = 1, . . . , n, are generated anew; paths at θ and θi , i = 1, · · · , k, are simulated using new random inputs and for each path Xi (ωj ) and Y (ωj ; θ) are calculated; finally, using these values, the controlled estimator is evaluated. 3.3 Statistical properties & computational efficiency The promise of the approach is the following: by anchoring estimation via CV at a few high quality estimates (at θ1 , · · · , θk ), it is possible to obtain high quality estimates at other locations in the parameter space (at other θ) with far fewer samples. The actual statistical properties of the resulting estimators, and the computational efficiency of generating them, reflect choices made in implementing each given problem. For example, how much computation should be “invested” in the exploration phase, and which points θi in the parameter space should be explored are two important questions that need further investigation. Such choices generally involve problem dependent tradeoffs, and we leave them to future studies. 9 Instead, the analysis that follows is meant to provide a general and qualitative understanding of the statistical properties, computational efficiency and the tradeoffs involved. The discussion is as general as possible, but consistent with the numerical study described in section 4, where such implementation choices were made utilizing only a basic familiarity with the problem. For further discussion, see [8]. 3.3.1 Statistical properties We give the analysis for implementation (I1). In other words, assume we are re-sampling from the database. Analysis of implementation (I2) shows similar estimator statistical properties. To simplify the discussion consider a single control, say X1 . Let J(θ) = E[Y ], 2 J(θ1 ) = E[X1 ], σY2 = Var(Y ), σX = Var(X1 ). Assume a database of input 1 ∗ variables are generated and let Y and X1∗ be random variables corresponding to Y and X1 that are generated by re-sampling (uniformly, with replacement) 2 from the database. Let J ∗ (θ), σY2 ∗ , J ∗ (θ1 ), σX ∗ denote the means and variances 1 ∗ ∗ of the re-sampled variables Y and X . Conditioned on the database, the controlled estimator is exactly the classical CV estimator and all results from classical CV apply. For example, for any scalar β, Z ∗ = Y ∗ + β(X ∗ − J ∗ (θ1 )) is an unbiased estimator of J ∗ (θ), J ∗ (θ1 ) is known, and the optimal β o is what is prescribed by classical CV if we take all random variables as those defined on the database. A measure of variance reduction due to using a controlled estimator is V RR = σY2 ∗ σZ2 ∗ (4) We use the controlled estimator Z ∗ as an estimator for J(θ). Assume optimal β o is used to define Z ∗ and assume E[Z ∗ ] = J ∗ (θ) 2 . In general J ∗ (θ) 6= J(θ). Therefore, Z ∗ is a biased estimator of J(θ) where the bias is introduced by sampling from the database, i.e., from Y ∗ , as opposed to from Y . We have some probabilistic assessment of this bias and we can reduce it by increasing the size of the database. Specifically, for this bias we can obtain an approximate 1 − α probability confidence interval: zα/2 σY P (\|J ∗ (θ) − J(θ)\| ≤ √ )≈1−α N 2 i.e. we ignore the low order bias that results from the typical CV procedure of estimating the optimal β o e.g. [3], not to be confused with the resampling bias discussed in this section 10 where zα/2 is the 1 − α/2 quantile from the standard normal distribution. In other words, with high probability the bias is of the order of O(N −1/2 ). We assume that for large N the bias is sufficiently small to be disregarded and that we can focus on V RR in (4) as the key measure of computational gain in using the controlled estimator to estimate J(θ). 3.3.2 Computational efficiency Generating the above large database, as we pointed out earlier, corresponds to an initial “setup” cost. Let C be the computational cost of generating a sample of Y (θ). This cost involves generating an ω, simulating the path, and evaluating Y (ω, θ). A reasonable assumption for many problems is that this cost is about the same for all θ and ω. Then the set-up cost of generating the database and obtaining averages of the controls is approximately N ×k×C. Let V RR(θ) denote the variance reduction ratio at θ, i.e., the ratio of the variance of an uncontrolled sample and that of a controlled sample at θ. Then, the statistical error of a controlled estimator based on n samples is approximately the same as that of n × V RR(θ) samples of an uncontrolled estimator. Thus, the ratios of the computational costs of the two estimators (to arrive at the same statistical accuracy) is (n×V RR(θ)×C)/(n×C) = V RR(θ). Therefore, V RR(·) can serve as a measure of benefit of the DBMC approach. The setup cost of the DBMC approach can be justified in two types of applications. The first type are those applications that require solving many instances of the estimation problem, at many θ’s. If the total number of instances is sufficiently large, and some variance reduction is achieved on the average on those instances, then the large fixed set-up cost can be dwarfed by the total computational savings from the many estimations. The second type are real-time applications where the setup cost can be viewed as an off-line cost enabling significant efficiency gains in the critical task of real-time estimation. Typically, the “cost” of delay in such real-time estimation is higher and not merely computational, justifying even a much larger computational effort off-line. 4 Numerical results The numerical results in this section are intended to give a qualitative illustration of the efficiency gains that can be achieved using the DBMC approach. Specifically, we estimate the variance reduction that can be achieved over regular (crude) sampling, when estimating the three quantities of interest (a point magnetization, total magnetization at a specific time t and the total time-space magnetization) at a range of the parameter θ. Our choices of the 11 size of the database, number of samples used for estimation, range of parameter values, and the controls are simply for illustration purposes. However, we expect that the numerical results are, qualitatively, quite representative. We simulate the TDGL dynamics on a 40 × 40 lattice (lattice spacing δxk = 1, k = 1, 2) with fixed χ = 1. On each path, we evolve the system for a total of 5000 time steps (δt = 0.01) which is sufficient for the system to exhibit behavior that is specific to its temperature region. The critical point for this system is θc = 1.265 [11], and our parameter range of interest (1.0 to 1.5) extends to both sides of that critical point. To build a database, we simulate N = 214 = 16384 paths and evaluate point magnetization, total magnetization at a specific time, and total space-time magnetization at two nominal values of θ, 1.2 and 1.35. For each quantity of interest, we consider three control variate estimators. The first two estimators, CV1.2 and CV1.35, use single controls corresponding to θ = 1.2 and θ = 1.35, respectively. We chose to anchor our estimators at those two nominal values for θ because they are located on opposite sides of the phase transition line θc . The third estimator, CV2C, uses both controls simultaneously. We use n = 28 = 256 samples for crude and CV estimators. To estimate the variance of these estimators, following the micro-macro simulation approach (see, e.g., [14]), we use 40 independent macro simulations consisting of 256 independent micro simulations. We obtain variance estimates from each macro simulation and average the resulting 40 values to obtain an overall variance estimate. We report the ratios of the variance estimates (crude/controlled, as in (4)) as V RR. A sampling of VRR results for the total space-time magnetization (problem P3) is given in Table 1 and the corresponding graph is given in Figure 3. The graph for point magnetization (problem P1) are given in Fig. 4, and the results for the total magnetization at a time t (problem P2) are quite similar and are excluded. Table 1 Variance reduction ratios of the estimators applied to space-time integral of the P P magnetization, x t φ(x, t), at several values of θ. Estimator\θ 1.150 1.175 1.225 1.250 1.265 1.300 1.325 1.375 1.400 CV1.2 63 236 219 55 33 15 10 6 5 CV1.35 5 6 11 16 21 59 231 245 67 CV2C 170 709 947 332 259 300 761 513 129 Based on these results, we draw the following conclusions: • Controlled estimators produce dramatic variance reduction for parameter 12 Variance Reduction Ratios for the time integral of the magnetization CV12 CV135 CV2C 2 VRR 10 1 10 1 1.05 1.1 1.15 1.2 1.25 θ 1.3 1.35 1.4 1.45 1.5 Fig. 3. VarianceP reduction ratios of the estimators of the space-time integral of the P magnetization, x t φ(x, t), over a range of values of θ (log scale). values very close to the nominal parameters and substantial variance reduction at values moderately close to the nominal. • For all the estimation problems, adding the second control consistently improves performance, in some cases leading to substantial reduction in variance (compared to single controls). Of course, by incorporating information from points on both sides of the critical temperature, CV2C is expected to give better coverage than either of the single control estimators. However, CV2C does better than either of the single control estimators even in their own regions, which suggests that each control provides relevant information to the estimation problem in the opposite region. • VRR values for the total space-time magnetization are somewhat larger than those for the point and total magnetization at a specific time t – we expect this to be true more generally for path integrals when compared with values at specific time instances. 13 Variance Reduction Ratios for the point magnetization CV12 CV135 CV2C 2 VRR 10 1 10 1 1.05 1.1 1.15 1.2 1.25 θ 1.3 1.35 1.4 1.45 1.5 Fig. 4. Variance reduction ratios of the estimators of the point magnetization over a range of values of θ, in log scale. 5 Conclusions In this paper we described a new strategy, DataBase Monte Carlo (DBMC), for improving computational efficiency of Ensemble Monte Carlo. For a specific time-dependent nonlinear dynamics we showed that the approach can lead to significant efficiency gains for a range of estimation problems. Our selection of the controls has been ad-hoc and for illustration purposes. Further work is required to better understand the options available and the computational tradeoffs involved. To this end, our current research is focused on (i) derivation of more specific guidelines for the selection of effective control variates, (ii) implementation of the DBMC strategy in conjunction with other variance reduction techniques, for example, stratification and importance sampling, and (iii) application of the method in some specific domains, for example, estimation problems in geophysical fluids and biochemical systems. 14 References [1] K. K. Binder, D. Heermann, Monte Carlo Simulation in Statistical Physics: An Introduction, Springer, 2002. [2] G. Evensen, Data Assimilation, The Ensemble Kalman Filter, Springer, 2006. [3] P. Glasserman, Monte Carlo Methods in Financial Engineering, Springer-Verlag New York, Inc, 2004. [4] D. J. Wilkinson, Stochastic Modelling for Systems Biology, CRC Press, 2006. [5] J. M. Hammersley, D. C. Handscomb, Monte Carlo Methods, John Wiley, 1964. [6] S. Asmussen, P. W. Glynn, Stochastic Simulation: Algorithms and Analysis, Springer, 2007. [7] P. Vakili, G. Zhao, T. Borogovac, Database monte carlo: A new strategy for efficient simulation, Tech. rep., Boston University College of Engineering (2008). [8] T. Borogovac, P. Vakili, Database monte carlo approach to effective control variates, Tech. rep., Boston University College of Engineering (2008). [9] A. M. Ferrenberg, R. H. Swendsen, New monte carlo technique for studying phase transitions, Phys. Rev. Lett. 61 (23) (1988) 2635–2638. [10] N. Gulbahce, F. J. Alexander, G. Johnson, Statistical mechanics of histories: A cluster monte carlo algorithm, Phys. Rev. E 73 (2006) 026701. [11] A. L. Ferreira, R. Toral, Hybrid monte carlo method for conserved-orderparameter systems, Phys. Rev. E 47 (6) (1993) R3848–R3851. [12] C. P. Robert, G. Casella, Monte Carlo Statistical Methods, Springer Science & Business Media Inc., 2004. [13] B. Schmeiser, M. R. Taaffe, J. Wang, Biased control-variate estimation, IIE Transactions 33 (2001) 219–228. [14] B. Schmeiser, Chapter 7: Simulation experiments, in: D. P. Heyman, M. J. Sobel (Eds.), Handbooks in OR and MS 2: Stochastic Models, Elsevier B. V., 1990, Ch. 7, pp. 295–330. 15	5cs.CE
Threshold Selection for Multivariate Heavy-Tailed Data Phyllis Wan and Richard A. Davis arXiv:1707.00464v2 [math.ST] 26 Mar 2018 Department of Statistics Columbia University 1255 Amsterdam Avenue, MC 4690 New York, NY 10027 e-mail: phyllis@stat.columbia.edu; rdavis@stat.columbia.edu Abstract: Regular variation is often used as the starting point for modeling multivariate heavy-tailed data. A random vector is regularly varying if and only if its radial part R is regularly varying and is asymptotically independent of the angular part Θ as R goes to infinity. The conditional limiting distribution of Θ given R is large characterizes the tail dependence of the random vector and hence its estimation is the primary goal of applications. A typical strategy is to look at the angular components of the data for which the radial parts exceed some threshold. While a large class of methods has been proposed to model the angular distribution from these exceedances, the choice of threshold has been scarcely discussed in the literature. In this paper, we describe a procedure for choosing the threshold by formally testing the independence of R and Θ using a measure of dependence called distance covariance. We generalize the limit theorem for distance covariance to our unique setting and propose an algorithm which selects the threshold for R. This algorithm incorporates a subsampling scheme that is also applicable to weakly dependent data. Moreover, it avoids the heavy computation in the calculation of the distance covariance, a typical limitation for this measure. The performance of our method is illustrated on both simulated and real data. Keywords and phrases: distance covariance, heavy-tailed data, multivariate regular variation, threshold selection. 1. Introduction For multivariate heavy-tailed data, the principal objective is often to study dependence in the ‘tail’ of the distribution. To achieve this goal, the assumption of multivariate regular variation is typically used as a starting point. A random vector X ∈ Rd is said to be multivariate regularly varying if the polar coordinates (R, Θ) = (kXk, X/kXk), where k · k is some norm, satisfy the conditions (a) R is univariate regularly varying, i.e., P(R > r) = L(r)r−α , where L(·) is a slowly varying function at infinity; (b) P(Θ ∈ ·\|R > r) converges weakly to a measure S(·) as r → ∞. The α is referred to as the index of the regular variation, while the S is called the angular distribution and characterizes the limiting tail dependence. There are other equivalent definitions of regular variation (Resnick, 2002), but this one is the most convenient for our purposes. Given observations {Xi }ni=1 and their corresponding polar coordinates {(Ri , Θi )}ni=1 , a straightforward procedure for estimating S is to look at angular components of the data for which the radii are greater than a large threshold r0 , that is, Θi for which Ri > r0 . In most studies, one takes r0 to be a large empirical quantile of R. While there has been extensive research on choosing a threshold for which the distribution of R is regularly varying (i.e., limit condition (a)), little research has been devoted to ensuring the threshold is large enough for the independence of Θ and R to be reasonable (i.e., limit condition (b)). To this end, de Haan and de Ronde (1998) fit a parametric extreme value distribution model to each marginal and examined the parameter stability plot of each coordinate. The Stǎricǎ plot (Stǎricǎ, 1999) looked at the joint tail empirical measure, but was, in some way, equivalent to only examining the extremal behavior of R. Resnick (2007) suggested an automatic threshold selection from the Stǎricǎ plot but observed that the thresholds were sometimes systematically underestimated. In their study of the threshold based inference for parametric max-stable processes, Jeon and Smith (2014) suggested choosing the threshold by minimizing the MSE of the estimated parameters. In this paper, we propose an algorithm which selects the threshold for modeling S. Our motivation is the implied property that R and Θ given R > r become independent as r → ∞. Given a sequence of candidate 1 imsart-generic ver. 2014/10/16 file: thres_arxiv.tex date: March 28, 2018 2 Wan and Davis threshold levels, we test the degree of dependence between R and Θ for the truncated data above each level. The dependence measure we use is the distance covariance introduced by Székely et al. (2007). This measure has the ability to account for various types of dependence and to be applicable to data in higher dimensions. The resulting test statistics are given in the form of p-values and are compared across all levels through a subsampling scheme. This enables us to extract more accurate information from the test statistics while not overloading the computational burden. The remainder of this paper is organized as follows. We first provide some theoretical backgrounds on multivariate regular variation in Section 2. The distance covariance and its theoretical properties are introduced in Section 3. Applying this dependence measure in our conditioning setting, we propose a test statistic and prove relevant theoretical results in Section 4. Our proposed algorithm for threshold selection is presented in Section 5, and illustrated on simulated and real examples in Section 6. The paper concludes with a discussion. 2. Multivariate regular variation and problem set-up One way to approach multivariate heavy-tailed data is through the notion of multivariate regular variation. For a detailed review, see, for example, Chapter 6 of Resnick (2007). Let X = (X1 , . . . , Xd ) be a d-dimensional random variable defined on the cone Rd+ = [0, ∞)\{0}. Define the polar coordinate transformation T (X) = (kXk, X/kXk) =: (R, Θ), (2.1) where k · k denotes some norm. Then X is regularly varying if and only if there exists a probability measure S(·) on Sd−1 , the unit sphere in Rd , and a function b(t) → ∞, such that v tP [(R/b(t), Θ) ∈ ·] → να × S, t → ∞, (2.2) v where → denotes vague convergence, and να is a measure defined on (0, ∞] such that να (x, ∞] = x−α , x > 0. Here b(t) can be chosen as the 1 − t−1 -quantile, i.e., b(t) = inf{s\|P(R ≤ s) ≥ 1 − t−1 }. The convergence (2.2) implies that R w , Θ ∈ · R > r → να × S, P r r → ∞, on [1, ∞) × Sd−1 , (2.3) w where → denotes weak convergence. In other words, given that R > r for r large, the conditional distribution of R/r and Θ are independent in the limit. In view of (2.3), we restrict the measure να to [1, ∞) throughout the remainder of the paper. The angular measure S characterizes the tail dependence structure of X. If S is concentrated on {ei , i = 1, . . . d}, where ei = (0, . . . , 0, 1, 0, . . . , 0), then the components of X are asymptotically independent in the tail, a case known as asymptotic independence. If S has mass lying in the subspace {(t1 , . . . , td ) ∈ Sd−1 \|ti > 0, tj > 0, i 6= j}, then an extreme observation in the Xi direction implicates a positive probability of an extreme observation in the Xj direction, a case known as asymptotic dependence. Hence the estimation of S from observations is an important problem, and often the primary goal, in multivariate heavy-tailed modeling. The following convergence is implied from (2.3): w P(Θ ∈ ·\|R > r) → S(·), r → ∞. (2.4) This suggests estimating S using the angular data (Θi ) whose radial parts satisfy Ri > r0 for r0 large. The motivation behind our method is to seek r0 such that when R > r0 , R and Θ are virtually independent. Given a candidate threshold sequence {rk }, we formally test the independence between (Ri , Θi ) on the index set {i\|Ri > rk }. The use of Pearson’s correlation as the dependence measure is unsuitable in this case, for two reasons. First, correlation is only applicable to univariate random variables, whereas Θ lies on the sphere of dimension d − 1. Second, correlation only describes the linear relationship between two random variables, thus having zero correlation is not a sufficient condition for independence. Instead, we use a more powerful dependence measure, the distance covariance, which is introduced in the next section. Threshold Selection for Multivariate Heavy-Tailed Data 3 3. Distance covariance In this section, we briefly review the definition and some properties of the distance covariance. More detailed descriptions and proofs can be found in Székely et al. (2007) and Davis et al. (2017). Let X ∈ Rp and Y ∈ Rq be two random vectors, then the distance covariance between X and Y is defined as Z 2 T (X, Y ; µ) = ϕX,Y (s, t) − ϕX (s) ϕY (t) µ(ds, dt) , (s, t) ∈ Rp+q (3.1) Rp+q where ϕX,Y (s, t), ϕX (s), ϕY (t) denote the joint and marginal characteristic functions of (X, Y ) and µ is a suitable measure on Rp+q . In order to ensure that T (X, Y ; µ) is well-defined, one of the following conditions is assumed to be satisfied throughout the paper (Davis et al., 2017): 1. µ is a finite measure on Rp+q ; 2. µ is an infinite measure on Rp+q such that Z (1 ∧ \|s\|α )(1 ∧ \|t\|α )µ(ds, dt) < ∞ Rp+q and E[\|XY \|α + \|X\|α + \|Y \|α ] < ∞ for some α ∈ (0, 2]. One advantage of distance covariance over, say, Pearson’s covariance, is that, if µ has a positive Lebesgue density on Rp+q , then X and Y are independent if and only if T (X, Y ; µ) = 0. Another attractive property of this dependence measure is that it readily applies to random vectors of different dimensions. To estimate T (X, Y ; µ) from observations (X1 , Y1 ), . . . , (Xn , Yn ), define the empirical distance covariance Z 2 ϕ̂X,Y (s, t) − ϕ̂X (s) ϕ̂Y (t) µ(ds, dt) , Tn (X, Y ; µ) = Rp+q 1 Pn e i hs,Xj i+i ht,Yj i and ϕ̂X (s) = ϕ̂X,Y (s, 0), ϕ̂Y (t) = ϕ̂X,Y (0, t) are the respective where ϕ̂X,Y (s, t) = n j=1 empirical characteristic functions. If we assume that µ = µ1 × µ2 and is symmetric about the origin, then under the conditions where T (X, Y ; µ) exists, Tn (X, Y ; µ) also has the computable form Tn (X, Y ; µ) = n 1 X µ̃1 (Xi − Xj ) µ̃2 (Yi − Yj ) n2 i,j=1 + 1 n4 n X µ̃1 (Xi − Xj )µ̃2 (Yk − Yl ) − i,j,k,l=1 2 n3 n X µ̃1 (Xi − Xj )µ̃2 (Yi − Yk ), i,j,k=1 R where µ̃(x) = (1 − coshs, xi) µ(ds) (Davis et al., 2017). The most popular choice of µ, first mentioned by Feuerverger (1993) and then more extensively studied by Székely et al. (2007), is µ(ds, dt) = cp,q \|s\|−κ−p \|t\|−κ−q ds dt . (3.2) where cp,q is as defined in Lemma 1 of Székely et al. (2007). This choice of µ gives µ̃(x)µ̃(y) = \|x\|κ \|y\|κ . Moreover, this is the only choice of µ for which the distance covariance is invariant relative to scale and orthogonal transformations. Note that in order for the integral (3.1) to exist, it is required that E[\|X\|κ \|Y \|κ + \|X\|κ + \|Y \|κ ] < ∞. We will utilize the described weight measure (3.2) with κ = 1 in our simulations and data analyses in Section 6, but applied to the log transformation on R to ensure that the moment condition is satisfied. As detailed in Davis et al. (2017), if the sequence {(Xi , Yi )} is stationary and ergodic, then a.s. Tn (X, Y ; µ) → T (X, Y ; µ). (3.3) 4 Wan and Davis Further, if X and Y are independent, then under an α-mixing condition, Z d n Tn (X, Y ; µ) → \|G(s, t)\|2 µ(s, t) (3.4) Rp+q for some centered Gaussian field G. On the other hand, if X and Y are dependent, then √ d n(Tn (X, Y ; µ) − T (X, Y ; µ)) → G0µ for some non-trivial limit G0µ , implying that n Tn (X, Y ; µ) diverges as n → ∞. Naturally one can devise a test of independence between X and Y using the statistic n Tn (X, Y ; µ): the null R hypothesis of independence is rejected at level χ if n Tn (X, YR; µ) > c, where c is the upper χ-quantile of Rp+q \|G(s, t)\|2 µ(s, t). In practice, the distribution Rp+q \|G(s, t)\|2 µ(s, t) is intractable and is typically approximated through bootstrap. Hence the main drawback of using distance covariance is the computation burden it brings for large sample size: the computation of a single distance covariance statistic requires O(n2 ) operations, while finding the cut-off values via resampling requires much more additional computation. Our method, however, overcomes this problem through subsampling the data, as will be described in Section 5. 4. Theoretical results Let {Xi }ni=1 be iid observations in Rd from a multivariate regularly varying distribution X satisfying (2.1) and (2.3), and {(Ri , Θi )}ni=1 be their polar coordinate transformations. Given a threshold rn , we measure the dependence between R/rn and Θ conditional on R > rn by the empirical distance covariance of (R/rn , Θ) on the set {R > rn }: Z \|Cn (s, t)\|2 µ(ds, dt), Tn := (4.1) Rd+1 with Cn (s, t) := ϕ̂ R ,Θ\|rn (s, t) − ϕ̂ R \|rn (s)ϕ̂Θ\|rn (t), rn rn where ϕ̂ R ,Θ\|rn is the conditional empirical characteristic function of (R/rn , Θ), rn n X T 1 eisRj /rn +it Θj 1{Rj >rn } , s ∈ R, t = (t1 , . . . , td )T ∈ Rd , j=1 1{Rj >rn } j=1 ϕ̂ R ,Θ\|rn (s, t) = Pn rn and ϕ̂ R \|rn , ϕ̂Θ\|rn are the corresponding empirical conditional marginal characteristic functions, rn ϕ̂ R \|rn (s) = ϕ̂ R ,Θ\|rn (s, 0), rn rn ϕ̂Θ\|rn (t) = ϕ̂ R ,Θ\|rn (0, t). rn In this section, we establish the limiting results (3.3) and (3.4) adapted to the conditional distance covariance. For ease of notation, let n pn := P(R > rn ) , p̂n := 1X 1{Rj >rn } n j=1 be the theoretical and empirical probability of exceedance, and let i h isR/rn +itT Θ h i E e 1 R>r n T ϕ R ,Θ\|rn (s, t) := E eisR/rn +it Θ \|R > rn = , rn pn and ϕ R \|rn (s) := ϕ R ,Θ\|rn (s, 0), rn rn ϕΘ\|rn (t) := ϕ R ,Θ\|rn (0, t), rn be the theoretical conditional joint and marginal characteristic functions. Threshold Selection for Multivariate Heavy-Tailed Data 5 Recall from (2.3) that as n → ∞, R/rn and Θ become asymptotically independent and converge to να and S respectively. Denote the characteristic functions of the corresponding limit distributions by Z ∞ ϕR (s) := exp(isr)αr−α−1 dr = lim ϕ R \|rn (s), (4.2) n→∞ rn 1 Z ϕΘ (t) := exp(itθ)S(dθ) = lim ϕΘ\|rn (t). (4.3) n→∞ Sd−1 We have the following results. Theorem 4.1. 1. Let X1 , . . . , Xn be iid observations generated from X, where X is multivariate regularly varying with index α > 1. Let Tn be the conditional empirical distance covariance between the angular and radial component defined in (4.1). Further assume that npn → ∞ and the weight measure µ satisfies Z (1 ∧ \|s\|β )(1 ∧ \|t\|2 )µ(ds, dt) < ∞, (4.4) Rd+1 for some 1 < β < 2 ∧ α. Then p Tn → 0. 2. In addition, if {rn } satisfies Z npn Rd+1 \|ϕ R ,Θ\|rn (s, t) − ϕ R \|rn (s)ϕΘ\|rn (t)\|2 µ(ds, dt) → 0, rn rn then d Z np̂n Tn → \|Q(s, t)\|2 µ(ds, dt), (4.5) (4.6) Rd+1 where Q is a centered Gaussian process with covariance function cov(Q(s, t), Q(s0 , t0 )) = (ϕR (s − s0 ) − ϕR (s)ϕR (−s0 ))(ϕΘ (t − t0 ) − ϕΘ (t)ϕΘ (−t0 )) (4.7) with ϕR , ϕΘ as defined in (4.2) and (4.3). Remark 4.2. In the case where X is regularly varying with index α ≤ 1, similar results hold if we replace R/rn with log(R/rn ) for which all moments exist. The proof of the theorem is delayed to Appendix A. In the following remark, we discuss certain sufficient conditions for assumption (4.5). Remark 4.3. Assume that µ = µ1 ×µ2 , where µ1 , µ2 are measures on R and Rd , respectively, and symmetric about the origin. From Section 2.2 of Davis et al. (2017), condition (4.5) is equivalent to R R0 npn E[µ̃1 ( − ) µ̃2 (Θ − Θ0 )\|R, R0 > rn ] rn rn R R0 +E[µ̃1 ( − )\|R, R0 > rn ] E[µ̃2 (Θ − Θ0 )\|R, R0 > rn ] rn rn R R0 −2 E[µ̃1 ( − )µ̃2 (Θ − Θ00 )\|R, R0 > rn ] → 0, (4.8) rn rn where Z µ̃i (x) = (1 − cos(xT s)) µi (ds), i = 1, 2. Let P R ,Θ\|rn denote the conditional joint distribution of (R/rn , Θ) given R > rn and P R \|rn , PΘ\|rn be the rn rn respective conditional marginals. Then (4.8) can be expressed as Z Z npn µ̃1 (T − T 0 ) µ̃2 (Θ − Θ0 ) (1,∞)×Sd−1 (1,∞)×Sd−1 6 Wan and Davis P R ,Θ\|rn (dT, dΘ)P R ,Θ\|rn (dT 0 , dΘ0 ) rn rn +P R \|rn (dT )PΘ\|rn (dΘ)P R \|rn (dT 0 )PΘ\|rn (dΘ0 ) rn rn −2P R ,Θ\|rn (dT, dΘ)P R \|rn (dT 0 )PΘ\|rn (dΘ0 ) r rn Zn µ̃1 (T − T 0 ) µ̃2 (Θ − Θ0 ) Z = (1,∞)×Sd−1 √ (1,∞)×Sd−1 npn P R ,Θ\|rn (dT, dΘ) − P R \|rn (dT )PΘ\|rn (dΘ) r rn n √ 0 0 npn P R ,Θ\|rn (dT , dΘ ) − P R \|rn (dT 0 )PΘ\|rn (dΘ0 ) → 0, rn rn (4.9) where (R0 , Θ0 ), (R00 , Θ00 ) are iid copies of (R, Θ). One way to verify (4.9) is to assume a second-order like condition on the distribution of (R, Θ). For example, assume that P R ,Θ\|rn − να × S w rn → χ, A(rn ) where χ is a signed measure such that χ([r, ∞] × B) is finite for all r ≥ 1 and B Borel set in Sd−1 , the unit sphere in Rd , and the scalar function A(t) → 0 as t → ∞. When the components of X are asymptotically independent, this is equivalent to the second order condition for multivariate regular variation (Resnick, √ √ 2002). If we choose the sequence rn such that npn → ∞ and npn A(rn ) → 0, then √ = √ npn A(rn ) npn A(rn ) P R ,Θ\|rn ((·, ·)) − P R \|rn (·) × PΘ\|rn (·) rn rn A(rn ) P R ,Θ\|rn ((·, ·)) − να × S((·, ·)) rn A(rn ) να (·) × (PΘ\|rn (·) − S(·)) w → 0 − A(rn ) − P R \|rn (·) − να (·) × PΘ\|rn (·) rn A(rn ) on [1, ∞] × Sd−1 . In the case where µ1 , µ2 are finite measures, µ̃1 , µ̃2 are bounded and (4.9) is satisfied since the integrand can be written as "Z Z µ̃1 (T − T 0 ) µ̃2 (Θ − Θ0 ) npn A2 (rn ) (1,∞)×Sd−1 (1,∞)×Sd−1 P R ,Θ\|rn (dT, dΘ) − P R \|rn (dT )PΘ\|rn (dΘ) rn # rn A(rn ) 0 P R ,Θ\|rn (dT , dΘ0 ) − P R \|rn (dT 0 )PΘ\|rn (dΘ0 ) rn rn A(rn ) → 0. In the special case that \|A\| ∈ RVρ for ρ < 0, (4.9) is met provided rn is chosen such that 1 1 O(n α+2\|ρ\| + ) ≤ rn ≤ o(n α ), for some > 0. When the measures µ1 , µ2 are infinite, (4.5) can be verified in specific cases. This is illustrated in the following example. Example 4.4. Let X follow a bivariate logistic distribution, i.e., X has cdf −1/γ P(X1 < x1 , X2 < x2 ) = exp(−(x1 −1/γ γ + x2 ) ), γ ∈ (0, 1). (4.10) Then X has asymptotically independent components if and only if γ = 1. It can be shown that X is regularly varying with index α = 1, i.e., pn = P(R > rn ) ∼ rn−1 as rn → ∞. Using the L1 -norm, k(x1 , x2 )k = \|x1 \|+\|x2 \|, Threshold Selection for Multivariate Heavy-Tailed Data 7 the pseudo-polar coordinate transform is (R, Θ) = (X1 + X2 , X1 /(X1 + X2 ))∈ (0, ∞) × [0, 1] and the pdf of (R, Θ) is fR,Θ (r, θ) = r − γ+1 γ −2 − γ1 − γ1 γ−2 θ (θ(1 − θ)) + (1 − θ) 1 γ γ − 1 1 − r−1 θ− γ + (1 − θ)− γ . γ e γ −1 −1 −r −1 θ γ +(1−θ) γ We now consider the case of the infinite weight measure µ given in (3.2) with κ = 1 and derive the condition on the sequence {rk } for which the conditions of Theorem 4.1 hold. First observe that f R ,Θ\|rn (t, θ) rn − γ+1 γ −2 − γ1 − γ1 γ −1 −1 −1 −1 −rn t θ γ +(1−θ) γ γ−2 (θ(1 − θ)) θ + (1 − θ) e γ γ − 1 1 1 − rn−1 t−1 θ− γ + (1 − θ)− γ γ 1 γ−2 γ+1 1 1 − γ − → t−2 (θ(1 − θ)) γ θ− γ + (1 − θ)− γ , γ =: fT (t)fΘ (θ), = t as n → ∞, (4.11) and rn f R ,Θ\|rn (t, θ) − fT (t)fΘ (θ) rn ≤ fT (t)fΘ (θ) rn e γ −1 −1 −1 −1 −rn t θ γ +(1−θ) γ γ −1 −1 −1 −1 −rn t θ γ +(1−θ) γ +e t−1 θ − γ1 −1 − γ1 + (1 − θ) γ γ 1−γ ! ≤ 1 γ γ 1 1 1 fT (t)fΘ (θ) t−1 θ− γ + (1 − θ)− γ + t−1 θ− γ + (1 − θ)− γ ≤ 1 2γ−2 1 t−3 θ− γ + (1 − θ)− γ ≤ ct−3 , γ 1−γ 1 1−γ for t ≥ 1 and θ ∈ [0, 1], where c denotes a generic a constant whose value may change from line to line throughout the paper, and the last inequality comes from the facts that 1 θ(1 − θ) ≤ 4 and θ − γ1 − γ1 + (1 − θ) 2γ−2 2−2γ γ 1 ≤ < ∞. 2 Letting hn (t, θ) := f R ,Θ\|rn (t, θ) − f R \|rn (t)fΘ\|rn (θ) rn rn rn−1 , we have 1 Z Z ∞ Z 1 Z \|hn (t, θ)\|dtdθ, max 0 Z 1 1Z ∞ ≤ ∞ \| log(t)hn (t, θ)\|dtdθ 0 1 \|thn (t, θ)\|dtdθ 0 Z 1 1 Z ≤ 0 1 ∞ f R ,Θ\|rn (t, θ) − fT (t)fΘ (θ) rn t−1 rn−1 dtdθ 8 Wan and Davis 1 Z ∞ Z fT (t) + dtdθ t−1 rn−1 f R \|rn (t) − fT (t) fΘ\|rn (θ) rn −1 −1 dtdθ, t rn 1 0 1 Z ∞ Z + 1 0 fΘ\|rn (θ) − fΘ (θ) where the first term can be bounded by 1 Z 0 Z ∞ ct−2 dtdθ < ∞, 1 and the other terms can be bounded in the same way. Since R has infinite first moment, we apply the distance correlation to log R and Θ. The integral in (4.9) is bounded by npn rn2 Z n c 3 rn Z 1 0 1 Z Z ∞ 1 ∞ Z Z 1 0 1 Z Z ∞ \| log t − log t0 \|\|θ − θ0 \|\|hn (t, θ)\|\|hn (t0 , θ0 )\|dtdθdt0 dθ0 1 ∞ (\| log t\| + \| log t0 \|)\|hn (t, θ)\|\|hn (t0 , θ0 )\|dtdθdt0 dθ0 1 0 1 0 Z 1 Z ∞ Z 1 Z ∞ n n ≤ c 3 \| log(t)hn (t, θ)\|dtdθ \|hn (t, θ)\|dtdθ ≤ c 3 , rn r 0 1 0 1 n ≤ which converges to zero if n = o(rn3 ). Therefore if {rn } is chosen such that rn = o(n) and n = o(rn3 ), then Theorem 4.1 holds. The result in Theorem 4.1 can be generalized from iid to a regularly varying time series setting, which we present in the next theorem. For a multivariate stationary time series {Xt } and h ≥ 1, set Yh = (X0 , . . . , Xh ). Then {Xt } is regularly varying if P(x−1 Yh ∈ ·) v ∗ → µh (·), P(x−1 kX0 k > 1) (h+1)d for some non-null measure µ∗h on R0 =R (h+1)d x → ∞, \{0}, R = R ∪ {±∞}, with the property that µ∗h (tC) = (h+1)d t−α µ∗h (C) for any t > 0 and Borel set C ⊂ R0 . See, for example, page 979 of Davis and Mikosch (2009). It follows easily that P(x−1 (X0 , Xh ) ∈ ·) v → µh (·), (4.12) P(kX0 k > x) where µh (D) = C · µ∗h ({s ∈ R (h+1)d : (s1 , sh ) ∈ D}). Assume that {Xt } is α-mixing. We assume the following conditions between {Xt } and the sequence of threshold {rn }, which can be verified for various time series models (Davis and Mikosch, 2009). −1 (M) Assume p−1 (kX1 k > rn ) = o(n1/3 ) and that there exists a sequence {ln } such that ln → ∞, n =P ln pn → 0, and i) δ X ∞ 1 αhδ → 0 for some δ ∈ (0, 1); (4.13) pn h=ln ii) lim lim sup h→∞ n→∞ ln 1 X P(kX0 k > rn , kXj k > rn ) = 0; pn (4.14) j=h iii) npn αln → 0. (4.15) Threshold Selection for Multivariate Heavy-Tailed Data 9 Theorem 4.5. Let {Xt } be a multivariate regularly varying time series with tail index α > 1 and αmixing with coefficients {αh }h≥0 . Assume the same conditions for the weight measure µ and the sequence of thresholds {rn } in Theorem 4.1, i.e., (4.4), (4.5) hold, and that condition (M) holds. Then Z d np̂n Tn → \|Q0 (s, t)\|2 µ(ds, dt), Rd+1 where Q0 is a centered Gaussian process. In particular, p Tn → 0. The proof of Theorem 4.5 is given in Appendix B. Note that the limiting distributions Q in Theorem 4.1 and Q0 in Theorem 4.5 are both intractable. In practice, quantiles of the distributions are calculated using resampling methods. While in the iid case this can be done straightforwardly, in the weakly dependent case one needs to apply the block bootstrap or stationary bootstrap to obtain the desired result (see Davis et al. (2012)). In the following section, we present a threshold selection framework with a subsampling scheme that does not require independence between the observations. 5. Threshold selection In this section, we propose a procedure to select the threshold for estimating the spectral measure S from observations X1 , · · · , Xn . Let us first consider the case where a specific threshold rn is given. Then (4.1) specifies the empirical distance covariance between R/rn and Θ conditional on R > rn . Under the assumption (4.5), we have from Theorem 4.1, Z np̂n Tn → \|Q(s, t)\|2 µ(ds, dt), Rd+1 R where np̂n is the number of observations such that Ri > rn . In practice, the limit distribution \|Q\|2 µ(s, t) is intractable, but one can resort to bootstrapping. Consider the hypothesis testing framework: H0 : R/rn and Θ are independent given R > rn H1 : R/rn and Θ are not independent given R > rn . Define the p-value for testing H0 versus H1 to be Z 2 pv = P \|Q(s, t)\| µ(ds, dt) > u Rd+1 . (5.1) u=np̂n Tn Then under H0 , pv follows U (0, 1), while under H1 , pv should be sufficiently small. Now consider a decreasing sequence of candidate thresholds {rk }. From (5.1), a sequence of p-values {pvk }, each corresponding to a threshold rk , can be obtained. Our goal is to find the smallest threshold r∗ such that conditional on R > r∗ , Θ can reasonably be considered independent of R. Note that the pvk ’s are not independent for each k since they are computed from the same set of data. Conventional multiple testing procedures, such as Bonferroni correction, are problematic to implement for dependent p-values. To counter these limitations, we propose an intuitive and direct method based on subsampling. The idea is outlined as follows: For a fixed level rk , we choose a subsample of size nk from the conditional empirical cdf F̂ R ,Θ\|rk of (Ri /rk , Θi ) with Ri > rk , i = 1, . . . , n. For this subsample, we compute the rn distance covariance Tn,k . To compute a p-value of Tn,k under the assumption that the conditional empirical distribution is a product of the conditional marginals, we take a large number (L) of subsamples of size nk from F̃ R ,Θ\|rk (dθ, dr) = F̂Θ\|rk (dθ)F̂ R \|rk (dr), rn rn (l) and calculate the value T̃n,k , l = 1, . . . , L for each subsample. The p-value of Tn,k , pvk , is then the empirical p(l) value of Tn,k relative the {T̃n,k }l=1,...,L . This process, starting with an initial subsample of nk from F̂ R ,Θ\|rk is rn 10 Wan and Davis (j) repeated m times, which produces m estimates {pvk }j=1,...,m of the pvk , which are independent conditional on the original sample. These are then averaged m pv k = 1 X (j) pv . m j=1 k So for the sequence of levels {rk }, we produce a sequence of independent p-values {pv k }. Our choice of threshold r at which (Θ, R)\|R > r are independent (and dependent otherwise) will be based on an examination of the path of the mean p-values, {pv k }. Note the following two observations: (1) (m) • If R and Θ are independent given R > rk , then the pvk , . . . , pvk will be iid and approximately U (0, 1)-distributed, so that pv k should center around 0.5. (j) • If R and Θ are dependent given R > rk , then the pvk ’s will be well below 0.5 (closer to 0), and so will pv k . By studying the sequence {pv k }, which we call the mean p-value path, we choose the threshold to be the smallest rk such that pv l is around 0.5 for l < k. A well-suited change-point method for our situation is the CUSUM algorithm, by Page (1954), which detects the changes in mean in a sequence by looking at mean-corrected partial sums. In our algorithm, we use a spline fitting method that is based on the CUSUM approach called wild binary segmentation (WBS), proposed by Fryzlewicz (2014). The WBS procedure uses the CUSUM statistics of subsamples and fits a piecewise constant spline to {pv k }. In our setting, we may choose rk to be the knot of the spline after which the fitted value is comfortably below 0.5. There are several advantages to using the subsampling scheme. First, recall that the p-value path {pvk }, which is obtained from the whole data set, has complicated serial structure and varies greatly from each realization. In contrast, the mean p-values pv k from subsampling are conditionally independent and will center around 0.5 with small variance when the total sample size n and the number of subsample m is large. This, in turns, helps to present a justifiable estimation for the threshold. Second, the calculation of distance covariance can be extremely slow for moderate sample size. Using smaller sample sizes for the subsamples, our computational burden is greatly reduced. In addition, this procedure is amenable to parallel computing, reducing the computation time even further. Third, the subsampling makes it possible to accommodate stationary but dependent data, waiving the stringent independent assumption. The idea of looking at the mean p-value path is inspired by Mallik et al. (2011), which used the mean of p-values from multiple independent tests to detect change points in population means. 6. Data Illustration In this section, we demonstrate our threshold selection method through simulated and real data examples. In practice, we set the sequence of thresholds {rk } to be the corresponding upper quantiles to {qk }, a pre-specified sequence of quantile levels. The subsample size nk at each threshold rk is set as nk = n0 · qk for some n0 << n. This is designed such that for any rk , each subsample is a n0 /n fraction of all the eligible data points with R > rk . Then the choice of {nk } boils down to the choice of n0 , which should reflect the following considerations: i) n0 should be large enough to ensure good resolution of p-values at all levels; ii) n0 /n should be sufficient small such that the subsamples do not contain too much overlap in observations; iii) larger n0 requires heavier computation for the distance correlation. In our examples, where the total sample size n ranges from 3000 to 20000, we find n0 between 500 and 1000 to be a suitable choice. The number of subsamples m can be set as large as computation capacity allows. In our examples, we take m = 60. For all the examples, we choose the weight function µ for distance covariance to be (3.2) with κ = 1, and the number of replications used to calculate each p-value is L = 200. To ensure that the moment conditions are met, the distance correlation is applied to the log of the radial part R in all examples. 6.1. Simulated data with known threshold To illustrate our methodology, we simulate observations from a distribution with a known threshold for which R and Θ become independent. Threshold Selection for Multivariate Heavy-Tailed Data 11 Let R be the absolute value of a t-distribution with 2 degrees of freedom and Θ1 , Θ2 be independent random variables such that Θ1 ∼ U (0, 1), Θ2 ∼ Beta(3, 3). Set ( Θ1 , if R > r0.2 , Θ\|R = Θ2 , if R ≤ r0.2 , where r0.2 is the upper 20%-quantile of R. Then R and Θ are independent given R > r if and only if r ≥ r0.2 . Let (Xi1 , Xi2 ) = (Ri Θi , Ri (1 − Θi )), i = 1, . . . , n, be the simulated observations. We generate n = 10000 iid observations from this distribution. Figures 1(a), 1(b) and 1(c) show the data in Cartesian and polar coordinates. Our goal is to recover the tail angular distribution by choosing the appropriate threshold. A sequence of candidate thresholds {rk } is selected to be the empirical upper quantiles of R corresponding to {qk }, 150 equidistant points between 0.01 and 0.4. We apply the procedure described in Section 5 to the data. For each rk , the mean p-value pv k is calculated using m = 60 random subsamples, each of size nk = 500 · qk , from the observations with Ri > rk . Figure 1(d) shows the mean p-value path. For the WBS algorithm, we set the threshold to be the largest rk such that for all thresholds r (quantile level q) such that r < rk (q > qk ), the fitted spline of the p-value stays below 0.451 . The threshold levels chosen is 20.4%, which are in good agreement with the true independence level 0.2. The empirical cdfs of the truncated Θi ’s corresponding to the chosen thresholds is shown in Figure 1(e). We can see that the true tail angular cdf (i.e., U (0, 1)) is accurately recovered. 6.2. Simulated logistic data We simulate data from a bivariate logistic distribution, which is bivariate regularly varying. Recall from Example 4.4 that (X1 , X2 ) follows a bivariate logistic distribution if it has cdf (4.10). In this example, we set γ = 0.8 and generate n = 10000 iid observations from this distribution. Similar to the previous example, for each threshold rk corresponding to the upper qk quantile, where {qk } is chosen to be the 150 equidistant points between 0.01 and 0.3. The mean p-value pv k is calculated using m = 60 random subsamples of size nk = 500 · qk from the observations with Ri > rk . Figures 2(a) , 2(b) and 2(c) show the scatterplots of the data. Here the L1 -norm is used to transform the data into polar coordinates. Our algorithms suggests using 7.4% of the data to estimate the angular distribution. The estimated cdf of the angular distribution is shown with the theoretical limiting cdf, derived from (4.11), in Figure 2(e). So even though R and Θ are not independent for any threshold rk , our procedures produce good estimates of the limiting distribution of Θ. 6.3. Real data In this example, we look at the following exchange rate returns relative to the US dollar: Deutsche mark (DEM), British pound (GBP), Canadian dollar (CAD), and Swiss franc (CHF). The time spans for the data are 1990-01-01 to 1998-12-31 with a total of 3287 days of observations. We examine the pairs GBP/CHF, CAD/CHF, DEM/CHF and estimate the angular density in the tail for each pair. Figures 3(a)–3(c) present the scatter plots of the data. The marginals of the observations are standardized using the rank transformation proposed in Joe et al. (1992): Zi = 1/ log{n/(Rank(Xi ) − .5)}, i = 1, . . . , n. Again {qk } is chosen to be the 150 equidistant points between 0.01 and 0.3, and the mean p-value pv k is calculated using m = 60 random subsamples of size nk = 500 · qk from the observations with Ri > rk . Note that while it may not be reasonable to view the observations as iid, the subsampling scheme can still be applied to choose the threshold of independence between R and Θ. 1 Of course, other selection rules can be used. For example, a more conservative approach would be choosing the threshold as the largest rk such that for r > rk , the fitted spline of the p-value stays above 0.45. ● ● ●● ●● ● ●●● ● ● ● ●●●●● ●●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ●● ● ●● ●● ● ● ● ●● ●● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●●● ●●●● ●● ● ● ● ● ●●●● ● ● ● ●● ● ●●● ● ●● ● ●● ●● ●● ●● ●● ● ● ● ● ●● ●● ●● ● ● ●●● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ●● ●● ●● ●● ● ● ●● ● ● ● ● ●● ● ●● ● ●●● ●●● ● ●● ● ● ● ● ●● ●● ●●● ●● ● ● ●● ● ● ●●● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ●●● ●● ● ● 0 ●● ● ● ● ● ●● ● ● ● 20 ● 40 60 80 ● ● ● ●● ● ●● ● ● ● ● ● ● radius ● ● ● ● ● ● 1e−04 1e−02 1e+00 ● 1e+02 ● ● ● 0.0 0.2 ● ● ● ● ● 0.6 0.8 1.0 angle (c) 1.0 1.0 ● ● 0.4 (b) estimated truth 0.2 0.2 0.4 0.4 cdf 0.6 0.6 0.8 mean−pv change point fitted spline 0.8 ● ● ● ● x1 (a) ● 0.0 0.0 ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●●●● ● ●● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ●●●● ●● ● ● ● ● ● ● ● ● ● ●● ●●● ●● ●● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ●● ●●● ●●●●● ●●● ● ● ● ●●● ● ● ●●●● ●● ●● ●● ● ●● ● ● ●●●● ●● ● ●● ● ●●● ● ● ●● ●● ● ●● ● ●● ●● ● ●● ● ● ● ● ●●● ●● ●● ● ● ●● ●● ● ●●● ●●●●● ● ● ●● ● ● ● ● ●● ●●●● ●● ● ●● ●● ● ●●●● ● ●● ● ●● ●●● ● ●● ●● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ●●● ● ● ● ● ● ● ●● ●● ● ●●● ●● ●● ● ●● ●● ● ● ● ● ●●● ● ● ● ●● ●● ● ●● ● ● ● ●● ●● ● ●● ●● ● ● ●● ● ●● ● ● ●●●● ●●●●● ● ● ●● ● ●● ● ● ●● ● ●● ● ● ● ● ● ●● ●● ●● ● ● ●● ●● ● ● ● ● ●●●● ●● ● ● ●● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ●● ●●● ●●● ●●● ●●●● ● ● ●● ● ●● ●● ● ●●● ● ● ● ●● ● ● ●● ● ● ● ● ●●● ● ● ●● ● ● ●● ●●● ●● ● ● ●● ● ● ● ● ●●● ● ● ● ● ●●●●● ● ● ●● ●●● ●● ●● ● ●● ●●● ● ●● ● ● ● ● ●● ● ●● ● ●●● ● ●● ● ● ● ●● ●●●●● ● ●●●●● ● ●●● ●● ● ●●● ● ●● ● ● ● ● ● ●● ● ● ●●● ● ●● ● ● ●● ● ●● ● ●● ● ● ● ● ●● ● ●● ●● ●● ● ●● ● ● ● ●● ● ●● ● ●●● ● ● ● ● ● ● ● ●●● ● ● ● ●●● ●●●● ● ●●●●● ●● ● ●● ●● ● ●● ●●● ●●● ● ●● ●● ●●● ●● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ●● ● ●●● ● ● ●● ●●●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●●● ● ●●● ●●● ● ● ● ●● ●● ● ●● ● ●● ●●●●● ● ● ●●● ●● ● ● ● ● ●● ● ● ● ●● ●●●● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ●● ● ●● ● ●●● ● ● ●● ● ● ● ● ●●● ●●● ●● ● ●●● ● ● ● ●● ● ●● ● ● ● ●● ● ● ● ●● ● ● ● ●●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ●●●● ● ● ● ●●● ● ●● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ●●● ●● ●● ●● ● ● ● ● ● ●● ●●●● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●●● ● ●● ●● ● ● ●● ●● ●● ●● ● ●● ● ● ● ●● ●●● ●● ● ● ●● ● ●● ●●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●●● ●● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ●● ● ● ●● ● ●● ●● ● ● ● ●●●●● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ●● ●●●● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ●● ●● ● ● ● ●●● ● ● ● ● ●● ● ●● ●● ●● ●● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ●●● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●●●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ●●●●●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ●● ● ● ● ● ●●●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ●● ● ● ● ●● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ●● ● ●● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ●● ●●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●●● ●● ● ● ●● ●● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●●●● ● ● ●● ● ● ● ●● ●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ●● ●●●●● ● ● ●● ● ●● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ●●●● ● ● ●● ● ●●● ●● ● ● ● ● ● ●● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●●●● ●●● ● ●●● ●● ● ● ● ●● ● ● ●●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ●●● ● ●● ● ● ●● ● ●● ● ●● ●●● ● ● ● ● ● ● ●● ●● ●● ● ● ●● ● ● ● ● ● ● ●●●● ● ● ●● ●● ● ●●● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●●●● ● ● ● ●●● ● ●●● ● ● ●●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ●●● ●● ● ● ● ● ● ● ●● ● ● ● ●●● ●●●●● ● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ●● ●●●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●●● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ●●● ● ● ● ● ●●● ● ●●● ● ● ● ● ● ●●● ●● ● ● ● ●●● ● ● ●●●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ●● ● ●● ● ● ●● ● ● ● ● ● ● ●● ●●●● ● ● ● ●● ● ● ●● ●● ●● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●●● ● ● ●● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ●●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●●● ●● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ●● ●● ● ● ● ●●● ● ● ●● ● ● ●● ● ●●● ●●● ●●● ●● ● ●●● ● ● ● ● ●● ● ●● ● ●● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ●●● ●● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●●● ● ● ● ● ●●●● ●●● ● ●●●● ●● ● ● ● ● ● ●● ●● ●● ● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ●●● ● ●● ● ● ●● ● ● ● ●●● ● ●● ●●● ●● ●●● ●● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ●● ●● ● ● ● ● ●●● ●● ● ●● ● ●● ● ● ● ●● ●● ●●● ● ● ●● ● ● ● ● ● ●● ● ●●●● ●●●● ●● ● ● ● ●● ●●●● ● ● ●● ● ● ● ●● ● ● ● ●● ●● ●● ● ●● ● ● ●● ● ●●●●● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ●●● ●●● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●●● ●●●●●●● ● ● ●● ● ●●●●● ● ●●● ● ● ●●●●● ● ●●●● ●● ● ●●●●● ● ●● ● ●● ●●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ●● ● ● ●● ● ●● ● ●●● ●● ● ●● ●● ●● ● ● ●●● ● ● ●● ● ● ●●● ●● ●● ●●●● ● ● ●●● ●● ● ● ●●● ● ● ● ●● ● ●●● ●● ● ●● ●●● ●● ● ● ● ● ●●●●●● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ●● ●●● ● ● ●● ● ● ●●● ●● ● ● ● ● ●● ● ●● ●● ● ●●●● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●●●●● ●● ● ● ● ●● ●● ● ●● ● ● ● ●● ● ● ●● ●● ● ● ●● ●● ●● ● ● ●●● ●●● ● ●●● ● ● ● ● ● ● ●● ●● ●●●●● ● ● ● ●● ● ●●●● ●● ● ● ●●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ●● ● ●●● ● ●●● ●● ● ● ●● ●● ● ●● ● ● ●● ● ● ● ●●● ● ● ● ●●● ● ● ● ● ● ●● ● ●●● ● ● ● ● ●● ●●● ●● ●● ● ● ● ●● ● ●●●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ●● ● ● ●●● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ●●●● ● ●●● ●●● ● ● ● ●●●●● ●● ●● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ●● ● ●●● ●● ●●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 1e+01 ● ● ● ● ● ● 1e−01 ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ●● ● ● ●●● ●● ● ●●●● ●● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ●● ● ●●● ● ● ●● ● ●● ● ● ●● ● ●●● ● ●● ● ● ● ● ●●● ● ● ●● ● ●● ● ● ● ●●● ●●●● ● ● ● ● ●● ●● ●● ●● ● ●●● ●● ● ● ●● ● ● ● ●● ● ●●●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ●● ●●●● ● ● ● ●● ●● ●●● ● ● ● ● ● ● ●●● ●●● ●●● ● ●● ● ●●● ● ● ● ● ●●●● ● ●● ● ● ● ●● ●●● ●●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ●●● ●●● ● ● ●● ● ●● ●●● ●● ●● ● ● ● ● ● ●●● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ●●● ● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●●● ● ●● ● ●● ●● ●● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●●●●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●●● ● ● ● ● ● ●● ● ●● ● ● ●●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ●●● ●● ● ● ● ●● ●● ● ● ●● ● ●●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ●●●● ● ●● ● ● ● ● ●●● ● ● ●●●● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●●●● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ●●● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ●●● ● ● ● ●● ● ●● ● ●● ●● ● ●● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ●● ● ● ●●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ●● ● ●● ●● ●● ● ● ● ● ●● ●● ● ● ● ●●● ● ● ●●● ● ● ● ● ● ● ● ●●●●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ●● ● ● ● ● ●● ● ● ● ●●●● ●● ● ● ● ●●● ● ● ●● ●● ●● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ●●● ●● ●● ●● ●●● ● ●●● ● ● ● ● ● ● ● ●●● ● ●●● ● ● ●● ● ● ●●● ●● ● ● ● ● ● ● ●● ● ●●●● ● ●● ● ● ● ●●● ● ●● ● ● ● ● ●● ● ● ● ●●● ● ● ●● ●● ● ● ● ● ● ●● ● ●● ●● ● ●● ● ●●● ● ● ● ●● ●● ● ● ● ●●● ●●● ● ● ● ● ● ● ● ●●● ●● ●● ● ● ● ●●● ● ●● ● ● ● ●● ●● ● ●●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ●● ● ●● ● ●● ● ● ● ● ●● ● ●●● ● ● ●● ●● ● ● ● ● ●● ● ●● ● ● ●●● ● ● ●●●●● ●● ●● ●●● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ●● ● ●●● ●●● ● ●● ●●● ●● ● ●● ●● ●● ●● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ●● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●●●● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ●● ● ●● ● ● ●● ●● ● ● ●● ●● ● ● ●● ● x1 p−value ● ● 1e−03 x2 300 0 100 x2 500 ● 1e+03 Wan and Davis 1e−04 1e−02 1e+00 1e+02 12 0.0 0.1 0.2 0.3 0.4 0.0 upper quantile 0.2 0.4 0.6 0.8 1.0 Theta (d) (e) Fig 1. Example 6.1. (a) scatterplot of (Xi1 , Xi2 ); (b) scatterplot of (Xi1 , Xi2 ) in log-log scale; (c) scatterplot of (Ri , Θi ); (d) mean p-value path (black triangles), fitted WBS spline (blue line), and the chosen threshold quantile (red vertical line); (e) estimated cdf of Θ using the threshold chosen, compared with the truth (black dotted). The mean p-value paths are shown in Figures 4(a)–4(c). The threshold levels selected for the three pairs are 9.6%, 7.4%, 16%, respectively. Figures 3(d)–3(f) show the shape of the estimated angular densities for each pairs. As expected, the tails of the two central European exchange rates, DEM and CHF, are highly dependent. In contrast, that of CAD and CHF are almost independent. 6.4. Simulated non-regularly varying data In this example, we generate data from a model which is not regularly varying. Let R be a random variable from the standard Pareto distribution: P(R > r) = r−1 , r ≥ 1. Let Θ1 , Θ2 be independent random variables such that Θ1 ∼ U (0, 0.5), Θ2 ∼ U (0.5, 1). Set ( Θ1 , if log R ∈ (2k, 2k + 1] for some integer k, Θ\|R ∼ Θ2 , if log R ∈ (2k + 1, 2k + 2] for some integer k. For any positive integer k, it can be verify that P(Θ ∈ (0, 0.5)\|R > e2k ) = 1 − e−1 , 1 − e−2 ● ● ● ● ● ● ● ● ● 5000 15000 25000 ● ● 1e−01 ● ● ● ● ● ● ● 1e+01 ● ● ● ● ● ● 1e+03 0.0 0.2 0.4 1.0 0.8 0.6 estimated truth cdf 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.10 0.15 0.20 0.25 ● 0.8 1.0 (c) 1.0 (b) ● 0.05 ● angle mean−pv change point fitted spline 0.00 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●●● ● ● ● ●● ● ● ● ● ●● ● ●●● ● ● ● ● ●● ● ● ●●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ●● ● ●● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ●● ●● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ●● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●●● ●● ● ● ● ●● ● ● ● ●●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ●●● ●● ● ●● ●● ● ● ● ●●● ●● ● ●● ●● ●● ● ● ●●● ● ● ● ● ●● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●●● ● ●● ●● ●● ●● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ●●● ●● ● ● ● ● ●●● ●● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ●●●● ● ● ●●●●●● ● ●● ● ● ●●● ● ●● ● ●● ● ● ●● ●● ●● ● ● ●● ● ●● ●● ●● ● ●● ●● ● ● ● ●● ● ● ●● ● ● ●● ●● ● ● ● ●●● ● ● ● ●●● ● ● ●●● ● ● ● ●● ● ● ●● ● ● ● ● ●●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●●● ●●● ●●●●● ● ● ● ● ●●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ●●●● ● ●● ● ●● ●● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ●● ● ●●●● ●●●● ●●●● ● ●● ● ● ● ●●● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ●● ●● ●●● ●●● ● ● ●● ● ●● ● ● ● ●● ● ●●● ● ● ●● ●●● ●● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ●●● ● ● ●●● ● ●●● ● ● ● ● ●●●●● ●● ● ● ● ● ●●● ●●●●●● ●●●● ● ●●● ● ● ●●● ●●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ●● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ●●● ●● ● ●● ●● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ●●● ●● ●●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ●● ●● ●● ● ● ● ● ●● ●● ●● ● ● ● ●●●● ●● ● ● ● ● ● ●●● ● ●● ● ●● ●●●● ● ● ● ●● ●● ● ●● ● ● ● ●● ● ●●●● ●●●● ●●● ●● ●● ● ●● ●●●● ● ● ●●● ●●●● ● ●● ● ●● ● ● ● ●●● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ●●●●●●●●●● ●●● ● ●● ●●●●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●●● ●● ● ●● ●● ● ● ●● ● ● ●●● ●● ●● ●●●●●● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●●● ●● ●●● ●●●●●● ●● ● ●● ●● ● ●● ● ● ● ● ●● ● ●● ● ●● ●● ●●● ● ● ●● ● ● ● ● ● ● ●● ●● ● ●● ● ● ●●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ●● ●● ● ● ●●● ● ● ● ● ●● ● ●●● ●●● ● ●● ● ● ● ●●●● ● ● ● ● ● ●●● ● ● ●●●● ●●● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ●● ● ●●● ● ● ● ● ●● ● ●●●● ● ●●●● ● ● ●● ● ● ● ●● ● ●●● ● ●● ●●● ● ●● ● ●● ● ●● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ●● ● ● ●●● ● ● ●● ●● ● ●● ● ● ● ●● ● ● ●● ●● ● ●● ● ●● ● ● ● ●● ●●● ●●● ● ●● ● ● ● ● ●● ●● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ●● ●● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ●●●● ● ● ●● ● ●● ● ● ●● ●● ● ●●● ● ●● ●● ●●●● ● ●● ●●●● ● ●● ● ● ●● ● ● ● ●● ●● ●●● ● ● ● ● ● ●● ●● ●● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ●● ●● ● ●●●● ● ●● ●● ● ●● ● ● ●● ●● ● ●● ●● ● ●● ●● ● ●● ● ● ● ● ● ●●●● ●●●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●●● ● ●● ● ● ●● ● ● ●● ●● ●● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●●● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ●●● ●● ● ● ●●●● ●● ●●● ● ● ●●●● ● ● ● ● ●● ● ● ● ●●●● ● ● ●● ● ● ● ● ● ●● ● ●●●●● ● ● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ●●● ● ● ●● ● ● ● ●● ●●● ●● ●● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ●● ●●● ● ● ● ● ●●●● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ●● ● ●● ●● ● ● ● ● ●● ● ● ●● ● ● ●● ● ●●● ● ● ● ● ●●● ● ● ● ●● ●●● ● ● ● ●● ● ●● ● ● ● ●● ● ●● ●● ● ● ●● ●●● ● ● ● ● ●● ● ● ●● ●●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●●● ●●● ● ● ● ●● ● ● ● ● ● ●●● ●● ●● ● ●● ● ● ● ●● ● ● ● ● ●●●● ●● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ●● ●● ●●● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ●●● ●● ● ●● ● ●● ● ● ● ● ●●● ● ● ●● ● ●●●● ● ●● ● ●● ● ● ●● ●● ●● ● ● ●●● ● ● ●●● ●●● ●● ●● ● ●● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ●● ●● ● ● ●● ●● ● ●● ● ●●● ● ● ● ● ● ● ● ●● ● ●● ●● ● ●● ● ●●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ●●● ● ● ●● ● ● ●● ●●● ● ●● ● ● ●● ●●● ● ● ● ● ● ●● ● ●●●●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ●● ● ● ● ● ● ●● ● ●●● ●● ● ● ● ● ● ● ● ● ●● ●●● ● ● ●● ● ●●● ● ● ● ● ● ●●●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ●●● ●●● ●● ●● ● ●●● ●● ●● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ●● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ●●●● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ●● ●● ●● ● ● ● ●● ●● ● ● ●●● ● ●●● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ●●● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ●●● ●●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ●● ●● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ●●●● ●● ●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ●● ●● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ●●●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ●●●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ●●● ● ● ●● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●●● ● ● ●● ●●● ● ● ● ● ●● ● ●● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ●●● ● ●● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ●● ● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●●●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ●●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ●●●● ● ● ●● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ●● ● ● ● ● ●● ●● ●● ● ● ●●●● ● ●● ● ●●● ●● ● ●●●● ●●● ● ● ●●●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●●●● ●● ● ●● ● ●●● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ●●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ●●●● ● ●● ● ● ●● ● ● ● ●● ● ● ● ●● ●●●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ●● ● ● ● ●● ●● ● ● ●● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ●● ● ●●● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●●●● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ●● ● ●● ●● ● ● ● ● ● ●● ● ●●●● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●●● ● ● ●● ● ●●● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ●● ●● ● ● ● ● ● ● ●● ●●● ●● ● ● ●● ● ● ● ● ● ● ●● ●● ●● ●●● ●●●● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ●● ●● ●●● ●● ● ●● ● ● ●●●● ● ●● ● ● ●●●● ● ●● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ●●● ●● ● ● ● ● ●● ● ●●●●● ●●●● ● ●● ●● ●●● ● ●● ● ●● ● ●● ● ●● ● ● ● ● ●● ● ●●● ●● ● ● ●● ● ●● ● ●● ●●●● ●● ● ● ●● ● ●●● ● ● ● ● ● ●● ●● ● ●●● ● ● ●● ● ●● ●● ● ●● ●● ● ●●●● ●●● ● ● ● ●● ● ●●●●● ●● ● ●●● ● ● ● ● ● ● ● ●●● ●●● ● ● ● ●● ● ● ● ●● ●● ● ●● ● ● ●● ● ●● ● ● ●● ●● ● ● ● ● ●● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● x1 (a) 0.0 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● x1 p−value ● ● ● ●● ● ● 10 ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ●●●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●●●● ● ● ● ●●● ● ● ●● ● ●● ● ● ● ●●● ● ●●● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ●●● ● ● ● ● ● ●● ●●●●● ● ●● ●●●● ● ● ●● ● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ●●●●● ● ● ●●●● ● ● ●● ● ● ● ● ●● ●● ● ●●●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ●● ● ●●●● ● ● ●● ●●● ●● ●● ●● ●● ●● ●● ● ●● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ●●●● ●● ●● ●● ● ● ● ●●●● ●● ●● ● ● ● ●●● ● ● ●● ●●●●● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ●●● ●● ●● ●● ● ●● ● ● ● ● ●● ●● ● ●● ● ● ● ● ●● ● ●● ● ● ●●●●●● ● ● ● ● ● ●●● ● ● ●●●● ●●● ●● ● ● ●● ●●●●● ●●●● ● ● ●●●● ● ● ●● ● ●● ●● ● ● ● ● ●● ● ● ● ●● ●●●●● ● ● ●● ●● ● ● ● ● ● ● ● ●●●● ●● ● ●● ● ● ● ● ● ● ●●●● ●● ● ● ●●●● ● ● ●● ● ●● ● ●● ●●● ● ● ●● ● ● ●● ● ● ●● ●●●● ● ● ● ●● ●●●●●●● ● ●● ●●● ●● ●● ● ● ● ● ● ● ●● ● ●●● ● ● ●● ●●● ● ● ● ● ● ●●● ●● ● ●● ● ●●●● ● ●● ●● ● ● ● ● ●●● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ●● ●● ●●●● ●● ●●● ● ●● ● ● ● ● ●● ● ● ● ●●● ● ● ●●● ●● ●● ● ● ●● ● ● ● ● ●●● ● ● ●● ●● ●● ●●● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●●● ●● ● ● ●● ● ●●●● ● ● ● ●● ● ● ● ●● ● ● ●●● ●● ● ● ● ● ● ●●● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ●●●● ● ●●● ● ● ● ●● ● ●●● ● ● ● ●● ●● ● ● ● ● ● ●●● ● ● ● ●●●●● ●● ●●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ●● ●●● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ●● ● ●● ● ●●● ●● ●● ●● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ●● ●●●● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ●●● ● ● ● ● ●● ● ●●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ●●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●●● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ●●● ●● ● ● ● ● ● ● ●● ●● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ●●●● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●●●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ●● ● ●●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●●●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ●● ● ● ● ●● ● ●● ●● ● ● ● ● ● ●● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ●●● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ●●● ●● ● ●●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ●●●● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ●● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ●●●● ● ●● ● ● ●● ● ●●● ● ● ●●● ●● ● ●● ●● ● ● ●●● ●●● ●●● ● ● ● ● ● ● ●● ● ● ●●● ● ● ●● ● ● ●●● ● ● ● ● ● ● ●●●●● ● ● ● ● ● ● ● ● ● 100 ● radius ● ● ● 0.8 0 ● ● ● ● ● 1 ● 1e−01 0 ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● 1e+01 x2 10000 x2 1e+03 25000 ● 13 10000 Threshold Selection for Multivariate Heavy-Tailed Data ● ● ● ● ● ●● ●● ● ● ●● ● ●● ● ● ●●● ● ●●● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●●●● ●● ●●●● ● ●● ● ● ● ●● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ●● ● ●● ● ●● ● ●● ● ●● ● ●●● ●● ● ● ●● ● ●● ●● ●● ●● ●● ●●●●● ● ● ● ● ●● ●●● ● ● ●●● ● ● ● ● ● ● ● ●● ●● ● ● ●● ●● ●● ● ●● ● ● ●● ● ● ●●● ● ● ● ● ● ●● ● ● ●● ● ● ● ●●● ● ● ●● ● ● ●● ● ●● ●● ● ●●● ●● ●●● ● ●●● ● ●● ● ●● ●● ● ● ● ●●● ● ● ● ●● ●● ●● ● ● ● ●● ● ● ● ●● ● ●● ●● ●● ●● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● 0.30 0.0 upper quantile (d) 0.2 0.4 0.6 0.8 1.0 Theta (e) Fig 2. Example 6.2. (a) scatterplot of (Xi1 , Xi2 ); (b) scatterplot of (Xi1 , Xi2 ) in log-log scale; (c) scatterplot of (Ri , Θi ); (d) mean p-value path (black triangles), fitted WBS spline (blue line), and the chosen threshold quantile (red vertical line); (e) estimated cdf of Θ using the threshold chosen, compared with the theoretical limiting cdf (black dotted). while e−1 − e−2 . 1 − e−2 Hence P(Θ ∈ ·\|R > r) does not convergence as r → ∞ and X = (RΘ, R(1 − Θ)) is not regularly varying. Let (Xi1 , Xi2 ) = (Ri Θi , Ri (1 − Θi )), i = 1, . . . , n, be iid observations from this distribution, where n = 20000. Figures 5(a), 5(b) and 5(c) show the data in Cartesian and polar coordinates. We apply our threshold selection algorithm to the data, with the threshold upper quantile levels qk chosen as the 150 equidistant points between 0.01 and 0.2. The mean p-value pv k is calculated using m = 60 random subsamples of size nk = 500 · qk from the observations with Ri > rk . This is shown in Figure 5(d). In this model, the radial part R is regularly varying, but Θ and R are dependent given R > r for any r. We expect the mean p-values to be well below 0.5, as are observed. No threshold is selected by the algorithm. This suggests that our technique can potentially be used to detect misspecified models from the regular variation assumption, especially in the scenario where the heavy-tailedness of R is observed but dependence between R and Θ is suspected. P(Θ ∈ (0, 0.5)\|R > e2k+1 ) = 7. Discussion In this paper, we propose a threshold selection procedure for multivariate regular variation, for which R and Θ are approximately independent for R beyond the threshold. While our problem is set in the multivariate heavy-tailed setting and we utilize distance covariance as our measure of dependence, our algorithm is 14 Wan and Davis ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ●● ●●● ●● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ●●●● ● ●●●●● ● ● ● ● ●●● ●● ●● ●● ● ● ● ●● ● ● ●● ● ● ●●●● ● ●●●● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ●● ● ● ●●●● ●●●● ● ● ● ●●● ●●● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ●● ● ●● ● ● ● ● ●● ● ● ●● ●●● ●●●●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ●●● ●●● ● ● ● ●● ● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ●●●● ● ●● ●●● ● ● ●●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ●●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●●● ●● ●● ● ● ● ● ● ● ● ●●● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●●● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●● ●●● ● ●●● ● ● ● ● ● ●● 0.00 0.01 0.02 0.01 ● ● ● 0.03 0.04 0.05 ● ● ● ● ● ● 0.03 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●●●● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ●● ●● ● ●●●● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ●●●●●● ● ● ●● ●● ●● ● ●● ●●● ● ● ● ● ●● ● ●●●● ●● ● ● ● ● ● ●● ●● ●● ● ● ● ●● ●●● ● ●● ●●● ● ● ●● ●● ●● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ●● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●●●● ● ● ● ● ●● ●●● ● ●● ●●● ● ●● ●● ●●● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ●●●● ●●●● ● ● ● ● ● ●● ● ● ● ●● ●●● ● ● ● ● ● ● ●● ● ●●●● ● ●● ● ●● ● ●● ● ●● ●● ● ●●●● ● ●●●● ● ● ● ●● ●● ●● ● ● ●● ●●●●● ● ● ● ●●● ● ● ● ●●● ●● ● ● ● ● ● ● ●● ● ●● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ●●●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●●●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ●●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0.005 GBP ● ● ● ● ● 0.010 ● ● 0.000 ● ● 0.010 0.020 ● 0.030 DEM (c) DEM/CHF 5 1.2 CAD/CHF 0.8 Density 0.4 3 2 0.0 0 0.0 1 0.5 1.0 Density 4 1.5 ● ● ● ● (b) GBP/CHF ●● ● ● ● 0.015 ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●●● ● ● ● ● ● ●● ●●● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ●● ● ●●●● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ●● ●● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●●● ● ●●● ●●● ● ● ● ●● ● ●● ● ● ● ●●● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ●●●● ●● ● ● ● ● ●● ●● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ●● ● ●● ●● ●● ●●● ● ● ●● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ●●●●● ● ● ● ● ●● ●●●● ● ● ●●● ● ●●● ●● ● ● ●● ●● ● ● ● ● ●●●● ● ● ●● ● ●● ● ● ●●● ●●●● ●●● ● ● ● ●●● ● ●●● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ●● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ●●● ● ● ● ● ●●● ●●●● ● ●●●● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ●●● ● ●● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●●● ●●● ●●● ● ● ●●● ● ● ● ●● ● ●● ● ● ● ●● ● ●● CAD (a) Density ● ● ● ● ● 0.000 ● ● ● ● ● ●●● ● ● ● 0.02 ● ● ● ● ● CHF ● 0.02 ● 0.00 0.02 0.01 ● ● CHF ● ● ● ● ● ● ● ● 0.00 ● ● ● ● ● 0.01 ● ● ● 0.00 ● ● ● CHF ● 0.03 0.03 ● ● ● ● 0.0 0.5 1.0 Theta (d) 1.5 0.0 0.5 1.0 Theta (e) 1.5 0.0 0.5 1.0 1.5 Theta (f) Fig 3. Example 6.3. Analysis of the paired exchange rate returns: CHF/DEM, CHF/GBP, CHF/CAD with respect to USD between 1990-01-01 to 1998-12-31. (a)–(c): Scatter plots of the standardized paired exchange rate returns; (d)–(f ): Estimated angular densities using the estimated thresholds chosen. essentially a change point detection method based on p-values generated through subsampling schemes. Hence this may be generalized to other problem settings and potentially incorporates other dependence measures. Though we have proposed an automatic selection for the threshold based on the fitted mean p-value path, we would like to emphasize that, like the Hill plot, this should be viewed as a visual tool rather than an optimal selection criterion. The final threshold should be based on the automatic procedure in conjunction with visual inspection of the p-value path. We note that the choice of norm in the polar coordinate transformation (2.1) may result in significant differences in the choice of thresholds, which indicates the rate of convergence to the limit spectral density. This is especially evident in the near ‘asymptotic independence’ case, where the mass of the angular distribution concentrates on the axes. As an illustration, we simulated iid observations {(Xi1 , Xi2 )}i=1,...,n from the bivariate logistic distribution, where the cdf is given in (4.10), with γ = 0.95 and n = 10000. We apply the polar coordinate transformation with respect to the Lp -norm for p = 0.2, 1, 5. Note that in the case of p = 0.2, Lp is only a quasi-norm as it does not satisfy the triangular inequality. However, it can be shown that (2.4) holds and the limiting angular distribution exists for bivariate logistic distribution. We compare the threshold selection results in Figure 6. Note that in the cases of the L1 and L5 -norms, the threshold levels are chosen to be upper 5% and 12%, respectively, while in the case of the L0.2 -norm, it is not possible to select the threshold as the dependence between R and Θ at all levels were shown to be significant. Indeed, this can be seen in Figure 7, where we compare the histogram of X1p /(X1p + X2p ) given kXkp is large across three levels of truncations, 2%, 5% and 12%, together with the theoretical limiting density curve. For the L0.2 -norm, the limiting angular density is poorly approximated by the truncated data for all levels. For the other two norms, the truncated Threshold Selection for Multivariate Heavy-Tailed Data 15 1.0 GBP/CHF 0.6 0.4 0.0 0.2 p−value 0.8 mean−pv change point fitted spline ● 0.00 0.05 0.10 0.15 0.20 0.25 0.30 upper quantile (a) 1.0 CAD/CHF 0.6 0.4 0.0 0.2 p−value 0.8 mean−pv change point fitted spline ● 0.00 0.05 0.10 0.15 0.20 0.25 0.30 upper quantile (b) 1.0 DEM/CHF 0.6 0.4 0.0 0.2 p−value 0.8 mean−pv change point fitted spline ● 0.00 0.05 0.10 0.15 0.20 0.25 0.30 upper quantile (c) Fig 4. Example 6.3 (cont.). Analysis of the paired exchange rate returns: CHF/DEM, CHF/GBP, CHF/CAD with respect to USD between 1990-01-01 to 1998-12-31. (a)–(c): mean p-value paths (black triangles), fitted WBS splines (blue lines) and the chosen threshold quantiles (red vertical line). ● ● ● ● ● ● 0 2000 4000 6000 ● ● ● ● ●● ● ● ● 8000 ● ● ● ● ● ●● ● ● ● ● ● 1e−04 1e−02 1e+00 1e+02 1e+04 ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ●● ● ●● ● ● ● ● ●●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ●●● ●● ● ● ●● ● ●● ● ●● ● ● ●●● ●● ● ●● ● ● ● ●●●● ● ● ● ● ●●● ● ●● ●● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ●● ● ● ●● ● ● ● ●●● ●●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●●● ● ● ●● ●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ●●●● ● ● ● ● ● ●●● ●● ● ●● ● ●● ●● ● ● ●● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ●● ●●●● ● ● ● ●● ●●● ●● ● ● ●● ● ● ●● ●● ● ●●●●●● ● ●●● ● ● ● ●● ●●● ● ●●● ● ● ●● ●●●● ● ● ●●● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ●● ●● ● ● ●● ● ● ●● ●● ● ●● ● ● ● ● ● ●● ●●●●●● ● ● ●● ● ● ● ●● ●● ●● ●● ●●● ● ●● ● ● ●●●● ●●●● ● ●● ● ●● ● ●● ●● ● ● ●● ● ●●● ●●● ●● ● ● ●●●● ● ●● ● ●● ● ● ●●● ● ● ●● ●● ● ● ● ●● ● ● ●● ● ●● ●●● ● ●●● ● ● ●● ●● ●● ● ●● ● ●●● ● ● ●● ●●● ● ● ●● ● ●● ● ● ● ● ● ● ●●●●● ●● ● ●● ●● ●●●●● ● ● ● ●● ●● ● ●● ●●●●● ● ●●●● ● ● ● ●● ● ●●● ●● ● ● ●●● ● ● ● ●● ●● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ●●● ● ● ●● ●●●● ● ●● ● ● ●● ● ●●● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●●● ● ●● ● ● ●● ●●●● ● ● ● ●● ● ● ● ● ●●● ● ● ●●●●● ● ● ●●● ●● ● ●●●● ●● ●● ● ● ●● ● ●● ● ●● ● ●● ● ● ●● ● ●● ● ●● ● ●●● ● ● ● ● ● ● ● ● ●● ●●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ●●●● ● ●● ● ● ● ● ●● ●● ● ●●● ●● ●●● ●● ● ● ● ● ● ● ●● ● ● ● ●●● ●● ●● ● ● ● ●● ●● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ●● ●● ● ●● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ●●●● ●●●● ● ● ● ● ●● ● ●●● ● ● ● ● ● ●● ●● ● ● ●● ●●● ●● ●● ●● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ●●●● ● ● ● ●● ●● ● ● ● ● ● ● ● ●●●● ●●●● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●●●● ●● ● ● ●● ●● ●●● ● ● ● ● ● ● ● ●● ● ●●● ●● ●● ●● ● ● ● ●● ● ●● ● ●● ● ●●● ● ● ●● ● ● ● ● ●●● ● ● ●● ● ● ●● ● ● ●● ● ●● ●● ●● ●● ● ● ●● ● ●● ● ● ● ● ●●● ● ●● ●● ●● ● ● ●● ● ● ●● ● ● ●●● ●● ● ● ● ●● ●●●●● ●● ●● ● ●●●●● ●●●●● ● ●● ● ●●● ● ●●●● ●● ● ● ● ●● ●●● ● ● ● ●● ● ● ●●● ● ● ● ● ● ●●● ● ●● ● ●● ●●● ● ●● ● ● ●● ●● ●● ● ● ●●● ●● ● ●●●● ● ●●● ● ● ●● ●● ● ● ● ● ● ● ●●● ● ● ●●●● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ●●●● ● ●●● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ●● ● ●● ● ●●●● ●● ● ● ● ●● ●● ● ●● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ●●● ● ● ● ● ●●●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ●● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ●● ● ● ●● ● ●● ● ●● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ●● ●●● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●●● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ●● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ●● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ●●● ● ● ● ● ● ● ● ●●● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ●● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● 00 x1 ● 02 04 06 08 10 ang e (b) (c) 10 (a) ● ● ● ● ● ● x1 ● ● ● ad us ● ● ● ● 1e−04 ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● 1 ● ●● ● ● ● ● ●● 100 1e+00 x2 5000 x2 ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ●●●● ●●● ● ● ●● ● ● ●●●● ● ● ●● ● ● ● ●● ● ● ● ● ●●● ● ●● ● ●● ●● ●● ● ●● ● ● ● ●●●● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●●● ● ● ●● ● ● ● ●● ● ●●● ●● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●●● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ● ●● ●● ●●● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ●●● ● ● ● ●● ●●● ●● ●●●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ●● ● ● ●● ●● ● ●●● ●● ●● ●● ● ●●●● ● ●● ●●● ●● ● ● ●● ● ● ● ●● ● ●● 0 ● ●● ● ● ● 10 ● 15000 ● 10000 Wan and Davis 1e+04 16 06 04 00 02 p−va ue 08 m 0 00 0 05 0 10 0 15 0 20 uppe quan e (d) F g 5 Examp e 6 1 (a) sca erp o o (X 1 X 2 ) (b) sca erp o o (X 1 X 2 ) n og- og sca e (c) sca erp o o (R Θ ) (d) mean p-va ue pa h (b ack r ang es) fi ed WBS sp ne (b ue ne) and he chosen hresho d quan e (red ver ca ne) observat ons accord ng to the se ected thresho d prov de decent approx mat ons to the true m t ng dens ty of the angu ar component One poss b e exp anat on for th s s that under the L0 2 -norm the thresho d s concave and hence observat ons on the d agona are much eas er to be c ass fied as “extremes” than those near the ax s As a resu t the est mator of the angu ar dens ty uses more observat ons near the d agona wh ch may not be n fact c ose enough to the m t Th s cho ce of norm s an nterest ng top c and s the subject of ongo ng research Acknowledgement The fore gn exchange rate data are obta ned from OANDA from through R-package qrmtoo s We wou d ke to thank Bodh sattva Sen for he pfu d scuss ons We wou d a so ke to thank the ed tor and referees for the r many construct ve and ns ghtfu comments References P J B cke and M J W chura Convergence cr ter a for mu t parameter stochast c processes and some app cat ons Ann. Statist. 42 1656–1670 1971 P B ngs ey Probabi ity and measure. W ey-Intersc ence 3rd ed t on 1995 R A Dav s and T M kosch The extremogram A corre ogram for extreme events Bernou i 15(4) 977 – 1009 2009 Threshold Selection for Multivariate Heavy-Tailed Data 17 1.0 L0.2−norm 0.6 0.4 0.0 0.2 p−value 0.8 mean−pv change point fitted spline ● 0.00 0.05 0.10 0.15 0.20 upper quantile 1.0 L1−norm 0.6 0.4 0.0 0.2 p−value 0.8 mean−pv change point fitted spline ● 0.00 0.05 0.10 0.15 0.20 upper quantile 1.0 L5−norm 0.6 0.4 0.0 0.2 p−value 0.8 mean−pv change point fitted spline ● 0.00 0.05 0.10 0.15 0.20 upper quantile Fig 6. Simulated logistic data of sample size n = 10000 with γ = 0.95. Threshold selection algorithm applied under the L0.2 -, L1 - and L5 -norms: mean p-value paths (black triangles), fitted WBS splines (blue lines) and the chosen threshold quantiles (red vertical line). R.A. Davis, T. Mikosch, and I. Cribben. Towards estimating extremal serial dependence via the bootstrapped extremogram. J. Econometrics, 170(1):142–152, 2012. R.A. Davis, M. Matsui, T. Mikosch, and P. Wan. Applications of distance covariance to time series. arXiv:1606.05481, 2017. L. de Haan and J. de Ronde. Sea and wind: Multivariate extremes at work. Extremes, 1:7–46, 1998. P. Doukhan. Mixing: Properties and Examples. Springer-Verlag, New York, 1994. A. Feuerverger. A consistent test for bivariate dependence. Internat. Statis. Rev., 61(3):419–433, 1993. P. Fryzlewicz. Wild binary segmentation for multiple change-point detection. Ann. Statist., 42(6):2243–2281, 2014. S. Jeon and R.L. Smith. Dependence structure of spatial extremes using threshold approach. arXiv:1209.6344v1, 2014. H. Joe, R.L. Smith, and I. Weissman. Bivariate threshold methods for extremes. JRSS. B., 54(1):171–183, 1992. A. Mallik, B. Sen, M. Banerjee, and G. Michailidis. Threshold estimation based on a p-value framework in dose-response and regression settings. Biometrika, 98(4):887–900, 2011. E.S. Page. Continuous inspection schemes. Biometrika, 41(1):100–115, 1954. S.I. Resnick. Hidden regular variation, second order regular variation and asymptotic independence. Ex- 18 Wan and Davis L0.2−norm, 95% L0.2−norm, 88% 2.5 0.4 0.6 0.8 1.0 2.0 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 Theta Theta Theta L1−norm, 98% L1−norm, 95% L1−norm, 88% 0.8 1.0 0.8 1.0 0.8 1.0 3 4 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 Theta Theta Theta L5−norm, 98% L5−norm, 95% L5−norm, 88% 0.2 0.4 0.6 0.8 1.0 Density 5 0 5 0 0.0 10 15 20 15 Density 10 20 15 10 0 5 Density 20 25 25 0.0 0 0 0 2 1 2 4 2 4 Density 6 Density 8 6 Density 10 8 5 12 10 14 6 0.2 0.0 0.5 0.5 0.0 0.5 0.0 0.0 1.5 Density 1.5 Density 1.0 1.5 1.0 Density 2.0 2.0 2.5 2.5 L0.2−norm, 98% 0.0 0.2 Theta 0.4 0.6 0.8 1.0 0.0 Theta 0.2 0.4 0.6 Theta Fig 7. Simulated logistic data of sample size n = 10000 with γ = 0.95. Histogram of X1p /(X1p + X2p ) for truncated levels 2%, 5% and 12% for p = 0.2, 1, 5. tremes, 5:303–336, 2002. S.I. Resnick. Heavy-tail phenomena: probabilistic and statistical modeling. Springer-Verlag, New York, 2007. C. Stǎricǎ. Multivariate extremes for models with constant conditional correlations. J. Empir. Finan., 6: 515–553, 1999. G.J. Székely, M.L. Rizzo, and N.K. Bakirov. Measuring and testing dependence by correlation of distances. Ann. Statist., 35:2769–2794, 2007. Appendix A: Proof of Theorem 4.1 Note from the definition of the empirical distance covariance in (4.1), the integrand can be expressed as Cn (s, t) = n 1 X isRj /rn +itT Θj e 1{Rj >rn } np̂n j=1 − n n 1 X isRj /rn 1 X itT Θk e 1{Rj >rn } e 1{Rk >rn } np̂n j=1 np̂n k=1 = n T 1 X isRj /rn e − ϕ R \|rn (s) eit Θj − ϕΘ\|rn (t) 1{Rj >rn } rn np̂n j=1 − n n 1 X isRj /rn 1 X itT Θk e − ϕ R \|rn (s) 1{Rj >rn } e − ϕΘ\|rn (t) 1{Rk >rn } . rn np̂n j=1 np̂n k=1 Threshold Selection for Multivariate Heavy-Tailed Data T Writing Ujn = eisRj /rn − ϕ R \|rn (s) 1{Rj >rn } , Vjn = eit Θj − ϕΘ\|rn (t) 1{Rj >rn } , we have 19 rn n pn 1 X Ujn Vjn Cn (s, t) = − p̂n n j=1 pn pn p̂n 2 n n 1 X Ujn 1 X Vkn . n j=1 pn n pn k=1 Since EUjn = EVjn = 0 and EUjn Vjn /pn = ϕ R ,Θ\|rn (s, t) − ϕ R \|rn (s)ϕΘ\|rn (t), it is convenient to mean rn rn correct the summands and obtain n n n p 2 1 X pn 1 X Ujn Vjn Ujn 1 X Vkn n − Cn (s, t) = − ϕ R ,Θ\|rn (s, t) − ϕ R \|rn (s)ϕΘ\|rn (t) rn rn p̂n n j=1 pn p̂n n j=1 pn n pn k=1 pn + (ϕ R ,Θ\|rn (s, t) − ϕ R \|rn (s)ϕΘ\|rn (t)) rn p̂n rn 2 pn pn pn =: Ẽ1 − Ẽ3 Ẽ21 Ẽ22 + p̂n p̂n p̂n 2 pn pn pn Ẽ1 − Ẽ3 =: Ẽ2 + p̂n p̂n p̂n Note that Ẽ1 , Ẽ21 , Ẽ22 are averages of iid zero-mean random variables and Ẽ3 is non-random. We first prove the second part of Theorem 4.1. The first part of Theorem 4.1 follows easily in a similar fashion. Proof of Theorem 4.1(2). In order to show (4.6), it suffices to establish that 2 Z Z pn d np̂n \|Ẽ1 \|2 µ(ds, dt) → \|Q(s, t)\|2 µ(ds, dt), p̂ d+1 d+1 n R R and Z np̂n Tn − np̂n Rd+1 pn p̂n 2 p \|Ẽ1 \|2 µ(ds, dt) → 0, where (A.2) can be implied by 2 2 Z Z pn pn p np̂n \|Ẽ2 \|2 µ(ds, dt) + np̂n \|Ẽ3 \|2 µ(ds, dt) → 0. p̂n p̂n Rd+1 Rd+1 (A.1) (A.2) (A.3) Notice that p̂n E −1 pn 2 n 1X =E n j=1 2 1{Rj >rn } −1 pn = 1 1{R1 >rn } E −1 n pn 2 ≤ 1 1 O(1) + O(1) → 0. npn n p Hence p̂n /pn → 1 and for (A.1) and (A.3), it is equivalent to prove that Z Z d npn \|Ẽ1 \|2 µ(ds, dt) → \|Q(s, t)\|2 µ(ds, dt) Rd+1 and Z Z 2 \|Ẽ2 \| µ(ds, dt) + npn npn (A.4) Rd+1 Rd+1 p \|Ẽ3 \|2 µ(ds, dt) → 0. (A.5) Rd+1 We will show the convergence (A.4) in Proposition A.1. By (4.5), Z npn \|Ẽ3 \|2 µ(ds, dt) → 0. Rd+1 So that (A.5) holds provided Z npn p \|Ẽ2 \|2 µ(ds, dt) → 0, Rd+1 which follows in a similar fashion as Proposition A.1. (A.6) 20 Wan and Davis Proposition A.1. Assume µ satisfies Z (1 ∧ \|s\|β )(1 ∧ \|t\|2 )µ(ds, dt) < ∞, Rd+1 and that npn → ∞ as n → ∞, then Z npn d \|Ẽ1 \|2 µ(ds, dt) → Z Rd+1 \|Q(s, t)\|2 µ(ds, dt), Rd+1 where Q is a centered Gaussian process with covariance function (4.7). Proof of Proposition A.1. We first show that √ d npn Ẽ1 → Q(s, t), on C(Rd+1 ) (A.7) √ which can be implied by the finite distributional convergence of npn Ẽ1 (s, t) and its tightness on C(Rd+1 ). Write n n 1 X 1 X Ujn Vjn √ √ npn Ẽ1 = √ Yjn , − pn (ϕ R ,Θ\|rn (s, t) − ϕ R \|rn (s)ϕΘ\|rn (t)) =: √ √ rn rn pn n j=1 n j=1 where Yjn ’s are iid random variables with mean 0. For fixed (s, t), note that Var(Y1n ) = E\|Y1n \|2 = E1{R1 >rn } E\|U1n V1n \|2 (1 + o(1)) = O(1) < ∞. pn pn On the other hand, any δ > 0, E\|Y1n \|2+δ = E\|U1n V1n \|2+δ 1+δ/2 pn (1 + o(1)) ≤ c E1{R1 >rn } 1+δ/2 pn (1 + o(1)) = O(p−δ/2 ) n Then we can apply the central limit theorem for triangular arrays by checking the Lyapounov condition (see, e.g., Billingsley (1995)) for the Yjn ’s: δ Pn − E\|Yjn \|2+δ O(npn 2 ) − δ2 = ) → 0. δ δ = O((npn ) P 2+δ 1+ 1+ 2 2 2 n Var(Y1n ) n Var Y jn j=1 j=1 It follows easily that for fixed (s, t), √ d npn Ẽ1 → Q(s, t). The finite-dimensional distribution can be obtained using the Cramér-Wold device and the covariance function can be verified through calculations. √ We now show the tightness of npn Ẽ1 . Note that T isRj /rn it Θj n R e − ϕ (s) e − ϕ (t) 1{Rj >rn } X Θ\|r \|r n 1 n rn Ẽ1 (s, t) = n j=1 pn − ϕ R ,Θ\|rn (s, t) − ϕ R \|rn (s)ϕΘ\|rn (t) rn rn   T n isRj /rn +it Θj X e 1 1 {Rj >rn } − ϕ R ,Θ\|rn (s, t) =  rn n j=1 pn   n isRj /rn X e 1 1 {Rj >rn } − − ϕ R \|rn (s) ϕΘ\|rn (t) rn n j=1 pn Threshold Selection for Multivariate Heavy-Tailed Data   T n X eit Θj 1{Rj >rn } 1 p̂ n − − ϕΘ\|rn (t)  ϕ R \|rn (s) rn n j=1 pn pn 21 =: Ẽ11 + Ẽ12 + Ẽ13 . √ √ √ Without loss of generality, we show the tightness for npn Ẽ11 and that of npn Ẽ12 and npn Ẽ13 follows from the same argument. First we introduce some notation following that from Bickel and Wichura (1971). Fix (s, t), (s0 , t0 ) ∈ Rd+1 where s < s0 and t < t0 . Let B be the subset of Rd+1 of the form B := ((s, t), (s0 , t0 )] = (s, s0 ] × d Y (tk , t0k ] ⊂ Rd+1 . k=1 For ease of notation, we suppress the dependence of B on (s, t), (s0 , t0 ). Define the increment of the stochastic process Ẽ11 on B to be n Ẽ11 (B) P X 1X X X (−1)d+1− j zj ··· n j=1 z =0,1 z =0,1 z =0,1 := 0 1 d 0 Ẽ11 (s + z0 (s − s), t1 + z1 (t01 − t1 ), . . . , td + zd (t0d − td )) . From a sufficient condition of Theorem 3 of Bickel and Wichura (1971), the tightness of if the following statement holds for any (s, t), (s0 , t0 ) and corresponding B, d Y √ \|tk − t0k \|β , E\| npn Ẽ11 (B)\|2 ≤ c\|s − s0 \|β √ npn Ẽ1 is implied for some β > 1. k=1 It follows that √ E npn Ẽ11 (B) = npn E X ··· z0 =0,1 2 X (−1)d+1− P j n d 1 X i(s+z0 (s0 −s))Rj /r Y i(tk +zk (t0k −tk ))Θjk 1{Rj >rn } e e n j=1 pn zj zd =0,1 k=1 " − X X ··· z0 =0,1 (−1) d+1− P j zj E e i(s+z0 (s0 −s))R/r d Y zd =0,1 k=1 n e i(tk +zk (t0k −tk ))Θk 1{R>rn } pn # 2 d Y 1{Rj >rn } 0 0 1 X isRj /rj = npn E (e − eis Rj /rj ) (eitk Θjk − eitk Θjk ) n j=1 pn k=1 " −E (e isR/r −e is0 R/r ) d Y (e k=1 0 = pn Var (eisR/rj − eis R/rj ) d Y (eitk Θk k=1  0 ≤ E  (eisR/r − eis R/r ) d Y 1{R>rn } −e ) pn ! it0k Θk 1{R>rn } ) −e pn  2 itk Θk 0 (eitk Θk − eitk Θk ) # 2 it0k Θk (A.8) R > rn  . k=1 From a Taylor series argument, 0 \|eix − eix \|2 ≤ c1 ∧ \|x − x0 \|2 ≤ c1 ∧ \|x − x0 \|β ≤ c\|x − x0 \|β , for any β ∈ (0, 2]. 22 Wan and Davis Hence for any β ∈ (1, 2 ∧ α), √ E 2 ≤ npn Ẽ11 (B) d Y 0 β c\|s − s \| " \|ti − t0i \|β E β (R/rn ) k=1 d Y c\|s − s0 \|β < d Y # β \|Θk \| \|R > rn k=1 \|ti − t0i \|β , k=1 since \|Θk \|β ’s are bounded and supn E[(R/rn )β \|R > rn ] < ∞ by the regular variation assumption. This proves the tightness. Define the bounded set Kδ = {(s, t)\| δ < \|s\| < 1/δ, δ < \|t\| < 1/δ}, for δ < .5. Then, using (A.7), we have from the continuous mapping theorem, Z Z d npn \|Ẽ1 \|2 µ(ds, dt) → \|Q(s, t)\|2 µ(ds, dt). Kδ (A.9) Kδ On the other hand, for any β < 2 ∧ α, we have √ E\| npn Ẽ1 \|2 2 n 1X = npn E n j=1 ≤ ≤ Ujn Vjn Ujn Vjn −E pn pn E\|Ujn Vjn − EUjn Vjn \|2 pn c E\|Ujn Vjn \|2 p n c E eisRj /rn − ϕ R \|rn (s) (A.10) 2 rn = T eit Θj 2 − ϕΘ\|rn (t) 1{Rj >rn } pn cE eisRj /rn − 1 2 2 + ϕ R \|rn (s) − 1 rn ≤ cE 1 ∧ sRj rn 2 +E 1∧ β +E 1∧ sRj rn 2 sRj rn sRj rn 1 ∧ \|s\| cE ≤ ≤ cE β Rj rn β +E R rn Θj 2 + ϕΘ\|rn (t) − 1 −1 2 1{Rj >rn } \| rRn \| rRn p n h i >1 1 ∧ \|tΘj \|2 + E 1 ∧ \|tΘj \|2 \| rRn > 1 1{Rj >rn } \| rRn p n 2 >1 1 ∧ \|t\| 1{Rj >rn } β ≤ T eit p n h i >1 1 ∧ \|tΘj \|2 + E 1 ∧ \|tΘj \|2 \| rRn > 1 1{Rj >rn } ≤ cE 1 ∧ β pn 1 ∧ \|s\|β (\|Rj /rn \|β + E[\|R/rn \|β \|R > rn ]) 1 ∧ \|t\|2 \|R > rn ≤ c(1 ∧ \|s\|β )(1 ∧ \|t\|2 ). Therefore for any > 0, " Z lim lim sup P npn δ→0 n→∞ # 2 \|Ẽ1 \| µ(ds, dt) > Kδc ≤ 1 lim lim sup δ→0 n→∞ Z Kδc √ E\| npn Ẽ1 \|2 µ(ds, dt) Threshold Selection for Multivariate Heavy-Tailed Data 23 1 lim lim sup δ→0 n→∞ ≤ Z c(1 ∧ \|s\|β )(1 ∧ \|t\|2 )µ(ds, dt) → 0 Kδc by Rthe dominated convergence theorem. This combined with (A.9) shows the convergence of npn to \|Q(s, t)\|2 µ(ds, dt), and hence completes the proof of the proposition. R \|Ẽ1 \|2 µ(ds, dt) Proof of Theorem 4.1(2) (cont.). Now it remains to show (A.6). Similar to the proof of Proposition A.1, we can show that √ d npn Ẽ21 → Q0 for a centered Gaussian process Q0 , and p Ẽ22 → 0. Hence √ npn Ẽ2 = √ p npn Ẽ21 Ẽ22 → 0. The argument then follows similarly from the continuous mapping theorem and bounding the tail integrals. Proof of Theorem 4.1(1). Similar to the proof of Theorem 4.1(2), it suffices to show that Z p \|Ẽi \|2 µ(ds, dt) → 0, i = 1, 2, 3. (A.11) The convergence (A.11) for i = 1, 2 follows trivially from the more general results (A.4) and (A.6) in the proof of Theorem 4.1(2). Hence it suffices to show Z \|Ẽ3 \|2 µ(ds, dt) → 0, (A.12) where we recall that Ẽ3 := ϕ R ,Θ\|rn (s, t) − ϕ R \|rn (s)ϕΘ\|rn (t) is non-random. h rn irn Let P R ,Θ\|rn (·) = P rRn , Θ ∈ ·\| rRn > 1 and P R \|rn , PΘ\|rn be the corresponding marginal measures. rn rn Then from (2.3), v P R ,Θ\|rn − P R \|rn PΘ\|rn → να × S − να × S = 0, rn rn and hence for fixed (s, t), Z Ẽ3 (s, t) = T eisT +it Θ (P R ,Θ\|rn − P R \|rn PΘ\|rn )(dT, dΘ) → 0. rn rn For any β < 2 ∧ α, using the same argument in (A.10), \|Ẽ3 \|2 = E\|Ujn Vjn \| pn 2 ≤ c(1 ∧ \|s\|β )(1 ∧ \|t\|2 ). Then (A.12) follows from (4.4) and dominated convergence. This concludes the proof. Appendix B: Proof of Theorem 4.5 Following the same notation and steps as the proof of Theorem 4.1 in Appendix A, it suffices to prove the following convergences for the mixing case: p̂n p → 1, (B.1) pn Z Z d npn \|Ẽ1 \|2 µ(ds, dt) → \|Q0 (s, t)\|2 µ(ds, dt) (B.2) Rd+1 Rd+1 24 Wan and Davis and Z npn p \|Ẽ2 \|2 µ(ds, dt) → 0. (B.3) Rd+1 We prove (B.1) and (B.2) in Propositions B.2 and B.3, respectively. The proof of (B.3) follows in a similar fashion. The proofs of both propositions rely on the following lemma. Throughout this proof we make use of the results that if {Zt } is stationary and α-mixing with coefficient {αh }, then 1−δ , for any δ ∈ (0, 1), (B.4) \|cov(Z0 , Zh )\| ≤ cαhδ E\|Z0 \|2/(1−δ) see Section 1.2.2, Theorem 3(a) of Doukhan (1994). Lemma B.1. Let {Xt } be a multivariate stationary time series that is regularly varying and α-mixing with mixing coefficient {αh }. For a sequence rn → ∞, set pn = P(kX0 k > rn ). Let f1 , f2 be bounded functions d which vanish outside R \B1 (0), where B1 (0) is the unit open ball {x\|kxk < 1}, with sets of discontinuity of measure zero. Set, n X X0 Xt (i) Sn = − Efi , i = 1, 2. fi rn rn t=1 Assume that condition (M) holds for {αh } and {rn }. Then √ 1 d (S (1) , Sn(2) )T → N (0, Σ), npn n (B.5) where the covariance matrix [Σij ]i,j=1,2 = [σ 2 (fi , fj )]i,j=1,2 with σ 2 (f1 , f2 ) := σ02 (f1 , f2 ) + 2 ∞ X σh2 (f1 , f2 ) (B.6) h ≥ 0. (B.7) h=1 and σh2 (f1 , f2 ) = Z f1 f2 dµh , In particular, 1 p (Sn(1) , Sn(2) )T → 0. npn The proof of Lemma B.1 is provided after the proofs of the propositions. Proposition B.2. Assume that condition (M) holds, then p̂n p → 1, pn Proof. We have n p̂n 1X −1= pn n j=1 n 1{Rj >rn } 1 X −1 = (1{Rj >rn } − pn ). pn npn j=1 Apply Lemma B.1 to f (x) = 1{kxk>1} and the result follows. Proposition B.3. Assume that condition (M) holds, and that µ and {rn } satisfies (4.4) and (4.5), respectively, then Z Z d npn \|Ẽ1 \|2 µ(ds, dt) → \|Q0 (s, t)\|2 µ(ds, dt), Rd+1 0 where Q is a centered Gaussian process. Rd+1 Threshold Selection for Multivariate Heavy-Tailed Data 25 √ Proof. Let us first establish the convergence of npn Ẽ1 (s, t) for fixed (s, t). Take n o f1 (x) = Re eiskxk − E[eiskxk \|kxk > 1] eitx/kxk − E[eitx/kxk \|kxk > 1] 1kxk>1 n o f2 (x) = Im eiskxk − E[eiskxk \|kxk > 1] eitx/kxk − E[eitx/kxk \|kxk > 1] 1kxk>1 . Then from Lemma B.1, √ 1 √ d (Sn(1) , Sn(2) )T = npn (Re{Ẽ1 (s, t)}, Im{Ẽ1 (s, t)}) → N (0, Σ), npn where the covariance structure Σ can be derived from (B.6) and (B.7). This implies that √ d npn Ẽ1 (s, t) → Q0 (s, t), where Q0 (s, t) is a zero-mean complex normal process with covariance matrix Σ11 + Σ22 and relation matrix Σ11 − Σ22 + i(Σ12 + Σ21 ). √ The finite-dimensional distributional convergence of np̂n Ẽ1 to a Q0 (s, t) can be generalized using the Cramér-Wold device and we omit the calculation of the covariance structure. The tightness condition for the functional convergence follows the same arguments in the proof of Proposition A.1 from Bickel and Wichura (1971), with equality (A.8) replaced by a variance calculation of the sum of α-mixing components using the inequality (B.4) and condition (A.9) is verified through the same argument. This completes the proof of Proposition B.3. of Lemma B.1. The proof follows from that of Theorem 3.2 in Davis and Mikosch (2009). Here we outline the sketch of the proof and detail only the parts that differ from their proof. By the vague convergence in (4.12), we have i) Z Z 1 1 X0 X0 2 E fi → fi dµ0 and E fi → fi2 dµ0 ; pn rn pn rn ii) 2 Z 1 X0 1 X0 1 X0 2 Var fi = E fi − pn E fi → fi2 dµ0 = σ02 (fi , fi ); pn rn pn rn pn rn iii) 1 X0 Xh cov fi , fj pn rn rn Z → fi fj dµh = σh2 (fi , fj ). (i) 1 Sn for i = 1, 2. Without loss of generality, we suppress Let us first consider the marginal convergence of √np n the dependency on i and set Xt X0 Ytn := f − Ef . rn rn Then 1 Var [Y1n ] → σ02 (f, f ) pn and 1 cov Y1n , Y(h+1)n → σh2 (f, f ). pn We also have the following two results for \|cov(Y1n , Y(h+1)n )\|: lim lim sup h→∞ n→∞ ln X 1 \|cov(Y1n , Y(j+1)n )\| pn j=h 2 ln ln X X 1 X0 Xj 1 X0 ≤ lim lim sup E f f + E f h→∞ n→∞ pn rn rn pn rn j=h j=h 26 Wan and Davis ln ln X X 2 c c ≤ lim lim sup E 1{kX0 k>rn } 1{kXj k>rn } + E1{kX0 k>rn } h→∞ n→∞ pn pn j=h j=h ln X c P (kX0 k > rn , kXj k > rn ) + cln pn ≤ lim lim sup h→∞ n→∞ pn j=h = 0 (B.8) from condition (4.14), and lim n→∞ ∞ X 1 \|cov(Y1n , Y(j+1)n )\| pn ∞ X 1 X0 Xj cov f ,f n→∞ pn rn rn j=ln 2/(1−δ) !1−δ ∞ X 1 δ X0 lim α E f n→∞ pn j rn ≤ lim j=ln ≤ ≤ j=ln ∞ X lim n→∞ ≤ lim n→∞ = j=ln ∞ X 2/(1−δ) 1−δ c δ αj E 1{kX0 k>rn } pn cαjδ p−δ n j=ln 0 (B.9) from condition (4.13). We apply the same technique of small/large blocks as used in Davis and Mikosch (2009). Let mn and ln be the sizes of big and small blocks, respectively, where ln mn n. Let Ikn = {(k − 1)mn + 1, . . . , kmn } and Jkn = {(k − 1)mn + 1, . . . , (k − 1)mn + ln }, k = 1, . . . , n/mn , be the index sets of big and small blocks respectively. Set I˜kn = Ikn \Jkn , i.e., I˜kn are the big blocks with the first ln observations removed. For simplicity, we set mn := 1/pn and assume that the number of big blocks n/mn = npn is integer-valued. The non-integer case can be generalized without additional difficulties. Denote X Sn (B) := Ytn , t∈B then n X Ytn = Sn (1 : n) = t=1 npn X Sn (Ikn ) = k=1 npn X Sn (I˜kn ) + k=1 npn X Sn (Jkn ). k=1 Let {S̃n (I˜kn )}k=1,...,npn be iid copies of S̃n (I˜1n ). To prove the convergence of show the following: √ 1 Sn (1 npn : n), it suffices to npn npn 1 X 1 X ˜ S̃ ( I ) and Sn (I˜kn ) has the same limiting distribution, √ √ n kn npn npn (B.10) npn 1 X p Sn (Jkn ) → 0, √ npn (B.11) k=1 k=1 k=1 and √ npn 1 X d S̃n (I˜kn ) → N (0, σ 2 (f, f )). npn (B.12) k=1 The statement (B.10) holds if npn αln → 0, as n → ∞. This follows from the same argument in equation (6.2) in Davis and Mikosch (2009). (B.13) Threshold Selection for Multivariate Heavy-Tailed Data 27 For condition (B.11), it suffices to show that ! npn X 1 Sn (Jkn ) → 0. Var npn k=1 Note that ! npn np n −1 X X 1 h Sn (Jkn ) ≤ Var(Sn (J1n )) + 2 Var )\|cov(Sn (J1n ), Sn (J(h+1)n ))\| =: P1 + P2 . (1 − npn npn k=1 h=1 We have lim sup P1 = n→∞   ln X lim sup Var  Yjn  n→∞ j=1   \|cov(Y , Y )\| Var(Y1n ) j 1n (j+1)n  ≤ lim sup ln pn  +2 (1 − ) pn l p n→∞ n n j=1 ≤ lim sup ln pn n→∞ lX n −1 h−1 X \|cov(Y1n , Y(j+1)n )\| Var(Y1n ) + lim lim sup 2ln pn h→∞ n→∞ pn pn j=1 + lim lim sup 2ln pn h→∞ n→∞ = lX n −1 \|cov(Y1n , Y(j+1)n )\| pn j=h 0 where the last step follows from dominated convergence and (B.8). And for the other term, P2 ≤ 2 np n −1 X X X \|cov(Ysn , Ytn )\| h=1 s∈J1n t∈J(h+1)n ≤ 2 np n −1 X h/pn X ln ∞ X ≤ 2ln pn k=1/pn −ln +1 ∞ X ≤ 2ln pn cov(Y1n , Y(k+1)n ) k=h/pn −ln +1 h=1 cov(Y1n , Y(k+1)n ) pn cov(Y1n , Y(k+1)n ) → 0. pn k=ln +1 Note that 1/pn = mn is the size of big blocks Ikn ’s and 1/pn − ln + 1 = mn − ln + 1 is the distance between consecutive small blocks (Jkn , J(k+1)n )’s. The last limit follows from (B.9). To finish the proof, we need to establish the central limit theorem in (B.12). Note the S̃n (I˜ln )’s are iid with ES̃n (I˜ln ) = 0. We now calculate its variance. Recall that 1/pn − ln is the size of I˜1n , the big block with small block removed. Then   1/pn −ln X Var S̃n (I˜1n ) = Var  Yjn  j=1 = ( 1 − ln )Var(Yjn ) + 2 pn 1/pn −ln −1 X  = ( (1/pn − ln − k)cov(Y1n , Y(k+1)n ) k=1 1 − ln )Var(Yjn ) + 2  pn h X k=1 + ln X k=h+1  1/pn −ln −1 + X k=ln +1  1 − ln + k 1 cov(Y1n , Y(k+1)n ) 1/pn pn 28 Wan and Davis := I0 + I1 + I2 + I3 . Here lim I0 = lim (1 − ln pn ) n→∞ n→∞ 1 Var(Yjn ) = σ02 (f, f ), pn and lim I1 = lim 2 n→∞ n→∞ h X h (1 − pn (ln + k)) k=1 X cov(Y1n , Y(k+1)n ) =2 σk2 (f, f ). pn k=1 We also have lim lim sup \|I2 \| ≤ lim lim sup h→∞ n→∞ h→∞ n→∞ ln X \|cov(Y1n , Y(k+1)n )\| =0 pn k=h+1 from (B.8), and lim \|I3 \| ≤ lim n→∞ n→∞ ∞ X \|cov(Y1n , Y(k+1)n )\| =0 pn k=ln from (B.9). Therefore ∞ X lim Var S̃n (I˜1n ) = lim I0 + lim lim I1 = σ02 (f, f ) + 2 σk2 (f, f ) =: σ 2 (f, f ) n→∞ n→∞ h→∞ n→∞ k=1 as defined. To show that this infinite sum converges, it suffices to show that ∞ X µh ({(x, x0 )\|kxk > 1, kx0 k > 1}) < ∞. h=1 This follows from (4.14) in condition (M), for if ∞ X µh ({(x, x0 )\|kxk > 1, kx0 k > 1}) = ∞, h=1 then lim sup n→∞ ln ln X 1 X P(kX0 k > rn , kXj k > rn ) ≥ lim inf P(kX0 k > rn , kXj k > rn \|kX0 k > rn ) n→∞ pn j=h j=h ≥ ∞ X µj ({(x, x0 )\|kxk > 1, kx0 k > 1}) = ∞, j=h which leads to a contradiction. To apply the central limit theorem, we verify the Lindeberg’s condition,   2 1/pn −ln h i X   ≤ E  Yjn  1{\|S̃n (I˜1n )\|>√npn }  E (S̃n (I˜1n ))2 1{\|S̃n (I˜1n )\|>√npn } j=1 h i ≤ E c(1/pn − ln )2 1{\|S̃n (I˜1n )\|>√npn } i 1 h √ ≤ c 2 P \|S̃n (I˜1n )\| > npn pn h i ˜ Var S̃ ( I ) n 1n 1 1 = O( 3 ) → 0. ≤ c 2 2 pn npn npn This completes the proof for the convergence of √ 1 Sn (1 npn : n). Threshold Selection for Multivariate Heavy-Tailed Data (1) 29 (2) 1 The joint convergence of √np (Sn , Sn )T follows from the same line of argument together with the n Crámer-Wold device. In particular, 1 cov Sn(i) , Sn(j) = σ 2 (fi , fj ), npn i, j = 1, 2. This completes the proof of the lemma. Remark B.4. Lemma B.1 itself is a more general result of independent interest. The result can be generalized d for functions fi defined on R \{0} with compact support. In this case, condition (4.14) should be modified to lim lim sup h→∞ n→∞ d ln 1 X P(kX0 k > rn , kXj k > rn ) = 0 pn j=h for some > 0, where support(f ) ⊆ R \B (0). Also, as seen during the proof of the lemma, the conditions on pn , ln , and αt can be further relaxed.	10math.ST

Subsets and Splits

Random Sampled ArXiv Labels

Provides a random sample of 100 rows, 20 from each of 5 labels, which is useful for a basic overview but does not reveal deep insights.

Random 20 Rows Per Class

Provides a basic sample of 20 rows per label, useful for a quick overview but does not offer deep insights.